memcpy.S 15 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529
  1. /* SPDX-License-Identifier: GPL-2.0-only */
  2. /*
  3. * Copyright (c) 2010-2011, The Linux Foundation. All rights reserved.
  4. */
  5. /*
  6. * Description
  7. *
  8. * library function for memcpy where length bytes are copied from
  9. * ptr_in to ptr_out. ptr_out is returned unchanged.
  10. * Allows any combination of alignment on input and output pointers
  11. * and length from 0 to 2^32-1
  12. *
  13. * Restrictions
  14. * The arrays should not overlap, the program will produce undefined output
  15. * if they do.
  16. * For blocks less than 16 bytes a byte by byte copy is performed. For
  17. * 8byte alignments, and length multiples, a dword copy is performed up to
  18. * 96bytes
  19. * History
  20. *
  21. * DJH 5/15/09 Initial version 1.0
  22. * DJH 6/ 1/09 Version 1.1 modified ABI to inlcude R16-R19
  23. * DJH 7/12/09 Version 1.2 optimized codesize down to 760 was 840
  24. * DJH 10/14/09 Version 1.3 added special loop for aligned case, was
  25. * overreading bloated codesize back up to 892
  26. * DJH 4/20/10 Version 1.4 fixed Ldword_loop_epilog loop to prevent loads
  27. * occurring if only 1 left outstanding, fixes bug
  28. * # 3888, corrected for all alignments. Peeled off
  29. * 1 32byte chunk from kernel loop and extended 8byte
  30. * loop at end to solve all combinations and prevent
  31. * over read. Fixed Ldword_loop_prolog to prevent
  32. * overread for blocks less than 48bytes. Reduced
  33. * codesize to 752 bytes
  34. * DJH 4/21/10 version 1.5 1.4 fix broke code for input block ends not
  35. * aligned to dword boundaries,underwriting by 1
  36. * byte, added detection for this and fixed. A
  37. * little bloat.
  38. * DJH 4/23/10 version 1.6 corrected stack error, R20 was not being restored
  39. * always, fixed the error of R20 being modified
  40. * before it was being saved
  41. * Natural c model
  42. * ===============
  43. * void * memcpy(char * ptr_out, char * ptr_in, int length) {
  44. * int i;
  45. * if(length) for(i=0; i < length; i++) { ptr_out[i] = ptr_in[i]; }
  46. * return(ptr_out);
  47. * }
  48. *
  49. * Optimized memcpy function
  50. * =========================
  51. * void * memcpy(char * ptr_out, char * ptr_in, int len) {
  52. * int i, prolog, kernel, epilog, mask;
  53. * u8 offset;
  54. * s64 data0, dataF8, data70;
  55. *
  56. * s64 * ptr8_in;
  57. * s64 * ptr8_out;
  58. * s32 * ptr4;
  59. * s16 * ptr2;
  60. *
  61. * offset = ((int) ptr_in) & 7;
  62. * ptr8_in = (s64 *) &ptr_in[-offset]; //read in the aligned pointers
  63. *
  64. * data70 = *ptr8_in++;
  65. * dataF8 = *ptr8_in++;
  66. *
  67. * data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
  68. *
  69. * prolog = 32 - ((int) ptr_out);
  70. * mask = 0x7fffffff >> HEXAGON_R_cl0_R(len);
  71. * prolog = prolog & mask;
  72. * kernel = len - prolog;
  73. * epilog = kernel & 0x1F;
  74. * kernel = kernel>>5;
  75. *
  76. * if (prolog & 1) { ptr_out[0] = (u8) data0; data0 >>= 8; ptr_out += 1;}
  77. * ptr2 = (s16 *) &ptr_out[0];
  78. * if (prolog & 2) { ptr2[0] = (u16) data0; data0 >>= 16; ptr_out += 2;}
  79. * ptr4 = (s32 *) &ptr_out[0];
  80. * if (prolog & 4) { ptr4[0] = (u32) data0; data0 >>= 32; ptr_out += 4;}
  81. *
  82. * offset = offset + (prolog & 7);
  83. * if (offset >= 8) {
  84. * data70 = dataF8;
  85. * dataF8 = *ptr8_in++;
  86. * }
  87. * offset = offset & 0x7;
  88. *
  89. * prolog = prolog >> 3;
  90. * if (prolog) for (i=0; i < prolog; i++) {
  91. * data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
  92. * ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
  93. * data70 = dataF8;
  94. * dataF8 = *ptr8_in++;
  95. * }
  96. * if(kernel) { kernel -= 1; epilog += 32; }
  97. * if(kernel) for(i=0; i < kernel; i++) {
  98. * data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
  99. * ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
  100. * data70 = *ptr8_in++;
  101. *
  102. * data0 = HEXAGON_P_valignb_PPp(data70, dataF8, offset);
  103. * ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
  104. * dataF8 = *ptr8_in++;
  105. *
  106. * data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
  107. * ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
  108. * data70 = *ptr8_in++;
  109. *
  110. * data0 = HEXAGON_P_valignb_PPp(data70, dataF8, offset);
  111. * ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
  112. * dataF8 = *ptr8_in++;
  113. * }
  114. * epilogdws = epilog >> 3;
  115. * if (epilogdws) for (i=0; i < epilogdws; i++) {
  116. * data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
  117. * ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
  118. * data70 = dataF8;
  119. * dataF8 = *ptr8_in++;
  120. * }
  121. * data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
  122. *
  123. * ptr4 = (s32 *) &ptr_out[0];
  124. * if (epilog & 4) { ptr4[0] = (u32) data0; data0 >>= 32; ptr_out += 4;}
  125. * ptr2 = (s16 *) &ptr_out[0];
  126. * if (epilog & 2) { ptr2[0] = (u16) data0; data0 >>= 16; ptr_out += 2;}
  127. * if (epilog & 1) { *ptr_out++ = (u8) data0; }
  128. *
  129. * return(ptr_out - length);
  130. * }
  131. *
  132. * Codesize : 784 bytes
  133. */
  134. #define ptr_out R0 /* destination pounter */
  135. #define ptr_in R1 /* source pointer */
  136. #define len R2 /* length of copy in bytes */
  137. #define data70 R13:12 /* lo 8 bytes of non-aligned transfer */
  138. #define dataF8 R11:10 /* hi 8 bytes of non-aligned transfer */
  139. #define ldata0 R7:6 /* even 8 bytes chunks */
  140. #define ldata1 R25:24 /* odd 8 bytes chunks */
  141. #define data1 R7 /* lower 8 bytes of ldata1 */
  142. #define data0 R6 /* lower 8 bytes of ldata0 */
  143. #define ifbyte p0 /* if transfer has bytes in epilog/prolog */
  144. #define ifhword p0 /* if transfer has shorts in epilog/prolog */
  145. #define ifword p0 /* if transfer has words in epilog/prolog */
  146. #define noprolog p0 /* no prolog, xfer starts at 32byte */
  147. #define nokernel p1 /* no 32byte multiple block in the transfer */
  148. #define noepilog p0 /* no epilog, xfer ends on 32byte boundary */
  149. #define align p2 /* alignment of input rel to 8byte boundary */
  150. #define kernel1 p0 /* kernel count == 1 */
  151. #define dalign R25 /* rel alignment of input to output data */
  152. #define star3 R16 /* number bytes in prolog - dwords */
  153. #define rest R8 /* length - prolog bytes */
  154. #define back R7 /* nr bytes > dword boundary in src block */
  155. #define epilog R3 /* bytes in epilog */
  156. #define inc R15:14 /* inc kernel by -1 and defetch ptr by 32 */
  157. #define kernel R4 /* number of 32byte chunks in kernel */
  158. #define ptr_in_p_128 R5 /* pointer for prefetch of input data */
  159. #define mask R8 /* mask used to determine prolog size */
  160. #define shift R8 /* used to work a shifter to extract bytes */
  161. #define shift2 R5 /* in epilog to workshifter to extract bytes */
  162. #define prolog R15 /* bytes in prolog */
  163. #define epilogdws R15 /* number dwords in epilog */
  164. #define shiftb R14 /* used to extract bytes */
  165. #define offset R9 /* same as align in reg */
  166. #define ptr_out_p_32 R17 /* pointer to output dczero */
  167. #define align888 R14 /* if simple dword loop can be used */
  168. #define len8 R9 /* number of dwords in length */
  169. #define over R20 /* nr of bytes > last inp buf dword boundary */
  170. #define ptr_in_p_128kernel R5:4 /* packed fetch pointer & kernel cnt */
  171. .section .text
  172. .p2align 4
  173. .global memcpy
  174. .type memcpy, @function
  175. memcpy:
  176. {
  177. p2 = cmp.eq(len, #0); /* =0 */
  178. align888 = or(ptr_in, ptr_out); /* %8 < 97 */
  179. p0 = cmp.gtu(len, #23); /* %1, <24 */
  180. p1 = cmp.eq(ptr_in, ptr_out); /* attempt to overwrite self */
  181. }
  182. {
  183. p1 = or(p2, p1);
  184. p3 = cmp.gtu(len, #95); /* %8 < 97 */
  185. align888 = or(align888, len); /* %8 < 97 */
  186. len8 = lsr(len, #3); /* %8 < 97 */
  187. }
  188. {
  189. dcfetch(ptr_in); /* zero/ptrin=ptrout causes fetch */
  190. p2 = bitsclr(align888, #7); /* %8 < 97 */
  191. if(p1) jumpr r31; /* =0 */
  192. }
  193. {
  194. p2 = and(p2,!p3); /* %8 < 97 */
  195. if (p2.new) len = add(len, #-8); /* %8 < 97 */
  196. if (p2.new) jump:NT .Ldwordaligned; /* %8 < 97 */
  197. }
  198. {
  199. if(!p0) jump .Lbytes23orless; /* %1, <24 */
  200. mask.l = #LO(0x7fffffff);
  201. /* all bytes before line multiples of data */
  202. prolog = sub(#0, ptr_out);
  203. }
  204. {
  205. /* save r31 on stack, decrement sp by 16 */
  206. allocframe(#24);
  207. mask.h = #HI(0x7fffffff);
  208. ptr_in_p_128 = add(ptr_in, #32);
  209. back = cl0(len);
  210. }
  211. {
  212. memd(sp+#0) = R17:16; /* save r16,r17 on stack6 */
  213. r31.l = #LO(.Lmemcpy_return); /* set up final return pointer */
  214. prolog &= lsr(mask, back);
  215. offset = and(ptr_in, #7);
  216. }
  217. {
  218. memd(sp+#8) = R25:24; /* save r25,r24 on stack */
  219. dalign = sub(ptr_out, ptr_in);
  220. r31.h = #HI(.Lmemcpy_return); /* set up final return pointer */
  221. }
  222. {
  223. /* see if there if input buffer end if aligned */
  224. over = add(len, ptr_in);
  225. back = add(len, offset);
  226. memd(sp+#16) = R21:20; /* save r20,r21 on stack */
  227. }
  228. {
  229. noprolog = bitsclr(prolog, #7);
  230. prolog = and(prolog, #31);
  231. dcfetch(ptr_in_p_128);
  232. ptr_in_p_128 = add(ptr_in_p_128, #32);
  233. }
  234. {
  235. kernel = sub(len, prolog);
  236. shift = asl(prolog, #3);
  237. star3 = and(prolog, #7);
  238. ptr_in = and(ptr_in, #-8);
  239. }
  240. {
  241. prolog = lsr(prolog, #3);
  242. epilog = and(kernel, #31);
  243. ptr_out_p_32 = add(ptr_out, prolog);
  244. over = and(over, #7);
  245. }
  246. {
  247. p3 = cmp.gtu(back, #8);
  248. kernel = lsr(kernel, #5);
  249. dcfetch(ptr_in_p_128);
  250. ptr_in_p_128 = add(ptr_in_p_128, #32);
  251. }
  252. {
  253. p1 = cmp.eq(prolog, #0);
  254. if(!p1.new) prolog = add(prolog, #1);
  255. dcfetch(ptr_in_p_128); /* reserve the line 64bytes on */
  256. ptr_in_p_128 = add(ptr_in_p_128, #32);
  257. }
  258. {
  259. nokernel = cmp.eq(kernel,#0);
  260. dcfetch(ptr_in_p_128); /* reserve the line 64bytes on */
  261. ptr_in_p_128 = add(ptr_in_p_128, #32);
  262. shiftb = and(shift, #8);
  263. }
  264. {
  265. dcfetch(ptr_in_p_128); /* reserve the line 64bytes on */
  266. ptr_in_p_128 = add(ptr_in_p_128, #32);
  267. if(nokernel) jump .Lskip64;
  268. p2 = cmp.eq(kernel, #1); /* skip ovr if kernel == 0 */
  269. }
  270. {
  271. dczeroa(ptr_out_p_32);
  272. /* don't advance pointer */
  273. if(!p2) ptr_out_p_32 = add(ptr_out_p_32, #32);
  274. }
  275. {
  276. dalign = and(dalign, #31);
  277. dczeroa(ptr_out_p_32);
  278. }
  279. .Lskip64:
  280. {
  281. data70 = memd(ptr_in++#16);
  282. if(p3) dataF8 = memd(ptr_in+#8);
  283. if(noprolog) jump .Lnoprolog32;
  284. align = offset;
  285. }
  286. /* upto initial 7 bytes */
  287. {
  288. ldata0 = valignb(dataF8, data70, align);
  289. ifbyte = tstbit(shift,#3);
  290. offset = add(offset, star3);
  291. }
  292. {
  293. if(ifbyte) memb(ptr_out++#1) = data0;
  294. ldata0 = lsr(ldata0, shiftb);
  295. shiftb = and(shift, #16);
  296. ifhword = tstbit(shift,#4);
  297. }
  298. {
  299. if(ifhword) memh(ptr_out++#2) = data0;
  300. ldata0 = lsr(ldata0, shiftb);
  301. ifword = tstbit(shift,#5);
  302. p2 = cmp.gtu(offset, #7);
  303. }
  304. {
  305. if(ifword) memw(ptr_out++#4) = data0;
  306. if(p2) data70 = dataF8;
  307. if(p2) dataF8 = memd(ptr_in++#8); /* another 8 bytes */
  308. align = offset;
  309. }
  310. .Lnoprolog32:
  311. {
  312. p3 = sp1loop0(.Ldword_loop_prolog, prolog)
  313. rest = sub(len, star3); /* whats left after the loop */
  314. p0 = cmp.gt(over, #0);
  315. }
  316. if(p0) rest = add(rest, #16);
  317. .Ldword_loop_prolog:
  318. {
  319. if(p3) memd(ptr_out++#8) = ldata0;
  320. ldata0 = valignb(dataF8, data70, align);
  321. p0 = cmp.gt(rest, #16);
  322. }
  323. {
  324. data70 = dataF8;
  325. if(p0) dataF8 = memd(ptr_in++#8);
  326. rest = add(rest, #-8);
  327. }:endloop0
  328. .Lkernel:
  329. {
  330. /* kernel is at least 32bytes */
  331. p3 = cmp.gtu(kernel, #0);
  332. /* last itn. remove edge effects */
  333. if(p3.new) kernel = add(kernel, #-1);
  334. /* dealt with in last dword loop */
  335. if(p3.new) epilog = add(epilog, #32);
  336. }
  337. {
  338. nokernel = cmp.eq(kernel, #0); /* after adjustment, recheck */
  339. if(nokernel.new) jump:NT .Lepilog; /* likely not taken */
  340. inc = combine(#32, #-1);
  341. p3 = cmp.gtu(dalign, #24);
  342. }
  343. {
  344. if(p3) jump .Lodd_alignment;
  345. }
  346. {
  347. loop0(.Loword_loop_25to31, kernel);
  348. kernel1 = cmp.gtu(kernel, #1);
  349. rest = kernel;
  350. }
  351. .falign
  352. .Loword_loop_25to31:
  353. {
  354. dcfetch(ptr_in_p_128); /* prefetch 4 lines ahead */
  355. if(kernel1) ptr_out_p_32 = add(ptr_out_p_32, #32);
  356. }
  357. {
  358. dczeroa(ptr_out_p_32); /* reserve the next 32bytes in cache */
  359. p3 = cmp.eq(kernel, rest);
  360. }
  361. {
  362. /* kernel -= 1 */
  363. ptr_in_p_128kernel = vaddw(ptr_in_p_128kernel, inc);
  364. /* kill write on first iteration */
  365. if(!p3) memd(ptr_out++#8) = ldata1;
  366. ldata1 = valignb(dataF8, data70, align);
  367. data70 = memd(ptr_in++#8);
  368. }
  369. {
  370. memd(ptr_out++#8) = ldata0;
  371. ldata0 = valignb(data70, dataF8, align);
  372. dataF8 = memd(ptr_in++#8);
  373. }
  374. {
  375. memd(ptr_out++#8) = ldata1;
  376. ldata1 = valignb(dataF8, data70, align);
  377. data70 = memd(ptr_in++#8);
  378. }
  379. {
  380. memd(ptr_out++#8) = ldata0;
  381. ldata0 = valignb(data70, dataF8, align);
  382. dataF8 = memd(ptr_in++#8);
  383. kernel1 = cmp.gtu(kernel, #1);
  384. }:endloop0
  385. {
  386. memd(ptr_out++#8) = ldata1;
  387. jump .Lepilog;
  388. }
  389. .Lodd_alignment:
  390. {
  391. loop0(.Loword_loop_00to24, kernel);
  392. kernel1 = cmp.gtu(kernel, #1);
  393. rest = add(kernel, #-1);
  394. }
  395. .falign
  396. .Loword_loop_00to24:
  397. {
  398. dcfetch(ptr_in_p_128); /* prefetch 4 lines ahead */
  399. ptr_in_p_128kernel = vaddw(ptr_in_p_128kernel, inc);
  400. if(kernel1) ptr_out_p_32 = add(ptr_out_p_32, #32);
  401. }
  402. {
  403. dczeroa(ptr_out_p_32); /* reserve the next 32bytes in cache */
  404. }
  405. {
  406. memd(ptr_out++#8) = ldata0;
  407. ldata0 = valignb(dataF8, data70, align);
  408. data70 = memd(ptr_in++#8);
  409. }
  410. {
  411. memd(ptr_out++#8) = ldata0;
  412. ldata0 = valignb(data70, dataF8, align);
  413. dataF8 = memd(ptr_in++#8);
  414. }
  415. {
  416. memd(ptr_out++#8) = ldata0;
  417. ldata0 = valignb(dataF8, data70, align);
  418. data70 = memd(ptr_in++#8);
  419. }
  420. {
  421. memd(ptr_out++#8) = ldata0;
  422. ldata0 = valignb(data70, dataF8, align);
  423. dataF8 = memd(ptr_in++#8);
  424. kernel1 = cmp.gtu(kernel, #1);
  425. }:endloop0
  426. .Lepilog:
  427. {
  428. noepilog = cmp.eq(epilog,#0);
  429. epilogdws = lsr(epilog, #3);
  430. kernel = and(epilog, #7);
  431. }
  432. {
  433. if(noepilog) jumpr r31;
  434. if(noepilog) ptr_out = sub(ptr_out, len);
  435. p3 = cmp.eq(epilogdws, #0);
  436. shift2 = asl(epilog, #3);
  437. }
  438. {
  439. shiftb = and(shift2, #32);
  440. ifword = tstbit(epilog,#2);
  441. if(p3) jump .Lepilog60;
  442. if(!p3) epilog = add(epilog, #-16);
  443. }
  444. {
  445. loop0(.Ldword_loop_epilog, epilogdws);
  446. /* stop criteria is lsbs unless = 0 then its 8 */
  447. p3 = cmp.eq(kernel, #0);
  448. if(p3.new) kernel= #8;
  449. p1 = cmp.gt(over, #0);
  450. }
  451. /* if not aligned to end of buffer execute 1 more iteration */
  452. if(p1) kernel= #0;
  453. .Ldword_loop_epilog:
  454. {
  455. memd(ptr_out++#8) = ldata0;
  456. ldata0 = valignb(dataF8, data70, align);
  457. p3 = cmp.gt(epilog, kernel);
  458. }
  459. {
  460. data70 = dataF8;
  461. if(p3) dataF8 = memd(ptr_in++#8);
  462. epilog = add(epilog, #-8);
  463. }:endloop0
  464. /* copy last 7 bytes */
  465. .Lepilog60:
  466. {
  467. if(ifword) memw(ptr_out++#4) = data0;
  468. ldata0 = lsr(ldata0, shiftb);
  469. ifhword = tstbit(epilog,#1);
  470. shiftb = and(shift2, #16);
  471. }
  472. {
  473. if(ifhword) memh(ptr_out++#2) = data0;
  474. ldata0 = lsr(ldata0, shiftb);
  475. ifbyte = tstbit(epilog,#0);
  476. if(ifbyte.new) len = add(len, #-1);
  477. }
  478. {
  479. if(ifbyte) memb(ptr_out) = data0;
  480. ptr_out = sub(ptr_out, len); /* return dest pointer */
  481. jumpr r31;
  482. }
  483. /* do byte copy for small n */
  484. .Lbytes23orless:
  485. {
  486. p3 = sp1loop0(.Lbyte_copy, len);
  487. len = add(len, #-1);
  488. }
  489. .Lbyte_copy:
  490. {
  491. data0 = memb(ptr_in++#1);
  492. if(p3) memb(ptr_out++#1) = data0;
  493. }:endloop0
  494. {
  495. memb(ptr_out) = data0;
  496. ptr_out = sub(ptr_out, len);
  497. jumpr r31;
  498. }
  499. /* do dword copies for aligned in, out and length */
  500. .Ldwordaligned:
  501. {
  502. p3 = sp1loop0(.Ldword_copy, len8);
  503. }
  504. .Ldword_copy:
  505. {
  506. if(p3) memd(ptr_out++#8) = ldata0;
  507. ldata0 = memd(ptr_in++#8);
  508. }:endloop0
  509. {
  510. memd(ptr_out) = ldata0;
  511. ptr_out = sub(ptr_out, len);
  512. jumpr r31; /* return to function caller */
  513. }
  514. .Lmemcpy_return:
  515. r21:20 = memd(sp+#16); /* restore r20+r21 */
  516. {
  517. r25:24 = memd(sp+#8); /* restore r24+r25 */
  518. r17:16 = memd(sp+#0); /* restore r16+r17 */
  519. }
  520. deallocframe; /* restore r31 and incrment stack by 16 */
  521. jumpr r31