1 /* NG2memcpy.S: Niagara-2 optimized memcpy.
3 * Copyright (C) 2007 David S. Miller (davem@davemloft.net)
7 #include <linux/linkage.h>
8 #include <asm/visasm.h>
10 #define GLOBAL_SPARE %g7
13 #define ASI_BLK_P 0xf0
14 #define ASI_BLK_INIT_QUAD_LDD_P 0xe2
17 #define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs; \
18 clr %g1; clr %g2; clr %g3; clr %g5; subcc %g0, %g0, %g0;
19 #define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
21 #define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs
22 #define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
24 #define GLOBAL_SPARE %g5
28 #ifndef SIMULATE_NIAGARA_ON_NON_NIAGARA
29 #define STORE_ASI ASI_BLK_INIT_QUAD_LDD_P
31 #define STORE_ASI 0x80 /* ASI_P */
39 #define EX_LD_FP(x,y) x
46 #define EX_ST_FP(x,y) x
50 #define LOAD(type,addr,dest) type [addr], dest
54 #define LOAD_BLK(addr,dest) ldda [addr] ASI_BLK_P, dest
59 #define STORE(type,src,addr) type src, [addr]
61 #define STORE(type,src,addr) type##a src, [addr] 0x80
66 #define STORE_BLK(src,addr) stda src, [addr] ASI_BLK_P
70 #define STORE_INIT(src,addr) stxa src, [addr] STORE_ASI
74 #define FUNC_NAME NG2memcpy
85 #define FREG_FROB(x0, x1, x2, x3, x4, x5, x6, x7, x8) \
86 faligndata %x0, %x1, %f0; \
87 faligndata %x1, %x2, %f2; \
88 faligndata %x2, %x3, %f4; \
89 faligndata %x3, %x4, %f6; \
90 faligndata %x4, %x5, %f8; \
91 faligndata %x5, %x6, %f10; \
92 faligndata %x6, %x7, %f12; \
93 faligndata %x7, %x8, %f14;
95 #define FREG_MOVE_1(x0) \
97 #define FREG_MOVE_2(x0, x1) \
100 #define FREG_MOVE_3(x0, x1, x2) \
104 #define FREG_MOVE_4(x0, x1, x2, x3) \
109 #define FREG_MOVE_5(x0, x1, x2, x3, x4) \
115 #define FREG_MOVE_6(x0, x1, x2, x3, x4, x5) \
122 #define FREG_MOVE_7(x0, x1, x2, x3, x4, x5, x6) \
130 #define FREG_MOVE_8(x0, x1, x2, x3, x4, x5, x6, x7) \
139 #define FREG_LOAD_1(base, x0) \
140 EX_LD_FP(LOAD(ldd, base + 0x00, %x0), NG2_retl_o2_plus_g1)
141 #define FREG_LOAD_2(base, x0, x1) \
142 EX_LD_FP(LOAD(ldd, base + 0x00, %x0), NG2_retl_o2_plus_g1); \
143 EX_LD_FP(LOAD(ldd, base + 0x08, %x1), NG2_retl_o2_plus_g1);
144 #define FREG_LOAD_3(base, x0, x1, x2) \
145 EX_LD_FP(LOAD(ldd, base + 0x00, %x0), NG2_retl_o2_plus_g1); \
146 EX_LD_FP(LOAD(ldd, base + 0x08, %x1), NG2_retl_o2_plus_g1); \
147 EX_LD_FP(LOAD(ldd, base + 0x10, %x2), NG2_retl_o2_plus_g1);
148 #define FREG_LOAD_4(base, x0, x1, x2, x3) \
149 EX_LD_FP(LOAD(ldd, base + 0x00, %x0), NG2_retl_o2_plus_g1); \
150 EX_LD_FP(LOAD(ldd, base + 0x08, %x1), NG2_retl_o2_plus_g1); \
151 EX_LD_FP(LOAD(ldd, base + 0x10, %x2), NG2_retl_o2_plus_g1); \
152 EX_LD_FP(LOAD(ldd, base + 0x18, %x3), NG2_retl_o2_plus_g1);
153 #define FREG_LOAD_5(base, x0, x1, x2, x3, x4) \
154 EX_LD_FP(LOAD(ldd, base + 0x00, %x0), NG2_retl_o2_plus_g1); \
155 EX_LD_FP(LOAD(ldd, base + 0x08, %x1), NG2_retl_o2_plus_g1); \
156 EX_LD_FP(LOAD(ldd, base + 0x10, %x2), NG2_retl_o2_plus_g1); \
157 EX_LD_FP(LOAD(ldd, base + 0x18, %x3), NG2_retl_o2_plus_g1); \
158 EX_LD_FP(LOAD(ldd, base + 0x20, %x4), NG2_retl_o2_plus_g1);
159 #define FREG_LOAD_6(base, x0, x1, x2, x3, x4, x5) \
160 EX_LD_FP(LOAD(ldd, base + 0x00, %x0), NG2_retl_o2_plus_g1); \
161 EX_LD_FP(LOAD(ldd, base + 0x08, %x1), NG2_retl_o2_plus_g1); \
162 EX_LD_FP(LOAD(ldd, base + 0x10, %x2), NG2_retl_o2_plus_g1); \
163 EX_LD_FP(LOAD(ldd, base + 0x18, %x3), NG2_retl_o2_plus_g1); \
164 EX_LD_FP(LOAD(ldd, base + 0x20, %x4), NG2_retl_o2_plus_g1); \
165 EX_LD_FP(LOAD(ldd, base + 0x28, %x5), NG2_retl_o2_plus_g1);
166 #define FREG_LOAD_7(base, x0, x1, x2, x3, x4, x5, x6) \
167 EX_LD_FP(LOAD(ldd, base + 0x00, %x0), NG2_retl_o2_plus_g1); \
168 EX_LD_FP(LOAD(ldd, base + 0x08, %x1), NG2_retl_o2_plus_g1); \
169 EX_LD_FP(LOAD(ldd, base + 0x10, %x2), NG2_retl_o2_plus_g1); \
170 EX_LD_FP(LOAD(ldd, base + 0x18, %x3), NG2_retl_o2_plus_g1); \
171 EX_LD_FP(LOAD(ldd, base + 0x20, %x4), NG2_retl_o2_plus_g1); \
172 EX_LD_FP(LOAD(ldd, base + 0x28, %x5), NG2_retl_o2_plus_g1); \
173 EX_LD_FP(LOAD(ldd, base + 0x30, %x6), NG2_retl_o2_plus_g1);
175 .register %g2,#scratch
176 .register %g3,#scratch
180 #define EX_RETVAL(x) x
185 wr %g0, ASI_AIUS, %asi
187 ba,pt %xcc, __restore_asi
190 ENTRY(NG2_retl_o2_plus_1)
191 ba,pt %xcc, __restore_asi
193 ENDPROC(NG2_retl_o2_plus_1)
194 ENTRY(NG2_retl_o2_plus_4)
195 ba,pt %xcc, __restore_asi
197 ENDPROC(NG2_retl_o2_plus_4)
198 ENTRY(NG2_retl_o2_plus_8)
199 ba,pt %xcc, __restore_asi
201 ENDPROC(NG2_retl_o2_plus_8)
202 ENTRY(NG2_retl_o2_plus_o4_plus_1)
204 ba,pt %xcc, __restore_asi
206 ENDPROC(NG2_retl_o2_plus_o4_plus_1)
207 ENTRY(NG2_retl_o2_plus_o4_plus_8)
209 ba,pt %xcc, __restore_asi
211 ENDPROC(NG2_retl_o2_plus_o4_plus_8)
212 ENTRY(NG2_retl_o2_plus_o4_plus_16)
214 ba,pt %xcc, __restore_asi
216 ENDPROC(NG2_retl_o2_plus_o4_plus_16)
217 ENTRY(NG2_retl_o2_plus_g1_fp)
218 ba,pt %xcc, __restore_fp
220 ENDPROC(NG2_retl_o2_plus_g1_fp)
221 ENTRY(NG2_retl_o2_plus_g1_plus_64_fp)
223 ba,pt %xcc, __restore_fp
225 ENDPROC(NG2_retl_o2_plus_g1_plus_64_fp)
226 ENTRY(NG2_retl_o2_plus_g1_plus_1)
228 ba,pt %xcc, __restore_asi
230 ENDPROC(NG2_retl_o2_plus_g1_plus_1)
231 ENTRY(NG2_retl_o2_and_7_plus_o4)
233 ba,pt %xcc, __restore_asi
235 ENDPROC(NG2_retl_o2_and_7_plus_o4)
236 ENTRY(NG2_retl_o2_and_7_plus_o4_plus_8)
239 ba,pt %xcc, __restore_asi
241 ENDPROC(NG2_retl_o2_and_7_plus_o4_plus_8)
247 .type FUNC_NAME,#function
248 FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
256 or %o0, %o1, GLOBAL_SPARE
259 or GLOBAL_SPARE, %o2, GLOBAL_SPARE
261 /* 2 blocks (128 bytes) is the minimum we can do the block
262 * copy with. We need to ensure that we'll iterate at least
263 * once in the block copy loop. At worst we'll need to align
264 * the destination to a 64-byte boundary which can chew up
265 * to (64 - 1) bytes from the length before we perform the
268 * However, the cut-off point, performance wise, is around
273 andcc GLOBAL_SPARE, 0x7, %g0
277 * %o2: len (known to be >= 128)
279 * The block copy loops can use %o4, %g2, %g3 as
280 * temporaries while copying the data. %o5 must
281 * be preserved between VISEntryHalf and VISExitHalf
284 LOAD(prefetch, %o1 + 0x000, #one_read)
285 LOAD(prefetch, %o1 + 0x040, #one_read)
286 LOAD(prefetch, %o1 + 0x080, #one_read)
288 /* Align destination on 64-byte boundary. */
289 andcc %o0, (64 - 1), %o4
292 sub %g0, %o4, %o4 ! bytes to align dst
295 EX_LD(LOAD(ldub, %o1, %g1), NG2_retl_o2_plus_o4_plus_1)
296 EX_ST(STORE(stb, %g1, %o0), NG2_retl_o2_plus_o4_plus_1)
302 /* Clobbers o5/g1/g2/g3/g7/icc/xcc. We must preserve
303 * o5 from here until we hit VISExitHalf.
308 alignaddr %o1, %g0, %g0
310 add %o1, (64 - 1), %o4
311 andn %o4, (64 - 1), %o4
312 andn %o2, (64 - 1), %g1
315 and %o1, (64 - 1), %g2
330 4: /* 32 <= low bits < 48 */
334 5: /* 0 < low bits < 32 */
341 6: /* 0 < low bits < 16 */
344 /* fall through for 0 < low bits < 8 */
345 110: sub %o4, 64, %g2
346 EX_LD_FP(LOAD_BLK(%g2, %f0), NG2_retl_o2_plus_g1)
347 1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3), NG2_retl_o2_plus_g1)
348 EX_LD_FP(LOAD_BLK(%o4, %f16), NG2_retl_o2_plus_g1)
349 FREG_FROB(f0, f2, f4, f6, f8, f10, f12, f14, f16)
350 EX_ST_FP(STORE_BLK(%f0, %o4 + %g3), NG2_retl_o2_plus_g1)
351 FREG_MOVE_8(f16, f18, f20, f22, f24, f26, f28, f30)
355 LOAD(prefetch, %o4 + 64, #one_read)
359 120: sub %o4, 56, %g2
360 FREG_LOAD_7(%g2, f0, f2, f4, f6, f8, f10, f12)
361 1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3), NG2_retl_o2_plus_g1)
362 EX_LD_FP(LOAD_BLK(%o4, %f16), NG2_retl_o2_plus_g1)
363 FREG_FROB(f0, f2, f4, f6, f8, f10, f12, f16, f18)
364 EX_ST_FP(STORE_BLK(%f0, %o4 + %g3), NG2_retl_o2_plus_g1)
365 FREG_MOVE_7(f18, f20, f22, f24, f26, f28, f30)
369 LOAD(prefetch, %o4 + 64, #one_read)
373 130: sub %o4, 48, %g2
374 FREG_LOAD_6(%g2, f0, f2, f4, f6, f8, f10)
375 1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3), NG2_retl_o2_plus_g1)
376 EX_LD_FP(LOAD_BLK(%o4, %f16), NG2_retl_o2_plus_g1)
377 FREG_FROB(f0, f2, f4, f6, f8, f10, f16, f18, f20)
378 EX_ST_FP(STORE_BLK(%f0, %o4 + %g3), NG2_retl_o2_plus_g1)
379 FREG_MOVE_6(f20, f22, f24, f26, f28, f30)
383 LOAD(prefetch, %o4 + 64, #one_read)
387 140: sub %o4, 40, %g2
388 FREG_LOAD_5(%g2, f0, f2, f4, f6, f8)
389 1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3), NG2_retl_o2_plus_g1)
390 EX_LD_FP(LOAD_BLK(%o4, %f16), NG2_retl_o2_plus_g1)
391 FREG_FROB(f0, f2, f4, f6, f8, f16, f18, f20, f22)
392 EX_ST_FP(STORE_BLK(%f0, %o4 + %g3), NG2_retl_o2_plus_g1)
393 FREG_MOVE_5(f22, f24, f26, f28, f30)
397 LOAD(prefetch, %o4 + 64, #one_read)
401 150: sub %o4, 32, %g2
402 FREG_LOAD_4(%g2, f0, f2, f4, f6)
403 1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3), NG2_retl_o2_plus_g1)
404 EX_LD_FP(LOAD_BLK(%o4, %f16), NG2_retl_o2_plus_g1)
405 FREG_FROB(f0, f2, f4, f6, f16, f18, f20, f22, f24)
406 EX_ST_FP(STORE_BLK(%f0, %o4 + %g3), NG2_retl_o2_plus_g1)
407 FREG_MOVE_4(f24, f26, f28, f30)
411 LOAD(prefetch, %o4 + 64, #one_read)
415 160: sub %o4, 24, %g2
416 FREG_LOAD_3(%g2, f0, f2, f4)
417 1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3), NG2_retl_o2_plus_g1)
418 EX_LD_FP(LOAD_BLK(%o4, %f16), NG2_retl_o2_plus_g1)
419 FREG_FROB(f0, f2, f4, f16, f18, f20, f22, f24, f26)
420 EX_ST_FP(STORE_BLK(%f0, %o4 + %g3), NG2_retl_o2_plus_g1)
421 FREG_MOVE_3(f26, f28, f30)
425 LOAD(prefetch, %o4 + 64, #one_read)
429 170: sub %o4, 16, %g2
430 FREG_LOAD_2(%g2, f0, f2)
431 1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3), NG2_retl_o2_plus_g1)
432 EX_LD_FP(LOAD_BLK(%o4, %f16), NG2_retl_o2_plus_g1)
433 FREG_FROB(f0, f2, f16, f18, f20, f22, f24, f26, f28)
434 EX_ST_FP(STORE_BLK(%f0, %o4 + %g3), NG2_retl_o2_plus_g1)
435 FREG_MOVE_2(f28, f30)
439 LOAD(prefetch, %o4 + 64, #one_read)
445 1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3), NG2_retl_o2_plus_g1)
446 EX_LD_FP(LOAD_BLK(%o4, %f16), NG2_retl_o2_plus_g1)
447 FREG_FROB(f0, f16, f18, f20, f22, f24, f26, f28, f30)
448 EX_ST_FP(STORE_BLK(%f0, %o4 + %g3), NG2_retl_o2_plus_g1)
453 LOAD(prefetch, %o4 + 64, #one_read)
458 1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3), NG2_retl_o2_plus_g1)
460 EX_LD_FP(LOAD_BLK(%o4, %f0), NG2_retl_o2_plus_g1_plus_64)
461 EX_ST_FP(STORE_BLK(%f0, %o4 + %g3), NG2_retl_o2_plus_g1_plus_64)
464 LOAD(prefetch, %o4 + 64, #one_read)
472 /* %o2 contains any final bytes still needed to be copied
473 * over. If anything is left, we copy it one byte at a time.
476 sub %o0, %o1, GLOBAL_SPARE
480 75: /* 16 < len <= 64 */
482 sub %o0, %o1, GLOBAL_SPARE
487 1: subcc %o4, 0x10, %o4
488 EX_LD(LOAD(ldx, %o1, %o5), NG2_retl_o2_plus_o4_plus_16)
490 EX_LD(LOAD(ldx, %o1, %g1), NG2_retl_o2_plus_o4_plus_16)
492 EX_ST(STORE(stx, %o5, %o1 + GLOBAL_SPARE), NG2_retl_o2_plus_o4_plus_16)
494 EX_ST(STORE(stx, %g1, %o1 + GLOBAL_SPARE), NG2_retl_o2_plus_o4_plus_8)
497 73: andcc %o2, 0x8, %g0
501 EX_LD(LOAD(ldx, %o1, %o5), NG2_retl_o2_plus_8)
502 EX_ST(STORE(stx, %o5, %o1 + GLOBAL_SPARE), NG2_retl_o2_plus_8)
504 1: andcc %o2, 0x4, %g0
508 EX_LD(LOAD(lduw, %o1, %o5), NG2_retl_o2_plus_4)
509 EX_ST(STORE(stw, %o5, %o1 + GLOBAL_SPARE), NG2_retl_o2_plus_4)
525 EX_LD(LOAD(ldub, %o1, %o5), NG2_retl_o2_plus_g1_plus_1)
526 EX_ST(STORE(stb, %o5, %o1 + GLOBAL_SPARE), NG2_retl_o2_plus_g1_plus_1)
530 2: add %o1, GLOBAL_SPARE, %o0
540 8: mov 64, GLOBAL_SPARE
542 EX_LD(LOAD(ldx, %o1, %g2), NG2_retl_o2)
543 sub GLOBAL_SPARE, %g1, GLOBAL_SPARE
547 EX_LD(LOAD(ldx, %o1, %g3), NG2_retl_o2_and_7_plus_o4)
549 srlx %g3, GLOBAL_SPARE, %o5
551 EX_ST(STORE(stx, %o5, %o0), NG2_retl_o2_and_7_plus_o4_plus_8)
561 sub %o0, %o1, GLOBAL_SPARE
564 80: /* 0 < len <= 16 */
565 andcc GLOBAL_SPARE, 0x3, %g0
567 sub %o0, %o1, GLOBAL_SPARE
571 EX_LD(LOAD(lduw, %o1, %g1), NG2_retl_o2_plus_4)
572 EX_ST(STORE(stw, %g1, %o1 + GLOBAL_SPARE), NG2_retl_o2_plus_4)
577 mov EX_RETVAL(%o3), %o0
582 EX_LD(LOAD(ldub, %o1, %g1), NG2_retl_o2_plus_1)
583 EX_ST(STORE(stb, %g1, %o1 + GLOBAL_SPARE), NG2_retl_o2_plus_1)
587 mov EX_RETVAL(%o3), %o0
589 .size FUNC_NAME, .-FUNC_NAME