1 /* checksum.S: Sparc optimized checksum code.
3 * Copyright(C) 1995 Linus Torvalds
4 * Copyright(C) 1995 Miguel de Icaza
5 * Copyright(C) 1996 David S. Miller
6 * Copyright(C) 1997 Jakub Jelinek
9 * Linux/Alpha checksum c-code
10 * Linux/ix86 inline checksum assembly
11 * RFC1071 Computing the Internet Checksum (esp. Jacobsons m68k code)
12 * David Mosberger-Tang for optimized reference c-code
13 * BSD4.4 portable checksum routine
16 #include <asm/errno.h>
18 #define CSUM_BIGCHUNK(buf, offset, sum, t0, t1, t2, t3, t4, t5) \
19 ldd [buf + offset + 0x00], t0; \
20 ldd [buf + offset + 0x08], t2; \
21 addxcc t0, sum, sum; \
22 addxcc t1, sum, sum; \
23 ldd [buf + offset + 0x10], t4; \
24 addxcc t2, sum, sum; \
25 addxcc t3, sum, sum; \
26 ldd [buf + offset + 0x18], t0; \
27 addxcc t4, sum, sum; \
28 addxcc t5, sum, sum; \
29 addxcc t0, sum, sum; \
32 #define CSUM_LASTCHUNK(buf, offset, sum, t0, t1, t2, t3) \
33 ldd [buf - offset - 0x08], t0; \
34 ldd [buf - offset - 0x00], t2; \
35 addxcc t0, sum, sum; \
36 addxcc t1, sum, sum; \
37 addxcc t2, sum, sum; \
40 /* Do end cruft out of band to get better cache patterns. */
41 csum_partial_end_cruft:
42 be 1f ! caller asks %o1 & 0x8
43 andcc %o1, 4, %g0 ! nope, check for word remaining
44 ldd [%o0], %g2 ! load two
45 addcc %g2, %o2, %o2 ! add first word to sum
46 addxcc %g3, %o2, %o2 ! add second word as well
47 add %o0, 8, %o0 ! advance buf ptr
48 addx %g0, %o2, %o2 ! add in final carry
49 andcc %o1, 4, %g0 ! check again for word remaining
50 1: be 1f ! nope, skip this code
51 andcc %o1, 3, %o1 ! check for trailing bytes
52 ld [%o0], %g2 ! load it
53 addcc %g2, %o2, %o2 ! add to sum
54 add %o0, 4, %o0 ! advance buf ptr
55 addx %g0, %o2, %o2 ! add in final carry
56 andcc %o1, 3, %g0 ! check again for trailing bytes
57 1: be 1f ! no trailing bytes, return
58 addcc %o1, -1, %g0 ! only one byte remains?
59 bne 2f ! at least two bytes more
60 subcc %o1, 2, %o1 ! only two bytes more?
61 b 4f ! only one byte remains
62 or %g0, %g0, %o4 ! clear fake hword value
63 2: lduh [%o0], %o4 ! get hword
64 be 6f ! jmp if only hword remains
65 add %o0, 2, %o0 ! advance buf ptr either way
66 sll %o4, 16, %o4 ! create upper hword
67 4: ldub [%o0], %o5 ! get final byte
68 sll %o5, 8, %o5 ! put into place
69 or %o5, %o4, %o4 ! coalese with hword (if any)
70 6: addcc %o4, %o2, %o2 ! add to sum
71 1: retl ! get outta here
72 addx %g0, %o2, %o0 ! add final carry into retval
74 /* Also do alignment out of band to get better cache patterns. */
75 csum_partial_fix_alignment:
81 lduh [%o0 + 0x00], %g2
94 andcc %o1, 0xffffff80, %o3
101 andcc %o1, 0xffffff80, %o3
103 /* The common case is to get called with a nicely aligned
104 * buffer of size 0x20. Follow the code path for that case.
107 csum_partial: /* %o0=buf, %o1=len, %o2=sum */
108 andcc %o0, 0x7, %g0 ! alignment problems?
109 bne csum_partial_fix_alignment ! yep, handle it
110 sethi %hi(cpte - 8), %g7 ! prepare table jmp ptr
111 andcc %o1, 0xffffff80, %o3 ! num loop iterations
112 cpa: be 3f ! none to do
113 andcc %o1, 0x70, %g1 ! clears carry flag too
114 5: CSUM_BIGCHUNK(%o0, 0x00, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
115 CSUM_BIGCHUNK(%o0, 0x20, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
116 CSUM_BIGCHUNK(%o0, 0x40, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
117 CSUM_BIGCHUNK(%o0, 0x60, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
118 addx %g0, %o2, %o2 ! sink in final carry
119 subcc %o3, 128, %o3 ! detract from loop iters
121 add %o0, 128, %o0 ! advance buf ptr
122 andcc %o1, 0x70, %g1 ! clears carry flag too
124 andcc %o1, 0xf, %g0 ! anything left at all?
125 srl %g1, 1, %o4 ! compute offset
126 sub %g7, %g1, %g7 ! adjust jmp ptr
127 sub %g7, %o4, %g7 ! final jmp ptr adjust
128 jmp %g7 + %lo(cpte - 8) ! enter the table
129 add %o0, %g1, %o0 ! advance buf ptr
130 cptbl: CSUM_LASTCHUNK(%o0, 0x68, %o2, %g2, %g3, %g4, %g5)
131 CSUM_LASTCHUNK(%o0, 0x58, %o2, %g2, %g3, %g4, %g5)
132 CSUM_LASTCHUNK(%o0, 0x48, %o2, %g2, %g3, %g4, %g5)
133 CSUM_LASTCHUNK(%o0, 0x38, %o2, %g2, %g3, %g4, %g5)
134 CSUM_LASTCHUNK(%o0, 0x28, %o2, %g2, %g3, %g4, %g5)
135 CSUM_LASTCHUNK(%o0, 0x18, %o2, %g2, %g3, %g4, %g5)
136 CSUM_LASTCHUNK(%o0, 0x08, %o2, %g2, %g3, %g4, %g5)
137 addx %g0, %o2, %o2 ! fetch final carry
138 andcc %o1, 0xf, %g0 ! anything left at all?
139 cpte: bne csum_partial_end_cruft ! yep, handle it
140 andcc %o1, 8, %g0 ! check how much
141 cpout: retl ! get outta here
142 mov %o2, %o0 ! return computed csum
144 .globl __csum_partial_copy_start, __csum_partial_copy_end
145 __csum_partial_copy_start:
148 #define EXECINSTR #execinstr
149 #define EX(x,y,a,b) \
151 .section .fixup,ALLOC,EXECINSTR; \
155 .section __ex_table,ALLOC; \
163 .section __ex_table,ALLOC; \
171 .section __ex_table,ALLOC; \
177 #define EXT(start,end,handler) \
178 .section __ex_table,ALLOC; \
180 .word start, 0, end, handler; \
184 /* This aligned version executes typically in 8.5 superscalar cycles, this
185 * is the best I can do. I say 8.5 because the final add will pair with
186 * the next ldd in the main unrolled loop. Thus the pipe is always full.
187 * If you change these macros (including order of instructions),
188 * please check the fixup code below as well.
190 #define CSUMCOPY_BIGCHUNK_ALIGNED(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7) \
191 ldd [src + off + 0x00], t0; \
192 ldd [src + off + 0x08], t2; \
193 addxcc t0, sum, sum; \
194 ldd [src + off + 0x10], t4; \
195 addxcc t1, sum, sum; \
196 ldd [src + off + 0x18], t6; \
197 addxcc t2, sum, sum; \
198 std t0, [dst + off + 0x00]; \
199 addxcc t3, sum, sum; \
200 std t2, [dst + off + 0x08]; \
201 addxcc t4, sum, sum; \
202 std t4, [dst + off + 0x10]; \
203 addxcc t5, sum, sum; \
204 std t6, [dst + off + 0x18]; \
205 addxcc t6, sum, sum; \
208 /* 12 superscalar cycles seems to be the limit for this case,
209 * because of this we thus do all the ldd's together to get
210 * Viking MXCC into streaming mode. Ho hum...
212 #define CSUMCOPY_BIGCHUNK(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7) \
213 ldd [src + off + 0x00], t0; \
214 ldd [src + off + 0x08], t2; \
215 ldd [src + off + 0x10], t4; \
216 ldd [src + off + 0x18], t6; \
217 st t0, [dst + off + 0x00]; \
218 addxcc t0, sum, sum; \
219 st t1, [dst + off + 0x04]; \
220 addxcc t1, sum, sum; \
221 st t2, [dst + off + 0x08]; \
222 addxcc t2, sum, sum; \
223 st t3, [dst + off + 0x0c]; \
224 addxcc t3, sum, sum; \
225 st t4, [dst + off + 0x10]; \
226 addxcc t4, sum, sum; \
227 st t5, [dst + off + 0x14]; \
228 addxcc t5, sum, sum; \
229 st t6, [dst + off + 0x18]; \
230 addxcc t6, sum, sum; \
231 st t7, [dst + off + 0x1c]; \
234 /* Yuck, 6 superscalar cycles... */
235 #define CSUMCOPY_LASTCHUNK(src, dst, sum, off, t0, t1, t2, t3) \
236 ldd [src - off - 0x08], t0; \
237 ldd [src - off - 0x00], t2; \
238 addxcc t0, sum, sum; \
239 st t0, [dst - off - 0x08]; \
240 addxcc t1, sum, sum; \
241 st t1, [dst - off - 0x04]; \
242 addxcc t2, sum, sum; \
243 st t2, [dst - off - 0x00]; \
244 addxcc t3, sum, sum; \
245 st t3, [dst - off + 0x04];
247 /* Handle the end cruft code out of band for better cache patterns. */
251 EX(ldd [%o0 + 0x00], %g2, and %o3, 0xf)
256 EX2(st %g2, [%o1 - 0x08])
259 EX2(st %g3, [%o1 - 0x04])
262 EX(ld [%o0 + 0x00], %g2, add %o3, 4)
265 EX2(st %g2, [%o1 - 0x04])
275 2: EX(lduh [%o0 + 0x00], %o4, add %o3, 2)
277 EX2(sth %o4, [%o1 + 0x00])
281 4: EX(ldub [%o0 + 0x00], %o5, add %g0, 1)
282 EX2(stb %o5, [%o1 + 0x00])
285 6: addcc %o4, %g7, %g7
289 /* Also, handle the alignment code out of band. */
299 EX(lduh [%o0 + 0x00], %g4, add %g1, 0)
301 EX2(sth %g4, [%o1 + 0x00])
314 andcc %g1, 0xffffff80, %g0
315 EX(ld [%o0 + 0x00], %g4, add %g1, 0)
317 EX2(st %g4, [%o1 + 0x00])
323 andcc %g1, 0xffffff80, %g0
325 /* Sun, you just can't beat me, you just can't. Stop trying,
326 * give up. I'm serious, I am going to kick the living shit
327 * out of you, game over, lights out.
330 .globl __csum_partial_copy_sparc_generic
331 __csum_partial_copy_sparc_generic:
332 /* %o0=src, %o1=dest, %g1=len, %g7=sum */
333 xor %o0, %o1, %o4 ! get changing bits
334 andcc %o4, 3, %g0 ! check for mismatched alignment
335 bne ccslow ! better this than unaligned/fixups
336 andcc %o0, 7, %g0 ! need to align things?
337 bne cc_dword_align ! yes, we check for short lengths there
338 andcc %g1, 0xffffff80, %g0 ! can we use unrolled loop?
339 3: be 3f ! nope, less than one loop remains
340 andcc %o1, 4, %g0 ! dest aligned on 4 or 8 byte boundary?
341 be ccdbl + 4 ! 8 byte aligned, kick ass
342 5: CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
343 CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
344 CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
345 CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
346 10: EXT(5b, 10b, 20f) ! note for exception handling
347 sub %g1, 128, %g1 ! detract from length
348 addx %g0, %g7, %g7 ! add in last carry bit
349 andcc %g1, 0xffffff80, %g0 ! more to csum?
350 add %o0, 128, %o0 ! advance src ptr
351 bne 5b ! we did not go negative, continue looping
352 add %o1, 128, %o1 ! advance dest ptr
353 3: andcc %g1, 0x70, %o2 ! can use table?
354 ccmerge:be ccte ! nope, go and check for end cruft
355 andcc %g1, 0xf, %o3 ! get low bits of length (clears carry btw)
356 srl %o2, 1, %o4 ! begin negative offset computation
357 sethi %hi(12f), %o5 ! set up table ptr end
358 add %o0, %o2, %o0 ! advance src ptr
359 sub %o5, %o4, %o5 ! continue table calculation
360 sll %o2, 1, %g2 ! constant multiplies are fun...
361 sub %o5, %g2, %o5 ! some more adjustments
362 jmp %o5 + %lo(12f) ! jump into it, duff style, wheee...
363 add %o1, %o2, %o1 ! advance dest ptr (carry is clear btw)
364 cctbl: CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x68,%g2,%g3,%g4,%g5)
365 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x58,%g2,%g3,%g4,%g5)
366 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x48,%g2,%g3,%g4,%g5)
367 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x38,%g2,%g3,%g4,%g5)
368 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x28,%g2,%g3,%g4,%g5)
369 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x18,%g2,%g3,%g4,%g5)
370 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x08,%g2,%g3,%g4,%g5)
371 12: EXT(cctbl, 12b, 22f) ! note for exception table handling
373 andcc %o3, 0xf, %g0 ! check for low bits set
374 ccte: bne cc_end_cruft ! something left, handle it out of band
375 andcc %o3, 8, %g0 ! begin checks for that code
377 mov %g7, %o0 ! give em the computed checksum
378 ccdbl: CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
379 CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
380 CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
381 CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
382 11: EXT(ccdbl, 11b, 21f) ! note for exception table handling
383 sub %g1, 128, %g1 ! detract from length
384 addx %g0, %g7, %g7 ! add in last carry bit
385 andcc %g1, 0xffffff80, %g0 ! more to csum?
386 add %o0, 128, %o0 ! advance src ptr
387 bne ccdbl ! we did not go negative, continue looping
388 add %o1, 128, %o1 ! advance dest ptr
389 b ccmerge ! finish it off, above
390 andcc %g1, 0x70, %o2 ! can use table? (clears carry btw)
399 EX(ldub [%o0], %g5, add %g1, 1)
410 EX(lduh [%o0], %o4, add %g1, 0)
416 EX2(stb %o4, [%o1 + 1])
428 EX2(stb %g3, [%o1 + 1])
430 EX2(stb %g2, [%o1 + 2])
432 EX2(stb %o4, [%o1 + 3])
433 addx %g5, %g0, %g5 ! I am now to lazy to optimize this (question it
434 add %o1, 4, %o1 ! is worthy). Maybe some day - with the sll/srl
435 subcc %g4, 1, %g4 ! tricks
445 EX(lduh [%o0], %o4, and %g1, 3)
451 EX2(stb %o4, [%o1 + 1])
455 EX(ldub [%o0], %g2, add %g0, 1)
460 1: addcc %o4, %g5, %g5
470 4: addcc %g7, %g5, %g7
473 __csum_partial_copy_end:
475 /* We do these strange calculations for the csum_*_from_user case only, ie.
476 * we only bother with faults on loads... */
478 /* o2 = ((g2%20)&3)*8
479 * o3 = g1 - (g2/20)*32 - o2 */
492 /* o2 = (!(g2 & 15) ? 0 : (((g2 & 15) + 1) & ~1)*8)
493 * o3 = g1 - (g2/16)*32 - o2 */
508 /* o0 += (g2/10)*16 - 0x70
509 * 01 += (g2/10)*16 - 0x70
510 * o2 = (g2 % 10) ? 8 : 0
511 * o3 += 0x70 - (g2/10)*16 - o2 */
537 * %o3 is # bytes to zero out
538 * %o4 is faulting address
539 * %o5 is %pc where fault occurred */
544 * %o2 is # of bytes to copy from src to dst
545 * %o3 is # bytes to zero out
546 * %o4 is faulting address
547 * %o5 is %pc where fault occurred */
556 add %g0, -EFAULT, %i5
574 ld [%sp + 168], %o2 ! struct_ptr of parent
579 .section __ex_table,#alloc