1 /* SPDX-License-Identifier: GPL-2.0-or-later * !! 1 /* checksum.S: Sparc optimized checksum code.
2 /* <<
3 * INET An implementation of the TCP/I <<
4 * operating system. INET is imp <<
5 * interface as the means of comm <<
6 * 2 *
7 * IP/TCP/UDP checksumming routin !! 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
8 * 7 *
9 * Xtensa version: Copyright (C) 2001 Tensili !! 8 * derived from:
10 * Optimized by Joe Taylor !! 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
11 */ 14 */
12 15
13 #include <linux/errno.h> !! 16 #include <asm/cprefix.h>
14 #include <linux/linkage.h> !! 17 #include <asm/errno.h>
15 #include <asm/asmmacro.h> <<
16 #include <asm/core.h> <<
17 18
18 /* !! 19 #define CSUM_BIGCHUNK(buf, offset, sum, t0, t1, t2, t3, t4, t5) \
19 * computes a partial checksum, e.g. for TCP/U !! 20 ldd [buf + offset + 0x00], t0; \
20 */ !! 21 ldd [buf + offset + 0x08], t2; \
21 !! 22 addxcc t0, sum, sum; \
22 /* !! 23 addxcc t1, sum, sum; \
23 * unsigned int csum_partial(const unsigned ch !! 24 ldd [buf + offset + 0x10], t4; \
24 * unsigned int sum) !! 25 addxcc t2, sum, sum; \
25 * a2 = buf !! 26 addxcc t3, sum, sum; \
26 * a3 = len !! 27 ldd [buf + offset + 0x18], t0; \
27 * a4 = sum !! 28 addxcc t4, sum, sum; \
28 * !! 29 addxcc t5, sum, sum; \
29 * This function assumes 2- or 4-byte alignmen !! 30 addxcc t0, sum, sum; \
30 */ !! 31 addxcc t1, sum, sum;
>> 32
>> 33 #define CSUM_LASTCHUNK(buf, offset, sum, t0, t1, t2, t3) \
>> 34 ldd [buf - offset - 0x08], t0; \
>> 35 ldd [buf - offset - 0x00], t2; \
>> 36 addxcc t0, sum, sum; \
>> 37 addxcc t1, sum, sum; \
>> 38 addxcc t2, sum, sum; \
>> 39 addxcc t3, sum, sum;
>> 40
>> 41 /* Do end cruft out of band to get better cache patterns. */
>> 42 csum_partial_end_cruft:
>> 43 be 1f ! caller asks %o1 & 0x8
>> 44 andcc %o1, 4, %g0 ! nope, check for word remaining
>> 45 ldd [%o0], %g2 ! load two
>> 46 addcc %g2, %o2, %o2 ! add first word to sum
>> 47 addxcc %g3, %o2, %o2 ! add second word as well
>> 48 add %o0, 8, %o0 ! advance buf ptr
>> 49 addx %g0, %o2, %o2 ! add in final carry
>> 50 andcc %o1, 4, %g0 ! check again for word remaining
>> 51 1: be 1f ! nope, skip this code
>> 52 andcc %o1, 3, %o1 ! check for trailing bytes
>> 53 ld [%o0], %g2 ! load it
>> 54 addcc %g2, %o2, %o2 ! add to sum
>> 55 add %o0, 4, %o0 ! advance buf ptr
>> 56 addx %g0, %o2, %o2 ! add in final carry
>> 57 andcc %o1, 3, %g0 ! check again for trailing bytes
>> 58 1: be 1f ! no trailing bytes, return
>> 59 addcc %o1, -1, %g0 ! only one byte remains?
>> 60 bne 2f ! at least two bytes more
>> 61 subcc %o1, 2, %o1 ! only two bytes more?
>> 62 b 4f ! only one byte remains
>> 63 or %g0, %g0, %o4 ! clear fake hword value
>> 64 2: lduh [%o0], %o4 ! get hword
>> 65 be 6f ! jmp if only hword remains
>> 66 add %o0, 2, %o0 ! advance buf ptr either way
>> 67 sll %o4, 16, %o4 ! create upper hword
>> 68 4: ldub [%o0], %o5 ! get final byte
>> 69 sll %o5, 8, %o5 ! put into place
>> 70 or %o5, %o4, %o4 ! coalese with hword (if any)
>> 71 6: addcc %o4, %o2, %o2 ! add to sum
>> 72 1: retl ! get outta here
>> 73 addx %g0, %o2, %o0 ! add final carry into retval
>> 74
>> 75 /* Also do alignment out of band to get better cache patterns. */
>> 76 csum_partial_fix_alignment:
>> 77 cmp %o1, 6
>> 78 bl cpte - 0x4
>> 79 andcc %o0, 0x2, %g0
>> 80 be 1f
>> 81 andcc %o0, 0x4, %g0
>> 82 lduh [%o0 + 0x00], %g2
>> 83 sub %o1, 2, %o1
>> 84 add %o0, 2, %o0
>> 85 sll %g2, 16, %g2
>> 86 addcc %g2, %o2, %o2
>> 87 srl %o2, 16, %g3
>> 88 addx %g0, %g3, %g2
>> 89 sll %o2, 16, %o2
>> 90 sll %g2, 16, %g3
>> 91 srl %o2, 16, %o2
>> 92 andcc %o0, 0x4, %g0
>> 93 or %g3, %o2, %o2
>> 94 1: be cpa
>> 95 andcc %o1, 0xffffff80, %o3
>> 96 ld [%o0 + 0x00], %g2
>> 97 sub %o1, 4, %o1
>> 98 addcc %g2, %o2, %o2
>> 99 add %o0, 4, %o0
>> 100 addx %g0, %o2, %o2
>> 101 b cpa
>> 102 andcc %o1, 0xffffff80, %o3
31 103
32 /* ONES_ADD converts twos-complement math to o !! 104 /* The common case is to get called with a nicely aligned
33 #define ONES_ADD(sum, val) \ !! 105 * buffer of size 0x20. Follow the code path for that case.
34 add sum, sum, val ; \ <<
35 bgeu sum, val, 99f ; \ <<
36 addi sum, sum, 1 ; \ <<
37 99: ; <<
38 <<
39 .text <<
40 ENTRY(csum_partial) <<
41 <<
42 /* <<
43 * Experiments with Ethernet and SLIP <<
44 * is aligned on either a 2-byte or 4- <<
45 */ 106 */
46 abi_entry_default !! 107 .globl C_LABEL(csum_partial)
47 extui a5, a2, 0, 2 !! 108 C_LABEL(csum_partial): /* %o0=buf, %o1=len, %o2=sum */
48 bnez a5, 8f /* branch if 2 !! 109 andcc %o0, 0x7, %g0 ! alignment problems?
49 /* Fall-through on common case, 4-byte !! 110 bne csum_partial_fix_alignment ! yep, handle it
50 1: !! 111 sethi %hi(cpte - 8), %g7 ! prepare table jmp ptr
51 srli a5, a3, 5 /* 32-byte chu !! 112 andcc %o1, 0xffffff80, %o3 ! num loop iterations
52 #if XCHAL_HAVE_LOOPS !! 113 cpa: be 3f ! none to do
53 loopgtz a5, 2f !! 114 andcc %o1, 0x70, %g1 ! clears carry flag too
54 #else !! 115 5: CSUM_BIGCHUNK(%o0, 0x00, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
55 beqz a5, 2f !! 116 CSUM_BIGCHUNK(%o0, 0x20, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
56 slli a5, a5, 5 !! 117 CSUM_BIGCHUNK(%o0, 0x40, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
57 add a5, a5, a2 /* a5 = end of !! 118 CSUM_BIGCHUNK(%o0, 0x60, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
58 .Loop1: !! 119 addx %g0, %o2, %o2 ! sink in final carry
59 #endif !! 120 subcc %o3, 128, %o3 ! detract from loop iters
60 l32i a6, a2, 0 !! 121 bne 5b ! more to do
61 l32i a7, a2, 4 !! 122 add %o0, 128, %o0 ! advance buf ptr
62 ONES_ADD(a4, a6) !! 123 andcc %o1, 0x70, %g1 ! clears carry flag too
63 ONES_ADD(a4, a7) !! 124 3: be cpte ! nope
64 l32i a6, a2, 8 !! 125 andcc %o1, 0xf, %g0 ! anything left at all?
65 l32i a7, a2, 12 !! 126 srl %g1, 1, %o4 ! compute offset
66 ONES_ADD(a4, a6) !! 127 sub %g7, %g1, %g7 ! adjust jmp ptr
67 ONES_ADD(a4, a7) !! 128 sub %g7, %o4, %g7 ! final jmp ptr adjust
68 l32i a6, a2, 16 !! 129 jmp %g7 + %lo(cpte - 8) ! enter the table
69 l32i a7, a2, 20 !! 130 add %o0, %g1, %o0 ! advance buf ptr
70 ONES_ADD(a4, a6) !! 131 cptbl: CSUM_LASTCHUNK(%o0, 0x68, %o2, %g2, %g3, %g4, %g5)
71 ONES_ADD(a4, a7) !! 132 CSUM_LASTCHUNK(%o0, 0x58, %o2, %g2, %g3, %g4, %g5)
72 l32i a6, a2, 24 !! 133 CSUM_LASTCHUNK(%o0, 0x48, %o2, %g2, %g3, %g4, %g5)
73 l32i a7, a2, 28 !! 134 CSUM_LASTCHUNK(%o0, 0x38, %o2, %g2, %g3, %g4, %g5)
74 ONES_ADD(a4, a6) !! 135 CSUM_LASTCHUNK(%o0, 0x28, %o2, %g2, %g3, %g4, %g5)
75 ONES_ADD(a4, a7) !! 136 CSUM_LASTCHUNK(%o0, 0x18, %o2, %g2, %g3, %g4, %g5)
76 addi a2, a2, 4*8 !! 137 CSUM_LASTCHUNK(%o0, 0x08, %o2, %g2, %g3, %g4, %g5)
77 #if !XCHAL_HAVE_LOOPS !! 138 addx %g0, %o2, %o2 ! fetch final carry
78 blt a2, a5, .Loop1 !! 139 andcc %o1, 0xf, %g0 ! anything left at all?
79 #endif !! 140 cpte: bne csum_partial_end_cruft ! yep, handle it
80 2: !! 141 andcc %o1, 8, %g0 ! check how much
81 extui a5, a3, 2, 3 /* remaining 4 !! 142 cpout: retl ! get outta here
82 #if XCHAL_HAVE_LOOPS !! 143 mov %o2, %o0 ! return computed csum
83 loopgtz a5, 3f !! 144
84 #else !! 145 .globl C_LABEL(__csum_partial_copy_start), C_LABEL(__csum_partial_copy_end)
85 beqz a5, 3f !! 146 C_LABEL(__csum_partial_copy_start):
86 slli a5, a5, 2 !! 147
87 add a5, a5, a2 /* a5 = end of !! 148 /* Work around cpp -rob */
88 .Loop2: !! 149 #define ALLOC #alloc
89 #endif !! 150 #define EXECINSTR #execinstr
90 l32i a6, a2, 0 !! 151 #define EX(x,y,a,b) \
91 ONES_ADD(a4, a6) !! 152 98: x,y; \
92 addi a2, a2, 4 !! 153 .section .fixup,ALLOC,EXECINSTR; \
93 #if !XCHAL_HAVE_LOOPS !! 154 .align 4; \
94 blt a2, a5, .Loop2 !! 155 99: ba 30f; \
95 #endif !! 156 a, b, %o3; \
96 3: !! 157 .section __ex_table,ALLOC; \
97 _bbci.l a3, 1, 5f /* remaining 2 !! 158 .align 4; \
98 l16ui a6, a2, 0 !! 159 .word 98b, 99b; \
99 ONES_ADD(a4, a6) !! 160 .text; \
100 addi a2, a2, 2 !! 161 .align 4
101 5: !! 162
102 _bbci.l a3, 0, 7f /* remaining 1 !! 163 #define EX2(x,y) \
103 6: l8ui a6, a2, 0 !! 164 98: x,y; \
104 #ifdef __XTENSA_EB__ !! 165 .section __ex_table,ALLOC; \
105 slli a6, a6, 8 /* load byte i !! 166 .align 4; \
106 #endif !! 167 .word 98b, 30f; \
107 ONES_ADD(a4, a6) !! 168 .text; \
108 7: !! 169 .align 4
109 mov a2, a4 !! 170
110 abi_ret_default !! 171 #define EX3(x,y) \
111 !! 172 98: x,y; \
112 /* uncommon case, buf is 2-byte aligne !! 173 .section __ex_table,ALLOC; \
113 8: !! 174 .align 4; \
114 beqz a3, 7b /* branch if l !! 175 .word 98b, 96f; \
115 beqi a3, 1, 6b /* branch if l !! 176 .text; \
116 !! 177 .align 4
117 extui a5, a2, 0, 1 !! 178
118 bnez a5, 8f /* branch if 1 !! 179 #define EXT(start,end,handler) \
119 !! 180 .section __ex_table,ALLOC; \
120 l16ui a6, a2, 0 /* common case !! 181 .align 4; \
121 ONES_ADD(a4, a6) !! 182 .word start, 0, end, handler; \
122 addi a2, a2, 2 /* adjust buf !! 183 .text; \
123 addi a3, a3, -2 /* adjust len !! 184 .align 4
124 j 1b /* now buf is !! 185
125 !! 186 /* This aligned version executes typically in 8.5 superscalar cycles, this
126 /* case: odd-byte aligned, len > 1 !! 187 * is the best I can do. I say 8.5 because the final add will pair with
127 * This case is dog slow, so don't giv !! 188 * the next ldd in the main unrolled loop. Thus the pipe is always full.
128 * (I don't think this ever happens, b !! 189 * If you change these macros (including order of instructions),
>> 190 * please check the fixup code below as well.
129 */ 191 */
130 8: !! 192 #define CSUMCOPY_BIGCHUNK_ALIGNED(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7) \
131 srli a5, a3, 2 /* 4-byte chun !! 193 ldd [src + off + 0x00], t0; \
132 #if XCHAL_HAVE_LOOPS !! 194 ldd [src + off + 0x08], t2; \
133 loopgtz a5, 2f !! 195 addxcc t0, sum, sum; \
134 #else !! 196 ldd [src + off + 0x10], t4; \
135 beqz a5, 2f !! 197 addxcc t1, sum, sum; \
136 slli a5, a5, 2 !! 198 ldd [src + off + 0x18], t6; \
137 add a5, a5, a2 /* a5 = end of !! 199 addxcc t2, sum, sum; \
138 .Loop3: !! 200 std t0, [dst + off + 0x00]; \
139 #endif !! 201 addxcc t3, sum, sum; \
140 l8ui a6, a2, 0 /* bits 24..31 !! 202 std t2, [dst + off + 0x08]; \
141 l16ui a7, a2, 1 /* bits 8..23 !! 203 addxcc t4, sum, sum; \
142 l8ui a8, a2, 3 /* bits 0.. 8 !! 204 std t4, [dst + off + 0x10]; \
143 #ifdef __XTENSA_EB__ !! 205 addxcc t5, sum, sum; \
144 slli a6, a6, 24 !! 206 std t6, [dst + off + 0x18]; \
145 #else !! 207 addxcc t6, sum, sum; \
146 slli a8, a8, 24 !! 208 addxcc t7, sum, sum;
147 #endif !! 209
148 slli a7, a7, 8 !! 210 /* 12 superscalar cycles seems to be the limit for this case,
149 or a7, a7, a6 !! 211 * because of this we thus do all the ldd's together to get
150 or a7, a7, a8 !! 212 * Viking MXCC into streaming mode. Ho hum...
151 ONES_ADD(a4, a7) !! 213 */
152 addi a2, a2, 4 !! 214 #define CSUMCOPY_BIGCHUNK(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7) \
153 #if !XCHAL_HAVE_LOOPS !! 215 ldd [src + off + 0x00], t0; \
154 blt a2, a5, .Loop3 !! 216 ldd [src + off + 0x08], t2; \
155 #endif !! 217 ldd [src + off + 0x10], t4; \
156 2: !! 218 ldd [src + off + 0x18], t6; \
157 _bbci.l a3, 1, 3f /* remaining 2 !! 219 st t0, [dst + off + 0x00]; \
158 l8ui a6, a2, 0 !! 220 addxcc t0, sum, sum; \
159 l8ui a7, a2, 1 !! 221 st t1, [dst + off + 0x04]; \
160 #ifdef __XTENSA_EB__ !! 222 addxcc t1, sum, sum; \
161 slli a6, a6, 8 !! 223 st t2, [dst + off + 0x08]; \
162 #else !! 224 addxcc t2, sum, sum; \
163 slli a7, a7, 8 !! 225 st t3, [dst + off + 0x0c]; \
164 #endif !! 226 addxcc t3, sum, sum; \
165 or a7, a7, a6 !! 227 st t4, [dst + off + 0x10]; \
166 ONES_ADD(a4, a7) !! 228 addxcc t4, sum, sum; \
167 addi a2, a2, 2 !! 229 st t5, [dst + off + 0x14]; \
168 3: !! 230 addxcc t5, sum, sum; \
169 j 5b /* branch to h !! 231 st t6, [dst + off + 0x18]; \
170 !! 232 addxcc t6, sum, sum; \
171 ENDPROC(csum_partial) !! 233 st t7, [dst + off + 0x1c]; \
172 EXPORT_SYMBOL(csum_partial) !! 234 addxcc t7, sum, sum;
173 !! 235
174 /* !! 236 /* Yuck, 6 superscalar cycles... */
175 * Copy from ds while checksumming, otherwise !! 237 #define CSUMCOPY_LASTCHUNK(src, dst, sum, off, t0, t1, t2, t3) \
176 */ !! 238 ldd [src - off - 0x08], t0; \
177 !! 239 ldd [src - off - 0x00], t2; \
178 /* !! 240 addxcc t0, sum, sum; \
179 unsigned int csum_partial_copy_generic (const !! 241 st t0, [dst - off - 0x08]; \
180 a2 = src !! 242 addxcc t1, sum, sum; \
181 a3 = dst !! 243 st t1, [dst - off - 0x04]; \
182 a4 = len !! 244 addxcc t2, sum, sum; \
183 a5 = sum !! 245 st t2, [dst - off - 0x00]; \
184 a8 = temp !! 246 addxcc t3, sum, sum; \
185 a9 = temp !! 247 st t3, [dst - off + 0x04];
186 a10 = temp !! 248
187 !! 249 /* Handle the end cruft code out of band for better cache patterns. */
188 This function is optimized for 4-byte alig !! 250 cc_end_cruft:
189 alignments work, but not nearly as efficie !! 251 be 1f
190 */ !! 252 andcc %o3, 4, %g0
191 !! 253 EX(ldd [%o0 + 0x00], %g2, and %o3, 0xf)
192 ENTRY(csum_partial_copy_generic) !! 254 add %o1, 8, %o1
193 !! 255 addcc %g2, %g7, %g7
194 abi_entry_default !! 256 add %o0, 8, %o0
195 movi a5, -1 !! 257 addxcc %g3, %g7, %g7
196 or a10, a2, a3 !! 258 EX2(st %g2, [%o1 - 0x08])
197 !! 259 addx %g0, %g7, %g7
198 /* We optimize the following alignment !! 260 andcc %o3, 4, %g0
199 aligned case. Two bbsi.l instructions !! 261 EX2(st %g3, [%o1 - 0x04])
200 (commented out below). However, both !! 262 1: be 1f
201 of the imm8 range, so the assembler re !! 263 andcc %o3, 3, %o3
202 equivalent bbci.l, j combinations, whi !! 264 EX(ld [%o0 + 0x00], %g2, add %o3, 4)
203 slower. */ !! 265 add %o1, 4, %o1
204 !! 266 addcc %g2, %g7, %g7
205 extui a9, a10, 0, 2 !! 267 EX2(st %g2, [%o1 - 0x04])
206 beqz a9, 1f /* branch if b !! 268 addx %g0, %g7, %g7
207 bbsi.l a10, 0, 5f /* branch if o !! 269 andcc %o3, 3, %g0
208 j 3f /* one address !! 270 add %o0, 4, %o0
209 !! 271 1: be 1f
210 /* _bbsi.l a10, 0, 5f */ /* branch if o !! 272 addcc %o3, -1, %g0
211 /* _bbsi.l a10, 1, 3f */ /* branch if 2 !! 273 bne 2f
212 !! 274 subcc %o3, 2, %o3
>> 275 b 4f
>> 276 or %g0, %g0, %o4
>> 277 2: EX(lduh [%o0 + 0x00], %o4, add %o3, 2)
>> 278 add %o0, 2, %o0
>> 279 EX2(sth %o4, [%o1 + 0x00])
>> 280 be 6f
>> 281 add %o1, 2, %o1
>> 282 sll %o4, 16, %o4
>> 283 4: EX(ldub [%o0 + 0x00], %o5, add %g0, 1)
>> 284 EX2(stb %o5, [%o1 + 0x00])
>> 285 sll %o5, 8, %o5
>> 286 or %o5, %o4, %o4
>> 287 6: addcc %o4, %g7, %g7
>> 288 1: retl
>> 289 addx %g0, %g7, %o0
>> 290
>> 291 /* Also, handle the alignment code out of band. */
>> 292 cc_dword_align:
>> 293 cmp %g1, 6
>> 294 bl,a ccte
>> 295 andcc %g1, 0xf, %o3
>> 296 andcc %o0, 0x1, %g0
>> 297 bne ccslow
>> 298 andcc %o0, 0x2, %g0
>> 299 be 1f
>> 300 andcc %o0, 0x4, %g0
>> 301 EX(lduh [%o0 + 0x00], %g4, add %g1, 0)
>> 302 sub %g1, 2, %g1
>> 303 EX2(sth %g4, [%o1 + 0x00])
>> 304 add %o0, 2, %o0
>> 305 sll %g4, 16, %g4
>> 306 addcc %g4, %g7, %g7
>> 307 add %o1, 2, %o1
>> 308 srl %g7, 16, %g3
>> 309 addx %g0, %g3, %g4
>> 310 sll %g7, 16, %g7
>> 311 sll %g4, 16, %g3
>> 312 srl %g7, 16, %g7
>> 313 andcc %o0, 0x4, %g0
>> 314 or %g3, %g7, %g7
>> 315 1: be 3f
>> 316 andcc %g1, 0xffffff80, %g0
>> 317 EX(ld [%o0 + 0x00], %g4, add %g1, 0)
>> 318 sub %g1, 4, %g1
>> 319 EX2(st %g4, [%o1 + 0x00])
>> 320 add %o0, 4, %o0
>> 321 addcc %g4, %g7, %g7
>> 322 add %o1, 4, %o1
>> 323 addx %g0, %g7, %g7
>> 324 b 3f
>> 325 andcc %g1, 0xffffff80, %g0
>> 326
>> 327 /* Sun, you just can't beat me, you just can't. Stop trying,
>> 328 * give up. I'm serious, I am going to kick the living shit
>> 329 * out of you, game over, lights out.
>> 330 */
>> 331 .align 8
>> 332 .globl C_LABEL(__csum_partial_copy_sparc_generic)
>> 333 C_LABEL(__csum_partial_copy_sparc_generic):
>> 334 /* %o0=src, %o1=dest, %g1=len, %g7=sum */
>> 335 xor %o0, %o1, %o4 ! get changing bits
>> 336 andcc %o4, 3, %g0 ! check for mismatched alignment
>> 337 bne ccslow ! better this than unaligned/fixups
>> 338 andcc %o0, 7, %g0 ! need to align things?
>> 339 bne cc_dword_align ! yes, we check for short lengths there
>> 340 andcc %g1, 0xffffff80, %g0 ! can we use unrolled loop?
>> 341 3: be 3f ! nope, less than one loop remains
>> 342 andcc %o1, 4, %g0 ! dest aligned on 4 or 8 byte boundry?
>> 343 be ccdbl + 4 ! 8 byte aligned, kick ass
>> 344 5: CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
>> 345 CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
>> 346 CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
>> 347 CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
>> 348 10: EXT(5b, 10b, 20f) ! note for exception handling
>> 349 sub %g1, 128, %g1 ! detract from length
>> 350 addx %g0, %g7, %g7 ! add in last carry bit
>> 351 andcc %g1, 0xffffff80, %g0 ! more to csum?
>> 352 add %o0, 128, %o0 ! advance src ptr
>> 353 bne 5b ! we did not go negative, continue looping
>> 354 add %o1, 128, %o1 ! advance dest ptr
>> 355 3: andcc %g1, 0x70, %o2 ! can use table?
>> 356 ccmerge:be ccte ! nope, go and check for end cruft
>> 357 andcc %g1, 0xf, %o3 ! get low bits of length (clears carry btw)
>> 358 srl %o2, 1, %o4 ! begin negative offset computation
>> 359 sethi %hi(12f), %o5 ! set up table ptr end
>> 360 add %o0, %o2, %o0 ! advance src ptr
>> 361 sub %o5, %o4, %o5 ! continue table calculation
>> 362 sll %o2, 1, %g2 ! constant multiplies are fun...
>> 363 sub %o5, %g2, %o5 ! some more adjustments
>> 364 jmp %o5 + %lo(12f) ! jump into it, duff style, wheee...
>> 365 add %o1, %o2, %o1 ! advance dest ptr (carry is clear btw)
>> 366 cctbl: CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x68,%g2,%g3,%g4,%g5)
>> 367 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x58,%g2,%g3,%g4,%g5)
>> 368 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x48,%g2,%g3,%g4,%g5)
>> 369 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x38,%g2,%g3,%g4,%g5)
>> 370 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x28,%g2,%g3,%g4,%g5)
>> 371 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x18,%g2,%g3,%g4,%g5)
>> 372 CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x08,%g2,%g3,%g4,%g5)
>> 373 12: EXT(cctbl, 12b, 22f) ! note for exception table handling
>> 374 addx %g0, %g7, %g7
>> 375 andcc %o3, 0xf, %g0 ! check for low bits set
>> 376 ccte: bne cc_end_cruft ! something left, handle it out of band
>> 377 andcc %o3, 8, %g0 ! begin checks for that code
>> 378 retl ! return
>> 379 mov %g7, %o0 ! give em the computed checksum
>> 380 ccdbl: CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
>> 381 CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
>> 382 CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
>> 383 CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
>> 384 11: EXT(ccdbl, 11b, 21f) ! note for exception table handling
>> 385 sub %g1, 128, %g1 ! detract from length
>> 386 addx %g0, %g7, %g7 ! add in last carry bit
>> 387 andcc %g1, 0xffffff80, %g0 ! more to csum?
>> 388 add %o0, 128, %o0 ! advance src ptr
>> 389 bne ccdbl ! we did not go negative, continue looping
>> 390 add %o1, 128, %o1 ! advance dest ptr
>> 391 b ccmerge ! finish it off, above
>> 392 andcc %g1, 0x70, %o2 ! can use table? (clears carry btw)
>> 393
>> 394 ccslow: cmp %g1, 0
>> 395 mov 0, %g5
>> 396 bleu 4f
>> 397 andcc %o0, 1, %o5
>> 398 be,a 1f
>> 399 srl %g1, 1, %g4
>> 400 sub %g1, 1, %g1
>> 401 EX(ldub [%o0], %g5, add %g1, 1)
>> 402 add %o0, 1, %o0
>> 403 EX2(stb %g5, [%o1])
>> 404 srl %g1, 1, %g4
>> 405 add %o1, 1, %o1
>> 406 1: cmp %g4, 0
>> 407 be,a 3f
>> 408 andcc %g1, 1, %g0
>> 409 andcc %o0, 2, %g0
>> 410 be,a 1f
>> 411 srl %g4, 1, %g4
>> 412 EX(lduh [%o0], %o4, add %g1, 0)
>> 413 sub %g1, 2, %g1
>> 414 srl %o4, 8, %g2
>> 415 sub %g4, 1, %g4
>> 416 EX2(stb %g2, [%o1])
>> 417 add %o4, %g5, %g5
>> 418 EX2(stb %o4, [%o1 + 1])
>> 419 add %o0, 2, %o0
>> 420 srl %g4, 1, %g4
>> 421 add %o1, 2, %o1
>> 422 1: cmp %g4, 0
>> 423 be,a 2f
>> 424 andcc %g1, 2, %g0
>> 425 EX3(ld [%o0], %o4)
>> 426 5: srl %o4, 24, %g2
>> 427 srl %o4, 16, %g3
>> 428 EX2(stb %g2, [%o1])
>> 429 srl %o4, 8, %g2
>> 430 EX2(stb %g3, [%o1 + 1])
>> 431 add %o0, 4, %o0
>> 432 EX2(stb %g2, [%o1 + 2])
>> 433 addcc %o4, %g5, %g5
>> 434 EX2(stb %o4, [%o1 + 3])
>> 435 addx %g5, %g0, %g5 ! I am now to lazy to optimize this (question it
>> 436 add %o1, 4, %o1 ! is worthy). Maybe some day - with the sll/srl
>> 437 subcc %g4, 1, %g4 ! tricks
>> 438 bne,a 5b
>> 439 EX3(ld [%o0], %o4)
>> 440 sll %g5, 16, %g2
>> 441 srl %g5, 16, %g5
>> 442 srl %g2, 16, %g2
>> 443 andcc %g1, 2, %g0
>> 444 add %g2, %g5, %g5
>> 445 2: be,a 3f
>> 446 andcc %g1, 1, %g0
>> 447 EX(lduh [%o0], %o4, and %g1, 3)
>> 448 andcc %g1, 1, %g0
>> 449 srl %o4, 8, %g2
>> 450 add %o0, 2, %o0
>> 451 EX2(stb %g2, [%o1])
>> 452 add %g5, %o4, %g5
>> 453 EX2(stb %o4, [%o1 + 1])
>> 454 add %o1, 2, %o1
>> 455 3: be,a 1f
>> 456 sll %g5, 16, %o4
>> 457 EX(ldub [%o0], %g2, add %g0, 1)
>> 458 sll %g2, 8, %o4
>> 459 EX2(stb %g2, [%o1])
>> 460 add %g5, %o4, %g5
>> 461 sll %g5, 16, %o4
>> 462 1: addcc %o4, %g5, %g5
>> 463 srl %g5, 16, %o4
>> 464 addx %g0, %o4, %g5
>> 465 orcc %o5, %g0, %g0
>> 466 be 4f
>> 467 srl %g5, 8, %o4
>> 468 and %g5, 0xff, %g2
>> 469 and %o4, 0xff, %o4
>> 470 sll %g2, 8, %g2
>> 471 or %g2, %o4, %g5
>> 472 4: addcc %g7, %g5, %g7
>> 473 retl
>> 474 addx %g0, %g7, %o0
>> 475 C_LABEL(__csum_partial_copy_end):
>> 476
>> 477 /* We do these strange calculations for the csum_*_from_user case only, ie.
>> 478 * we only bother with faults on loads... */
>> 479
>> 480 /* o2 = ((g2%20)&3)*8
>> 481 * o3 = g1 - (g2/20)*32 - o2 */
>> 482 20:
>> 483 cmp %g2, 20
>> 484 blu,a 1f
>> 485 and %g2, 3, %o2
>> 486 sub %g1, 32, %g1
>> 487 b 20b
>> 488 sub %g2, 20, %g2
213 1: 489 1:
214 /* src and dst are both 4-byte aligned !! 490 sll %o2, 3, %o2
215 srli a10, a4, 5 /* 32-byte chu !! 491 b 31f
216 #if XCHAL_HAVE_LOOPS !! 492 sub %g1, %o2, %o3
217 loopgtz a10, 2f !! 493
218 #else !! 494 /* o2 = (!(g2 & 15) ? 0 : (((g2 & 15) + 1) & ~1)*8)
219 beqz a10, 2f !! 495 * o3 = g1 - (g2/16)*32 - o2 */
220 slli a10, a10, 5 !! 496 21:
221 add a10, a10, a2 /* a10 = end o !! 497 andcc %g2, 15, %o3
222 .Loop5: !! 498 srl %g2, 4, %g2
223 #endif !! 499 be,a 1f
224 EX(10f) l32i a9, a2, 0 !! 500 clr %o2
225 EX(10f) l32i a8, a2, 4 !! 501 add %o3, 1, %o3
226 EX(10f) s32i a9, a3, 0 !! 502 and %o3, 14, %o3
227 EX(10f) s32i a8, a3, 4 !! 503 sll %o3, 3, %o2
228 ONES_ADD(a5, a9) !! 504 1:
229 ONES_ADD(a5, a8) !! 505 sll %g2, 5, %g2
230 EX(10f) l32i a9, a2, 8 !! 506 sub %g1, %g2, %o3
231 EX(10f) l32i a8, a2, 12 !! 507 b 31f
232 EX(10f) s32i a9, a3, 8 !! 508 sub %o3, %o2, %o3
233 EX(10f) s32i a8, a3, 12 !! 509
234 ONES_ADD(a5, a9) !! 510 /* o0 += (g2/10)*16 - 0x70
235 ONES_ADD(a5, a8) !! 511 * 01 += (g2/10)*16 - 0x70
236 EX(10f) l32i a9, a2, 16 !! 512 * o2 = (g2 % 10) ? 8 : 0
237 EX(10f) l32i a8, a2, 20 !! 513 * o3 += 0x70 - (g2/10)*16 - o2 */
238 EX(10f) s32i a9, a3, 16 !! 514 22:
239 EX(10f) s32i a8, a3, 20 !! 515 cmp %g2, 10
240 ONES_ADD(a5, a9) !! 516 blu,a 1f
241 ONES_ADD(a5, a8) !! 517 sub %o0, 0x70, %o0
242 EX(10f) l32i a9, a2, 24 !! 518 add %o0, 16, %o0
243 EX(10f) l32i a8, a2, 28 !! 519 add %o1, 16, %o1
244 EX(10f) s32i a9, a3, 24 !! 520 sub %o3, 16, %o3
245 EX(10f) s32i a8, a3, 28 !! 521 b 22b
246 ONES_ADD(a5, a9) !! 522 sub %g2, 10, %g2
247 ONES_ADD(a5, a8) !! 523 1:
248 addi a2, a2, 32 !! 524 sub %o1, 0x70, %o1
249 addi a3, a3, 32 !! 525 add %o3, 0x70, %o3
250 #if !XCHAL_HAVE_LOOPS !! 526 clr %o2
251 blt a2, a10, .Loop5 !! 527 tst %g2
252 #endif !! 528 bne,a 1f
253 2: !! 529 mov 8, %o2
254 extui a10, a4, 2, 3 /* remaining 4 !! 530 1:
255 extui a4, a4, 0, 2 /* reset len f !! 531 b 31f
256 #if XCHAL_HAVE_LOOPS !! 532 sub %o3, %o2, %o3
257 loopgtz a10, 3f !! 533 96:
258 #else !! 534 and %g1, 3, %g1
259 beqz a10, 3f !! 535 sll %g4, 2, %g4
260 slli a10, a10, 2 !! 536 add %g1, %g4, %o3
261 add a10, a10, a2 /* a10 = end o !! 537 30:
262 .Loop6: !! 538 /* %o1 is dst
263 #endif !! 539 * %o3 is # bytes to zero out
264 EX(10f) l32i a9, a2, 0 !! 540 * %o4 is faulting address
265 EX(10f) s32i a9, a3, 0 !! 541 * %o5 is %pc where fault occurred */
266 ONES_ADD(a5, a9) !! 542 clr %o2
267 addi a2, a2, 4 !! 543 31:
268 addi a3, a3, 4 !! 544 /* %o0 is src
269 #if !XCHAL_HAVE_LOOPS !! 545 * %o1 is dst
270 blt a2, a10, .Loop6 !! 546 * %o2 is # of bytes to copy from src to dst
271 #endif !! 547 * %o3 is # bytes to zero out
272 3: !! 548 * %o4 is faulting address
273 /* !! 549 * %o5 is %pc where fault occurred */
274 Control comes to here in two cases: (1 !! 550 save %sp, -104, %sp
275 to here from the 4-byte alignment case !! 551 mov %i5, %o0
276 one 2-byte chunk. (2) It branches to !! 552 mov %i7, %o1
277 either src or dst is 2-byte aligned, a !! 553 mov %i4, %o2
278 here, except for perhaps a trailing od !! 554 call C_LABEL(lookup_fault)
279 inefficient, so align your addresses t !! 555 mov %g7, %i4
280 !! 556 cmp %o0, 2
281 a2 = src !! 557 bne 1f
282 a3 = dst !! 558 add %g0, -EFAULT, %i5
283 a4 = len !! 559 tst %i2
284 a5 = sum !! 560 be 2f
285 */ !! 561 mov %i0, %o1
286 srli a10, a4, 1 /* 2-byte chun !! 562 mov %i1, %o0
287 #if XCHAL_HAVE_LOOPS <<
288 loopgtz a10, 4f <<
289 #else <<
290 beqz a10, 4f <<
291 slli a10, a10, 1 <<
292 add a10, a10, a2 /* a10 = end o <<
293 .Loop7: <<
294 #endif <<
295 EX(10f) l16ui a9, a2, 0 <<
296 EX(10f) s16i a9, a3, 0 <<
297 ONES_ADD(a5, a9) <<
298 addi a2, a2, 2 <<
299 addi a3, a3, 2 <<
300 #if !XCHAL_HAVE_LOOPS <<
301 blt a2, a10, .Loop7 <<
302 #endif <<
303 4: <<
304 /* This section processes a possible t <<
305 _bbci.l a4, 0, 8f /* 1-byte chun <<
306 EX(10f) l8ui a9, a2, 0 <<
307 EX(10f) s8i a9, a3, 0 <<
308 #ifdef __XTENSA_EB__ <<
309 slli a9, a9, 8 /* shift byte <<
310 #endif <<
311 ONES_ADD(a5, a9) <<
312 8: <<
313 mov a2, a5 <<
314 abi_ret_default <<
315 <<
316 5: 563 5:
317 /* Control branch to here when either !! 564 call C_LABEL(__memcpy)
318 process all bytes using 8-bit accesses !! 565 mov %i2, %o2
319 so don't feed us an odd address. */ !! 566 tst %o0
320 !! 567 bne,a 2f
321 srli a10, a4, 1 /* handle in p !! 568 add %i3, %i2, %i3
322 #if XCHAL_HAVE_LOOPS !! 569 add %i1, %i2, %i1
323 loopgtz a10, 6f !! 570 2:
324 #else !! 571 mov %i1, %o0
325 beqz a10, 6f <<
326 slli a10, a10, 1 <<
327 add a10, a10, a2 /* a10 = end o <<
328 .Loop8: <<
329 #endif <<
330 EX(10f) l8ui a9, a2, 0 <<
331 EX(10f) l8ui a8, a2, 1 <<
332 EX(10f) s8i a9, a3, 0 <<
333 EX(10f) s8i a8, a3, 1 <<
334 #ifdef __XTENSA_EB__ <<
335 slli a9, a9, 8 /* combine int <<
336 #else /* for checksu <<
337 slli a8, a8, 8 <<
338 #endif <<
339 or a9, a9, a8 <<
340 ONES_ADD(a5, a9) <<
341 addi a2, a2, 2 <<
342 addi a3, a3, 2 <<
343 #if !XCHAL_HAVE_LOOPS <<
344 blt a2, a10, .Loop8 <<
345 #endif <<
346 6: 572 6:
347 j 4b /* process the !! 573 call C_LABEL(__bzero)
348 !! 574 mov %i3, %o1
349 ENDPROC(csum_partial_copy_generic) !! 575 1:
350 EXPORT_SYMBOL(csum_partial_copy_generic) !! 576 ld [%sp + 168], %o2 ! struct_ptr of parent
351 !! 577 st %i5, [%o2]
352 !! 578 ret
353 # Exception handler: !! 579 restore
354 .section .fixup, "ax" !! 580
355 10: !! 581 .section __ex_table,#alloc
356 movi a2, 0 !! 582 .align 4
357 abi_ret_default !! 583 .word 5b,2
358 !! 584 .word 6b,2
359 .previous <<
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.