| 1 |
1275 |
phoenix |
/* $Id: VIScsumcopyusr.S,v 1.1.1.1 2004-04-15 01:33:52 phoenix Exp $
|
| 2 |
|
|
* VIScsumcopyusr.S: High bandwidth IP checksumming with simultaneous
|
| 3 |
|
|
* copying utilizing the UltraSparc Visual Instruction Set.
|
| 4 |
|
|
*
|
| 5 |
|
|
* Copyright (C) 1997, 1999 Jakub Jelinek (jj@ultra.linux.cz)
|
| 6 |
|
|
* Copyright (C) 2000 David S. Miller (davem@redhat.com)
|
| 7 |
|
|
*
|
| 8 |
|
|
* Based on older sparc32/sparc64 checksum.S, which is:
|
| 9 |
|
|
*
|
| 10 |
|
|
* Copyright(C) 1995 Linus Torvalds
|
| 11 |
|
|
* Copyright(C) 1995 Miguel de Icaza
|
| 12 |
|
|
* Copyright(C) 1996,1997 David S. Miller
|
| 13 |
|
|
* derived from:
|
| 14 |
|
|
* Linux/Alpha checksum c-code
|
| 15 |
|
|
* Linux/ix86 inline checksum assembly
|
| 16 |
|
|
* RFC1071 Computing the Internet Checksum (esp. Jacobsons m68k code)
|
| 17 |
|
|
* David Mosberger-Tang for optimized reference c-code
|
| 18 |
|
|
* BSD4.4 portable checksum routine
|
| 19 |
|
|
*/
|
| 20 |
|
|
|
| 21 |
|
|
#ifdef __sparc_v9__
|
| 22 |
|
|
#define STACKOFF 0x7ff+128
|
| 23 |
|
|
#else
|
| 24 |
|
|
#define STACKOFF 64
|
| 25 |
|
|
#endif
|
| 26 |
|
|
|
| 27 |
|
|
#ifdef __KERNEL__
|
| 28 |
|
|
#include
|
| 29 |
|
|
#include
|
| 30 |
|
|
#include
|
| 31 |
|
|
#include
|
| 32 |
|
|
#define ASI_BLK_XOR 0
|
| 33 |
|
|
#define ASI_BLK_XOR1 (ASI_BLK_P ^ (ASI_BLK_P >> 3) ^ ASI_P)
|
| 34 |
|
|
#define ASI_BLK_OR (ASI_BLK_P & ~ASI_P)
|
| 35 |
|
|
#else
|
| 36 |
|
|
#define ASI_P 0x80
|
| 37 |
|
|
#define ASI_BLK_P 0xf0
|
| 38 |
|
|
#define FRPS_FEF 0x04
|
| 39 |
|
|
#define FPRS_DU 0x02
|
| 40 |
|
|
#define FPRS_DL 0x01
|
| 41 |
|
|
#define ASI_BLK_XOR (ASI_BLK_P ^ ASI_P)
|
| 42 |
|
|
#endif
|
| 43 |
|
|
|
| 44 |
|
|
#define src o0
|
| 45 |
|
|
#define dst o1
|
| 46 |
|
|
#define len o2
|
| 47 |
|
|
#define sum o3
|
| 48 |
|
|
#define x1 g1
|
| 49 |
|
|
#define x2 g2
|
| 50 |
|
|
#define x3 o4
|
| 51 |
|
|
#define x4 g4
|
| 52 |
|
|
#define x5 g5
|
| 53 |
|
|
#define x6 g7
|
| 54 |
|
|
#define x7 g3
|
| 55 |
|
|
#define x8 o5
|
| 56 |
|
|
|
| 57 |
|
|
/* Dobrou noc, SunSoft engineers. Spete sladce.
|
| 58 |
|
|
* This has a couple of tricks in and those
|
| 59 |
|
|
* tricks are UltraLinux trade secrets :))
|
| 60 |
|
|
* Once AGAIN, the SunSoft engineers are caught
|
| 61 |
|
|
* asleep at the keyboard :)).
|
| 62 |
|
|
* The main loop does about 20 superscalar cycles
|
| 63 |
|
|
* per 64bytes checksummed/copied.
|
| 64 |
|
|
*/
|
| 65 |
|
|
|
| 66 |
|
|
#define LDBLK(O0) \
|
| 67 |
|
|
ldda [%src] ASI_BLK_P, %O0 /* Load Group */
|
| 68 |
|
|
|
| 69 |
|
|
#define STBLK \
|
| 70 |
|
|
stda %f48, [%dst] %asi /* Store */
|
| 71 |
|
|
|
| 72 |
|
|
#ifdef __KERNEL__
|
| 73 |
|
|
#define STBLK_XORASI(tmpreg1,tmpreg2) \
|
| 74 |
|
|
stda %f48, [%dst] %asi /* Store */; \
|
| 75 |
|
|
rd %asi, %tmpreg1; \
|
| 76 |
|
|
srl %tmpreg1, 3, %tmpreg2; \
|
| 77 |
|
|
xor %tmpreg1, ASI_BLK_XOR1, %tmpreg1; \
|
| 78 |
|
|
wr %tmpreg1, %tmpreg2, %asi;
|
| 79 |
|
|
#else
|
| 80 |
|
|
#define STBLK_XORASI(tmpreg1,tmpreg2) \
|
| 81 |
|
|
stda %f48, [%dst] %asi /* Store */; \
|
| 82 |
|
|
rd %asi, %tmpreg1; \
|
| 83 |
|
|
wr %tmpreg1, ASI_BLK_XOR, %asi;
|
| 84 |
|
|
#endif
|
| 85 |
|
|
|
| 86 |
|
|
#define ST(fx,off) \
|
| 87 |
|
|
stda %fx, [%dst + off] %asi /* Store */
|
| 88 |
|
|
|
| 89 |
|
|
#define SYNC \
|
| 90 |
|
|
membar #Sync
|
| 91 |
|
|
|
| 92 |
|
|
|
| 93 |
|
|
#define DO_THE_TRICK(f0,f2,f4,f6,f8,f10,f12,f14,F0,F2,F4,F6,F8,F10,F12,F14,DUMMY1,A0,A2,A4,A6,A8,A10,A12,A14,B14,DUMMY2,LOAD,STORE1,STORE2,STORE3,STORE4,STORE5,STORE6,STORE7,STORE8,DUMMY3,BRANCH...) \
|
| 94 |
|
|
LOAD /* Load (Group) */; \
|
| 95 |
|
|
faligndata %A14, %F0, %A14 /* FPA Group */; \
|
| 96 |
|
|
inc %x5 /* IEU0 */; \
|
| 97 |
|
|
STORE1 /* Store (optional) */; \
|
| 98 |
|
|
faligndata %F0, %F2, %A0 /* FPA Group */; \
|
| 99 |
|
|
srl %x5, 1, %x5 /* IEU0 */; \
|
| 100 |
|
|
add %sum, %x4, %sum /* IEU1 */; \
|
| 101 |
|
|
fpadd32 %F0, %f0, %F0 /* FPA Group */; \
|
| 102 |
|
|
inc %x6 /* IEU0 */; \
|
| 103 |
|
|
STORE2 /* Store (optional) */; \
|
| 104 |
|
|
faligndata %F2, %F4, %A2 /* FPA Group */; \
|
| 105 |
|
|
srl %x6, 1, %x6 /* IEU0 */; \
|
| 106 |
|
|
add %sum, %x5, %sum /* IEU1 */; \
|
| 107 |
|
|
fpadd32 %F2, %f2, %F2 /* FPA Group */; \
|
| 108 |
|
|
add %src, 64, %src /* IEU0 */; \
|
| 109 |
|
|
fcmpgt32 %f0, %F0, %x1 /* FPM */; \
|
| 110 |
|
|
add %dst, 64, %dst /* IEU1 Group */; \
|
| 111 |
|
|
inc %x7 /* IEU0 */; \
|
| 112 |
|
|
STORE3 /* Store (optional) */; \
|
| 113 |
|
|
faligndata %F4, %F6, %A4 /* FPA */; \
|
| 114 |
|
|
fpadd32 %F4, %f4, %F4 /* FPA Group */; \
|
| 115 |
|
|
add %sum, %x6, %sum /* IEU1 */; \
|
| 116 |
|
|
fcmpgt32 %f2, %F2, %x2 /* FPM */; \
|
| 117 |
|
|
srl %x7, 1, %x7 /* IEU0 Group */; \
|
| 118 |
|
|
inc %x8 /* IEU1 */; \
|
| 119 |
|
|
STORE4 /* Store (optional) */; \
|
| 120 |
|
|
faligndata %F6, %F8, %A6 /* FPA */; \
|
| 121 |
|
|
fpadd32 %F6, %f6, %F6 /* FPA Group */; \
|
| 122 |
|
|
srl %x8, 1, %x8 /* IEU0 */; \
|
| 123 |
|
|
fcmpgt32 %f4, %F4, %x3 /* FPM */; \
|
| 124 |
|
|
add %sum, %x7, %sum /* IEU0 Group */; \
|
| 125 |
|
|
inc %x1 /* IEU1 */; \
|
| 126 |
|
|
STORE5 /* Store (optional) */; \
|
| 127 |
|
|
faligndata %F8, %F10, %A8 /* FPA */; \
|
| 128 |
|
|
fpadd32 %F8, %f8, %F8 /* FPA Group */; \
|
| 129 |
|
|
srl %x1, 1, %x1 /* IEU0 */; \
|
| 130 |
|
|
fcmpgt32 %f6, %F6, %x4 /* FPM */; \
|
| 131 |
|
|
add %sum, %x8, %sum /* IEU0 Group */; \
|
| 132 |
|
|
inc %x2 /* IEU1 */; \
|
| 133 |
|
|
STORE6 /* Store (optional) */; \
|
| 134 |
|
|
faligndata %F10, %F12, %A10 /* FPA */; \
|
| 135 |
|
|
fpadd32 %F10, %f10, %F10 /* FPA Group */; \
|
| 136 |
|
|
srl %x2, 1, %x2 /* IEU0 */; \
|
| 137 |
|
|
fcmpgt32 %f8, %F8, %x5 /* FPM */; \
|
| 138 |
|
|
add %sum, %x1, %sum /* IEU0 Group */; \
|
| 139 |
|
|
inc %x3 /* IEU1 */; \
|
| 140 |
|
|
STORE7 /* Store (optional) */; \
|
| 141 |
|
|
faligndata %F12, %F14, %A12 /* FPA */; \
|
| 142 |
|
|
fpadd32 %F12, %f12, %F12 /* FPA Group */; \
|
| 143 |
|
|
srl %x3, 1, %x3 /* IEU0 */; \
|
| 144 |
|
|
fcmpgt32 %f10, %F10, %x6 /* FPM */; \
|
| 145 |
|
|
add %sum, %x2, %sum /* IEU0 Group */; \
|
| 146 |
|
|
inc %x4 /* IEU1 */; \
|
| 147 |
|
|
STORE8 /* Store (optional) */; \
|
| 148 |
|
|
fmovd %F14, %B14 /* FPA */; \
|
| 149 |
|
|
fpadd32 %F14, %f14, %F14 /* FPA Group */; \
|
| 150 |
|
|
srl %x4, 1, %x4 /* IEU0 */; \
|
| 151 |
|
|
fcmpgt32 %f12, %F12, %x7 /* FPM */; \
|
| 152 |
|
|
add %sum, %x3, %sum /* IEU0 Group */; \
|
| 153 |
|
|
subcc %len, 64, %len /* IEU1 */; \
|
| 154 |
|
|
BRANCH /* CTI */; \
|
| 155 |
|
|
fcmpgt32 %f14, %F14, %x8 /* FPM Group */;
|
| 156 |
|
|
|
| 157 |
|
|
#define END_THE_TRICK(f0,f2,f4,f6,f8,f10,f12,f14,FA,FB,S0,S1,S2,S3,T0,T1,U0,fz) \
|
| 158 |
|
|
inc %x5 /* IEU0 Group */; \
|
| 159 |
|
|
fpadd32 %f2, %f0, %S0 /* FPA */; \
|
| 160 |
|
|
add %sum, %x4, %sum /* IEU1 */; \
|
| 161 |
|
|
srl %x5, 1, %x5 /* IEU0 Group */; \
|
| 162 |
|
|
fpadd32 %f6, %f4, %S1 /* FPA */; \
|
| 163 |
|
|
inc %x6 /* IEU1 */; \
|
| 164 |
|
|
fpadd32 %f10, %f8, %S2 /* FPA Group */; \
|
| 165 |
|
|
add %sum, %x5, %sum /* IEU0 */; \
|
| 166 |
|
|
fcmpgt32 %f0, %S0, %x1 /* FPM */; \
|
| 167 |
|
|
fpadd32 %f14, %f12, %S3 /* FPA Group */; \
|
| 168 |
|
|
srl %x6, 1, %x6 /* IEU0 */; \
|
| 169 |
|
|
fcmpgt32 %f4, %S1, %x2 /* FPM */; \
|
| 170 |
|
|
add %sum, %x6, %sum /* IEU0 Group */; \
|
| 171 |
|
|
fzero %fz /* FPA */; \
|
| 172 |
|
|
fcmpgt32 %f8, %S2, %x3 /* FPM */; \
|
| 173 |
|
|
inc %x7 /* IEU0 Group */; \
|
| 174 |
|
|
inc %x8 /* IEU1 */; \
|
| 175 |
|
|
srl %x7, 1, %x7 /* IEU0 Group */; \
|
| 176 |
|
|
inc %x1 /* IEU1 */; \
|
| 177 |
|
|
fpadd32 %S0, %S1, %T0 /* FPA */; \
|
| 178 |
|
|
fpadd32 %S2, %S3, %T1 /* FPA Group */; \
|
| 179 |
|
|
add %sum, %x7, %sum /* IEU0 */; \
|
| 180 |
|
|
fcmpgt32 %f12, %S3, %x4 /* FPM */; \
|
| 181 |
|
|
srl %x8, 1, %x8 /* IEU0 Group */; \
|
| 182 |
|
|
inc %x2 /* IEU1 */; \
|
| 183 |
|
|
srl %x1, 1, %x1 /* IEU0 Group */; \
|
| 184 |
|
|
add %sum, %x8, %sum /* IEU1 */; \
|
| 185 |
|
|
add %sum, %x1, %sum /* IEU0 Group */; \
|
| 186 |
|
|
fcmpgt32 %S0, %T0, %x5 /* FPM */; \
|
| 187 |
|
|
srl %x2, 1, %x2 /* IEU0 Group */; \
|
| 188 |
|
|
fcmpgt32 %S2, %T1, %x6 /* FPM */; \
|
| 189 |
|
|
inc %x3 /* IEU0 Group */; \
|
| 190 |
|
|
add %sum, %x2, %sum /* IEU1 */; \
|
| 191 |
|
|
srl %x3, 1, %x3 /* IEU0 Group */; \
|
| 192 |
|
|
inc %x4 /* IEU1 */; \
|
| 193 |
|
|
fpadd32 %T0, %T1, %U0 /* FPA Group */; \
|
| 194 |
|
|
add %sum, %x3, %sum /* IEU0 */; \
|
| 195 |
|
|
fcmpgt32 %fz, %f2, %x7 /* FPM */; \
|
| 196 |
|
|
srl %x4, 1, %x4 /* IEU0 Group */; \
|
| 197 |
|
|
fcmpgt32 %fz, %f6, %x8 /* FPM */; \
|
| 198 |
|
|
inc %x5 /* IEU0 Group */; \
|
| 199 |
|
|
add %sum, %x4, %sum /* IEU1 */; \
|
| 200 |
|
|
srl %x5, 1, %x5 /* IEU0 Group */; \
|
| 201 |
|
|
fcmpgt32 %fz, %f10, %x1 /* FPM */; \
|
| 202 |
|
|
inc %x6 /* IEU0 Group */; \
|
| 203 |
|
|
add %sum, %x5, %sum /* IEU1 */; \
|
| 204 |
|
|
fmovd %FA, %FB /* FPA Group */; \
|
| 205 |
|
|
fcmpgt32 %fz, %f14, %x2 /* FPM */; \
|
| 206 |
|
|
srl %x6, 1, %x6 /* IEU0 Group */; \
|
| 207 |
|
|
ba,pt %xcc, ett /* CTI */; \
|
| 208 |
|
|
inc %x7 /* IEU1 */;
|
| 209 |
|
|
|
| 210 |
|
|
#define END_THE_TRICK1(f0,f2,f4,f6,f8,f10,f12,f14,FA,FB) \
|
| 211 |
|
|
END_THE_TRICK(f0,f2,f4,f6,f8,f10,f12,f14,FA,FB,f48,f50,f52,f54,f56,f58,f60,f62)
|
| 212 |
|
|
|
| 213 |
|
|
#define END_THE_TRICK2(S0,S1,S2,S3,T0,T1,U0,U1,V0,fz) \
|
| 214 |
|
|
fpadd32 %U0, %U1, %V0 /* FPA Group */; \
|
| 215 |
|
|
srl %x7, 1, %x7 /* IEU0 */; \
|
| 216 |
|
|
add %sum, %x6, %sum /* IEU1 */; \
|
| 217 |
|
|
std %V0, [%sp + STACKOFF] /* Store Group */; \
|
| 218 |
|
|
inc %x8 /* IEU0 */; \
|
| 219 |
|
|
sub %sum, %x7, %sum /* IEU1 */; \
|
| 220 |
|
|
srl %x8, 1, %x8 /* IEU0 Group */; \
|
| 221 |
|
|
fcmpgt32 %fz, %S1, %x3 /* FPM */; \
|
| 222 |
|
|
inc %x1 /* IEU0 Group */; \
|
| 223 |
|
|
fcmpgt32 %fz, %S3, %x4 /* FPM */; \
|
| 224 |
|
|
srl %x1, 1, %x1 /* IEU0 Group */; \
|
| 225 |
|
|
sub %sum, %x8, %sum /* IEU1 */; \
|
| 226 |
|
|
ldx [%sp + STACKOFF], %x8 /* Load Group */; \
|
| 227 |
|
|
inc %x2 /* IEU0 */; \
|
| 228 |
|
|
sub %sum, %x1, %sum /* IEU1 */; \
|
| 229 |
|
|
srl %x2, 1, %x2 /* IEU0 Group */; \
|
| 230 |
|
|
fcmpgt32 %fz, %T1, %x5 /* FPM */; \
|
| 231 |
|
|
inc %x3 /* IEU0 Group */; \
|
| 232 |
|
|
fcmpgt32 %T0, %U0, %x6 /* FPM */; \
|
| 233 |
|
|
srl %x3, 1, %x3 /* IEU0 Group */; \
|
| 234 |
|
|
sub %sum, %x2, %sum /* IEU1 */; \
|
| 235 |
|
|
inc %x4 /* IEU0 Group */; \
|
| 236 |
|
|
sub %sum, %x3, %sum /* IEU1 */; \
|
| 237 |
|
|
srl %x4, 1, %x4 /* IEU0 Group */; \
|
| 238 |
|
|
fcmpgt32 %fz, %U1, %x7 /* FPM */; \
|
| 239 |
|
|
inc %x5 /* IEU0 Group */; \
|
| 240 |
|
|
fcmpgt32 %U0, %V0, %x1 /* FPM */; \
|
| 241 |
|
|
srl %x5, 1, %x5 /* IEU0 Group */; \
|
| 242 |
|
|
sub %sum, %x4, %sum /* IEU1 */; \
|
| 243 |
|
|
sub %sum, %x5, %sum /* IEU0 Group */; \
|
| 244 |
|
|
fcmpgt32 %fz, %V0, %x2 /* FPM */; \
|
| 245 |
|
|
inc %x6 /* IEU0 Group */; \
|
| 246 |
|
|
inc %x7 /* IEU1 */; \
|
| 247 |
|
|
srl %x6, 1, %x6 /* IEU0 Group */; \
|
| 248 |
|
|
inc %x1 /* IEU1 */; \
|
| 249 |
|
|
srl %x7, 1, %x7 /* IEU0 Group */; \
|
| 250 |
|
|
add %sum, %x6, %sum /* IEU1 */; \
|
| 251 |
|
|
srl %x1, 1, %x1 /* IEU0 Group */; \
|
| 252 |
|
|
sub %sum, %x7, %sum /* IEU1 */; \
|
| 253 |
|
|
inc %x2 /* IEU0 Group */; \
|
| 254 |
|
|
add %sum, %x1, %sum /* IEU1 */; \
|
| 255 |
|
|
srl %x2, 1, %x2 /* IEU0 Group */; \
|
| 256 |
|
|
sub %sum, %x2, %sum /* IEU0 Group */; \
|
| 257 |
|
|
addcc %sum, %x8, %sum /* IEU1 Group */; \
|
| 258 |
|
|
bcs,a,pn %xcc, 33f /* CTI */; \
|
| 259 |
|
|
add %sum, 1, %sum /* IEU0 (Group) */; \
|
| 260 |
|
|
33: /* That's it */;
|
| 261 |
|
|
|
| 262 |
|
|
.text
|
| 263 |
|
|
.globl csum_partial_copy_user_vis
|
| 264 |
|
|
.align 32
|
| 265 |
|
|
/* %asi should be either ASI_P or ASI_AIUS for csum_partial_copy resp.
|
| 266 |
|
|
* csum_partial_copy_from_user
|
| 267 |
|
|
* This assumes that !((%src^%dst)&3) && !((%src|%dst)&1) && %len >= 256
|
| 268 |
|
|
*/
|
| 269 |
|
|
csum_partial_copy_user_vis:
|
| 270 |
|
|
andcc %dst, 7, %g0 /* IEU1 Group */
|
| 271 |
|
|
be,pt %icc, 4f /* CTI */
|
| 272 |
|
|
and %dst, 0x38, %o4 /* IEU0 */
|
| 273 |
|
|
mov 1, %g5 /* IEU0 Group */
|
| 274 |
|
|
andcc %dst, 2, %g0 /* IEU1 */
|
| 275 |
|
|
be,pt %icc, 1f /* CTI */
|
| 276 |
|
|
and %dst, 4, %g7 /* IEU0 Group */
|
| 277 |
|
|
lduh [%src], %g2 /* Load */
|
| 278 |
|
|
sub %len, 2, %len /* IEU0 Group */
|
| 279 |
|
|
add %dst, 2, %dst /* IEU1 */
|
| 280 |
|
|
andcc %dst, 4, %g7 /* IEU1 Group */
|
| 281 |
|
|
sll %g5, 16, %g5 /* IEU0 */
|
| 282 |
|
|
stha %g2, [%dst - 2] %asi /* Store Group */
|
| 283 |
|
|
sll %g2, 16, %g2 /* IEU0 */
|
| 284 |
|
|
add %src, 2, %src /* IEU1 */
|
| 285 |
|
|
addcc %g2, %sum, %sum /* IEU1 Group */
|
| 286 |
|
|
bcs,a,pn %icc, 1f /* CTI */
|
| 287 |
|
|
add %sum, %g5, %sum /* IEU0 */
|
| 288 |
|
|
1: lduw [%src], %g2 /* Load */
|
| 289 |
|
|
brz,a,pn %g7, 4f /* CTI+IEU1 Group */
|
| 290 |
|
|
and %dst, 0x38, %o4 /* IEU0 */
|
| 291 |
|
|
add %dst, 4, %dst /* IEU0 Group */
|
| 292 |
|
|
sub %len, 4, %len /* IEU1 */
|
| 293 |
|
|
addcc %g2, %sum, %sum /* IEU1 Group */
|
| 294 |
|
|
bcs,a,pn %icc, 1f /* CTI */
|
| 295 |
|
|
add %sum, 1, %sum /* IEU0 */
|
| 296 |
|
|
1: and %dst, 0x38, %o4 /* IEU0 Group */
|
| 297 |
|
|
stwa %g2, [%dst - 4] %asi /* Store */
|
| 298 |
|
|
add %src, 4, %src /* IEU1 */
|
| 299 |
|
|
4:
|
| 300 |
|
|
#ifdef __KERNEL__
|
| 301 |
|
|
VISEntry
|
| 302 |
|
|
#endif
|
| 303 |
|
|
mov %src, %g7 /* IEU1 Group */
|
| 304 |
|
|
fzero %f48 /* FPA */
|
| 305 |
|
|
alignaddr %src, %g0, %src /* Single Group */
|
| 306 |
|
|
subcc %g7, %src, %g7 /* IEU1 Group */
|
| 307 |
|
|
be,pt %xcc, 1f /* CTI */
|
| 308 |
|
|
mov 0x40, %g1 /* IEU0 */
|
| 309 |
|
|
lduw [%src], %g2 /* Load Group */
|
| 310 |
|
|
subcc %sum, %g2, %sum /* IEU1 Group+load stall*/
|
| 311 |
|
|
bcs,a,pn %icc, 1f /* CTI */
|
| 312 |
|
|
sub %sum, 1, %sum /* IEU0 */
|
| 313 |
|
|
1: srl %sum, 0, %sum /* IEU0 Group */
|
| 314 |
|
|
clr %g5 /* IEU1 */
|
| 315 |
|
|
brz,pn %o4, 3f /* CTI+IEU1 Group */
|
| 316 |
|
|
sub %g1, %o4, %g1 /* IEU0 */
|
| 317 |
|
|
ldd [%src], %f0 /* Load */
|
| 318 |
|
|
clr %o4 /* IEU0 Group */
|
| 319 |
|
|
andcc %dst, 8, %g0 /* IEU1 */
|
| 320 |
|
|
be,pn %icc, 1f /* CTI */
|
| 321 |
|
|
ldd [%src + 8], %f2 /* Load Group */
|
| 322 |
|
|
add %src, 8, %src /* IEU0 */
|
| 323 |
|
|
sub %len, 8, %len /* IEU1 */
|
| 324 |
|
|
fpadd32 %f0, %f48, %f50 /* FPA */
|
| 325 |
|
|
addcc %dst, 8, %dst /* IEU1 Group */
|
| 326 |
|
|
faligndata %f0, %f2, %f16 /* FPA */
|
| 327 |
|
|
fcmpgt32 %f48, %f50, %o4 /* FPM Group */
|
| 328 |
|
|
fmovd %f2, %f0 /* FPA Group */
|
| 329 |
|
|
ldd [%src + 8], %f2 /* Load */
|
| 330 |
|
|
stda %f16, [%dst - 8] %asi /* Store */
|
| 331 |
|
|
fmovd %f50, %f48 /* FPA */
|
| 332 |
|
|
1: andcc %g1, 0x10, %g0 /* IEU1 Group */
|
| 333 |
|
|
be,pn %icc, 1f /* CTI */
|
| 334 |
|
|
and %g1, 0x20, %g1 /* IEU0 */
|
| 335 |
|
|
fpadd32 %f0, %f48, %f50 /* FPA */
|
| 336 |
|
|
ldd [%src + 16], %f4 /* Load Group */
|
| 337 |
|
|
add %src, 16, %src /* IEU0 */
|
| 338 |
|
|
add %dst, 16, %dst /* IEU1 */
|
| 339 |
|
|
faligndata %f0, %f2, %f16 /* FPA */
|
| 340 |
|
|
fcmpgt32 %f48, %f50, %g5 /* FPM Group */
|
| 341 |
|
|
sub %len, 16, %len /* IEU0 */
|
| 342 |
|
|
inc %o4 /* IEU1 */
|
| 343 |
|
|
stda %f16, [%dst - 16] %asi /* Store Group */
|
| 344 |
|
|
fpadd32 %f2, %f50, %f48 /* FPA */
|
| 345 |
|
|
srl %o4, 1, %o5 /* IEU0 */
|
| 346 |
|
|
faligndata %f2, %f4, %f18 /* FPA Group */
|
| 347 |
|
|
stda %f18, [%dst - 8] %asi /* Store */
|
| 348 |
|
|
fcmpgt32 %f50, %f48, %o4 /* FPM Group */
|
| 349 |
|
|
add %o5, %sum, %sum /* IEU0 */
|
| 350 |
|
|
ldd [%src + 8], %f2 /* Load */
|
| 351 |
|
|
fmovd %f4, %f0 /* FPA */
|
| 352 |
|
|
1: brz,a,pn %g1, 4f /* CTI+IEU1 Group */
|
| 353 |
|
|
rd %asi, %g2 /* LSU Group + 4 bubbles*/
|
| 354 |
|
|
inc %g5 /* IEU0 */
|
| 355 |
|
|
fpadd32 %f0, %f48, %f50 /* FPA */
|
| 356 |
|
|
ldd [%src + 16], %f4 /* Load Group */
|
| 357 |
|
|
srl %g5, 1, %g5 /* IEU0 */
|
| 358 |
|
|
add %dst, 32, %dst /* IEU1 */
|
| 359 |
|
|
faligndata %f0, %f2, %f16 /* FPA */
|
| 360 |
|
|
fcmpgt32 %f48, %f50, %o5 /* FPM Group */
|
| 361 |
|
|
inc %o4 /* IEU0 */
|
| 362 |
|
|
ldd [%src + 24], %f6 /* Load */
|
| 363 |
|
|
srl %o4, 1, %o4 /* IEU0 Group */
|
| 364 |
|
|
add %g5, %sum, %sum /* IEU1 */
|
| 365 |
|
|
ldd [%src + 32], %f8 /* Load */
|
| 366 |
|
|
fpadd32 %f2, %f50, %f48 /* FPA */
|
| 367 |
|
|
faligndata %f2, %f4, %f18 /* FPA Group */
|
| 368 |
|
|
sub %len, 32, %len /* IEU0 */
|
| 369 |
|
|
stda %f16, [%dst - 32] %asi /* Store */
|
| 370 |
|
|
fcmpgt32 %f50, %f48, %g3 /* FPM Group */
|
| 371 |
|
|
inc %o5 /* IEU0 */
|
| 372 |
|
|
add %o4, %sum, %sum /* IEU1 */
|
| 373 |
|
|
fpadd32 %f4, %f48, %f50 /* FPA */
|
| 374 |
|
|
faligndata %f4, %f6, %f20 /* FPA Group */
|
| 375 |
|
|
srl %o5, 1, %o5 /* IEU0 */
|
| 376 |
|
|
fcmpgt32 %f48, %f50, %g5 /* FPM Group */
|
| 377 |
|
|
add %o5, %sum, %sum /* IEU0 */
|
| 378 |
|
|
stda %f18, [%dst - 24] %asi /* Store */
|
| 379 |
|
|
fpadd32 %f6, %f50, %f48 /* FPA */
|
| 380 |
|
|
inc %g3 /* IEU0 Group */
|
| 381 |
|
|
stda %f20, [%dst - 16] %asi /* Store */
|
| 382 |
|
|
add %src, 32, %src /* IEU1 */
|
| 383 |
|
|
faligndata %f6, %f8, %f22 /* FPA */
|
| 384 |
|
|
fcmpgt32 %f50, %f48, %o4 /* FPM Group */
|
| 385 |
|
|
srl %g3, 1, %g3 /* IEU0 */
|
| 386 |
|
|
stda %f22, [%dst - 8] %asi /* Store */
|
| 387 |
|
|
add %g3, %sum, %sum /* IEU0 Group */
|
| 388 |
|
|
3: rd %asi, %g2 /* LSU Group + 4 bubbles*/
|
| 389 |
|
|
#ifdef __KERNEL__
|
| 390 |
|
|
4: sethi %hi(vis0s), %g7 /* IEU0 Group */
|
| 391 |
|
|
or %g2, ASI_BLK_OR, %g2 /* IEU1 */
|
| 392 |
|
|
#else
|
| 393 |
|
|
4: rd %pc, %g7 /* LSU Group + 4 bubbles*/
|
| 394 |
|
|
#endif
|
| 395 |
|
|
inc %g5 /* IEU0 Group */
|
| 396 |
|
|
and %src, 0x38, %g3 /* IEU1 */
|
| 397 |
|
|
membar #StoreLoad /* LSU Group */
|
| 398 |
|
|
srl %g5, 1, %g5 /* IEU0 */
|
| 399 |
|
|
inc %o4 /* IEU1 */
|
| 400 |
|
|
sll %g3, 8, %g3 /* IEU0 Group */
|
| 401 |
|
|
sub %len, 0xc0, %len /* IEU1 */
|
| 402 |
|
|
addcc %g5, %sum, %sum /* IEU1 Group */
|
| 403 |
|
|
srl %o4, 1, %o4 /* IEU0 */
|
| 404 |
|
|
add %g7, %g3, %g7 /* IEU0 Group */
|
| 405 |
|
|
add %o4, %sum, %sum /* IEU1 */
|
| 406 |
|
|
#ifdef __KERNEL__
|
| 407 |
|
|
jmpl %g7 + %lo(vis0s), %g0 /* CTI+IEU1 Group */
|
| 408 |
|
|
#else
|
| 409 |
|
|
jmpl %g7 + (vis0s - 4b), %g0 /* CTI+IEU1 Group */
|
| 410 |
|
|
#endif
|
| 411 |
|
|
fzero %f32 /* FPA */
|
| 412 |
|
|
|
| 413 |
|
|
.align 2048
|
| 414 |
|
|
vis0s: wr %g2, ASI_BLK_XOR, %asi /* LSU Group */
|
| 415 |
|
|
ldda [%src] ASI_BLK_P, %f0 /* Load Group */
|
| 416 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 417 |
|
|
ldda [%src] ASI_BLK_P, %f16 /* Load Group */
|
| 418 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 419 |
|
|
fmovd %f48, %f62 /* FPA Group f0 available*/
|
| 420 |
|
|
faligndata %f0, %f2, %f48 /* FPA Group f2 available*/
|
| 421 |
|
|
fcmpgt32 %f32, %f2, %x1 /* FPM Group f4 available*/
|
| 422 |
|
|
fpadd32 %f0, %f62, %f0 /* FPA */
|
| 423 |
|
|
fcmpgt32 %f32, %f4, %x2 /* FPM Group f6 available*/
|
| 424 |
|
|
faligndata %f2, %f4, %f50 /* FPA */
|
| 425 |
|
|
fcmpgt32 %f62, %f0, %x3 /* FPM Group f8 available*/
|
| 426 |
|
|
faligndata %f4, %f6, %f52 /* FPA */
|
| 427 |
|
|
fcmpgt32 %f32, %f6, %x4 /* FPM Group f10 available*/
|
| 428 |
|
|
inc %x1 /* IEU0 */
|
| 429 |
|
|
faligndata %f6, %f8, %f54 /* FPA */
|
| 430 |
|
|
fcmpgt32 %f32, %f8, %x5 /* FPM Group f12 available*/
|
| 431 |
|
|
srl %x1, 1, %x1 /* IEU0 */
|
| 432 |
|
|
inc %x2 /* IEU1 */
|
| 433 |
|
|
faligndata %f8, %f10, %f56 /* FPA */
|
| 434 |
|
|
fcmpgt32 %f32, %f10, %x6 /* FPM Group f14 available*/
|
| 435 |
|
|
srl %x2, 1, %x2 /* IEU0 */
|
| 436 |
|
|
add %sum, %x1, %sum /* IEU1 */
|
| 437 |
|
|
faligndata %f10, %f12, %f58 /* FPA */
|
| 438 |
|
|
fcmpgt32 %f32, %f12, %x7 /* FPM Group */
|
| 439 |
|
|
inc %x3 /* IEU0 */
|
| 440 |
|
|
add %sum, %x2, %sum /* IEU1 */
|
| 441 |
|
|
faligndata %f12, %f14, %f60 /* FPA */
|
| 442 |
|
|
fcmpgt32 %f32, %f14, %x8 /* FPM Group */
|
| 443 |
|
|
srl %x3, 1, %x3 /* IEU0 */
|
| 444 |
|
|
inc %x4 /* IEU1 */
|
| 445 |
|
|
fmovd %f14, %f62 /* FPA */
|
| 446 |
|
|
srl %x4, 1, %x4 /* IEU0 Group */
|
| 447 |
|
|
add %sum, %x3, %sum /* IEU1 */
|
| 448 |
|
|
vis0: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 449 |
|
|
,f48,f50,f52,f54,f56,f58,f60,f62,f62,
|
| 450 |
|
|
,LDBLK(f32), STBLK,,,,,,,,
|
| 451 |
|
|
,bcs,pn %icc, vis0e1)
|
| 452 |
|
|
DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 453 |
|
|
,f48,f50,f52,f54,f56,f58,f60,f62,f62,
|
| 454 |
|
|
,LDBLK(f0), STBLK,,,,,,,,
|
| 455 |
|
|
,bcs,pn %icc, vis0e2)
|
| 456 |
|
|
DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 457 |
|
|
,f48,f50,f52,f54,f56,f58,f60,f62,f62,
|
| 458 |
|
|
,LDBLK(f16), STBLK,,,,,,,,
|
| 459 |
|
|
,bcc,pt %icc, vis0)
|
| 460 |
|
|
vis0e3: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 461 |
|
|
,f48,f50,f52,f54,f56,f58,f60,f62,f32,
|
| 462 |
|
|
,SYNC, STBLK_XORASI(x1,x2),ST(f48,64),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),ST(f60,48),
|
| 463 |
|
|
,add %dst, 56, %dst; add %len, 192 - 8*8, %len; ba,pt %icc, e2)
|
| 464 |
|
|
vis0e1: DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 465 |
|
|
,f48,f50,f52,f54,f56,f58,f60,f62,f0,
|
| 466 |
|
|
,SYNC, STBLK_XORASI(x1,x2),ST(f48,64),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),ST(f60,48),
|
| 467 |
|
|
,add %dst, 56, %dst; add %len, 192 - 8*8, %len; ba,pt %icc, e3)
|
| 468 |
|
|
vis0e2: DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 469 |
|
|
,f48,f50,f52,f54,f56,f58,f60,f62,f16,
|
| 470 |
|
|
,SYNC, STBLK_XORASI(x1,x2),ST(f48,64),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),ST(f60,48),
|
| 471 |
|
|
,add %dst, 56, %dst; add %len, 192 - 8*8, %len; ba,pt %icc, e1)
|
| 472 |
|
|
.align 2048
|
| 473 |
|
|
vis1s: wr %g2, ASI_BLK_XOR, %asi /* LSU Group */
|
| 474 |
|
|
sub %src, 8, %src /* IEU0 Group */
|
| 475 |
|
|
ldda [%src] ASI_BLK_P, %f0 /* Load Group */
|
| 476 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 477 |
|
|
ldda [%src] ASI_BLK_P, %f16 /* Load Group */
|
| 478 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 479 |
|
|
fmovd %f0, %f58 /* FPA Group */
|
| 480 |
|
|
fmovd %f48, %f0 /* FPA Group */
|
| 481 |
|
|
fcmpgt32 %f32, %f2, %x2 /* FPM Group */
|
| 482 |
|
|
faligndata %f2, %f4, %f48 /* FPA */
|
| 483 |
|
|
fcmpgt32 %f32, %f4, %x3 /* FPM Group */
|
| 484 |
|
|
faligndata %f4, %f6, %f50 /* FPA */
|
| 485 |
|
|
fcmpgt32 %f32, %f6, %x4 /* FPM Group */
|
| 486 |
|
|
faligndata %f6, %f8, %f52 /* FPA */
|
| 487 |
|
|
fcmpgt32 %f32, %f8, %x5 /* FPM Group */
|
| 488 |
|
|
inc %x2 /* IEU1 */
|
| 489 |
|
|
faligndata %f8, %f10, %f54 /* FPA */
|
| 490 |
|
|
fcmpgt32 %f32, %f10, %x6 /* FPM Group */
|
| 491 |
|
|
srl %x2, 1, %x2 /* IEU0 */
|
| 492 |
|
|
faligndata %f10, %f12, %f56 /* FPA */
|
| 493 |
|
|
fcmpgt32 %f32, %f12, %x7 /* FPM Group */
|
| 494 |
|
|
inc %x3 /* IEU0 */
|
| 495 |
|
|
add %sum, %x2, %sum /* IEU1 */
|
| 496 |
|
|
faligndata %f12, %f14, %f58 /* FPA */
|
| 497 |
|
|
fcmpgt32 %f32, %f14, %x8 /* FPM Group */
|
| 498 |
|
|
srl %x3, 1, %x3 /* IEU0 */
|
| 499 |
|
|
inc %x4 /* IEU1 */
|
| 500 |
|
|
fmovd %f14, %f60 /* FPA */
|
| 501 |
|
|
srl %x4, 1, %x4 /* IEU0 Group */
|
| 502 |
|
|
add %sum, %x3, %sum /* IEU1 */
|
| 503 |
|
|
vis1: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 504 |
|
|
,f62,f48,f50,f52,f54,f56,f58,f60,f60,
|
| 505 |
|
|
,LDBLK(f32), ,STBLK,,,,,,,
|
| 506 |
|
|
,bcs,pn %icc, vis1e1)
|
| 507 |
|
|
DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 508 |
|
|
,f62,f48,f50,f52,f54,f56,f58,f60,f60,
|
| 509 |
|
|
,LDBLK(f0), ,STBLK,,,,,,,
|
| 510 |
|
|
,bcs,pn %icc, vis1e2)
|
| 511 |
|
|
DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 512 |
|
|
,f62,f48,f50,f52,f54,f56,f58,f60,f60,
|
| 513 |
|
|
,LDBLK(f16), ,STBLK,,,,,,,
|
| 514 |
|
|
,bcc,pt %icc, vis1)
|
| 515 |
|
|
vis1e3: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 516 |
|
|
,f62,f48,f50,f52,f54,f56,f58,f60,f32,
|
| 517 |
|
|
,SYNC, ,STBLK_XORASI(x1,x2),ST(f48,0),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),
|
| 518 |
|
|
,add %dst, 48, %dst; add %len, 192 - 7*8, %len; ba,pt %icc, e2)
|
| 519 |
|
|
vis1e1: DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 520 |
|
|
,f62,f48,f50,f52,f54,f56,f58,f60,f0,
|
| 521 |
|
|
,SYNC, ,STBLK_XORASI(x1,x2),ST(f48,0),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),
|
| 522 |
|
|
,add %dst, 48, %dst; add %len, 192 - 7*8, %len; ba,pt %icc, e3)
|
| 523 |
|
|
vis1e2: DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 524 |
|
|
,f62,f48,f50,f52,f54,f56,f58,f60,f16,
|
| 525 |
|
|
,SYNC, ,STBLK_XORASI(x1,x2),ST(f48,0),ST(f50,8),ST(f52,16),ST(f54,24),ST(f56,32),ST(f58,40),
|
| 526 |
|
|
,add %dst, 48, %dst; add %len, 192 - 7*8, %len; ba,pt %icc, e1)
|
| 527 |
|
|
.align 2048
|
| 528 |
|
|
vis2s: wr %g2, ASI_BLK_XOR, %asi /* LSU Group */
|
| 529 |
|
|
sub %src, 16, %src /* IEU0 Group */
|
| 530 |
|
|
ldda [%src] ASI_BLK_P, %f0 /* Load Group */
|
| 531 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 532 |
|
|
ldda [%src] ASI_BLK_P, %f16 /* Load Group */
|
| 533 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 534 |
|
|
fmovd %f0, %f56 /* FPA Group */
|
| 535 |
|
|
fmovd %f48, %f0 /* FPA Group */
|
| 536 |
|
|
sub %dst, 64, %dst /* IEU0 */
|
| 537 |
|
|
fpsub32 %f2, %f2, %f2 /* FPA Group */
|
| 538 |
|
|
fcmpgt32 %f32, %f4, %x3 /* FPM Group */
|
| 539 |
|
|
faligndata %f4, %f6, %f48 /* FPA */
|
| 540 |
|
|
fcmpgt32 %f32, %f6, %x4 /* FPM Group */
|
| 541 |
|
|
faligndata %f6, %f8, %f50 /* FPA */
|
| 542 |
|
|
fcmpgt32 %f32, %f8, %x5 /* FPM Group */
|
| 543 |
|
|
faligndata %f8, %f10, %f52 /* FPA */
|
| 544 |
|
|
fcmpgt32 %f32, %f10, %x6 /* FPM Group */
|
| 545 |
|
|
faligndata %f10, %f12, %f54 /* FPA */
|
| 546 |
|
|
fcmpgt32 %f32, %f12, %x7 /* FPM Group */
|
| 547 |
|
|
inc %x3 /* IEU0 */
|
| 548 |
|
|
faligndata %f12, %f14, %f56 /* FPA */
|
| 549 |
|
|
fcmpgt32 %f32, %f14, %x8 /* FPM Group */
|
| 550 |
|
|
srl %x3, 1, %x3 /* IEU0 */
|
| 551 |
|
|
inc %x4 /* IEU1 */
|
| 552 |
|
|
fmovd %f14, %f58 /* FPA */
|
| 553 |
|
|
srl %x4, 1, %x4 /* IEU0 Group */
|
| 554 |
|
|
add %sum, %x3, %sum /* IEU1 */
|
| 555 |
|
|
vis2: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 556 |
|
|
,f60,f62,f48,f50,f52,f54,f56,f58,f58,
|
| 557 |
|
|
,LDBLK(f32), ,,STBLK,,,,,,
|
| 558 |
|
|
,bcs,pn %icc, vis2e1)
|
| 559 |
|
|
DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 560 |
|
|
,f60,f62,f48,f50,f52,f54,f56,f58,f58,
|
| 561 |
|
|
,LDBLK(f0), ,,STBLK,,,,,,
|
| 562 |
|
|
,bcs,pn %icc, vis2e2)
|
| 563 |
|
|
DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 564 |
|
|
,f60,f62,f48,f50,f52,f54,f56,f58,f58,
|
| 565 |
|
|
,LDBLK(f16), ,,STBLK,,,,,,
|
| 566 |
|
|
,bcc,pt %icc, vis2)
|
| 567 |
|
|
vis2e3: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 568 |
|
|
,f60,f62,f48,f50,f52,f54,f56,f58,f32,
|
| 569 |
|
|
,SYNC, ,,STBLK_XORASI(x2,x3),ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),ST(f56,96),
|
| 570 |
|
|
,add %dst, 104, %dst; add %len, 192 - 6*8, %len; ba,pt %icc, e2)
|
| 571 |
|
|
vis2e1: DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 572 |
|
|
,f60,f62,f48,f50,f52,f54,f56,f58,f0,
|
| 573 |
|
|
,SYNC, ,,STBLK_XORASI(x2,x3),ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),ST(f56,96),
|
| 574 |
|
|
,add %dst, 104, %dst; add %len, 192 - 6*8, %len; ba,pt %icc, e3)
|
| 575 |
|
|
vis2e2: DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 576 |
|
|
,f60,f62,f48,f50,f52,f54,f56,f58,f16,
|
| 577 |
|
|
,SYNC, ,,STBLK_XORASI(x2,x3),ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),ST(f56,96),
|
| 578 |
|
|
,add %dst, 104, %dst; add %len, 192 - 6*8, %len; ba,pt %icc, e1)
|
| 579 |
|
|
.align 2048
|
| 580 |
|
|
vis3s: wr %g2, ASI_BLK_XOR, %asi /* LSU Group */
|
| 581 |
|
|
sub %src, 24, %src /* IEU0 Group */
|
| 582 |
|
|
ldda [%src] ASI_BLK_P, %f0 /* Load Group */
|
| 583 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 584 |
|
|
ldda [%src] ASI_BLK_P, %f16 /* Load Group */
|
| 585 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 586 |
|
|
fmovd %f0, %f54 /* FPA Group */
|
| 587 |
|
|
fmovd %f48, %f0 /* FPA Group */
|
| 588 |
|
|
sub %dst, 64, %dst /* IEU0 */
|
| 589 |
|
|
fpsub32 %f2, %f2, %f2 /* FPA Group */
|
| 590 |
|
|
fpsub32 %f4, %f4, %f4 /* FPA Group */
|
| 591 |
|
|
fcmpgt32 %f32, %f6, %x4 /* FPM Group */
|
| 592 |
|
|
faligndata %f6, %f8, %f48 /* FPA */
|
| 593 |
|
|
fcmpgt32 %f32, %f8, %x5 /* FPM Group */
|
| 594 |
|
|
faligndata %f8, %f10, %f50 /* FPA */
|
| 595 |
|
|
fcmpgt32 %f32, %f10, %x6 /* FPM Group */
|
| 596 |
|
|
faligndata %f10, %f12, %f52 /* FPA */
|
| 597 |
|
|
fcmpgt32 %f32, %f12, %x7 /* FPM Group */
|
| 598 |
|
|
faligndata %f12, %f14, %f54 /* FPA */
|
| 599 |
|
|
fcmpgt32 %f32, %f14, %x8 /* FPM Group */
|
| 600 |
|
|
fmovd %f14, %f56 /* FPA */
|
| 601 |
|
|
inc %x4 /* IEU0 */
|
| 602 |
|
|
srl %x4, 1, %x4 /* IEU0 Group */
|
| 603 |
|
|
vis3: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 604 |
|
|
,f58,f60,f62,f48,f50,f52,f54,f56,f56,
|
| 605 |
|
|
,LDBLK(f32), ,,,STBLK,,,,,
|
| 606 |
|
|
,bcs,pn %icc, vis3e1)
|
| 607 |
|
|
DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 608 |
|
|
,f58,f60,f62,f48,f50,f52,f54,f56,f56,
|
| 609 |
|
|
,LDBLK(f0), ,,,STBLK,,,,,
|
| 610 |
|
|
,bcs,pn %icc, vis3e2)
|
| 611 |
|
|
DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 612 |
|
|
,f58,f60,f62,f48,f50,f52,f54,f56,f56,
|
| 613 |
|
|
,LDBLK(f16), ,,,STBLK,,,,,
|
| 614 |
|
|
,bcc,pt %icc, vis3)
|
| 615 |
|
|
vis3e3: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 616 |
|
|
,f58,f60,f62,f48,f50,f52,f54,f56,f32,
|
| 617 |
|
|
,SYNC, ,,,STBLK_XORASI(x3,x4),ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),
|
| 618 |
|
|
,add %dst, 96, %dst; add %len, 192 - 5*8, %len; ba,pt %icc, e2)
|
| 619 |
|
|
vis3e1: DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 620 |
|
|
,f58,f60,f62,f48,f50,f52,f54,f56,f0,
|
| 621 |
|
|
,SYNC, ,,,STBLK_XORASI(x3,x4),ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),
|
| 622 |
|
|
,add %dst, 96, %dst; add %len, 192 - 5*8, %len; ba,pt %icc, e3)
|
| 623 |
|
|
vis3e2: DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 624 |
|
|
,f58,f60,f62,f48,f50,f52,f54,f56,f16,
|
| 625 |
|
|
,SYNC, ,,,STBLK_XORASI(x3,x4),ST(f48,64),ST(f50,72),ST(f52,80),ST(f54,88),
|
| 626 |
|
|
,add %dst, 96, %dst; add %len, 192 - 5*8, %len; ba,pt %icc, e1)
|
| 627 |
|
|
.align 2048
|
| 628 |
|
|
vis4s: wr %g2, ASI_BLK_XOR, %asi /* LSU Group */
|
| 629 |
|
|
sub %src, 32, %src /* IEU0 Group */
|
| 630 |
|
|
ldda [%src] ASI_BLK_P, %f0 /* Load Group */
|
| 631 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 632 |
|
|
ldda [%src] ASI_BLK_P, %f16 /* Load Group */
|
| 633 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 634 |
|
|
fmovd %f0, %f52 /* FPA Group */
|
| 635 |
|
|
fmovd %f48, %f0 /* FPA Group */
|
| 636 |
|
|
sub %dst, 64, %dst /* IEU0 */
|
| 637 |
|
|
fpsub32 %f2, %f2, %f2 /* FPA Group */
|
| 638 |
|
|
fpsub32 %f4, %f4, %f4 /* FPA Group */
|
| 639 |
|
|
fpsub32 %f6, %f6, %f6 /* FPA Group */
|
| 640 |
|
|
clr %x4 /* IEU0 */
|
| 641 |
|
|
fcmpgt32 %f32, %f8, %x5 /* FPM Group */
|
| 642 |
|
|
faligndata %f8, %f10, %f48 /* FPA */
|
| 643 |
|
|
fcmpgt32 %f32, %f10, %x6 /* FPM Group */
|
| 644 |
|
|
faligndata %f10, %f12, %f50 /* FPA */
|
| 645 |
|
|
fcmpgt32 %f32, %f12, %x7 /* FPM Group */
|
| 646 |
|
|
faligndata %f12, %f14, %f52 /* FPA */
|
| 647 |
|
|
fcmpgt32 %f32, %f14, %x8 /* FPM Group */
|
| 648 |
|
|
fmovd %f14, %f54 /* FPA */
|
| 649 |
|
|
vis4: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 650 |
|
|
,f56,f58,f60,f62,f48,f50,f52,f54,f54,
|
| 651 |
|
|
,LDBLK(f32), ,,,,STBLK,,,,
|
| 652 |
|
|
,bcs,pn %icc, vis4e1)
|
| 653 |
|
|
DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 654 |
|
|
,f56,f58,f60,f62,f48,f50,f52,f54,f54,
|
| 655 |
|
|
,LDBLK(f0), ,,,,STBLK,,,,
|
| 656 |
|
|
,bcs,pn %icc, vis4e2)
|
| 657 |
|
|
DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 658 |
|
|
,f56,f58,f60,f62,f48,f50,f52,f54,f54,
|
| 659 |
|
|
,LDBLK(f16), ,,,,STBLK,,,,
|
| 660 |
|
|
,bcc,pt %icc, vis4)
|
| 661 |
|
|
vis4e3: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 662 |
|
|
,f56,f58,f60,f62,f48,f50,f52,f54,f32,
|
| 663 |
|
|
,SYNC, ,,,,STBLK_XORASI(x4,x5),ST(f48,64),ST(f50,72),ST(f52,80),
|
| 664 |
|
|
,add %dst, 88, %dst; add %len, 192 - 4*8, %len; ba,pt %icc, e2)
|
| 665 |
|
|
vis4e1: DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 666 |
|
|
,f56,f58,f60,f62,f48,f50,f52,f54,f0,
|
| 667 |
|
|
,SYNC, ,,,,STBLK_XORASI(x4,x5),ST(f48,64),ST(f50,72),ST(f52,80),
|
| 668 |
|
|
,add %dst, 88, %dst; add %len, 192 - 4*8, %len; ba,pt %icc, e3)
|
| 669 |
|
|
vis4e2: DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 670 |
|
|
,f56,f58,f60,f62,f48,f50,f52,f54,f16,
|
| 671 |
|
|
,SYNC, ,,,,STBLK_XORASI(x4,x5),ST(f48,64),ST(f50,72),ST(f52,80),
|
| 672 |
|
|
,add %dst, 88, %dst; add %len, 192 - 4*8, %len; ba,pt %icc, e1)
|
| 673 |
|
|
.align 2048
|
| 674 |
|
|
vis5s: ldd [%src+0], %f10 /* Load Group */
|
| 675 |
|
|
ldd [%src+8], %f12 /* Load Group */
|
| 676 |
|
|
ldd [%src+16], %f14 /* Load Group */
|
| 677 |
|
|
add %src, 24, %src /* IEU0 Group */
|
| 678 |
|
|
wr %g2, ASI_BLK_XOR, %asi /* LSU Group */
|
| 679 |
|
|
ldda [%src] ASI_BLK_P, %f16 /* Load Group */
|
| 680 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 681 |
|
|
fmovd %f48, %f0 /* FPA Group */
|
| 682 |
|
|
fmuld %f32, %f32, %f2 /* FPM */
|
| 683 |
|
|
clr %x4 /* IEU0 */
|
| 684 |
|
|
faddd %f32, %f32, %f4 /* FPA Group */
|
| 685 |
|
|
fmuld %f32, %f32, %f6 /* FPM */
|
| 686 |
|
|
clr %x5 /* IEU0 */
|
| 687 |
|
|
faddd %f32, %f32, %f8 /* FPA Group */
|
| 688 |
|
|
fcmpgt32 %f32, %f10, %x6 /* FPM Group */
|
| 689 |
|
|
sub %dst, 64, %dst /* IEU0 */
|
| 690 |
|
|
faligndata %f10, %f12, %f48 /* FPA */
|
| 691 |
|
|
fcmpgt32 %f32, %f12, %x7 /* FPM Group */
|
| 692 |
|
|
faligndata %f12, %f14, %f50 /* FPA */
|
| 693 |
|
|
fcmpgt32 %f32, %f14, %x8 /* FPM Group */
|
| 694 |
|
|
fmovd %f14, %f52 /* FPA */
|
| 695 |
|
|
vis5: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 696 |
|
|
,f54,f56,f58,f60,f62,f48,f50,f52,f52,
|
| 697 |
|
|
,LDBLK(f32), ,,,,,STBLK,,,
|
| 698 |
|
|
,bcs,pn %icc, vis5e1)
|
| 699 |
|
|
DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 700 |
|
|
,f54,f56,f58,f60,f62,f48,f50,f52,f52,
|
| 701 |
|
|
,LDBLK(f0), ,,,,,STBLK,,,
|
| 702 |
|
|
,bcs,pn %icc, vis5e2)
|
| 703 |
|
|
DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 704 |
|
|
,f54,f56,f58,f60,f62,f48,f50,f52,f52,
|
| 705 |
|
|
,LDBLK(f16), ,,,,,STBLK,,,
|
| 706 |
|
|
,bcc,pt %icc, vis5)
|
| 707 |
|
|
vis5e3: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 708 |
|
|
,f54,f56,f58,f60,f62,f48,f50,f52,f32,
|
| 709 |
|
|
,SYNC, ,,,,,STBLK_XORASI(x5,x6),ST(f48,64),ST(f50,72),
|
| 710 |
|
|
,add %dst, 80, %dst; add %len, 192 - 3*8, %len; ba,pt %icc, e2)
|
| 711 |
|
|
vis5e1: DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 712 |
|
|
,f54,f56,f58,f60,f62,f48,f50,f52,f0,
|
| 713 |
|
|
,SYNC, ,,,,,STBLK_XORASI(x5,x6),ST(f48,64),ST(f50,72),
|
| 714 |
|
|
,add %dst, 80, %dst; add %len, 192 - 3*8, %len; ba,pt %icc, e3)
|
| 715 |
|
|
vis5e2: DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 716 |
|
|
,f54,f56,f58,f60,f62,f48,f50,f52,f16,
|
| 717 |
|
|
,SYNC, ,,,,,STBLK_XORASI(x5,x6),ST(f48,64),ST(f50,72),
|
| 718 |
|
|
,add %dst, 80, %dst; add %len, 192 - 3*8, %len; ba,pt %icc, e1)
|
| 719 |
|
|
.align 2048
|
| 720 |
|
|
vis6s: ldd [%src+0], %f12 /* Load Group */
|
| 721 |
|
|
ldd [%src+8], %f14 /* Load Group */
|
| 722 |
|
|
add %src, 16, %src /* IEU0 Group */
|
| 723 |
|
|
wr %g2, ASI_BLK_XOR, %asi /* LSU Group */
|
| 724 |
|
|
ldda [%src] ASI_BLK_P, %f16 /* Load Group */
|
| 725 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 726 |
|
|
fmovd %f48, %f0 /* FPA Group */
|
| 727 |
|
|
fmuld %f32, %f32, %f2 /* FPM */
|
| 728 |
|
|
clr %x4 /* IEU0 */
|
| 729 |
|
|
faddd %f32, %f32, %f4 /* FPA Group */
|
| 730 |
|
|
fmuld %f32, %f32, %f6 /* FPM */
|
| 731 |
|
|
clr %x5 /* IEU0 */
|
| 732 |
|
|
faddd %f32, %f32, %f8 /* FPA Group */
|
| 733 |
|
|
fmuld %f32, %f32, %f10 /* FPM */
|
| 734 |
|
|
clr %x6 /* IEU0 */
|
| 735 |
|
|
fcmpgt32 %f32, %f12, %x7 /* FPM Group */
|
| 736 |
|
|
sub %dst, 64, %dst /* IEU0 */
|
| 737 |
|
|
fcmpgt32 %f32, %f14, %x8 /* FPM Group */
|
| 738 |
|
|
faligndata %f12, %f14, %f48 /* FPA */
|
| 739 |
|
|
fmovd %f14, %f50 /* FPA Group */
|
| 740 |
|
|
vis6: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 741 |
|
|
,f52,f54,f56,f58,f60,f62,f48,f50,f50,
|
| 742 |
|
|
,LDBLK(f32), ,,,,,,STBLK,,
|
| 743 |
|
|
,bcs,pn %icc, vis6e1)
|
| 744 |
|
|
DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 745 |
|
|
,f52,f54,f56,f58,f60,f62,f48,f50,f50,
|
| 746 |
|
|
,LDBLK(f0), ,,,,,,STBLK,,
|
| 747 |
|
|
,bcs,pn %icc, vis6e2)
|
| 748 |
|
|
DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 749 |
|
|
,f52,f54,f56,f58,f60,f62,f48,f50,f50,
|
| 750 |
|
|
,LDBLK(f16), ,,,,,,STBLK,,
|
| 751 |
|
|
,bcc,pt %icc, vis6)
|
| 752 |
|
|
vis6e3: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 753 |
|
|
,f52,f54,f56,f58,f60,f62,f48,f50,f32,
|
| 754 |
|
|
,SYNC, ,,,,,,STBLK_XORASI(x6,x7),ST(f48,64),
|
| 755 |
|
|
,add %dst, 72, %dst; add %len, 192 - 2*8, %len; ba,pt %icc, e2)
|
| 756 |
|
|
vis6e1: DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 757 |
|
|
,f52,f54,f56,f58,f60,f62,f48,f50,f0,
|
| 758 |
|
|
,SYNC, ,,,,,,STBLK_XORASI(x6,x7),ST(f48,64),
|
| 759 |
|
|
,add %dst, 72, %dst; add %len, 192 - 2*8, %len; ba,pt %icc, e3)
|
| 760 |
|
|
vis6e2: DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 761 |
|
|
,f52,f54,f56,f58,f60,f62,f48,f50,f16,
|
| 762 |
|
|
,SYNC, ,,,,,,STBLK_XORASI(x6,x7),ST(f48,64),
|
| 763 |
|
|
,add %dst, 72, %dst; add %len, 192 - 2*8, %len; ba,pt %icc, e1)
|
| 764 |
|
|
.align 2048
|
| 765 |
|
|
vis7s: ldd [%src+0], %f14 /* Load Group */
|
| 766 |
|
|
add %src, 8, %src /* IEU0 Group */
|
| 767 |
|
|
wr %g2, ASI_BLK_XOR, %asi /* LSU Group */
|
| 768 |
|
|
ldda [%src] ASI_BLK_P, %f16 /* Load Group */
|
| 769 |
|
|
add %src, 64, %src /* IEU0 Group */
|
| 770 |
|
|
fmovd %f48, %f0 /* FPA Group */
|
| 771 |
|
|
fmuld %f32, %f32, %f2 /* FPM */
|
| 772 |
|
|
clr %x4 /* IEU0 */
|
| 773 |
|
|
faddd %f32, %f32, %f4 /* FPA Group */
|
| 774 |
|
|
fmuld %f32, %f32, %f6 /* FPM */
|
| 775 |
|
|
clr %x5 /* IEU0 */
|
| 776 |
|
|
faddd %f32, %f32, %f8 /* FPA Group */
|
| 777 |
|
|
fmuld %f32, %f32, %f10 /* FPM */
|
| 778 |
|
|
clr %x6 /* IEU0 */
|
| 779 |
|
|
faddd %f32, %f32, %f12 /* FPA Group */
|
| 780 |
|
|
clr %x7 /* IEU0 */
|
| 781 |
|
|
fcmpgt32 %f32, %f14, %x8 /* FPM Group */
|
| 782 |
|
|
sub %dst, 64, %dst /* IEU0 */
|
| 783 |
|
|
fmovd %f14, %f48 /* FPA */
|
| 784 |
|
|
vis7: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 785 |
|
|
,f50,f52,f54,f56,f58,f60,f62,f48,f48,
|
| 786 |
|
|
,LDBLK(f32), ,,,,,,,STBLK,
|
| 787 |
|
|
,bcs,pn %icc, vis7e1)
|
| 788 |
|
|
DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 789 |
|
|
,f50,f52,f54,f56,f58,f60,f62,f48,f48,
|
| 790 |
|
|
,LDBLK(f0), ,,,,,,,STBLK,
|
| 791 |
|
|
,bcs,pn %icc, vis7e2)
|
| 792 |
|
|
DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 793 |
|
|
,f50,f52,f54,f56,f58,f60,f62,f48,f48,
|
| 794 |
|
|
,LDBLK(f16), ,,,,,,,STBLK,
|
| 795 |
|
|
,bcc,pt %icc, vis7)
|
| 796 |
|
|
vis7e3: DO_THE_TRICK( f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,
|
| 797 |
|
|
,f50,f52,f54,f56,f58,f60,f62,f48,f32,
|
| 798 |
|
|
,SYNC, ,,,,,,,STBLK_XORASI(x7,x8),
|
| 799 |
|
|
,add %dst, 64, %dst; add %len, 192 - 1*8, %len; ba,pt %icc, e2)
|
| 800 |
|
|
vis7e1: DO_THE_TRICK( f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,
|
| 801 |
|
|
,f50,f52,f54,f56,f58,f60,f62,f48,f0,
|
| 802 |
|
|
,SYNC, ,,,,,,,STBLK_XORASI(x7,x8),
|
| 803 |
|
|
,add %dst, 64, %dst; add %len, 192 - 1*8, %len; ba,pt %icc, e3)
|
| 804 |
|
|
vis7e2: DO_THE_TRICK( f32,f34,f36,f38,f40,f42,f44,f46,f0,f2,f4,f6,f8,f10,f12,f14,
|
| 805 |
|
|
,f50,f52,f54,f56,f58,f60,f62,f48,f16,
|
| 806 |
|
|
,SYNC, ,,,,,,,STBLK_XORASI(x7,x8),
|
| 807 |
|
|
,add %dst, 64, %dst; add %len, 192 - 1*8, %len; ba,pt %icc, e1)
|
| 808 |
|
|
e1: END_THE_TRICK1( f0,f2,f4,f6,f8,f10,f12,f14,f16,f6)
|
| 809 |
|
|
e2: END_THE_TRICK1( f16,f18,f20,f22,f24,f26,f28,f30,f32,f6)
|
| 810 |
|
|
e3: END_THE_TRICK1( f32,f34,f36,f38,f40,f42,f44,f46,f0,f6)
|
| 811 |
|
|
ett: rd %gsr, %x3 /* LSU Group+4bubbles */
|
| 812 |
|
|
andcc %x3, 7, %x3 /* IEU1 Group */
|
| 813 |
|
|
add %dst, 8, %dst /* IEU0 */
|
| 814 |
|
|
bne,pn %icc, 1f /* CTI */
|
| 815 |
|
|
fzero %f10 /* FPA */
|
| 816 |
|
|
brz,a,pn %len, 2f /* CTI+IEU1 Group */
|
| 817 |
|
|
stda %f6, [%dst - 8] %asi /* Store */
|
| 818 |
|
|
1: cmp %len, 8 /* IEU1 */
|
| 819 |
|
|
blu,pn %icc, 3f /* CTI */
|
| 820 |
|
|
sub %src, 64, %src /* IEU0 Group */
|
| 821 |
|
|
1: ldd [%src], %f2 /* Load Group */
|
| 822 |
|
|
fpadd32 %f10, %f2, %f12 /* FPA Group+load stall*/
|
| 823 |
|
|
add %src, 8, %src /* IEU0 */
|
| 824 |
|
|
add %dst, 8, %dst /* IEU1 */
|
| 825 |
|
|
faligndata %f6, %f2, %f14 /* FPA Group */
|
| 826 |
|
|
fcmpgt32 %f10, %f12, %x5 /* FPM Group */
|
| 827 |
|
|
stda %f14, [%dst - 16] %asi /* Store */
|
| 828 |
|
|
fmovd %f2, %f6 /* FPA */
|
| 829 |
|
|
fmovd %f12, %f10 /* FPA Group */
|
| 830 |
|
|
sub %len, 8, %len /* IEU1 */
|
| 831 |
|
|
fzero %f16 /* FPA Group - FPU nop */
|
| 832 |
|
|
fzero %f18 /* FPA Group - FPU nop */
|
| 833 |
|
|
inc %x5 /* IEU0 */
|
| 834 |
|
|
srl %x5, 1, %x5 /* IEU0 Group (regdep) */
|
| 835 |
|
|
cmp %len, 8 /* IEU1 */
|
| 836 |
|
|
bgeu,pt %icc, 1b /* CTI */
|
| 837 |
|
|
add %x5, %sum, %sum /* IEU0 Group */
|
| 838 |
|
|
3: brz,a,pt %x3, 2f /* CTI+IEU1 */
|
| 839 |
|
|
stda %f6, [%dst - 8] %asi /* Store Group */
|
| 840 |
|
|
sta %f7, [%dst - 8] %asi /* Store Group */
|
| 841 |
|
|
sub %dst, 4, %dst /* IEU0 */
|
| 842 |
|
|
add %len, 4, %len /* IEU1 */
|
| 843 |
|
|
2:
|
| 844 |
|
|
#ifdef __KERNEL__
|
| 845 |
|
|
sub %sp, 8, %sp /* IEU0 Group */
|
| 846 |
|
|
#endif
|
| 847 |
|
|
END_THE_TRICK2( f48,f50,f52,f54,f56,f58,f60,f10,f12,f62)
|
| 848 |
|
|
membar #Sync /* LSU Group */
|
| 849 |
|
|
#ifdef __KERNEL__
|
| 850 |
|
|
VISExit
|
| 851 |
|
|
add %sp, 8, %sp /* IEU0 Group */
|
| 852 |
|
|
#endif
|
| 853 |
|
|
23: brnz,pn %len, 26f /* CTI+IEU1 Group */
|
| 854 |
|
|
24: sllx %sum, 32, %g1 /* IEU0 */
|
| 855 |
|
|
25: addcc %sum, %g1, %src /* IEU1 Group */
|
| 856 |
|
|
srlx %src, 32, %src /* IEU0 Group (regdep) */
|
| 857 |
|
|
bcs,a,pn %xcc, 1f /* CTI */
|
| 858 |
|
|
add %src, 1, %src /* IEU1 */
|
| 859 |
|
|
#ifndef __KERNEL__
|
| 860 |
|
|
1: retl /* CTI Group brk forced*/
|
| 861 |
|
|
srl %src, 0, %src /* IEU0 */
|
| 862 |
|
|
#else
|
| 863 |
|
|
1: sethi %uhi(PAGE_OFFSET), %g4 /* IEU0 Group */
|
| 864 |
|
|
retl /* CTI Group brk forced*/
|
| 865 |
|
|
sllx %g4, 32, %g4 /* IEU0 */
|
| 866 |
|
|
#endif
|
| 867 |
|
|
26: andcc %len, 8, %g0 /* IEU1 Group */
|
| 868 |
|
|
be,pn %icc, 1f /* CTI */
|
| 869 |
|
|
lduw [%src], %o4 /* Load */
|
| 870 |
|
|
lduw [%src+4], %g2 /* Load Group */
|
| 871 |
|
|
add %src, 8, %src /* IEU0 */
|
| 872 |
|
|
add %dst, 8, %dst /* IEU1 */
|
| 873 |
|
|
sllx %o4, 32, %g5 /* IEU0 Group */
|
| 874 |
|
|
stwa %o4, [%dst - 8] %asi /* Store */
|
| 875 |
|
|
or %g5, %g2, %g5 /* IEU0 Group */
|
| 876 |
|
|
stwa %g2, [%dst - 4] %asi /* Store */
|
| 877 |
|
|
addcc %g5, %sum, %sum /* IEU1 Group */
|
| 878 |
|
|
bcs,a,pn %xcc, 1f /* CTI */
|
| 879 |
|
|
add %sum, 1, %sum /* IEU0 */
|
| 880 |
|
|
1: andcc %len, 4, %g0 /* IEU1 Group */
|
| 881 |
|
|
be,a,pn %icc, 1f /* CTI */
|
| 882 |
|
|
clr %g2 /* IEU0 */
|
| 883 |
|
|
lduw [%src], %g7 /* Load */
|
| 884 |
|
|
add %src, 4, %src /* IEU0 Group */
|
| 885 |
|
|
add %dst, 4, %dst /* IEU1 */
|
| 886 |
|
|
sllx %g7, 32, %g2 /* IEU0 Group */
|
| 887 |
|
|
stwa %g7, [%dst - 4] %asi /* Store */
|
| 888 |
|
|
1: andcc %len, 2, %g0 /* IEU1 */
|
| 889 |
|
|
be,a,pn %icc, 1f /* CTI */
|
| 890 |
|
|
clr %g3 /* IEU0 Group */
|
| 891 |
|
|
lduh [%src], %g7 /* Load */
|
| 892 |
|
|
add %src, 2, %src /* IEU1 */
|
| 893 |
|
|
add %dst, 2, %dst /* IEU0 Group */
|
| 894 |
|
|
sll %g7, 16, %g3 /* IEU0 Group */
|
| 895 |
|
|
stha %g7, [%dst - 2] %asi /* Store */
|
| 896 |
|
|
1: andcc %len, 1, %g0 /* IEU1 */
|
| 897 |
|
|
be,a,pn %icc, 1f /* CTI */
|
| 898 |
|
|
clr %o5 /* IEU0 Group */
|
| 899 |
|
|
ldub [%src], %g7 /* Load */
|
| 900 |
|
|
sll %g7, 8, %o5 /* IEU0 Group */
|
| 901 |
|
|
stba %g7, [%dst] %asi /* Store */
|
| 902 |
|
|
1: or %g2, %g3, %g3 /* IEU1 */
|
| 903 |
|
|
or %o5, %g3, %g3 /* IEU0 Group (regdep) */
|
| 904 |
|
|
addcc %g3, %sum, %sum /* IEU1 Group (regdep) */
|
| 905 |
|
|
bcs,a,pn %xcc, 1f /* CTI */
|
| 906 |
|
|
add %sum, 1, %sum /* IEU0 */
|
| 907 |
|
|
1: ba,pt %xcc, 25b /* CTI Group */
|
| 908 |
|
|
sllx %sum, 32, %g1 /* IEU0 */
|
| 909 |
|
|
|
| 910 |
|
|
#ifdef __KERNEL__
|
| 911 |
|
|
end:
|
| 912 |
|
|
|
| 913 |
|
|
.section __ex_table
|
| 914 |
|
|
.align 4
|
| 915 |
|
|
.word csum_partial_copy_user_vis, 0, end, cpc_handler
|
| 916 |
|
|
#endif
|
