0001
0002
0003
0004
0005
0006
0007
0008
0009
0010
0011
0012 #include <linux/linkage.h>
0013
0014 .file "serpent-sse2-i586-asm_32.S"
0015 .text
0016
0017 #define arg_ctx 4
0018 #define arg_dst 8
0019 #define arg_src 12
0020 #define arg_xor 16
0021
0022
0023
0024
0025 #define CTX %edx
0026
0027 #define RA %xmm0
0028 #define RB %xmm1
0029 #define RC %xmm2
0030 #define RD %xmm3
0031 #define RE %xmm4
0032
0033 #define RT0 %xmm5
0034 #define RT1 %xmm6
0035
0036 #define RNOT %xmm7
0037
0038 #define get_key(i, j, t) \
0039 movd (4*(i)+(j))*4(CTX), t; \
0040 pshufd $0, t, t;
0041
0042 #define K(x0, x1, x2, x3, x4, i) \
0043 get_key(i, 0, x4); \
0044 get_key(i, 1, RT0); \
0045 get_key(i, 2, RT1); \
0046 pxor x4, x0; \
0047 pxor RT0, x1; \
0048 pxor RT1, x2; \
0049 get_key(i, 3, x4); \
0050 pxor x4, x3;
0051
0052 #define LK(x0, x1, x2, x3, x4, i) \
0053 movdqa x0, x4; \
0054 pslld $13, x0; \
0055 psrld $(32 - 13), x4; \
0056 por x4, x0; \
0057 pxor x0, x1; \
0058 movdqa x2, x4; \
0059 pslld $3, x2; \
0060 psrld $(32 - 3), x4; \
0061 por x4, x2; \
0062 pxor x2, x1; \
0063 movdqa x1, x4; \
0064 pslld $1, x1; \
0065 psrld $(32 - 1), x4; \
0066 por x4, x1; \
0067 movdqa x0, x4; \
0068 pslld $3, x4; \
0069 pxor x2, x3; \
0070 pxor x4, x3; \
0071 movdqa x3, x4; \
0072 pslld $7, x3; \
0073 psrld $(32 - 7), x4; \
0074 por x4, x3; \
0075 movdqa x1, x4; \
0076 pslld $7, x4; \
0077 pxor x1, x0; \
0078 pxor x3, x0; \
0079 pxor x3, x2; \
0080 pxor x4, x2; \
0081 movdqa x0, x4; \
0082 get_key(i, 1, RT0); \
0083 pxor RT0, x1; \
0084 get_key(i, 3, RT0); \
0085 pxor RT0, x3; \
0086 pslld $5, x0; \
0087 psrld $(32 - 5), x4; \
0088 por x4, x0; \
0089 movdqa x2, x4; \
0090 pslld $22, x2; \
0091 psrld $(32 - 22), x4; \
0092 por x4, x2; \
0093 get_key(i, 0, RT0); \
0094 pxor RT0, x0; \
0095 get_key(i, 2, RT0); \
0096 pxor RT0, x2;
0097
0098 #define KL(x0, x1, x2, x3, x4, i) \
0099 K(x0, x1, x2, x3, x4, i); \
0100 movdqa x0, x4; \
0101 psrld $5, x0; \
0102 pslld $(32 - 5), x4; \
0103 por x4, x0; \
0104 movdqa x2, x4; \
0105 psrld $22, x2; \
0106 pslld $(32 - 22), x4; \
0107 por x4, x2; \
0108 pxor x3, x2; \
0109 pxor x3, x0; \
0110 movdqa x1, x4; \
0111 pslld $7, x4; \
0112 pxor x1, x0; \
0113 pxor x4, x2; \
0114 movdqa x1, x4; \
0115 psrld $1, x1; \
0116 pslld $(32 - 1), x4; \
0117 por x4, x1; \
0118 movdqa x3, x4; \
0119 psrld $7, x3; \
0120 pslld $(32 - 7), x4; \
0121 por x4, x3; \
0122 pxor x0, x1; \
0123 movdqa x0, x4; \
0124 pslld $3, x4; \
0125 pxor x4, x3; \
0126 movdqa x0, x4; \
0127 psrld $13, x0; \
0128 pslld $(32 - 13), x4; \
0129 por x4, x0; \
0130 pxor x2, x1; \
0131 pxor x2, x3; \
0132 movdqa x2, x4; \
0133 psrld $3, x2; \
0134 pslld $(32 - 3), x4; \
0135 por x4, x2;
0136
0137 #define S0(x0, x1, x2, x3, x4) \
0138 movdqa x3, x4; \
0139 por x0, x3; \
0140 pxor x4, x0; \
0141 pxor x2, x4; \
0142 pxor RNOT, x4; \
0143 pxor x1, x3; \
0144 pand x0, x1; \
0145 pxor x4, x1; \
0146 pxor x0, x2; \
0147 pxor x3, x0; \
0148 por x0, x4; \
0149 pxor x2, x0; \
0150 pand x1, x2; \
0151 pxor x2, x3; \
0152 pxor RNOT, x1; \
0153 pxor x4, x2; \
0154 pxor x2, x1;
0155
0156 #define S1(x0, x1, x2, x3, x4) \
0157 movdqa x1, x4; \
0158 pxor x0, x1; \
0159 pxor x3, x0; \
0160 pxor RNOT, x3; \
0161 pand x1, x4; \
0162 por x1, x0; \
0163 pxor x2, x3; \
0164 pxor x3, x0; \
0165 pxor x3, x1; \
0166 pxor x4, x3; \
0167 por x4, x1; \
0168 pxor x2, x4; \
0169 pand x0, x2; \
0170 pxor x1, x2; \
0171 por x0, x1; \
0172 pxor RNOT, x0; \
0173 pxor x2, x0; \
0174 pxor x1, x4;
0175
0176 #define S2(x0, x1, x2, x3, x4) \
0177 pxor RNOT, x3; \
0178 pxor x0, x1; \
0179 movdqa x0, x4; \
0180 pand x2, x0; \
0181 pxor x3, x0; \
0182 por x4, x3; \
0183 pxor x1, x2; \
0184 pxor x1, x3; \
0185 pand x0, x1; \
0186 pxor x2, x0; \
0187 pand x3, x2; \
0188 por x1, x3; \
0189 pxor RNOT, x0; \
0190 pxor x0, x3; \
0191 pxor x0, x4; \
0192 pxor x2, x0; \
0193 por x2, x1;
0194
0195 #define S3(x0, x1, x2, x3, x4) \
0196 movdqa x1, x4; \
0197 pxor x3, x1; \
0198 por x0, x3; \
0199 pand x0, x4; \
0200 pxor x2, x0; \
0201 pxor x1, x2; \
0202 pand x3, x1; \
0203 pxor x3, x2; \
0204 por x4, x0; \
0205 pxor x3, x4; \
0206 pxor x0, x1; \
0207 pand x3, x0; \
0208 pand x4, x3; \
0209 pxor x2, x3; \
0210 por x1, x4; \
0211 pand x1, x2; \
0212 pxor x3, x4; \
0213 pxor x3, x0; \
0214 pxor x2, x3;
0215
0216 #define S4(x0, x1, x2, x3, x4) \
0217 movdqa x3, x4; \
0218 pand x0, x3; \
0219 pxor x4, x0; \
0220 pxor x2, x3; \
0221 por x4, x2; \
0222 pxor x1, x0; \
0223 pxor x3, x4; \
0224 por x0, x2; \
0225 pxor x1, x2; \
0226 pand x0, x1; \
0227 pxor x4, x1; \
0228 pand x2, x4; \
0229 pxor x3, x2; \
0230 pxor x0, x4; \
0231 por x1, x3; \
0232 pxor RNOT, x1; \
0233 pxor x0, x3;
0234
0235 #define S5(x0, x1, x2, x3, x4) \
0236 movdqa x1, x4; \
0237 por x0, x1; \
0238 pxor x1, x2; \
0239 pxor RNOT, x3; \
0240 pxor x0, x4; \
0241 pxor x2, x0; \
0242 pand x4, x1; \
0243 por x3, x4; \
0244 pxor x0, x4; \
0245 pand x3, x0; \
0246 pxor x3, x1; \
0247 pxor x2, x3; \
0248 pxor x1, x0; \
0249 pand x4, x2; \
0250 pxor x2, x1; \
0251 pand x0, x2; \
0252 pxor x2, x3;
0253
0254 #define S6(x0, x1, x2, x3, x4) \
0255 movdqa x1, x4; \
0256 pxor x0, x3; \
0257 pxor x2, x1; \
0258 pxor x0, x2; \
0259 pand x3, x0; \
0260 por x3, x1; \
0261 pxor RNOT, x4; \
0262 pxor x1, x0; \
0263 pxor x2, x1; \
0264 pxor x4, x3; \
0265 pxor x0, x4; \
0266 pand x0, x2; \
0267 pxor x1, x4; \
0268 pxor x3, x2; \
0269 pand x1, x3; \
0270 pxor x0, x3; \
0271 pxor x2, x1;
0272
0273 #define S7(x0, x1, x2, x3, x4) \
0274 pxor RNOT, x1; \
0275 movdqa x1, x4; \
0276 pxor RNOT, x0; \
0277 pand x2, x1; \
0278 pxor x3, x1; \
0279 por x4, x3; \
0280 pxor x2, x4; \
0281 pxor x3, x2; \
0282 pxor x0, x3; \
0283 por x1, x0; \
0284 pand x0, x2; \
0285 pxor x4, x0; \
0286 pxor x3, x4; \
0287 pand x0, x3; \
0288 pxor x1, x4; \
0289 pxor x4, x2; \
0290 pxor x1, x3; \
0291 por x0, x4; \
0292 pxor x1, x4;
0293
0294 #define SI0(x0, x1, x2, x3, x4) \
0295 movdqa x3, x4; \
0296 pxor x0, x1; \
0297 por x1, x3; \
0298 pxor x1, x4; \
0299 pxor RNOT, x0; \
0300 pxor x3, x2; \
0301 pxor x0, x3; \
0302 pand x1, x0; \
0303 pxor x2, x0; \
0304 pand x3, x2; \
0305 pxor x4, x3; \
0306 pxor x3, x2; \
0307 pxor x3, x1; \
0308 pand x0, x3; \
0309 pxor x0, x1; \
0310 pxor x2, x0; \
0311 pxor x3, x4;
0312
0313 #define SI1(x0, x1, x2, x3, x4) \
0314 pxor x3, x1; \
0315 movdqa x0, x4; \
0316 pxor x2, x0; \
0317 pxor RNOT, x2; \
0318 por x1, x4; \
0319 pxor x3, x4; \
0320 pand x1, x3; \
0321 pxor x2, x1; \
0322 pand x4, x2; \
0323 pxor x1, x4; \
0324 por x3, x1; \
0325 pxor x0, x3; \
0326 pxor x0, x2; \
0327 por x4, x0; \
0328 pxor x4, x2; \
0329 pxor x0, x1; \
0330 pxor x1, x4;
0331
0332 #define SI2(x0, x1, x2, x3, x4) \
0333 pxor x1, x2; \
0334 movdqa x3, x4; \
0335 pxor RNOT, x3; \
0336 por x2, x3; \
0337 pxor x4, x2; \
0338 pxor x0, x4; \
0339 pxor x1, x3; \
0340 por x2, x1; \
0341 pxor x0, x2; \
0342 pxor x4, x1; \
0343 por x3, x4; \
0344 pxor x3, x2; \
0345 pxor x2, x4; \
0346 pand x1, x2; \
0347 pxor x3, x2; \
0348 pxor x4, x3; \
0349 pxor x0, x4;
0350
0351 #define SI3(x0, x1, x2, x3, x4) \
0352 pxor x1, x2; \
0353 movdqa x1, x4; \
0354 pand x2, x1; \
0355 pxor x0, x1; \
0356 por x4, x0; \
0357 pxor x3, x4; \
0358 pxor x3, x0; \
0359 por x1, x3; \
0360 pxor x2, x1; \
0361 pxor x3, x1; \
0362 pxor x2, x0; \
0363 pxor x3, x2; \
0364 pand x1, x3; \
0365 pxor x0, x1; \
0366 pand x2, x0; \
0367 pxor x3, x4; \
0368 pxor x0, x3; \
0369 pxor x1, x0;
0370
0371 #define SI4(x0, x1, x2, x3, x4) \
0372 pxor x3, x2; \
0373 movdqa x0, x4; \
0374 pand x1, x0; \
0375 pxor x2, x0; \
0376 por x3, x2; \
0377 pxor RNOT, x4; \
0378 pxor x0, x1; \
0379 pxor x2, x0; \
0380 pand x4, x2; \
0381 pxor x0, x2; \
0382 por x4, x0; \
0383 pxor x3, x0; \
0384 pand x2, x3; \
0385 pxor x3, x4; \
0386 pxor x1, x3; \
0387 pand x0, x1; \
0388 pxor x1, x4; \
0389 pxor x3, x0;
0390
0391 #define SI5(x0, x1, x2, x3, x4) \
0392 movdqa x1, x4; \
0393 por x2, x1; \
0394 pxor x4, x2; \
0395 pxor x3, x1; \
0396 pand x4, x3; \
0397 pxor x3, x2; \
0398 por x0, x3; \
0399 pxor RNOT, x0; \
0400 pxor x2, x3; \
0401 por x0, x2; \
0402 pxor x1, x4; \
0403 pxor x4, x2; \
0404 pand x0, x4; \
0405 pxor x1, x0; \
0406 pxor x3, x1; \
0407 pand x2, x0; \
0408 pxor x3, x2; \
0409 pxor x2, x0; \
0410 pxor x4, x2; \
0411 pxor x3, x4;
0412
0413 #define SI6(x0, x1, x2, x3, x4) \
0414 pxor x2, x0; \
0415 movdqa x0, x4; \
0416 pand x3, x0; \
0417 pxor x3, x2; \
0418 pxor x2, x0; \
0419 pxor x1, x3; \
0420 por x4, x2; \
0421 pxor x3, x2; \
0422 pand x0, x3; \
0423 pxor RNOT, x0; \
0424 pxor x1, x3; \
0425 pand x2, x1; \
0426 pxor x0, x4; \
0427 pxor x4, x3; \
0428 pxor x2, x4; \
0429 pxor x1, x0; \
0430 pxor x0, x2;
0431
0432 #define SI7(x0, x1, x2, x3, x4) \
0433 movdqa x3, x4; \
0434 pand x0, x3; \
0435 pxor x2, x0; \
0436 por x4, x2; \
0437 pxor x1, x4; \
0438 pxor RNOT, x0; \
0439 por x3, x1; \
0440 pxor x0, x4; \
0441 pand x2, x0; \
0442 pxor x1, x0; \
0443 pand x2, x1; \
0444 pxor x2, x3; \
0445 pxor x3, x4; \
0446 pand x3, x2; \
0447 por x0, x3; \
0448 pxor x4, x1; \
0449 pxor x4, x3; \
0450 pand x0, x4; \
0451 pxor x2, x4;
0452
0453 #define transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
0454 movdqa x0, t2; \
0455 punpckldq x1, x0; \
0456 punpckhdq x1, t2; \
0457 movdqa x2, t1; \
0458 punpckhdq x3, x2; \
0459 punpckldq x3, t1; \
0460 movdqa x0, x1; \
0461 punpcklqdq t1, x0; \
0462 punpckhqdq t1, x1; \
0463 movdqa t2, x3; \
0464 punpcklqdq x2, t2; \
0465 punpckhqdq x2, x3; \
0466 movdqa t2, x2;
0467
0468 #define read_blocks(in, x0, x1, x2, x3, t0, t1, t2) \
0469 movdqu (0*4*4)(in), x0; \
0470 movdqu (1*4*4)(in), x1; \
0471 movdqu (2*4*4)(in), x2; \
0472 movdqu (3*4*4)(in), x3; \
0473 \
0474 transpose_4x4(x0, x1, x2, x3, t0, t1, t2)
0475
0476 #define write_blocks(out, x0, x1, x2, x3, t0, t1, t2) \
0477 transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
0478 \
0479 movdqu x0, (0*4*4)(out); \
0480 movdqu x1, (1*4*4)(out); \
0481 movdqu x2, (2*4*4)(out); \
0482 movdqu x3, (3*4*4)(out);
0483
0484 #define xor_blocks(out, x0, x1, x2, x3, t0, t1, t2) \
0485 transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
0486 \
0487 movdqu (0*4*4)(out), t0; \
0488 pxor t0, x0; \
0489 movdqu x0, (0*4*4)(out); \
0490 movdqu (1*4*4)(out), t0; \
0491 pxor t0, x1; \
0492 movdqu x1, (1*4*4)(out); \
0493 movdqu (2*4*4)(out), t0; \
0494 pxor t0, x2; \
0495 movdqu x2, (2*4*4)(out); \
0496 movdqu (3*4*4)(out), t0; \
0497 pxor t0, x3; \
0498 movdqu x3, (3*4*4)(out);
0499
0500 SYM_FUNC_START(__serpent_enc_blk_4way)
0501
0502
0503
0504
0505
0506
0507
0508 pcmpeqd RNOT, RNOT;
0509
0510 movl arg_ctx(%esp), CTX;
0511
0512 movl arg_src(%esp), %eax;
0513 read_blocks(%eax, RA, RB, RC, RD, RT0, RT1, RE);
0514
0515 K(RA, RB, RC, RD, RE, 0);
0516 S0(RA, RB, RC, RD, RE); LK(RC, RB, RD, RA, RE, 1);
0517 S1(RC, RB, RD, RA, RE); LK(RE, RD, RA, RC, RB, 2);
0518 S2(RE, RD, RA, RC, RB); LK(RB, RD, RE, RC, RA, 3);
0519 S3(RB, RD, RE, RC, RA); LK(RC, RA, RD, RB, RE, 4);
0520 S4(RC, RA, RD, RB, RE); LK(RA, RD, RB, RE, RC, 5);
0521 S5(RA, RD, RB, RE, RC); LK(RC, RA, RD, RE, RB, 6);
0522 S6(RC, RA, RD, RE, RB); LK(RD, RB, RA, RE, RC, 7);
0523 S7(RD, RB, RA, RE, RC); LK(RC, RA, RE, RD, RB, 8);
0524 S0(RC, RA, RE, RD, RB); LK(RE, RA, RD, RC, RB, 9);
0525 S1(RE, RA, RD, RC, RB); LK(RB, RD, RC, RE, RA, 10);
0526 S2(RB, RD, RC, RE, RA); LK(RA, RD, RB, RE, RC, 11);
0527 S3(RA, RD, RB, RE, RC); LK(RE, RC, RD, RA, RB, 12);
0528 S4(RE, RC, RD, RA, RB); LK(RC, RD, RA, RB, RE, 13);
0529 S5(RC, RD, RA, RB, RE); LK(RE, RC, RD, RB, RA, 14);
0530 S6(RE, RC, RD, RB, RA); LK(RD, RA, RC, RB, RE, 15);
0531 S7(RD, RA, RC, RB, RE); LK(RE, RC, RB, RD, RA, 16);
0532 S0(RE, RC, RB, RD, RA); LK(RB, RC, RD, RE, RA, 17);
0533 S1(RB, RC, RD, RE, RA); LK(RA, RD, RE, RB, RC, 18);
0534 S2(RA, RD, RE, RB, RC); LK(RC, RD, RA, RB, RE, 19);
0535 S3(RC, RD, RA, RB, RE); LK(RB, RE, RD, RC, RA, 20);
0536 S4(RB, RE, RD, RC, RA); LK(RE, RD, RC, RA, RB, 21);
0537 S5(RE, RD, RC, RA, RB); LK(RB, RE, RD, RA, RC, 22);
0538 S6(RB, RE, RD, RA, RC); LK(RD, RC, RE, RA, RB, 23);
0539 S7(RD, RC, RE, RA, RB); LK(RB, RE, RA, RD, RC, 24);
0540 S0(RB, RE, RA, RD, RC); LK(RA, RE, RD, RB, RC, 25);
0541 S1(RA, RE, RD, RB, RC); LK(RC, RD, RB, RA, RE, 26);
0542 S2(RC, RD, RB, RA, RE); LK(RE, RD, RC, RA, RB, 27);
0543 S3(RE, RD, RC, RA, RB); LK(RA, RB, RD, RE, RC, 28);
0544 S4(RA, RB, RD, RE, RC); LK(RB, RD, RE, RC, RA, 29);
0545 S5(RB, RD, RE, RC, RA); LK(RA, RB, RD, RC, RE, 30);
0546 S6(RA, RB, RD, RC, RE); LK(RD, RE, RB, RC, RA, 31);
0547 S7(RD, RE, RB, RC, RA); K(RA, RB, RC, RD, RE, 32);
0548
0549 movl arg_dst(%esp), %eax;
0550
0551 cmpb $0, arg_xor(%esp);
0552 jnz .L__enc_xor4;
0553
0554 write_blocks(%eax, RA, RB, RC, RD, RT0, RT1, RE);
0555
0556 RET;
0557
0558 .L__enc_xor4:
0559 xor_blocks(%eax, RA, RB, RC, RD, RT0, RT1, RE);
0560
0561 RET;
0562 SYM_FUNC_END(__serpent_enc_blk_4way)
0563
0564 SYM_FUNC_START(serpent_dec_blk_4way)
0565
0566
0567
0568
0569
0570
0571 pcmpeqd RNOT, RNOT;
0572
0573 movl arg_ctx(%esp), CTX;
0574
0575 movl arg_src(%esp), %eax;
0576 read_blocks(%eax, RA, RB, RC, RD, RT0, RT1, RE);
0577
0578 K(RA, RB, RC, RD, RE, 32);
0579 SI7(RA, RB, RC, RD, RE); KL(RB, RD, RA, RE, RC, 31);
0580 SI6(RB, RD, RA, RE, RC); KL(RA, RC, RE, RB, RD, 30);
0581 SI5(RA, RC, RE, RB, RD); KL(RC, RD, RA, RE, RB, 29);
0582 SI4(RC, RD, RA, RE, RB); KL(RC, RA, RB, RE, RD, 28);
0583 SI3(RC, RA, RB, RE, RD); KL(RB, RC, RD, RE, RA, 27);
0584 SI2(RB, RC, RD, RE, RA); KL(RC, RA, RE, RD, RB, 26);
0585 SI1(RC, RA, RE, RD, RB); KL(RB, RA, RE, RD, RC, 25);
0586 SI0(RB, RA, RE, RD, RC); KL(RE, RC, RA, RB, RD, 24);
0587 SI7(RE, RC, RA, RB, RD); KL(RC, RB, RE, RD, RA, 23);
0588 SI6(RC, RB, RE, RD, RA); KL(RE, RA, RD, RC, RB, 22);
0589 SI5(RE, RA, RD, RC, RB); KL(RA, RB, RE, RD, RC, 21);
0590 SI4(RA, RB, RE, RD, RC); KL(RA, RE, RC, RD, RB, 20);
0591 SI3(RA, RE, RC, RD, RB); KL(RC, RA, RB, RD, RE, 19);
0592 SI2(RC, RA, RB, RD, RE); KL(RA, RE, RD, RB, RC, 18);
0593 SI1(RA, RE, RD, RB, RC); KL(RC, RE, RD, RB, RA, 17);
0594 SI0(RC, RE, RD, RB, RA); KL(RD, RA, RE, RC, RB, 16);
0595 SI7(RD, RA, RE, RC, RB); KL(RA, RC, RD, RB, RE, 15);
0596 SI6(RA, RC, RD, RB, RE); KL(RD, RE, RB, RA, RC, 14);
0597 SI5(RD, RE, RB, RA, RC); KL(RE, RC, RD, RB, RA, 13);
0598 SI4(RE, RC, RD, RB, RA); KL(RE, RD, RA, RB, RC, 12);
0599 SI3(RE, RD, RA, RB, RC); KL(RA, RE, RC, RB, RD, 11);
0600 SI2(RA, RE, RC, RB, RD); KL(RE, RD, RB, RC, RA, 10);
0601 SI1(RE, RD, RB, RC, RA); KL(RA, RD, RB, RC, RE, 9);
0602 SI0(RA, RD, RB, RC, RE); KL(RB, RE, RD, RA, RC, 8);
0603 SI7(RB, RE, RD, RA, RC); KL(RE, RA, RB, RC, RD, 7);
0604 SI6(RE, RA, RB, RC, RD); KL(RB, RD, RC, RE, RA, 6);
0605 SI5(RB, RD, RC, RE, RA); KL(RD, RA, RB, RC, RE, 5);
0606 SI4(RD, RA, RB, RC, RE); KL(RD, RB, RE, RC, RA, 4);
0607 SI3(RD, RB, RE, RC, RA); KL(RE, RD, RA, RC, RB, 3);
0608 SI2(RE, RD, RA, RC, RB); KL(RD, RB, RC, RA, RE, 2);
0609 SI1(RD, RB, RC, RA, RE); KL(RE, RB, RC, RA, RD, 1);
0610 SI0(RE, RB, RC, RA, RD); K(RC, RD, RB, RE, RA, 0);
0611
0612 movl arg_dst(%esp), %eax;
0613 write_blocks(%eax, RC, RD, RB, RE, RT0, RT1, RA);
0614
0615 RET;
0616 SYM_FUNC_END(serpent_dec_blk_4way)