www.gitorious.org Git - gnutls:gnutls.git/blob - lib/accelerated/x86/coff/padlock-x86-coff.s

3 #

4 # Redistribution and use in source and binary forms, with or without

5 # modification, are permitted provided that the following conditions

6 # are met:

7 #

8 # * Redistributions of source code must retain copyright notices,

9 # this list of conditions and the following disclaimer.

10 #

11 # * Redistributions in binary form must reproduce the above

12 # copyright notice, this list of conditions and the following

13 # disclaimer in the documentation and/or other materials

14 # provided with the distribution.

15 #

16 # * Neither the name of the Andy Polyakov nor the names of its

17 # copyright holder and contributors may be used to endorse or

18 # promote products derived from this software without specific

19 # prior written permission.

20 #

21 # ALTERNATIVELY, provided that this notice is retained in full, this

22 # product may be distributed under the terms of the GNU General Public

23 # License (GPL), in which case the provisions of the GPL apply INSTEAD OF

24 # those given above.

25 #

26 # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER AND CONTRIBUTORS

27 # "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT

28 # LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR

29 # A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT

30 # OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,

31 # SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT

32 # LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,

33 # DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY

34 # THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT

35 # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE

36 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

38 .file "./engines/asm/e_padlock-x86.s"

39 .text

40 .globl _padlock_capability

41 .def _padlock_capability; .scl 2; .type 32; .endef

42 .align 16

43 _padlock_capability:

44 .L_padlock_capability_begin:

45 pushl %ebx

46 pushfl

47 popl %eax

48 movl %eax,%ecx

49 xorl $2097152,%eax

50 pushl %eax

51 popfl

52 pushfl

53 popl %eax

54 xorl %eax,%ecx

55 xorl %eax,%eax

56 btl $21,%ecx

57 jnc .L000noluck

58 .byte 0x0f,0xa2

59 xorl %eax,%eax

60 cmpl $0x746e6543,%ebx

61 jne .L000noluck

62 cmpl $0x48727561,%edx

63 jne .L000noluck

64 cmpl $0x736c7561,%ecx

65 jne .L000noluck

66 movl $3221225472,%eax

67 .byte 0x0f,0xa2

68 movl %eax,%edx

69 xorl %eax,%eax

70 cmpl $3221225473,%edx

71 jb .L000noluck

72 movl $1,%eax

73 .byte 0x0f,0xa2

74 orl $15,%eax

75 xorl %ebx,%ebx

76 andl $4095,%eax

77 cmpl $1791,%eax

78 sete %bl

79 movl $3221225473,%eax

80 pushl %ebx

81 .byte 0x0f,0xa2

82 popl %ebx

83 movl %edx,%eax

84 shll $4,%ebx

85 andl $4294967279,%eax

86 orl %ebx,%eax

87 .L000noluck:

88 popl %ebx

89 ret

90 .globl _padlock_key_bswap

91 .def _padlock_key_bswap; .scl 2; .type 32; .endef

92 .align 16

93 _padlock_key_bswap:

94 .L_padlock_key_bswap_begin:

95 movl 4(%esp),%edx

96 movl 240(%edx),%ecx

97 .L001bswap_loop:

98 movl (%edx),%eax

99 bswap %eax

100 movl %eax,(%edx)

101 leal 4(%edx),%edx

102 subl $1,%ecx

103 jnz .L001bswap_loop

104 ret

105 .globl _padlock_verify_context

106 .def _padlock_verify_context; .scl 2; .type 32; .endef

107 .align 16

108 _padlock_verify_context:

109 .L_padlock_verify_context_begin:

110 movl 4(%esp),%edx

111 leal .Lpadlock_saved_context,%eax

112 pushfl

113 call __padlock_verify_ctx

114 .L002verify_pic_point:

115 leal 4(%esp),%esp

116 ret

117 .def __padlock_verify_ctx; .scl 3; .type 32; .endef

118 .align 16

119 __padlock_verify_ctx:

120 btl $30,4(%esp)

121 jnc .L003verified

122 cmpl (%eax),%edx

123 je .L003verified

124 pushfl

125 popfl

126 .L003verified:

127 movl %edx,(%eax)

128 ret

129 .globl _padlock_reload_key

130 .def _padlock_reload_key; .scl 2; .type 32; .endef

131 .align 16

132 _padlock_reload_key:

133 .L_padlock_reload_key_begin:

134 pushfl

135 popfl

136 ret

137 .globl _padlock_aes_block

138 .def _padlock_aes_block; .scl 2; .type 32; .endef

139 .align 16

140 _padlock_aes_block:

141 .L_padlock_aes_block_begin:

142 pushl %edi

143 pushl %esi

144 pushl %ebx

145 movl 16(%esp),%edi

146 movl 20(%esp),%esi

147 movl 24(%esp),%edx

148 movl $1,%ecx

149 leal 32(%edx),%ebx

150 leal 16(%edx),%edx

151 .byte 243,15,167,200

152 popl %ebx

153 popl %esi

154 popl %edi

155 ret

156 .globl _padlock_ecb_encrypt

157 .def _padlock_ecb_encrypt; .scl 2; .type 32; .endef

158 .align 16

159 _padlock_ecb_encrypt:

160 .L_padlock_ecb_encrypt_begin:

161 pushl %ebp

162 pushl %ebx

163 pushl %esi

164 pushl %edi

165 movl 20(%esp),%edi

166 movl 24(%esp),%esi

167 movl 28(%esp),%edx

168 movl 32(%esp),%ecx

169 testl $15,%edx

170 jnz .L004ecb_abort

171 testl $15,%ecx

172 jnz .L004ecb_abort

173 leal .Lpadlock_saved_context,%eax

174 pushfl

175 cld

176 call __padlock_verify_ctx

177 .L005ecb_pic_point:

178 leal 16(%edx),%edx

179 xorl %eax,%eax

180 xorl %ebx,%ebx

181 cmpl $128,%ecx

182 jbe .L006ecb_short

183 testl $32,(%edx)

184 jnz .L007ecb_aligned

185 testl $15,%edi

186 setz %al

187 testl $15,%esi

188 setz %bl

189 testl %ebx,%eax

190 jnz .L007ecb_aligned

191 negl %eax

192 movl $512,%ebx

193 notl %eax

194 leal -24(%esp),%ebp

195 cmpl %ebx,%ecx

196 cmovcl %ecx,%ebx

197 andl %ebx,%eax

198 movl %ecx,%ebx

199 negl %eax

200 andl $511,%ebx

201 leal (%eax,%ebp,1),%esp

202 andl $-16,%esp

203 jmp .L008ecb_loop

204 .align 16

205 .L008ecb_loop:

206 movl %edi,(%ebp)

207 movl %esi,4(%ebp)

208 movl %ecx,8(%ebp)

209 movl %ebx,%ecx

210 movl %ebx,12(%ebp)

211 testl $15,%edi

212 cmovnzl %esp,%edi

213 testl $15,%esi

214 jz .L009ecb_inp_aligned

215 shrl $2,%ecx

216 .byte 243,165

217 subl %ebx,%edi

218 movl %ebx,%ecx

219 movl %edi,%esi

220 .L009ecb_inp_aligned:

221 leal -16(%edx),%eax

222 leal 16(%edx),%ebx

223 shrl $4,%ecx

224 .byte 243,15,167,200

225 movl (%ebp),%edi

226 movl 12(%ebp),%ebx

227 testl $15,%edi

228 jz .L010ecb_out_aligned

229 movl %ebx,%ecx

230 shrl $2,%ecx

231 leal (%esp),%esi

232 .byte 243,165

233 subl %ebx,%edi

234 .L010ecb_out_aligned:

235 movl 4(%ebp),%esi

236 movl 8(%ebp),%ecx

237 addl %ebx,%edi

238 addl %ebx,%esi

239 subl %ebx,%ecx

240 movl $512,%ebx

241 jnz .L008ecb_loop

242 cmpl %ebp,%esp

243 je .L011ecb_done

244 pxor %xmm0,%xmm0

245 leal (%esp),%eax

246 .L012ecb_bzero:

247 movaps %xmm0,(%eax)

248 leal 16(%eax),%eax

249 cmpl %eax,%ebp

250 ja .L012ecb_bzero

251 .L011ecb_done:

252 leal 24(%ebp),%esp

253 jmp .L013ecb_exit

254 .align 16

255 .L006ecb_short:

256 xorl %eax,%eax

257 leal -24(%esp),%ebp

258 subl %ecx,%eax

259 leal (%eax,%ebp,1),%esp

260 andl $-16,%esp

261 xorl %ebx,%ebx

262 .L014ecb_short_copy:

263 movups (%esi,%ebx,1),%xmm0

264 leal 16(%ebx),%ebx

265 cmpl %ebx,%ecx

266 movaps %xmm0,-16(%esp,%ebx,1)

267 ja .L014ecb_short_copy

268 movl %esp,%esi

269 movl %ecx,%ebx

270 jmp .L008ecb_loop

271 .align 16

272 .L007ecb_aligned:

273 leal -16(%edx),%eax

274 leal 16(%edx),%ebx

275 shrl $4,%ecx

276 .byte 243,15,167,200

277 .L013ecb_exit:

278 movl $1,%eax

279 leal 4(%esp),%esp

280 .L004ecb_abort:

281 popl %edi

282 popl %esi

283 popl %ebx

284 popl %ebp

285 ret

286 .globl _padlock_cbc_encrypt

287 .def _padlock_cbc_encrypt; .scl 2; .type 32; .endef

288 .align 16

289 _padlock_cbc_encrypt:

290 .L_padlock_cbc_encrypt_begin:

291 pushl %ebp

292 pushl %ebx

293 pushl %esi

294 pushl %edi

295 movl 20(%esp),%edi

296 movl 24(%esp),%esi

297 movl 28(%esp),%edx

298 movl 32(%esp),%ecx

299 testl $15,%edx

300 jnz .L015cbc_abort

301 testl $15,%ecx

302 jnz .L015cbc_abort

303 leal .Lpadlock_saved_context,%eax

304 pushfl

305 cld

306 call __padlock_verify_ctx

307 .L016cbc_pic_point:

308 leal 16(%edx),%edx

309 xorl %eax,%eax

310 xorl %ebx,%ebx

311 cmpl $64,%ecx

312 jbe .L017cbc_short

313 testl $32,(%edx)

314 jnz .L018cbc_aligned

315 testl $15,%edi

316 setz %al

317 testl $15,%esi

318 setz %bl

319 testl %ebx,%eax

320 jnz .L018cbc_aligned

321 negl %eax

322 movl $512,%ebx

323 notl %eax

324 leal -24(%esp),%ebp

325 cmpl %ebx,%ecx

326 cmovcl %ecx,%ebx

327 andl %ebx,%eax

328 movl %ecx,%ebx

329 negl %eax

330 andl $511,%ebx

331 leal (%eax,%ebp,1),%esp

332 andl $-16,%esp

333 jmp .L019cbc_loop

334 .align 16

335 .L019cbc_loop:

336 movl %edi,(%ebp)

337 movl %esi,4(%ebp)

338 movl %ecx,8(%ebp)

339 movl %ebx,%ecx

340 movl %ebx,12(%ebp)

341 testl $15,%edi

342 cmovnzl %esp,%edi

343 testl $15,%esi

344 jz .L020cbc_inp_aligned

345 shrl $2,%ecx

346 .byte 243,165

347 subl %ebx,%edi

348 movl %ebx,%ecx

349 movl %edi,%esi

350 .L020cbc_inp_aligned:

351 leal -16(%edx),%eax

352 leal 16(%edx),%ebx

353 shrl $4,%ecx

354 .byte 243,15,167,208

355 movaps (%eax),%xmm0

356 movaps %xmm0,-16(%edx)

357 movl (%ebp),%edi

358 movl 12(%ebp),%ebx

359 testl $15,%edi

360 jz .L021cbc_out_aligned

361 movl %ebx,%ecx

362 shrl $2,%ecx

363 leal (%esp),%esi

364 .byte 243,165

365 subl %ebx,%edi

366 .L021cbc_out_aligned:

367 movl 4(%ebp),%esi

368 movl 8(%ebp),%ecx

369 addl %ebx,%edi

370 addl %ebx,%esi

371 subl %ebx,%ecx

372 movl $512,%ebx

373 jnz .L019cbc_loop

374 cmpl %ebp,%esp

375 je .L022cbc_done

376 pxor %xmm0,%xmm0

377 leal (%esp),%eax

378 .L023cbc_bzero:

379 movaps %xmm0,(%eax)

380 leal 16(%eax),%eax

381 cmpl %eax,%ebp

382 ja .L023cbc_bzero

383 .L022cbc_done:

384 leal 24(%ebp),%esp

385 jmp .L024cbc_exit

386 .align 16

387 .L017cbc_short:

388 xorl %eax,%eax

389 leal -24(%esp),%ebp

390 subl %ecx,%eax

391 leal (%eax,%ebp,1),%esp

392 andl $-16,%esp

393 xorl %ebx,%ebx

394 .L025cbc_short_copy:

395 movups (%esi,%ebx,1),%xmm0

396 leal 16(%ebx),%ebx

397 cmpl %ebx,%ecx

398 movaps %xmm0,-16(%esp,%ebx,1)

399 ja .L025cbc_short_copy

400 movl %esp,%esi

401 movl %ecx,%ebx

402 jmp .L019cbc_loop

403 .align 16

404 .L018cbc_aligned:

405 leal -16(%edx),%eax

406 leal 16(%edx),%ebx

407 shrl $4,%ecx

408 .byte 243,15,167,208

409 movaps (%eax),%xmm0

410 movaps %xmm0,-16(%edx)

411 .L024cbc_exit:

412 movl $1,%eax

413 leal 4(%esp),%esp

414 .L015cbc_abort:

415 popl %edi

416 popl %esi

417 popl %ebx

418 popl %ebp

419 ret

420 .globl _padlock_xstore

421 .def _padlock_xstore; .scl 2; .type 32; .endef

422 .align 16

423 _padlock_xstore:

424 .L_padlock_xstore_begin:

425 pushl %edi

426 movl 8(%esp),%edi

427 movl 12(%esp),%edx

428 .byte 15,167,192

429 popl %edi

430 ret

431 .def __win32_segv_handler; .scl 3; .type 32; .endef

432 .align 16

433 __win32_segv_handler:

434 movl $1,%eax

435 movl 4(%esp),%edx

436 movl 12(%esp),%ecx

437 cmpl $3221225477,(%edx)

438 jne .L026ret

439 addl $4,184(%ecx)

440 movl $0,%eax

441 .L026ret:

442 ret

443 .globl _padlock_sha1_oneshot

444 .def _padlock_sha1_oneshot; .scl 2; .type 32; .endef

445 .align 16

446 _padlock_sha1_oneshot:

447 .L_padlock_sha1_oneshot_begin:

448 pushl %edi

449 pushl %esi

450 xorl %eax,%eax

451 movl 12(%esp),%edi

452 movl 16(%esp),%esi

453 movl 20(%esp),%ecx

454 pushl __win32_segv_handler

455 .byte 100,255,48

456 .byte 100,137,32

457 movl %esp,%edx

458 addl $-128,%esp

459 movups (%edi),%xmm0

460 andl $-16,%esp

461 movl 16(%edi),%eax

462 movaps %xmm0,(%esp)

463 movl %esp,%edi

464 movl %eax,16(%esp)

465 xorl %eax,%eax

466 .byte 243,15,166,200

467 movaps (%esp),%xmm0

468 movl 16(%esp),%eax

469 movl %edx,%esp

470 .byte 100,143,5,0,0,0,0

471 leal 4(%esp),%esp

472 movl 16(%esp),%edi

473 movups %xmm0,(%edi)

474 movl %eax,16(%edi)

475 popl %esi

476 popl %edi

477 ret

478 .globl _padlock_sha1_blocks

479 .def _padlock_sha1_blocks; .scl 2; .type 32; .endef

480 .align 16

481 _padlock_sha1_blocks:

482 .L_padlock_sha1_blocks_begin:

483 pushl %edi

484 pushl %esi

485 movl 12(%esp),%edi

486 movl 16(%esp),%esi

487 movl %esp,%edx

488 movl 20(%esp),%ecx

489 addl $-128,%esp

490 movups (%edi),%xmm0

491 andl $-16,%esp

492 movl 16(%edi),%eax

493 movaps %xmm0,(%esp)

494 movl %esp,%edi

495 movl %eax,16(%esp)

496 movl $-1,%eax

497 .byte 243,15,166,200

498 movaps (%esp),%xmm0

499 movl 16(%esp),%eax

500 movl %edx,%esp

501 movl 12(%esp),%edi

502 movups %xmm0,(%edi)

503 movl %eax,16(%edi)

504 popl %esi

505 popl %edi

506 ret

507 .globl _padlock_sha256_oneshot

508 .def _padlock_sha256_oneshot; .scl 2; .type 32; .endef

509 .align 16

510 _padlock_sha256_oneshot:

511 .L_padlock_sha256_oneshot_begin:

512 pushl %edi

513 pushl %esi

514 xorl %eax,%eax

515 movl 12(%esp),%edi

516 movl 16(%esp),%esi

517 movl 20(%esp),%ecx

518 pushl __win32_segv_handler

519 .byte 100,255,48

520 .byte 100,137,32

521 movl %esp,%edx

522 addl $-128,%esp

523 movups (%edi),%xmm0

524 andl $-16,%esp

525 movups 16(%edi),%xmm1

526 movaps %xmm0,(%esp)

527 movl %esp,%edi

528 movaps %xmm1,16(%esp)

529 xorl %eax,%eax

530 .byte 243,15,166,208

531 movaps (%esp),%xmm0

532 movaps 16(%esp),%xmm1

533 movl %edx,%esp

534 .byte 100,143,5,0,0,0,0

535 leal 4(%esp),%esp

536 movl 16(%esp),%edi

537 movups %xmm0,(%edi)

538 movups %xmm1,16(%edi)

539 popl %esi

540 popl %edi

541 ret

542 .globl _padlock_sha256_blocks

543 .def _padlock_sha256_blocks; .scl 2; .type 32; .endef

544 .align 16

545 _padlock_sha256_blocks:

546 .L_padlock_sha256_blocks_begin:

547 pushl %edi

548 pushl %esi

549 movl 12(%esp),%edi

550 movl 16(%esp),%esi

551 movl 20(%esp),%ecx

552 movl %esp,%edx

553 addl $-128,%esp

554 movups (%edi),%xmm0

555 andl $-16,%esp

556 movups 16(%edi),%xmm1

557 movaps %xmm0,(%esp)

558 movl %esp,%edi

559 movaps %xmm1,16(%esp)

560 movl $-1,%eax

561 .byte 243,15,166,208

562 movaps (%esp),%xmm0

563 movaps 16(%esp),%xmm1

564 movl %edx,%esp

565 movl 12(%esp),%edi

566 movups %xmm0,(%edi)

567 movups %xmm1,16(%edi)

568 popl %esi

569 popl %edi

570 ret

571 .globl _padlock_sha512_blocks

572 .def _padlock_sha512_blocks; .scl 2; .type 32; .endef

573 .align 16

574 _padlock_sha512_blocks:

575 .L_padlock_sha512_blocks_begin:

576 pushl %edi

577 pushl %esi

578 movl 12(%esp),%edi

579 movl 16(%esp),%esi

580 movl 20(%esp),%ecx

581 movl %esp,%edx

582 addl $-128,%esp

583 movups (%edi),%xmm0

584 andl $-16,%esp

585 movups 16(%edi),%xmm1

586 movups 32(%edi),%xmm2

587 movups 48(%edi),%xmm3

588 movaps %xmm0,(%esp)

589 movl %esp,%edi

590 movaps %xmm1,16(%esp)

591 movaps %xmm2,32(%esp)

592 movaps %xmm3,48(%esp)

593 .byte 243,15,166,224

594 movaps (%esp),%xmm0

595 movaps 16(%esp),%xmm1

596 movaps 32(%esp),%xmm2

597 movaps 48(%esp),%xmm3

598 movl %edx,%esp

599 movl 12(%esp),%edi

600 movups %xmm0,(%edi)

601 movups %xmm1,16(%edi)

602 movups %xmm2,32(%edi)

603 movups %xmm3,48(%edi)

604 popl %esi

605 popl %edi

606 ret

607 .byte 86,73,65,32,80,97,100,108,111,99,107,32,120,56,54,32

608 .byte 109,111,100,117,108,101,44,32,67,82,89,80,84,79,71,65

609 .byte 77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101

610 .byte 110,115,115,108,46,111,114,103,62,0

611 .align 16

612 .data

613 .align 4

614 .Lpadlock_saved_context:

615 .long 0