aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--arch/x86/crypto/cast5-avx-x86_64-asm_64.S48
1 files changed, 18 insertions, 30 deletions
diff --git a/arch/x86/crypto/cast5-avx-x86_64-asm_64.S b/arch/x86/crypto/cast5-avx-x86_64-asm_64.S
index 15b00ac7cbd3..c35fd5d6ecd2 100644
--- a/arch/x86/crypto/cast5-avx-x86_64-asm_64.S
+++ b/arch/x86/crypto/cast5-avx-x86_64-asm_64.S
@@ -23,6 +23,8 @@
23 * 23 *
24 */ 24 */
25 25
26#include <linux/linkage.h>
27
26.file "cast5-avx-x86_64-asm_64.S" 28.file "cast5-avx-x86_64-asm_64.S"
27 29
28.extern cast_s1 30.extern cast_s1
@@ -211,8 +213,6 @@
211.text 213.text
212 214
213.align 16 215.align 16
214.type __cast5_enc_blk16,@function;
215
216__cast5_enc_blk16: 216__cast5_enc_blk16:
217 /* input: 217 /* input:
218 * %rdi: ctx, CTX 218 * %rdi: ctx, CTX
@@ -263,14 +263,14 @@ __cast5_enc_blk16:
263 263
264 movzbl rr(CTX), %eax; 264 movzbl rr(CTX), %eax;
265 testl %eax, %eax; 265 testl %eax, %eax;
266 jnz __skip_enc; 266 jnz .L__skip_enc;
267 267
268 round(RL, RR, 12, 1); 268 round(RL, RR, 12, 1);
269 round(RR, RL, 13, 2); 269 round(RR, RL, 13, 2);
270 round(RL, RR, 14, 3); 270 round(RL, RR, 14, 3);
271 round(RR, RL, 15, 1); 271 round(RR, RL, 15, 1);
272 272
273__skip_enc: 273.L__skip_enc:
274 popq %rbx; 274 popq %rbx;
275 popq %rbp; 275 popq %rbp;
276 276
@@ -282,10 +282,9 @@ __skip_enc:
282 outunpack_blocks(RR4, RL4, RTMP, RX, RKM); 282 outunpack_blocks(RR4, RL4, RTMP, RX, RKM);
283 283
284 ret; 284 ret;
285ENDPROC(__cast5_enc_blk16)
285 286
286.align 16 287.align 16
287.type __cast5_dec_blk16,@function;
288
289__cast5_dec_blk16: 288__cast5_dec_blk16:
290 /* input: 289 /* input:
291 * %rdi: ctx, CTX 290 * %rdi: ctx, CTX
@@ -323,14 +322,14 @@ __cast5_dec_blk16:
323 322
324 movzbl rr(CTX), %eax; 323 movzbl rr(CTX), %eax;
325 testl %eax, %eax; 324 testl %eax, %eax;
326 jnz __skip_dec; 325 jnz .L__skip_dec;
327 326
328 round(RL, RR, 15, 1); 327 round(RL, RR, 15, 1);
329 round(RR, RL, 14, 3); 328 round(RR, RL, 14, 3);
330 round(RL, RR, 13, 2); 329 round(RL, RR, 13, 2);
331 round(RR, RL, 12, 1); 330 round(RR, RL, 12, 1);
332 331
333__dec_tail: 332.L__dec_tail:
334 round(RL, RR, 11, 3); 333 round(RL, RR, 11, 3);
335 round(RR, RL, 10, 2); 334 round(RR, RL, 10, 2);
336 round(RL, RR, 9, 1); 335 round(RL, RR, 9, 1);
@@ -355,15 +354,12 @@ __dec_tail:
355 354
356 ret; 355 ret;
357 356
358__skip_dec: 357.L__skip_dec:
359 vpsrldq $4, RKR, RKR; 358 vpsrldq $4, RKR, RKR;
360 jmp __dec_tail; 359 jmp .L__dec_tail;
360ENDPROC(__cast5_dec_blk16)
361 361
362.align 16 362ENTRY(cast5_ecb_enc_16way)
363.global cast5_ecb_enc_16way
364.type cast5_ecb_enc_16way,@function;
365
366cast5_ecb_enc_16way:
367 /* input: 363 /* input:
368 * %rdi: ctx, CTX 364 * %rdi: ctx, CTX
369 * %rsi: dst 365 * %rsi: dst
@@ -393,12 +389,9 @@ cast5_ecb_enc_16way:
393 vmovdqu RL4, (7*4*4)(%r11); 389 vmovdqu RL4, (7*4*4)(%r11);
394 390
395 ret; 391 ret;
392ENDPROC(cast5_ecb_enc_16way)
396 393
397.align 16 394ENTRY(cast5_ecb_dec_16way)
398.global cast5_ecb_dec_16way
399.type cast5_ecb_dec_16way,@function;
400
401cast5_ecb_dec_16way:
402 /* input: 395 /* input:
403 * %rdi: ctx, CTX 396 * %rdi: ctx, CTX
404 * %rsi: dst 397 * %rsi: dst
@@ -428,12 +421,9 @@ cast5_ecb_dec_16way:
428 vmovdqu RL4, (7*4*4)(%r11); 421 vmovdqu RL4, (7*4*4)(%r11);
429 422
430 ret; 423 ret;
424ENDPROC(cast5_ecb_dec_16way)
431 425
432.align 16 426ENTRY(cast5_cbc_dec_16way)
433.global cast5_cbc_dec_16way
434.type cast5_cbc_dec_16way,@function;
435
436cast5_cbc_dec_16way:
437 /* input: 427 /* input:
438 * %rdi: ctx, CTX 428 * %rdi: ctx, CTX
439 * %rsi: dst 429 * %rsi: dst
@@ -480,12 +470,9 @@ cast5_cbc_dec_16way:
480 popq %r12; 470 popq %r12;
481 471
482 ret; 472 ret;
473ENDPROC(cast5_cbc_dec_16way)
483 474
484.align 16 475ENTRY(cast5_ctr_16way)
485.global cast5_ctr_16way
486.type cast5_ctr_16way,@function;
487
488cast5_ctr_16way:
489 /* input: 476 /* input:
490 * %rdi: ctx, CTX 477 * %rdi: ctx, CTX
491 * %rsi: dst 478 * %rsi: dst
@@ -556,3 +543,4 @@ cast5_ctr_16way:
556 popq %r12; 543 popq %r12;
557 544
558 ret; 545 ret;
546ENDPROC(cast5_ctr_16way)