Linux Audio

Check our new training course

Loading...
v3.5.6
 
  1/*
  2 * This is a SIMD SHA-1 implementation. It requires the Intel(R) Supplemental
  3 * SSE3 instruction set extensions introduced in Intel Core Microarchitecture
  4 * processors. CPUs supporting Intel(R) AVX extensions will get an additional
  5 * boost.
  6 *
  7 * This work was inspired by the vectorized implementation of Dean Gaudet.
  8 * Additional information on it can be found at:
  9 *    http://www.arctic.org/~dean/crypto/sha1.html
 10 *
 11 * It was improved upon with more efficient vectorization of the message
 12 * scheduling. This implementation has also been optimized for all current and
 13 * several future generations of Intel CPUs.
 14 *
 15 * See this article for more information about the implementation details:
 16 *   http://software.intel.com/en-us/articles/improving-the-performance-of-the-secure-hash-algorithm-1/
 17 *
 18 * Copyright (C) 2010, Intel Corp.
 19 *   Authors: Maxim Locktyukhin <maxim.locktyukhin@intel.com>
 20 *            Ronen Zohar <ronen.zohar@intel.com>
 21 *
 22 * Converted to AT&T syntax and adapted for inclusion in the Linux kernel:
 23 *   Author: Mathias Krause <minipli@googlemail.com>
 24 *
 25 * This program is free software; you can redistribute it and/or modify
 26 * it under the terms of the GNU General Public License as published by
 27 * the Free Software Foundation; either version 2 of the License, or
 28 * (at your option) any later version.
 29 */
 30
 
 
 31#define CTX	%rdi	// arg1
 32#define BUF	%rsi	// arg2
 33#define CNT	%rdx	// arg3
 34
 35#define REG_A	%ecx
 36#define REG_B	%esi
 37#define REG_C	%edi
 38#define REG_D	%ebp
 39#define REG_E	%edx
 40
 41#define REG_T1	%eax
 42#define REG_T2	%ebx
 43
 44#define K_BASE		%r8
 45#define HASH_PTR	%r9
 46#define BUFFER_PTR	%r10
 47#define BUFFER_END	%r11
 48
 49#define W_TMP1	%xmm0
 50#define W_TMP2	%xmm9
 51
 52#define W0	%xmm1
 53#define W4	%xmm2
 54#define W8	%xmm3
 55#define W12	%xmm4
 56#define W16	%xmm5
 57#define W20	%xmm6
 58#define W24	%xmm7
 59#define W28	%xmm8
 60
 61#define XMM_SHUFB_BSWAP	%xmm10
 62
 63/* we keep window of 64 w[i]+K pre-calculated values in a circular buffer */
 64#define WK(t)	(((t) & 15) * 4)(%rsp)
 65#define W_PRECALC_AHEAD	16
 66
 67/*
 68 * This macro implements the SHA-1 function's body for single 64-byte block
 69 * param: function's name
 70 */
 71.macro SHA1_VECTOR_ASM  name
 72	.global	\name
 73	.type	\name, @function
 74	.align 32
 75\name:
 76	push	%rbx
 77	push	%rbp
 78	push	%r12
 
 
 79
 80	mov	%rsp, %r12
 81	sub	$64, %rsp		# allocate workspace
 82	and	$~15, %rsp		# align stack
 83
 84	mov	CTX, HASH_PTR
 85	mov	BUF, BUFFER_PTR
 86
 87	shl	$6, CNT			# multiply by 64
 88	add	BUF, CNT
 89	mov	CNT, BUFFER_END
 90
 91	lea	K_XMM_AR(%rip), K_BASE
 92	xmm_mov	BSWAP_SHUFB_CTL(%rip), XMM_SHUFB_BSWAP
 93
 94	SHA1_PIPELINED_MAIN_BODY
 95
 96	# cleanup workspace
 97	mov	$8, %ecx
 98	mov	%rsp, %rdi
 99	xor	%rax, %rax
100	rep stosq
101
102	mov	%r12, %rsp		# deallocate workspace
103
104	pop	%r12
105	pop	%rbp
 
106	pop	%rbx
107	ret
108
109	.size	\name, .-\name
110.endm
111
112/*
113 * This macro implements 80 rounds of SHA-1 for one 64-byte block
114 */
115.macro SHA1_PIPELINED_MAIN_BODY
116	INIT_REGALLOC
117
118	mov	  (HASH_PTR), A
119	mov	 4(HASH_PTR), B
120	mov	 8(HASH_PTR), C
121	mov	12(HASH_PTR), D
122	mov	16(HASH_PTR), E
123
124  .set i, 0
125  .rept W_PRECALC_AHEAD
126	W_PRECALC i
127    .set i, (i+1)
128  .endr
129
130.align 4
1311:
132	RR F1,A,B,C,D,E,0
133	RR F1,D,E,A,B,C,2
134	RR F1,B,C,D,E,A,4
135	RR F1,E,A,B,C,D,6
136	RR F1,C,D,E,A,B,8
137
138	RR F1,A,B,C,D,E,10
139	RR F1,D,E,A,B,C,12
140	RR F1,B,C,D,E,A,14
141	RR F1,E,A,B,C,D,16
142	RR F1,C,D,E,A,B,18
143
144	RR F2,A,B,C,D,E,20
145	RR F2,D,E,A,B,C,22
146	RR F2,B,C,D,E,A,24
147	RR F2,E,A,B,C,D,26
148	RR F2,C,D,E,A,B,28
149
150	RR F2,A,B,C,D,E,30
151	RR F2,D,E,A,B,C,32
152	RR F2,B,C,D,E,A,34
153	RR F2,E,A,B,C,D,36
154	RR F2,C,D,E,A,B,38
155
156	RR F3,A,B,C,D,E,40
157	RR F3,D,E,A,B,C,42
158	RR F3,B,C,D,E,A,44
159	RR F3,E,A,B,C,D,46
160	RR F3,C,D,E,A,B,48
161
162	RR F3,A,B,C,D,E,50
163	RR F3,D,E,A,B,C,52
164	RR F3,B,C,D,E,A,54
165	RR F3,E,A,B,C,D,56
166	RR F3,C,D,E,A,B,58
167
168	add	$64, BUFFER_PTR		# move to the next 64-byte block
169	cmp	BUFFER_END, BUFFER_PTR	# if the current is the last one use
170	cmovae	K_BASE, BUFFER_PTR	# dummy source to avoid buffer overrun
171
172	RR F4,A,B,C,D,E,60
173	RR F4,D,E,A,B,C,62
174	RR F4,B,C,D,E,A,64
175	RR F4,E,A,B,C,D,66
176	RR F4,C,D,E,A,B,68
177
178	RR F4,A,B,C,D,E,70
179	RR F4,D,E,A,B,C,72
180	RR F4,B,C,D,E,A,74
181	RR F4,E,A,B,C,D,76
182	RR F4,C,D,E,A,B,78
183
184	UPDATE_HASH   (HASH_PTR), A
185	UPDATE_HASH  4(HASH_PTR), B
186	UPDATE_HASH  8(HASH_PTR), C
187	UPDATE_HASH 12(HASH_PTR), D
188	UPDATE_HASH 16(HASH_PTR), E
189
190	RESTORE_RENAMED_REGS
191	cmp	K_BASE, BUFFER_PTR	# K_BASE means, we reached the end
192	jne	1b
193.endm
194
195.macro INIT_REGALLOC
196  .set A, REG_A
197  .set B, REG_B
198  .set C, REG_C
199  .set D, REG_D
200  .set E, REG_E
201  .set T1, REG_T1
202  .set T2, REG_T2
203.endm
204
205.macro RESTORE_RENAMED_REGS
206	# order is important (REG_C is where it should be)
207	mov	B, REG_B
208	mov	D, REG_D
209	mov	A, REG_A
210	mov	E, REG_E
211.endm
212
213.macro SWAP_REG_NAMES  a, b
214  .set _T, \a
215  .set \a, \b
216  .set \b, _T
217.endm
218
219.macro F1  b, c, d
220	mov	\c, T1
221	SWAP_REG_NAMES \c, T1
222	xor	\d, T1
223	and	\b, T1
224	xor	\d, T1
225.endm
226
227.macro F2  b, c, d
228	mov	\d, T1
229	SWAP_REG_NAMES \d, T1
230	xor	\c, T1
231	xor	\b, T1
232.endm
233
234.macro F3  b, c ,d
235	mov	\c, T1
236	SWAP_REG_NAMES \c, T1
237	mov	\b, T2
238	or	\b, T1
239	and	\c, T2
240	and	\d, T1
241	or	T2, T1
242.endm
243
244.macro F4  b, c, d
245	F2 \b, \c, \d
246.endm
247
248.macro UPDATE_HASH  hash, val
249	add	\hash, \val
250	mov	\val, \hash
251.endm
252
253/*
254 * RR does two rounds of SHA-1 back to back with W[] pre-calc
255 *   t1 = F(b, c, d);   e += w(i)
256 *   e += t1;           b <<= 30;   d  += w(i+1);
257 *   t1 = F(a, b, c);
258 *   d += t1;           a <<= 5;
259 *   e += a;
260 *   t1 = e;            a >>= 7;
261 *   t1 <<= 5;
262 *   d += t1;
263 */
264.macro RR  F, a, b, c, d, e, round
265	add	WK(\round), \e
266	\F   \b, \c, \d		# t1 = F(b, c, d);
267	W_PRECALC (\round + W_PRECALC_AHEAD)
268	rol	$30, \b
269	add	T1, \e
270	add	WK(\round + 1), \d
271
272	\F   \a, \b, \c
273	W_PRECALC (\round + W_PRECALC_AHEAD + 1)
274	rol	$5, \a
275	add	\a, \e
276	add	T1, \d
277	ror	$7, \a		# (a <<r 5) >>r 7) => a <<r 30)
278
279	mov	\e, T1
280	SWAP_REG_NAMES \e, T1
281
282	rol	$5, T1
283	add	T1, \d
284
285	# write:  \a, \b
286	# rotate: \a<=\d, \b<=\e, \c<=\a, \d<=\b, \e<=\c
287.endm
288
289.macro W_PRECALC  r
290  .set i, \r
291
292  .if (i < 20)
293    .set K_XMM, 0
294  .elseif (i < 40)
295    .set K_XMM, 16
296  .elseif (i < 60)
297    .set K_XMM, 32
298  .elseif (i < 80)
299    .set K_XMM, 48
300  .endif
301
302  .if ((i < 16) || ((i >= 80) && (i < (80 + W_PRECALC_AHEAD))))
303    .set i, ((\r) % 80)	    # pre-compute for the next iteration
304    .if (i == 0)
305	W_PRECALC_RESET
306    .endif
307	W_PRECALC_00_15
308  .elseif (i<32)
309	W_PRECALC_16_31
310  .elseif (i < 80)   // rounds 32-79
311	W_PRECALC_32_79
312  .endif
313.endm
314
315.macro W_PRECALC_RESET
316  .set W,          W0
317  .set W_minus_04, W4
318  .set W_minus_08, W8
319  .set W_minus_12, W12
320  .set W_minus_16, W16
321  .set W_minus_20, W20
322  .set W_minus_24, W24
323  .set W_minus_28, W28
324  .set W_minus_32, W
325.endm
326
327.macro W_PRECALC_ROTATE
328  .set W_minus_32, W_minus_28
329  .set W_minus_28, W_minus_24
330  .set W_minus_24, W_minus_20
331  .set W_minus_20, W_minus_16
332  .set W_minus_16, W_minus_12
333  .set W_minus_12, W_minus_08
334  .set W_minus_08, W_minus_04
335  .set W_minus_04, W
336  .set W,          W_minus_32
337.endm
338
339.macro W_PRECALC_SSSE3
340
341.macro W_PRECALC_00_15
342	W_PRECALC_00_15_SSSE3
343.endm
344.macro W_PRECALC_16_31
345	W_PRECALC_16_31_SSSE3
346.endm
347.macro W_PRECALC_32_79
348	W_PRECALC_32_79_SSSE3
349.endm
350
351/* message scheduling pre-compute for rounds 0-15 */
352.macro W_PRECALC_00_15_SSSE3
353  .if ((i & 3) == 0)
354	movdqu	(i*4)(BUFFER_PTR), W_TMP1
355  .elseif ((i & 3) == 1)
356	pshufb	XMM_SHUFB_BSWAP, W_TMP1
357	movdqa	W_TMP1, W
358  .elseif ((i & 3) == 2)
359	paddd	(K_BASE), W_TMP1
360  .elseif ((i & 3) == 3)
361	movdqa  W_TMP1, WK(i&~3)
362	W_PRECALC_ROTATE
363  .endif
364.endm
365
366/* message scheduling pre-compute for rounds 16-31
367 *
368 * - calculating last 32 w[i] values in 8 XMM registers
369 * - pre-calculate K+w[i] values and store to mem, for later load by ALU add
370 *   instruction
371 *
372 * some "heavy-lifting" vectorization for rounds 16-31 due to w[i]->w[i-3]
373 * dependency, but improves for 32-79
374 */
375.macro W_PRECALC_16_31_SSSE3
376  # blended scheduling of vector and scalar instruction streams, one 4-wide
377  # vector iteration / 4 scalar rounds
378  .if ((i & 3) == 0)
379	movdqa	W_minus_12, W
380	palignr	$8, W_minus_16, W	# w[i-14]
381	movdqa	W_minus_04, W_TMP1
382	psrldq	$4, W_TMP1		# w[i-3]
383	pxor	W_minus_08, W
384  .elseif ((i & 3) == 1)
385	pxor	W_minus_16, W_TMP1
386	pxor	W_TMP1, W
387	movdqa	W, W_TMP2
388	movdqa	W, W_TMP1
389	pslldq	$12, W_TMP2
390  .elseif ((i & 3) == 2)
391	psrld	$31, W
392	pslld	$1, W_TMP1
393	por	W, W_TMP1
394	movdqa	W_TMP2, W
395	psrld	$30, W_TMP2
396	pslld	$2, W
397  .elseif ((i & 3) == 3)
398	pxor	W, W_TMP1
399	pxor	W_TMP2, W_TMP1
400	movdqa	W_TMP1, W
401	paddd	K_XMM(K_BASE), W_TMP1
402	movdqa	W_TMP1, WK(i&~3)
403	W_PRECALC_ROTATE
404  .endif
405.endm
406
407/* message scheduling pre-compute for rounds 32-79
408 *
409 * in SHA-1 specification: w[i] = (w[i-3] ^ w[i-8]  ^ w[i-14] ^ w[i-16]) rol 1
410 * instead we do equal:    w[i] = (w[i-6] ^ w[i-16] ^ w[i-28] ^ w[i-32]) rol 2
411 * allows more efficient vectorization since w[i]=>w[i-3] dependency is broken
412 */
413.macro W_PRECALC_32_79_SSSE3
414  .if ((i & 3) == 0)
415	movdqa	W_minus_04, W_TMP1
416	pxor	W_minus_28, W		# W is W_minus_32 before xor
417	palignr	$8, W_minus_08, W_TMP1
418  .elseif ((i & 3) == 1)
419	pxor	W_minus_16, W
420	pxor	W_TMP1, W
421	movdqa	W, W_TMP1
422  .elseif ((i & 3) == 2)
423	psrld	$30, W
424	pslld	$2, W_TMP1
425	por	W, W_TMP1
426  .elseif ((i & 3) == 3)
427	movdqa	W_TMP1, W
428	paddd	K_XMM(K_BASE), W_TMP1
429	movdqa	W_TMP1, WK(i&~3)
430	W_PRECALC_ROTATE
431  .endif
432.endm
433
434.endm		// W_PRECALC_SSSE3
435
436
437#define K1	0x5a827999
438#define K2	0x6ed9eba1
439#define K3	0x8f1bbcdc
440#define K4	0xca62c1d6
441
442.section .rodata
443.align 16
444
445K_XMM_AR:
446	.long K1, K1, K1, K1
447	.long K2, K2, K2, K2
448	.long K3, K3, K3, K3
449	.long K4, K4, K4, K4
450
451BSWAP_SHUFB_CTL:
452	.long 0x00010203
453	.long 0x04050607
454	.long 0x08090a0b
455	.long 0x0c0d0e0f
456
457
458.section .text
459
460W_PRECALC_SSSE3
461.macro xmm_mov a, b
462	movdqu	\a,\b
463.endm
464
465/* SSSE3 optimized implementation:
466 *  extern "C" void sha1_transform_ssse3(u32 *digest, const char *data, u32 *ws,
467 *                                       unsigned int rounds);
 
 
 
 
468 */
469SHA1_VECTOR_ASM     sha1_transform_ssse3
470
471#ifdef SHA1_ENABLE_AVX_SUPPORT
472
473.macro W_PRECALC_AVX
474
475.purgem W_PRECALC_00_15
476.macro  W_PRECALC_00_15
477    W_PRECALC_00_15_AVX
478.endm
479.purgem W_PRECALC_16_31
480.macro  W_PRECALC_16_31
481    W_PRECALC_16_31_AVX
482.endm
483.purgem W_PRECALC_32_79
484.macro  W_PRECALC_32_79
485    W_PRECALC_32_79_AVX
486.endm
487
488.macro W_PRECALC_00_15_AVX
489  .if ((i & 3) == 0)
490	vmovdqu	(i*4)(BUFFER_PTR), W_TMP1
491  .elseif ((i & 3) == 1)
492	vpshufb	XMM_SHUFB_BSWAP, W_TMP1, W
493  .elseif ((i & 3) == 2)
494	vpaddd	(K_BASE), W, W_TMP1
495  .elseif ((i & 3) == 3)
496	vmovdqa	W_TMP1, WK(i&~3)
497	W_PRECALC_ROTATE
498  .endif
499.endm
500
501.macro W_PRECALC_16_31_AVX
502  .if ((i & 3) == 0)
503	vpalignr $8, W_minus_16, W_minus_12, W	# w[i-14]
504	vpsrldq	$4, W_minus_04, W_TMP1		# w[i-3]
505	vpxor	W_minus_08, W, W
506	vpxor	W_minus_16, W_TMP1, W_TMP1
507  .elseif ((i & 3) == 1)
508	vpxor	W_TMP1, W, W
509	vpslldq	$12, W, W_TMP2
510	vpslld	$1, W, W_TMP1
511  .elseif ((i & 3) == 2)
512	vpsrld	$31, W, W
513	vpor	W, W_TMP1, W_TMP1
514	vpslld	$2, W_TMP2, W
515	vpsrld	$30, W_TMP2, W_TMP2
516  .elseif ((i & 3) == 3)
517	vpxor	W, W_TMP1, W_TMP1
518	vpxor	W_TMP2, W_TMP1, W
519	vpaddd	K_XMM(K_BASE), W, W_TMP1
520	vmovdqu	W_TMP1, WK(i&~3)
521	W_PRECALC_ROTATE
522  .endif
523.endm
524
525.macro W_PRECALC_32_79_AVX
526  .if ((i & 3) == 0)
527	vpalignr $8, W_minus_08, W_minus_04, W_TMP1
528	vpxor	W_minus_28, W, W		# W is W_minus_32 before xor
529  .elseif ((i & 3) == 1)
530	vpxor	W_minus_16, W_TMP1, W_TMP1
531	vpxor	W_TMP1, W, W
532  .elseif ((i & 3) == 2)
533	vpslld	$2, W, W_TMP1
534	vpsrld	$30, W, W
535	vpor	W, W_TMP1, W
536  .elseif ((i & 3) == 3)
537	vpaddd	K_XMM(K_BASE), W, W_TMP1
538	vmovdqu	W_TMP1, WK(i&~3)
539	W_PRECALC_ROTATE
540  .endif
541.endm
542
543.endm    // W_PRECALC_AVX
544
545W_PRECALC_AVX
546.purgem xmm_mov
547.macro xmm_mov a, b
548	vmovdqu	\a,\b
549.endm
550
551
552/* AVX optimized implementation:
553 *  extern "C" void sha1_transform_avx(u32 *digest, const char *data, u32 *ws,
554 *                                     unsigned int rounds);
555 */
556SHA1_VECTOR_ASM     sha1_transform_avx
557
558#endif
v5.14.15
  1/* SPDX-License-Identifier: GPL-2.0-or-later */
  2/*
  3 * This is a SIMD SHA-1 implementation. It requires the Intel(R) Supplemental
  4 * SSE3 instruction set extensions introduced in Intel Core Microarchitecture
  5 * processors. CPUs supporting Intel(R) AVX extensions will get an additional
  6 * boost.
  7 *
  8 * This work was inspired by the vectorized implementation of Dean Gaudet.
  9 * Additional information on it can be found at:
 10 *    http://www.arctic.org/~dean/crypto/sha1.html
 11 *
 12 * It was improved upon with more efficient vectorization of the message
 13 * scheduling. This implementation has also been optimized for all current and
 14 * several future generations of Intel CPUs.
 15 *
 16 * See this article for more information about the implementation details:
 17 *   http://software.intel.com/en-us/articles/improving-the-performance-of-the-secure-hash-algorithm-1/
 18 *
 19 * Copyright (C) 2010, Intel Corp.
 20 *   Authors: Maxim Locktyukhin <maxim.locktyukhin@intel.com>
 21 *            Ronen Zohar <ronen.zohar@intel.com>
 22 *
 23 * Converted to AT&T syntax and adapted for inclusion in the Linux kernel:
 24 *   Author: Mathias Krause <minipli@googlemail.com>
 
 
 
 
 
 25 */
 26
 27#include <linux/linkage.h>
 28
 29#define CTX	%rdi	// arg1
 30#define BUF	%rsi	// arg2
 31#define CNT	%rdx	// arg3
 32
 33#define REG_A	%ecx
 34#define REG_B	%esi
 35#define REG_C	%edi
 36#define REG_D	%r12d
 37#define REG_E	%edx
 38
 39#define REG_T1	%eax
 40#define REG_T2	%ebx
 41
 42#define K_BASE		%r8
 43#define HASH_PTR	%r9
 44#define BUFFER_PTR	%r10
 45#define BUFFER_END	%r11
 46
 47#define W_TMP1	%xmm0
 48#define W_TMP2	%xmm9
 49
 50#define W0	%xmm1
 51#define W4	%xmm2
 52#define W8	%xmm3
 53#define W12	%xmm4
 54#define W16	%xmm5
 55#define W20	%xmm6
 56#define W24	%xmm7
 57#define W28	%xmm8
 58
 59#define XMM_SHUFB_BSWAP	%xmm10
 60
 61/* we keep window of 64 w[i]+K pre-calculated values in a circular buffer */
 62#define WK(t)	(((t) & 15) * 4)(%rsp)
 63#define W_PRECALC_AHEAD	16
 64
 65/*
 66 * This macro implements the SHA-1 function's body for single 64-byte block
 67 * param: function's name
 68 */
 69.macro SHA1_VECTOR_ASM  name
 70	SYM_FUNC_START(\name)
 71
 
 
 72	push	%rbx
 
 73	push	%r12
 74	push	%rbp
 75	mov	%rsp, %rbp
 76
 
 77	sub	$64, %rsp		# allocate workspace
 78	and	$~15, %rsp		# align stack
 79
 80	mov	CTX, HASH_PTR
 81	mov	BUF, BUFFER_PTR
 82
 83	shl	$6, CNT			# multiply by 64
 84	add	BUF, CNT
 85	mov	CNT, BUFFER_END
 86
 87	lea	K_XMM_AR(%rip), K_BASE
 88	xmm_mov	BSWAP_SHUFB_CTL(%rip), XMM_SHUFB_BSWAP
 89
 90	SHA1_PIPELINED_MAIN_BODY
 91
 92	# cleanup workspace
 93	mov	$8, %ecx
 94	mov	%rsp, %rdi
 95	xor	%eax, %eax
 96	rep stosq
 97
 98	mov	%rbp, %rsp		# deallocate workspace
 
 
 99	pop	%rbp
100	pop	%r12
101	pop	%rbx
102	ret
103
104	SYM_FUNC_END(\name)
105.endm
106
107/*
108 * This macro implements 80 rounds of SHA-1 for one 64-byte block
109 */
110.macro SHA1_PIPELINED_MAIN_BODY
111	INIT_REGALLOC
112
113	mov	  (HASH_PTR), A
114	mov	 4(HASH_PTR), B
115	mov	 8(HASH_PTR), C
116	mov	12(HASH_PTR), D
117	mov	16(HASH_PTR), E
118
119  .set i, 0
120  .rept W_PRECALC_AHEAD
121	W_PRECALC i
122    .set i, (i+1)
123  .endr
124
125.align 4
1261:
127	RR F1,A,B,C,D,E,0
128	RR F1,D,E,A,B,C,2
129	RR F1,B,C,D,E,A,4
130	RR F1,E,A,B,C,D,6
131	RR F1,C,D,E,A,B,8
132
133	RR F1,A,B,C,D,E,10
134	RR F1,D,E,A,B,C,12
135	RR F1,B,C,D,E,A,14
136	RR F1,E,A,B,C,D,16
137	RR F1,C,D,E,A,B,18
138
139	RR F2,A,B,C,D,E,20
140	RR F2,D,E,A,B,C,22
141	RR F2,B,C,D,E,A,24
142	RR F2,E,A,B,C,D,26
143	RR F2,C,D,E,A,B,28
144
145	RR F2,A,B,C,D,E,30
146	RR F2,D,E,A,B,C,32
147	RR F2,B,C,D,E,A,34
148	RR F2,E,A,B,C,D,36
149	RR F2,C,D,E,A,B,38
150
151	RR F3,A,B,C,D,E,40
152	RR F3,D,E,A,B,C,42
153	RR F3,B,C,D,E,A,44
154	RR F3,E,A,B,C,D,46
155	RR F3,C,D,E,A,B,48
156
157	RR F3,A,B,C,D,E,50
158	RR F3,D,E,A,B,C,52
159	RR F3,B,C,D,E,A,54
160	RR F3,E,A,B,C,D,56
161	RR F3,C,D,E,A,B,58
162
163	add	$64, BUFFER_PTR		# move to the next 64-byte block
164	cmp	BUFFER_END, BUFFER_PTR	# if the current is the last one use
165	cmovae	K_BASE, BUFFER_PTR	# dummy source to avoid buffer overrun
166
167	RR F4,A,B,C,D,E,60
168	RR F4,D,E,A,B,C,62
169	RR F4,B,C,D,E,A,64
170	RR F4,E,A,B,C,D,66
171	RR F4,C,D,E,A,B,68
172
173	RR F4,A,B,C,D,E,70
174	RR F4,D,E,A,B,C,72
175	RR F4,B,C,D,E,A,74
176	RR F4,E,A,B,C,D,76
177	RR F4,C,D,E,A,B,78
178
179	UPDATE_HASH   (HASH_PTR), A
180	UPDATE_HASH  4(HASH_PTR), B
181	UPDATE_HASH  8(HASH_PTR), C
182	UPDATE_HASH 12(HASH_PTR), D
183	UPDATE_HASH 16(HASH_PTR), E
184
185	RESTORE_RENAMED_REGS
186	cmp	K_BASE, BUFFER_PTR	# K_BASE means, we reached the end
187	jne	1b
188.endm
189
190.macro INIT_REGALLOC
191  .set A, REG_A
192  .set B, REG_B
193  .set C, REG_C
194  .set D, REG_D
195  .set E, REG_E
196  .set T1, REG_T1
197  .set T2, REG_T2
198.endm
199
200.macro RESTORE_RENAMED_REGS
201	# order is important (REG_C is where it should be)
202	mov	B, REG_B
203	mov	D, REG_D
204	mov	A, REG_A
205	mov	E, REG_E
206.endm
207
208.macro SWAP_REG_NAMES  a, b
209  .set _T, \a
210  .set \a, \b
211  .set \b, _T
212.endm
213
214.macro F1  b, c, d
215	mov	\c, T1
216	SWAP_REG_NAMES \c, T1
217	xor	\d, T1
218	and	\b, T1
219	xor	\d, T1
220.endm
221
222.macro F2  b, c, d
223	mov	\d, T1
224	SWAP_REG_NAMES \d, T1
225	xor	\c, T1
226	xor	\b, T1
227.endm
228
229.macro F3  b, c ,d
230	mov	\c, T1
231	SWAP_REG_NAMES \c, T1
232	mov	\b, T2
233	or	\b, T1
234	and	\c, T2
235	and	\d, T1
236	or	T2, T1
237.endm
238
239.macro F4  b, c, d
240	F2 \b, \c, \d
241.endm
242
243.macro UPDATE_HASH  hash, val
244	add	\hash, \val
245	mov	\val, \hash
246.endm
247
248/*
249 * RR does two rounds of SHA-1 back to back with W[] pre-calc
250 *   t1 = F(b, c, d);   e += w(i)
251 *   e += t1;           b <<= 30;   d  += w(i+1);
252 *   t1 = F(a, b, c);
253 *   d += t1;           a <<= 5;
254 *   e += a;
255 *   t1 = e;            a >>= 7;
256 *   t1 <<= 5;
257 *   d += t1;
258 */
259.macro RR  F, a, b, c, d, e, round
260	add	WK(\round), \e
261	\F   \b, \c, \d		# t1 = F(b, c, d);
262	W_PRECALC (\round + W_PRECALC_AHEAD)
263	rol	$30, \b
264	add	T1, \e
265	add	WK(\round + 1), \d
266
267	\F   \a, \b, \c
268	W_PRECALC (\round + W_PRECALC_AHEAD + 1)
269	rol	$5, \a
270	add	\a, \e
271	add	T1, \d
272	ror	$7, \a		# (a <<r 5) >>r 7) => a <<r 30)
273
274	mov	\e, T1
275	SWAP_REG_NAMES \e, T1
276
277	rol	$5, T1
278	add	T1, \d
279
280	# write:  \a, \b
281	# rotate: \a<=\d, \b<=\e, \c<=\a, \d<=\b, \e<=\c
282.endm
283
284.macro W_PRECALC  r
285  .set i, \r
286
287  .if (i < 20)
288    .set K_XMM, 0
289  .elseif (i < 40)
290    .set K_XMM, 16
291  .elseif (i < 60)
292    .set K_XMM, 32
293  .elseif (i < 80)
294    .set K_XMM, 48
295  .endif
296
297  .if ((i < 16) || ((i >= 80) && (i < (80 + W_PRECALC_AHEAD))))
298    .set i, ((\r) % 80)	    # pre-compute for the next iteration
299    .if (i == 0)
300	W_PRECALC_RESET
301    .endif
302	W_PRECALC_00_15
303  .elseif (i<32)
304	W_PRECALC_16_31
305  .elseif (i < 80)   // rounds 32-79
306	W_PRECALC_32_79
307  .endif
308.endm
309
310.macro W_PRECALC_RESET
311  .set W,          W0
312  .set W_minus_04, W4
313  .set W_minus_08, W8
314  .set W_minus_12, W12
315  .set W_minus_16, W16
316  .set W_minus_20, W20
317  .set W_minus_24, W24
318  .set W_minus_28, W28
319  .set W_minus_32, W
320.endm
321
322.macro W_PRECALC_ROTATE
323  .set W_minus_32, W_minus_28
324  .set W_minus_28, W_minus_24
325  .set W_minus_24, W_minus_20
326  .set W_minus_20, W_minus_16
327  .set W_minus_16, W_minus_12
328  .set W_minus_12, W_minus_08
329  .set W_minus_08, W_minus_04
330  .set W_minus_04, W
331  .set W,          W_minus_32
332.endm
333
334.macro W_PRECALC_SSSE3
335
336.macro W_PRECALC_00_15
337	W_PRECALC_00_15_SSSE3
338.endm
339.macro W_PRECALC_16_31
340	W_PRECALC_16_31_SSSE3
341.endm
342.macro W_PRECALC_32_79
343	W_PRECALC_32_79_SSSE3
344.endm
345
346/* message scheduling pre-compute for rounds 0-15 */
347.macro W_PRECALC_00_15_SSSE3
348  .if ((i & 3) == 0)
349	movdqu	(i*4)(BUFFER_PTR), W_TMP1
350  .elseif ((i & 3) == 1)
351	pshufb	XMM_SHUFB_BSWAP, W_TMP1
352	movdqa	W_TMP1, W
353  .elseif ((i & 3) == 2)
354	paddd	(K_BASE), W_TMP1
355  .elseif ((i & 3) == 3)
356	movdqa  W_TMP1, WK(i&~3)
357	W_PRECALC_ROTATE
358  .endif
359.endm
360
361/* message scheduling pre-compute for rounds 16-31
362 *
363 * - calculating last 32 w[i] values in 8 XMM registers
364 * - pre-calculate K+w[i] values and store to mem, for later load by ALU add
365 *   instruction
366 *
367 * some "heavy-lifting" vectorization for rounds 16-31 due to w[i]->w[i-3]
368 * dependency, but improves for 32-79
369 */
370.macro W_PRECALC_16_31_SSSE3
371  # blended scheduling of vector and scalar instruction streams, one 4-wide
372  # vector iteration / 4 scalar rounds
373  .if ((i & 3) == 0)
374	movdqa	W_minus_12, W
375	palignr	$8, W_minus_16, W	# w[i-14]
376	movdqa	W_minus_04, W_TMP1
377	psrldq	$4, W_TMP1		# w[i-3]
378	pxor	W_minus_08, W
379  .elseif ((i & 3) == 1)
380	pxor	W_minus_16, W_TMP1
381	pxor	W_TMP1, W
382	movdqa	W, W_TMP2
383	movdqa	W, W_TMP1
384	pslldq	$12, W_TMP2
385  .elseif ((i & 3) == 2)
386	psrld	$31, W
387	pslld	$1, W_TMP1
388	por	W, W_TMP1
389	movdqa	W_TMP2, W
390	psrld	$30, W_TMP2
391	pslld	$2, W
392  .elseif ((i & 3) == 3)
393	pxor	W, W_TMP1
394	pxor	W_TMP2, W_TMP1
395	movdqa	W_TMP1, W
396	paddd	K_XMM(K_BASE), W_TMP1
397	movdqa	W_TMP1, WK(i&~3)
398	W_PRECALC_ROTATE
399  .endif
400.endm
401
402/* message scheduling pre-compute for rounds 32-79
403 *
404 * in SHA-1 specification: w[i] = (w[i-3] ^ w[i-8]  ^ w[i-14] ^ w[i-16]) rol 1
405 * instead we do equal:    w[i] = (w[i-6] ^ w[i-16] ^ w[i-28] ^ w[i-32]) rol 2
406 * allows more efficient vectorization since w[i]=>w[i-3] dependency is broken
407 */
408.macro W_PRECALC_32_79_SSSE3
409  .if ((i & 3) == 0)
410	movdqa	W_minus_04, W_TMP1
411	pxor	W_minus_28, W		# W is W_minus_32 before xor
412	palignr	$8, W_minus_08, W_TMP1
413  .elseif ((i & 3) == 1)
414	pxor	W_minus_16, W
415	pxor	W_TMP1, W
416	movdqa	W, W_TMP1
417  .elseif ((i & 3) == 2)
418	psrld	$30, W
419	pslld	$2, W_TMP1
420	por	W, W_TMP1
421  .elseif ((i & 3) == 3)
422	movdqa	W_TMP1, W
423	paddd	K_XMM(K_BASE), W_TMP1
424	movdqa	W_TMP1, WK(i&~3)
425	W_PRECALC_ROTATE
426  .endif
427.endm
428
429.endm		// W_PRECALC_SSSE3
430
431
432#define K1	0x5a827999
433#define K2	0x6ed9eba1
434#define K3	0x8f1bbcdc
435#define K4	0xca62c1d6
436
437.section .rodata
438.align 16
439
440K_XMM_AR:
441	.long K1, K1, K1, K1
442	.long K2, K2, K2, K2
443	.long K3, K3, K3, K3
444	.long K4, K4, K4, K4
445
446BSWAP_SHUFB_CTL:
447	.long 0x00010203
448	.long 0x04050607
449	.long 0x08090a0b
450	.long 0x0c0d0e0f
451
452
453.section .text
454
455W_PRECALC_SSSE3
456.macro xmm_mov a, b
457	movdqu	\a,\b
458.endm
459
460/*
461 * SSSE3 optimized implementation:
462 *
463 * extern "C" void sha1_transform_ssse3(struct sha1_state *state,
464 *					const u8 *data, int blocks);
465 *
466 * Note that struct sha1_state is assumed to begin with u32 state[5].
467 */
468SHA1_VECTOR_ASM     sha1_transform_ssse3
469
 
 
470.macro W_PRECALC_AVX
471
472.purgem W_PRECALC_00_15
473.macro  W_PRECALC_00_15
474    W_PRECALC_00_15_AVX
475.endm
476.purgem W_PRECALC_16_31
477.macro  W_PRECALC_16_31
478    W_PRECALC_16_31_AVX
479.endm
480.purgem W_PRECALC_32_79
481.macro  W_PRECALC_32_79
482    W_PRECALC_32_79_AVX
483.endm
484
485.macro W_PRECALC_00_15_AVX
486  .if ((i & 3) == 0)
487	vmovdqu	(i*4)(BUFFER_PTR), W_TMP1
488  .elseif ((i & 3) == 1)
489	vpshufb	XMM_SHUFB_BSWAP, W_TMP1, W
490  .elseif ((i & 3) == 2)
491	vpaddd	(K_BASE), W, W_TMP1
492  .elseif ((i & 3) == 3)
493	vmovdqa	W_TMP1, WK(i&~3)
494	W_PRECALC_ROTATE
495  .endif
496.endm
497
498.macro W_PRECALC_16_31_AVX
499  .if ((i & 3) == 0)
500	vpalignr $8, W_minus_16, W_minus_12, W	# w[i-14]
501	vpsrldq	$4, W_minus_04, W_TMP1		# w[i-3]
502	vpxor	W_minus_08, W, W
503	vpxor	W_minus_16, W_TMP1, W_TMP1
504  .elseif ((i & 3) == 1)
505	vpxor	W_TMP1, W, W
506	vpslldq	$12, W, W_TMP2
507	vpslld	$1, W, W_TMP1
508  .elseif ((i & 3) == 2)
509	vpsrld	$31, W, W
510	vpor	W, W_TMP1, W_TMP1
511	vpslld	$2, W_TMP2, W
512	vpsrld	$30, W_TMP2, W_TMP2
513  .elseif ((i & 3) == 3)
514	vpxor	W, W_TMP1, W_TMP1
515	vpxor	W_TMP2, W_TMP1, W
516	vpaddd	K_XMM(K_BASE), W, W_TMP1
517	vmovdqu	W_TMP1, WK(i&~3)
518	W_PRECALC_ROTATE
519  .endif
520.endm
521
522.macro W_PRECALC_32_79_AVX
523  .if ((i & 3) == 0)
524	vpalignr $8, W_minus_08, W_minus_04, W_TMP1
525	vpxor	W_minus_28, W, W		# W is W_minus_32 before xor
526  .elseif ((i & 3) == 1)
527	vpxor	W_minus_16, W_TMP1, W_TMP1
528	vpxor	W_TMP1, W, W
529  .elseif ((i & 3) == 2)
530	vpslld	$2, W, W_TMP1
531	vpsrld	$30, W, W
532	vpor	W, W_TMP1, W
533  .elseif ((i & 3) == 3)
534	vpaddd	K_XMM(K_BASE), W, W_TMP1
535	vmovdqu	W_TMP1, WK(i&~3)
536	W_PRECALC_ROTATE
537  .endif
538.endm
539
540.endm    // W_PRECALC_AVX
541
542W_PRECALC_AVX
543.purgem xmm_mov
544.macro xmm_mov a, b
545	vmovdqu	\a,\b
546.endm
547
548
549/* AVX optimized implementation:
550 *  extern "C" void sha1_transform_avx(struct sha1_state *state,
551 *				       const u8 *data, int blocks);
552 */
553SHA1_VECTOR_ASM     sha1_transform_avx