xref: /linux/arch/x86/crypto/sha512-avx-asm.S (revision 048c9acca90ca7da42b92745445fe008a48add88)
1########################################################################
2# Implement fast SHA-512 with AVX instructions. (x86_64)
3#
4# Copyright (C) 2013 Intel Corporation.
5#
6# Authors:
7#     James Guilford <james.guilford@intel.com>
8#     Kirk Yap <kirk.s.yap@intel.com>
9#     David Cote <david.m.cote@intel.com>
10#     Tim Chen <tim.c.chen@linux.intel.com>
11#
12# This software is available to you under a choice of one of two
13# licenses.  You may choose to be licensed under the terms of the GNU
14# General Public License (GPL) Version 2, available from the file
15# COPYING in the main directory of this source tree, or the
16# OpenIB.org BSD license below:
17#
18#     Redistribution and use in source and binary forms, with or
19#     without modification, are permitted provided that the following
20#     conditions are met:
21#
22#      - Redistributions of source code must retain the above
23#        copyright notice, this list of conditions and the following
24#        disclaimer.
25#
26#      - Redistributions in binary form must reproduce the above
27#        copyright notice, this list of conditions and the following
28#        disclaimer in the documentation and/or other materials
29#        provided with the distribution.
30#
31# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
32# EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
33# MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
34# NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
35# BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
36# ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
37# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
38# SOFTWARE.
39#
40########################################################################
41#
42# This code is described in an Intel White-Paper:
43# "Fast SHA-512 Implementations on Intel Architecture Processors"
44#
45# To find it, surf to http://www.intel.com/p/en_US/embedded
46# and search for that title.
47#
48########################################################################
49
50#ifdef CONFIG_AS_AVX
51#include <linux/linkage.h>
52
53.text
54
55# Virtual Registers
56# ARG1
57msg	= %rdi
58# ARG2
59digest	= %rsi
60# ARG3
61msglen	= %rdx
62T1	= %rcx
63T2	= %r8
64a_64	= %r9
65b_64	= %r10
66c_64	= %r11
67d_64	= %r12
68e_64	= %r13
69f_64	= %r14
70g_64	= %r15
71h_64	= %rbx
72tmp0	= %rax
73
74# Local variables (stack frame)
75
76# Message Schedule
77W_SIZE = 80*8
78# W[t] + K[t] | W[t+1] + K[t+1]
79WK_SIZE = 2*8
80RSPSAVE_SIZE = 1*8
81GPRSAVE_SIZE = 5*8
82
83frame_W = 0
84frame_WK = frame_W + W_SIZE
85frame_RSPSAVE = frame_WK + WK_SIZE
86frame_GPRSAVE = frame_RSPSAVE + RSPSAVE_SIZE
87frame_size = frame_GPRSAVE + GPRSAVE_SIZE
88
89# Useful QWORD "arrays" for simpler memory references
90# MSG, DIGEST, K_t, W_t are arrays
91# WK_2(t) points to 1 of 2 qwords at frame.WK depdending on t being odd/even
92
93# Input message (arg1)
94#define MSG(i)    8*i(msg)
95
96# Output Digest (arg2)
97#define DIGEST(i) 8*i(digest)
98
99# SHA Constants (static mem)
100#define K_t(i)    8*i+K512(%rip)
101
102# Message Schedule (stack frame)
103#define W_t(i)    8*i+frame_W(%rsp)
104
105# W[t]+K[t] (stack frame)
106#define WK_2(i)   8*((i%2))+frame_WK(%rsp)
107
108.macro RotateState
109	# Rotate symbols a..h right
110	TMP   = h_64
111	h_64  = g_64
112	g_64  = f_64
113	f_64  = e_64
114	e_64  = d_64
115	d_64  = c_64
116	c_64  = b_64
117	b_64  = a_64
118	a_64  = TMP
119.endm
120
121.macro RORQ p1 p2
122	# shld is faster than ror on Sandybridge
123	shld	$(64-\p2), \p1, \p1
124.endm
125
126.macro SHA512_Round rnd
127	# Compute Round %%t
128	mov     f_64, T1          # T1 = f
129	mov     e_64, tmp0        # tmp = e
130	xor     g_64, T1          # T1 = f ^ g
131	RORQ    tmp0, 23   # 41    # tmp = e ror 23
132	and     e_64, T1          # T1 = (f ^ g) & e
133	xor     e_64, tmp0        # tmp = (e ror 23) ^ e
134	xor     g_64, T1          # T1 = ((f ^ g) & e) ^ g = CH(e,f,g)
135	idx = \rnd
136	add     WK_2(idx), T1     # W[t] + K[t] from message scheduler
137	RORQ    tmp0, 4   # 18    # tmp = ((e ror 23) ^ e) ror 4
138	xor     e_64, tmp0        # tmp = (((e ror 23) ^ e) ror 4) ^ e
139	mov     a_64, T2          # T2 = a
140	add     h_64, T1          # T1 = CH(e,f,g) + W[t] + K[t] + h
141	RORQ    tmp0, 14  # 14    # tmp = ((((e ror23)^e)ror4)^e)ror14 = S1(e)
142	add     tmp0, T1          # T1 = CH(e,f,g) + W[t] + K[t] + S1(e)
143	mov     a_64, tmp0        # tmp = a
144	xor     c_64, T2          # T2 = a ^ c
145	and     c_64, tmp0        # tmp = a & c
146	and     b_64, T2          # T2 = (a ^ c) & b
147	xor     tmp0, T2          # T2 = ((a ^ c) & b) ^ (a & c) = Maj(a,b,c)
148	mov     a_64, tmp0        # tmp = a
149	RORQ    tmp0, 5  # 39     # tmp = a ror 5
150	xor     a_64, tmp0        # tmp = (a ror 5) ^ a
151	add     T1, d_64          # e(next_state) = d + T1
152	RORQ    tmp0, 6  # 34     # tmp = ((a ror 5) ^ a) ror 6
153	xor     a_64, tmp0        # tmp = (((a ror 5) ^ a) ror 6) ^ a
154	lea     (T1, T2), h_64    # a(next_state) = T1 + Maj(a,b,c)
155	RORQ    tmp0, 28  # 28    # tmp = ((((a ror5)^a)ror6)^a)ror28 = S0(a)
156	add     tmp0, h_64        # a(next_state) = T1 + Maj(a,b,c) S0(a)
157	RotateState
158.endm
159
160.macro SHA512_2Sched_2Round_avx rnd
161	# Compute rounds t-2 and t-1
162	# Compute message schedule QWORDS t and t+1
163
164	#   Two rounds are computed based on the values for K[t-2]+W[t-2] and
165	# K[t-1]+W[t-1] which were previously stored at WK_2 by the message
166	# scheduler.
167	#   The two new schedule QWORDS are stored at [W_t(t)] and [W_t(t+1)].
168	# They are then added to their respective SHA512 constants at
169	# [K_t(t)] and [K_t(t+1)] and stored at dqword [WK_2(t)]
170	#   For brievity, the comments following vectored instructions only refer to
171	# the first of a pair of QWORDS.
172	# Eg. XMM4=W[t-2] really means XMM4={W[t-2]|W[t-1]}
173	#   The computation of the message schedule and the rounds are tightly
174	# stitched to take advantage of instruction-level parallelism.
175
176	idx = \rnd - 2
177	vmovdqa	W_t(idx), %xmm4		# XMM4 = W[t-2]
178	idx = \rnd - 15
179	vmovdqu	W_t(idx), %xmm5		# XMM5 = W[t-15]
180	mov	f_64, T1
181	vpsrlq	$61, %xmm4, %xmm0	# XMM0 = W[t-2]>>61
182	mov	e_64, tmp0
183	vpsrlq	$1, %xmm5, %xmm6	# XMM6 = W[t-15]>>1
184	xor	g_64, T1
185	RORQ	tmp0, 23 # 41
186	vpsrlq	$19, %xmm4, %xmm1	# XMM1 = W[t-2]>>19
187	and	e_64, T1
188	xor	e_64, tmp0
189	vpxor	%xmm1, %xmm0, %xmm0	# XMM0 = W[t-2]>>61 ^ W[t-2]>>19
190	xor	g_64, T1
191	idx = \rnd
192	add	WK_2(idx), T1#
193	vpsrlq	$8, %xmm5, %xmm7	# XMM7 = W[t-15]>>8
194	RORQ	tmp0, 4 # 18
195	vpsrlq	$6, %xmm4, %xmm2	# XMM2 = W[t-2]>>6
196	xor	e_64, tmp0
197	mov	a_64, T2
198	add	h_64, T1
199	vpxor	%xmm7, %xmm6, %xmm6	# XMM6 = W[t-15]>>1 ^ W[t-15]>>8
200	RORQ	tmp0, 14 # 14
201	add	tmp0, T1
202	vpsrlq	$7, %xmm5, %xmm8	# XMM8 = W[t-15]>>7
203	mov	a_64, tmp0
204	xor	c_64, T2
205	vpsllq	$(64-61), %xmm4, %xmm3  # XMM3 = W[t-2]<<3
206	and	c_64, tmp0
207	and	b_64, T2
208	vpxor	%xmm3, %xmm2, %xmm2	# XMM2 = W[t-2]>>6 ^ W[t-2]<<3
209	xor	tmp0, T2
210	mov	a_64, tmp0
211	vpsllq	$(64-1), %xmm5, %xmm9	# XMM9 = W[t-15]<<63
212	RORQ	tmp0, 5 # 39
213	vpxor	%xmm9, %xmm8, %xmm8	# XMM8 = W[t-15]>>7 ^ W[t-15]<<63
214	xor	a_64, tmp0
215	add	T1, d_64
216	RORQ	tmp0, 6 # 34
217	xor	a_64, tmp0
218	vpxor	%xmm8, %xmm6, %xmm6	# XMM6 = W[t-15]>>1 ^ W[t-15]>>8 ^
219					#  W[t-15]>>7 ^ W[t-15]<<63
220	lea	(T1, T2), h_64
221	RORQ	tmp0, 28 # 28
222	vpsllq	$(64-19), %xmm4, %xmm4  # XMM4 = W[t-2]<<25
223	add	tmp0, h_64
224	RotateState
225	vpxor	%xmm4, %xmm0, %xmm0     # XMM0 = W[t-2]>>61 ^ W[t-2]>>19 ^
226					#        W[t-2]<<25
227	mov	f_64, T1
228	vpxor	%xmm2, %xmm0, %xmm0     # XMM0 = s1(W[t-2])
229	mov	e_64, tmp0
230	xor	g_64, T1
231	idx = \rnd - 16
232	vpaddq	W_t(idx), %xmm0, %xmm0  # XMM0 = s1(W[t-2]) + W[t-16]
233	idx = \rnd - 7
234	vmovdqu	W_t(idx), %xmm1		# XMM1 = W[t-7]
235	RORQ	tmp0, 23 # 41
236	and	e_64, T1
237	xor	e_64, tmp0
238	xor	g_64, T1
239	vpsllq	$(64-8), %xmm5, %xmm5   # XMM5 = W[t-15]<<56
240	idx = \rnd + 1
241	add	WK_2(idx), T1
242	vpxor	%xmm5, %xmm6, %xmm6     # XMM6 = s0(W[t-15])
243	RORQ	tmp0, 4 # 18
244	vpaddq	%xmm6, %xmm0, %xmm0     # XMM0 = s1(W[t-2]) + W[t-16] + s0(W[t-15])
245	xor	e_64, tmp0
246	vpaddq	%xmm1, %xmm0, %xmm0     # XMM0 = W[t] = s1(W[t-2]) + W[t-7] +
247					#               s0(W[t-15]) + W[t-16]
248	mov	a_64, T2
249	add	h_64, T1
250	RORQ	tmp0, 14 # 14
251	add	tmp0, T1
252	idx = \rnd
253	vmovdqa	%xmm0, W_t(idx)		# Store W[t]
254	vpaddq	K_t(idx), %xmm0, %xmm0  # Compute W[t]+K[t]
255	vmovdqa	%xmm0, WK_2(idx)	# Store W[t]+K[t] for next rounds
256	mov	a_64, tmp0
257	xor	c_64, T2
258	and	c_64, tmp0
259	and	b_64, T2
260	xor	tmp0, T2
261	mov	a_64, tmp0
262	RORQ	tmp0, 5 # 39
263	xor	a_64, tmp0
264	add	T1, d_64
265	RORQ	tmp0, 6 # 34
266	xor	a_64, tmp0
267	lea	(T1, T2), h_64
268	RORQ	tmp0, 28 # 28
269	add	tmp0, h_64
270	RotateState
271.endm
272
273########################################################################
274# void sha512_transform_avx(const void* M, void* D, u64 L)
275# Purpose: Updates the SHA512 digest stored at D with the message stored in M.
276# The size of the message pointed to by M must be an integer multiple of SHA512
277# message blocks.
278# L is the message length in SHA512 blocks
279########################################################################
280ENTRY(sha512_transform_avx)
281	cmp $0, msglen
282	je nowork
283
284	# Allocate Stack Space
285	mov	%rsp, %rax
286	sub     $frame_size, %rsp
287	and	$~(0x20 - 1), %rsp
288	mov	%rax, frame_RSPSAVE(%rsp)
289
290	# Save GPRs
291	mov     %rbx, frame_GPRSAVE(%rsp)
292	mov     %r12, frame_GPRSAVE +8*1(%rsp)
293	mov     %r13, frame_GPRSAVE +8*2(%rsp)
294	mov     %r14, frame_GPRSAVE +8*3(%rsp)
295	mov     %r15, frame_GPRSAVE +8*4(%rsp)
296
297updateblock:
298
299	# Load state variables
300	mov     DIGEST(0), a_64
301	mov     DIGEST(1), b_64
302	mov     DIGEST(2), c_64
303	mov     DIGEST(3), d_64
304	mov     DIGEST(4), e_64
305	mov     DIGEST(5), f_64
306	mov     DIGEST(6), g_64
307	mov     DIGEST(7), h_64
308
309	t = 0
310	.rept 80/2 + 1
311	# (80 rounds) / (2 rounds/iteration) + (1 iteration)
312	# +1 iteration because the scheduler leads hashing by 1 iteration
313		.if t < 2
314			# BSWAP 2 QWORDS
315			vmovdqa  XMM_QWORD_BSWAP(%rip), %xmm1
316			vmovdqu  MSG(t), %xmm0
317			vpshufb  %xmm1, %xmm0, %xmm0    # BSWAP
318			vmovdqa  %xmm0, W_t(t) # Store Scheduled Pair
319			vpaddq   K_t(t), %xmm0, %xmm0 # Compute W[t]+K[t]
320			vmovdqa  %xmm0, WK_2(t) # Store into WK for rounds
321		.elseif t < 16
322			# BSWAP 2 QWORDS# Compute 2 Rounds
323			vmovdqu  MSG(t), %xmm0
324			vpshufb  %xmm1, %xmm0, %xmm0    # BSWAP
325			SHA512_Round t-2    # Round t-2
326			vmovdqa  %xmm0, W_t(t) # Store Scheduled Pair
327			vpaddq   K_t(t), %xmm0, %xmm0 # Compute W[t]+K[t]
328			SHA512_Round t-1    # Round t-1
329			vmovdqa  %xmm0, WK_2(t)# Store W[t]+K[t] into WK
330		.elseif t < 79
331			# Schedule 2 QWORDS# Compute 2 Rounds
332			SHA512_2Sched_2Round_avx t
333		.else
334			# Compute 2 Rounds
335			SHA512_Round t-2
336			SHA512_Round t-1
337		.endif
338		t = t+2
339	.endr
340
341	# Update digest
342	add     a_64, DIGEST(0)
343	add     b_64, DIGEST(1)
344	add     c_64, DIGEST(2)
345	add     d_64, DIGEST(3)
346	add     e_64, DIGEST(4)
347	add     f_64, DIGEST(5)
348	add     g_64, DIGEST(6)
349	add     h_64, DIGEST(7)
350
351	# Advance to next message block
352	add     $16*8, msg
353	dec     msglen
354	jnz     updateblock
355
356	# Restore GPRs
357	mov     frame_GPRSAVE(%rsp),      %rbx
358	mov     frame_GPRSAVE +8*1(%rsp), %r12
359	mov     frame_GPRSAVE +8*2(%rsp), %r13
360	mov     frame_GPRSAVE +8*3(%rsp), %r14
361	mov     frame_GPRSAVE +8*4(%rsp), %r15
362
363	# Restore Stack Pointer
364	mov	frame_RSPSAVE(%rsp), %rsp
365
366nowork:
367	ret
368ENDPROC(sha512_transform_avx)
369
370########################################################################
371### Binary Data
372
373.data
374
375.align 16
376
377# Mask for byte-swapping a couple of qwords in an XMM register using (v)pshufb.
378XMM_QWORD_BSWAP:
379	.octa 0x08090a0b0c0d0e0f0001020304050607
380
381# K[t] used in SHA512 hashing
382K512:
383	.quad 0x428a2f98d728ae22,0x7137449123ef65cd
384	.quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
385	.quad 0x3956c25bf348b538,0x59f111f1b605d019
386	.quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118
387	.quad 0xd807aa98a3030242,0x12835b0145706fbe
388	.quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
389	.quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1
390	.quad 0x9bdc06a725c71235,0xc19bf174cf692694
391	.quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3
392	.quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
393	.quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483
394	.quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5
395	.quad 0x983e5152ee66dfab,0xa831c66d2db43210
396	.quad 0xb00327c898fb213f,0xbf597fc7beef0ee4
397	.quad 0xc6e00bf33da88fc2,0xd5a79147930aa725
398	.quad 0x06ca6351e003826f,0x142929670a0e6e70
399	.quad 0x27b70a8546d22ffc,0x2e1b21385c26c926
400	.quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df
401	.quad 0x650a73548baf63de,0x766a0abb3c77b2a8
402	.quad 0x81c2c92e47edaee6,0x92722c851482353b
403	.quad 0xa2bfe8a14cf10364,0xa81a664bbc423001
404	.quad 0xc24b8b70d0f89791,0xc76c51a30654be30
405	.quad 0xd192e819d6ef5218,0xd69906245565a910
406	.quad 0xf40e35855771202a,0x106aa07032bbd1b8
407	.quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53
408	.quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
409	.quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
410	.quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
411	.quad 0x748f82ee5defb2fc,0x78a5636f43172f60
412	.quad 0x84c87814a1f0ab72,0x8cc702081a6439ec
413	.quad 0x90befffa23631e28,0xa4506cebde82bde9
414	.quad 0xbef9a3f7b2c67915,0xc67178f2e372532b
415	.quad 0xca273eceea26619c,0xd186b8c721c0c207
416	.quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
417	.quad 0x06f067aa72176fba,0x0a637dc5a2c898a6
418	.quad 0x113f9804bef90dae,0x1b710b35131c471b
419	.quad 0x28db77f523047d84,0x32caab7b40c72493
420	.quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
421	.quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a
422	.quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817
423#endif
424