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#include <linux/linkage.h>
51#include <linux/cfi_types.h>
52
53.text
54
55# Virtual Registers
56# ARG1
57digest = %rdi
58# ARG2
59msg = %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
80
81frame_W = 0
82frame_WK = frame_W + W_SIZE
83frame_size = frame_WK + WK_SIZE
84
85# Useful QWORD "arrays" for simpler memory references
86# MSG, DIGEST, K_t, W_t are arrays
87# WK_2(t) points to 1 of 2 qwords at frame.WK depending on t being odd/even
88
89# Input message (arg1)
90#define MSG(i) 8*i(msg)
91
92# Output Digest (arg2)
93#define DIGEST(i) 8*i(digest)
94
95# SHA Constants (static mem)
96#define K_t(i) 8*i+K512(%rip)
97
98# Message Schedule (stack frame)
99#define W_t(i) 8*i+frame_W(%rsp)
100
101# W[t]+K[t] (stack frame)
102#define WK_2(i) 8*((i%2))+frame_WK(%rsp)
103
104.macro RotateState
105 # Rotate symbols a..h right
106 TMP = h_64
107 h_64 = g_64
108 g_64 = f_64
109 f_64 = e_64
110 e_64 = d_64
111 d_64 = c_64
112 c_64 = b_64
113 b_64 = a_64
114 a_64 = TMP
115.endm
116
117.macro RORQ p1 p2
118 # shld is faster than ror on Sandybridge
119 shld $(64-\p2), \p1, \p1
120.endm
121
122.macro SHA512_Round rnd
123 # Compute Round %%t
124 mov f_64, T1 # T1 = f
125 mov e_64, tmp0 # tmp = e
126 xor g_64, T1 # T1 = f ^ g
127 RORQ tmp0, 23 # 41 # tmp = e ror 23
128 and e_64, T1 # T1 = (f ^ g) & e
129 xor e_64, tmp0 # tmp = (e ror 23) ^ e
130 xor g_64, T1 # T1 = ((f ^ g) & e) ^ g = CH(e,f,g)
131 idx = \rnd
132 add WK_2(idx), T1 # W[t] + K[t] from message scheduler
133 RORQ tmp0, 4 # 18 # tmp = ((e ror 23) ^ e) ror 4
134 xor e_64, tmp0 # tmp = (((e ror 23) ^ e) ror 4) ^ e
135 mov a_64, T2 # T2 = a
136 add h_64, T1 # T1 = CH(e,f,g) + W[t] + K[t] + h
137 RORQ tmp0, 14 # 14 # tmp = ((((e ror23)^e)ror4)^e)ror14 = S1(e)
138 add tmp0, T1 # T1 = CH(e,f,g) + W[t] + K[t] + S1(e)
139 mov a_64, tmp0 # tmp = a
140 xor c_64, T2 # T2 = a ^ c
141 and c_64, tmp0 # tmp = a & c
142 and b_64, T2 # T2 = (a ^ c) & b
143 xor tmp0, T2 # T2 = ((a ^ c) & b) ^ (a & c) = Maj(a,b,c)
144 mov a_64, tmp0 # tmp = a
145 RORQ tmp0, 5 # 39 # tmp = a ror 5
146 xor a_64, tmp0 # tmp = (a ror 5) ^ a
147 add T1, d_64 # e(next_state) = d + T1
148 RORQ tmp0, 6 # 34 # tmp = ((a ror 5) ^ a) ror 6
149 xor a_64, tmp0 # tmp = (((a ror 5) ^ a) ror 6) ^ a
150 lea (T1, T2), h_64 # a(next_state) = T1 + Maj(a,b,c)
151 RORQ tmp0, 28 # 28 # tmp = ((((a ror5)^a)ror6)^a)ror28 = S0(a)
152 add tmp0, h_64 # a(next_state) = T1 + Maj(a,b,c) S0(a)
153 RotateState
154.endm
155
156.macro SHA512_2Sched_2Round_avx rnd
157 # Compute rounds t-2 and t-1
158 # Compute message schedule QWORDS t and t+1
159
160 # Two rounds are computed based on the values for K[t-2]+W[t-2] and
161 # K[t-1]+W[t-1] which were previously stored at WK_2 by the message
162 # scheduler.
163 # The two new schedule QWORDS are stored at [W_t(t)] and [W_t(t+1)].
164 # They are then added to their respective SHA512 constants at
165 # [K_t(t)] and [K_t(t+1)] and stored at dqword [WK_2(t)]
166 # For brievity, the comments following vectored instructions only refer to
167 # the first of a pair of QWORDS.
168 # Eg. XMM4=W[t-2] really means XMM4={W[t-2]|W[t-1]}
169 # The computation of the message schedule and the rounds are tightly
170 # stitched to take advantage of instruction-level parallelism.
171
172 idx = \rnd - 2
173 vmovdqa W_t(idx), %xmm4 # XMM4 = W[t-2]
174 idx = \rnd - 15
175 vmovdqu W_t(idx), %xmm5 # XMM5 = W[t-15]
176 mov f_64, T1
177 vpsrlq $61, %xmm4, %xmm0 # XMM0 = W[t-2]>>61
178 mov e_64, tmp0
179 vpsrlq $1, %xmm5, %xmm6 # XMM6 = W[t-15]>>1
180 xor g_64, T1
181 RORQ tmp0, 23 # 41
182 vpsrlq $19, %xmm4, %xmm1 # XMM1 = W[t-2]>>19
183 and e_64, T1
184 xor e_64, tmp0
185 vpxor %xmm1, %xmm0, %xmm0 # XMM0 = W[t-2]>>61 ^ W[t-2]>>19
186 xor g_64, T1
187 idx = \rnd
188 add WK_2(idx), T1#
189 vpsrlq $8, %xmm5, %xmm7 # XMM7 = W[t-15]>>8
190 RORQ tmp0, 4 # 18
191 vpsrlq $6, %xmm4, %xmm2 # XMM2 = W[t-2]>>6
192 xor e_64, tmp0
193 mov a_64, T2
194 add h_64, T1
195 vpxor %xmm7, %xmm6, %xmm6 # XMM6 = W[t-15]>>1 ^ W[t-15]>>8
196 RORQ tmp0, 14 # 14
197 add tmp0, T1
198 vpsrlq $7, %xmm5, %xmm8 # XMM8 = W[t-15]>>7
199 mov a_64, tmp0
200 xor c_64, T2
201 vpsllq $(64-61), %xmm4, %xmm3 # XMM3 = W[t-2]<<3
202 and c_64, tmp0
203 and b_64, T2
204 vpxor %xmm3, %xmm2, %xmm2 # XMM2 = W[t-2]>>6 ^ W[t-2]<<3
205 xor tmp0, T2
206 mov a_64, tmp0
207 vpsllq $(64-1), %xmm5, %xmm9 # XMM9 = W[t-15]<<63
208 RORQ tmp0, 5 # 39
209 vpxor %xmm9, %xmm8, %xmm8 # XMM8 = W[t-15]>>7 ^ W[t-15]<<63
210 xor a_64, tmp0
211 add T1, d_64
212 RORQ tmp0, 6 # 34
213 xor a_64, tmp0
214 vpxor %xmm8, %xmm6, %xmm6 # XMM6 = W[t-15]>>1 ^ W[t-15]>>8 ^
215 # W[t-15]>>7 ^ W[t-15]<<63
216 lea (T1, T2), h_64
217 RORQ tmp0, 28 # 28
218 vpsllq $(64-19), %xmm4, %xmm4 # XMM4 = W[t-2]<<25
219 add tmp0, h_64
220 RotateState
221 vpxor %xmm4, %xmm0, %xmm0 # XMM0 = W[t-2]>>61 ^ W[t-2]>>19 ^
222 # W[t-2]<<25
223 mov f_64, T1
224 vpxor %xmm2, %xmm0, %xmm0 # XMM0 = s1(W[t-2])
225 mov e_64, tmp0
226 xor g_64, T1
227 idx = \rnd - 16
228 vpaddq W_t(idx), %xmm0, %xmm0 # XMM0 = s1(W[t-2]) + W[t-16]
229 idx = \rnd - 7
230 vmovdqu W_t(idx), %xmm1 # XMM1 = W[t-7]
231 RORQ tmp0, 23 # 41
232 and e_64, T1
233 xor e_64, tmp0
234 xor g_64, T1
235 vpsllq $(64-8), %xmm5, %xmm5 # XMM5 = W[t-15]<<56
236 idx = \rnd + 1
237 add WK_2(idx), T1
238 vpxor %xmm5, %xmm6, %xmm6 # XMM6 = s0(W[t-15])
239 RORQ tmp0, 4 # 18
240 vpaddq %xmm6, %xmm0, %xmm0 # XMM0 = s1(W[t-2]) + W[t-16] + s0(W[t-15])
241 xor e_64, tmp0
242 vpaddq %xmm1, %xmm0, %xmm0 # XMM0 = W[t] = s1(W[t-2]) + W[t-7] +
243 # s0(W[t-15]) + W[t-16]
244 mov a_64, T2
245 add h_64, T1
246 RORQ tmp0, 14 # 14
247 add tmp0, T1
248 idx = \rnd
249 vmovdqa %xmm0, W_t(idx) # Store W[t]
250 vpaddq K_t(idx), %xmm0, %xmm0 # Compute W[t]+K[t]
251 vmovdqa %xmm0, WK_2(idx) # Store W[t]+K[t] for next rounds
252 mov a_64, tmp0
253 xor c_64, T2
254 and c_64, tmp0
255 and b_64, T2
256 xor tmp0, T2
257 mov a_64, tmp0
258 RORQ tmp0, 5 # 39
259 xor a_64, tmp0
260 add T1, d_64
261 RORQ tmp0, 6 # 34
262 xor a_64, tmp0
263 lea (T1, T2), h_64
264 RORQ tmp0, 28 # 28
265 add tmp0, h_64
266 RotateState
267.endm
268
269########################################################################
270# void sha512_transform_avx(sha512_state *state, const u8 *data, int blocks)
271# Purpose: Updates the SHA512 digest stored at "state" with the message
272# stored in "data".
273# The size of the message pointed to by "data" must be an integer multiple
274# of SHA512 message blocks.
275# "blocks" is the message length in SHA512 blocks
276########################################################################
277SYM_TYPED_FUNC_START(sha512_transform_avx)
278 test msglen, msglen
279 je .Lnowork
280
281 # Save GPRs
282 push %rbx
283 push %r12
284 push %r13
285 push %r14
286 push %r15
287
288 # Allocate Stack Space
289 push %rbp
290 mov %rsp, %rbp
291 sub $frame_size, %rsp
292 and $~(0x20 - 1), %rsp
293
294.Lupdateblock:
295
296 # Load state variables
297 mov DIGEST(0), a_64
298 mov DIGEST(1), b_64
299 mov DIGEST(2), c_64
300 mov DIGEST(3), d_64
301 mov DIGEST(4), e_64
302 mov DIGEST(5), f_64
303 mov DIGEST(6), g_64
304 mov DIGEST(7), h_64
305
306 t = 0
307 .rept 80/2 + 1
308 # (80 rounds) / (2 rounds/iteration) + (1 iteration)
309 # +1 iteration because the scheduler leads hashing by 1 iteration
310 .if t < 2
311 # BSWAP 2 QWORDS
312 vmovdqa XMM_QWORD_BSWAP(%rip), %xmm1
313 vmovdqu MSG(t), %xmm0
314 vpshufb %xmm1, %xmm0, %xmm0 # BSWAP
315 vmovdqa %xmm0, W_t(t) # Store Scheduled Pair
316 vpaddq K_t(t), %xmm0, %xmm0 # Compute W[t]+K[t]
317 vmovdqa %xmm0, WK_2(t) # Store into WK for rounds
318 .elseif t < 16
319 # BSWAP 2 QWORDS# Compute 2 Rounds
320 vmovdqu MSG(t), %xmm0
321 vpshufb %xmm1, %xmm0, %xmm0 # BSWAP
322 SHA512_Round t-2 # Round t-2
323 vmovdqa %xmm0, W_t(t) # Store Scheduled Pair
324 vpaddq K_t(t), %xmm0, %xmm0 # Compute W[t]+K[t]
325 SHA512_Round t-1 # Round t-1
326 vmovdqa %xmm0, WK_2(t)# Store W[t]+K[t] into WK
327 .elseif t < 79
328 # Schedule 2 QWORDS# Compute 2 Rounds
329 SHA512_2Sched_2Round_avx t
330 .else
331 # Compute 2 Rounds
332 SHA512_Round t-2
333 SHA512_Round t-1
334 .endif
335 t = t+2
336 .endr
337
338 # Update digest
339 add a_64, DIGEST(0)
340 add b_64, DIGEST(1)
341 add c_64, DIGEST(2)
342 add d_64, DIGEST(3)
343 add e_64, DIGEST(4)
344 add f_64, DIGEST(5)
345 add g_64, DIGEST(6)
346 add h_64, DIGEST(7)
347
348 # Advance to next message block
349 add $16*8, msg
350 dec msglen
351 jnz .Lupdateblock
352
353 # Restore Stack Pointer
354 mov %rbp, %rsp
355 pop %rbp
356
357 # Restore GPRs
358 pop %r15
359 pop %r14
360 pop %r13
361 pop %r12
362 pop %rbx
363
364.Lnowork:
365 RET
366SYM_FUNC_END(sha512_transform_avx)
367
368########################################################################
369### Binary Data
370
371.section .rodata.cst16.XMM_QWORD_BSWAP, "aM", @progbits, 16
372.align 16
373# Mask for byte-swapping a couple of qwords in an XMM register using (v)pshufb.
374XMM_QWORD_BSWAP:
375 .octa 0x08090a0b0c0d0e0f0001020304050607
376
377# Mergeable 640-byte rodata section. This allows linker to merge the table
378# with other, exactly the same 640-byte fragment of another rodata section
379# (if such section exists).
380.section .rodata.cst640.K512, "aM", @progbits, 640
381.align 64
382# K[t] used in SHA512 hashing
383K512:
384 .quad 0x428a2f98d728ae22,0x7137449123ef65cd
385 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
386 .quad 0x3956c25bf348b538,0x59f111f1b605d019
387 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118
388 .quad 0xd807aa98a3030242,0x12835b0145706fbe
389 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
390 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1
391 .quad 0x9bdc06a725c71235,0xc19bf174cf692694
392 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3
393 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
394 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483
395 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5
396 .quad 0x983e5152ee66dfab,0xa831c66d2db43210
397 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4
398 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725
399 .quad 0x06ca6351e003826f,0x142929670a0e6e70
400 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926
401 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df
402 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8
403 .quad 0x81c2c92e47edaee6,0x92722c851482353b
404 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001
405 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30
406 .quad 0xd192e819d6ef5218,0xd69906245565a910
407 .quad 0xf40e35855771202a,0x106aa07032bbd1b8
408 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53
409 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
410 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
411 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
412 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60
413 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec
414 .quad 0x90befffa23631e28,0xa4506cebde82bde9
415 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b
416 .quad 0xca273eceea26619c,0xd186b8c721c0c207
417 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
418 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6
419 .quad 0x113f9804bef90dae,0x1b710b35131c471b
420 .quad 0x28db77f523047d84,0x32caab7b40c72493
421 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
422 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a
423 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817
424

source code of linux/arch/x86/crypto/sha512-avx-asm.S