uncle_blake3 0.0.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +7 -0
- data/LICENSE.md +27 -0
- data/README.md +89 -0
- data/ext/Rakefile +55 -0
- data/ext/binding/uncle_blake3.c +41 -0
- data/ext/blake3/c/Makefile.testing +82 -0
- data/ext/blake3/c/README.md +316 -0
- data/ext/blake3/c/blake3.c +616 -0
- data/ext/blake3/c/blake3.h +60 -0
- data/ext/blake3/c/blake3_avx2.c +326 -0
- data/ext/blake3/c/blake3_avx2_x86-64_unix.S +1815 -0
- data/ext/blake3/c/blake3_avx2_x86-64_windows_gnu.S +1817 -0
- data/ext/blake3/c/blake3_avx2_x86-64_windows_msvc.asm +1828 -0
- data/ext/blake3/c/blake3_avx512.c +1207 -0
- data/ext/blake3/c/blake3_avx512_x86-64_unix.S +2585 -0
- data/ext/blake3/c/blake3_avx512_x86-64_windows_gnu.S +2615 -0
- data/ext/blake3/c/blake3_avx512_x86-64_windows_msvc.asm +2634 -0
- data/ext/blake3/c/blake3_dispatch.c +276 -0
- data/ext/blake3/c/blake3_impl.h +282 -0
- data/ext/blake3/c/blake3_neon.c +351 -0
- data/ext/blake3/c/blake3_portable.c +160 -0
- data/ext/blake3/c/blake3_sse2.c +566 -0
- data/ext/blake3/c/blake3_sse2_x86-64_unix.S +2291 -0
- data/ext/blake3/c/blake3_sse2_x86-64_windows_gnu.S +2332 -0
- data/ext/blake3/c/blake3_sse2_x86-64_windows_msvc.asm +2350 -0
- data/ext/blake3/c/blake3_sse41.c +560 -0
- data/ext/blake3/c/blake3_sse41_x86-64_unix.S +2028 -0
- data/ext/blake3/c/blake3_sse41_x86-64_windows_gnu.S +2069 -0
- data/ext/blake3/c/blake3_sse41_x86-64_windows_msvc.asm +2089 -0
- data/ext/blake3/c/example.c +37 -0
- data/ext/blake3/c/main.c +166 -0
- data/ext/blake3/c/test.py +97 -0
- data/lib/uncle_blake3/binding.rb +20 -0
- data/lib/uncle_blake3/build/loader.rb +40 -0
- data/lib/uncle_blake3/build/platform.rb +37 -0
- data/lib/uncle_blake3/build.rb +4 -0
- data/lib/uncle_blake3/digest.rb +119 -0
- data/lib/uncle_blake3/version.rb +5 -0
- data/lib/uncle_blake3.rb +7 -0
- metadata +112 -0
|
@@ -0,0 +1,1828 @@
|
|
|
1
|
+
public _blake3_hash_many_avx2
|
|
2
|
+
public blake3_hash_many_avx2
|
|
3
|
+
|
|
4
|
+
_TEXT SEGMENT ALIGN(16) 'CODE'
|
|
5
|
+
|
|
6
|
+
ALIGN 16
|
|
7
|
+
blake3_hash_many_avx2 PROC
|
|
8
|
+
_blake3_hash_many_avx2 PROC
|
|
9
|
+
push r15
|
|
10
|
+
push r14
|
|
11
|
+
push r13
|
|
12
|
+
push r12
|
|
13
|
+
push rsi
|
|
14
|
+
push rdi
|
|
15
|
+
push rbx
|
|
16
|
+
push rbp
|
|
17
|
+
mov rbp, rsp
|
|
18
|
+
sub rsp, 880
|
|
19
|
+
and rsp, 0FFFFFFFFFFFFFFC0H
|
|
20
|
+
vmovdqa xmmword ptr [rsp+2D0H], xmm6
|
|
21
|
+
vmovdqa xmmword ptr [rsp+2E0H], xmm7
|
|
22
|
+
vmovdqa xmmword ptr [rsp+2F0H], xmm8
|
|
23
|
+
vmovdqa xmmword ptr [rsp+300H], xmm9
|
|
24
|
+
vmovdqa xmmword ptr [rsp+310H], xmm10
|
|
25
|
+
vmovdqa xmmword ptr [rsp+320H], xmm11
|
|
26
|
+
vmovdqa xmmword ptr [rsp+330H], xmm12
|
|
27
|
+
vmovdqa xmmword ptr [rsp+340H], xmm13
|
|
28
|
+
vmovdqa xmmword ptr [rsp+350H], xmm14
|
|
29
|
+
vmovdqa xmmword ptr [rsp+360H], xmm15
|
|
30
|
+
mov rdi, rcx
|
|
31
|
+
mov rsi, rdx
|
|
32
|
+
mov rdx, r8
|
|
33
|
+
mov rcx, r9
|
|
34
|
+
mov r8, qword ptr [rbp+68H]
|
|
35
|
+
movzx r9, byte ptr [rbp+70H]
|
|
36
|
+
neg r9d
|
|
37
|
+
vmovd xmm0, r9d
|
|
38
|
+
vpbroadcastd ymm0, xmm0
|
|
39
|
+
vmovdqa ymmword ptr [rsp+260H], ymm0
|
|
40
|
+
vpand ymm1, ymm0, ymmword ptr [ADD0]
|
|
41
|
+
vpand ymm2, ymm0, ymmword ptr [ADD1]
|
|
42
|
+
vmovdqa ymmword ptr [rsp+2A0H], ymm2
|
|
43
|
+
vmovd xmm2, r8d
|
|
44
|
+
vpbroadcastd ymm2, xmm2
|
|
45
|
+
vpaddd ymm2, ymm2, ymm1
|
|
46
|
+
vmovdqa ymmword ptr [rsp+220H], ymm2
|
|
47
|
+
vpxor ymm1, ymm1, ymmword ptr [CMP_MSB_MASK]
|
|
48
|
+
vpxor ymm2, ymm2, ymmword ptr [CMP_MSB_MASK]
|
|
49
|
+
vpcmpgtd ymm2, ymm1, ymm2
|
|
50
|
+
shr r8, 32
|
|
51
|
+
vmovd xmm3, r8d
|
|
52
|
+
vpbroadcastd ymm3, xmm3
|
|
53
|
+
vpsubd ymm3, ymm3, ymm2
|
|
54
|
+
vmovdqa ymmword ptr [rsp+240H], ymm3
|
|
55
|
+
shl rdx, 6
|
|
56
|
+
mov qword ptr [rsp+2C0H], rdx
|
|
57
|
+
cmp rsi, 8
|
|
58
|
+
jc final7blocks
|
|
59
|
+
outerloop8:
|
|
60
|
+
vpbroadcastd ymm0, dword ptr [rcx]
|
|
61
|
+
vpbroadcastd ymm1, dword ptr [rcx+4H]
|
|
62
|
+
vpbroadcastd ymm2, dword ptr [rcx+8H]
|
|
63
|
+
vpbroadcastd ymm3, dword ptr [rcx+0CH]
|
|
64
|
+
vpbroadcastd ymm4, dword ptr [rcx+10H]
|
|
65
|
+
vpbroadcastd ymm5, dword ptr [rcx+14H]
|
|
66
|
+
vpbroadcastd ymm6, dword ptr [rcx+18H]
|
|
67
|
+
vpbroadcastd ymm7, dword ptr [rcx+1CH]
|
|
68
|
+
mov r8, qword ptr [rdi]
|
|
69
|
+
mov r9, qword ptr [rdi+8H]
|
|
70
|
+
mov r10, qword ptr [rdi+10H]
|
|
71
|
+
mov r11, qword ptr [rdi+18H]
|
|
72
|
+
mov r12, qword ptr [rdi+20H]
|
|
73
|
+
mov r13, qword ptr [rdi+28H]
|
|
74
|
+
mov r14, qword ptr [rdi+30H]
|
|
75
|
+
mov r15, qword ptr [rdi+38H]
|
|
76
|
+
movzx eax, byte ptr [rbp+78H]
|
|
77
|
+
movzx ebx, byte ptr [rbp+80H]
|
|
78
|
+
or eax, ebx
|
|
79
|
+
xor edx, edx
|
|
80
|
+
ALIGN 16
|
|
81
|
+
innerloop8:
|
|
82
|
+
movzx ebx, byte ptr [rbp+88H]
|
|
83
|
+
or ebx, eax
|
|
84
|
+
add rdx, 64
|
|
85
|
+
cmp rdx, qword ptr [rsp+2C0H]
|
|
86
|
+
cmove eax, ebx
|
|
87
|
+
mov dword ptr [rsp+200H], eax
|
|
88
|
+
vmovups xmm8, xmmword ptr [r8+rdx-40H]
|
|
89
|
+
vinsertf128 ymm8, ymm8, xmmword ptr [r12+rdx-40H], 01H
|
|
90
|
+
vmovups xmm9, xmmword ptr [r9+rdx-40H]
|
|
91
|
+
vinsertf128 ymm9, ymm9, xmmword ptr [r13+rdx-40H], 01H
|
|
92
|
+
vunpcklpd ymm12, ymm8, ymm9
|
|
93
|
+
vunpckhpd ymm13, ymm8, ymm9
|
|
94
|
+
vmovups xmm10, xmmword ptr [r10+rdx-40H]
|
|
95
|
+
vinsertf128 ymm10, ymm10, xmmword ptr [r14+rdx-40H], 01H
|
|
96
|
+
vmovups xmm11, xmmword ptr [r11+rdx-40H]
|
|
97
|
+
vinsertf128 ymm11, ymm11, xmmword ptr [r15+rdx-40H], 01H
|
|
98
|
+
vunpcklpd ymm14, ymm10, ymm11
|
|
99
|
+
vunpckhpd ymm15, ymm10, ymm11
|
|
100
|
+
vshufps ymm8, ymm12, ymm14, 136
|
|
101
|
+
vmovaps ymmword ptr [rsp], ymm8
|
|
102
|
+
vshufps ymm9, ymm12, ymm14, 221
|
|
103
|
+
vmovaps ymmword ptr [rsp+20H], ymm9
|
|
104
|
+
vshufps ymm10, ymm13, ymm15, 136
|
|
105
|
+
vmovaps ymmword ptr [rsp+40H], ymm10
|
|
106
|
+
vshufps ymm11, ymm13, ymm15, 221
|
|
107
|
+
vmovaps ymmword ptr [rsp+60H], ymm11
|
|
108
|
+
vmovups xmm8, xmmword ptr [r8+rdx-30H]
|
|
109
|
+
vinsertf128 ymm8, ymm8, xmmword ptr [r12+rdx-30H], 01H
|
|
110
|
+
vmovups xmm9, xmmword ptr [r9+rdx-30H]
|
|
111
|
+
vinsertf128 ymm9, ymm9, xmmword ptr [r13+rdx-30H], 01H
|
|
112
|
+
vunpcklpd ymm12, ymm8, ymm9
|
|
113
|
+
vunpckhpd ymm13, ymm8, ymm9
|
|
114
|
+
vmovups xmm10, xmmword ptr [r10+rdx-30H]
|
|
115
|
+
vinsertf128 ymm10, ymm10, xmmword ptr [r14+rdx-30H], 01H
|
|
116
|
+
vmovups xmm11, xmmword ptr [r11+rdx-30H]
|
|
117
|
+
vinsertf128 ymm11, ymm11, xmmword ptr [r15+rdx-30H], 01H
|
|
118
|
+
vunpcklpd ymm14, ymm10, ymm11
|
|
119
|
+
vunpckhpd ymm15, ymm10, ymm11
|
|
120
|
+
vshufps ymm8, ymm12, ymm14, 136
|
|
121
|
+
vmovaps ymmword ptr [rsp+80H], ymm8
|
|
122
|
+
vshufps ymm9, ymm12, ymm14, 221
|
|
123
|
+
vmovaps ymmword ptr [rsp+0A0H], ymm9
|
|
124
|
+
vshufps ymm10, ymm13, ymm15, 136
|
|
125
|
+
vmovaps ymmword ptr [rsp+0C0H], ymm10
|
|
126
|
+
vshufps ymm11, ymm13, ymm15, 221
|
|
127
|
+
vmovaps ymmword ptr [rsp+0E0H], ymm11
|
|
128
|
+
vmovups xmm8, xmmword ptr [r8+rdx-20H]
|
|
129
|
+
vinsertf128 ymm8, ymm8, xmmword ptr [r12+rdx-20H], 01H
|
|
130
|
+
vmovups xmm9, xmmword ptr [r9+rdx-20H]
|
|
131
|
+
vinsertf128 ymm9, ymm9, xmmword ptr [r13+rdx-20H], 01H
|
|
132
|
+
vunpcklpd ymm12, ymm8, ymm9
|
|
133
|
+
vunpckhpd ymm13, ymm8, ymm9
|
|
134
|
+
vmovups xmm10, xmmword ptr [r10+rdx-20H]
|
|
135
|
+
vinsertf128 ymm10, ymm10, xmmword ptr [r14+rdx-20H], 01H
|
|
136
|
+
vmovups xmm11, xmmword ptr [r11+rdx-20H]
|
|
137
|
+
vinsertf128 ymm11, ymm11, xmmword ptr [r15+rdx-20H], 01H
|
|
138
|
+
vunpcklpd ymm14, ymm10, ymm11
|
|
139
|
+
vunpckhpd ymm15, ymm10, ymm11
|
|
140
|
+
vshufps ymm8, ymm12, ymm14, 136
|
|
141
|
+
vmovaps ymmword ptr [rsp+100H], ymm8
|
|
142
|
+
vshufps ymm9, ymm12, ymm14, 221
|
|
143
|
+
vmovaps ymmword ptr [rsp+120H], ymm9
|
|
144
|
+
vshufps ymm10, ymm13, ymm15, 136
|
|
145
|
+
vmovaps ymmword ptr [rsp+140H], ymm10
|
|
146
|
+
vshufps ymm11, ymm13, ymm15, 221
|
|
147
|
+
vmovaps ymmword ptr [rsp+160H], ymm11
|
|
148
|
+
vmovups xmm8, xmmword ptr [r8+rdx-10H]
|
|
149
|
+
vinsertf128 ymm8, ymm8, xmmword ptr [r12+rdx-10H], 01H
|
|
150
|
+
vmovups xmm9, xmmword ptr [r9+rdx-10H]
|
|
151
|
+
vinsertf128 ymm9, ymm9, xmmword ptr [r13+rdx-10H], 01H
|
|
152
|
+
vunpcklpd ymm12, ymm8, ymm9
|
|
153
|
+
vunpckhpd ymm13, ymm8, ymm9
|
|
154
|
+
vmovups xmm10, xmmword ptr [r10+rdx-10H]
|
|
155
|
+
vinsertf128 ymm10, ymm10, xmmword ptr [r14+rdx-10H], 01H
|
|
156
|
+
vmovups xmm11, xmmword ptr [r11+rdx-10H]
|
|
157
|
+
vinsertf128 ymm11, ymm11, xmmword ptr [r15+rdx-10H], 01H
|
|
158
|
+
vunpcklpd ymm14, ymm10, ymm11
|
|
159
|
+
vunpckhpd ymm15, ymm10, ymm11
|
|
160
|
+
vshufps ymm8, ymm12, ymm14, 136
|
|
161
|
+
vmovaps ymmword ptr [rsp+180H], ymm8
|
|
162
|
+
vshufps ymm9, ymm12, ymm14, 221
|
|
163
|
+
vmovaps ymmword ptr [rsp+1A0H], ymm9
|
|
164
|
+
vshufps ymm10, ymm13, ymm15, 136
|
|
165
|
+
vmovaps ymmword ptr [rsp+1C0H], ymm10
|
|
166
|
+
vshufps ymm11, ymm13, ymm15, 221
|
|
167
|
+
vmovaps ymmword ptr [rsp+1E0H], ymm11
|
|
168
|
+
vpbroadcastd ymm15, dword ptr [rsp+200H]
|
|
169
|
+
prefetcht0 byte ptr [r8+rdx+80H]
|
|
170
|
+
prefetcht0 byte ptr [r12+rdx+80H]
|
|
171
|
+
prefetcht0 byte ptr [r9+rdx+80H]
|
|
172
|
+
prefetcht0 byte ptr [r13+rdx+80H]
|
|
173
|
+
prefetcht0 byte ptr [r10+rdx+80H]
|
|
174
|
+
prefetcht0 byte ptr [r14+rdx+80H]
|
|
175
|
+
prefetcht0 byte ptr [r11+rdx+80H]
|
|
176
|
+
prefetcht0 byte ptr [r15+rdx+80H]
|
|
177
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp]
|
|
178
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+40H]
|
|
179
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+80H]
|
|
180
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+0C0H]
|
|
181
|
+
vpaddd ymm0, ymm0, ymm4
|
|
182
|
+
vpaddd ymm1, ymm1, ymm5
|
|
183
|
+
vpaddd ymm2, ymm2, ymm6
|
|
184
|
+
vpaddd ymm3, ymm3, ymm7
|
|
185
|
+
vpxor ymm12, ymm0, ymmword ptr [rsp+220H]
|
|
186
|
+
vpxor ymm13, ymm1, ymmword ptr [rsp+240H]
|
|
187
|
+
vpxor ymm14, ymm2, ymmword ptr [BLAKE3_BLOCK_LEN]
|
|
188
|
+
vpxor ymm15, ymm3, ymm15
|
|
189
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
190
|
+
vpshufb ymm12, ymm12, ymm8
|
|
191
|
+
vpshufb ymm13, ymm13, ymm8
|
|
192
|
+
vpshufb ymm14, ymm14, ymm8
|
|
193
|
+
vpshufb ymm15, ymm15, ymm8
|
|
194
|
+
vpaddd ymm8, ymm12, ymmword ptr [BLAKE3_IV_0]
|
|
195
|
+
vpaddd ymm9, ymm13, ymmword ptr [BLAKE3_IV_1]
|
|
196
|
+
vpaddd ymm10, ymm14, ymmword ptr [BLAKE3_IV_2]
|
|
197
|
+
vpaddd ymm11, ymm15, ymmword ptr [BLAKE3_IV_3]
|
|
198
|
+
vpxor ymm4, ymm4, ymm8
|
|
199
|
+
vpxor ymm5, ymm5, ymm9
|
|
200
|
+
vpxor ymm6, ymm6, ymm10
|
|
201
|
+
vpxor ymm7, ymm7, ymm11
|
|
202
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
203
|
+
vpsrld ymm8, ymm4, 12
|
|
204
|
+
vpslld ymm4, ymm4, 20
|
|
205
|
+
vpor ymm4, ymm4, ymm8
|
|
206
|
+
vpsrld ymm8, ymm5, 12
|
|
207
|
+
vpslld ymm5, ymm5, 20
|
|
208
|
+
vpor ymm5, ymm5, ymm8
|
|
209
|
+
vpsrld ymm8, ymm6, 12
|
|
210
|
+
vpslld ymm6, ymm6, 20
|
|
211
|
+
vpor ymm6, ymm6, ymm8
|
|
212
|
+
vpsrld ymm8, ymm7, 12
|
|
213
|
+
vpslld ymm7, ymm7, 20
|
|
214
|
+
vpor ymm7, ymm7, ymm8
|
|
215
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+20H]
|
|
216
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+60H]
|
|
217
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+0A0H]
|
|
218
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+0E0H]
|
|
219
|
+
vpaddd ymm0, ymm0, ymm4
|
|
220
|
+
vpaddd ymm1, ymm1, ymm5
|
|
221
|
+
vpaddd ymm2, ymm2, ymm6
|
|
222
|
+
vpaddd ymm3, ymm3, ymm7
|
|
223
|
+
vpxor ymm12, ymm12, ymm0
|
|
224
|
+
vpxor ymm13, ymm13, ymm1
|
|
225
|
+
vpxor ymm14, ymm14, ymm2
|
|
226
|
+
vpxor ymm15, ymm15, ymm3
|
|
227
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
228
|
+
vpshufb ymm12, ymm12, ymm8
|
|
229
|
+
vpshufb ymm13, ymm13, ymm8
|
|
230
|
+
vpshufb ymm14, ymm14, ymm8
|
|
231
|
+
vpshufb ymm15, ymm15, ymm8
|
|
232
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
233
|
+
vpaddd ymm9, ymm9, ymm13
|
|
234
|
+
vpaddd ymm10, ymm10, ymm14
|
|
235
|
+
vpaddd ymm11, ymm11, ymm15
|
|
236
|
+
vpxor ymm4, ymm4, ymm8
|
|
237
|
+
vpxor ymm5, ymm5, ymm9
|
|
238
|
+
vpxor ymm6, ymm6, ymm10
|
|
239
|
+
vpxor ymm7, ymm7, ymm11
|
|
240
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
241
|
+
vpsrld ymm8, ymm4, 7
|
|
242
|
+
vpslld ymm4, ymm4, 25
|
|
243
|
+
vpor ymm4, ymm4, ymm8
|
|
244
|
+
vpsrld ymm8, ymm5, 7
|
|
245
|
+
vpslld ymm5, ymm5, 25
|
|
246
|
+
vpor ymm5, ymm5, ymm8
|
|
247
|
+
vpsrld ymm8, ymm6, 7
|
|
248
|
+
vpslld ymm6, ymm6, 25
|
|
249
|
+
vpor ymm6, ymm6, ymm8
|
|
250
|
+
vpsrld ymm8, ymm7, 7
|
|
251
|
+
vpslld ymm7, ymm7, 25
|
|
252
|
+
vpor ymm7, ymm7, ymm8
|
|
253
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+100H]
|
|
254
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+140H]
|
|
255
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+180H]
|
|
256
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1C0H]
|
|
257
|
+
vpaddd ymm0, ymm0, ymm5
|
|
258
|
+
vpaddd ymm1, ymm1, ymm6
|
|
259
|
+
vpaddd ymm2, ymm2, ymm7
|
|
260
|
+
vpaddd ymm3, ymm3, ymm4
|
|
261
|
+
vpxor ymm15, ymm15, ymm0
|
|
262
|
+
vpxor ymm12, ymm12, ymm1
|
|
263
|
+
vpxor ymm13, ymm13, ymm2
|
|
264
|
+
vpxor ymm14, ymm14, ymm3
|
|
265
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
266
|
+
vpshufb ymm15, ymm15, ymm8
|
|
267
|
+
vpshufb ymm12, ymm12, ymm8
|
|
268
|
+
vpshufb ymm13, ymm13, ymm8
|
|
269
|
+
vpshufb ymm14, ymm14, ymm8
|
|
270
|
+
vpaddd ymm10, ymm10, ymm15
|
|
271
|
+
vpaddd ymm11, ymm11, ymm12
|
|
272
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
273
|
+
vpaddd ymm9, ymm9, ymm14
|
|
274
|
+
vpxor ymm5, ymm5, ymm10
|
|
275
|
+
vpxor ymm6, ymm6, ymm11
|
|
276
|
+
vpxor ymm7, ymm7, ymm8
|
|
277
|
+
vpxor ymm4, ymm4, ymm9
|
|
278
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
279
|
+
vpsrld ymm8, ymm5, 12
|
|
280
|
+
vpslld ymm5, ymm5, 20
|
|
281
|
+
vpor ymm5, ymm5, ymm8
|
|
282
|
+
vpsrld ymm8, ymm6, 12
|
|
283
|
+
vpslld ymm6, ymm6, 20
|
|
284
|
+
vpor ymm6, ymm6, ymm8
|
|
285
|
+
vpsrld ymm8, ymm7, 12
|
|
286
|
+
vpslld ymm7, ymm7, 20
|
|
287
|
+
vpor ymm7, ymm7, ymm8
|
|
288
|
+
vpsrld ymm8, ymm4, 12
|
|
289
|
+
vpslld ymm4, ymm4, 20
|
|
290
|
+
vpor ymm4, ymm4, ymm8
|
|
291
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+120H]
|
|
292
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+160H]
|
|
293
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+1A0H]
|
|
294
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1E0H]
|
|
295
|
+
vpaddd ymm0, ymm0, ymm5
|
|
296
|
+
vpaddd ymm1, ymm1, ymm6
|
|
297
|
+
vpaddd ymm2, ymm2, ymm7
|
|
298
|
+
vpaddd ymm3, ymm3, ymm4
|
|
299
|
+
vpxor ymm15, ymm15, ymm0
|
|
300
|
+
vpxor ymm12, ymm12, ymm1
|
|
301
|
+
vpxor ymm13, ymm13, ymm2
|
|
302
|
+
vpxor ymm14, ymm14, ymm3
|
|
303
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
304
|
+
vpshufb ymm15, ymm15, ymm8
|
|
305
|
+
vpshufb ymm12, ymm12, ymm8
|
|
306
|
+
vpshufb ymm13, ymm13, ymm8
|
|
307
|
+
vpshufb ymm14, ymm14, ymm8
|
|
308
|
+
vpaddd ymm10, ymm10, ymm15
|
|
309
|
+
vpaddd ymm11, ymm11, ymm12
|
|
310
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
311
|
+
vpaddd ymm9, ymm9, ymm14
|
|
312
|
+
vpxor ymm5, ymm5, ymm10
|
|
313
|
+
vpxor ymm6, ymm6, ymm11
|
|
314
|
+
vpxor ymm7, ymm7, ymm8
|
|
315
|
+
vpxor ymm4, ymm4, ymm9
|
|
316
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
317
|
+
vpsrld ymm8, ymm5, 7
|
|
318
|
+
vpslld ymm5, ymm5, 25
|
|
319
|
+
vpor ymm5, ymm5, ymm8
|
|
320
|
+
vpsrld ymm8, ymm6, 7
|
|
321
|
+
vpslld ymm6, ymm6, 25
|
|
322
|
+
vpor ymm6, ymm6, ymm8
|
|
323
|
+
vpsrld ymm8, ymm7, 7
|
|
324
|
+
vpslld ymm7, ymm7, 25
|
|
325
|
+
vpor ymm7, ymm7, ymm8
|
|
326
|
+
vpsrld ymm8, ymm4, 7
|
|
327
|
+
vpslld ymm4, ymm4, 25
|
|
328
|
+
vpor ymm4, ymm4, ymm8
|
|
329
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+40H]
|
|
330
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+60H]
|
|
331
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+0E0H]
|
|
332
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+80H]
|
|
333
|
+
vpaddd ymm0, ymm0, ymm4
|
|
334
|
+
vpaddd ymm1, ymm1, ymm5
|
|
335
|
+
vpaddd ymm2, ymm2, ymm6
|
|
336
|
+
vpaddd ymm3, ymm3, ymm7
|
|
337
|
+
vpxor ymm12, ymm12, ymm0
|
|
338
|
+
vpxor ymm13, ymm13, ymm1
|
|
339
|
+
vpxor ymm14, ymm14, ymm2
|
|
340
|
+
vpxor ymm15, ymm15, ymm3
|
|
341
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
342
|
+
vpshufb ymm12, ymm12, ymm8
|
|
343
|
+
vpshufb ymm13, ymm13, ymm8
|
|
344
|
+
vpshufb ymm14, ymm14, ymm8
|
|
345
|
+
vpshufb ymm15, ymm15, ymm8
|
|
346
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
347
|
+
vpaddd ymm9, ymm9, ymm13
|
|
348
|
+
vpaddd ymm10, ymm10, ymm14
|
|
349
|
+
vpaddd ymm11, ymm11, ymm15
|
|
350
|
+
vpxor ymm4, ymm4, ymm8
|
|
351
|
+
vpxor ymm5, ymm5, ymm9
|
|
352
|
+
vpxor ymm6, ymm6, ymm10
|
|
353
|
+
vpxor ymm7, ymm7, ymm11
|
|
354
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
355
|
+
vpsrld ymm8, ymm4, 12
|
|
356
|
+
vpslld ymm4, ymm4, 20
|
|
357
|
+
vpor ymm4, ymm4, ymm8
|
|
358
|
+
vpsrld ymm8, ymm5, 12
|
|
359
|
+
vpslld ymm5, ymm5, 20
|
|
360
|
+
vpor ymm5, ymm5, ymm8
|
|
361
|
+
vpsrld ymm8, ymm6, 12
|
|
362
|
+
vpslld ymm6, ymm6, 20
|
|
363
|
+
vpor ymm6, ymm6, ymm8
|
|
364
|
+
vpsrld ymm8, ymm7, 12
|
|
365
|
+
vpslld ymm7, ymm7, 20
|
|
366
|
+
vpor ymm7, ymm7, ymm8
|
|
367
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+0C0H]
|
|
368
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+140H]
|
|
369
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp]
|
|
370
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1A0H]
|
|
371
|
+
vpaddd ymm0, ymm0, ymm4
|
|
372
|
+
vpaddd ymm1, ymm1, ymm5
|
|
373
|
+
vpaddd ymm2, ymm2, ymm6
|
|
374
|
+
vpaddd ymm3, ymm3, ymm7
|
|
375
|
+
vpxor ymm12, ymm12, ymm0
|
|
376
|
+
vpxor ymm13, ymm13, ymm1
|
|
377
|
+
vpxor ymm14, ymm14, ymm2
|
|
378
|
+
vpxor ymm15, ymm15, ymm3
|
|
379
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
380
|
+
vpshufb ymm12, ymm12, ymm8
|
|
381
|
+
vpshufb ymm13, ymm13, ymm8
|
|
382
|
+
vpshufb ymm14, ymm14, ymm8
|
|
383
|
+
vpshufb ymm15, ymm15, ymm8
|
|
384
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
385
|
+
vpaddd ymm9, ymm9, ymm13
|
|
386
|
+
vpaddd ymm10, ymm10, ymm14
|
|
387
|
+
vpaddd ymm11, ymm11, ymm15
|
|
388
|
+
vpxor ymm4, ymm4, ymm8
|
|
389
|
+
vpxor ymm5, ymm5, ymm9
|
|
390
|
+
vpxor ymm6, ymm6, ymm10
|
|
391
|
+
vpxor ymm7, ymm7, ymm11
|
|
392
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
393
|
+
vpsrld ymm8, ymm4, 7
|
|
394
|
+
vpslld ymm4, ymm4, 25
|
|
395
|
+
vpor ymm4, ymm4, ymm8
|
|
396
|
+
vpsrld ymm8, ymm5, 7
|
|
397
|
+
vpslld ymm5, ymm5, 25
|
|
398
|
+
vpor ymm5, ymm5, ymm8
|
|
399
|
+
vpsrld ymm8, ymm6, 7
|
|
400
|
+
vpslld ymm6, ymm6, 25
|
|
401
|
+
vpor ymm6, ymm6, ymm8
|
|
402
|
+
vpsrld ymm8, ymm7, 7
|
|
403
|
+
vpslld ymm7, ymm7, 25
|
|
404
|
+
vpor ymm7, ymm7, ymm8
|
|
405
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+20H]
|
|
406
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+180H]
|
|
407
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+120H]
|
|
408
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1E0H]
|
|
409
|
+
vpaddd ymm0, ymm0, ymm5
|
|
410
|
+
vpaddd ymm1, ymm1, ymm6
|
|
411
|
+
vpaddd ymm2, ymm2, ymm7
|
|
412
|
+
vpaddd ymm3, ymm3, ymm4
|
|
413
|
+
vpxor ymm15, ymm15, ymm0
|
|
414
|
+
vpxor ymm12, ymm12, ymm1
|
|
415
|
+
vpxor ymm13, ymm13, ymm2
|
|
416
|
+
vpxor ymm14, ymm14, ymm3
|
|
417
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
418
|
+
vpshufb ymm15, ymm15, ymm8
|
|
419
|
+
vpshufb ymm12, ymm12, ymm8
|
|
420
|
+
vpshufb ymm13, ymm13, ymm8
|
|
421
|
+
vpshufb ymm14, ymm14, ymm8
|
|
422
|
+
vpaddd ymm10, ymm10, ymm15
|
|
423
|
+
vpaddd ymm11, ymm11, ymm12
|
|
424
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
425
|
+
vpaddd ymm9, ymm9, ymm14
|
|
426
|
+
vpxor ymm5, ymm5, ymm10
|
|
427
|
+
vpxor ymm6, ymm6, ymm11
|
|
428
|
+
vpxor ymm7, ymm7, ymm8
|
|
429
|
+
vpxor ymm4, ymm4, ymm9
|
|
430
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
431
|
+
vpsrld ymm8, ymm5, 12
|
|
432
|
+
vpslld ymm5, ymm5, 20
|
|
433
|
+
vpor ymm5, ymm5, ymm8
|
|
434
|
+
vpsrld ymm8, ymm6, 12
|
|
435
|
+
vpslld ymm6, ymm6, 20
|
|
436
|
+
vpor ymm6, ymm6, ymm8
|
|
437
|
+
vpsrld ymm8, ymm7, 12
|
|
438
|
+
vpslld ymm7, ymm7, 20
|
|
439
|
+
vpor ymm7, ymm7, ymm8
|
|
440
|
+
vpsrld ymm8, ymm4, 12
|
|
441
|
+
vpslld ymm4, ymm4, 20
|
|
442
|
+
vpor ymm4, ymm4, ymm8
|
|
443
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+160H]
|
|
444
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+0A0H]
|
|
445
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+1C0H]
|
|
446
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+100H]
|
|
447
|
+
vpaddd ymm0, ymm0, ymm5
|
|
448
|
+
vpaddd ymm1, ymm1, ymm6
|
|
449
|
+
vpaddd ymm2, ymm2, ymm7
|
|
450
|
+
vpaddd ymm3, ymm3, ymm4
|
|
451
|
+
vpxor ymm15, ymm15, ymm0
|
|
452
|
+
vpxor ymm12, ymm12, ymm1
|
|
453
|
+
vpxor ymm13, ymm13, ymm2
|
|
454
|
+
vpxor ymm14, ymm14, ymm3
|
|
455
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
456
|
+
vpshufb ymm15, ymm15, ymm8
|
|
457
|
+
vpshufb ymm12, ymm12, ymm8
|
|
458
|
+
vpshufb ymm13, ymm13, ymm8
|
|
459
|
+
vpshufb ymm14, ymm14, ymm8
|
|
460
|
+
vpaddd ymm10, ymm10, ymm15
|
|
461
|
+
vpaddd ymm11, ymm11, ymm12
|
|
462
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
463
|
+
vpaddd ymm9, ymm9, ymm14
|
|
464
|
+
vpxor ymm5, ymm5, ymm10
|
|
465
|
+
vpxor ymm6, ymm6, ymm11
|
|
466
|
+
vpxor ymm7, ymm7, ymm8
|
|
467
|
+
vpxor ymm4, ymm4, ymm9
|
|
468
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
469
|
+
vpsrld ymm8, ymm5, 7
|
|
470
|
+
vpslld ymm5, ymm5, 25
|
|
471
|
+
vpor ymm5, ymm5, ymm8
|
|
472
|
+
vpsrld ymm8, ymm6, 7
|
|
473
|
+
vpslld ymm6, ymm6, 25
|
|
474
|
+
vpor ymm6, ymm6, ymm8
|
|
475
|
+
vpsrld ymm8, ymm7, 7
|
|
476
|
+
vpslld ymm7, ymm7, 25
|
|
477
|
+
vpor ymm7, ymm7, ymm8
|
|
478
|
+
vpsrld ymm8, ymm4, 7
|
|
479
|
+
vpslld ymm4, ymm4, 25
|
|
480
|
+
vpor ymm4, ymm4, ymm8
|
|
481
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+60H]
|
|
482
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+140H]
|
|
483
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+1A0H]
|
|
484
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+0E0H]
|
|
485
|
+
vpaddd ymm0, ymm0, ymm4
|
|
486
|
+
vpaddd ymm1, ymm1, ymm5
|
|
487
|
+
vpaddd ymm2, ymm2, ymm6
|
|
488
|
+
vpaddd ymm3, ymm3, ymm7
|
|
489
|
+
vpxor ymm12, ymm12, ymm0
|
|
490
|
+
vpxor ymm13, ymm13, ymm1
|
|
491
|
+
vpxor ymm14, ymm14, ymm2
|
|
492
|
+
vpxor ymm15, ymm15, ymm3
|
|
493
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
494
|
+
vpshufb ymm12, ymm12, ymm8
|
|
495
|
+
vpshufb ymm13, ymm13, ymm8
|
|
496
|
+
vpshufb ymm14, ymm14, ymm8
|
|
497
|
+
vpshufb ymm15, ymm15, ymm8
|
|
498
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
499
|
+
vpaddd ymm9, ymm9, ymm13
|
|
500
|
+
vpaddd ymm10, ymm10, ymm14
|
|
501
|
+
vpaddd ymm11, ymm11, ymm15
|
|
502
|
+
vpxor ymm4, ymm4, ymm8
|
|
503
|
+
vpxor ymm5, ymm5, ymm9
|
|
504
|
+
vpxor ymm6, ymm6, ymm10
|
|
505
|
+
vpxor ymm7, ymm7, ymm11
|
|
506
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
507
|
+
vpsrld ymm8, ymm4, 12
|
|
508
|
+
vpslld ymm4, ymm4, 20
|
|
509
|
+
vpor ymm4, ymm4, ymm8
|
|
510
|
+
vpsrld ymm8, ymm5, 12
|
|
511
|
+
vpslld ymm5, ymm5, 20
|
|
512
|
+
vpor ymm5, ymm5, ymm8
|
|
513
|
+
vpsrld ymm8, ymm6, 12
|
|
514
|
+
vpslld ymm6, ymm6, 20
|
|
515
|
+
vpor ymm6, ymm6, ymm8
|
|
516
|
+
vpsrld ymm8, ymm7, 12
|
|
517
|
+
vpslld ymm7, ymm7, 20
|
|
518
|
+
vpor ymm7, ymm7, ymm8
|
|
519
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+80H]
|
|
520
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+180H]
|
|
521
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+40H]
|
|
522
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1C0H]
|
|
523
|
+
vpaddd ymm0, ymm0, ymm4
|
|
524
|
+
vpaddd ymm1, ymm1, ymm5
|
|
525
|
+
vpaddd ymm2, ymm2, ymm6
|
|
526
|
+
vpaddd ymm3, ymm3, ymm7
|
|
527
|
+
vpxor ymm12, ymm12, ymm0
|
|
528
|
+
vpxor ymm13, ymm13, ymm1
|
|
529
|
+
vpxor ymm14, ymm14, ymm2
|
|
530
|
+
vpxor ymm15, ymm15, ymm3
|
|
531
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
532
|
+
vpshufb ymm12, ymm12, ymm8
|
|
533
|
+
vpshufb ymm13, ymm13, ymm8
|
|
534
|
+
vpshufb ymm14, ymm14, ymm8
|
|
535
|
+
vpshufb ymm15, ymm15, ymm8
|
|
536
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
537
|
+
vpaddd ymm9, ymm9, ymm13
|
|
538
|
+
vpaddd ymm10, ymm10, ymm14
|
|
539
|
+
vpaddd ymm11, ymm11, ymm15
|
|
540
|
+
vpxor ymm4, ymm4, ymm8
|
|
541
|
+
vpxor ymm5, ymm5, ymm9
|
|
542
|
+
vpxor ymm6, ymm6, ymm10
|
|
543
|
+
vpxor ymm7, ymm7, ymm11
|
|
544
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
545
|
+
vpsrld ymm8, ymm4, 7
|
|
546
|
+
vpslld ymm4, ymm4, 25
|
|
547
|
+
vpor ymm4, ymm4, ymm8
|
|
548
|
+
vpsrld ymm8, ymm5, 7
|
|
549
|
+
vpslld ymm5, ymm5, 25
|
|
550
|
+
vpor ymm5, ymm5, ymm8
|
|
551
|
+
vpsrld ymm8, ymm6, 7
|
|
552
|
+
vpslld ymm6, ymm6, 25
|
|
553
|
+
vpor ymm6, ymm6, ymm8
|
|
554
|
+
vpsrld ymm8, ymm7, 7
|
|
555
|
+
vpslld ymm7, ymm7, 25
|
|
556
|
+
vpor ymm7, ymm7, ymm8
|
|
557
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+0C0H]
|
|
558
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+120H]
|
|
559
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+160H]
|
|
560
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+100H]
|
|
561
|
+
vpaddd ymm0, ymm0, ymm5
|
|
562
|
+
vpaddd ymm1, ymm1, ymm6
|
|
563
|
+
vpaddd ymm2, ymm2, ymm7
|
|
564
|
+
vpaddd ymm3, ymm3, ymm4
|
|
565
|
+
vpxor ymm15, ymm15, ymm0
|
|
566
|
+
vpxor ymm12, ymm12, ymm1
|
|
567
|
+
vpxor ymm13, ymm13, ymm2
|
|
568
|
+
vpxor ymm14, ymm14, ymm3
|
|
569
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
570
|
+
vpshufb ymm15, ymm15, ymm8
|
|
571
|
+
vpshufb ymm12, ymm12, ymm8
|
|
572
|
+
vpshufb ymm13, ymm13, ymm8
|
|
573
|
+
vpshufb ymm14, ymm14, ymm8
|
|
574
|
+
vpaddd ymm10, ymm10, ymm15
|
|
575
|
+
vpaddd ymm11, ymm11, ymm12
|
|
576
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
577
|
+
vpaddd ymm9, ymm9, ymm14
|
|
578
|
+
vpxor ymm5, ymm5, ymm10
|
|
579
|
+
vpxor ymm6, ymm6, ymm11
|
|
580
|
+
vpxor ymm7, ymm7, ymm8
|
|
581
|
+
vpxor ymm4, ymm4, ymm9
|
|
582
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
583
|
+
vpsrld ymm8, ymm5, 12
|
|
584
|
+
vpslld ymm5, ymm5, 20
|
|
585
|
+
vpor ymm5, ymm5, ymm8
|
|
586
|
+
vpsrld ymm8, ymm6, 12
|
|
587
|
+
vpslld ymm6, ymm6, 20
|
|
588
|
+
vpor ymm6, ymm6, ymm8
|
|
589
|
+
vpsrld ymm8, ymm7, 12
|
|
590
|
+
vpslld ymm7, ymm7, 20
|
|
591
|
+
vpor ymm7, ymm7, ymm8
|
|
592
|
+
vpsrld ymm8, ymm4, 12
|
|
593
|
+
vpslld ymm4, ymm4, 20
|
|
594
|
+
vpor ymm4, ymm4, ymm8
|
|
595
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+0A0H]
|
|
596
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp]
|
|
597
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+1E0H]
|
|
598
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+20H]
|
|
599
|
+
vpaddd ymm0, ymm0, ymm5
|
|
600
|
+
vpaddd ymm1, ymm1, ymm6
|
|
601
|
+
vpaddd ymm2, ymm2, ymm7
|
|
602
|
+
vpaddd ymm3, ymm3, ymm4
|
|
603
|
+
vpxor ymm15, ymm15, ymm0
|
|
604
|
+
vpxor ymm12, ymm12, ymm1
|
|
605
|
+
vpxor ymm13, ymm13, ymm2
|
|
606
|
+
vpxor ymm14, ymm14, ymm3
|
|
607
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
608
|
+
vpshufb ymm15, ymm15, ymm8
|
|
609
|
+
vpshufb ymm12, ymm12, ymm8
|
|
610
|
+
vpshufb ymm13, ymm13, ymm8
|
|
611
|
+
vpshufb ymm14, ymm14, ymm8
|
|
612
|
+
vpaddd ymm10, ymm10, ymm15
|
|
613
|
+
vpaddd ymm11, ymm11, ymm12
|
|
614
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
615
|
+
vpaddd ymm9, ymm9, ymm14
|
|
616
|
+
vpxor ymm5, ymm5, ymm10
|
|
617
|
+
vpxor ymm6, ymm6, ymm11
|
|
618
|
+
vpxor ymm7, ymm7, ymm8
|
|
619
|
+
vpxor ymm4, ymm4, ymm9
|
|
620
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
621
|
+
vpsrld ymm8, ymm5, 7
|
|
622
|
+
vpslld ymm5, ymm5, 25
|
|
623
|
+
vpor ymm5, ymm5, ymm8
|
|
624
|
+
vpsrld ymm8, ymm6, 7
|
|
625
|
+
vpslld ymm6, ymm6, 25
|
|
626
|
+
vpor ymm6, ymm6, ymm8
|
|
627
|
+
vpsrld ymm8, ymm7, 7
|
|
628
|
+
vpslld ymm7, ymm7, 25
|
|
629
|
+
vpor ymm7, ymm7, ymm8
|
|
630
|
+
vpsrld ymm8, ymm4, 7
|
|
631
|
+
vpslld ymm4, ymm4, 25
|
|
632
|
+
vpor ymm4, ymm4, ymm8
|
|
633
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+140H]
|
|
634
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+180H]
|
|
635
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+1C0H]
|
|
636
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1A0H]
|
|
637
|
+
vpaddd ymm0, ymm0, ymm4
|
|
638
|
+
vpaddd ymm1, ymm1, ymm5
|
|
639
|
+
vpaddd ymm2, ymm2, ymm6
|
|
640
|
+
vpaddd ymm3, ymm3, ymm7
|
|
641
|
+
vpxor ymm12, ymm12, ymm0
|
|
642
|
+
vpxor ymm13, ymm13, ymm1
|
|
643
|
+
vpxor ymm14, ymm14, ymm2
|
|
644
|
+
vpxor ymm15, ymm15, ymm3
|
|
645
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
646
|
+
vpshufb ymm12, ymm12, ymm8
|
|
647
|
+
vpshufb ymm13, ymm13, ymm8
|
|
648
|
+
vpshufb ymm14, ymm14, ymm8
|
|
649
|
+
vpshufb ymm15, ymm15, ymm8
|
|
650
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
651
|
+
vpaddd ymm9, ymm9, ymm13
|
|
652
|
+
vpaddd ymm10, ymm10, ymm14
|
|
653
|
+
vpaddd ymm11, ymm11, ymm15
|
|
654
|
+
vpxor ymm4, ymm4, ymm8
|
|
655
|
+
vpxor ymm5, ymm5, ymm9
|
|
656
|
+
vpxor ymm6, ymm6, ymm10
|
|
657
|
+
vpxor ymm7, ymm7, ymm11
|
|
658
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
659
|
+
vpsrld ymm8, ymm4, 12
|
|
660
|
+
vpslld ymm4, ymm4, 20
|
|
661
|
+
vpor ymm4, ymm4, ymm8
|
|
662
|
+
vpsrld ymm8, ymm5, 12
|
|
663
|
+
vpslld ymm5, ymm5, 20
|
|
664
|
+
vpor ymm5, ymm5, ymm8
|
|
665
|
+
vpsrld ymm8, ymm6, 12
|
|
666
|
+
vpslld ymm6, ymm6, 20
|
|
667
|
+
vpor ymm6, ymm6, ymm8
|
|
668
|
+
vpsrld ymm8, ymm7, 12
|
|
669
|
+
vpslld ymm7, ymm7, 20
|
|
670
|
+
vpor ymm7, ymm7, ymm8
|
|
671
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+0E0H]
|
|
672
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+120H]
|
|
673
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+60H]
|
|
674
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1E0H]
|
|
675
|
+
vpaddd ymm0, ymm0, ymm4
|
|
676
|
+
vpaddd ymm1, ymm1, ymm5
|
|
677
|
+
vpaddd ymm2, ymm2, ymm6
|
|
678
|
+
vpaddd ymm3, ymm3, ymm7
|
|
679
|
+
vpxor ymm12, ymm12, ymm0
|
|
680
|
+
vpxor ymm13, ymm13, ymm1
|
|
681
|
+
vpxor ymm14, ymm14, ymm2
|
|
682
|
+
vpxor ymm15, ymm15, ymm3
|
|
683
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
684
|
+
vpshufb ymm12, ymm12, ymm8
|
|
685
|
+
vpshufb ymm13, ymm13, ymm8
|
|
686
|
+
vpshufb ymm14, ymm14, ymm8
|
|
687
|
+
vpshufb ymm15, ymm15, ymm8
|
|
688
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
689
|
+
vpaddd ymm9, ymm9, ymm13
|
|
690
|
+
vpaddd ymm10, ymm10, ymm14
|
|
691
|
+
vpaddd ymm11, ymm11, ymm15
|
|
692
|
+
vpxor ymm4, ymm4, ymm8
|
|
693
|
+
vpxor ymm5, ymm5, ymm9
|
|
694
|
+
vpxor ymm6, ymm6, ymm10
|
|
695
|
+
vpxor ymm7, ymm7, ymm11
|
|
696
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
697
|
+
vpsrld ymm8, ymm4, 7
|
|
698
|
+
vpslld ymm4, ymm4, 25
|
|
699
|
+
vpor ymm4, ymm4, ymm8
|
|
700
|
+
vpsrld ymm8, ymm5, 7
|
|
701
|
+
vpslld ymm5, ymm5, 25
|
|
702
|
+
vpor ymm5, ymm5, ymm8
|
|
703
|
+
vpsrld ymm8, ymm6, 7
|
|
704
|
+
vpslld ymm6, ymm6, 25
|
|
705
|
+
vpor ymm6, ymm6, ymm8
|
|
706
|
+
vpsrld ymm8, ymm7, 7
|
|
707
|
+
vpslld ymm7, ymm7, 25
|
|
708
|
+
vpor ymm7, ymm7, ymm8
|
|
709
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+80H]
|
|
710
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+160H]
|
|
711
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+0A0H]
|
|
712
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+20H]
|
|
713
|
+
vpaddd ymm0, ymm0, ymm5
|
|
714
|
+
vpaddd ymm1, ymm1, ymm6
|
|
715
|
+
vpaddd ymm2, ymm2, ymm7
|
|
716
|
+
vpaddd ymm3, ymm3, ymm4
|
|
717
|
+
vpxor ymm15, ymm15, ymm0
|
|
718
|
+
vpxor ymm12, ymm12, ymm1
|
|
719
|
+
vpxor ymm13, ymm13, ymm2
|
|
720
|
+
vpxor ymm14, ymm14, ymm3
|
|
721
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
722
|
+
vpshufb ymm15, ymm15, ymm8
|
|
723
|
+
vpshufb ymm12, ymm12, ymm8
|
|
724
|
+
vpshufb ymm13, ymm13, ymm8
|
|
725
|
+
vpshufb ymm14, ymm14, ymm8
|
|
726
|
+
vpaddd ymm10, ymm10, ymm15
|
|
727
|
+
vpaddd ymm11, ymm11, ymm12
|
|
728
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
729
|
+
vpaddd ymm9, ymm9, ymm14
|
|
730
|
+
vpxor ymm5, ymm5, ymm10
|
|
731
|
+
vpxor ymm6, ymm6, ymm11
|
|
732
|
+
vpxor ymm7, ymm7, ymm8
|
|
733
|
+
vpxor ymm4, ymm4, ymm9
|
|
734
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
735
|
+
vpsrld ymm8, ymm5, 12
|
|
736
|
+
vpslld ymm5, ymm5, 20
|
|
737
|
+
vpor ymm5, ymm5, ymm8
|
|
738
|
+
vpsrld ymm8, ymm6, 12
|
|
739
|
+
vpslld ymm6, ymm6, 20
|
|
740
|
+
vpor ymm6, ymm6, ymm8
|
|
741
|
+
vpsrld ymm8, ymm7, 12
|
|
742
|
+
vpslld ymm7, ymm7, 20
|
|
743
|
+
vpor ymm7, ymm7, ymm8
|
|
744
|
+
vpsrld ymm8, ymm4, 12
|
|
745
|
+
vpslld ymm4, ymm4, 20
|
|
746
|
+
vpor ymm4, ymm4, ymm8
|
|
747
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp]
|
|
748
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+40H]
|
|
749
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+100H]
|
|
750
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+0C0H]
|
|
751
|
+
vpaddd ymm0, ymm0, ymm5
|
|
752
|
+
vpaddd ymm1, ymm1, ymm6
|
|
753
|
+
vpaddd ymm2, ymm2, ymm7
|
|
754
|
+
vpaddd ymm3, ymm3, ymm4
|
|
755
|
+
vpxor ymm15, ymm15, ymm0
|
|
756
|
+
vpxor ymm12, ymm12, ymm1
|
|
757
|
+
vpxor ymm13, ymm13, ymm2
|
|
758
|
+
vpxor ymm14, ymm14, ymm3
|
|
759
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
760
|
+
vpshufb ymm15, ymm15, ymm8
|
|
761
|
+
vpshufb ymm12, ymm12, ymm8
|
|
762
|
+
vpshufb ymm13, ymm13, ymm8
|
|
763
|
+
vpshufb ymm14, ymm14, ymm8
|
|
764
|
+
vpaddd ymm10, ymm10, ymm15
|
|
765
|
+
vpaddd ymm11, ymm11, ymm12
|
|
766
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
767
|
+
vpaddd ymm9, ymm9, ymm14
|
|
768
|
+
vpxor ymm5, ymm5, ymm10
|
|
769
|
+
vpxor ymm6, ymm6, ymm11
|
|
770
|
+
vpxor ymm7, ymm7, ymm8
|
|
771
|
+
vpxor ymm4, ymm4, ymm9
|
|
772
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
773
|
+
vpsrld ymm8, ymm5, 7
|
|
774
|
+
vpslld ymm5, ymm5, 25
|
|
775
|
+
vpor ymm5, ymm5, ymm8
|
|
776
|
+
vpsrld ymm8, ymm6, 7
|
|
777
|
+
vpslld ymm6, ymm6, 25
|
|
778
|
+
vpor ymm6, ymm6, ymm8
|
|
779
|
+
vpsrld ymm8, ymm7, 7
|
|
780
|
+
vpslld ymm7, ymm7, 25
|
|
781
|
+
vpor ymm7, ymm7, ymm8
|
|
782
|
+
vpsrld ymm8, ymm4, 7
|
|
783
|
+
vpslld ymm4, ymm4, 25
|
|
784
|
+
vpor ymm4, ymm4, ymm8
|
|
785
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+180H]
|
|
786
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+120H]
|
|
787
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+1E0H]
|
|
788
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1C0H]
|
|
789
|
+
vpaddd ymm0, ymm0, ymm4
|
|
790
|
+
vpaddd ymm1, ymm1, ymm5
|
|
791
|
+
vpaddd ymm2, ymm2, ymm6
|
|
792
|
+
vpaddd ymm3, ymm3, ymm7
|
|
793
|
+
vpxor ymm12, ymm12, ymm0
|
|
794
|
+
vpxor ymm13, ymm13, ymm1
|
|
795
|
+
vpxor ymm14, ymm14, ymm2
|
|
796
|
+
vpxor ymm15, ymm15, ymm3
|
|
797
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
798
|
+
vpshufb ymm12, ymm12, ymm8
|
|
799
|
+
vpshufb ymm13, ymm13, ymm8
|
|
800
|
+
vpshufb ymm14, ymm14, ymm8
|
|
801
|
+
vpshufb ymm15, ymm15, ymm8
|
|
802
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
803
|
+
vpaddd ymm9, ymm9, ymm13
|
|
804
|
+
vpaddd ymm10, ymm10, ymm14
|
|
805
|
+
vpaddd ymm11, ymm11, ymm15
|
|
806
|
+
vpxor ymm4, ymm4, ymm8
|
|
807
|
+
vpxor ymm5, ymm5, ymm9
|
|
808
|
+
vpxor ymm6, ymm6, ymm10
|
|
809
|
+
vpxor ymm7, ymm7, ymm11
|
|
810
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
811
|
+
vpsrld ymm8, ymm4, 12
|
|
812
|
+
vpslld ymm4, ymm4, 20
|
|
813
|
+
vpor ymm4, ymm4, ymm8
|
|
814
|
+
vpsrld ymm8, ymm5, 12
|
|
815
|
+
vpslld ymm5, ymm5, 20
|
|
816
|
+
vpor ymm5, ymm5, ymm8
|
|
817
|
+
vpsrld ymm8, ymm6, 12
|
|
818
|
+
vpslld ymm6, ymm6, 20
|
|
819
|
+
vpor ymm6, ymm6, ymm8
|
|
820
|
+
vpsrld ymm8, ymm7, 12
|
|
821
|
+
vpslld ymm7, ymm7, 20
|
|
822
|
+
vpor ymm7, ymm7, ymm8
|
|
823
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+1A0H]
|
|
824
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+160H]
|
|
825
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+140H]
|
|
826
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+100H]
|
|
827
|
+
vpaddd ymm0, ymm0, ymm4
|
|
828
|
+
vpaddd ymm1, ymm1, ymm5
|
|
829
|
+
vpaddd ymm2, ymm2, ymm6
|
|
830
|
+
vpaddd ymm3, ymm3, ymm7
|
|
831
|
+
vpxor ymm12, ymm12, ymm0
|
|
832
|
+
vpxor ymm13, ymm13, ymm1
|
|
833
|
+
vpxor ymm14, ymm14, ymm2
|
|
834
|
+
vpxor ymm15, ymm15, ymm3
|
|
835
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
836
|
+
vpshufb ymm12, ymm12, ymm8
|
|
837
|
+
vpshufb ymm13, ymm13, ymm8
|
|
838
|
+
vpshufb ymm14, ymm14, ymm8
|
|
839
|
+
vpshufb ymm15, ymm15, ymm8
|
|
840
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
841
|
+
vpaddd ymm9, ymm9, ymm13
|
|
842
|
+
vpaddd ymm10, ymm10, ymm14
|
|
843
|
+
vpaddd ymm11, ymm11, ymm15
|
|
844
|
+
vpxor ymm4, ymm4, ymm8
|
|
845
|
+
vpxor ymm5, ymm5, ymm9
|
|
846
|
+
vpxor ymm6, ymm6, ymm10
|
|
847
|
+
vpxor ymm7, ymm7, ymm11
|
|
848
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
849
|
+
vpsrld ymm8, ymm4, 7
|
|
850
|
+
vpslld ymm4, ymm4, 25
|
|
851
|
+
vpor ymm4, ymm4, ymm8
|
|
852
|
+
vpsrld ymm8, ymm5, 7
|
|
853
|
+
vpslld ymm5, ymm5, 25
|
|
854
|
+
vpor ymm5, ymm5, ymm8
|
|
855
|
+
vpsrld ymm8, ymm6, 7
|
|
856
|
+
vpslld ymm6, ymm6, 25
|
|
857
|
+
vpor ymm6, ymm6, ymm8
|
|
858
|
+
vpsrld ymm8, ymm7, 7
|
|
859
|
+
vpslld ymm7, ymm7, 25
|
|
860
|
+
vpor ymm7, ymm7, ymm8
|
|
861
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+0E0H]
|
|
862
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+0A0H]
|
|
863
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp]
|
|
864
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+0C0H]
|
|
865
|
+
vpaddd ymm0, ymm0, ymm5
|
|
866
|
+
vpaddd ymm1, ymm1, ymm6
|
|
867
|
+
vpaddd ymm2, ymm2, ymm7
|
|
868
|
+
vpaddd ymm3, ymm3, ymm4
|
|
869
|
+
vpxor ymm15, ymm15, ymm0
|
|
870
|
+
vpxor ymm12, ymm12, ymm1
|
|
871
|
+
vpxor ymm13, ymm13, ymm2
|
|
872
|
+
vpxor ymm14, ymm14, ymm3
|
|
873
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
874
|
+
vpshufb ymm15, ymm15, ymm8
|
|
875
|
+
vpshufb ymm12, ymm12, ymm8
|
|
876
|
+
vpshufb ymm13, ymm13, ymm8
|
|
877
|
+
vpshufb ymm14, ymm14, ymm8
|
|
878
|
+
vpaddd ymm10, ymm10, ymm15
|
|
879
|
+
vpaddd ymm11, ymm11, ymm12
|
|
880
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
881
|
+
vpaddd ymm9, ymm9, ymm14
|
|
882
|
+
vpxor ymm5, ymm5, ymm10
|
|
883
|
+
vpxor ymm6, ymm6, ymm11
|
|
884
|
+
vpxor ymm7, ymm7, ymm8
|
|
885
|
+
vpxor ymm4, ymm4, ymm9
|
|
886
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
887
|
+
vpsrld ymm8, ymm5, 12
|
|
888
|
+
vpslld ymm5, ymm5, 20
|
|
889
|
+
vpor ymm5, ymm5, ymm8
|
|
890
|
+
vpsrld ymm8, ymm6, 12
|
|
891
|
+
vpslld ymm6, ymm6, 20
|
|
892
|
+
vpor ymm6, ymm6, ymm8
|
|
893
|
+
vpsrld ymm8, ymm7, 12
|
|
894
|
+
vpslld ymm7, ymm7, 20
|
|
895
|
+
vpor ymm7, ymm7, ymm8
|
|
896
|
+
vpsrld ymm8, ymm4, 12
|
|
897
|
+
vpslld ymm4, ymm4, 20
|
|
898
|
+
vpor ymm4, ymm4, ymm8
|
|
899
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+40H]
|
|
900
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+60H]
|
|
901
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+20H]
|
|
902
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+80H]
|
|
903
|
+
vpaddd ymm0, ymm0, ymm5
|
|
904
|
+
vpaddd ymm1, ymm1, ymm6
|
|
905
|
+
vpaddd ymm2, ymm2, ymm7
|
|
906
|
+
vpaddd ymm3, ymm3, ymm4
|
|
907
|
+
vpxor ymm15, ymm15, ymm0
|
|
908
|
+
vpxor ymm12, ymm12, ymm1
|
|
909
|
+
vpxor ymm13, ymm13, ymm2
|
|
910
|
+
vpxor ymm14, ymm14, ymm3
|
|
911
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
912
|
+
vpshufb ymm15, ymm15, ymm8
|
|
913
|
+
vpshufb ymm12, ymm12, ymm8
|
|
914
|
+
vpshufb ymm13, ymm13, ymm8
|
|
915
|
+
vpshufb ymm14, ymm14, ymm8
|
|
916
|
+
vpaddd ymm10, ymm10, ymm15
|
|
917
|
+
vpaddd ymm11, ymm11, ymm12
|
|
918
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
919
|
+
vpaddd ymm9, ymm9, ymm14
|
|
920
|
+
vpxor ymm5, ymm5, ymm10
|
|
921
|
+
vpxor ymm6, ymm6, ymm11
|
|
922
|
+
vpxor ymm7, ymm7, ymm8
|
|
923
|
+
vpxor ymm4, ymm4, ymm9
|
|
924
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
925
|
+
vpsrld ymm8, ymm5, 7
|
|
926
|
+
vpslld ymm5, ymm5, 25
|
|
927
|
+
vpor ymm5, ymm5, ymm8
|
|
928
|
+
vpsrld ymm8, ymm6, 7
|
|
929
|
+
vpslld ymm6, ymm6, 25
|
|
930
|
+
vpor ymm6, ymm6, ymm8
|
|
931
|
+
vpsrld ymm8, ymm7, 7
|
|
932
|
+
vpslld ymm7, ymm7, 25
|
|
933
|
+
vpor ymm7, ymm7, ymm8
|
|
934
|
+
vpsrld ymm8, ymm4, 7
|
|
935
|
+
vpslld ymm4, ymm4, 25
|
|
936
|
+
vpor ymm4, ymm4, ymm8
|
|
937
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+120H]
|
|
938
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+160H]
|
|
939
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+100H]
|
|
940
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1E0H]
|
|
941
|
+
vpaddd ymm0, ymm0, ymm4
|
|
942
|
+
vpaddd ymm1, ymm1, ymm5
|
|
943
|
+
vpaddd ymm2, ymm2, ymm6
|
|
944
|
+
vpaddd ymm3, ymm3, ymm7
|
|
945
|
+
vpxor ymm12, ymm12, ymm0
|
|
946
|
+
vpxor ymm13, ymm13, ymm1
|
|
947
|
+
vpxor ymm14, ymm14, ymm2
|
|
948
|
+
vpxor ymm15, ymm15, ymm3
|
|
949
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
950
|
+
vpshufb ymm12, ymm12, ymm8
|
|
951
|
+
vpshufb ymm13, ymm13, ymm8
|
|
952
|
+
vpshufb ymm14, ymm14, ymm8
|
|
953
|
+
vpshufb ymm15, ymm15, ymm8
|
|
954
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
955
|
+
vpaddd ymm9, ymm9, ymm13
|
|
956
|
+
vpaddd ymm10, ymm10, ymm14
|
|
957
|
+
vpaddd ymm11, ymm11, ymm15
|
|
958
|
+
vpxor ymm4, ymm4, ymm8
|
|
959
|
+
vpxor ymm5, ymm5, ymm9
|
|
960
|
+
vpxor ymm6, ymm6, ymm10
|
|
961
|
+
vpxor ymm7, ymm7, ymm11
|
|
962
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
963
|
+
vpsrld ymm8, ymm4, 12
|
|
964
|
+
vpslld ymm4, ymm4, 20
|
|
965
|
+
vpor ymm4, ymm4, ymm8
|
|
966
|
+
vpsrld ymm8, ymm5, 12
|
|
967
|
+
vpslld ymm5, ymm5, 20
|
|
968
|
+
vpor ymm5, ymm5, ymm8
|
|
969
|
+
vpsrld ymm8, ymm6, 12
|
|
970
|
+
vpslld ymm6, ymm6, 20
|
|
971
|
+
vpor ymm6, ymm6, ymm8
|
|
972
|
+
vpsrld ymm8, ymm7, 12
|
|
973
|
+
vpslld ymm7, ymm7, 20
|
|
974
|
+
vpor ymm7, ymm7, ymm8
|
|
975
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+1C0H]
|
|
976
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+0A0H]
|
|
977
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+180H]
|
|
978
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+20H]
|
|
979
|
+
vpaddd ymm0, ymm0, ymm4
|
|
980
|
+
vpaddd ymm1, ymm1, ymm5
|
|
981
|
+
vpaddd ymm2, ymm2, ymm6
|
|
982
|
+
vpaddd ymm3, ymm3, ymm7
|
|
983
|
+
vpxor ymm12, ymm12, ymm0
|
|
984
|
+
vpxor ymm13, ymm13, ymm1
|
|
985
|
+
vpxor ymm14, ymm14, ymm2
|
|
986
|
+
vpxor ymm15, ymm15, ymm3
|
|
987
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
988
|
+
vpshufb ymm12, ymm12, ymm8
|
|
989
|
+
vpshufb ymm13, ymm13, ymm8
|
|
990
|
+
vpshufb ymm14, ymm14, ymm8
|
|
991
|
+
vpshufb ymm15, ymm15, ymm8
|
|
992
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
993
|
+
vpaddd ymm9, ymm9, ymm13
|
|
994
|
+
vpaddd ymm10, ymm10, ymm14
|
|
995
|
+
vpaddd ymm11, ymm11, ymm15
|
|
996
|
+
vpxor ymm4, ymm4, ymm8
|
|
997
|
+
vpxor ymm5, ymm5, ymm9
|
|
998
|
+
vpxor ymm6, ymm6, ymm10
|
|
999
|
+
vpxor ymm7, ymm7, ymm11
|
|
1000
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
1001
|
+
vpsrld ymm8, ymm4, 7
|
|
1002
|
+
vpslld ymm4, ymm4, 25
|
|
1003
|
+
vpor ymm4, ymm4, ymm8
|
|
1004
|
+
vpsrld ymm8, ymm5, 7
|
|
1005
|
+
vpslld ymm5, ymm5, 25
|
|
1006
|
+
vpor ymm5, ymm5, ymm8
|
|
1007
|
+
vpsrld ymm8, ymm6, 7
|
|
1008
|
+
vpslld ymm6, ymm6, 25
|
|
1009
|
+
vpor ymm6, ymm6, ymm8
|
|
1010
|
+
vpsrld ymm8, ymm7, 7
|
|
1011
|
+
vpslld ymm7, ymm7, 25
|
|
1012
|
+
vpor ymm7, ymm7, ymm8
|
|
1013
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+1A0H]
|
|
1014
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp]
|
|
1015
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+40H]
|
|
1016
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+80H]
|
|
1017
|
+
vpaddd ymm0, ymm0, ymm5
|
|
1018
|
+
vpaddd ymm1, ymm1, ymm6
|
|
1019
|
+
vpaddd ymm2, ymm2, ymm7
|
|
1020
|
+
vpaddd ymm3, ymm3, ymm4
|
|
1021
|
+
vpxor ymm15, ymm15, ymm0
|
|
1022
|
+
vpxor ymm12, ymm12, ymm1
|
|
1023
|
+
vpxor ymm13, ymm13, ymm2
|
|
1024
|
+
vpxor ymm14, ymm14, ymm3
|
|
1025
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
1026
|
+
vpshufb ymm15, ymm15, ymm8
|
|
1027
|
+
vpshufb ymm12, ymm12, ymm8
|
|
1028
|
+
vpshufb ymm13, ymm13, ymm8
|
|
1029
|
+
vpshufb ymm14, ymm14, ymm8
|
|
1030
|
+
vpaddd ymm10, ymm10, ymm15
|
|
1031
|
+
vpaddd ymm11, ymm11, ymm12
|
|
1032
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
1033
|
+
vpaddd ymm9, ymm9, ymm14
|
|
1034
|
+
vpxor ymm5, ymm5, ymm10
|
|
1035
|
+
vpxor ymm6, ymm6, ymm11
|
|
1036
|
+
vpxor ymm7, ymm7, ymm8
|
|
1037
|
+
vpxor ymm4, ymm4, ymm9
|
|
1038
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
1039
|
+
vpsrld ymm8, ymm5, 12
|
|
1040
|
+
vpslld ymm5, ymm5, 20
|
|
1041
|
+
vpor ymm5, ymm5, ymm8
|
|
1042
|
+
vpsrld ymm8, ymm6, 12
|
|
1043
|
+
vpslld ymm6, ymm6, 20
|
|
1044
|
+
vpor ymm6, ymm6, ymm8
|
|
1045
|
+
vpsrld ymm8, ymm7, 12
|
|
1046
|
+
vpslld ymm7, ymm7, 20
|
|
1047
|
+
vpor ymm7, ymm7, ymm8
|
|
1048
|
+
vpsrld ymm8, ymm4, 12
|
|
1049
|
+
vpslld ymm4, ymm4, 20
|
|
1050
|
+
vpor ymm4, ymm4, ymm8
|
|
1051
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+60H]
|
|
1052
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+140H]
|
|
1053
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+0C0H]
|
|
1054
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+0E0H]
|
|
1055
|
+
vpaddd ymm0, ymm0, ymm5
|
|
1056
|
+
vpaddd ymm1, ymm1, ymm6
|
|
1057
|
+
vpaddd ymm2, ymm2, ymm7
|
|
1058
|
+
vpaddd ymm3, ymm3, ymm4
|
|
1059
|
+
vpxor ymm15, ymm15, ymm0
|
|
1060
|
+
vpxor ymm12, ymm12, ymm1
|
|
1061
|
+
vpxor ymm13, ymm13, ymm2
|
|
1062
|
+
vpxor ymm14, ymm14, ymm3
|
|
1063
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
1064
|
+
vpshufb ymm15, ymm15, ymm8
|
|
1065
|
+
vpshufb ymm12, ymm12, ymm8
|
|
1066
|
+
vpshufb ymm13, ymm13, ymm8
|
|
1067
|
+
vpshufb ymm14, ymm14, ymm8
|
|
1068
|
+
vpaddd ymm10, ymm10, ymm15
|
|
1069
|
+
vpaddd ymm11, ymm11, ymm12
|
|
1070
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
1071
|
+
vpaddd ymm9, ymm9, ymm14
|
|
1072
|
+
vpxor ymm5, ymm5, ymm10
|
|
1073
|
+
vpxor ymm6, ymm6, ymm11
|
|
1074
|
+
vpxor ymm7, ymm7, ymm8
|
|
1075
|
+
vpxor ymm4, ymm4, ymm9
|
|
1076
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
1077
|
+
vpsrld ymm8, ymm5, 7
|
|
1078
|
+
vpslld ymm5, ymm5, 25
|
|
1079
|
+
vpor ymm5, ymm5, ymm8
|
|
1080
|
+
vpsrld ymm8, ymm6, 7
|
|
1081
|
+
vpslld ymm6, ymm6, 25
|
|
1082
|
+
vpor ymm6, ymm6, ymm8
|
|
1083
|
+
vpsrld ymm8, ymm7, 7
|
|
1084
|
+
vpslld ymm7, ymm7, 25
|
|
1085
|
+
vpor ymm7, ymm7, ymm8
|
|
1086
|
+
vpsrld ymm8, ymm4, 7
|
|
1087
|
+
vpslld ymm4, ymm4, 25
|
|
1088
|
+
vpor ymm4, ymm4, ymm8
|
|
1089
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+160H]
|
|
1090
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+0A0H]
|
|
1091
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+20H]
|
|
1092
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+100H]
|
|
1093
|
+
vpaddd ymm0, ymm0, ymm4
|
|
1094
|
+
vpaddd ymm1, ymm1, ymm5
|
|
1095
|
+
vpaddd ymm2, ymm2, ymm6
|
|
1096
|
+
vpaddd ymm3, ymm3, ymm7
|
|
1097
|
+
vpxor ymm12, ymm12, ymm0
|
|
1098
|
+
vpxor ymm13, ymm13, ymm1
|
|
1099
|
+
vpxor ymm14, ymm14, ymm2
|
|
1100
|
+
vpxor ymm15, ymm15, ymm3
|
|
1101
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
1102
|
+
vpshufb ymm12, ymm12, ymm8
|
|
1103
|
+
vpshufb ymm13, ymm13, ymm8
|
|
1104
|
+
vpshufb ymm14, ymm14, ymm8
|
|
1105
|
+
vpshufb ymm15, ymm15, ymm8
|
|
1106
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
1107
|
+
vpaddd ymm9, ymm9, ymm13
|
|
1108
|
+
vpaddd ymm10, ymm10, ymm14
|
|
1109
|
+
vpaddd ymm11, ymm11, ymm15
|
|
1110
|
+
vpxor ymm4, ymm4, ymm8
|
|
1111
|
+
vpxor ymm5, ymm5, ymm9
|
|
1112
|
+
vpxor ymm6, ymm6, ymm10
|
|
1113
|
+
vpxor ymm7, ymm7, ymm11
|
|
1114
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
1115
|
+
vpsrld ymm8, ymm4, 12
|
|
1116
|
+
vpslld ymm4, ymm4, 20
|
|
1117
|
+
vpor ymm4, ymm4, ymm8
|
|
1118
|
+
vpsrld ymm8, ymm5, 12
|
|
1119
|
+
vpslld ymm5, ymm5, 20
|
|
1120
|
+
vpor ymm5, ymm5, ymm8
|
|
1121
|
+
vpsrld ymm8, ymm6, 12
|
|
1122
|
+
vpslld ymm6, ymm6, 20
|
|
1123
|
+
vpor ymm6, ymm6, ymm8
|
|
1124
|
+
vpsrld ymm8, ymm7, 12
|
|
1125
|
+
vpslld ymm7, ymm7, 20
|
|
1126
|
+
vpor ymm7, ymm7, ymm8
|
|
1127
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+1E0H]
|
|
1128
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp]
|
|
1129
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+120H]
|
|
1130
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+0C0H]
|
|
1131
|
+
vpaddd ymm0, ymm0, ymm4
|
|
1132
|
+
vpaddd ymm1, ymm1, ymm5
|
|
1133
|
+
vpaddd ymm2, ymm2, ymm6
|
|
1134
|
+
vpaddd ymm3, ymm3, ymm7
|
|
1135
|
+
vpxor ymm12, ymm12, ymm0
|
|
1136
|
+
vpxor ymm13, ymm13, ymm1
|
|
1137
|
+
vpxor ymm14, ymm14, ymm2
|
|
1138
|
+
vpxor ymm15, ymm15, ymm3
|
|
1139
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
1140
|
+
vpshufb ymm12, ymm12, ymm8
|
|
1141
|
+
vpshufb ymm13, ymm13, ymm8
|
|
1142
|
+
vpshufb ymm14, ymm14, ymm8
|
|
1143
|
+
vpshufb ymm15, ymm15, ymm8
|
|
1144
|
+
vpaddd ymm8, ymm12, ymmword ptr [rsp+200H]
|
|
1145
|
+
vpaddd ymm9, ymm9, ymm13
|
|
1146
|
+
vpaddd ymm10, ymm10, ymm14
|
|
1147
|
+
vpaddd ymm11, ymm11, ymm15
|
|
1148
|
+
vpxor ymm4, ymm4, ymm8
|
|
1149
|
+
vpxor ymm5, ymm5, ymm9
|
|
1150
|
+
vpxor ymm6, ymm6, ymm10
|
|
1151
|
+
vpxor ymm7, ymm7, ymm11
|
|
1152
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
1153
|
+
vpsrld ymm8, ymm4, 7
|
|
1154
|
+
vpslld ymm4, ymm4, 25
|
|
1155
|
+
vpor ymm4, ymm4, ymm8
|
|
1156
|
+
vpsrld ymm8, ymm5, 7
|
|
1157
|
+
vpslld ymm5, ymm5, 25
|
|
1158
|
+
vpor ymm5, ymm5, ymm8
|
|
1159
|
+
vpsrld ymm8, ymm6, 7
|
|
1160
|
+
vpslld ymm6, ymm6, 25
|
|
1161
|
+
vpor ymm6, ymm6, ymm8
|
|
1162
|
+
vpsrld ymm8, ymm7, 7
|
|
1163
|
+
vpslld ymm7, ymm7, 25
|
|
1164
|
+
vpor ymm7, ymm7, ymm8
|
|
1165
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+1C0H]
|
|
1166
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+40H]
|
|
1167
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+60H]
|
|
1168
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+0E0H]
|
|
1169
|
+
vpaddd ymm0, ymm0, ymm5
|
|
1170
|
+
vpaddd ymm1, ymm1, ymm6
|
|
1171
|
+
vpaddd ymm2, ymm2, ymm7
|
|
1172
|
+
vpaddd ymm3, ymm3, ymm4
|
|
1173
|
+
vpxor ymm15, ymm15, ymm0
|
|
1174
|
+
vpxor ymm12, ymm12, ymm1
|
|
1175
|
+
vpxor ymm13, ymm13, ymm2
|
|
1176
|
+
vpxor ymm14, ymm14, ymm3
|
|
1177
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT16]
|
|
1178
|
+
vpshufb ymm15, ymm15, ymm8
|
|
1179
|
+
vpshufb ymm12, ymm12, ymm8
|
|
1180
|
+
vpshufb ymm13, ymm13, ymm8
|
|
1181
|
+
vpshufb ymm14, ymm14, ymm8
|
|
1182
|
+
vpaddd ymm10, ymm10, ymm15
|
|
1183
|
+
vpaddd ymm11, ymm11, ymm12
|
|
1184
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
1185
|
+
vpaddd ymm9, ymm9, ymm14
|
|
1186
|
+
vpxor ymm5, ymm5, ymm10
|
|
1187
|
+
vpxor ymm6, ymm6, ymm11
|
|
1188
|
+
vpxor ymm7, ymm7, ymm8
|
|
1189
|
+
vpxor ymm4, ymm4, ymm9
|
|
1190
|
+
vmovdqa ymmword ptr [rsp+200H], ymm8
|
|
1191
|
+
vpsrld ymm8, ymm5, 12
|
|
1192
|
+
vpslld ymm5, ymm5, 20
|
|
1193
|
+
vpor ymm5, ymm5, ymm8
|
|
1194
|
+
vpsrld ymm8, ymm6, 12
|
|
1195
|
+
vpslld ymm6, ymm6, 20
|
|
1196
|
+
vpor ymm6, ymm6, ymm8
|
|
1197
|
+
vpsrld ymm8, ymm7, 12
|
|
1198
|
+
vpslld ymm7, ymm7, 20
|
|
1199
|
+
vpor ymm7, ymm7, ymm8
|
|
1200
|
+
vpsrld ymm8, ymm4, 12
|
|
1201
|
+
vpslld ymm4, ymm4, 20
|
|
1202
|
+
vpor ymm4, ymm4, ymm8
|
|
1203
|
+
vpaddd ymm0, ymm0, ymmword ptr [rsp+140H]
|
|
1204
|
+
vpaddd ymm1, ymm1, ymmword ptr [rsp+180H]
|
|
1205
|
+
vpaddd ymm2, ymm2, ymmword ptr [rsp+80H]
|
|
1206
|
+
vpaddd ymm3, ymm3, ymmword ptr [rsp+1A0H]
|
|
1207
|
+
vpaddd ymm0, ymm0, ymm5
|
|
1208
|
+
vpaddd ymm1, ymm1, ymm6
|
|
1209
|
+
vpaddd ymm2, ymm2, ymm7
|
|
1210
|
+
vpaddd ymm3, ymm3, ymm4
|
|
1211
|
+
vpxor ymm15, ymm15, ymm0
|
|
1212
|
+
vpxor ymm12, ymm12, ymm1
|
|
1213
|
+
vpxor ymm13, ymm13, ymm2
|
|
1214
|
+
vpxor ymm14, ymm14, ymm3
|
|
1215
|
+
vbroadcasti128 ymm8, xmmword ptr [ROT8]
|
|
1216
|
+
vpshufb ymm15, ymm15, ymm8
|
|
1217
|
+
vpshufb ymm12, ymm12, ymm8
|
|
1218
|
+
vpshufb ymm13, ymm13, ymm8
|
|
1219
|
+
vpshufb ymm14, ymm14, ymm8
|
|
1220
|
+
vpaddd ymm10, ymm10, ymm15
|
|
1221
|
+
vpaddd ymm11, ymm11, ymm12
|
|
1222
|
+
vpaddd ymm8, ymm13, ymmword ptr [rsp+200H]
|
|
1223
|
+
vpaddd ymm9, ymm9, ymm14
|
|
1224
|
+
vpxor ymm5, ymm5, ymm10
|
|
1225
|
+
vpxor ymm6, ymm6, ymm11
|
|
1226
|
+
vpxor ymm7, ymm7, ymm8
|
|
1227
|
+
vpxor ymm4, ymm4, ymm9
|
|
1228
|
+
vpxor ymm0, ymm0, ymm8
|
|
1229
|
+
vpxor ymm1, ymm1, ymm9
|
|
1230
|
+
vpxor ymm2, ymm2, ymm10
|
|
1231
|
+
vpxor ymm3, ymm3, ymm11
|
|
1232
|
+
vpsrld ymm8, ymm5, 7
|
|
1233
|
+
vpslld ymm5, ymm5, 25
|
|
1234
|
+
vpor ymm5, ymm5, ymm8
|
|
1235
|
+
vpsrld ymm8, ymm6, 7
|
|
1236
|
+
vpslld ymm6, ymm6, 25
|
|
1237
|
+
vpor ymm6, ymm6, ymm8
|
|
1238
|
+
vpsrld ymm8, ymm7, 7
|
|
1239
|
+
vpslld ymm7, ymm7, 25
|
|
1240
|
+
vpor ymm7, ymm7, ymm8
|
|
1241
|
+
vpsrld ymm8, ymm4, 7
|
|
1242
|
+
vpslld ymm4, ymm4, 25
|
|
1243
|
+
vpor ymm4, ymm4, ymm8
|
|
1244
|
+
vpxor ymm4, ymm4, ymm12
|
|
1245
|
+
vpxor ymm5, ymm5, ymm13
|
|
1246
|
+
vpxor ymm6, ymm6, ymm14
|
|
1247
|
+
vpxor ymm7, ymm7, ymm15
|
|
1248
|
+
movzx eax, byte ptr [rbp+78H]
|
|
1249
|
+
jne innerloop8
|
|
1250
|
+
mov rbx, qword ptr [rbp+90H]
|
|
1251
|
+
vunpcklps ymm8, ymm0, ymm1
|
|
1252
|
+
vunpcklps ymm9, ymm2, ymm3
|
|
1253
|
+
vunpckhps ymm10, ymm0, ymm1
|
|
1254
|
+
vunpcklps ymm11, ymm4, ymm5
|
|
1255
|
+
vunpcklps ymm0, ymm6, ymm7
|
|
1256
|
+
vshufps ymm12, ymm8, ymm9, 78
|
|
1257
|
+
vblendps ymm1, ymm8, ymm12, 0CCH
|
|
1258
|
+
vshufps ymm8, ymm11, ymm0, 78
|
|
1259
|
+
vunpckhps ymm13, ymm2, ymm3
|
|
1260
|
+
vblendps ymm2, ymm11, ymm8, 0CCH
|
|
1261
|
+
vblendps ymm3, ymm12, ymm9, 0CCH
|
|
1262
|
+
vperm2f128 ymm12, ymm1, ymm2, 20H
|
|
1263
|
+
vmovups ymmword ptr [rbx], ymm12
|
|
1264
|
+
vunpckhps ymm14, ymm4, ymm5
|
|
1265
|
+
vblendps ymm4, ymm8, ymm0, 0CCH
|
|
1266
|
+
vunpckhps ymm15, ymm6, ymm7
|
|
1267
|
+
vperm2f128 ymm7, ymm3, ymm4, 20H
|
|
1268
|
+
vmovups ymmword ptr [rbx+20H], ymm7
|
|
1269
|
+
vshufps ymm5, ymm10, ymm13, 78
|
|
1270
|
+
vblendps ymm6, ymm5, ymm13, 0CCH
|
|
1271
|
+
vshufps ymm13, ymm14, ymm15, 78
|
|
1272
|
+
vblendps ymm10, ymm10, ymm5, 0CCH
|
|
1273
|
+
vblendps ymm14, ymm14, ymm13, 0CCH
|
|
1274
|
+
vperm2f128 ymm8, ymm10, ymm14, 20H
|
|
1275
|
+
vmovups ymmword ptr [rbx+40H], ymm8
|
|
1276
|
+
vblendps ymm15, ymm13, ymm15, 0CCH
|
|
1277
|
+
vperm2f128 ymm13, ymm6, ymm15, 20H
|
|
1278
|
+
vmovups ymmword ptr [rbx+60H], ymm13
|
|
1279
|
+
vperm2f128 ymm9, ymm1, ymm2, 31H
|
|
1280
|
+
vperm2f128 ymm11, ymm3, ymm4, 31H
|
|
1281
|
+
vmovups ymmword ptr [rbx+80H], ymm9
|
|
1282
|
+
vperm2f128 ymm14, ymm10, ymm14, 31H
|
|
1283
|
+
vperm2f128 ymm15, ymm6, ymm15, 31H
|
|
1284
|
+
vmovups ymmword ptr [rbx+0A0H], ymm11
|
|
1285
|
+
vmovups ymmword ptr [rbx+0C0H], ymm14
|
|
1286
|
+
vmovups ymmword ptr [rbx+0E0H], ymm15
|
|
1287
|
+
vmovdqa ymm0, ymmword ptr [rsp+2A0H]
|
|
1288
|
+
vpaddd ymm1, ymm0, ymmword ptr [rsp+220H]
|
|
1289
|
+
vmovdqa ymmword ptr [rsp+220H], ymm1
|
|
1290
|
+
vpxor ymm0, ymm0, ymmword ptr [CMP_MSB_MASK]
|
|
1291
|
+
vpxor ymm2, ymm1, ymmword ptr [CMP_MSB_MASK]
|
|
1292
|
+
vpcmpgtd ymm2, ymm0, ymm2
|
|
1293
|
+
vmovdqa ymm0, ymmword ptr [rsp+240H]
|
|
1294
|
+
vpsubd ymm2, ymm0, ymm2
|
|
1295
|
+
vmovdqa ymmword ptr [rsp+240H], ymm2
|
|
1296
|
+
add rdi, 64
|
|
1297
|
+
add rbx, 256
|
|
1298
|
+
mov qword ptr [rbp+90H], rbx
|
|
1299
|
+
sub rsi, 8
|
|
1300
|
+
cmp rsi, 8
|
|
1301
|
+
jnc outerloop8
|
|
1302
|
+
test rsi, rsi
|
|
1303
|
+
jnz final7blocks
|
|
1304
|
+
unwind:
|
|
1305
|
+
vzeroupper
|
|
1306
|
+
vmovdqa xmm6, xmmword ptr [rsp+2D0H]
|
|
1307
|
+
vmovdqa xmm7, xmmword ptr [rsp+2E0H]
|
|
1308
|
+
vmovdqa xmm8, xmmword ptr [rsp+2F0H]
|
|
1309
|
+
vmovdqa xmm9, xmmword ptr [rsp+300H]
|
|
1310
|
+
vmovdqa xmm10, xmmword ptr [rsp+310H]
|
|
1311
|
+
vmovdqa xmm11, xmmword ptr [rsp+320H]
|
|
1312
|
+
vmovdqa xmm12, xmmword ptr [rsp+330H]
|
|
1313
|
+
vmovdqa xmm13, xmmword ptr [rsp+340H]
|
|
1314
|
+
vmovdqa xmm14, xmmword ptr [rsp+350H]
|
|
1315
|
+
vmovdqa xmm15, xmmword ptr [rsp+360H]
|
|
1316
|
+
mov rsp, rbp
|
|
1317
|
+
pop rbp
|
|
1318
|
+
pop rbx
|
|
1319
|
+
pop rdi
|
|
1320
|
+
pop rsi
|
|
1321
|
+
pop r12
|
|
1322
|
+
pop r13
|
|
1323
|
+
pop r14
|
|
1324
|
+
pop r15
|
|
1325
|
+
ret
|
|
1326
|
+
ALIGN 16
|
|
1327
|
+
final7blocks:
|
|
1328
|
+
mov rbx, qword ptr [rbp+90H]
|
|
1329
|
+
mov r15, qword ptr [rsp+2C0H]
|
|
1330
|
+
movzx r13d, byte ptr [rbp+78H]
|
|
1331
|
+
movzx r12d, byte ptr [rbp+88H]
|
|
1332
|
+
test rsi, 4H
|
|
1333
|
+
je final3blocks
|
|
1334
|
+
vbroadcasti128 ymm0, xmmword ptr [rcx]
|
|
1335
|
+
vbroadcasti128 ymm1, xmmword ptr [rcx+10H]
|
|
1336
|
+
vmovdqa ymm8, ymm0
|
|
1337
|
+
vmovdqa ymm9, ymm1
|
|
1338
|
+
vbroadcasti128 ymm12, xmmword ptr [rsp+220H]
|
|
1339
|
+
vbroadcasti128 ymm13, xmmword ptr [rsp+240H]
|
|
1340
|
+
vpunpckldq ymm14, ymm12, ymm13
|
|
1341
|
+
vpunpckhdq ymm15, ymm12, ymm13
|
|
1342
|
+
vpermq ymm14, ymm14, 50H
|
|
1343
|
+
vpermq ymm15, ymm15, 50H
|
|
1344
|
+
vbroadcasti128 ymm12, xmmword ptr [BLAKE3_BLOCK_LEN]
|
|
1345
|
+
vpblendd ymm14, ymm14, ymm12, 44H
|
|
1346
|
+
vpblendd ymm15, ymm15, ymm12, 44H
|
|
1347
|
+
vmovdqa ymmword ptr [rsp], ymm14
|
|
1348
|
+
vmovdqa ymmword ptr [rsp+20H], ymm15
|
|
1349
|
+
mov r8, qword ptr [rdi]
|
|
1350
|
+
mov r9, qword ptr [rdi+8H]
|
|
1351
|
+
mov r10, qword ptr [rdi+10H]
|
|
1352
|
+
mov r11, qword ptr [rdi+18H]
|
|
1353
|
+
movzx eax, byte ptr [rbp+80H]
|
|
1354
|
+
or eax, r13d
|
|
1355
|
+
xor edx, edx
|
|
1356
|
+
ALIGN 16
|
|
1357
|
+
innerloop4:
|
|
1358
|
+
mov r14d, eax
|
|
1359
|
+
or eax, r12d
|
|
1360
|
+
add rdx, 64
|
|
1361
|
+
cmp rdx, r15
|
|
1362
|
+
cmovne eax, r14d
|
|
1363
|
+
mov dword ptr [rsp+200H], eax
|
|
1364
|
+
vmovups ymm2, ymmword ptr [r8+rdx-40H]
|
|
1365
|
+
vinsertf128 ymm2, ymm2, xmmword ptr [r9+rdx-40H], 01H
|
|
1366
|
+
vmovups ymm3, ymmword ptr [r8+rdx-30H]
|
|
1367
|
+
vinsertf128 ymm3, ymm3, xmmword ptr [r9+rdx-30H], 01H
|
|
1368
|
+
vshufps ymm4, ymm2, ymm3, 136
|
|
1369
|
+
vshufps ymm5, ymm2, ymm3, 221
|
|
1370
|
+
vmovups ymm2, ymmword ptr [r8+rdx-20H]
|
|
1371
|
+
vinsertf128 ymm2, ymm2, xmmword ptr [r9+rdx-20H], 01H
|
|
1372
|
+
vmovups ymm3, ymmword ptr [r8+rdx-10H]
|
|
1373
|
+
vinsertf128 ymm3, ymm3, xmmword ptr [r9+rdx-10H], 01H
|
|
1374
|
+
vshufps ymm6, ymm2, ymm3, 136
|
|
1375
|
+
vshufps ymm7, ymm2, ymm3, 221
|
|
1376
|
+
vpshufd ymm6, ymm6, 93H
|
|
1377
|
+
vpshufd ymm7, ymm7, 93H
|
|
1378
|
+
vmovups ymm10, ymmword ptr [r10+rdx-40H]
|
|
1379
|
+
vinsertf128 ymm10, ymm10, xmmword ptr [r11+rdx-40H], 01H
|
|
1380
|
+
vmovups ymm11, ymmword ptr [r10+rdx-30H]
|
|
1381
|
+
vinsertf128 ymm11, ymm11, xmmword ptr [r11+rdx-30H], 01H
|
|
1382
|
+
vshufps ymm12, ymm10, ymm11, 136
|
|
1383
|
+
vshufps ymm13, ymm10, ymm11, 221
|
|
1384
|
+
vmovups ymm10, ymmword ptr [r10+rdx-20H]
|
|
1385
|
+
vinsertf128 ymm10, ymm10, xmmword ptr [r11+rdx-20H], 01H
|
|
1386
|
+
vmovups ymm11, ymmword ptr [r10+rdx-10H]
|
|
1387
|
+
vinsertf128 ymm11, ymm11, xmmword ptr [r11+rdx-10H], 01H
|
|
1388
|
+
vshufps ymm14, ymm10, ymm11, 136
|
|
1389
|
+
vshufps ymm15, ymm10, ymm11, 221
|
|
1390
|
+
vpshufd ymm14, ymm14, 93H
|
|
1391
|
+
vpshufd ymm15, ymm15, 93H
|
|
1392
|
+
vpbroadcastd ymm2, dword ptr [rsp+200H]
|
|
1393
|
+
vmovdqa ymm3, ymmword ptr [rsp]
|
|
1394
|
+
vmovdqa ymm11, ymmword ptr [rsp+20H]
|
|
1395
|
+
vpblendd ymm3, ymm3, ymm2, 88H
|
|
1396
|
+
vpblendd ymm11, ymm11, ymm2, 88H
|
|
1397
|
+
vbroadcasti128 ymm2, xmmword ptr [BLAKE3_IV]
|
|
1398
|
+
vmovdqa ymm10, ymm2
|
|
1399
|
+
mov al, 7
|
|
1400
|
+
roundloop4:
|
|
1401
|
+
vpaddd ymm0, ymm0, ymm4
|
|
1402
|
+
vpaddd ymm8, ymm8, ymm12
|
|
1403
|
+
vmovdqa ymmword ptr [rsp+40H], ymm4
|
|
1404
|
+
nop
|
|
1405
|
+
vmovdqa ymmword ptr [rsp+60H], ymm12
|
|
1406
|
+
nop
|
|
1407
|
+
vpaddd ymm0, ymm0, ymm1
|
|
1408
|
+
vpaddd ymm8, ymm8, ymm9
|
|
1409
|
+
vpxor ymm3, ymm3, ymm0
|
|
1410
|
+
vpxor ymm11, ymm11, ymm8
|
|
1411
|
+
vbroadcasti128 ymm4, xmmword ptr [ROT16]
|
|
1412
|
+
vpshufb ymm3, ymm3, ymm4
|
|
1413
|
+
vpshufb ymm11, ymm11, ymm4
|
|
1414
|
+
vpaddd ymm2, ymm2, ymm3
|
|
1415
|
+
vpaddd ymm10, ymm10, ymm11
|
|
1416
|
+
vpxor ymm1, ymm1, ymm2
|
|
1417
|
+
vpxor ymm9, ymm9, ymm10
|
|
1418
|
+
vpsrld ymm4, ymm1, 12
|
|
1419
|
+
vpslld ymm1, ymm1, 20
|
|
1420
|
+
vpor ymm1, ymm1, ymm4
|
|
1421
|
+
vpsrld ymm4, ymm9, 12
|
|
1422
|
+
vpslld ymm9, ymm9, 20
|
|
1423
|
+
vpor ymm9, ymm9, ymm4
|
|
1424
|
+
vpaddd ymm0, ymm0, ymm5
|
|
1425
|
+
vpaddd ymm8, ymm8, ymm13
|
|
1426
|
+
vpaddd ymm0, ymm0, ymm1
|
|
1427
|
+
vpaddd ymm8, ymm8, ymm9
|
|
1428
|
+
vmovdqa ymmword ptr [rsp+80H], ymm5
|
|
1429
|
+
vmovdqa ymmword ptr [rsp+0A0H], ymm13
|
|
1430
|
+
vpxor ymm3, ymm3, ymm0
|
|
1431
|
+
vpxor ymm11, ymm11, ymm8
|
|
1432
|
+
vbroadcasti128 ymm4, xmmword ptr [ROT8]
|
|
1433
|
+
vpshufb ymm3, ymm3, ymm4
|
|
1434
|
+
vpshufb ymm11, ymm11, ymm4
|
|
1435
|
+
vpaddd ymm2, ymm2, ymm3
|
|
1436
|
+
vpaddd ymm10, ymm10, ymm11
|
|
1437
|
+
vpxor ymm1, ymm1, ymm2
|
|
1438
|
+
vpxor ymm9, ymm9, ymm10
|
|
1439
|
+
vpsrld ymm4, ymm1, 7
|
|
1440
|
+
vpslld ymm1, ymm1, 25
|
|
1441
|
+
vpor ymm1, ymm1, ymm4
|
|
1442
|
+
vpsrld ymm4, ymm9, 7
|
|
1443
|
+
vpslld ymm9, ymm9, 25
|
|
1444
|
+
vpor ymm9, ymm9, ymm4
|
|
1445
|
+
vpshufd ymm0, ymm0, 93H
|
|
1446
|
+
vpshufd ymm8, ymm8, 93H
|
|
1447
|
+
vpshufd ymm3, ymm3, 4EH
|
|
1448
|
+
vpshufd ymm11, ymm11, 4EH
|
|
1449
|
+
vpshufd ymm2, ymm2, 39H
|
|
1450
|
+
vpshufd ymm10, ymm10, 39H
|
|
1451
|
+
vpaddd ymm0, ymm0, ymm6
|
|
1452
|
+
vpaddd ymm8, ymm8, ymm14
|
|
1453
|
+
vpaddd ymm0, ymm0, ymm1
|
|
1454
|
+
vpaddd ymm8, ymm8, ymm9
|
|
1455
|
+
vpxor ymm3, ymm3, ymm0
|
|
1456
|
+
vpxor ymm11, ymm11, ymm8
|
|
1457
|
+
vbroadcasti128 ymm4, xmmword ptr [ROT16]
|
|
1458
|
+
vpshufb ymm3, ymm3, ymm4
|
|
1459
|
+
vpshufb ymm11, ymm11, ymm4
|
|
1460
|
+
vpaddd ymm2, ymm2, ymm3
|
|
1461
|
+
vpaddd ymm10, ymm10, ymm11
|
|
1462
|
+
vpxor ymm1, ymm1, ymm2
|
|
1463
|
+
vpxor ymm9, ymm9, ymm10
|
|
1464
|
+
vpsrld ymm4, ymm1, 12
|
|
1465
|
+
vpslld ymm1, ymm1, 20
|
|
1466
|
+
vpor ymm1, ymm1, ymm4
|
|
1467
|
+
vpsrld ymm4, ymm9, 12
|
|
1468
|
+
vpslld ymm9, ymm9, 20
|
|
1469
|
+
vpor ymm9, ymm9, ymm4
|
|
1470
|
+
vpaddd ymm0, ymm0, ymm7
|
|
1471
|
+
vpaddd ymm8, ymm8, ymm15
|
|
1472
|
+
vpaddd ymm0, ymm0, ymm1
|
|
1473
|
+
vpaddd ymm8, ymm8, ymm9
|
|
1474
|
+
vpxor ymm3, ymm3, ymm0
|
|
1475
|
+
vpxor ymm11, ymm11, ymm8
|
|
1476
|
+
vbroadcasti128 ymm4, xmmword ptr [ROT8]
|
|
1477
|
+
vpshufb ymm3, ymm3, ymm4
|
|
1478
|
+
vpshufb ymm11, ymm11, ymm4
|
|
1479
|
+
vpaddd ymm2, ymm2, ymm3
|
|
1480
|
+
vpaddd ymm10, ymm10, ymm11
|
|
1481
|
+
vpxor ymm1, ymm1, ymm2
|
|
1482
|
+
vpxor ymm9, ymm9, ymm10
|
|
1483
|
+
vpsrld ymm4, ymm1, 7
|
|
1484
|
+
vpslld ymm1, ymm1, 25
|
|
1485
|
+
vpor ymm1, ymm1, ymm4
|
|
1486
|
+
vpsrld ymm4, ymm9, 7
|
|
1487
|
+
vpslld ymm9, ymm9, 25
|
|
1488
|
+
vpor ymm9, ymm9, ymm4
|
|
1489
|
+
vpshufd ymm0, ymm0, 39H
|
|
1490
|
+
vpshufd ymm8, ymm8, 39H
|
|
1491
|
+
vpshufd ymm3, ymm3, 4EH
|
|
1492
|
+
vpshufd ymm11, ymm11, 4EH
|
|
1493
|
+
vpshufd ymm2, ymm2, 93H
|
|
1494
|
+
vpshufd ymm10, ymm10, 93H
|
|
1495
|
+
dec al
|
|
1496
|
+
je endroundloop4
|
|
1497
|
+
vmovdqa ymm4, ymmword ptr [rsp+40H]
|
|
1498
|
+
vmovdqa ymm5, ymmword ptr [rsp+80H]
|
|
1499
|
+
vshufps ymm12, ymm4, ymm5, 214
|
|
1500
|
+
vpshufd ymm13, ymm4, 0FH
|
|
1501
|
+
vpshufd ymm4, ymm12, 39H
|
|
1502
|
+
vshufps ymm12, ymm6, ymm7, 250
|
|
1503
|
+
vpblendd ymm13, ymm13, ymm12, 0AAH
|
|
1504
|
+
vpunpcklqdq ymm12, ymm7, ymm5
|
|
1505
|
+
vpblendd ymm12, ymm12, ymm6, 88H
|
|
1506
|
+
vpshufd ymm12, ymm12, 78H
|
|
1507
|
+
vpunpckhdq ymm5, ymm5, ymm7
|
|
1508
|
+
vpunpckldq ymm6, ymm6, ymm5
|
|
1509
|
+
vpshufd ymm7, ymm6, 1EH
|
|
1510
|
+
vmovdqa ymmword ptr [rsp+40H], ymm13
|
|
1511
|
+
vmovdqa ymmword ptr [rsp+80H], ymm12
|
|
1512
|
+
vmovdqa ymm12, ymmword ptr [rsp+60H]
|
|
1513
|
+
vmovdqa ymm13, ymmword ptr [rsp+0A0H]
|
|
1514
|
+
vshufps ymm5, ymm12, ymm13, 214
|
|
1515
|
+
vpshufd ymm6, ymm12, 0FH
|
|
1516
|
+
vpshufd ymm12, ymm5, 39H
|
|
1517
|
+
vshufps ymm5, ymm14, ymm15, 250
|
|
1518
|
+
vpblendd ymm6, ymm6, ymm5, 0AAH
|
|
1519
|
+
vpunpcklqdq ymm5, ymm15, ymm13
|
|
1520
|
+
vpblendd ymm5, ymm5, ymm14, 88H
|
|
1521
|
+
vpshufd ymm5, ymm5, 78H
|
|
1522
|
+
vpunpckhdq ymm13, ymm13, ymm15
|
|
1523
|
+
vpunpckldq ymm14, ymm14, ymm13
|
|
1524
|
+
vpshufd ymm15, ymm14, 1EH
|
|
1525
|
+
vmovdqa ymm13, ymm6
|
|
1526
|
+
vmovdqa ymm14, ymm5
|
|
1527
|
+
vmovdqa ymm5, ymmword ptr [rsp+40H]
|
|
1528
|
+
vmovdqa ymm6, ymmword ptr [rsp+80H]
|
|
1529
|
+
jmp roundloop4
|
|
1530
|
+
endroundloop4:
|
|
1531
|
+
vpxor ymm0, ymm0, ymm2
|
|
1532
|
+
vpxor ymm1, ymm1, ymm3
|
|
1533
|
+
vpxor ymm8, ymm8, ymm10
|
|
1534
|
+
vpxor ymm9, ymm9, ymm11
|
|
1535
|
+
mov eax, r13d
|
|
1536
|
+
cmp rdx, r15
|
|
1537
|
+
jne innerloop4
|
|
1538
|
+
vmovdqu xmmword ptr [rbx], xmm0
|
|
1539
|
+
vmovdqu xmmword ptr [rbx+10H], xmm1
|
|
1540
|
+
vextracti128 xmmword ptr [rbx+20H], ymm0, 01H
|
|
1541
|
+
vextracti128 xmmword ptr [rbx+30H], ymm1, 01H
|
|
1542
|
+
vmovdqu xmmword ptr [rbx+40H], xmm8
|
|
1543
|
+
vmovdqu xmmword ptr [rbx+50H], xmm9
|
|
1544
|
+
vextracti128 xmmword ptr [rbx+60H], ymm8, 01H
|
|
1545
|
+
vextracti128 xmmword ptr [rbx+70H], ymm9, 01H
|
|
1546
|
+
vmovaps xmm8, xmmword ptr [rsp+260H]
|
|
1547
|
+
vmovaps xmm0, xmmword ptr [rsp+220H]
|
|
1548
|
+
vmovaps xmm1, xmmword ptr [rsp+230H]
|
|
1549
|
+
vmovaps xmm2, xmmword ptr [rsp+240H]
|
|
1550
|
+
vmovaps xmm3, xmmword ptr [rsp+250H]
|
|
1551
|
+
vblendvps xmm0, xmm0, xmm1, xmm8
|
|
1552
|
+
vblendvps xmm2, xmm2, xmm3, xmm8
|
|
1553
|
+
vmovaps xmmword ptr [rsp+220H], xmm0
|
|
1554
|
+
vmovaps xmmword ptr [rsp+240H], xmm2
|
|
1555
|
+
add rbx, 128
|
|
1556
|
+
add rdi, 32
|
|
1557
|
+
sub rsi, 4
|
|
1558
|
+
final3blocks:
|
|
1559
|
+
test rsi, 2H
|
|
1560
|
+
je final1blocks
|
|
1561
|
+
vbroadcasti128 ymm0, xmmword ptr [rcx]
|
|
1562
|
+
vbroadcasti128 ymm1, xmmword ptr [rcx+10H]
|
|
1563
|
+
vmovd xmm13, dword ptr [rsp+220H]
|
|
1564
|
+
vpinsrd xmm13, xmm13, dword ptr [rsp+240H], 1
|
|
1565
|
+
vpinsrd xmm13, xmm13, dword ptr [BLAKE3_BLOCK_LEN], 2
|
|
1566
|
+
vmovd xmm14, dword ptr [rsp+224H]
|
|
1567
|
+
vpinsrd xmm14, xmm14, dword ptr [rsp+244H], 1
|
|
1568
|
+
vpinsrd xmm14, xmm14, dword ptr [BLAKE3_BLOCK_LEN], 2
|
|
1569
|
+
vinserti128 ymm13, ymm13, xmm14, 01H
|
|
1570
|
+
vbroadcasti128 ymm14, xmmword ptr [ROT16]
|
|
1571
|
+
vbroadcasti128 ymm15, xmmword ptr [ROT8]
|
|
1572
|
+
mov r8, qword ptr [rdi]
|
|
1573
|
+
mov r9, qword ptr [rdi+8H]
|
|
1574
|
+
movzx eax, byte ptr [rbp+80H]
|
|
1575
|
+
or eax, r13d
|
|
1576
|
+
xor edx, edx
|
|
1577
|
+
ALIGN 16
|
|
1578
|
+
innerloop2:
|
|
1579
|
+
mov r14d, eax
|
|
1580
|
+
or eax, r12d
|
|
1581
|
+
add rdx, 64
|
|
1582
|
+
cmp rdx, r15
|
|
1583
|
+
cmovne eax, r14d
|
|
1584
|
+
mov dword ptr [rsp+200H], eax
|
|
1585
|
+
vbroadcasti128 ymm2, xmmword ptr [BLAKE3_IV]
|
|
1586
|
+
vpbroadcastd ymm8, dword ptr [rsp+200H]
|
|
1587
|
+
vpblendd ymm3, ymm13, ymm8, 88H
|
|
1588
|
+
vmovups ymm8, ymmword ptr [r8+rdx-40H]
|
|
1589
|
+
vinsertf128 ymm8, ymm8, xmmword ptr [r9+rdx-40H], 01H
|
|
1590
|
+
vmovups ymm9, ymmword ptr [r8+rdx-30H]
|
|
1591
|
+
vinsertf128 ymm9, ymm9, xmmword ptr [r9+rdx-30H], 01H
|
|
1592
|
+
vshufps ymm4, ymm8, ymm9, 136
|
|
1593
|
+
vshufps ymm5, ymm8, ymm9, 221
|
|
1594
|
+
vmovups ymm8, ymmword ptr [r8+rdx-20H]
|
|
1595
|
+
vinsertf128 ymm8, ymm8, xmmword ptr [r9+rdx-20H], 01H
|
|
1596
|
+
vmovups ymm9, ymmword ptr [r8+rdx-10H]
|
|
1597
|
+
vinsertf128 ymm9, ymm9, xmmword ptr [r9+rdx-10H], 01H
|
|
1598
|
+
vshufps ymm6, ymm8, ymm9, 136
|
|
1599
|
+
vshufps ymm7, ymm8, ymm9, 221
|
|
1600
|
+
vpshufd ymm6, ymm6, 93H
|
|
1601
|
+
vpshufd ymm7, ymm7, 93H
|
|
1602
|
+
mov al, 7
|
|
1603
|
+
roundloop2:
|
|
1604
|
+
vpaddd ymm0, ymm0, ymm4
|
|
1605
|
+
vpaddd ymm0, ymm0, ymm1
|
|
1606
|
+
vpxor ymm3, ymm3, ymm0
|
|
1607
|
+
vpshufb ymm3, ymm3, ymm14
|
|
1608
|
+
vpaddd ymm2, ymm2, ymm3
|
|
1609
|
+
vpxor ymm1, ymm1, ymm2
|
|
1610
|
+
vpsrld ymm8, ymm1, 12
|
|
1611
|
+
vpslld ymm1, ymm1, 20
|
|
1612
|
+
vpor ymm1, ymm1, ymm8
|
|
1613
|
+
vpaddd ymm0, ymm0, ymm5
|
|
1614
|
+
vpaddd ymm0, ymm0, ymm1
|
|
1615
|
+
vpxor ymm3, ymm3, ymm0
|
|
1616
|
+
vpshufb ymm3, ymm3, ymm15
|
|
1617
|
+
vpaddd ymm2, ymm2, ymm3
|
|
1618
|
+
vpxor ymm1, ymm1, ymm2
|
|
1619
|
+
vpsrld ymm8, ymm1, 7
|
|
1620
|
+
vpslld ymm1, ymm1, 25
|
|
1621
|
+
vpor ymm1, ymm1, ymm8
|
|
1622
|
+
vpshufd ymm0, ymm0, 93H
|
|
1623
|
+
vpshufd ymm3, ymm3, 4EH
|
|
1624
|
+
vpshufd ymm2, ymm2, 39H
|
|
1625
|
+
vpaddd ymm0, ymm0, ymm6
|
|
1626
|
+
vpaddd ymm0, ymm0, ymm1
|
|
1627
|
+
vpxor ymm3, ymm3, ymm0
|
|
1628
|
+
vpshufb ymm3, ymm3, ymm14
|
|
1629
|
+
vpaddd ymm2, ymm2, ymm3
|
|
1630
|
+
vpxor ymm1, ymm1, ymm2
|
|
1631
|
+
vpsrld ymm8, ymm1, 12
|
|
1632
|
+
vpslld ymm1, ymm1, 20
|
|
1633
|
+
vpor ymm1, ymm1, ymm8
|
|
1634
|
+
vpaddd ymm0, ymm0, ymm7
|
|
1635
|
+
vpaddd ymm0, ymm0, ymm1
|
|
1636
|
+
vpxor ymm3, ymm3, ymm0
|
|
1637
|
+
vpshufb ymm3, ymm3, ymm15
|
|
1638
|
+
vpaddd ymm2, ymm2, ymm3
|
|
1639
|
+
vpxor ymm1, ymm1, ymm2
|
|
1640
|
+
vpsrld ymm8, ymm1, 7
|
|
1641
|
+
vpslld ymm1, ymm1, 25
|
|
1642
|
+
vpor ymm1, ymm1, ymm8
|
|
1643
|
+
vpshufd ymm0, ymm0, 39H
|
|
1644
|
+
vpshufd ymm3, ymm3, 4EH
|
|
1645
|
+
vpshufd ymm2, ymm2, 93H
|
|
1646
|
+
dec al
|
|
1647
|
+
jz endroundloop2
|
|
1648
|
+
vshufps ymm8, ymm4, ymm5, 214
|
|
1649
|
+
vpshufd ymm9, ymm4, 0FH
|
|
1650
|
+
vpshufd ymm4, ymm8, 39H
|
|
1651
|
+
vshufps ymm8, ymm6, ymm7, 250
|
|
1652
|
+
vpblendd ymm9, ymm9, ymm8, 0AAH
|
|
1653
|
+
vpunpcklqdq ymm8, ymm7, ymm5
|
|
1654
|
+
vpblendd ymm8, ymm8, ymm6, 88H
|
|
1655
|
+
vpshufd ymm8, ymm8, 78H
|
|
1656
|
+
vpunpckhdq ymm5, ymm5, ymm7
|
|
1657
|
+
vpunpckldq ymm6, ymm6, ymm5
|
|
1658
|
+
vpshufd ymm7, ymm6, 1EH
|
|
1659
|
+
vmovdqa ymm5, ymm9
|
|
1660
|
+
vmovdqa ymm6, ymm8
|
|
1661
|
+
jmp roundloop2
|
|
1662
|
+
endroundloop2:
|
|
1663
|
+
vpxor ymm0, ymm0, ymm2
|
|
1664
|
+
vpxor ymm1, ymm1, ymm3
|
|
1665
|
+
mov eax, r13d
|
|
1666
|
+
cmp rdx, r15
|
|
1667
|
+
jne innerloop2
|
|
1668
|
+
vmovdqu xmmword ptr [rbx], xmm0
|
|
1669
|
+
vmovdqu xmmword ptr [rbx+10H], xmm1
|
|
1670
|
+
vextracti128 xmmword ptr [rbx+20H], ymm0, 01H
|
|
1671
|
+
vextracti128 xmmword ptr [rbx+30H], ymm1, 01H
|
|
1672
|
+
vmovaps ymm8, ymmword ptr [rsp+260H]
|
|
1673
|
+
vmovaps ymm0, ymmword ptr [rsp+220H]
|
|
1674
|
+
vmovups ymm1, ymmword ptr [rsp+228H]
|
|
1675
|
+
vmovaps ymm2, ymmword ptr [rsp+240H]
|
|
1676
|
+
vmovups ymm3, ymmword ptr [rsp+248H]
|
|
1677
|
+
vblendvps ymm0, ymm0, ymm1, ymm8
|
|
1678
|
+
vblendvps ymm2, ymm2, ymm3, ymm8
|
|
1679
|
+
vmovaps ymmword ptr [rsp+220H], ymm0
|
|
1680
|
+
vmovaps ymmword ptr [rsp+240H], ymm2
|
|
1681
|
+
add rbx, 64
|
|
1682
|
+
add rdi, 16
|
|
1683
|
+
sub rsi, 2
|
|
1684
|
+
final1blocks:
|
|
1685
|
+
test rsi, 1H
|
|
1686
|
+
je unwind
|
|
1687
|
+
vmovdqu xmm0, xmmword ptr [rcx]
|
|
1688
|
+
vmovdqu xmm1, xmmword ptr [rcx+10H]
|
|
1689
|
+
vmovd xmm3, dword ptr [rsp+220H]
|
|
1690
|
+
vpinsrd xmm3, xmm3, dword ptr [rsp+240H], 1
|
|
1691
|
+
vpinsrd xmm13, xmm3, dword ptr [BLAKE3_BLOCK_LEN], 2
|
|
1692
|
+
vmovdqa xmm14, xmmword ptr [ROT16]
|
|
1693
|
+
vmovdqa xmm15, xmmword ptr [ROT8]
|
|
1694
|
+
mov r8, qword ptr [rdi]
|
|
1695
|
+
movzx eax, byte ptr [rbp+80H]
|
|
1696
|
+
or eax, r13d
|
|
1697
|
+
xor edx, edx
|
|
1698
|
+
ALIGN 16
|
|
1699
|
+
innerloop1:
|
|
1700
|
+
mov r14d, eax
|
|
1701
|
+
or eax, r12d
|
|
1702
|
+
add rdx, 64
|
|
1703
|
+
cmp rdx, r15
|
|
1704
|
+
cmovne eax, r14d
|
|
1705
|
+
vmovdqa xmm2, xmmword ptr [BLAKE3_IV]
|
|
1706
|
+
vmovdqa xmm3, xmm13
|
|
1707
|
+
vpinsrd xmm3, xmm3, eax, 3
|
|
1708
|
+
vmovups xmm8, xmmword ptr [r8+rdx-40H]
|
|
1709
|
+
vmovups xmm9, xmmword ptr [r8+rdx-30H]
|
|
1710
|
+
vshufps xmm4, xmm8, xmm9, 136
|
|
1711
|
+
vshufps xmm5, xmm8, xmm9, 221
|
|
1712
|
+
vmovups xmm8, xmmword ptr [r8+rdx-20H]
|
|
1713
|
+
vmovups xmm9, xmmword ptr [r8+rdx-10H]
|
|
1714
|
+
vshufps xmm6, xmm8, xmm9, 136
|
|
1715
|
+
vshufps xmm7, xmm8, xmm9, 221
|
|
1716
|
+
vpshufd xmm6, xmm6, 93H
|
|
1717
|
+
vpshufd xmm7, xmm7, 93H
|
|
1718
|
+
mov al, 7
|
|
1719
|
+
roundloop1:
|
|
1720
|
+
vpaddd xmm0, xmm0, xmm4
|
|
1721
|
+
vpaddd xmm0, xmm0, xmm1
|
|
1722
|
+
vpxor xmm3, xmm3, xmm0
|
|
1723
|
+
vpshufb xmm3, xmm3, xmm14
|
|
1724
|
+
vpaddd xmm2, xmm2, xmm3
|
|
1725
|
+
vpxor xmm1, xmm1, xmm2
|
|
1726
|
+
vpsrld xmm8, xmm1, 12
|
|
1727
|
+
vpslld xmm1, xmm1, 20
|
|
1728
|
+
vpor xmm1, xmm1, xmm8
|
|
1729
|
+
vpaddd xmm0, xmm0, xmm5
|
|
1730
|
+
vpaddd xmm0, xmm0, xmm1
|
|
1731
|
+
vpxor xmm3, xmm3, xmm0
|
|
1732
|
+
vpshufb xmm3, xmm3, xmm15
|
|
1733
|
+
vpaddd xmm2, xmm2, xmm3
|
|
1734
|
+
vpxor xmm1, xmm1, xmm2
|
|
1735
|
+
vpsrld xmm8, xmm1, 7
|
|
1736
|
+
vpslld xmm1, xmm1, 25
|
|
1737
|
+
vpor xmm1, xmm1, xmm8
|
|
1738
|
+
vpshufd xmm0, xmm0, 93H
|
|
1739
|
+
vpshufd xmm3, xmm3, 4EH
|
|
1740
|
+
vpshufd xmm2, xmm2, 39H
|
|
1741
|
+
vpaddd xmm0, xmm0, xmm6
|
|
1742
|
+
vpaddd xmm0, xmm0, xmm1
|
|
1743
|
+
vpxor xmm3, xmm3, xmm0
|
|
1744
|
+
vpshufb xmm3, xmm3, xmm14
|
|
1745
|
+
vpaddd xmm2, xmm2, xmm3
|
|
1746
|
+
vpxor xmm1, xmm1, xmm2
|
|
1747
|
+
vpsrld xmm8, xmm1, 12
|
|
1748
|
+
vpslld xmm1, xmm1, 20
|
|
1749
|
+
vpor xmm1, xmm1, xmm8
|
|
1750
|
+
vpaddd xmm0, xmm0, xmm7
|
|
1751
|
+
vpaddd xmm0, xmm0, xmm1
|
|
1752
|
+
vpxor xmm3, xmm3, xmm0
|
|
1753
|
+
vpshufb xmm3, xmm3, xmm15
|
|
1754
|
+
vpaddd xmm2, xmm2, xmm3
|
|
1755
|
+
vpxor xmm1, xmm1, xmm2
|
|
1756
|
+
vpsrld xmm8, xmm1, 7
|
|
1757
|
+
vpslld xmm1, xmm1, 25
|
|
1758
|
+
vpor xmm1, xmm1, xmm8
|
|
1759
|
+
vpshufd xmm0, xmm0, 39H
|
|
1760
|
+
vpshufd xmm3, xmm3, 4EH
|
|
1761
|
+
vpshufd xmm2, xmm2, 93H
|
|
1762
|
+
dec al
|
|
1763
|
+
jz endroundloop1
|
|
1764
|
+
vshufps xmm8, xmm4, xmm5, 214
|
|
1765
|
+
vpshufd xmm9, xmm4, 0FH
|
|
1766
|
+
vpshufd xmm4, xmm8, 39H
|
|
1767
|
+
vshufps xmm8, xmm6, xmm7, 250
|
|
1768
|
+
vpblendd xmm9, xmm9, xmm8, 0AAH
|
|
1769
|
+
vpunpcklqdq xmm8, xmm7, xmm5
|
|
1770
|
+
vpblendd xmm8, xmm8, xmm6, 88H
|
|
1771
|
+
vpshufd xmm8, xmm8, 78H
|
|
1772
|
+
vpunpckhdq xmm5, xmm5, xmm7
|
|
1773
|
+
vpunpckldq xmm6, xmm6, xmm5
|
|
1774
|
+
vpshufd xmm7, xmm6, 1EH
|
|
1775
|
+
vmovdqa xmm5, xmm9
|
|
1776
|
+
vmovdqa xmm6, xmm8
|
|
1777
|
+
jmp roundloop1
|
|
1778
|
+
endroundloop1:
|
|
1779
|
+
vpxor xmm0, xmm0, xmm2
|
|
1780
|
+
vpxor xmm1, xmm1, xmm3
|
|
1781
|
+
mov eax, r13d
|
|
1782
|
+
cmp rdx, r15
|
|
1783
|
+
jne innerloop1
|
|
1784
|
+
vmovdqu xmmword ptr [rbx], xmm0
|
|
1785
|
+
vmovdqu xmmword ptr [rbx+10H], xmm1
|
|
1786
|
+
jmp unwind
|
|
1787
|
+
|
|
1788
|
+
_blake3_hash_many_avx2 ENDP
|
|
1789
|
+
blake3_hash_many_avx2 ENDP
|
|
1790
|
+
_TEXT ENDS
|
|
1791
|
+
|
|
1792
|
+
_RDATA SEGMENT READONLY PAGE ALIAS(".rdata") 'CONST'
|
|
1793
|
+
ALIGN 64
|
|
1794
|
+
ADD0:
|
|
1795
|
+
dd 0, 1, 2, 3, 4, 5, 6, 7
|
|
1796
|
+
|
|
1797
|
+
ADD1:
|
|
1798
|
+
dd 8 dup (8)
|
|
1799
|
+
|
|
1800
|
+
BLAKE3_IV_0:
|
|
1801
|
+
dd 8 dup (6A09E667H)
|
|
1802
|
+
|
|
1803
|
+
BLAKE3_IV_1:
|
|
1804
|
+
dd 8 dup (0BB67AE85H)
|
|
1805
|
+
|
|
1806
|
+
BLAKE3_IV_2:
|
|
1807
|
+
dd 8 dup (3C6EF372H)
|
|
1808
|
+
|
|
1809
|
+
BLAKE3_IV_3:
|
|
1810
|
+
dd 8 dup (0A54FF53AH)
|
|
1811
|
+
|
|
1812
|
+
BLAKE3_BLOCK_LEN:
|
|
1813
|
+
dd 8 dup (64)
|
|
1814
|
+
|
|
1815
|
+
ROT16:
|
|
1816
|
+
db 2, 3, 0, 1, 6, 7, 4, 5, 10, 11, 8, 9, 14, 15, 12, 13
|
|
1817
|
+
|
|
1818
|
+
ROT8:
|
|
1819
|
+
db 1, 2, 3, 0, 5, 6, 7, 4, 9, 10, 11, 8, 13, 14, 15, 12
|
|
1820
|
+
|
|
1821
|
+
CMP_MSB_MASK:
|
|
1822
|
+
dd 8 dup(80000000H)
|
|
1823
|
+
|
|
1824
|
+
BLAKE3_IV:
|
|
1825
|
+
dd 6A09E667H, 0BB67AE85H, 3C6EF372H, 0A54FF53AH
|
|
1826
|
+
|
|
1827
|
+
_RDATA ENDS
|
|
1828
|
+
END
|