1 |
|
/* $MidnightBSD$ */ |
2 |
< |
# $FreeBSD: stable/10/secure/lib/libcrypto/amd64/aesni-sha1-x86_64.S 299966 2016-05-16 19:30:27Z jkim $ |
3 |
< |
# Do not modify. This file is auto-generated from aesni-sha1-x86_64.pl. |
2 |
> |
/* $FreeBSD: stable/11/secure/lib/libcrypto/amd64/aesni-sha1-x86_64.S 325335 2017-11-02 18:22:53Z jkim $ */ |
3 |
> |
/* Do not modify. This file is auto-generated from aesni-sha1-x86_64.pl. */ |
4 |
|
.text |
5 |
|
|
6 |
|
|
7 |
|
.globl aesni_cbc_sha1_enc |
8 |
|
.type aesni_cbc_sha1_enc,@function |
9 |
< |
.align 16 |
9 |
> |
.align 32 |
10 |
|
aesni_cbc_sha1_enc: |
11 |
|
|
12 |
|
movl OPENSSL_ia32cap_P+0(%rip),%r10d |
13 |
< |
movl OPENSSL_ia32cap_P+4(%rip),%r11d |
13 |
> |
movq OPENSSL_ia32cap_P+4(%rip),%r11 |
14 |
> |
btq $61,%r11 |
15 |
> |
jc aesni_cbc_sha1_enc_shaext |
16 |
|
andl $268435456,%r11d |
17 |
|
andl $1073741824,%r10d |
18 |
|
orl %r11d,%r10d |
22 |
|
.byte 0xf3,0xc3 |
23 |
|
.size aesni_cbc_sha1_enc,.-aesni_cbc_sha1_enc |
24 |
|
.type aesni_cbc_sha1_enc_ssse3,@function |
25 |
< |
.align 16 |
25 |
> |
.align 32 |
26 |
|
aesni_cbc_sha1_enc_ssse3: |
27 |
|
movq 8(%rsp),%r10 |
28 |
|
|
39 |
|
movq %rdi,%r12 |
40 |
|
movq %rsi,%r13 |
41 |
|
movq %rdx,%r14 |
42 |
< |
movq %rcx,%r15 |
43 |
< |
movdqu (%r8),%xmm11 |
42 |
> |
leaq 112(%rcx),%r15 |
43 |
> |
movdqu (%r8),%xmm2 |
44 |
|
movq %r8,88(%rsp) |
45 |
|
shlq $6,%r14 |
46 |
|
subq %r12,%r13 |
47 |
< |
movl 240(%r15),%r8d |
47 |
> |
movl 240-112(%r15),%r8d |
48 |
|
addq %r10,%r14 |
49 |
|
|
50 |
|
leaq K_XX_XX(%rip),%r11 |
54 |
|
movl 12(%r9),%edx |
55 |
|
movl %ebx,%esi |
56 |
|
movl 16(%r9),%ebp |
57 |
+ |
movl %ecx,%edi |
58 |
+ |
xorl %edx,%edi |
59 |
+ |
andl %edi,%esi |
60 |
|
|
61 |
< |
movdqa 64(%r11),%xmm6 |
62 |
< |
movdqa 0(%r11),%xmm9 |
63 |
< |
movdqu 0(%r10),%xmm0 |
64 |
< |
movdqu 16(%r10),%xmm1 |
65 |
< |
movdqu 32(%r10),%xmm2 |
66 |
< |
movdqu 48(%r10),%xmm3 |
67 |
< |
.byte 102,15,56,0,198 |
61 |
> |
movdqa 64(%r11),%xmm3 |
62 |
> |
movdqa 0(%r11),%xmm13 |
63 |
> |
movdqu 0(%r10),%xmm4 |
64 |
> |
movdqu 16(%r10),%xmm5 |
65 |
> |
movdqu 32(%r10),%xmm6 |
66 |
> |
movdqu 48(%r10),%xmm7 |
67 |
> |
.byte 102,15,56,0,227 |
68 |
> |
.byte 102,15,56,0,235 |
69 |
> |
.byte 102,15,56,0,243 |
70 |
|
addq $64,%r10 |
71 |
< |
.byte 102,15,56,0,206 |
72 |
< |
.byte 102,15,56,0,214 |
73 |
< |
.byte 102,15,56,0,222 |
74 |
< |
paddd %xmm9,%xmm0 |
75 |
< |
paddd %xmm9,%xmm1 |
76 |
< |
paddd %xmm9,%xmm2 |
77 |
< |
movdqa %xmm0,0(%rsp) |
78 |
< |
psubd %xmm9,%xmm0 |
79 |
< |
movdqa %xmm1,16(%rsp) |
80 |
< |
psubd %xmm9,%xmm1 |
81 |
< |
movdqa %xmm2,32(%rsp) |
82 |
< |
psubd %xmm9,%xmm2 |
76 |
< |
movups (%r15),%xmm13 |
77 |
< |
movups 16(%r15),%xmm14 |
71 |
> |
paddd %xmm13,%xmm4 |
72 |
> |
.byte 102,15,56,0,251 |
73 |
> |
paddd %xmm13,%xmm5 |
74 |
> |
paddd %xmm13,%xmm6 |
75 |
> |
movdqa %xmm4,0(%rsp) |
76 |
> |
psubd %xmm13,%xmm4 |
77 |
> |
movdqa %xmm5,16(%rsp) |
78 |
> |
psubd %xmm13,%xmm5 |
79 |
> |
movdqa %xmm6,32(%rsp) |
80 |
> |
psubd %xmm13,%xmm6 |
81 |
> |
movups -112(%r15),%xmm15 |
82 |
> |
movups 16-112(%r15),%xmm0 |
83 |
|
jmp .Loop_ssse3 |
84 |
< |
.align 16 |
84 |
> |
.align 32 |
85 |
|
.Loop_ssse3: |
86 |
< |
movdqa %xmm1,%xmm4 |
87 |
< |
addl 0(%rsp),%ebp |
88 |
< |
movups 0(%r12),%xmm12 |
89 |
< |
xorps %xmm13,%xmm12 |
90 |
< |
xorps %xmm12,%xmm11 |
91 |
< |
.byte 102,69,15,56,220,222 |
92 |
< |
movups 32(%r15),%xmm15 |
93 |
< |
xorl %edx,%ecx |
94 |
< |
movdqa %xmm3,%xmm8 |
95 |
< |
.byte 102,15,58,15,224,8 |
86 |
> |
rorl $2,%ebx |
87 |
> |
movups 0(%r12),%xmm14 |
88 |
> |
xorps %xmm15,%xmm14 |
89 |
> |
xorps %xmm14,%xmm2 |
90 |
> |
movups -80(%r15),%xmm1 |
91 |
> |
.byte 102,15,56,220,208 |
92 |
> |
pshufd $238,%xmm4,%xmm8 |
93 |
> |
xorl %edx,%esi |
94 |
> |
movdqa %xmm7,%xmm12 |
95 |
> |
paddd %xmm7,%xmm13 |
96 |
|
movl %eax,%edi |
97 |
+ |
addl 0(%rsp),%ebp |
98 |
+ |
punpcklqdq %xmm5,%xmm8 |
99 |
+ |
xorl %ecx,%ebx |
100 |
|
roll $5,%eax |
93 |
– |
paddd %xmm3,%xmm9 |
94 |
– |
andl %ecx,%esi |
95 |
– |
xorl %edx,%ecx |
96 |
– |
psrldq $4,%xmm8 |
97 |
– |
xorl %edx,%esi |
98 |
– |
addl %eax,%ebp |
99 |
– |
pxor %xmm0,%xmm4 |
100 |
– |
rorl $2,%ebx |
101 |
|
addl %esi,%ebp |
102 |
< |
pxor %xmm2,%xmm8 |
103 |
< |
addl 4(%rsp),%edx |
104 |
< |
xorl %ecx,%ebx |
105 |
< |
movl %ebp,%esi |
106 |
< |
roll $5,%ebp |
107 |
< |
pxor %xmm8,%xmm4 |
102 |
> |
psrldq $4,%xmm12 |
103 |
|
andl %ebx,%edi |
104 |
|
xorl %ecx,%ebx |
105 |
< |
movdqa %xmm9,48(%rsp) |
106 |
< |
xorl %ecx,%edi |
112 |
< |
.byte 102,69,15,56,220,223 |
113 |
< |
movups 48(%r15),%xmm14 |
114 |
< |
addl %ebp,%edx |
115 |
< |
movdqa %xmm4,%xmm10 |
116 |
< |
movdqa %xmm4,%xmm8 |
105 |
> |
pxor %xmm4,%xmm8 |
106 |
> |
addl %eax,%ebp |
107 |
|
rorl $7,%eax |
108 |
< |
addl %edi,%edx |
109 |
< |
addl 8(%rsp),%ecx |
108 |
> |
pxor %xmm6,%xmm12 |
109 |
> |
xorl %ecx,%edi |
110 |
> |
movl %ebp,%esi |
111 |
> |
addl 4(%rsp),%edx |
112 |
> |
pxor %xmm12,%xmm8 |
113 |
|
xorl %ebx,%eax |
114 |
< |
pslldq $12,%xmm10 |
115 |
< |
paddd %xmm4,%xmm4 |
116 |
< |
movl %edx,%edi |
117 |
< |
roll $5,%edx |
114 |
> |
roll $5,%ebp |
115 |
> |
movdqa %xmm13,48(%rsp) |
116 |
> |
addl %edi,%edx |
117 |
> |
movups -64(%r15),%xmm0 |
118 |
> |
.byte 102,15,56,220,209 |
119 |
|
andl %eax,%esi |
120 |
+ |
movdqa %xmm8,%xmm3 |
121 |
|
xorl %ebx,%eax |
122 |
< |
psrld $31,%xmm8 |
128 |
< |
xorl %ebx,%esi |
129 |
< |
addl %edx,%ecx |
130 |
< |
movdqa %xmm10,%xmm9 |
122 |
> |
addl %ebp,%edx |
123 |
|
rorl $7,%ebp |
124 |
< |
addl %esi,%ecx |
125 |
< |
psrld $30,%xmm10 |
126 |
< |
por %xmm8,%xmm4 |
127 |
< |
addl 12(%rsp),%ebx |
124 |
> |
movdqa %xmm8,%xmm12 |
125 |
> |
xorl %ebx,%esi |
126 |
> |
pslldq $12,%xmm3 |
127 |
> |
paddd %xmm8,%xmm8 |
128 |
> |
movl %edx,%edi |
129 |
> |
addl 8(%rsp),%ecx |
130 |
> |
psrld $31,%xmm12 |
131 |
|
xorl %eax,%ebp |
132 |
< |
movl %ecx,%esi |
133 |
< |
roll $5,%ecx |
134 |
< |
.byte 102,69,15,56,220,222 |
140 |
< |
movups 64(%r15),%xmm15 |
141 |
< |
pslld $2,%xmm9 |
142 |
< |
pxor %xmm10,%xmm4 |
132 |
> |
roll $5,%edx |
133 |
> |
addl %esi,%ecx |
134 |
> |
movdqa %xmm3,%xmm13 |
135 |
|
andl %ebp,%edi |
136 |
|
xorl %eax,%ebp |
137 |
< |
movdqa 0(%r11),%xmm10 |
138 |
< |
xorl %eax,%edi |
147 |
< |
addl %ecx,%ebx |
148 |
< |
pxor %xmm9,%xmm4 |
137 |
> |
psrld $30,%xmm3 |
138 |
> |
addl %edx,%ecx |
139 |
|
rorl $7,%edx |
140 |
< |
addl %edi,%ebx |
141 |
< |
movdqa %xmm2,%xmm5 |
142 |
< |
addl 16(%rsp),%eax |
140 |
> |
por %xmm12,%xmm8 |
141 |
> |
xorl %eax,%edi |
142 |
> |
movl %ecx,%esi |
143 |
> |
addl 12(%rsp),%ebx |
144 |
> |
movups -48(%r15),%xmm1 |
145 |
> |
.byte 102,15,56,220,208 |
146 |
> |
pslld $2,%xmm13 |
147 |
> |
pxor %xmm3,%xmm8 |
148 |
|
xorl %ebp,%edx |
149 |
< |
movdqa %xmm4,%xmm9 |
150 |
< |
.byte 102,15,58,15,233,8 |
151 |
< |
movl %ebx,%edi |
157 |
< |
roll $5,%ebx |
158 |
< |
paddd %xmm4,%xmm10 |
149 |
> |
movdqa 0(%r11),%xmm3 |
150 |
> |
roll $5,%ecx |
151 |
> |
addl %edi,%ebx |
152 |
|
andl %edx,%esi |
153 |
+ |
pxor %xmm13,%xmm8 |
154 |
|
xorl %ebp,%edx |
155 |
< |
psrldq $4,%xmm9 |
162 |
< |
xorl %ebp,%esi |
163 |
< |
addl %ebx,%eax |
164 |
< |
pxor %xmm1,%xmm5 |
155 |
> |
addl %ecx,%ebx |
156 |
|
rorl $7,%ecx |
157 |
< |
addl %esi,%eax |
158 |
< |
pxor %xmm3,%xmm9 |
159 |
< |
addl 20(%rsp),%ebp |
160 |
< |
.byte 102,69,15,56,220,223 |
161 |
< |
movups 80(%r15),%xmm14 |
157 |
> |
pshufd $238,%xmm5,%xmm9 |
158 |
> |
xorl %ebp,%esi |
159 |
> |
movdqa %xmm8,%xmm13 |
160 |
> |
paddd %xmm8,%xmm3 |
161 |
> |
movl %ebx,%edi |
162 |
> |
addl 16(%rsp),%eax |
163 |
> |
punpcklqdq %xmm6,%xmm9 |
164 |
|
xorl %edx,%ecx |
165 |
< |
movl %eax,%esi |
166 |
< |
roll $5,%eax |
167 |
< |
pxor %xmm9,%xmm5 |
165 |
> |
roll $5,%ebx |
166 |
> |
addl %esi,%eax |
167 |
> |
psrldq $4,%xmm13 |
168 |
|
andl %ecx,%edi |
169 |
|
xorl %edx,%ecx |
170 |
< |
movdqa %xmm10,0(%rsp) |
171 |
< |
xorl %edx,%edi |
179 |
< |
addl %eax,%ebp |
180 |
< |
movdqa %xmm5,%xmm8 |
181 |
< |
movdqa %xmm5,%xmm9 |
170 |
> |
pxor %xmm5,%xmm9 |
171 |
> |
addl %ebx,%eax |
172 |
|
rorl $7,%ebx |
173 |
< |
addl %edi,%ebp |
174 |
< |
addl 24(%rsp),%edx |
173 |
> |
movups -32(%r15),%xmm0 |
174 |
> |
.byte 102,15,56,220,209 |
175 |
> |
pxor %xmm7,%xmm13 |
176 |
> |
xorl %edx,%edi |
177 |
> |
movl %eax,%esi |
178 |
> |
addl 20(%rsp),%ebp |
179 |
> |
pxor %xmm13,%xmm9 |
180 |
|
xorl %ecx,%ebx |
181 |
< |
pslldq $12,%xmm8 |
182 |
< |
paddd %xmm5,%xmm5 |
183 |
< |
movl %ebp,%edi |
189 |
< |
roll $5,%ebp |
181 |
> |
roll $5,%eax |
182 |
> |
movdqa %xmm3,0(%rsp) |
183 |
> |
addl %edi,%ebp |
184 |
|
andl %ebx,%esi |
185 |
+ |
movdqa %xmm9,%xmm12 |
186 |
|
xorl %ecx,%ebx |
187 |
< |
psrld $31,%xmm9 |
193 |
< |
xorl %ecx,%esi |
194 |
< |
.byte 102,69,15,56,220,222 |
195 |
< |
movups 96(%r15),%xmm15 |
196 |
< |
addl %ebp,%edx |
197 |
< |
movdqa %xmm8,%xmm10 |
187 |
> |
addl %eax,%ebp |
188 |
|
rorl $7,%eax |
189 |
< |
addl %esi,%edx |
190 |
< |
psrld $30,%xmm8 |
191 |
< |
por %xmm9,%xmm5 |
192 |
< |
addl 28(%rsp),%ecx |
189 |
> |
movdqa %xmm9,%xmm13 |
190 |
> |
xorl %ecx,%esi |
191 |
> |
pslldq $12,%xmm12 |
192 |
> |
paddd %xmm9,%xmm9 |
193 |
> |
movl %ebp,%edi |
194 |
> |
addl 24(%rsp),%edx |
195 |
> |
psrld $31,%xmm13 |
196 |
|
xorl %ebx,%eax |
197 |
< |
movl %edx,%esi |
198 |
< |
roll $5,%edx |
199 |
< |
pslld $2,%xmm10 |
200 |
< |
pxor %xmm8,%xmm5 |
197 |
> |
roll $5,%ebp |
198 |
> |
addl %esi,%edx |
199 |
> |
movups -16(%r15),%xmm1 |
200 |
> |
.byte 102,15,56,220,208 |
201 |
> |
movdqa %xmm12,%xmm3 |
202 |
|
andl %eax,%edi |
203 |
|
xorl %ebx,%eax |
204 |
< |
movdqa 16(%r11),%xmm8 |
205 |
< |
xorl %ebx,%edi |
212 |
< |
addl %edx,%ecx |
213 |
< |
pxor %xmm10,%xmm5 |
204 |
> |
psrld $30,%xmm12 |
205 |
> |
addl %ebp,%edx |
206 |
|
rorl $7,%ebp |
207 |
< |
addl %edi,%ecx |
208 |
< |
movdqa %xmm3,%xmm6 |
209 |
< |
addl 32(%rsp),%ebx |
207 |
> |
por %xmm13,%xmm9 |
208 |
> |
xorl %ebx,%edi |
209 |
> |
movl %edx,%esi |
210 |
> |
addl 28(%rsp),%ecx |
211 |
> |
pslld $2,%xmm3 |
212 |
> |
pxor %xmm12,%xmm9 |
213 |
|
xorl %eax,%ebp |
214 |
< |
movdqa %xmm5,%xmm10 |
215 |
< |
.byte 102,15,58,15,242,8 |
216 |
< |
movl %ecx,%edi |
222 |
< |
roll $5,%ecx |
223 |
< |
.byte 102,69,15,56,220,223 |
224 |
< |
movups 112(%r15),%xmm14 |
225 |
< |
paddd %xmm5,%xmm8 |
214 |
> |
movdqa 16(%r11),%xmm12 |
215 |
> |
roll $5,%edx |
216 |
> |
addl %edi,%ecx |
217 |
|
andl %ebp,%esi |
218 |
+ |
pxor %xmm3,%xmm9 |
219 |
|
xorl %eax,%ebp |
220 |
< |
psrldq $4,%xmm10 |
229 |
< |
xorl %eax,%esi |
230 |
< |
addl %ecx,%ebx |
231 |
< |
pxor %xmm2,%xmm6 |
220 |
> |
addl %edx,%ecx |
221 |
|
rorl $7,%edx |
222 |
< |
addl %esi,%ebx |
223 |
< |
pxor %xmm4,%xmm10 |
224 |
< |
addl 36(%rsp),%eax |
222 |
> |
pshufd $238,%xmm6,%xmm10 |
223 |
> |
xorl %eax,%esi |
224 |
> |
movdqa %xmm9,%xmm3 |
225 |
> |
paddd %xmm9,%xmm12 |
226 |
> |
movl %ecx,%edi |
227 |
> |
addl 32(%rsp),%ebx |
228 |
> |
movups 0(%r15),%xmm0 |
229 |
> |
.byte 102,15,56,220,209 |
230 |
> |
punpcklqdq %xmm7,%xmm10 |
231 |
|
xorl %ebp,%edx |
232 |
< |
movl %ebx,%esi |
233 |
< |
roll $5,%ebx |
234 |
< |
pxor %xmm10,%xmm6 |
232 |
> |
roll $5,%ecx |
233 |
> |
addl %esi,%ebx |
234 |
> |
psrldq $4,%xmm3 |
235 |
|
andl %edx,%edi |
236 |
|
xorl %ebp,%edx |
237 |
< |
movdqa %xmm8,16(%rsp) |
238 |
< |
xorl %ebp,%edi |
244 |
< |
addl %ebx,%eax |
245 |
< |
movdqa %xmm6,%xmm9 |
246 |
< |
movdqa %xmm6,%xmm10 |
237 |
> |
pxor %xmm6,%xmm10 |
238 |
> |
addl %ecx,%ebx |
239 |
|
rorl $7,%ecx |
240 |
< |
addl %edi,%eax |
241 |
< |
addl 40(%rsp),%ebp |
242 |
< |
.byte 102,69,15,56,220,222 |
243 |
< |
movups 128(%r15),%xmm15 |
240 |
> |
pxor %xmm8,%xmm3 |
241 |
> |
xorl %ebp,%edi |
242 |
> |
movl %ebx,%esi |
243 |
> |
addl 36(%rsp),%eax |
244 |
> |
pxor %xmm3,%xmm10 |
245 |
|
xorl %edx,%ecx |
246 |
< |
pslldq $12,%xmm9 |
247 |
< |
paddd %xmm6,%xmm6 |
248 |
< |
movl %eax,%edi |
256 |
< |
roll $5,%eax |
246 |
> |
roll $5,%ebx |
247 |
> |
movdqa %xmm12,16(%rsp) |
248 |
> |
addl %edi,%eax |
249 |
|
andl %ecx,%esi |
250 |
+ |
movdqa %xmm10,%xmm13 |
251 |
|
xorl %edx,%ecx |
252 |
< |
psrld $31,%xmm10 |
260 |
< |
xorl %edx,%esi |
261 |
< |
addl %eax,%ebp |
262 |
< |
movdqa %xmm9,%xmm8 |
252 |
> |
addl %ebx,%eax |
253 |
|
rorl $7,%ebx |
254 |
< |
addl %esi,%ebp |
255 |
< |
psrld $30,%xmm9 |
256 |
< |
por %xmm10,%xmm6 |
257 |
< |
addl 44(%rsp),%edx |
254 |
> |
movups 16(%r15),%xmm1 |
255 |
> |
.byte 102,15,56,220,208 |
256 |
> |
movdqa %xmm10,%xmm3 |
257 |
> |
xorl %edx,%esi |
258 |
> |
pslldq $12,%xmm13 |
259 |
> |
paddd %xmm10,%xmm10 |
260 |
> |
movl %eax,%edi |
261 |
> |
addl 40(%rsp),%ebp |
262 |
> |
psrld $31,%xmm3 |
263 |
|
xorl %ecx,%ebx |
264 |
< |
movl %ebp,%esi |
265 |
< |
roll $5,%ebp |
266 |
< |
pslld $2,%xmm8 |
272 |
< |
pxor %xmm9,%xmm6 |
264 |
> |
roll $5,%eax |
265 |
> |
addl %esi,%ebp |
266 |
> |
movdqa %xmm13,%xmm12 |
267 |
|
andl %ebx,%edi |
268 |
|
xorl %ecx,%ebx |
269 |
< |
movdqa 16(%r11),%xmm9 |
270 |
< |
xorl %ecx,%edi |
277 |
< |
.byte 102,69,15,56,220,223 |
278 |
< |
movups 144(%r15),%xmm14 |
279 |
< |
addl %ebp,%edx |
280 |
< |
pxor %xmm8,%xmm6 |
269 |
> |
psrld $30,%xmm13 |
270 |
> |
addl %eax,%ebp |
271 |
|
rorl $7,%eax |
272 |
< |
addl %edi,%edx |
273 |
< |
movdqa %xmm4,%xmm7 |
274 |
< |
addl 48(%rsp),%ecx |
272 |
> |
por %xmm3,%xmm10 |
273 |
> |
xorl %ecx,%edi |
274 |
> |
movl %ebp,%esi |
275 |
> |
addl 44(%rsp),%edx |
276 |
> |
pslld $2,%xmm12 |
277 |
> |
pxor %xmm13,%xmm10 |
278 |
|
xorl %ebx,%eax |
279 |
< |
movdqa %xmm6,%xmm8 |
280 |
< |
.byte 102,15,58,15,251,8 |
281 |
< |
movl %edx,%edi |
282 |
< |
roll $5,%edx |
283 |
< |
paddd %xmm6,%xmm9 |
279 |
> |
movdqa 16(%r11),%xmm13 |
280 |
> |
roll $5,%ebp |
281 |
> |
addl %edi,%edx |
282 |
> |
movups 32(%r15),%xmm0 |
283 |
> |
.byte 102,15,56,220,209 |
284 |
|
andl %eax,%esi |
285 |
+ |
pxor %xmm12,%xmm10 |
286 |
|
xorl %ebx,%eax |
287 |
< |
psrldq $4,%xmm8 |
294 |
< |
xorl %ebx,%esi |
295 |
< |
addl %edx,%ecx |
296 |
< |
pxor %xmm3,%xmm7 |
287 |
> |
addl %ebp,%edx |
288 |
|
rorl $7,%ebp |
289 |
< |
addl %esi,%ecx |
290 |
< |
pxor %xmm5,%xmm8 |
291 |
< |
addl 52(%rsp),%ebx |
289 |
> |
pshufd $238,%xmm7,%xmm11 |
290 |
> |
xorl %ebx,%esi |
291 |
> |
movdqa %xmm10,%xmm12 |
292 |
> |
paddd %xmm10,%xmm13 |
293 |
> |
movl %edx,%edi |
294 |
> |
addl 48(%rsp),%ecx |
295 |
> |
punpcklqdq %xmm8,%xmm11 |
296 |
|
xorl %eax,%ebp |
297 |
< |
movl %ecx,%esi |
298 |
< |
roll $5,%ecx |
299 |
< |
.byte 102,69,15,56,220,222 |
305 |
< |
movups 160(%r15),%xmm15 |
306 |
< |
pxor %xmm8,%xmm7 |
297 |
> |
roll $5,%edx |
298 |
> |
addl %esi,%ecx |
299 |
> |
psrldq $4,%xmm12 |
300 |
|
andl %ebp,%edi |
301 |
|
xorl %eax,%ebp |
302 |
< |
movdqa %xmm9,32(%rsp) |
303 |
< |
xorl %eax,%edi |
311 |
< |
addl %ecx,%ebx |
312 |
< |
movdqa %xmm7,%xmm10 |
313 |
< |
movdqa %xmm7,%xmm8 |
302 |
> |
pxor %xmm7,%xmm11 |
303 |
> |
addl %edx,%ecx |
304 |
|
rorl $7,%edx |
305 |
< |
addl %edi,%ebx |
306 |
< |
addl 56(%rsp),%eax |
305 |
> |
pxor %xmm9,%xmm12 |
306 |
> |
xorl %eax,%edi |
307 |
> |
movl %ecx,%esi |
308 |
> |
addl 52(%rsp),%ebx |
309 |
> |
movups 48(%r15),%xmm1 |
310 |
> |
.byte 102,15,56,220,208 |
311 |
> |
pxor %xmm12,%xmm11 |
312 |
|
xorl %ebp,%edx |
313 |
< |
pslldq $12,%xmm10 |
314 |
< |
paddd %xmm7,%xmm7 |
315 |
< |
movl %ebx,%edi |
321 |
< |
roll $5,%ebx |
313 |
> |
roll $5,%ecx |
314 |
> |
movdqa %xmm13,32(%rsp) |
315 |
> |
addl %edi,%ebx |
316 |
|
andl %edx,%esi |
317 |
+ |
movdqa %xmm11,%xmm3 |
318 |
|
xorl %ebp,%edx |
319 |
< |
psrld $31,%xmm8 |
325 |
< |
xorl %ebp,%esi |
326 |
< |
addl %ebx,%eax |
327 |
< |
movdqa %xmm10,%xmm9 |
319 |
> |
addl %ecx,%ebx |
320 |
|
rorl $7,%ecx |
321 |
+ |
movdqa %xmm11,%xmm12 |
322 |
+ |
xorl %ebp,%esi |
323 |
+ |
pslldq $12,%xmm3 |
324 |
+ |
paddd %xmm11,%xmm11 |
325 |
+ |
movl %ebx,%edi |
326 |
+ |
addl 56(%rsp),%eax |
327 |
+ |
psrld $31,%xmm12 |
328 |
+ |
xorl %edx,%ecx |
329 |
+ |
roll $5,%ebx |
330 |
|
addl %esi,%eax |
331 |
< |
psrld $30,%xmm10 |
332 |
< |
por %xmm8,%xmm7 |
333 |
< |
addl 60(%rsp),%ebp |
331 |
> |
movdqa %xmm3,%xmm13 |
332 |
> |
andl %ecx,%edi |
333 |
> |
xorl %edx,%ecx |
334 |
> |
psrld $30,%xmm3 |
335 |
> |
addl %ebx,%eax |
336 |
> |
rorl $7,%ebx |
337 |
|
cmpl $11,%r8d |
338 |
|
jb .Laesenclast1 |
339 |
< |
movups 176(%r15),%xmm14 |
340 |
< |
.byte 102,69,15,56,220,223 |
341 |
< |
movups 192(%r15),%xmm15 |
342 |
< |
.byte 102,69,15,56,220,222 |
339 |
> |
movups 64(%r15),%xmm0 |
340 |
> |
.byte 102,15,56,220,209 |
341 |
> |
movups 80(%r15),%xmm1 |
342 |
> |
.byte 102,15,56,220,208 |
343 |
|
je .Laesenclast1 |
344 |
< |
movups 208(%r15),%xmm14 |
345 |
< |
.byte 102,69,15,56,220,223 |
346 |
< |
movups 224(%r15),%xmm15 |
347 |
< |
.byte 102,69,15,56,220,222 |
344 |
> |
movups 96(%r15),%xmm0 |
345 |
> |
.byte 102,15,56,220,209 |
346 |
> |
movups 112(%r15),%xmm1 |
347 |
> |
.byte 102,15,56,220,208 |
348 |
|
.Laesenclast1: |
349 |
< |
.byte 102,69,15,56,221,223 |
350 |
< |
movups 16(%r15),%xmm14 |
351 |
< |
xorl %edx,%ecx |
349 |
> |
.byte 102,15,56,221,209 |
350 |
> |
movups 16-112(%r15),%xmm0 |
351 |
> |
por %xmm12,%xmm11 |
352 |
> |
xorl %edx,%edi |
353 |
|
movl %eax,%esi |
354 |
+ |
addl 60(%rsp),%ebp |
355 |
+ |
pslld $2,%xmm13 |
356 |
+ |
pxor %xmm3,%xmm11 |
357 |
+ |
xorl %ecx,%ebx |
358 |
+ |
movdqa 16(%r11),%xmm3 |
359 |
|
roll $5,%eax |
350 |
– |
pslld $2,%xmm9 |
351 |
– |
pxor %xmm10,%xmm7 |
352 |
– |
andl %ecx,%edi |
353 |
– |
xorl %edx,%ecx |
354 |
– |
movdqa 16(%r11),%xmm10 |
355 |
– |
xorl %edx,%edi |
356 |
– |
addl %eax,%ebp |
357 |
– |
pxor %xmm9,%xmm7 |
358 |
– |
rorl $7,%ebx |
360 |
|
addl %edi,%ebp |
360 |
– |
movdqa %xmm7,%xmm9 |
361 |
– |
addl 0(%rsp),%edx |
362 |
– |
pxor %xmm4,%xmm0 |
363 |
– |
.byte 102,68,15,58,15,206,8 |
364 |
– |
xorl %ecx,%ebx |
365 |
– |
movl %ebp,%edi |
366 |
– |
roll $5,%ebp |
367 |
– |
pxor %xmm1,%xmm0 |
361 |
|
andl %ebx,%esi |
362 |
+ |
pxor %xmm13,%xmm11 |
363 |
+ |
pshufd $238,%xmm10,%xmm13 |
364 |
|
xorl %ecx,%ebx |
365 |
< |
movdqa %xmm10,%xmm8 |
371 |
< |
paddd %xmm7,%xmm10 |
372 |
< |
xorl %ecx,%esi |
373 |
< |
movups 16(%r12),%xmm12 |
374 |
< |
xorps %xmm13,%xmm12 |
375 |
< |
movups %xmm11,0(%r13,%r12,1) |
376 |
< |
xorps %xmm12,%xmm11 |
377 |
< |
.byte 102,69,15,56,220,222 |
378 |
< |
movups 32(%r15),%xmm15 |
379 |
< |
addl %ebp,%edx |
380 |
< |
pxor %xmm9,%xmm0 |
365 |
> |
addl %eax,%ebp |
366 |
|
rorl $7,%eax |
367 |
< |
addl %esi,%edx |
368 |
< |
addl 4(%rsp),%ecx |
367 |
> |
pxor %xmm8,%xmm4 |
368 |
> |
xorl %ecx,%esi |
369 |
> |
movl %ebp,%edi |
370 |
> |
addl 0(%rsp),%edx |
371 |
> |
punpcklqdq %xmm11,%xmm13 |
372 |
|
xorl %ebx,%eax |
373 |
< |
movdqa %xmm0,%xmm9 |
374 |
< |
movdqa %xmm10,48(%rsp) |
375 |
< |
movl %edx,%esi |
376 |
< |
roll $5,%edx |
373 |
> |
roll $5,%ebp |
374 |
> |
pxor %xmm5,%xmm4 |
375 |
> |
addl %esi,%edx |
376 |
> |
movups 16(%r12),%xmm14 |
377 |
> |
xorps %xmm15,%xmm14 |
378 |
> |
movups %xmm2,0(%r12,%r13,1) |
379 |
> |
xorps %xmm14,%xmm2 |
380 |
> |
movups -80(%r15),%xmm1 |
381 |
> |
.byte 102,15,56,220,208 |
382 |
|
andl %eax,%edi |
383 |
+ |
movdqa %xmm3,%xmm12 |
384 |
|
xorl %ebx,%eax |
385 |
< |
pslld $2,%xmm0 |
386 |
< |
xorl %ebx,%edi |
387 |
< |
addl %edx,%ecx |
394 |
< |
psrld $30,%xmm9 |
385 |
> |
paddd %xmm11,%xmm3 |
386 |
> |
addl %ebp,%edx |
387 |
> |
pxor %xmm13,%xmm4 |
388 |
|
rorl $7,%ebp |
389 |
< |
addl %edi,%ecx |
390 |
< |
addl 8(%rsp),%ebx |
389 |
> |
xorl %ebx,%edi |
390 |
> |
movl %edx,%esi |
391 |
> |
addl 4(%rsp),%ecx |
392 |
> |
movdqa %xmm4,%xmm13 |
393 |
|
xorl %eax,%ebp |
394 |
< |
movl %ecx,%edi |
395 |
< |
roll $5,%ecx |
396 |
< |
.byte 102,69,15,56,220,223 |
402 |
< |
movups 48(%r15),%xmm14 |
403 |
< |
por %xmm9,%xmm0 |
394 |
> |
roll $5,%edx |
395 |
> |
movdqa %xmm3,48(%rsp) |
396 |
> |
addl %edi,%ecx |
397 |
|
andl %ebp,%esi |
398 |
|
xorl %eax,%ebp |
399 |
< |
movdqa %xmm0,%xmm10 |
400 |
< |
xorl %eax,%esi |
408 |
< |
addl %ecx,%ebx |
399 |
> |
pslld $2,%xmm4 |
400 |
> |
addl %edx,%ecx |
401 |
|
rorl $7,%edx |
402 |
< |
addl %esi,%ebx |
403 |
< |
addl 12(%rsp),%eax |
402 |
> |
psrld $30,%xmm13 |
403 |
> |
xorl %eax,%esi |
404 |
> |
movl %ecx,%edi |
405 |
> |
addl 8(%rsp),%ebx |
406 |
> |
movups -64(%r15),%xmm0 |
407 |
> |
.byte 102,15,56,220,209 |
408 |
> |
por %xmm13,%xmm4 |
409 |
|
xorl %ebp,%edx |
410 |
< |
movl %ebx,%esi |
411 |
< |
roll $5,%ebx |
410 |
> |
roll $5,%ecx |
411 |
> |
pshufd $238,%xmm11,%xmm3 |
412 |
> |
addl %esi,%ebx |
413 |
|
andl %edx,%edi |
414 |
|
xorl %ebp,%edx |
415 |
+ |
addl %ecx,%ebx |
416 |
+ |
addl 12(%rsp),%eax |
417 |
|
xorl %ebp,%edi |
418 |
< |
addl %ebx,%eax |
419 |
< |
rorl $7,%ecx |
418 |
> |
movl %ebx,%esi |
419 |
> |
roll $5,%ebx |
420 |
|
addl %edi,%eax |
421 |
– |
addl 16(%rsp),%ebp |
422 |
– |
.byte 102,69,15,56,220,222 |
423 |
– |
movups 64(%r15),%xmm15 |
424 |
– |
pxor %xmm5,%xmm1 |
425 |
– |
.byte 102,68,15,58,15,215,8 |
421 |
|
xorl %edx,%esi |
422 |
+ |
rorl $7,%ecx |
423 |
+ |
addl %ebx,%eax |
424 |
+ |
pxor %xmm9,%xmm5 |
425 |
+ |
addl 16(%rsp),%ebp |
426 |
+ |
movups -48(%r15),%xmm1 |
427 |
+ |
.byte 102,15,56,220,208 |
428 |
+ |
xorl %ecx,%esi |
429 |
+ |
punpcklqdq %xmm4,%xmm3 |
430 |
|
movl %eax,%edi |
431 |
|
roll $5,%eax |
432 |
< |
pxor %xmm2,%xmm1 |
430 |
< |
xorl %ecx,%esi |
431 |
< |
addl %eax,%ebp |
432 |
< |
movdqa %xmm8,%xmm9 |
433 |
< |
paddd %xmm0,%xmm8 |
434 |
< |
rorl $7,%ebx |
432 |
> |
pxor %xmm6,%xmm5 |
433 |
|
addl %esi,%ebp |
436 |
– |
pxor %xmm10,%xmm1 |
437 |
– |
addl 20(%rsp),%edx |
434 |
|
xorl %ecx,%edi |
435 |
+ |
movdqa %xmm12,%xmm13 |
436 |
+ |
rorl $7,%ebx |
437 |
+ |
paddd %xmm4,%xmm12 |
438 |
+ |
addl %eax,%ebp |
439 |
+ |
pxor %xmm3,%xmm5 |
440 |
+ |
addl 20(%rsp),%edx |
441 |
+ |
xorl %ebx,%edi |
442 |
|
movl %ebp,%esi |
443 |
|
roll $5,%ebp |
444 |
< |
movdqa %xmm1,%xmm10 |
442 |
< |
movdqa %xmm8,0(%rsp) |
443 |
< |
xorl %ebx,%edi |
444 |
< |
addl %ebp,%edx |
445 |
< |
rorl $7,%eax |
444 |
> |
movdqa %xmm5,%xmm3 |
445 |
|
addl %edi,%edx |
447 |
– |
pslld $2,%xmm1 |
448 |
– |
addl 24(%rsp),%ecx |
446 |
|
xorl %ebx,%esi |
447 |
< |
psrld $30,%xmm10 |
447 |
> |
movdqa %xmm12,0(%rsp) |
448 |
> |
rorl $7,%eax |
449 |
> |
addl %ebp,%edx |
450 |
> |
addl 24(%rsp),%ecx |
451 |
> |
pslld $2,%xmm5 |
452 |
> |
xorl %eax,%esi |
453 |
|
movl %edx,%edi |
454 |
+ |
psrld $30,%xmm3 |
455 |
|
roll $5,%edx |
453 |
– |
xorl %eax,%esi |
454 |
– |
.byte 102,69,15,56,220,223 |
455 |
– |
movups 80(%r15),%xmm14 |
456 |
– |
addl %edx,%ecx |
457 |
– |
rorl $7,%ebp |
456 |
|
addl %esi,%ecx |
457 |
< |
por %xmm10,%xmm1 |
458 |
< |
addl 28(%rsp),%ebx |
457 |
> |
movups -32(%r15),%xmm0 |
458 |
> |
.byte 102,15,56,220,209 |
459 |
|
xorl %eax,%edi |
460 |
< |
movdqa %xmm1,%xmm8 |
460 |
> |
rorl $7,%ebp |
461 |
> |
por %xmm3,%xmm5 |
462 |
> |
addl %edx,%ecx |
463 |
> |
addl 28(%rsp),%ebx |
464 |
> |
pshufd $238,%xmm4,%xmm12 |
465 |
> |
xorl %ebp,%edi |
466 |
|
movl %ecx,%esi |
467 |
|
roll $5,%ecx |
465 |
– |
xorl %ebp,%edi |
466 |
– |
addl %ecx,%ebx |
467 |
– |
rorl $7,%edx |
468 |
|
addl %edi,%ebx |
469 |
– |
addl 32(%rsp),%eax |
470 |
– |
pxor %xmm6,%xmm2 |
471 |
– |
.byte 102,68,15,58,15,192,8 |
469 |
|
xorl %ebp,%esi |
470 |
+ |
rorl $7,%edx |
471 |
+ |
addl %ecx,%ebx |
472 |
+ |
pxor %xmm10,%xmm6 |
473 |
+ |
addl 32(%rsp),%eax |
474 |
+ |
xorl %edx,%esi |
475 |
+ |
punpcklqdq %xmm5,%xmm12 |
476 |
|
movl %ebx,%edi |
477 |
|
roll $5,%ebx |
478 |
< |
pxor %xmm3,%xmm2 |
476 |
< |
xorl %edx,%esi |
477 |
< |
addl %ebx,%eax |
478 |
< |
movdqa 32(%r11),%xmm10 |
479 |
< |
paddd %xmm1,%xmm9 |
480 |
< |
rorl $7,%ecx |
478 |
> |
pxor %xmm7,%xmm6 |
479 |
|
addl %esi,%eax |
482 |
– |
pxor %xmm8,%xmm2 |
483 |
– |
addl 36(%rsp),%ebp |
484 |
– |
.byte 102,69,15,56,220,222 |
485 |
– |
movups 96(%r15),%xmm15 |
480 |
|
xorl %edx,%edi |
481 |
+ |
movdqa 32(%r11),%xmm3 |
482 |
+ |
rorl $7,%ecx |
483 |
+ |
paddd %xmm5,%xmm13 |
484 |
+ |
addl %ebx,%eax |
485 |
+ |
pxor %xmm12,%xmm6 |
486 |
+ |
addl 36(%rsp),%ebp |
487 |
+ |
movups -16(%r15),%xmm1 |
488 |
+ |
.byte 102,15,56,220,208 |
489 |
+ |
xorl %ecx,%edi |
490 |
|
movl %eax,%esi |
491 |
|
roll $5,%eax |
492 |
< |
movdqa %xmm2,%xmm8 |
490 |
< |
movdqa %xmm9,16(%rsp) |
491 |
< |
xorl %ecx,%edi |
492 |
< |
addl %eax,%ebp |
493 |
< |
rorl $7,%ebx |
492 |
> |
movdqa %xmm6,%xmm12 |
493 |
|
addl %edi,%ebp |
495 |
– |
pslld $2,%xmm2 |
496 |
– |
addl 40(%rsp),%edx |
494 |
|
xorl %ecx,%esi |
495 |
< |
psrld $30,%xmm8 |
495 |
> |
movdqa %xmm13,16(%rsp) |
496 |
> |
rorl $7,%ebx |
497 |
> |
addl %eax,%ebp |
498 |
> |
addl 40(%rsp),%edx |
499 |
> |
pslld $2,%xmm6 |
500 |
> |
xorl %ebx,%esi |
501 |
|
movl %ebp,%edi |
502 |
+ |
psrld $30,%xmm12 |
503 |
|
roll $5,%ebp |
501 |
– |
xorl %ebx,%esi |
502 |
– |
addl %ebp,%edx |
503 |
– |
rorl $7,%eax |
504 |
|
addl %esi,%edx |
505 |
– |
por %xmm8,%xmm2 |
506 |
– |
addl 44(%rsp),%ecx |
505 |
|
xorl %ebx,%edi |
506 |
< |
movdqa %xmm2,%xmm9 |
506 |
> |
rorl $7,%eax |
507 |
> |
por %xmm12,%xmm6 |
508 |
> |
addl %ebp,%edx |
509 |
> |
addl 44(%rsp),%ecx |
510 |
> |
pshufd $238,%xmm5,%xmm13 |
511 |
> |
xorl %eax,%edi |
512 |
|
movl %edx,%esi |
513 |
|
roll $5,%edx |
511 |
– |
xorl %eax,%edi |
512 |
– |
.byte 102,69,15,56,220,223 |
513 |
– |
movups 112(%r15),%xmm14 |
514 |
– |
addl %edx,%ecx |
515 |
– |
rorl $7,%ebp |
514 |
|
addl %edi,%ecx |
515 |
< |
addl 48(%rsp),%ebx |
516 |
< |
pxor %xmm7,%xmm3 |
519 |
< |
.byte 102,68,15,58,15,201,8 |
515 |
> |
movups 0(%r15),%xmm0 |
516 |
> |
.byte 102,15,56,220,209 |
517 |
|
xorl %eax,%esi |
518 |
+ |
rorl $7,%ebp |
519 |
+ |
addl %edx,%ecx |
520 |
+ |
pxor %xmm11,%xmm7 |
521 |
+ |
addl 48(%rsp),%ebx |
522 |
+ |
xorl %ebp,%esi |
523 |
+ |
punpcklqdq %xmm6,%xmm13 |
524 |
|
movl %ecx,%edi |
525 |
|
roll $5,%ecx |
526 |
< |
pxor %xmm4,%xmm3 |
524 |
< |
xorl %ebp,%esi |
525 |
< |
addl %ecx,%ebx |
526 |
< |
movdqa %xmm10,%xmm8 |
527 |
< |
paddd %xmm2,%xmm10 |
528 |
< |
rorl $7,%edx |
526 |
> |
pxor %xmm8,%xmm7 |
527 |
|
addl %esi,%ebx |
530 |
– |
pxor %xmm9,%xmm3 |
531 |
– |
addl 52(%rsp),%eax |
528 |
|
xorl %ebp,%edi |
529 |
+ |
movdqa %xmm3,%xmm12 |
530 |
+ |
rorl $7,%edx |
531 |
+ |
paddd %xmm6,%xmm3 |
532 |
+ |
addl %ecx,%ebx |
533 |
+ |
pxor %xmm13,%xmm7 |
534 |
+ |
addl 52(%rsp),%eax |
535 |
+ |
xorl %edx,%edi |
536 |
|
movl %ebx,%esi |
537 |
|
roll $5,%ebx |
538 |
< |
movdqa %xmm3,%xmm9 |
536 |
< |
movdqa %xmm10,32(%rsp) |
537 |
< |
xorl %edx,%edi |
538 |
< |
addl %ebx,%eax |
539 |
< |
rorl $7,%ecx |
538 |
> |
movdqa %xmm7,%xmm13 |
539 |
|
addl %edi,%eax |
541 |
– |
pslld $2,%xmm3 |
542 |
– |
addl 56(%rsp),%ebp |
543 |
– |
.byte 102,69,15,56,220,222 |
544 |
– |
movups 128(%r15),%xmm15 |
540 |
|
xorl %edx,%esi |
541 |
< |
psrld $30,%xmm9 |
541 |
> |
movdqa %xmm3,32(%rsp) |
542 |
> |
rorl $7,%ecx |
543 |
> |
addl %ebx,%eax |
544 |
> |
addl 56(%rsp),%ebp |
545 |
> |
movups 16(%r15),%xmm1 |
546 |
> |
.byte 102,15,56,220,208 |
547 |
> |
pslld $2,%xmm7 |
548 |
> |
xorl %ecx,%esi |
549 |
|
movl %eax,%edi |
550 |
+ |
psrld $30,%xmm13 |
551 |
|
roll $5,%eax |
549 |
– |
xorl %ecx,%esi |
550 |
– |
addl %eax,%ebp |
551 |
– |
rorl $7,%ebx |
552 |
|
addl %esi,%ebp |
553 |
– |
por %xmm9,%xmm3 |
554 |
– |
addl 60(%rsp),%edx |
553 |
|
xorl %ecx,%edi |
554 |
< |
movdqa %xmm3,%xmm10 |
554 |
> |
rorl $7,%ebx |
555 |
> |
por %xmm13,%xmm7 |
556 |
> |
addl %eax,%ebp |
557 |
> |
addl 60(%rsp),%edx |
558 |
> |
pshufd $238,%xmm6,%xmm3 |
559 |
> |
xorl %ebx,%edi |
560 |
|
movl %ebp,%esi |
561 |
|
roll $5,%ebp |
559 |
– |
xorl %ebx,%edi |
560 |
– |
addl %ebp,%edx |
561 |
– |
rorl $7,%eax |
562 |
|
addl %edi,%edx |
563 |
– |
addl 0(%rsp),%ecx |
564 |
– |
pxor %xmm0,%xmm4 |
565 |
– |
.byte 102,68,15,58,15,210,8 |
563 |
|
xorl %ebx,%esi |
564 |
+ |
rorl $7,%eax |
565 |
+ |
addl %ebp,%edx |
566 |
+ |
pxor %xmm4,%xmm8 |
567 |
+ |
addl 0(%rsp),%ecx |
568 |
+ |
xorl %eax,%esi |
569 |
+ |
punpcklqdq %xmm7,%xmm3 |
570 |
|
movl %edx,%edi |
571 |
|
roll $5,%edx |
572 |
< |
pxor %xmm5,%xmm4 |
570 |
< |
xorl %eax,%esi |
571 |
< |
.byte 102,69,15,56,220,223 |
572 |
< |
movups 144(%r15),%xmm14 |
573 |
< |
addl %edx,%ecx |
574 |
< |
movdqa %xmm8,%xmm9 |
575 |
< |
paddd %xmm3,%xmm8 |
576 |
< |
rorl $7,%ebp |
572 |
> |
pxor %xmm9,%xmm8 |
573 |
|
addl %esi,%ecx |
574 |
< |
pxor %xmm10,%xmm4 |
575 |
< |
addl 4(%rsp),%ebx |
574 |
> |
movups 32(%r15),%xmm0 |
575 |
> |
.byte 102,15,56,220,209 |
576 |
|
xorl %eax,%edi |
577 |
+ |
movdqa %xmm12,%xmm13 |
578 |
+ |
rorl $7,%ebp |
579 |
+ |
paddd %xmm7,%xmm12 |
580 |
+ |
addl %edx,%ecx |
581 |
+ |
pxor %xmm3,%xmm8 |
582 |
+ |
addl 4(%rsp),%ebx |
583 |
+ |
xorl %ebp,%edi |
584 |
|
movl %ecx,%esi |
585 |
|
roll $5,%ecx |
586 |
< |
movdqa %xmm4,%xmm10 |
584 |
< |
movdqa %xmm8,48(%rsp) |
585 |
< |
xorl %ebp,%edi |
586 |
< |
addl %ecx,%ebx |
587 |
< |
rorl $7,%edx |
586 |
> |
movdqa %xmm8,%xmm3 |
587 |
|
addl %edi,%ebx |
589 |
– |
pslld $2,%xmm4 |
590 |
– |
addl 8(%rsp),%eax |
588 |
|
xorl %ebp,%esi |
589 |
< |
psrld $30,%xmm10 |
589 |
> |
movdqa %xmm12,48(%rsp) |
590 |
> |
rorl $7,%edx |
591 |
> |
addl %ecx,%ebx |
592 |
> |
addl 8(%rsp),%eax |
593 |
> |
pslld $2,%xmm8 |
594 |
> |
xorl %edx,%esi |
595 |
|
movl %ebx,%edi |
596 |
+ |
psrld $30,%xmm3 |
597 |
|
roll $5,%ebx |
595 |
– |
xorl %edx,%esi |
596 |
– |
addl %ebx,%eax |
597 |
– |
rorl $7,%ecx |
598 |
|
addl %esi,%eax |
599 |
– |
por %xmm10,%xmm4 |
600 |
– |
addl 12(%rsp),%ebp |
601 |
– |
.byte 102,69,15,56,220,222 |
602 |
– |
movups 160(%r15),%xmm15 |
599 |
|
xorl %edx,%edi |
600 |
< |
movdqa %xmm4,%xmm8 |
600 |
> |
rorl $7,%ecx |
601 |
> |
por %xmm3,%xmm8 |
602 |
> |
addl %ebx,%eax |
603 |
> |
addl 12(%rsp),%ebp |
604 |
> |
movups 48(%r15),%xmm1 |
605 |
> |
.byte 102,15,56,220,208 |
606 |
> |
pshufd $238,%xmm7,%xmm12 |
607 |
> |
xorl %ecx,%edi |
608 |
|
movl %eax,%esi |
609 |
|
roll $5,%eax |
607 |
– |
xorl %ecx,%edi |
608 |
– |
addl %eax,%ebp |
609 |
– |
rorl $7,%ebx |
610 |
|
addl %edi,%ebp |
611 |
– |
addl 16(%rsp),%edx |
612 |
– |
pxor %xmm1,%xmm5 |
613 |
– |
.byte 102,68,15,58,15,195,8 |
611 |
|
xorl %ecx,%esi |
612 |
+ |
rorl $7,%ebx |
613 |
+ |
addl %eax,%ebp |
614 |
+ |
pxor %xmm5,%xmm9 |
615 |
+ |
addl 16(%rsp),%edx |
616 |
+ |
xorl %ebx,%esi |
617 |
+ |
punpcklqdq %xmm8,%xmm12 |
618 |
|
movl %ebp,%edi |
619 |
|
roll $5,%ebp |
620 |
< |
pxor %xmm6,%xmm5 |
618 |
< |
xorl %ebx,%esi |
619 |
< |
addl %ebp,%edx |
620 |
< |
movdqa %xmm9,%xmm10 |
621 |
< |
paddd %xmm4,%xmm9 |
622 |
< |
rorl $7,%eax |
620 |
> |
pxor %xmm10,%xmm9 |
621 |
|
addl %esi,%edx |
624 |
– |
pxor %xmm8,%xmm5 |
625 |
– |
addl 20(%rsp),%ecx |
622 |
|
xorl %ebx,%edi |
623 |
+ |
movdqa %xmm13,%xmm3 |
624 |
+ |
rorl $7,%eax |
625 |
+ |
paddd %xmm8,%xmm13 |
626 |
+ |
addl %ebp,%edx |
627 |
+ |
pxor %xmm12,%xmm9 |
628 |
+ |
addl 20(%rsp),%ecx |
629 |
+ |
xorl %eax,%edi |
630 |
|
movl %edx,%esi |
631 |
|
roll $5,%edx |
632 |
< |
movdqa %xmm5,%xmm8 |
633 |
< |
movdqa %xmm9,0(%rsp) |
631 |
< |
xorl %eax,%edi |
632 |
> |
movdqa %xmm9,%xmm12 |
633 |
> |
addl %edi,%ecx |
634 |
|
cmpl $11,%r8d |
635 |
|
jb .Laesenclast2 |
636 |
< |
movups 176(%r15),%xmm14 |
637 |
< |
.byte 102,69,15,56,220,223 |
638 |
< |
movups 192(%r15),%xmm15 |
639 |
< |
.byte 102,69,15,56,220,222 |
636 |
> |
movups 64(%r15),%xmm0 |
637 |
> |
.byte 102,15,56,220,209 |
638 |
> |
movups 80(%r15),%xmm1 |
639 |
> |
.byte 102,15,56,220,208 |
640 |
|
je .Laesenclast2 |
641 |
< |
movups 208(%r15),%xmm14 |
642 |
< |
.byte 102,69,15,56,220,223 |
643 |
< |
movups 224(%r15),%xmm15 |
644 |
< |
.byte 102,69,15,56,220,222 |
641 |
> |
movups 96(%r15),%xmm0 |
642 |
> |
.byte 102,15,56,220,209 |
643 |
> |
movups 112(%r15),%xmm1 |
644 |
> |
.byte 102,15,56,220,208 |
645 |
|
.Laesenclast2: |
646 |
< |
.byte 102,69,15,56,221,223 |
647 |
< |
movups 16(%r15),%xmm14 |
648 |
< |
addl %edx,%ecx |
646 |
> |
.byte 102,15,56,221,209 |
647 |
> |
movups 16-112(%r15),%xmm0 |
648 |
> |
xorl %eax,%esi |
649 |
> |
movdqa %xmm13,0(%rsp) |
650 |
|
rorl $7,%ebp |
651 |
< |
addl %edi,%ecx |
649 |
< |
pslld $2,%xmm5 |
651 |
> |
addl %edx,%ecx |
652 |
|
addl 24(%rsp),%ebx |
653 |
< |
xorl %eax,%esi |
654 |
< |
psrld $30,%xmm8 |
653 |
> |
pslld $2,%xmm9 |
654 |
> |
xorl %ebp,%esi |
655 |
|
movl %ecx,%edi |
656 |
+ |
psrld $30,%xmm12 |
657 |
|
roll $5,%ecx |
655 |
– |
xorl %ebp,%esi |
656 |
– |
addl %ecx,%ebx |
657 |
– |
rorl $7,%edx |
658 |
|
addl %esi,%ebx |
659 |
– |
por %xmm8,%xmm5 |
660 |
– |
addl 28(%rsp),%eax |
659 |
|
xorl %ebp,%edi |
660 |
< |
movdqa %xmm5,%xmm9 |
660 |
> |
rorl $7,%edx |
661 |
> |
por %xmm12,%xmm9 |
662 |
> |
addl %ecx,%ebx |
663 |
> |
addl 28(%rsp),%eax |
664 |
> |
pshufd $238,%xmm8,%xmm13 |
665 |
> |
rorl $7,%ecx |
666 |
|
movl %ebx,%esi |
664 |
– |
roll $5,%ebx |
667 |
|
xorl %edx,%edi |
668 |
< |
addl %ebx,%eax |
667 |
< |
rorl $7,%ecx |
668 |
> |
roll $5,%ebx |
669 |
|
addl %edi,%eax |
670 |
< |
movl %ecx,%edi |
670 |
< |
movups 32(%r12),%xmm12 |
671 |
< |
xorps %xmm13,%xmm12 |
672 |
< |
movups %xmm11,16(%r13,%r12,1) |
673 |
< |
xorps %xmm12,%xmm11 |
674 |
< |
.byte 102,69,15,56,220,222 |
675 |
< |
movups 32(%r15),%xmm15 |
676 |
< |
pxor %xmm2,%xmm6 |
677 |
< |
.byte 102,68,15,58,15,204,8 |
670 |
> |
xorl %ecx,%esi |
671 |
|
xorl %edx,%ecx |
672 |
+ |
addl %ebx,%eax |
673 |
+ |
pxor %xmm6,%xmm10 |
674 |
|
addl 32(%rsp),%ebp |
675 |
< |
andl %edx,%edi |
676 |
< |
pxor %xmm7,%xmm6 |
675 |
> |
movups 32(%r12),%xmm14 |
676 |
> |
xorps %xmm15,%xmm14 |
677 |
> |
movups %xmm2,16(%r13,%r12,1) |
678 |
> |
xorps %xmm14,%xmm2 |
679 |
> |
movups -80(%r15),%xmm1 |
680 |
> |
.byte 102,15,56,220,208 |
681 |
|
andl %ecx,%esi |
682 |
+ |
xorl %edx,%ecx |
683 |
|
rorl $7,%ebx |
684 |
< |
movdqa %xmm10,%xmm8 |
685 |
< |
paddd %xmm5,%xmm10 |
686 |
< |
addl %edi,%ebp |
684 |
> |
punpcklqdq %xmm9,%xmm13 |
685 |
|
movl %eax,%edi |
686 |
< |
pxor %xmm9,%xmm6 |
686 |
> |
xorl %ecx,%esi |
687 |
> |
pxor %xmm11,%xmm10 |
688 |
|
roll $5,%eax |
689 |
|
addl %esi,%ebp |
690 |
< |
xorl %edx,%ecx |
691 |
< |
addl %eax,%ebp |
692 |
< |
movdqa %xmm6,%xmm9 |
694 |
< |
movdqa %xmm10,16(%rsp) |
695 |
< |
movl %ebx,%esi |
690 |
> |
movdqa %xmm3,%xmm12 |
691 |
> |
xorl %ebx,%edi |
692 |
> |
paddd %xmm9,%xmm3 |
693 |
|
xorl %ecx,%ebx |
694 |
+ |
pxor %xmm13,%xmm10 |
695 |
+ |
addl %eax,%ebp |
696 |
|
addl 36(%rsp),%edx |
698 |
– |
andl %ecx,%esi |
699 |
– |
pslld $2,%xmm6 |
697 |
|
andl %ebx,%edi |
698 |
+ |
xorl %ecx,%ebx |
699 |
|
rorl $7,%eax |
700 |
< |
psrld $30,%xmm9 |
703 |
< |
addl %esi,%edx |
700 |
> |
movdqa %xmm10,%xmm13 |
701 |
|
movl %ebp,%esi |
702 |
+ |
xorl %ebx,%edi |
703 |
+ |
movdqa %xmm3,16(%rsp) |
704 |
|
roll $5,%ebp |
706 |
– |
.byte 102,69,15,56,220,223 |
707 |
– |
movups 48(%r15),%xmm14 |
705 |
|
addl %edi,%edx |
706 |
< |
xorl %ecx,%ebx |
707 |
< |
addl %ebp,%edx |
708 |
< |
por %xmm9,%xmm6 |
709 |
< |
movl %eax,%edi |
706 |
> |
movups -64(%r15),%xmm0 |
707 |
> |
.byte 102,15,56,220,209 |
708 |
> |
xorl %eax,%esi |
709 |
> |
pslld $2,%xmm10 |
710 |
|
xorl %ebx,%eax |
711 |
< |
movdqa %xmm6,%xmm10 |
711 |
> |
addl %ebp,%edx |
712 |
> |
psrld $30,%xmm13 |
713 |
|
addl 40(%rsp),%ecx |
716 |
– |
andl %ebx,%edi |
714 |
|
andl %eax,%esi |
715 |
+ |
xorl %ebx,%eax |
716 |
+ |
por %xmm13,%xmm10 |
717 |
|
rorl $7,%ebp |
719 |
– |
addl %edi,%ecx |
718 |
|
movl %edx,%edi |
719 |
+ |
xorl %eax,%esi |
720 |
|
roll $5,%edx |
721 |
+ |
pshufd $238,%xmm9,%xmm3 |
722 |
|
addl %esi,%ecx |
723 |
< |
xorl %ebx,%eax |
724 |
< |
addl %edx,%ecx |
725 |
< |
movl %ebp,%esi |
723 |
> |
xorl %ebp,%edi |
724 |
|
xorl %eax,%ebp |
725 |
+ |
addl %edx,%ecx |
726 |
|
addl 44(%rsp),%ebx |
728 |
– |
andl %eax,%esi |
727 |
|
andl %ebp,%edi |
728 |
< |
.byte 102,69,15,56,220,222 |
731 |
< |
movups 64(%r15),%xmm15 |
728 |
> |
xorl %eax,%ebp |
729 |
|
rorl $7,%edx |
730 |
< |
addl %esi,%ebx |
730 |
> |
movups -48(%r15),%xmm1 |
731 |
> |
.byte 102,15,56,220,208 |
732 |
|
movl %ecx,%esi |
733 |
+ |
xorl %ebp,%edi |
734 |
|
roll $5,%ecx |
735 |
|
addl %edi,%ebx |
736 |
< |
xorl %eax,%ebp |
738 |
< |
addl %ecx,%ebx |
739 |
< |
movl %edx,%edi |
740 |
< |
pxor %xmm3,%xmm7 |
741 |
< |
.byte 102,68,15,58,15,213,8 |
736 |
> |
xorl %edx,%esi |
737 |
|
xorl %ebp,%edx |
738 |
+ |
addl %ecx,%ebx |
739 |
+ |
pxor %xmm7,%xmm11 |
740 |
|
addl 48(%rsp),%eax |
744 |
– |
andl %ebp,%edi |
745 |
– |
pxor %xmm0,%xmm7 |
741 |
|
andl %edx,%esi |
742 |
+ |
xorl %ebp,%edx |
743 |
|
rorl $7,%ecx |
744 |
< |
movdqa 48(%r11),%xmm9 |
749 |
< |
paddd %xmm6,%xmm8 |
750 |
< |
addl %edi,%eax |
744 |
> |
punpcklqdq %xmm10,%xmm3 |
745 |
|
movl %ebx,%edi |
746 |
< |
pxor %xmm10,%xmm7 |
746 |
> |
xorl %edx,%esi |
747 |
> |
pxor %xmm4,%xmm11 |
748 |
|
roll $5,%ebx |
749 |
|
addl %esi,%eax |
750 |
< |
xorl %ebp,%edx |
751 |
< |
addl %ebx,%eax |
752 |
< |
movdqa %xmm7,%xmm10 |
758 |
< |
movdqa %xmm8,32(%rsp) |
759 |
< |
movl %ecx,%esi |
760 |
< |
.byte 102,69,15,56,220,223 |
761 |
< |
movups 80(%r15),%xmm14 |
750 |
> |
movdqa 48(%r11),%xmm13 |
751 |
> |
xorl %ecx,%edi |
752 |
> |
paddd %xmm10,%xmm12 |
753 |
|
xorl %edx,%ecx |
754 |
+ |
pxor %xmm3,%xmm11 |
755 |
+ |
addl %ebx,%eax |
756 |
|
addl 52(%rsp),%ebp |
757 |
< |
andl %edx,%esi |
758 |
< |
pslld $2,%xmm7 |
757 |
> |
movups -32(%r15),%xmm0 |
758 |
> |
.byte 102,15,56,220,209 |
759 |
|
andl %ecx,%edi |
760 |
+ |
xorl %edx,%ecx |
761 |
|
rorl $7,%ebx |
762 |
< |
psrld $30,%xmm10 |
769 |
< |
addl %esi,%ebp |
762 |
> |
movdqa %xmm11,%xmm3 |
763 |
|
movl %eax,%esi |
764 |
+ |
xorl %ecx,%edi |
765 |
+ |
movdqa %xmm12,32(%rsp) |
766 |
|
roll $5,%eax |
767 |
|
addl %edi,%ebp |
768 |
< |
xorl %edx,%ecx |
769 |
< |
addl %eax,%ebp |
775 |
< |
por %xmm10,%xmm7 |
776 |
< |
movl %ebx,%edi |
768 |
> |
xorl %ebx,%esi |
769 |
> |
pslld $2,%xmm11 |
770 |
|
xorl %ecx,%ebx |
771 |
< |
movdqa %xmm7,%xmm8 |
771 |
> |
addl %eax,%ebp |
772 |
> |
psrld $30,%xmm3 |
773 |
|
addl 56(%rsp),%edx |
780 |
– |
andl %ecx,%edi |
774 |
|
andl %ebx,%esi |
775 |
+ |
xorl %ecx,%ebx |
776 |
+ |
por %xmm3,%xmm11 |
777 |
|
rorl $7,%eax |
783 |
– |
addl %edi,%edx |
778 |
|
movl %ebp,%edi |
779 |
+ |
xorl %ebx,%esi |
780 |
|
roll $5,%ebp |
781 |
< |
.byte 102,69,15,56,220,222 |
787 |
< |
movups 96(%r15),%xmm15 |
781 |
> |
pshufd $238,%xmm10,%xmm12 |
782 |
|
addl %esi,%edx |
783 |
< |
xorl %ecx,%ebx |
784 |
< |
addl %ebp,%edx |
785 |
< |
movl %eax,%esi |
783 |
> |
movups -16(%r15),%xmm1 |
784 |
> |
.byte 102,15,56,220,208 |
785 |
> |
xorl %eax,%edi |
786 |
|
xorl %ebx,%eax |
787 |
+ |
addl %ebp,%edx |
788 |
|
addl 60(%rsp),%ecx |
794 |
– |
andl %ebx,%esi |
789 |
|
andl %eax,%edi |
790 |
+ |
xorl %ebx,%eax |
791 |
|
rorl $7,%ebp |
797 |
– |
addl %esi,%ecx |
792 |
|
movl %edx,%esi |
793 |
+ |
xorl %eax,%edi |
794 |
|
roll $5,%edx |
795 |
|
addl %edi,%ecx |
796 |
< |
xorl %ebx,%eax |
802 |
< |
addl %edx,%ecx |
803 |
< |
movl %ebp,%edi |
804 |
< |
pxor %xmm4,%xmm0 |
805 |
< |
.byte 102,68,15,58,15,198,8 |
796 |
> |
xorl %ebp,%esi |
797 |
|
xorl %eax,%ebp |
798 |
+ |
addl %edx,%ecx |
799 |
+ |
pxor %xmm8,%xmm4 |
800 |
|
addl 0(%rsp),%ebx |
808 |
– |
andl %eax,%edi |
809 |
– |
pxor %xmm1,%xmm0 |
801 |
|
andl %ebp,%esi |
802 |
< |
.byte 102,69,15,56,220,223 |
812 |
< |
movups 112(%r15),%xmm14 |
802 |
> |
xorl %eax,%ebp |
803 |
|
rorl $7,%edx |
804 |
< |
movdqa %xmm9,%xmm10 |
805 |
< |
paddd %xmm7,%xmm9 |
806 |
< |
addl %edi,%ebx |
804 |
> |
movups 0(%r15),%xmm0 |
805 |
> |
.byte 102,15,56,220,209 |
806 |
> |
punpcklqdq %xmm11,%xmm12 |
807 |
|
movl %ecx,%edi |
808 |
< |
pxor %xmm8,%xmm0 |
808 |
> |
xorl %ebp,%esi |
809 |
> |
pxor %xmm5,%xmm4 |
810 |
|
roll $5,%ecx |
811 |
|
addl %esi,%ebx |
812 |
< |
xorl %eax,%ebp |
813 |
< |
addl %ecx,%ebx |
814 |
< |
movdqa %xmm0,%xmm8 |
824 |
< |
movdqa %xmm9,48(%rsp) |
825 |
< |
movl %edx,%esi |
812 |
> |
movdqa %xmm13,%xmm3 |
813 |
> |
xorl %edx,%edi |
814 |
> |
paddd %xmm11,%xmm13 |
815 |
|
xorl %ebp,%edx |
816 |
+ |
pxor %xmm12,%xmm4 |
817 |
+ |
addl %ecx,%ebx |
818 |
|
addl 4(%rsp),%eax |
828 |
– |
andl %ebp,%esi |
829 |
– |
pslld $2,%xmm0 |
819 |
|
andl %edx,%edi |
820 |
+ |
xorl %ebp,%edx |
821 |
|
rorl $7,%ecx |
822 |
< |
psrld $30,%xmm8 |
833 |
< |
addl %esi,%eax |
822 |
> |
movdqa %xmm4,%xmm12 |
823 |
|
movl %ebx,%esi |
824 |
+ |
xorl %edx,%edi |
825 |
+ |
movdqa %xmm13,48(%rsp) |
826 |
|
roll $5,%ebx |
827 |
|
addl %edi,%eax |
828 |
< |
xorl %ebp,%edx |
829 |
< |
addl %ebx,%eax |
839 |
< |
por %xmm8,%xmm0 |
840 |
< |
movl %ecx,%edi |
841 |
< |
.byte 102,69,15,56,220,222 |
842 |
< |
movups 128(%r15),%xmm15 |
828 |
> |
xorl %ecx,%esi |
829 |
> |
pslld $2,%xmm4 |
830 |
|
xorl %edx,%ecx |
831 |
< |
movdqa %xmm0,%xmm9 |
831 |
> |
addl %ebx,%eax |
832 |
> |
psrld $30,%xmm12 |
833 |
|
addl 8(%rsp),%ebp |
834 |
< |
andl %edx,%edi |
834 |
> |
movups 16(%r15),%xmm1 |
835 |
> |
.byte 102,15,56,220,208 |
836 |
|
andl %ecx,%esi |
837 |
+ |
xorl %edx,%ecx |
838 |
+ |
por %xmm12,%xmm4 |
839 |
|
rorl $7,%ebx |
849 |
– |
addl %edi,%ebp |
840 |
|
movl %eax,%edi |
841 |
+ |
xorl %ecx,%esi |
842 |
|
roll $5,%eax |
843 |
+ |
pshufd $238,%xmm11,%xmm13 |
844 |
|
addl %esi,%ebp |
845 |
< |
xorl %edx,%ecx |
854 |
< |
addl %eax,%ebp |
855 |
< |
movl %ebx,%esi |
845 |
> |
xorl %ebx,%edi |
846 |
|
xorl %ecx,%ebx |
847 |
+ |
addl %eax,%ebp |
848 |
|
addl 12(%rsp),%edx |
858 |
– |
andl %ecx,%esi |
849 |
|
andl %ebx,%edi |
850 |
+ |
xorl %ecx,%ebx |
851 |
|
rorl $7,%eax |
861 |
– |
addl %esi,%edx |
852 |
|
movl %ebp,%esi |
853 |
+ |
xorl %ebx,%edi |
854 |
|
roll $5,%ebp |
864 |
– |
.byte 102,69,15,56,220,223 |
865 |
– |
movups 144(%r15),%xmm14 |
855 |
|
addl %edi,%edx |
856 |
< |
xorl %ecx,%ebx |
857 |
< |
addl %ebp,%edx |
858 |
< |
movl %eax,%edi |
870 |
< |
pxor %xmm5,%xmm1 |
871 |
< |
.byte 102,68,15,58,15,207,8 |
856 |
> |
movups 32(%r15),%xmm0 |
857 |
> |
.byte 102,15,56,220,209 |
858 |
> |
xorl %eax,%esi |
859 |
|
xorl %ebx,%eax |
860 |
+ |
addl %ebp,%edx |
861 |
+ |
pxor %xmm9,%xmm5 |
862 |
|
addl 16(%rsp),%ecx |
874 |
– |
andl %ebx,%edi |
875 |
– |
pxor %xmm2,%xmm1 |
863 |
|
andl %eax,%esi |
864 |
+ |
xorl %ebx,%eax |
865 |
|
rorl $7,%ebp |
866 |
< |
movdqa %xmm10,%xmm8 |
879 |
< |
paddd %xmm0,%xmm10 |
880 |
< |
addl %edi,%ecx |
866 |
> |
punpcklqdq %xmm4,%xmm13 |
867 |
|
movl %edx,%edi |
868 |
< |
pxor %xmm9,%xmm1 |
868 |
> |
xorl %eax,%esi |
869 |
> |
pxor %xmm6,%xmm5 |
870 |
|
roll $5,%edx |
871 |
|
addl %esi,%ecx |
872 |
< |
xorl %ebx,%eax |
873 |
< |
addl %edx,%ecx |
874 |
< |
movdqa %xmm1,%xmm9 |
888 |
< |
movdqa %xmm10,0(%rsp) |
889 |
< |
movl %ebp,%esi |
872 |
> |
movdqa %xmm3,%xmm12 |
873 |
> |
xorl %ebp,%edi |
874 |
> |
paddd %xmm4,%xmm3 |
875 |
|
xorl %eax,%ebp |
876 |
+ |
pxor %xmm13,%xmm5 |
877 |
+ |
addl %edx,%ecx |
878 |
|
addl 20(%rsp),%ebx |
892 |
– |
andl %eax,%esi |
893 |
– |
pslld $2,%xmm1 |
879 |
|
andl %ebp,%edi |
880 |
< |
.byte 102,69,15,56,220,222 |
896 |
< |
movups 160(%r15),%xmm15 |
880 |
> |
xorl %eax,%ebp |
881 |
|
rorl $7,%edx |
882 |
< |
psrld $30,%xmm9 |
883 |
< |
addl %esi,%ebx |
882 |
> |
movups 48(%r15),%xmm1 |
883 |
> |
.byte 102,15,56,220,208 |
884 |
> |
movdqa %xmm5,%xmm13 |
885 |
|
movl %ecx,%esi |
886 |
+ |
xorl %ebp,%edi |
887 |
+ |
movdqa %xmm3,0(%rsp) |
888 |
|
roll $5,%ecx |
889 |
|
addl %edi,%ebx |
890 |
< |
xorl %eax,%ebp |
891 |
< |
addl %ecx,%ebx |
905 |
< |
por %xmm9,%xmm1 |
906 |
< |
movl %edx,%edi |
890 |
> |
xorl %edx,%esi |
891 |
> |
pslld $2,%xmm5 |
892 |
|
xorl %ebp,%edx |
893 |
< |
movdqa %xmm1,%xmm10 |
893 |
> |
addl %ecx,%ebx |
894 |
> |
psrld $30,%xmm13 |
895 |
|
addl 24(%rsp),%eax |
910 |
– |
andl %ebp,%edi |
896 |
|
andl %edx,%esi |
897 |
+ |
xorl %ebp,%edx |
898 |
+ |
por %xmm13,%xmm5 |
899 |
|
rorl $7,%ecx |
913 |
– |
addl %edi,%eax |
900 |
|
movl %ebx,%edi |
901 |
+ |
xorl %edx,%esi |
902 |
|
roll $5,%ebx |
903 |
+ |
pshufd $238,%xmm4,%xmm3 |
904 |
|
addl %esi,%eax |
905 |
< |
xorl %ebp,%edx |
905 |
> |
xorl %ecx,%edi |
906 |
> |
xorl %edx,%ecx |
907 |
|
addl %ebx,%eax |
908 |
< |
movl %ecx,%esi |
908 |
> |
addl 28(%rsp),%ebp |
909 |
|
cmpl $11,%r8d |
910 |
|
jb .Laesenclast3 |
911 |
< |
movups 176(%r15),%xmm14 |
912 |
< |
.byte 102,69,15,56,220,223 |
913 |
< |
movups 192(%r15),%xmm15 |
914 |
< |
.byte 102,69,15,56,220,222 |
911 |
> |
movups 64(%r15),%xmm0 |
912 |
> |
.byte 102,15,56,220,209 |
913 |
> |
movups 80(%r15),%xmm1 |
914 |
> |
.byte 102,15,56,220,208 |
915 |
|
je .Laesenclast3 |
916 |
< |
movups 208(%r15),%xmm14 |
917 |
< |
.byte 102,69,15,56,220,223 |
918 |
< |
movups 224(%r15),%xmm15 |
919 |
< |
.byte 102,69,15,56,220,222 |
916 |
> |
movups 96(%r15),%xmm0 |
917 |
> |
.byte 102,15,56,220,209 |
918 |
> |
movups 112(%r15),%xmm1 |
919 |
> |
.byte 102,15,56,220,208 |
920 |
|
.Laesenclast3: |
921 |
< |
.byte 102,69,15,56,221,223 |
922 |
< |
movups 16(%r15),%xmm14 |
934 |
< |
xorl %edx,%ecx |
935 |
< |
addl 28(%rsp),%ebp |
936 |
< |
andl %edx,%esi |
921 |
> |
.byte 102,15,56,221,209 |
922 |
> |
movups 16-112(%r15),%xmm0 |
923 |
|
andl %ecx,%edi |
924 |
+ |
xorl %edx,%ecx |
925 |
|
rorl $7,%ebx |
939 |
– |
addl %esi,%ebp |
926 |
|
movl %eax,%esi |
927 |
+ |
xorl %ecx,%edi |
928 |
|
roll $5,%eax |
929 |
|
addl %edi,%ebp |
930 |
< |
xorl %edx,%ecx |
944 |
< |
addl %eax,%ebp |
945 |
< |
movl %ebx,%edi |
946 |
< |
pxor %xmm6,%xmm2 |
947 |
< |
.byte 102,68,15,58,15,208,8 |
930 |
> |
xorl %ebx,%esi |
931 |
|
xorl %ecx,%ebx |
932 |
+ |
addl %eax,%ebp |
933 |
+ |
pxor %xmm10,%xmm6 |
934 |
|
addl 32(%rsp),%edx |
950 |
– |
andl %ecx,%edi |
951 |
– |
pxor %xmm3,%xmm2 |
935 |
|
andl %ebx,%esi |
936 |
+ |
xorl %ecx,%ebx |
937 |
|
rorl $7,%eax |
938 |
< |
movdqa %xmm8,%xmm9 |
955 |
< |
paddd %xmm1,%xmm8 |
956 |
< |
addl %edi,%edx |
938 |
> |
punpcklqdq %xmm5,%xmm3 |
939 |
|
movl %ebp,%edi |
940 |
< |
pxor %xmm10,%xmm2 |
940 |
> |
xorl %ebx,%esi |
941 |
> |
pxor %xmm7,%xmm6 |
942 |
|
roll $5,%ebp |
960 |
– |
movups 48(%r12),%xmm12 |
961 |
– |
xorps %xmm13,%xmm12 |
962 |
– |
movups %xmm11,32(%r13,%r12,1) |
963 |
– |
xorps %xmm12,%xmm11 |
964 |
– |
.byte 102,69,15,56,220,222 |
965 |
– |
movups 32(%r15),%xmm15 |
943 |
|
addl %esi,%edx |
944 |
< |
xorl %ecx,%ebx |
945 |
< |
addl %ebp,%edx |
946 |
< |
movdqa %xmm2,%xmm10 |
947 |
< |
movdqa %xmm8,16(%rsp) |
948 |
< |
movl %eax,%esi |
944 |
> |
movups 48(%r12),%xmm14 |
945 |
> |
xorps %xmm15,%xmm14 |
946 |
> |
movups %xmm2,32(%r13,%r12,1) |
947 |
> |
xorps %xmm14,%xmm2 |
948 |
> |
movups -80(%r15),%xmm1 |
949 |
> |
.byte 102,15,56,220,208 |
950 |
> |
movdqa %xmm12,%xmm13 |
951 |
> |
xorl %eax,%edi |
952 |
> |
paddd %xmm5,%xmm12 |
953 |
|
xorl %ebx,%eax |
954 |
+ |
pxor %xmm3,%xmm6 |
955 |
+ |
addl %ebp,%edx |
956 |
|
addl 36(%rsp),%ecx |
974 |
– |
andl %ebx,%esi |
975 |
– |
pslld $2,%xmm2 |
957 |
|
andl %eax,%edi |
958 |
+ |
xorl %ebx,%eax |
959 |
|
rorl $7,%ebp |
960 |
< |
psrld $30,%xmm10 |
979 |
< |
addl %esi,%ecx |
960 |
> |
movdqa %xmm6,%xmm3 |
961 |
|
movl %edx,%esi |
962 |
+ |
xorl %eax,%edi |
963 |
+ |
movdqa %xmm12,16(%rsp) |
964 |
|
roll $5,%edx |
965 |
|
addl %edi,%ecx |
966 |
< |
xorl %ebx,%eax |
967 |
< |
addl %edx,%ecx |
985 |
< |
por %xmm10,%xmm2 |
986 |
< |
movl %ebp,%edi |
966 |
> |
xorl %ebp,%esi |
967 |
> |
pslld $2,%xmm6 |
968 |
|
xorl %eax,%ebp |
969 |
< |
movdqa %xmm2,%xmm8 |
969 |
> |
addl %edx,%ecx |
970 |
> |
psrld $30,%xmm3 |
971 |
|
addl 40(%rsp),%ebx |
990 |
– |
andl %eax,%edi |
972 |
|
andl %ebp,%esi |
973 |
< |
.byte 102,69,15,56,220,223 |
974 |
< |
movups 48(%r15),%xmm14 |
973 |
> |
xorl %eax,%ebp |
974 |
> |
por %xmm3,%xmm6 |
975 |
|
rorl $7,%edx |
976 |
< |
addl %edi,%ebx |
976 |
> |
movups -64(%r15),%xmm0 |
977 |
> |
.byte 102,15,56,220,209 |
978 |
|
movl %ecx,%edi |
979 |
+ |
xorl %ebp,%esi |
980 |
|
roll $5,%ecx |
981 |
+ |
pshufd $238,%xmm5,%xmm12 |
982 |
|
addl %esi,%ebx |
983 |
< |
xorl %eax,%ebp |
1000 |
< |
addl %ecx,%ebx |
1001 |
< |
movl %edx,%esi |
983 |
> |
xorl %edx,%edi |
984 |
|
xorl %ebp,%edx |
985 |
+ |
addl %ecx,%ebx |
986 |
|
addl 44(%rsp),%eax |
1004 |
– |
andl %ebp,%esi |
987 |
|
andl %edx,%edi |
988 |
+ |
xorl %ebp,%edx |
989 |
|
rorl $7,%ecx |
1007 |
– |
addl %esi,%eax |
990 |
|
movl %ebx,%esi |
991 |
+ |
xorl %edx,%edi |
992 |
|
roll $5,%ebx |
993 |
|
addl %edi,%eax |
994 |
< |
xorl %ebp,%edx |
994 |
> |
xorl %edx,%esi |
995 |
|
addl %ebx,%eax |
996 |
+ |
pxor %xmm11,%xmm7 |
997 |
|
addl 48(%rsp),%ebp |
998 |
< |
.byte 102,69,15,56,220,222 |
999 |
< |
movups 64(%r15),%xmm15 |
1000 |
< |
pxor %xmm7,%xmm3 |
1001 |
< |
.byte 102,68,15,58,15,193,8 |
1018 |
< |
xorl %edx,%esi |
998 |
> |
movups -48(%r15),%xmm1 |
999 |
> |
.byte 102,15,56,220,208 |
1000 |
> |
xorl %ecx,%esi |
1001 |
> |
punpcklqdq %xmm6,%xmm12 |
1002 |
|
movl %eax,%edi |
1003 |
|
roll $5,%eax |
1004 |
< |
pxor %xmm4,%xmm3 |
1022 |
< |
xorl %ecx,%esi |
1023 |
< |
addl %eax,%ebp |
1024 |
< |
movdqa %xmm9,%xmm10 |
1025 |
< |
paddd %xmm2,%xmm9 |
1026 |
< |
rorl $7,%ebx |
1004 |
> |
pxor %xmm8,%xmm7 |
1005 |
|
addl %esi,%ebp |
1028 |
– |
pxor %xmm8,%xmm3 |
1029 |
– |
addl 52(%rsp),%edx |
1006 |
|
xorl %ecx,%edi |
1007 |
+ |
movdqa %xmm13,%xmm3 |
1008 |
+ |
rorl $7,%ebx |
1009 |
+ |
paddd %xmm6,%xmm13 |
1010 |
+ |
addl %eax,%ebp |
1011 |
+ |
pxor %xmm12,%xmm7 |
1012 |
+ |
addl 52(%rsp),%edx |
1013 |
+ |
xorl %ebx,%edi |
1014 |
|
movl %ebp,%esi |
1015 |
|
roll $5,%ebp |
1016 |
< |
movdqa %xmm3,%xmm8 |
1034 |
< |
movdqa %xmm9,32(%rsp) |
1035 |
< |
xorl %ebx,%edi |
1036 |
< |
addl %ebp,%edx |
1037 |
< |
rorl $7,%eax |
1016 |
> |
movdqa %xmm7,%xmm12 |
1017 |
|
addl %edi,%edx |
1039 |
– |
pslld $2,%xmm3 |
1040 |
– |
addl 56(%rsp),%ecx |
1018 |
|
xorl %ebx,%esi |
1019 |
< |
psrld $30,%xmm8 |
1019 |
> |
movdqa %xmm13,32(%rsp) |
1020 |
> |
rorl $7,%eax |
1021 |
> |
addl %ebp,%edx |
1022 |
> |
addl 56(%rsp),%ecx |
1023 |
> |
pslld $2,%xmm7 |
1024 |
> |
xorl %eax,%esi |
1025 |
|
movl %edx,%edi |
1026 |
+ |
psrld $30,%xmm12 |
1027 |
|
roll $5,%edx |
1045 |
– |
xorl %eax,%esi |
1046 |
– |
.byte 102,69,15,56,220,223 |
1047 |
– |
movups 80(%r15),%xmm14 |
1048 |
– |
addl %edx,%ecx |
1049 |
– |
rorl $7,%ebp |
1028 |
|
addl %esi,%ecx |
1029 |
< |
por %xmm8,%xmm3 |
1030 |
< |
addl 60(%rsp),%ebx |
1029 |
> |
movups -32(%r15),%xmm0 |
1030 |
> |
.byte 102,15,56,220,209 |
1031 |
|
xorl %eax,%edi |
1032 |
+ |
rorl $7,%ebp |
1033 |
+ |
por %xmm12,%xmm7 |
1034 |
+ |
addl %edx,%ecx |
1035 |
+ |
addl 60(%rsp),%ebx |
1036 |
+ |
xorl %ebp,%edi |
1037 |
|
movl %ecx,%esi |
1038 |
|
roll $5,%ecx |
1056 |
– |
xorl %ebp,%edi |
1057 |
– |
addl %ecx,%ebx |
1058 |
– |
rorl $7,%edx |
1039 |
|
addl %edi,%ebx |
1060 |
– |
addl 0(%rsp),%eax |
1061 |
– |
paddd %xmm3,%xmm10 |
1040 |
|
xorl %ebp,%esi |
1041 |
+ |
rorl $7,%edx |
1042 |
+ |
addl %ecx,%ebx |
1043 |
+ |
addl 0(%rsp),%eax |
1044 |
+ |
xorl %edx,%esi |
1045 |
|
movl %ebx,%edi |
1046 |
|
roll $5,%ebx |
1047 |
< |
xorl %edx,%esi |
1066 |
< |
movdqa %xmm10,48(%rsp) |
1067 |
< |
addl %ebx,%eax |
1068 |
< |
rorl $7,%ecx |
1047 |
> |
paddd %xmm7,%xmm3 |
1048 |
|
addl %esi,%eax |
1070 |
– |
addl 4(%rsp),%ebp |
1071 |
– |
.byte 102,69,15,56,220,222 |
1072 |
– |
movups 96(%r15),%xmm15 |
1049 |
|
xorl %edx,%edi |
1050 |
+ |
movdqa %xmm3,48(%rsp) |
1051 |
+ |
rorl $7,%ecx |
1052 |
+ |
addl %ebx,%eax |
1053 |
+ |
addl 4(%rsp),%ebp |
1054 |
+ |
movups -16(%r15),%xmm1 |
1055 |
+ |
.byte 102,15,56,220,208 |
1056 |
+ |
xorl %ecx,%edi |
1057 |
|
movl %eax,%esi |
1058 |
|
roll $5,%eax |
1076 |
– |
xorl %ecx,%edi |
1077 |
– |
addl %eax,%ebp |
1078 |
– |
rorl $7,%ebx |
1059 |
|
addl %edi,%ebp |
1080 |
– |
addl 8(%rsp),%edx |
1060 |
|
xorl %ecx,%esi |
1061 |
+ |
rorl $7,%ebx |
1062 |
+ |
addl %eax,%ebp |
1063 |
+ |
addl 8(%rsp),%edx |
1064 |
+ |
xorl %ebx,%esi |
1065 |
|
movl %ebp,%edi |
1066 |
|
roll $5,%ebp |
1084 |
– |
xorl %ebx,%esi |
1085 |
– |
addl %ebp,%edx |
1086 |
– |
rorl $7,%eax |
1067 |
|
addl %esi,%edx |
1088 |
– |
addl 12(%rsp),%ecx |
1068 |
|
xorl %ebx,%edi |
1069 |
+ |
rorl $7,%eax |
1070 |
+ |
addl %ebp,%edx |
1071 |
+ |
addl 12(%rsp),%ecx |
1072 |
+ |
xorl %eax,%edi |
1073 |
|
movl %edx,%esi |
1074 |
|
roll $5,%edx |
1092 |
– |
xorl %eax,%edi |
1093 |
– |
.byte 102,69,15,56,220,223 |
1094 |
– |
movups 112(%r15),%xmm14 |
1095 |
– |
addl %edx,%ecx |
1096 |
– |
rorl $7,%ebp |
1075 |
|
addl %edi,%ecx |
1076 |
+ |
movups 0(%r15),%xmm0 |
1077 |
+ |
.byte 102,15,56,220,209 |
1078 |
+ |
xorl %eax,%esi |
1079 |
+ |
rorl $7,%ebp |
1080 |
+ |
addl %edx,%ecx |
1081 |
|
cmpq %r14,%r10 |
1082 |
|
je .Ldone_ssse3 |
1083 |
< |
movdqa 64(%r11),%xmm6 |
1084 |
< |
movdqa 0(%r11),%xmm9 |
1085 |
< |
movdqu 0(%r10),%xmm0 |
1086 |
< |
movdqu 16(%r10),%xmm1 |
1087 |
< |
movdqu 32(%r10),%xmm2 |
1088 |
< |
movdqu 48(%r10),%xmm3 |
1089 |
< |
.byte 102,15,56,0,198 |
1083 |
> |
movdqa 64(%r11),%xmm3 |
1084 |
> |
movdqa 0(%r11),%xmm13 |
1085 |
> |
movdqu 0(%r10),%xmm4 |
1086 |
> |
movdqu 16(%r10),%xmm5 |
1087 |
> |
movdqu 32(%r10),%xmm6 |
1088 |
> |
movdqu 48(%r10),%xmm7 |
1089 |
> |
.byte 102,15,56,0,227 |
1090 |
|
addq $64,%r10 |
1091 |
|
addl 16(%rsp),%ebx |
1092 |
< |
xorl %eax,%esi |
1110 |
< |
.byte 102,15,56,0,206 |
1092 |
> |
xorl %ebp,%esi |
1093 |
|
movl %ecx,%edi |
1094 |
+ |
.byte 102,15,56,0,235 |
1095 |
|
roll $5,%ecx |
1113 |
– |
paddd %xmm9,%xmm0 |
1114 |
– |
xorl %ebp,%esi |
1115 |
– |
addl %ecx,%ebx |
1116 |
– |
rorl $7,%edx |
1096 |
|
addl %esi,%ebx |
1118 |
– |
movdqa %xmm0,0(%rsp) |
1119 |
– |
addl 20(%rsp),%eax |
1097 |
|
xorl %ebp,%edi |
1098 |
< |
psubd %xmm9,%xmm0 |
1098 |
> |
rorl $7,%edx |
1099 |
> |
paddd %xmm13,%xmm4 |
1100 |
> |
addl %ecx,%ebx |
1101 |
> |
addl 20(%rsp),%eax |
1102 |
> |
xorl %edx,%edi |
1103 |
|
movl %ebx,%esi |
1104 |
+ |
movdqa %xmm4,0(%rsp) |
1105 |
|
roll $5,%ebx |
1124 |
– |
xorl %edx,%edi |
1125 |
– |
addl %ebx,%eax |
1126 |
– |
rorl $7,%ecx |
1106 |
|
addl %edi,%eax |
1128 |
– |
addl 24(%rsp),%ebp |
1129 |
– |
.byte 102,69,15,56,220,222 |
1130 |
– |
movups 128(%r15),%xmm15 |
1107 |
|
xorl %edx,%esi |
1108 |
+ |
rorl $7,%ecx |
1109 |
+ |
psubd %xmm13,%xmm4 |
1110 |
+ |
addl %ebx,%eax |
1111 |
+ |
addl 24(%rsp),%ebp |
1112 |
+ |
movups 16(%r15),%xmm1 |
1113 |
+ |
.byte 102,15,56,220,208 |
1114 |
+ |
xorl %ecx,%esi |
1115 |
|
movl %eax,%edi |
1116 |
|
roll $5,%eax |
1134 |
– |
xorl %ecx,%esi |
1135 |
– |
addl %eax,%ebp |
1136 |
– |
rorl $7,%ebx |
1117 |
|
addl %esi,%ebp |
1138 |
– |
addl 28(%rsp),%edx |
1118 |
|
xorl %ecx,%edi |
1119 |
+ |
rorl $7,%ebx |
1120 |
+ |
addl %eax,%ebp |
1121 |
+ |
addl 28(%rsp),%edx |
1122 |
+ |
xorl %ebx,%edi |
1123 |
|
movl %ebp,%esi |
1124 |
|
roll $5,%ebp |
1142 |
– |
xorl %ebx,%edi |
1143 |
– |
addl %ebp,%edx |
1144 |
– |
rorl $7,%eax |
1125 |
|
addl %edi,%edx |
1146 |
– |
addl 32(%rsp),%ecx |
1126 |
|
xorl %ebx,%esi |
1127 |
< |
.byte 102,15,56,0,214 |
1127 |
> |
rorl $7,%eax |
1128 |
> |
addl %ebp,%edx |
1129 |
> |
addl 32(%rsp),%ecx |
1130 |
> |
xorl %eax,%esi |
1131 |
|
movl %edx,%edi |
1132 |
+ |
.byte 102,15,56,0,243 |
1133 |
|
roll $5,%edx |
1151 |
– |
paddd %xmm9,%xmm1 |
1152 |
– |
xorl %eax,%esi |
1153 |
– |
.byte 102,69,15,56,220,223 |
1154 |
– |
movups 144(%r15),%xmm14 |
1155 |
– |
addl %edx,%ecx |
1156 |
– |
rorl $7,%ebp |
1134 |
|
addl %esi,%ecx |
1135 |
< |
movdqa %xmm1,16(%rsp) |
1136 |
< |
addl 36(%rsp),%ebx |
1135 |
> |
movups 32(%r15),%xmm0 |
1136 |
> |
.byte 102,15,56,220,209 |
1137 |
|
xorl %eax,%edi |
1138 |
< |
psubd %xmm9,%xmm1 |
1138 |
> |
rorl $7,%ebp |
1139 |
> |
paddd %xmm13,%xmm5 |
1140 |
> |
addl %edx,%ecx |
1141 |
> |
addl 36(%rsp),%ebx |
1142 |
> |
xorl %ebp,%edi |
1143 |
|
movl %ecx,%esi |
1144 |
+ |
movdqa %xmm5,16(%rsp) |
1145 |
|
roll $5,%ecx |
1164 |
– |
xorl %ebp,%edi |
1165 |
– |
addl %ecx,%ebx |
1166 |
– |
rorl $7,%edx |
1146 |
|
addl %edi,%ebx |
1168 |
– |
addl 40(%rsp),%eax |
1147 |
|
xorl %ebp,%esi |
1148 |
+ |
rorl $7,%edx |
1149 |
+ |
psubd %xmm13,%xmm5 |
1150 |
+ |
addl %ecx,%ebx |
1151 |
+ |
addl 40(%rsp),%eax |
1152 |
+ |
xorl %edx,%esi |
1153 |
|
movl %ebx,%edi |
1154 |
|
roll $5,%ebx |
1172 |
– |
xorl %edx,%esi |
1173 |
– |
addl %ebx,%eax |
1174 |
– |
rorl $7,%ecx |
1155 |
|
addl %esi,%eax |
1176 |
– |
addl 44(%rsp),%ebp |
1177 |
– |
.byte 102,69,15,56,220,222 |
1178 |
– |
movups 160(%r15),%xmm15 |
1156 |
|
xorl %edx,%edi |
1157 |
+ |
rorl $7,%ecx |
1158 |
+ |
addl %ebx,%eax |
1159 |
+ |
addl 44(%rsp),%ebp |
1160 |
+ |
movups 48(%r15),%xmm1 |
1161 |
+ |
.byte 102,15,56,220,208 |
1162 |
+ |
xorl %ecx,%edi |
1163 |
|
movl %eax,%esi |
1164 |
|
roll $5,%eax |
1182 |
– |
xorl %ecx,%edi |
1183 |
– |
addl %eax,%ebp |
1184 |
– |
rorl $7,%ebx |
1165 |
|
addl %edi,%ebp |
1186 |
– |
addl 48(%rsp),%edx |
1166 |
|
xorl %ecx,%esi |
1167 |
< |
.byte 102,15,56,0,222 |
1167 |
> |
rorl $7,%ebx |
1168 |
> |
addl %eax,%ebp |
1169 |
> |
addl 48(%rsp),%edx |
1170 |
> |
xorl %ebx,%esi |
1171 |
|
movl %ebp,%edi |
1172 |
+ |
.byte 102,15,56,0,251 |
1173 |
|
roll $5,%ebp |
1191 |
– |
paddd %xmm9,%xmm2 |
1192 |
– |
xorl %ebx,%esi |
1193 |
– |
addl %ebp,%edx |
1194 |
– |
rorl $7,%eax |
1174 |
|
addl %esi,%edx |
1196 |
– |
movdqa %xmm2,32(%rsp) |
1197 |
– |
addl 52(%rsp),%ecx |
1175 |
|
xorl %ebx,%edi |
1176 |
< |
psubd %xmm9,%xmm2 |
1176 |
> |
rorl $7,%eax |
1177 |
> |
paddd %xmm13,%xmm6 |
1178 |
> |
addl %ebp,%edx |
1179 |
> |
addl 52(%rsp),%ecx |
1180 |
> |
xorl %eax,%edi |
1181 |
|
movl %edx,%esi |
1182 |
+ |
movdqa %xmm6,32(%rsp) |
1183 |
|
roll $5,%edx |
1184 |
< |
xorl %eax,%edi |
1184 |
> |
addl %edi,%ecx |
1185 |
|
cmpl $11,%r8d |
1186 |
|
jb .Laesenclast4 |
1187 |
< |
movups 176(%r15),%xmm14 |
1188 |
< |
.byte 102,69,15,56,220,223 |
1189 |
< |
movups 192(%r15),%xmm15 |
1190 |
< |
.byte 102,69,15,56,220,222 |
1187 |
> |
movups 64(%r15),%xmm0 |
1188 |
> |
.byte 102,15,56,220,209 |
1189 |
> |
movups 80(%r15),%xmm1 |
1190 |
> |
.byte 102,15,56,220,208 |
1191 |
|
je .Laesenclast4 |
1192 |
< |
movups 208(%r15),%xmm14 |
1193 |
< |
.byte 102,69,15,56,220,223 |
1194 |
< |
movups 224(%r15),%xmm15 |
1195 |
< |
.byte 102,69,15,56,220,222 |
1192 |
> |
movups 96(%r15),%xmm0 |
1193 |
> |
.byte 102,15,56,220,209 |
1194 |
> |
movups 112(%r15),%xmm1 |
1195 |
> |
.byte 102,15,56,220,208 |
1196 |
|
.Laesenclast4: |
1197 |
< |
.byte 102,69,15,56,221,223 |
1198 |
< |
movups 16(%r15),%xmm14 |
1199 |
< |
addl %edx,%ecx |
1197 |
> |
.byte 102,15,56,221,209 |
1198 |
> |
movups 16-112(%r15),%xmm0 |
1199 |
> |
xorl %eax,%esi |
1200 |
|
rorl $7,%ebp |
1201 |
< |
addl %edi,%ecx |
1201 |
> |
psubd %xmm13,%xmm6 |
1202 |
> |
addl %edx,%ecx |
1203 |
|
addl 56(%rsp),%ebx |
1204 |
< |
xorl %eax,%esi |
1204 |
> |
xorl %ebp,%esi |
1205 |
|
movl %ecx,%edi |
1206 |
|
roll $5,%ecx |
1224 |
– |
xorl %ebp,%esi |
1225 |
– |
addl %ecx,%ebx |
1226 |
– |
rorl $7,%edx |
1207 |
|
addl %esi,%ebx |
1228 |
– |
addl 60(%rsp),%eax |
1208 |
|
xorl %ebp,%edi |
1209 |
+ |
rorl $7,%edx |
1210 |
+ |
addl %ecx,%ebx |
1211 |
+ |
addl 60(%rsp),%eax |
1212 |
+ |
xorl %edx,%edi |
1213 |
|
movl %ebx,%esi |
1214 |
|
roll $5,%ebx |
1232 |
– |
xorl %edx,%edi |
1233 |
– |
addl %ebx,%eax |
1234 |
– |
rorl $7,%ecx |
1215 |
|
addl %edi,%eax |
1216 |
< |
movups %xmm11,48(%r13,%r12,1) |
1216 |
> |
rorl $7,%ecx |
1217 |
> |
addl %ebx,%eax |
1218 |
> |
movups %xmm2,48(%r13,%r12,1) |
1219 |
|
leaq 64(%r12),%r12 |
1220 |
|
|
1221 |
|
addl 0(%r9),%eax |
1227 |
|
movl %esi,4(%r9) |
1228 |
|
movl %esi,%ebx |
1229 |
|
movl %ecx,8(%r9) |
1230 |
+ |
movl %ecx,%edi |
1231 |
|
movl %edx,12(%r9) |
1232 |
+ |
xorl %edx,%edi |
1233 |
|
movl %ebp,16(%r9) |
1234 |
+ |
andl %edi,%esi |
1235 |
|
jmp .Loop_ssse3 |
1236 |
|
|
1252 |
– |
.align 16 |
1237 |
|
.Ldone_ssse3: |
1238 |
|
addl 16(%rsp),%ebx |
1239 |
< |
xorl %eax,%esi |
1239 |
> |
xorl %ebp,%esi |
1240 |
|
movl %ecx,%edi |
1241 |
|
roll $5,%ecx |
1258 |
– |
xorl %ebp,%esi |
1259 |
– |
addl %ecx,%ebx |
1260 |
– |
rorl $7,%edx |
1242 |
|
addl %esi,%ebx |
1262 |
– |
addl 20(%rsp),%eax |
1243 |
|
xorl %ebp,%edi |
1244 |
+ |
rorl $7,%edx |
1245 |
+ |
addl %ecx,%ebx |
1246 |
+ |
addl 20(%rsp),%eax |
1247 |
+ |
xorl %edx,%edi |
1248 |
|
movl %ebx,%esi |
1249 |
|
roll $5,%ebx |
1266 |
– |
xorl %edx,%edi |
1267 |
– |
addl %ebx,%eax |
1268 |
– |
rorl $7,%ecx |
1250 |
|
addl %edi,%eax |
1270 |
– |
addl 24(%rsp),%ebp |
1271 |
– |
.byte 102,69,15,56,220,222 |
1272 |
– |
movups 128(%r15),%xmm15 |
1251 |
|
xorl %edx,%esi |
1252 |
+ |
rorl $7,%ecx |
1253 |
+ |
addl %ebx,%eax |
1254 |
+ |
addl 24(%rsp),%ebp |
1255 |
+ |
movups 16(%r15),%xmm1 |
1256 |
+ |
.byte 102,15,56,220,208 |
1257 |
+ |
xorl %ecx,%esi |
1258 |
|
movl %eax,%edi |
1259 |
|
roll $5,%eax |
1276 |
– |
xorl %ecx,%esi |
1277 |
– |
addl %eax,%ebp |
1278 |
– |
rorl $7,%ebx |
1260 |
|
addl %esi,%ebp |
1280 |
– |
addl 28(%rsp),%edx |
1261 |
|
xorl %ecx,%edi |
1262 |
+ |
rorl $7,%ebx |
1263 |
+ |
addl %eax,%ebp |
1264 |
+ |
addl 28(%rsp),%edx |
1265 |
+ |
xorl %ebx,%edi |
1266 |
|
movl %ebp,%esi |
1267 |
|
roll $5,%ebp |
1284 |
– |
xorl %ebx,%edi |
1285 |
– |
addl %ebp,%edx |
1286 |
– |
rorl $7,%eax |
1268 |
|
addl %edi,%edx |
1288 |
– |
addl 32(%rsp),%ecx |
1269 |
|
xorl %ebx,%esi |
1270 |
+ |
rorl $7,%eax |
1271 |
+ |
addl %ebp,%edx |
1272 |
+ |
addl 32(%rsp),%ecx |
1273 |
+ |
xorl %eax,%esi |
1274 |
|
movl %edx,%edi |
1275 |
|
roll $5,%edx |
1292 |
– |
xorl %eax,%esi |
1293 |
– |
.byte 102,69,15,56,220,223 |
1294 |
– |
movups 144(%r15),%xmm14 |
1295 |
– |
addl %edx,%ecx |
1296 |
– |
rorl $7,%ebp |
1276 |
|
addl %esi,%ecx |
1277 |
< |
addl 36(%rsp),%ebx |
1277 |
> |
movups 32(%r15),%xmm0 |
1278 |
> |
.byte 102,15,56,220,209 |
1279 |
|
xorl %eax,%edi |
1280 |
+ |
rorl $7,%ebp |
1281 |
+ |
addl %edx,%ecx |
1282 |
+ |
addl 36(%rsp),%ebx |
1283 |
+ |
xorl %ebp,%edi |
1284 |
|
movl %ecx,%esi |
1285 |
|
roll $5,%ecx |
1302 |
– |
xorl %ebp,%edi |
1303 |
– |
addl %ecx,%ebx |
1304 |
– |
rorl $7,%edx |
1286 |
|
addl %edi,%ebx |
1306 |
– |
addl 40(%rsp),%eax |
1287 |
|
xorl %ebp,%esi |
1288 |
+ |
rorl $7,%edx |
1289 |
+ |
addl %ecx,%ebx |
1290 |
+ |
addl 40(%rsp),%eax |
1291 |
+ |
xorl %edx,%esi |
1292 |
|
movl %ebx,%edi |
1293 |
|
roll $5,%ebx |
1310 |
– |
xorl %edx,%esi |
1311 |
– |
addl %ebx,%eax |
1312 |
– |
rorl $7,%ecx |
1294 |
|
addl %esi,%eax |
1314 |
– |
addl 44(%rsp),%ebp |
1315 |
– |
.byte 102,69,15,56,220,222 |
1316 |
– |
movups 160(%r15),%xmm15 |
1295 |
|
xorl %edx,%edi |
1296 |
+ |
rorl $7,%ecx |
1297 |
+ |
addl %ebx,%eax |
1298 |
+ |
addl 44(%rsp),%ebp |
1299 |
+ |
movups 48(%r15),%xmm1 |
1300 |
+ |
.byte 102,15,56,220,208 |
1301 |
+ |
xorl %ecx,%edi |
1302 |
|
movl %eax,%esi |
1303 |
|
roll $5,%eax |
1320 |
– |
xorl %ecx,%edi |
1321 |
– |
addl %eax,%ebp |
1322 |
– |
rorl $7,%ebx |
1304 |
|
addl %edi,%ebp |
1324 |
– |
addl 48(%rsp),%edx |
1305 |
|
xorl %ecx,%esi |
1306 |
+ |
rorl $7,%ebx |
1307 |
+ |
addl %eax,%ebp |
1308 |
+ |
addl 48(%rsp),%edx |
1309 |
+ |
xorl %ebx,%esi |
1310 |
|
movl %ebp,%edi |
1311 |
|
roll $5,%ebp |
1328 |
– |
xorl %ebx,%esi |
1329 |
– |
addl %ebp,%edx |
1330 |
– |
rorl $7,%eax |
1312 |
|
addl %esi,%edx |
1332 |
– |
addl 52(%rsp),%ecx |
1313 |
|
xorl %ebx,%edi |
1314 |
+ |
rorl $7,%eax |
1315 |
+ |
addl %ebp,%edx |
1316 |
+ |
addl 52(%rsp),%ecx |
1317 |
+ |
xorl %eax,%edi |
1318 |
|
movl %edx,%esi |
1319 |
|
roll $5,%edx |
1320 |
< |
xorl %eax,%edi |
1320 |
> |
addl %edi,%ecx |
1321 |
|
cmpl $11,%r8d |
1322 |
|
jb .Laesenclast5 |
1323 |
< |
movups 176(%r15),%xmm14 |
1324 |
< |
.byte 102,69,15,56,220,223 |
1325 |
< |
movups 192(%r15),%xmm15 |
1326 |
< |
.byte 102,69,15,56,220,222 |
1323 |
> |
movups 64(%r15),%xmm0 |
1324 |
> |
.byte 102,15,56,220,209 |
1325 |
> |
movups 80(%r15),%xmm1 |
1326 |
> |
.byte 102,15,56,220,208 |
1327 |
|
je .Laesenclast5 |
1328 |
< |
movups 208(%r15),%xmm14 |
1329 |
< |
.byte 102,69,15,56,220,223 |
1330 |
< |
movups 224(%r15),%xmm15 |
1331 |
< |
.byte 102,69,15,56,220,222 |
1328 |
> |
movups 96(%r15),%xmm0 |
1329 |
> |
.byte 102,15,56,220,209 |
1330 |
> |
movups 112(%r15),%xmm1 |
1331 |
> |
.byte 102,15,56,220,208 |
1332 |
|
.Laesenclast5: |
1333 |
< |
.byte 102,69,15,56,221,223 |
1334 |
< |
movups 16(%r15),%xmm14 |
1335 |
< |
addl %edx,%ecx |
1333 |
> |
.byte 102,15,56,221,209 |
1334 |
> |
movups 16-112(%r15),%xmm0 |
1335 |
> |
xorl %eax,%esi |
1336 |
|
rorl $7,%ebp |
1337 |
< |
addl %edi,%ecx |
1337 |
> |
addl %edx,%ecx |
1338 |
|
addl 56(%rsp),%ebx |
1339 |
< |
xorl %eax,%esi |
1339 |
> |
xorl %ebp,%esi |
1340 |
|
movl %ecx,%edi |
1341 |
|
roll $5,%ecx |
1358 |
– |
xorl %ebp,%esi |
1359 |
– |
addl %ecx,%ebx |
1360 |
– |
rorl $7,%edx |
1342 |
|
addl %esi,%ebx |
1362 |
– |
addl 60(%rsp),%eax |
1343 |
|
xorl %ebp,%edi |
1344 |
+ |
rorl $7,%edx |
1345 |
+ |
addl %ecx,%ebx |
1346 |
+ |
addl 60(%rsp),%eax |
1347 |
+ |
xorl %edx,%edi |
1348 |
|
movl %ebx,%esi |
1349 |
|
roll $5,%ebx |
1366 |
– |
xorl %edx,%edi |
1367 |
– |
addl %ebx,%eax |
1368 |
– |
rorl $7,%ecx |
1350 |
|
addl %edi,%eax |
1351 |
< |
movups %xmm11,48(%r13,%r12,1) |
1351 |
> |
rorl $7,%ecx |
1352 |
> |
addl %ebx,%eax |
1353 |
> |
movups %xmm2,48(%r13,%r12,1) |
1354 |
|
movq 88(%rsp),%r8 |
1355 |
|
|
1356 |
|
addl 0(%r9),%eax |
1363 |
|
movl %ecx,8(%r9) |
1364 |
|
movl %edx,12(%r9) |
1365 |
|
movl %ebp,16(%r9) |
1366 |
< |
movups %xmm11,(%r8) |
1366 |
> |
movups %xmm2,(%r8) |
1367 |
|
leaq 104(%rsp),%rsi |
1368 |
|
movq 0(%rsi),%r15 |
1369 |
|
movq 8(%rsi),%r14 |
1376 |
|
.byte 0xf3,0xc3 |
1377 |
|
.size aesni_cbc_sha1_enc_ssse3,.-aesni_cbc_sha1_enc_ssse3 |
1378 |
|
.type aesni_cbc_sha1_enc_avx,@function |
1379 |
< |
.align 16 |
1379 |
> |
.align 32 |
1380 |
|
aesni_cbc_sha1_enc_avx: |
1381 |
|
movq 8(%rsp),%r10 |
1382 |
|
|
1394 |
|
movq %rdi,%r12 |
1395 |
|
movq %rsi,%r13 |
1396 |
|
movq %rdx,%r14 |
1397 |
< |
movq %rcx,%r15 |
1398 |
< |
vmovdqu (%r8),%xmm11 |
1397 |
> |
leaq 112(%rcx),%r15 |
1398 |
> |
vmovdqu (%r8),%xmm12 |
1399 |
|
movq %r8,88(%rsp) |
1400 |
|
shlq $6,%r14 |
1401 |
|
subq %r12,%r13 |
1402 |
< |
movl 240(%r15),%r8d |
1420 |
< |
addq $112,%r15 |
1402 |
> |
movl 240-112(%r15),%r8d |
1403 |
|
addq %r10,%r14 |
1404 |
|
|
1405 |
|
leaq K_XX_XX(%rip),%r11 |
1409 |
|
movl 12(%r9),%edx |
1410 |
|
movl %ebx,%esi |
1411 |
|
movl 16(%r9),%ebp |
1412 |
+ |
movl %ecx,%edi |
1413 |
+ |
xorl %edx,%edi |
1414 |
+ |
andl %edi,%esi |
1415 |
|
|
1416 |
|
vmovdqa 64(%r11),%xmm6 |
1417 |
< |
vmovdqa 0(%r11),%xmm9 |
1417 |
> |
vmovdqa 0(%r11),%xmm10 |
1418 |
|
vmovdqu 0(%r10),%xmm0 |
1419 |
|
vmovdqu 16(%r10),%xmm1 |
1420 |
|
vmovdqu 32(%r10),%xmm2 |
1424 |
|
vpshufb %xmm6,%xmm1,%xmm1 |
1425 |
|
vpshufb %xmm6,%xmm2,%xmm2 |
1426 |
|
vpshufb %xmm6,%xmm3,%xmm3 |
1427 |
< |
vpaddd %xmm9,%xmm0,%xmm4 |
1428 |
< |
vpaddd %xmm9,%xmm1,%xmm5 |
1429 |
< |
vpaddd %xmm9,%xmm2,%xmm6 |
1427 |
> |
vpaddd %xmm10,%xmm0,%xmm4 |
1428 |
> |
vpaddd %xmm10,%xmm1,%xmm5 |
1429 |
> |
vpaddd %xmm10,%xmm2,%xmm6 |
1430 |
|
vmovdqa %xmm4,0(%rsp) |
1431 |
|
vmovdqa %xmm5,16(%rsp) |
1432 |
|
vmovdqa %xmm6,32(%rsp) |
1433 |
< |
vmovups -112(%r15),%xmm13 |
1433 |
> |
vmovups -112(%r15),%xmm15 |
1434 |
|
vmovups 16-112(%r15),%xmm14 |
1435 |
|
jmp .Loop_avx |
1436 |
< |
.align 16 |
1436 |
> |
.align 32 |
1437 |
|
.Loop_avx: |
1438 |
< |
addl 0(%rsp),%ebp |
1439 |
< |
vmovups 0(%r12),%xmm12 |
1440 |
< |
vxorps %xmm13,%xmm12,%xmm12 |
1441 |
< |
vxorps %xmm12,%xmm11,%xmm11 |
1442 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1438 |
> |
shrdl $2,%ebx,%ebx |
1439 |
> |
vmovdqu 0(%r12),%xmm13 |
1440 |
> |
vpxor %xmm15,%xmm13,%xmm13 |
1441 |
> |
vpxor %xmm13,%xmm12,%xmm12 |
1442 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1443 |
|
vmovups -80(%r15),%xmm15 |
1444 |
< |
xorl %edx,%ecx |
1444 |
> |
xorl %edx,%esi |
1445 |
|
vpalignr $8,%xmm0,%xmm1,%xmm4 |
1446 |
|
movl %eax,%edi |
1447 |
+ |
addl 0(%rsp),%ebp |
1448 |
+ |
vpaddd %xmm3,%xmm10,%xmm9 |
1449 |
+ |
xorl %ecx,%ebx |
1450 |
|
shldl $5,%eax,%eax |
1463 |
– |
vpaddd %xmm3,%xmm9,%xmm9 |
1464 |
– |
andl %ecx,%esi |
1465 |
– |
xorl %edx,%ecx |
1451 |
|
vpsrldq $4,%xmm3,%xmm8 |
1467 |
– |
xorl %edx,%esi |
1468 |
– |
addl %eax,%ebp |
1469 |
– |
vpxor %xmm0,%xmm4,%xmm4 |
1470 |
– |
shrdl $2,%ebx,%ebx |
1452 |
|
addl %esi,%ebp |
1453 |
< |
vpxor %xmm2,%xmm8,%xmm8 |
1454 |
< |
addl 4(%rsp),%edx |
1453 |
> |
andl %ebx,%edi |
1454 |
> |
vpxor %xmm0,%xmm4,%xmm4 |
1455 |
|
xorl %ecx,%ebx |
1456 |
+ |
addl %eax,%ebp |
1457 |
+ |
vpxor %xmm2,%xmm8,%xmm8 |
1458 |
+ |
shrdl $7,%eax,%eax |
1459 |
+ |
xorl %ecx,%edi |
1460 |
|
movl %ebp,%esi |
1461 |
< |
shldl $5,%ebp,%ebp |
1461 |
> |
addl 4(%rsp),%edx |
1462 |
|
vpxor %xmm8,%xmm4,%xmm4 |
1463 |
< |
andl %ebx,%edi |
1464 |
< |
xorl %ecx,%ebx |
1463 |
> |
xorl %ebx,%eax |
1464 |
> |
shldl $5,%ebp,%ebp |
1465 |
|
vmovdqa %xmm9,48(%rsp) |
1466 |
< |
xorl %ecx,%edi |
1467 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1466 |
> |
addl %edi,%edx |
1467 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1468 |
|
vmovups -64(%r15),%xmm14 |
1469 |
< |
addl %ebp,%edx |
1469 |
> |
andl %eax,%esi |
1470 |
|
vpsrld $31,%xmm4,%xmm8 |
1486 |
– |
shrdl $7,%eax,%eax |
1487 |
– |
addl %edi,%edx |
1488 |
– |
addl 8(%rsp),%ecx |
1471 |
|
xorl %ebx,%eax |
1472 |
< |
vpslldq $12,%xmm4,%xmm10 |
1472 |
> |
addl %ebp,%edx |
1473 |
> |
shrdl $7,%ebp,%ebp |
1474 |
> |
xorl %ebx,%esi |
1475 |
> |
vpslldq $12,%xmm4,%xmm9 |
1476 |
|
vpaddd %xmm4,%xmm4,%xmm4 |
1477 |
|
movl %edx,%edi |
1478 |
+ |
addl 8(%rsp),%ecx |
1479 |
+ |
xorl %eax,%ebp |
1480 |
|
shldl $5,%edx,%edx |
1494 |
– |
andl %eax,%esi |
1495 |
– |
xorl %ebx,%eax |
1496 |
– |
vpsrld $30,%xmm10,%xmm9 |
1481 |
|
vpor %xmm8,%xmm4,%xmm4 |
1482 |
< |
xorl %ebx,%esi |
1499 |
< |
addl %edx,%ecx |
1500 |
< |
shrdl $7,%ebp,%ebp |
1482 |
> |
vpsrld $30,%xmm9,%xmm8 |
1483 |
|
addl %esi,%ecx |
1502 |
– |
vpslld $2,%xmm10,%xmm10 |
1503 |
– |
vpxor %xmm9,%xmm4,%xmm4 |
1504 |
– |
addl 12(%rsp),%ebx |
1505 |
– |
xorl %eax,%ebp |
1506 |
– |
movl %ecx,%esi |
1507 |
– |
shldl $5,%ecx,%ecx |
1508 |
– |
vaesenc %xmm14,%xmm11,%xmm11 |
1509 |
– |
vmovups -48(%r15),%xmm15 |
1510 |
– |
vpxor %xmm10,%xmm4,%xmm4 |
1484 |
|
andl %ebp,%edi |
1485 |
|
xorl %eax,%ebp |
1486 |
< |
vmovdqa 0(%r11),%xmm10 |
1487 |
< |
xorl %eax,%edi |
1488 |
< |
addl %ecx,%ebx |
1486 |
> |
addl %edx,%ecx |
1487 |
> |
vpslld $2,%xmm9,%xmm9 |
1488 |
> |
vpxor %xmm8,%xmm4,%xmm4 |
1489 |
|
shrdl $7,%edx,%edx |
1490 |
+ |
xorl %eax,%edi |
1491 |
+ |
movl %ecx,%esi |
1492 |
+ |
addl 12(%rsp),%ebx |
1493 |
+ |
vaesenc %xmm14,%xmm12,%xmm12 |
1494 |
+ |
vmovups -48(%r15),%xmm15 |
1495 |
+ |
vpxor %xmm9,%xmm4,%xmm4 |
1496 |
+ |
xorl %ebp,%edx |
1497 |
+ |
shldl $5,%ecx,%ecx |
1498 |
|
addl %edi,%ebx |
1499 |
< |
addl 16(%rsp),%eax |
1499 |
> |
andl %edx,%esi |
1500 |
|
xorl %ebp,%edx |
1501 |
+ |
addl %ecx,%ebx |
1502 |
+ |
shrdl $7,%ecx,%ecx |
1503 |
+ |
xorl %ebp,%esi |
1504 |
|
vpalignr $8,%xmm1,%xmm2,%xmm5 |
1505 |
|
movl %ebx,%edi |
1506 |
+ |
addl 16(%rsp),%eax |
1507 |
+ |
vpaddd %xmm4,%xmm10,%xmm9 |
1508 |
+ |
xorl %edx,%ecx |
1509 |
|
shldl $5,%ebx,%ebx |
1510 |
< |
vpaddd %xmm4,%xmm10,%xmm10 |
1524 |
< |
andl %edx,%esi |
1525 |
< |
xorl %ebp,%edx |
1526 |
< |
vpsrldq $4,%xmm4,%xmm9 |
1527 |
< |
xorl %ebp,%esi |
1528 |
< |
addl %ebx,%eax |
1529 |
< |
vpxor %xmm1,%xmm5,%xmm5 |
1530 |
< |
shrdl $7,%ecx,%ecx |
1510 |
> |
vpsrldq $4,%xmm4,%xmm8 |
1511 |
|
addl %esi,%eax |
1532 |
– |
vpxor %xmm3,%xmm9,%xmm9 |
1533 |
– |
addl 20(%rsp),%ebp |
1534 |
– |
vaesenc %xmm15,%xmm11,%xmm11 |
1535 |
– |
vmovups -32(%r15),%xmm14 |
1536 |
– |
xorl %edx,%ecx |
1537 |
– |
movl %eax,%esi |
1538 |
– |
shldl $5,%eax,%eax |
1539 |
– |
vpxor %xmm9,%xmm5,%xmm5 |
1512 |
|
andl %ecx,%edi |
1513 |
+ |
vpxor %xmm1,%xmm5,%xmm5 |
1514 |
|
xorl %edx,%ecx |
1515 |
< |
vmovdqa %xmm10,0(%rsp) |
1516 |
< |
xorl %edx,%edi |
1544 |
< |
addl %eax,%ebp |
1545 |
< |
vpsrld $31,%xmm5,%xmm9 |
1515 |
> |
addl %ebx,%eax |
1516 |
> |
vpxor %xmm3,%xmm8,%xmm8 |
1517 |
|
shrdl $7,%ebx,%ebx |
1518 |
+ |
vaesenc %xmm15,%xmm12,%xmm12 |
1519 |
+ |
vmovups -32(%r15),%xmm14 |
1520 |
+ |
xorl %edx,%edi |
1521 |
+ |
movl %eax,%esi |
1522 |
+ |
addl 20(%rsp),%ebp |
1523 |
+ |
vpxor %xmm8,%xmm5,%xmm5 |
1524 |
+ |
xorl %ecx,%ebx |
1525 |
+ |
shldl $5,%eax,%eax |
1526 |
+ |
vmovdqa %xmm9,0(%rsp) |
1527 |
|
addl %edi,%ebp |
1528 |
< |
addl 24(%rsp),%edx |
1528 |
> |
andl %ebx,%esi |
1529 |
> |
vpsrld $31,%xmm5,%xmm8 |
1530 |
|
xorl %ecx,%ebx |
1531 |
< |
vpslldq $12,%xmm5,%xmm8 |
1531 |
> |
addl %eax,%ebp |
1532 |
> |
shrdl $7,%eax,%eax |
1533 |
> |
xorl %ecx,%esi |
1534 |
> |
vpslldq $12,%xmm5,%xmm9 |
1535 |
|
vpaddd %xmm5,%xmm5,%xmm5 |
1536 |
|
movl %ebp,%edi |
1537 |
+ |
addl 24(%rsp),%edx |
1538 |
+ |
xorl %ebx,%eax |
1539 |
|
shldl $5,%ebp,%ebp |
1540 |
< |
andl %ebx,%esi |
1541 |
< |
xorl %ecx,%ebx |
1556 |
< |
vpsrld $30,%xmm8,%xmm10 |
1557 |
< |
vpor %xmm9,%xmm5,%xmm5 |
1558 |
< |
xorl %ecx,%esi |
1559 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1560 |
< |
vmovups -16(%r15),%xmm15 |
1561 |
< |
addl %ebp,%edx |
1562 |
< |
shrdl $7,%eax,%eax |
1540 |
> |
vpor %xmm8,%xmm5,%xmm5 |
1541 |
> |
vpsrld $30,%xmm9,%xmm8 |
1542 |
|
addl %esi,%edx |
1543 |
< |
vpslld $2,%xmm8,%xmm8 |
1544 |
< |
vpxor %xmm10,%xmm5,%xmm5 |
1566 |
< |
addl 28(%rsp),%ecx |
1567 |
< |
xorl %ebx,%eax |
1568 |
< |
movl %edx,%esi |
1569 |
< |
shldl $5,%edx,%edx |
1570 |
< |
vpxor %xmm8,%xmm5,%xmm5 |
1543 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1544 |
> |
vmovups -16(%r15),%xmm15 |
1545 |
|
andl %eax,%edi |
1546 |
|
xorl %ebx,%eax |
1547 |
< |
vmovdqa 16(%r11),%xmm8 |
1548 |
< |
xorl %ebx,%edi |
1549 |
< |
addl %edx,%ecx |
1547 |
> |
addl %ebp,%edx |
1548 |
> |
vpslld $2,%xmm9,%xmm9 |
1549 |
> |
vpxor %xmm8,%xmm5,%xmm5 |
1550 |
|
shrdl $7,%ebp,%ebp |
1551 |
+ |
xorl %ebx,%edi |
1552 |
+ |
movl %edx,%esi |
1553 |
+ |
addl 28(%rsp),%ecx |
1554 |
+ |
vpxor %xmm9,%xmm5,%xmm5 |
1555 |
+ |
xorl %eax,%ebp |
1556 |
+ |
shldl $5,%edx,%edx |
1557 |
+ |
vmovdqa 16(%r11),%xmm10 |
1558 |
|
addl %edi,%ecx |
1559 |
< |
addl 32(%rsp),%ebx |
1559 |
> |
andl %ebp,%esi |
1560 |
|
xorl %eax,%ebp |
1561 |
+ |
addl %edx,%ecx |
1562 |
+ |
shrdl $7,%edx,%edx |
1563 |
+ |
xorl %eax,%esi |
1564 |
|
vpalignr $8,%xmm2,%xmm3,%xmm6 |
1565 |
|
movl %ecx,%edi |
1566 |
< |
shldl $5,%ecx,%ecx |
1567 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1566 |
> |
addl 32(%rsp),%ebx |
1567 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1568 |
|
vmovups 0(%r15),%xmm14 |
1569 |
< |
vpaddd %xmm5,%xmm8,%xmm8 |
1586 |
< |
andl %ebp,%esi |
1587 |
< |
xorl %eax,%ebp |
1588 |
< |
vpsrldq $4,%xmm5,%xmm10 |
1589 |
< |
xorl %eax,%esi |
1590 |
< |
addl %ecx,%ebx |
1591 |
< |
vpxor %xmm2,%xmm6,%xmm6 |
1592 |
< |
shrdl $7,%edx,%edx |
1593 |
< |
addl %esi,%ebx |
1594 |
< |
vpxor %xmm4,%xmm10,%xmm10 |
1595 |
< |
addl 36(%rsp),%eax |
1569 |
> |
vpaddd %xmm5,%xmm10,%xmm9 |
1570 |
|
xorl %ebp,%edx |
1571 |
< |
movl %ebx,%esi |
1572 |
< |
shldl $5,%ebx,%ebx |
1573 |
< |
vpxor %xmm10,%xmm6,%xmm6 |
1571 |
> |
shldl $5,%ecx,%ecx |
1572 |
> |
vpsrldq $4,%xmm5,%xmm8 |
1573 |
> |
addl %esi,%ebx |
1574 |
|
andl %edx,%edi |
1575 |
+ |
vpxor %xmm2,%xmm6,%xmm6 |
1576 |
|
xorl %ebp,%edx |
1577 |
< |
vmovdqa %xmm8,16(%rsp) |
1578 |
< |
xorl %ebp,%edi |
1604 |
< |
addl %ebx,%eax |
1605 |
< |
vpsrld $31,%xmm6,%xmm10 |
1577 |
> |
addl %ecx,%ebx |
1578 |
> |
vpxor %xmm4,%xmm8,%xmm8 |
1579 |
|
shrdl $7,%ecx,%ecx |
1580 |
+ |
xorl %ebp,%edi |
1581 |
+ |
movl %ebx,%esi |
1582 |
+ |
addl 36(%rsp),%eax |
1583 |
+ |
vpxor %xmm8,%xmm6,%xmm6 |
1584 |
+ |
xorl %edx,%ecx |
1585 |
+ |
shldl $5,%ebx,%ebx |
1586 |
+ |
vmovdqa %xmm9,16(%rsp) |
1587 |
|
addl %edi,%eax |
1588 |
< |
addl 40(%rsp),%ebp |
1589 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1610 |
< |
vmovups 16(%r15),%xmm15 |
1588 |
> |
andl %ecx,%esi |
1589 |
> |
vpsrld $31,%xmm6,%xmm8 |
1590 |
|
xorl %edx,%ecx |
1591 |
+ |
addl %ebx,%eax |
1592 |
+ |
shrdl $7,%ebx,%ebx |
1593 |
+ |
vaesenc %xmm14,%xmm12,%xmm12 |
1594 |
+ |
vmovups 16(%r15),%xmm15 |
1595 |
+ |
xorl %edx,%esi |
1596 |
|
vpslldq $12,%xmm6,%xmm9 |
1597 |
|
vpaddd %xmm6,%xmm6,%xmm6 |
1598 |
|
movl %eax,%edi |
1599 |
+ |
addl 40(%rsp),%ebp |
1600 |
+ |
xorl %ecx,%ebx |
1601 |
|
shldl $5,%eax,%eax |
1602 |
< |
andl %ecx,%esi |
1617 |
< |
xorl %edx,%ecx |
1602 |
> |
vpor %xmm8,%xmm6,%xmm6 |
1603 |
|
vpsrld $30,%xmm9,%xmm8 |
1619 |
– |
vpor %xmm10,%xmm6,%xmm6 |
1620 |
– |
xorl %edx,%esi |
1621 |
– |
addl %eax,%ebp |
1622 |
– |
shrdl $7,%ebx,%ebx |
1604 |
|
addl %esi,%ebp |
1605 |
+ |
andl %ebx,%edi |
1606 |
+ |
xorl %ecx,%ebx |
1607 |
+ |
addl %eax,%ebp |
1608 |
|
vpslld $2,%xmm9,%xmm9 |
1609 |
|
vpxor %xmm8,%xmm6,%xmm6 |
1610 |
< |
addl 44(%rsp),%edx |
1611 |
< |
xorl %ecx,%ebx |
1610 |
> |
shrdl $7,%eax,%eax |
1611 |
> |
xorl %ecx,%edi |
1612 |
|
movl %ebp,%esi |
1613 |
< |
shldl $5,%ebp,%ebp |
1613 |
> |
addl 44(%rsp),%edx |
1614 |
|
vpxor %xmm9,%xmm6,%xmm6 |
1615 |
< |
andl %ebx,%edi |
1616 |
< |
xorl %ecx,%ebx |
1633 |
< |
vmovdqa 16(%r11),%xmm9 |
1634 |
< |
xorl %ecx,%edi |
1635 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1636 |
< |
vmovups 32(%r15),%xmm14 |
1637 |
< |
addl %ebp,%edx |
1638 |
< |
shrdl $7,%eax,%eax |
1615 |
> |
xorl %ebx,%eax |
1616 |
> |
shldl $5,%ebp,%ebp |
1617 |
|
addl %edi,%edx |
1618 |
< |
addl 48(%rsp),%ecx |
1618 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1619 |
> |
vmovups 32(%r15),%xmm14 |
1620 |
> |
andl %eax,%esi |
1621 |
|
xorl %ebx,%eax |
1622 |
+ |
addl %ebp,%edx |
1623 |
+ |
shrdl $7,%ebp,%ebp |
1624 |
+ |
xorl %ebx,%esi |
1625 |
|
vpalignr $8,%xmm3,%xmm4,%xmm7 |
1626 |
|
movl %edx,%edi |
1627 |
+ |
addl 48(%rsp),%ecx |
1628 |
+ |
vpaddd %xmm6,%xmm10,%xmm9 |
1629 |
+ |
xorl %eax,%ebp |
1630 |
|
shldl $5,%edx,%edx |
1645 |
– |
vpaddd %xmm6,%xmm9,%xmm9 |
1646 |
– |
andl %eax,%esi |
1647 |
– |
xorl %ebx,%eax |
1631 |
|
vpsrldq $4,%xmm6,%xmm8 |
1649 |
– |
xorl %ebx,%esi |
1650 |
– |
addl %edx,%ecx |
1651 |
– |
vpxor %xmm3,%xmm7,%xmm7 |
1652 |
– |
shrdl $7,%ebp,%ebp |
1632 |
|
addl %esi,%ecx |
1633 |
< |
vpxor %xmm5,%xmm8,%xmm8 |
1634 |
< |
addl 52(%rsp),%ebx |
1633 |
> |
andl %ebp,%edi |
1634 |
> |
vpxor %xmm3,%xmm7,%xmm7 |
1635 |
|
xorl %eax,%ebp |
1636 |
+ |
addl %edx,%ecx |
1637 |
+ |
vpxor %xmm5,%xmm8,%xmm8 |
1638 |
+ |
shrdl $7,%edx,%edx |
1639 |
+ |
xorl %eax,%edi |
1640 |
|
movl %ecx,%esi |
1641 |
< |
shldl $5,%ecx,%ecx |
1642 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1641 |
> |
addl 52(%rsp),%ebx |
1642 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1643 |
|
vmovups 48(%r15),%xmm15 |
1644 |
|
vpxor %xmm8,%xmm7,%xmm7 |
1645 |
< |
andl %ebp,%edi |
1646 |
< |
xorl %eax,%ebp |
1645 |
> |
xorl %ebp,%edx |
1646 |
> |
shldl $5,%ecx,%ecx |
1647 |
|
vmovdqa %xmm9,32(%rsp) |
1665 |
– |
xorl %eax,%edi |
1666 |
– |
addl %ecx,%ebx |
1667 |
– |
vpsrld $31,%xmm7,%xmm8 |
1668 |
– |
shrdl $7,%edx,%edx |
1648 |
|
addl %edi,%ebx |
1649 |
< |
addl 56(%rsp),%eax |
1649 |
> |
andl %edx,%esi |
1650 |
> |
vpsrld $31,%xmm7,%xmm8 |
1651 |
|
xorl %ebp,%edx |
1652 |
< |
vpslldq $12,%xmm7,%xmm10 |
1652 |
> |
addl %ecx,%ebx |
1653 |
> |
shrdl $7,%ecx,%ecx |
1654 |
> |
xorl %ebp,%esi |
1655 |
> |
vpslldq $12,%xmm7,%xmm9 |
1656 |
|
vpaddd %xmm7,%xmm7,%xmm7 |
1657 |
|
movl %ebx,%edi |
1658 |
+ |
addl 56(%rsp),%eax |
1659 |
+ |
xorl %edx,%ecx |
1660 |
|
shldl $5,%ebx,%ebx |
1676 |
– |
andl %edx,%esi |
1677 |
– |
xorl %ebp,%edx |
1678 |
– |
vpsrld $30,%xmm10,%xmm9 |
1661 |
|
vpor %xmm8,%xmm7,%xmm7 |
1662 |
< |
xorl %ebp,%esi |
1681 |
< |
addl %ebx,%eax |
1682 |
< |
shrdl $7,%ecx,%ecx |
1662 |
> |
vpsrld $30,%xmm9,%xmm8 |
1663 |
|
addl %esi,%eax |
1664 |
< |
vpslld $2,%xmm10,%xmm10 |
1665 |
< |
vpxor %xmm9,%xmm7,%xmm7 |
1666 |
< |
addl 60(%rsp),%ebp |
1664 |
> |
andl %ecx,%edi |
1665 |
> |
xorl %edx,%ecx |
1666 |
> |
addl %ebx,%eax |
1667 |
> |
vpslld $2,%xmm9,%xmm9 |
1668 |
> |
vpxor %xmm8,%xmm7,%xmm7 |
1669 |
> |
shrdl $7,%ebx,%ebx |
1670 |
|
cmpl $11,%r8d |
1671 |
< |
jb .Lvaesenclast1 |
1672 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1671 |
> |
jb .Lvaesenclast6 |
1672 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1673 |
|
vmovups 64(%r15),%xmm14 |
1674 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1674 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1675 |
|
vmovups 80(%r15),%xmm15 |
1676 |
< |
je .Lvaesenclast1 |
1677 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1676 |
> |
je .Lvaesenclast6 |
1677 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1678 |
|
vmovups 96(%r15),%xmm14 |
1679 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1679 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1680 |
|
vmovups 112(%r15),%xmm15 |
1681 |
< |
.Lvaesenclast1: |
1682 |
< |
vaesenclast %xmm15,%xmm11,%xmm11 |
1681 |
> |
.Lvaesenclast6: |
1682 |
> |
vaesenclast %xmm15,%xmm12,%xmm12 |
1683 |
> |
vmovups -112(%r15),%xmm15 |
1684 |
|
vmovups 16-112(%r15),%xmm14 |
1685 |
< |
xorl %edx,%ecx |
1685 |
> |
xorl %edx,%edi |
1686 |
|
movl %eax,%esi |
1687 |
+ |
addl 60(%rsp),%ebp |
1688 |
+ |
vpxor %xmm9,%xmm7,%xmm7 |
1689 |
+ |
xorl %ecx,%ebx |
1690 |
|
shldl $5,%eax,%eax |
1704 |
– |
vpxor %xmm10,%xmm7,%xmm7 |
1705 |
– |
andl %ecx,%edi |
1706 |
– |
xorl %edx,%ecx |
1707 |
– |
vmovdqa 16(%r11),%xmm10 |
1708 |
– |
xorl %edx,%edi |
1709 |
– |
addl %eax,%ebp |
1710 |
– |
shrdl $7,%ebx,%ebx |
1691 |
|
addl %edi,%ebp |
1712 |
– |
vpalignr $8,%xmm6,%xmm7,%xmm9 |
1713 |
– |
vpxor %xmm4,%xmm0,%xmm0 |
1714 |
– |
addl 0(%rsp),%edx |
1715 |
– |
xorl %ecx,%ebx |
1716 |
– |
movl %ebp,%edi |
1717 |
– |
shldl $5,%ebp,%ebp |
1718 |
– |
vpxor %xmm1,%xmm0,%xmm0 |
1692 |
|
andl %ebx,%esi |
1693 |
|
xorl %ecx,%ebx |
1694 |
< |
vmovdqa %xmm10,%xmm8 |
1695 |
< |
vpaddd %xmm7,%xmm10,%xmm10 |
1696 |
< |
xorl %ecx,%esi |
1724 |
< |
vmovups 16(%r12),%xmm12 |
1725 |
< |
vxorps %xmm13,%xmm12,%xmm12 |
1726 |
< |
vmovups %xmm11,0(%r13,%r12,1) |
1727 |
< |
vxorps %xmm12,%xmm11,%xmm11 |
1728 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1729 |
< |
vmovups -80(%r15),%xmm15 |
1730 |
< |
addl %ebp,%edx |
1731 |
< |
vpxor %xmm9,%xmm0,%xmm0 |
1694 |
> |
addl %eax,%ebp |
1695 |
> |
vpalignr $8,%xmm6,%xmm7,%xmm8 |
1696 |
> |
vpxor %xmm4,%xmm0,%xmm0 |
1697 |
|
shrdl $7,%eax,%eax |
1698 |
+ |
xorl %ecx,%esi |
1699 |
+ |
movl %ebp,%edi |
1700 |
+ |
addl 0(%rsp),%edx |
1701 |
+ |
vpxor %xmm1,%xmm0,%xmm0 |
1702 |
+ |
xorl %ebx,%eax |
1703 |
+ |
shldl $5,%ebp,%ebp |
1704 |
+ |
vpaddd %xmm7,%xmm10,%xmm9 |
1705 |
|
addl %esi,%edx |
1706 |
< |
addl 4(%rsp),%ecx |
1706 |
> |
vmovdqu 16(%r12),%xmm13 |
1707 |
> |
vpxor %xmm15,%xmm13,%xmm13 |
1708 |
> |
vmovups %xmm12,0(%r12,%r13,1) |
1709 |
> |
vpxor %xmm13,%xmm12,%xmm12 |
1710 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1711 |
> |
vmovups -80(%r15),%xmm15 |
1712 |
> |
andl %eax,%edi |
1713 |
> |
vpxor %xmm8,%xmm0,%xmm0 |
1714 |
|
xorl %ebx,%eax |
1715 |
< |
vpsrld $30,%xmm0,%xmm9 |
1716 |
< |
vmovdqa %xmm10,48(%rsp) |
1715 |
> |
addl %ebp,%edx |
1716 |
> |
shrdl $7,%ebp,%ebp |
1717 |
> |
xorl %ebx,%edi |
1718 |
> |
vpsrld $30,%xmm0,%xmm8 |
1719 |
> |
vmovdqa %xmm9,48(%rsp) |
1720 |
|
movl %edx,%esi |
1721 |
+ |
addl 4(%rsp),%ecx |
1722 |
+ |
xorl %eax,%ebp |
1723 |
|
shldl $5,%edx,%edx |
1740 |
– |
andl %eax,%edi |
1741 |
– |
xorl %ebx,%eax |
1724 |
|
vpslld $2,%xmm0,%xmm0 |
1743 |
– |
xorl %ebx,%edi |
1744 |
– |
addl %edx,%ecx |
1745 |
– |
shrdl $7,%ebp,%ebp |
1725 |
|
addl %edi,%ecx |
1747 |
– |
addl 8(%rsp),%ebx |
1748 |
– |
xorl %eax,%ebp |
1749 |
– |
movl %ecx,%edi |
1750 |
– |
shldl $5,%ecx,%ecx |
1751 |
– |
vaesenc %xmm15,%xmm11,%xmm11 |
1752 |
– |
vmovups -64(%r15),%xmm14 |
1753 |
– |
vpor %xmm9,%xmm0,%xmm0 |
1726 |
|
andl %ebp,%esi |
1727 |
|
xorl %eax,%ebp |
1728 |
< |
vmovdqa %xmm0,%xmm10 |
1757 |
< |
xorl %eax,%esi |
1758 |
< |
addl %ecx,%ebx |
1728 |
> |
addl %edx,%ecx |
1729 |
|
shrdl $7,%edx,%edx |
1730 |
< |
addl %esi,%ebx |
1731 |
< |
addl 12(%rsp),%eax |
1730 |
> |
xorl %eax,%esi |
1731 |
> |
movl %ecx,%edi |
1732 |
> |
addl 8(%rsp),%ebx |
1733 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1734 |
> |
vmovups -64(%r15),%xmm14 |
1735 |
> |
vpor %xmm8,%xmm0,%xmm0 |
1736 |
|
xorl %ebp,%edx |
1737 |
< |
movl %ebx,%esi |
1738 |
< |
shldl $5,%ebx,%ebx |
1737 |
> |
shldl $5,%ecx,%ecx |
1738 |
> |
addl %esi,%ebx |
1739 |
|
andl %edx,%edi |
1740 |
|
xorl %ebp,%edx |
1741 |
+ |
addl %ecx,%ebx |
1742 |
+ |
addl 12(%rsp),%eax |
1743 |
|
xorl %ebp,%edi |
1744 |
< |
addl %ebx,%eax |
1745 |
< |
shrdl $7,%ecx,%ecx |
1744 |
> |
movl %ebx,%esi |
1745 |
> |
shldl $5,%ebx,%ebx |
1746 |
|
addl %edi,%eax |
1747 |
< |
vpalignr $8,%xmm7,%xmm0,%xmm10 |
1747 |
> |
xorl %edx,%esi |
1748 |
> |
shrdl $7,%ecx,%ecx |
1749 |
> |
addl %ebx,%eax |
1750 |
> |
vpalignr $8,%xmm7,%xmm0,%xmm8 |
1751 |
|
vpxor %xmm5,%xmm1,%xmm1 |
1752 |
|
addl 16(%rsp),%ebp |
1753 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1753 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1754 |
|
vmovups -48(%r15),%xmm15 |
1755 |
< |
xorl %edx,%esi |
1755 |
> |
xorl %ecx,%esi |
1756 |
|
movl %eax,%edi |
1757 |
|
shldl $5,%eax,%eax |
1758 |
|
vpxor %xmm2,%xmm1,%xmm1 |
1780 |
– |
xorl %ecx,%esi |
1781 |
– |
addl %eax,%ebp |
1782 |
– |
vmovdqa %xmm8,%xmm9 |
1783 |
– |
vpaddd %xmm0,%xmm8,%xmm8 |
1784 |
– |
shrdl $7,%ebx,%ebx |
1759 |
|
addl %esi,%ebp |
1786 |
– |
vpxor %xmm10,%xmm1,%xmm1 |
1787 |
– |
addl 20(%rsp),%edx |
1760 |
|
xorl %ecx,%edi |
1761 |
+ |
vpaddd %xmm0,%xmm10,%xmm9 |
1762 |
+ |
shrdl $7,%ebx,%ebx |
1763 |
+ |
addl %eax,%ebp |
1764 |
+ |
vpxor %xmm8,%xmm1,%xmm1 |
1765 |
+ |
addl 20(%rsp),%edx |
1766 |
+ |
xorl %ebx,%edi |
1767 |
|
movl %ebp,%esi |
1768 |
|
shldl $5,%ebp,%ebp |
1769 |
< |
vpsrld $30,%xmm1,%xmm10 |
1770 |
< |
vmovdqa %xmm8,0(%rsp) |
1793 |
< |
xorl %ebx,%edi |
1794 |
< |
addl %ebp,%edx |
1795 |
< |
shrdl $7,%eax,%eax |
1769 |
> |
vpsrld $30,%xmm1,%xmm8 |
1770 |
> |
vmovdqa %xmm9,0(%rsp) |
1771 |
|
addl %edi,%edx |
1772 |
+ |
xorl %ebx,%esi |
1773 |
+ |
shrdl $7,%eax,%eax |
1774 |
+ |
addl %ebp,%edx |
1775 |
|
vpslld $2,%xmm1,%xmm1 |
1776 |
|
addl 24(%rsp),%ecx |
1777 |
< |
xorl %ebx,%esi |
1777 |
> |
xorl %eax,%esi |
1778 |
|
movl %edx,%edi |
1779 |
|
shldl $5,%edx,%edx |
1780 |
< |
xorl %eax,%esi |
1781 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1780 |
> |
addl %esi,%ecx |
1781 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1782 |
|
vmovups -32(%r15),%xmm14 |
1783 |
< |
addl %edx,%ecx |
1783 |
> |
xorl %eax,%edi |
1784 |
|
shrdl $7,%ebp,%ebp |
1785 |
< |
addl %esi,%ecx |
1786 |
< |
vpor %xmm10,%xmm1,%xmm1 |
1785 |
> |
addl %edx,%ecx |
1786 |
> |
vpor %xmm8,%xmm1,%xmm1 |
1787 |
|
addl 28(%rsp),%ebx |
1788 |
< |
xorl %eax,%edi |
1811 |
< |
vmovdqa %xmm1,%xmm8 |
1788 |
> |
xorl %ebp,%edi |
1789 |
|
movl %ecx,%esi |
1790 |
|
shldl $5,%ecx,%ecx |
1814 |
– |
xorl %ebp,%edi |
1815 |
– |
addl %ecx,%ebx |
1816 |
– |
shrdl $7,%edx,%edx |
1791 |
|
addl %edi,%ebx |
1792 |
+ |
xorl %ebp,%esi |
1793 |
+ |
shrdl $7,%edx,%edx |
1794 |
+ |
addl %ecx,%ebx |
1795 |
|
vpalignr $8,%xmm0,%xmm1,%xmm8 |
1796 |
|
vpxor %xmm6,%xmm2,%xmm2 |
1797 |
|
addl 32(%rsp),%eax |
1798 |
< |
xorl %ebp,%esi |
1798 |
> |
xorl %edx,%esi |
1799 |
|
movl %ebx,%edi |
1800 |
|
shldl $5,%ebx,%ebx |
1801 |
|
vpxor %xmm3,%xmm2,%xmm2 |
1802 |
< |
xorl %edx,%esi |
1803 |
< |
addl %ebx,%eax |
1802 |
> |
addl %esi,%eax |
1803 |
> |
xorl %edx,%edi |
1804 |
> |
vpaddd %xmm1,%xmm10,%xmm9 |
1805 |
|
vmovdqa 32(%r11),%xmm10 |
1828 |
– |
vpaddd %xmm1,%xmm9,%xmm9 |
1806 |
|
shrdl $7,%ecx,%ecx |
1807 |
< |
addl %esi,%eax |
1807 |
> |
addl %ebx,%eax |
1808 |
|
vpxor %xmm8,%xmm2,%xmm2 |
1809 |
|
addl 36(%rsp),%ebp |
1810 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1810 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1811 |
|
vmovups -16(%r15),%xmm15 |
1812 |
< |
xorl %edx,%edi |
1812 |
> |
xorl %ecx,%edi |
1813 |
|
movl %eax,%esi |
1814 |
|
shldl $5,%eax,%eax |
1815 |
|
vpsrld $30,%xmm2,%xmm8 |
1816 |
|
vmovdqa %xmm9,16(%rsp) |
1840 |
– |
xorl %ecx,%edi |
1841 |
– |
addl %eax,%ebp |
1842 |
– |
shrdl $7,%ebx,%ebx |
1817 |
|
addl %edi,%ebp |
1818 |
+ |
xorl %ecx,%esi |
1819 |
+ |
shrdl $7,%ebx,%ebx |
1820 |
+ |
addl %eax,%ebp |
1821 |
|
vpslld $2,%xmm2,%xmm2 |
1822 |
|
addl 40(%rsp),%edx |
1823 |
< |
xorl %ecx,%esi |
1823 |
> |
xorl %ebx,%esi |
1824 |
|
movl %ebp,%edi |
1825 |
|
shldl $5,%ebp,%ebp |
1849 |
– |
xorl %ebx,%esi |
1850 |
– |
addl %ebp,%edx |
1851 |
– |
shrdl $7,%eax,%eax |
1826 |
|
addl %esi,%edx |
1827 |
+ |
xorl %ebx,%edi |
1828 |
+ |
shrdl $7,%eax,%eax |
1829 |
+ |
addl %ebp,%edx |
1830 |
|
vpor %xmm8,%xmm2,%xmm2 |
1831 |
|
addl 44(%rsp),%ecx |
1832 |
< |
xorl %ebx,%edi |
1856 |
< |
vmovdqa %xmm2,%xmm9 |
1832 |
> |
xorl %eax,%edi |
1833 |
|
movl %edx,%esi |
1834 |
|
shldl $5,%edx,%edx |
1835 |
< |
xorl %eax,%edi |
1836 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1835 |
> |
addl %edi,%ecx |
1836 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1837 |
|
vmovups 0(%r15),%xmm14 |
1838 |
< |
addl %edx,%ecx |
1838 |
> |
xorl %eax,%esi |
1839 |
|
shrdl $7,%ebp,%ebp |
1840 |
< |
addl %edi,%ecx |
1841 |
< |
vpalignr $8,%xmm1,%xmm2,%xmm9 |
1840 |
> |
addl %edx,%ecx |
1841 |
> |
vpalignr $8,%xmm1,%xmm2,%xmm8 |
1842 |
|
vpxor %xmm7,%xmm3,%xmm3 |
1843 |
|
addl 48(%rsp),%ebx |
1844 |
< |
xorl %eax,%esi |
1844 |
> |
xorl %ebp,%esi |
1845 |
|
movl %ecx,%edi |
1846 |
|
shldl $5,%ecx,%ecx |
1847 |
|
vpxor %xmm4,%xmm3,%xmm3 |
1872 |
– |
xorl %ebp,%esi |
1873 |
– |
addl %ecx,%ebx |
1874 |
– |
vmovdqa %xmm10,%xmm8 |
1875 |
– |
vpaddd %xmm2,%xmm10,%xmm10 |
1876 |
– |
shrdl $7,%edx,%edx |
1848 |
|
addl %esi,%ebx |
1878 |
– |
vpxor %xmm9,%xmm3,%xmm3 |
1879 |
– |
addl 52(%rsp),%eax |
1849 |
|
xorl %ebp,%edi |
1850 |
+ |
vpaddd %xmm2,%xmm10,%xmm9 |
1851 |
+ |
shrdl $7,%edx,%edx |
1852 |
+ |
addl %ecx,%ebx |
1853 |
+ |
vpxor %xmm8,%xmm3,%xmm3 |
1854 |
+ |
addl 52(%rsp),%eax |
1855 |
+ |
xorl %edx,%edi |
1856 |
|
movl %ebx,%esi |
1857 |
|
shldl $5,%ebx,%ebx |
1858 |
< |
vpsrld $30,%xmm3,%xmm9 |
1859 |
< |
vmovdqa %xmm10,32(%rsp) |
1885 |
< |
xorl %edx,%edi |
1886 |
< |
addl %ebx,%eax |
1887 |
< |
shrdl $7,%ecx,%ecx |
1858 |
> |
vpsrld $30,%xmm3,%xmm8 |
1859 |
> |
vmovdqa %xmm9,32(%rsp) |
1860 |
|
addl %edi,%eax |
1861 |
+ |
xorl %edx,%esi |
1862 |
+ |
shrdl $7,%ecx,%ecx |
1863 |
+ |
addl %ebx,%eax |
1864 |
|
vpslld $2,%xmm3,%xmm3 |
1865 |
|
addl 56(%rsp),%ebp |
1866 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1866 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1867 |
|
vmovups 16(%r15),%xmm15 |
1868 |
< |
xorl %edx,%esi |
1868 |
> |
xorl %ecx,%esi |
1869 |
|
movl %eax,%edi |
1870 |
|
shldl $5,%eax,%eax |
1896 |
– |
xorl %ecx,%esi |
1897 |
– |
addl %eax,%ebp |
1898 |
– |
shrdl $7,%ebx,%ebx |
1871 |
|
addl %esi,%ebp |
1900 |
– |
vpor %xmm9,%xmm3,%xmm3 |
1901 |
– |
addl 60(%rsp),%edx |
1872 |
|
xorl %ecx,%edi |
1873 |
< |
vmovdqa %xmm3,%xmm10 |
1873 |
> |
shrdl $7,%ebx,%ebx |
1874 |
> |
addl %eax,%ebp |
1875 |
> |
vpor %xmm8,%xmm3,%xmm3 |
1876 |
> |
addl 60(%rsp),%edx |
1877 |
> |
xorl %ebx,%edi |
1878 |
|
movl %ebp,%esi |
1879 |
|
shldl $5,%ebp,%ebp |
1906 |
– |
xorl %ebx,%edi |
1907 |
– |
addl %ebp,%edx |
1908 |
– |
shrdl $7,%eax,%eax |
1880 |
|
addl %edi,%edx |
1881 |
< |
vpalignr $8,%xmm2,%xmm3,%xmm10 |
1881 |
> |
xorl %ebx,%esi |
1882 |
> |
shrdl $7,%eax,%eax |
1883 |
> |
addl %ebp,%edx |
1884 |
> |
vpalignr $8,%xmm2,%xmm3,%xmm8 |
1885 |
|
vpxor %xmm0,%xmm4,%xmm4 |
1886 |
|
addl 0(%rsp),%ecx |
1887 |
< |
xorl %ebx,%esi |
1887 |
> |
xorl %eax,%esi |
1888 |
|
movl %edx,%edi |
1889 |
|
shldl $5,%edx,%edx |
1890 |
|
vpxor %xmm5,%xmm4,%xmm4 |
1891 |
< |
xorl %eax,%esi |
1892 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1891 |
> |
addl %esi,%ecx |
1892 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1893 |
|
vmovups 32(%r15),%xmm14 |
1894 |
< |
addl %edx,%ecx |
1895 |
< |
vmovdqa %xmm8,%xmm9 |
1922 |
< |
vpaddd %xmm3,%xmm8,%xmm8 |
1894 |
> |
xorl %eax,%edi |
1895 |
> |
vpaddd %xmm3,%xmm10,%xmm9 |
1896 |
|
shrdl $7,%ebp,%ebp |
1897 |
< |
addl %esi,%ecx |
1898 |
< |
vpxor %xmm10,%xmm4,%xmm4 |
1897 |
> |
addl %edx,%ecx |
1898 |
> |
vpxor %xmm8,%xmm4,%xmm4 |
1899 |
|
addl 4(%rsp),%ebx |
1900 |
< |
xorl %eax,%edi |
1900 |
> |
xorl %ebp,%edi |
1901 |
|
movl %ecx,%esi |
1902 |
|
shldl $5,%ecx,%ecx |
1903 |
< |
vpsrld $30,%xmm4,%xmm10 |
1904 |
< |
vmovdqa %xmm8,48(%rsp) |
1932 |
< |
xorl %ebp,%edi |
1933 |
< |
addl %ecx,%ebx |
1934 |
< |
shrdl $7,%edx,%edx |
1903 |
> |
vpsrld $30,%xmm4,%xmm8 |
1904 |
> |
vmovdqa %xmm9,48(%rsp) |
1905 |
|
addl %edi,%ebx |
1906 |
+ |
xorl %ebp,%esi |
1907 |
+ |
shrdl $7,%edx,%edx |
1908 |
+ |
addl %ecx,%ebx |
1909 |
|
vpslld $2,%xmm4,%xmm4 |
1910 |
|
addl 8(%rsp),%eax |
1911 |
< |
xorl %ebp,%esi |
1911 |
> |
xorl %edx,%esi |
1912 |
|
movl %ebx,%edi |
1913 |
|
shldl $5,%ebx,%ebx |
1941 |
– |
xorl %edx,%esi |
1942 |
– |
addl %ebx,%eax |
1943 |
– |
shrdl $7,%ecx,%ecx |
1914 |
|
addl %esi,%eax |
1915 |
< |
vpor %xmm10,%xmm4,%xmm4 |
1915 |
> |
xorl %edx,%edi |
1916 |
> |
shrdl $7,%ecx,%ecx |
1917 |
> |
addl %ebx,%eax |
1918 |
> |
vpor %xmm8,%xmm4,%xmm4 |
1919 |
|
addl 12(%rsp),%ebp |
1920 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1920 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1921 |
|
vmovups 48(%r15),%xmm15 |
1922 |
< |
xorl %edx,%edi |
1950 |
< |
vmovdqa %xmm4,%xmm8 |
1922 |
> |
xorl %ecx,%edi |
1923 |
|
movl %eax,%esi |
1924 |
|
shldl $5,%eax,%eax |
1953 |
– |
xorl %ecx,%edi |
1954 |
– |
addl %eax,%ebp |
1955 |
– |
shrdl $7,%ebx,%ebx |
1925 |
|
addl %edi,%ebp |
1926 |
+ |
xorl %ecx,%esi |
1927 |
+ |
shrdl $7,%ebx,%ebx |
1928 |
+ |
addl %eax,%ebp |
1929 |
|
vpalignr $8,%xmm3,%xmm4,%xmm8 |
1930 |
|
vpxor %xmm1,%xmm5,%xmm5 |
1931 |
|
addl 16(%rsp),%edx |
1932 |
< |
xorl %ecx,%esi |
1932 |
> |
xorl %ebx,%esi |
1933 |
|
movl %ebp,%edi |
1934 |
|
shldl $5,%ebp,%ebp |
1935 |
|
vpxor %xmm6,%xmm5,%xmm5 |
1964 |
– |
xorl %ebx,%esi |
1965 |
– |
addl %ebp,%edx |
1966 |
– |
vmovdqa %xmm9,%xmm10 |
1967 |
– |
vpaddd %xmm4,%xmm9,%xmm9 |
1968 |
– |
shrdl $7,%eax,%eax |
1936 |
|
addl %esi,%edx |
1937 |
+ |
xorl %ebx,%edi |
1938 |
+ |
vpaddd %xmm4,%xmm10,%xmm9 |
1939 |
+ |
shrdl $7,%eax,%eax |
1940 |
+ |
addl %ebp,%edx |
1941 |
|
vpxor %xmm8,%xmm5,%xmm5 |
1942 |
|
addl 20(%rsp),%ecx |
1943 |
< |
xorl %ebx,%edi |
1943 |
> |
xorl %eax,%edi |
1944 |
|
movl %edx,%esi |
1945 |
|
shldl $5,%edx,%edx |
1946 |
|
vpsrld $30,%xmm5,%xmm8 |
1947 |
|
vmovdqa %xmm9,0(%rsp) |
1948 |
< |
xorl %eax,%edi |
1948 |
> |
addl %edi,%ecx |
1949 |
|
cmpl $11,%r8d |
1950 |
< |
jb .Lvaesenclast2 |
1951 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1950 |
> |
jb .Lvaesenclast7 |
1951 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1952 |
|
vmovups 64(%r15),%xmm14 |
1953 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1953 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1954 |
|
vmovups 80(%r15),%xmm15 |
1955 |
< |
je .Lvaesenclast2 |
1956 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
1955 |
> |
je .Lvaesenclast7 |
1956 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
1957 |
|
vmovups 96(%r15),%xmm14 |
1958 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
1958 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1959 |
|
vmovups 112(%r15),%xmm15 |
1960 |
< |
.Lvaesenclast2: |
1961 |
< |
vaesenclast %xmm15,%xmm11,%xmm11 |
1960 |
> |
.Lvaesenclast7: |
1961 |
> |
vaesenclast %xmm15,%xmm12,%xmm12 |
1962 |
> |
vmovups -112(%r15),%xmm15 |
1963 |
|
vmovups 16-112(%r15),%xmm14 |
1964 |
< |
addl %edx,%ecx |
1964 |
> |
xorl %eax,%esi |
1965 |
|
shrdl $7,%ebp,%ebp |
1966 |
< |
addl %edi,%ecx |
1966 |
> |
addl %edx,%ecx |
1967 |
|
vpslld $2,%xmm5,%xmm5 |
1968 |
|
addl 24(%rsp),%ebx |
1969 |
< |
xorl %eax,%esi |
1969 |
> |
xorl %ebp,%esi |
1970 |
|
movl %ecx,%edi |
1971 |
|
shldl $5,%ecx,%ecx |
2000 |
– |
xorl %ebp,%esi |
2001 |
– |
addl %ecx,%ebx |
2002 |
– |
shrdl $7,%edx,%edx |
1972 |
|
addl %esi,%ebx |
1973 |
+ |
xorl %ebp,%edi |
1974 |
+ |
shrdl $7,%edx,%edx |
1975 |
+ |
addl %ecx,%ebx |
1976 |
|
vpor %xmm8,%xmm5,%xmm5 |
1977 |
|
addl 28(%rsp),%eax |
1978 |
< |
xorl %ebp,%edi |
2007 |
< |
vmovdqa %xmm5,%xmm9 |
1978 |
> |
shrdl $7,%ecx,%ecx |
1979 |
|
movl %ebx,%esi |
2009 |
– |
shldl $5,%ebx,%ebx |
1980 |
|
xorl %edx,%edi |
1981 |
< |
addl %ebx,%eax |
2012 |
< |
shrdl $7,%ecx,%ecx |
1981 |
> |
shldl $5,%ebx,%ebx |
1982 |
|
addl %edi,%eax |
1983 |
< |
vpalignr $8,%xmm4,%xmm5,%xmm9 |
2015 |
< |
vpxor %xmm2,%xmm6,%xmm6 |
2016 |
< |
movl %ecx,%edi |
2017 |
< |
vmovups 32(%r12),%xmm12 |
2018 |
< |
vxorps %xmm13,%xmm12,%xmm12 |
2019 |
< |
vmovups %xmm11,16(%r13,%r12,1) |
2020 |
< |
vxorps %xmm12,%xmm11,%xmm11 |
2021 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2022 |
< |
vmovups -80(%r15),%xmm15 |
1983 |
> |
xorl %ecx,%esi |
1984 |
|
xorl %edx,%ecx |
1985 |
+ |
addl %ebx,%eax |
1986 |
+ |
vpalignr $8,%xmm4,%xmm5,%xmm8 |
1987 |
+ |
vpxor %xmm2,%xmm6,%xmm6 |
1988 |
|
addl 32(%rsp),%ebp |
1989 |
< |
andl %edx,%edi |
1990 |
< |
vpxor %xmm7,%xmm6,%xmm6 |
1989 |
> |
vmovdqu 32(%r12),%xmm13 |
1990 |
> |
vpxor %xmm15,%xmm13,%xmm13 |
1991 |
> |
vmovups %xmm12,16(%r13,%r12,1) |
1992 |
> |
vpxor %xmm13,%xmm12,%xmm12 |
1993 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
1994 |
> |
vmovups -80(%r15),%xmm15 |
1995 |
|
andl %ecx,%esi |
1996 |
+ |
xorl %edx,%ecx |
1997 |
|
shrdl $7,%ebx,%ebx |
1998 |
< |
vmovdqa %xmm10,%xmm8 |
2030 |
< |
vpaddd %xmm5,%xmm10,%xmm10 |
2031 |
< |
addl %edi,%ebp |
1998 |
> |
vpxor %xmm7,%xmm6,%xmm6 |
1999 |
|
movl %eax,%edi |
2000 |
< |
vpxor %xmm9,%xmm6,%xmm6 |
2000 |
> |
xorl %ecx,%esi |
2001 |
> |
vpaddd %xmm5,%xmm10,%xmm9 |
2002 |
|
shldl $5,%eax,%eax |
2003 |
|
addl %esi,%ebp |
2004 |
< |
xorl %edx,%ecx |
2005 |
< |
addl %eax,%ebp |
2038 |
< |
vpsrld $30,%xmm6,%xmm9 |
2039 |
< |
vmovdqa %xmm10,16(%rsp) |
2040 |
< |
movl %ebx,%esi |
2004 |
> |
vpxor %xmm8,%xmm6,%xmm6 |
2005 |
> |
xorl %ebx,%edi |
2006 |
|
xorl %ecx,%ebx |
2007 |
+ |
addl %eax,%ebp |
2008 |
|
addl 36(%rsp),%edx |
2009 |
< |
andl %ecx,%esi |
2010 |
< |
vpslld $2,%xmm6,%xmm6 |
2009 |
> |
vpsrld $30,%xmm6,%xmm8 |
2010 |
> |
vmovdqa %xmm9,16(%rsp) |
2011 |
|
andl %ebx,%edi |
2012 |
+ |
xorl %ecx,%ebx |
2013 |
|
shrdl $7,%eax,%eax |
2047 |
– |
addl %esi,%edx |
2014 |
|
movl %ebp,%esi |
2015 |
+ |
vpslld $2,%xmm6,%xmm6 |
2016 |
+ |
xorl %ebx,%edi |
2017 |
|
shldl $5,%ebp,%ebp |
2050 |
– |
vaesenc %xmm15,%xmm11,%xmm11 |
2051 |
– |
vmovups -64(%r15),%xmm14 |
2018 |
|
addl %edi,%edx |
2019 |
< |
xorl %ecx,%ebx |
2020 |
< |
addl %ebp,%edx |
2021 |
< |
vpor %xmm9,%xmm6,%xmm6 |
2056 |
< |
movl %eax,%edi |
2019 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2020 |
> |
vmovups -64(%r15),%xmm14 |
2021 |
> |
xorl %eax,%esi |
2022 |
|
xorl %ebx,%eax |
2023 |
< |
vmovdqa %xmm6,%xmm10 |
2023 |
> |
addl %ebp,%edx |
2024 |
|
addl 40(%rsp),%ecx |
2060 |
– |
andl %ebx,%edi |
2025 |
|
andl %eax,%esi |
2026 |
+ |
vpor %xmm8,%xmm6,%xmm6 |
2027 |
+ |
xorl %ebx,%eax |
2028 |
|
shrdl $7,%ebp,%ebp |
2063 |
– |
addl %edi,%ecx |
2029 |
|
movl %edx,%edi |
2030 |
+ |
xorl %eax,%esi |
2031 |
|
shldl $5,%edx,%edx |
2032 |
|
addl %esi,%ecx |
2033 |
< |
xorl %ebx,%eax |
2068 |
< |
addl %edx,%ecx |
2069 |
< |
movl %ebp,%esi |
2033 |
> |
xorl %ebp,%edi |
2034 |
|
xorl %eax,%ebp |
2035 |
+ |
addl %edx,%ecx |
2036 |
|
addl 44(%rsp),%ebx |
2072 |
– |
andl %eax,%esi |
2037 |
|
andl %ebp,%edi |
2038 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2075 |
< |
vmovups -48(%r15),%xmm15 |
2038 |
> |
xorl %eax,%ebp |
2039 |
|
shrdl $7,%edx,%edx |
2040 |
< |
addl %esi,%ebx |
2040 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2041 |
> |
vmovups -48(%r15),%xmm15 |
2042 |
|
movl %ecx,%esi |
2043 |
+ |
xorl %ebp,%edi |
2044 |
|
shldl $5,%ecx,%ecx |
2045 |
|
addl %edi,%ebx |
2046 |
< |
xorl %eax,%ebp |
2046 |
> |
xorl %edx,%esi |
2047 |
> |
xorl %ebp,%edx |
2048 |
|
addl %ecx,%ebx |
2049 |
< |
vpalignr $8,%xmm5,%xmm6,%xmm10 |
2049 |
> |
vpalignr $8,%xmm5,%xmm6,%xmm8 |
2050 |
|
vpxor %xmm3,%xmm7,%xmm7 |
2085 |
– |
movl %edx,%edi |
2086 |
– |
xorl %ebp,%edx |
2051 |
|
addl 48(%rsp),%eax |
2088 |
– |
andl %ebp,%edi |
2089 |
– |
vpxor %xmm0,%xmm7,%xmm7 |
2052 |
|
andl %edx,%esi |
2053 |
+ |
xorl %ebp,%edx |
2054 |
|
shrdl $7,%ecx,%ecx |
2055 |
< |
vmovdqa 48(%r11),%xmm9 |
2093 |
< |
vpaddd %xmm6,%xmm8,%xmm8 |
2094 |
< |
addl %edi,%eax |
2055 |
> |
vpxor %xmm0,%xmm7,%xmm7 |
2056 |
|
movl %ebx,%edi |
2057 |
< |
vpxor %xmm10,%xmm7,%xmm7 |
2057 |
> |
xorl %edx,%esi |
2058 |
> |
vpaddd %xmm6,%xmm10,%xmm9 |
2059 |
> |
vmovdqa 48(%r11),%xmm10 |
2060 |
|
shldl $5,%ebx,%ebx |
2061 |
|
addl %esi,%eax |
2062 |
< |
xorl %ebp,%edx |
2063 |
< |
addl %ebx,%eax |
2101 |
< |
vpsrld $30,%xmm7,%xmm10 |
2102 |
< |
vmovdqa %xmm8,32(%rsp) |
2103 |
< |
movl %ecx,%esi |
2104 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2105 |
< |
vmovups -32(%r15),%xmm14 |
2062 |
> |
vpxor %xmm8,%xmm7,%xmm7 |
2063 |
> |
xorl %ecx,%edi |
2064 |
|
xorl %edx,%ecx |
2065 |
+ |
addl %ebx,%eax |
2066 |
|
addl 52(%rsp),%ebp |
2067 |
< |
andl %edx,%esi |
2068 |
< |
vpslld $2,%xmm7,%xmm7 |
2067 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2068 |
> |
vmovups -32(%r15),%xmm14 |
2069 |
> |
vpsrld $30,%xmm7,%xmm8 |
2070 |
> |
vmovdqa %xmm9,32(%rsp) |
2071 |
|
andl %ecx,%edi |
2072 |
+ |
xorl %edx,%ecx |
2073 |
|
shrdl $7,%ebx,%ebx |
2112 |
– |
addl %esi,%ebp |
2074 |
|
movl %eax,%esi |
2075 |
+ |
vpslld $2,%xmm7,%xmm7 |
2076 |
+ |
xorl %ecx,%edi |
2077 |
|
shldl $5,%eax,%eax |
2078 |
|
addl %edi,%ebp |
2079 |
< |
xorl %edx,%ecx |
2117 |
< |
addl %eax,%ebp |
2118 |
< |
vpor %xmm10,%xmm7,%xmm7 |
2119 |
< |
movl %ebx,%edi |
2079 |
> |
xorl %ebx,%esi |
2080 |
|
xorl %ecx,%ebx |
2081 |
< |
vmovdqa %xmm7,%xmm8 |
2081 |
> |
addl %eax,%ebp |
2082 |
|
addl 56(%rsp),%edx |
2123 |
– |
andl %ecx,%edi |
2083 |
|
andl %ebx,%esi |
2084 |
+ |
vpor %xmm8,%xmm7,%xmm7 |
2085 |
+ |
xorl %ecx,%ebx |
2086 |
|
shrdl $7,%eax,%eax |
2126 |
– |
addl %edi,%edx |
2087 |
|
movl %ebp,%edi |
2088 |
+ |
xorl %ebx,%esi |
2089 |
|
shldl $5,%ebp,%ebp |
2129 |
– |
vaesenc %xmm14,%xmm11,%xmm11 |
2130 |
– |
vmovups -16(%r15),%xmm15 |
2090 |
|
addl %esi,%edx |
2091 |
< |
xorl %ecx,%ebx |
2092 |
< |
addl %ebp,%edx |
2093 |
< |
movl %eax,%esi |
2091 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2092 |
> |
vmovups -16(%r15),%xmm15 |
2093 |
> |
xorl %eax,%edi |
2094 |
|
xorl %ebx,%eax |
2095 |
+ |
addl %ebp,%edx |
2096 |
|
addl 60(%rsp),%ecx |
2137 |
– |
andl %ebx,%esi |
2097 |
|
andl %eax,%edi |
2098 |
+ |
xorl %ebx,%eax |
2099 |
|
shrdl $7,%ebp,%ebp |
2140 |
– |
addl %esi,%ecx |
2100 |
|
movl %edx,%esi |
2101 |
+ |
xorl %eax,%edi |
2102 |
|
shldl $5,%edx,%edx |
2103 |
|
addl %edi,%ecx |
2104 |
< |
xorl %ebx,%eax |
2104 |
> |
xorl %ebp,%esi |
2105 |
> |
xorl %eax,%ebp |
2106 |
|
addl %edx,%ecx |
2107 |
|
vpalignr $8,%xmm6,%xmm7,%xmm8 |
2108 |
|
vpxor %xmm4,%xmm0,%xmm0 |
2148 |
– |
movl %ebp,%edi |
2149 |
– |
xorl %eax,%ebp |
2109 |
|
addl 0(%rsp),%ebx |
2151 |
– |
andl %eax,%edi |
2152 |
– |
vpxor %xmm1,%xmm0,%xmm0 |
2110 |
|
andl %ebp,%esi |
2111 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2155 |
< |
vmovups 0(%r15),%xmm14 |
2111 |
> |
xorl %eax,%ebp |
2112 |
|
shrdl $7,%edx,%edx |
2113 |
< |
vmovdqa %xmm9,%xmm10 |
2114 |
< |
vpaddd %xmm7,%xmm9,%xmm9 |
2115 |
< |
addl %edi,%ebx |
2113 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2114 |
> |
vmovups 0(%r15),%xmm14 |
2115 |
> |
vpxor %xmm1,%xmm0,%xmm0 |
2116 |
|
movl %ecx,%edi |
2117 |
< |
vpxor %xmm8,%xmm0,%xmm0 |
2117 |
> |
xorl %ebp,%esi |
2118 |
> |
vpaddd %xmm7,%xmm10,%xmm9 |
2119 |
|
shldl $5,%ecx,%ecx |
2120 |
|
addl %esi,%ebx |
2121 |
< |
xorl %eax,%ebp |
2121 |
> |
vpxor %xmm8,%xmm0,%xmm0 |
2122 |
> |
xorl %edx,%edi |
2123 |
> |
xorl %ebp,%edx |
2124 |
|
addl %ecx,%ebx |
2125 |
+ |
addl 4(%rsp),%eax |
2126 |
|
vpsrld $30,%xmm0,%xmm8 |
2127 |
|
vmovdqa %xmm9,48(%rsp) |
2168 |
– |
movl %edx,%esi |
2169 |
– |
xorl %ebp,%edx |
2170 |
– |
addl 4(%rsp),%eax |
2171 |
– |
andl %ebp,%esi |
2172 |
– |
vpslld $2,%xmm0,%xmm0 |
2128 |
|
andl %edx,%edi |
2129 |
+ |
xorl %ebp,%edx |
2130 |
|
shrdl $7,%ecx,%ecx |
2175 |
– |
addl %esi,%eax |
2131 |
|
movl %ebx,%esi |
2132 |
+ |
vpslld $2,%xmm0,%xmm0 |
2133 |
+ |
xorl %edx,%edi |
2134 |
|
shldl $5,%ebx,%ebx |
2135 |
|
addl %edi,%eax |
2136 |
< |
xorl %ebp,%edx |
2180 |
< |
addl %ebx,%eax |
2181 |
< |
vpor %xmm8,%xmm0,%xmm0 |
2182 |
< |
movl %ecx,%edi |
2183 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2184 |
< |
vmovups 16(%r15),%xmm15 |
2136 |
> |
xorl %ecx,%esi |
2137 |
|
xorl %edx,%ecx |
2138 |
< |
vmovdqa %xmm0,%xmm9 |
2138 |
> |
addl %ebx,%eax |
2139 |
|
addl 8(%rsp),%ebp |
2140 |
< |
andl %edx,%edi |
2140 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2141 |
> |
vmovups 16(%r15),%xmm15 |
2142 |
|
andl %ecx,%esi |
2143 |
+ |
vpor %xmm8,%xmm0,%xmm0 |
2144 |
+ |
xorl %edx,%ecx |
2145 |
|
shrdl $7,%ebx,%ebx |
2191 |
– |
addl %edi,%ebp |
2146 |
|
movl %eax,%edi |
2147 |
+ |
xorl %ecx,%esi |
2148 |
|
shldl $5,%eax,%eax |
2149 |
|
addl %esi,%ebp |
2150 |
< |
xorl %edx,%ecx |
2196 |
< |
addl %eax,%ebp |
2197 |
< |
movl %ebx,%esi |
2150 |
> |
xorl %ebx,%edi |
2151 |
|
xorl %ecx,%ebx |
2152 |
+ |
addl %eax,%ebp |
2153 |
|
addl 12(%rsp),%edx |
2200 |
– |
andl %ecx,%esi |
2154 |
|
andl %ebx,%edi |
2155 |
+ |
xorl %ecx,%ebx |
2156 |
|
shrdl $7,%eax,%eax |
2203 |
– |
addl %esi,%edx |
2157 |
|
movl %ebp,%esi |
2158 |
+ |
xorl %ebx,%edi |
2159 |
|
shldl $5,%ebp,%ebp |
2206 |
– |
vaesenc %xmm15,%xmm11,%xmm11 |
2207 |
– |
vmovups 32(%r15),%xmm14 |
2160 |
|
addl %edi,%edx |
2161 |
< |
xorl %ecx,%ebx |
2161 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2162 |
> |
vmovups 32(%r15),%xmm14 |
2163 |
> |
xorl %eax,%esi |
2164 |
> |
xorl %ebx,%eax |
2165 |
|
addl %ebp,%edx |
2166 |
< |
vpalignr $8,%xmm7,%xmm0,%xmm9 |
2166 |
> |
vpalignr $8,%xmm7,%xmm0,%xmm8 |
2167 |
|
vpxor %xmm5,%xmm1,%xmm1 |
2213 |
– |
movl %eax,%edi |
2214 |
– |
xorl %ebx,%eax |
2168 |
|
addl 16(%rsp),%ecx |
2216 |
– |
andl %ebx,%edi |
2217 |
– |
vpxor %xmm2,%xmm1,%xmm1 |
2169 |
|
andl %eax,%esi |
2170 |
+ |
xorl %ebx,%eax |
2171 |
|
shrdl $7,%ebp,%ebp |
2172 |
< |
vmovdqa %xmm10,%xmm8 |
2221 |
< |
vpaddd %xmm0,%xmm10,%xmm10 |
2222 |
< |
addl %edi,%ecx |
2172 |
> |
vpxor %xmm2,%xmm1,%xmm1 |
2173 |
|
movl %edx,%edi |
2174 |
< |
vpxor %xmm9,%xmm1,%xmm1 |
2174 |
> |
xorl %eax,%esi |
2175 |
> |
vpaddd %xmm0,%xmm10,%xmm9 |
2176 |
|
shldl $5,%edx,%edx |
2177 |
|
addl %esi,%ecx |
2178 |
< |
xorl %ebx,%eax |
2179 |
< |
addl %edx,%ecx |
2229 |
< |
vpsrld $30,%xmm1,%xmm9 |
2230 |
< |
vmovdqa %xmm10,0(%rsp) |
2231 |
< |
movl %ebp,%esi |
2178 |
> |
vpxor %xmm8,%xmm1,%xmm1 |
2179 |
> |
xorl %ebp,%edi |
2180 |
|
xorl %eax,%ebp |
2181 |
+ |
addl %edx,%ecx |
2182 |
|
addl 20(%rsp),%ebx |
2183 |
< |
andl %eax,%esi |
2184 |
< |
vpslld $2,%xmm1,%xmm1 |
2183 |
> |
vpsrld $30,%xmm1,%xmm8 |
2184 |
> |
vmovdqa %xmm9,0(%rsp) |
2185 |
|
andl %ebp,%edi |
2186 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2238 |
< |
vmovups 48(%r15),%xmm15 |
2186 |
> |
xorl %eax,%ebp |
2187 |
|
shrdl $7,%edx,%edx |
2188 |
< |
addl %esi,%ebx |
2188 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2189 |
> |
vmovups 48(%r15),%xmm15 |
2190 |
|
movl %ecx,%esi |
2191 |
+ |
vpslld $2,%xmm1,%xmm1 |
2192 |
+ |
xorl %ebp,%edi |
2193 |
|
shldl $5,%ecx,%ecx |
2194 |
|
addl %edi,%ebx |
2195 |
< |
xorl %eax,%ebp |
2245 |
< |
addl %ecx,%ebx |
2246 |
< |
vpor %xmm9,%xmm1,%xmm1 |
2247 |
< |
movl %edx,%edi |
2195 |
> |
xorl %edx,%esi |
2196 |
|
xorl %ebp,%edx |
2197 |
< |
vmovdqa %xmm1,%xmm10 |
2197 |
> |
addl %ecx,%ebx |
2198 |
|
addl 24(%rsp),%eax |
2251 |
– |
andl %ebp,%edi |
2199 |
|
andl %edx,%esi |
2200 |
+ |
vpor %xmm8,%xmm1,%xmm1 |
2201 |
+ |
xorl %ebp,%edx |
2202 |
|
shrdl $7,%ecx,%ecx |
2254 |
– |
addl %edi,%eax |
2203 |
|
movl %ebx,%edi |
2204 |
+ |
xorl %edx,%esi |
2205 |
|
shldl $5,%ebx,%ebx |
2206 |
|
addl %esi,%eax |
2207 |
< |
xorl %ebp,%edx |
2207 |
> |
xorl %ecx,%edi |
2208 |
> |
xorl %edx,%ecx |
2209 |
|
addl %ebx,%eax |
2210 |
< |
movl %ecx,%esi |
2210 |
> |
addl 28(%rsp),%ebp |
2211 |
|
cmpl $11,%r8d |
2212 |
< |
jb .Lvaesenclast3 |
2213 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2212 |
> |
jb .Lvaesenclast8 |
2213 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2214 |
|
vmovups 64(%r15),%xmm14 |
2215 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2215 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2216 |
|
vmovups 80(%r15),%xmm15 |
2217 |
< |
je .Lvaesenclast3 |
2218 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2217 |
> |
je .Lvaesenclast8 |
2218 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2219 |
|
vmovups 96(%r15),%xmm14 |
2220 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2220 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2221 |
|
vmovups 112(%r15),%xmm15 |
2222 |
< |
.Lvaesenclast3: |
2223 |
< |
vaesenclast %xmm15,%xmm11,%xmm11 |
2222 |
> |
.Lvaesenclast8: |
2223 |
> |
vaesenclast %xmm15,%xmm12,%xmm12 |
2224 |
> |
vmovups -112(%r15),%xmm15 |
2225 |
|
vmovups 16-112(%r15),%xmm14 |
2275 |
– |
xorl %edx,%ecx |
2276 |
– |
addl 28(%rsp),%ebp |
2277 |
– |
andl %edx,%esi |
2226 |
|
andl %ecx,%edi |
2227 |
+ |
xorl %edx,%ecx |
2228 |
|
shrdl $7,%ebx,%ebx |
2280 |
– |
addl %esi,%ebp |
2229 |
|
movl %eax,%esi |
2230 |
+ |
xorl %ecx,%edi |
2231 |
|
shldl $5,%eax,%eax |
2232 |
|
addl %edi,%ebp |
2233 |
< |
xorl %edx,%ecx |
2233 |
> |
xorl %ebx,%esi |
2234 |
> |
xorl %ecx,%ebx |
2235 |
|
addl %eax,%ebp |
2236 |
< |
vpalignr $8,%xmm0,%xmm1,%xmm10 |
2236 |
> |
vpalignr $8,%xmm0,%xmm1,%xmm8 |
2237 |
|
vpxor %xmm6,%xmm2,%xmm2 |
2288 |
– |
movl %ebx,%edi |
2289 |
– |
xorl %ecx,%ebx |
2238 |
|
addl 32(%rsp),%edx |
2291 |
– |
andl %ecx,%edi |
2292 |
– |
vpxor %xmm3,%xmm2,%xmm2 |
2239 |
|
andl %ebx,%esi |
2240 |
+ |
xorl %ecx,%ebx |
2241 |
|
shrdl $7,%eax,%eax |
2242 |
< |
vmovdqa %xmm8,%xmm9 |
2296 |
< |
vpaddd %xmm1,%xmm8,%xmm8 |
2297 |
< |
addl %edi,%edx |
2242 |
> |
vpxor %xmm3,%xmm2,%xmm2 |
2243 |
|
movl %ebp,%edi |
2244 |
< |
vpxor %xmm10,%xmm2,%xmm2 |
2244 |
> |
xorl %ebx,%esi |
2245 |
> |
vpaddd %xmm1,%xmm10,%xmm9 |
2246 |
|
shldl $5,%ebp,%ebp |
2301 |
– |
vmovups 48(%r12),%xmm12 |
2302 |
– |
vxorps %xmm13,%xmm12,%xmm12 |
2303 |
– |
vmovups %xmm11,32(%r13,%r12,1) |
2304 |
– |
vxorps %xmm12,%xmm11,%xmm11 |
2305 |
– |
vaesenc %xmm14,%xmm11,%xmm11 |
2306 |
– |
vmovups -80(%r15),%xmm15 |
2247 |
|
addl %esi,%edx |
2248 |
< |
xorl %ecx,%ebx |
2249 |
< |
addl %ebp,%edx |
2250 |
< |
vpsrld $30,%xmm2,%xmm10 |
2251 |
< |
vmovdqa %xmm8,16(%rsp) |
2252 |
< |
movl %eax,%esi |
2248 |
> |
vmovdqu 48(%r12),%xmm13 |
2249 |
> |
vpxor %xmm15,%xmm13,%xmm13 |
2250 |
> |
vmovups %xmm12,32(%r13,%r12,1) |
2251 |
> |
vpxor %xmm13,%xmm12,%xmm12 |
2252 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2253 |
> |
vmovups -80(%r15),%xmm15 |
2254 |
> |
vpxor %xmm8,%xmm2,%xmm2 |
2255 |
> |
xorl %eax,%edi |
2256 |
|
xorl %ebx,%eax |
2257 |
+ |
addl %ebp,%edx |
2258 |
|
addl 36(%rsp),%ecx |
2259 |
< |
andl %ebx,%esi |
2260 |
< |
vpslld $2,%xmm2,%xmm2 |
2259 |
> |
vpsrld $30,%xmm2,%xmm8 |
2260 |
> |
vmovdqa %xmm9,16(%rsp) |
2261 |
|
andl %eax,%edi |
2262 |
+ |
xorl %ebx,%eax |
2263 |
|
shrdl $7,%ebp,%ebp |
2319 |
– |
addl %esi,%ecx |
2264 |
|
movl %edx,%esi |
2265 |
+ |
vpslld $2,%xmm2,%xmm2 |
2266 |
+ |
xorl %eax,%edi |
2267 |
|
shldl $5,%edx,%edx |
2268 |
|
addl %edi,%ecx |
2269 |
< |
xorl %ebx,%eax |
2324 |
< |
addl %edx,%ecx |
2325 |
< |
vpor %xmm10,%xmm2,%xmm2 |
2326 |
< |
movl %ebp,%edi |
2269 |
> |
xorl %ebp,%esi |
2270 |
|
xorl %eax,%ebp |
2271 |
< |
vmovdqa %xmm2,%xmm8 |
2271 |
> |
addl %edx,%ecx |
2272 |
|
addl 40(%rsp),%ebx |
2330 |
– |
andl %eax,%edi |
2273 |
|
andl %ebp,%esi |
2274 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2275 |
< |
vmovups -64(%r15),%xmm14 |
2274 |
> |
vpor %xmm8,%xmm2,%xmm2 |
2275 |
> |
xorl %eax,%ebp |
2276 |
|
shrdl $7,%edx,%edx |
2277 |
< |
addl %edi,%ebx |
2277 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2278 |
> |
vmovups -64(%r15),%xmm14 |
2279 |
|
movl %ecx,%edi |
2280 |
+ |
xorl %ebp,%esi |
2281 |
|
shldl $5,%ecx,%ecx |
2282 |
|
addl %esi,%ebx |
2283 |
< |
xorl %eax,%ebp |
2340 |
< |
addl %ecx,%ebx |
2341 |
< |
movl %edx,%esi |
2283 |
> |
xorl %edx,%edi |
2284 |
|
xorl %ebp,%edx |
2285 |
+ |
addl %ecx,%ebx |
2286 |
|
addl 44(%rsp),%eax |
2344 |
– |
andl %ebp,%esi |
2287 |
|
andl %edx,%edi |
2288 |
+ |
xorl %ebp,%edx |
2289 |
|
shrdl $7,%ecx,%ecx |
2347 |
– |
addl %esi,%eax |
2290 |
|
movl %ebx,%esi |
2291 |
+ |
xorl %edx,%edi |
2292 |
|
shldl $5,%ebx,%ebx |
2293 |
|
addl %edi,%eax |
2294 |
< |
xorl %ebp,%edx |
2294 |
> |
xorl %edx,%esi |
2295 |
|
addl %ebx,%eax |
2296 |
|
vpalignr $8,%xmm1,%xmm2,%xmm8 |
2297 |
|
vpxor %xmm7,%xmm3,%xmm3 |
2298 |
|
addl 48(%rsp),%ebp |
2299 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2299 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2300 |
|
vmovups -48(%r15),%xmm15 |
2301 |
< |
xorl %edx,%esi |
2301 |
> |
xorl %ecx,%esi |
2302 |
|
movl %eax,%edi |
2303 |
|
shldl $5,%eax,%eax |
2304 |
|
vpxor %xmm4,%xmm3,%xmm3 |
2362 |
– |
xorl %ecx,%esi |
2363 |
– |
addl %eax,%ebp |
2364 |
– |
vmovdqa %xmm9,%xmm10 |
2365 |
– |
vpaddd %xmm2,%xmm9,%xmm9 |
2366 |
– |
shrdl $7,%ebx,%ebx |
2305 |
|
addl %esi,%ebp |
2306 |
+ |
xorl %ecx,%edi |
2307 |
+ |
vpaddd %xmm2,%xmm10,%xmm9 |
2308 |
+ |
shrdl $7,%ebx,%ebx |
2309 |
+ |
addl %eax,%ebp |
2310 |
|
vpxor %xmm8,%xmm3,%xmm3 |
2311 |
|
addl 52(%rsp),%edx |
2312 |
< |
xorl %ecx,%edi |
2312 |
> |
xorl %ebx,%edi |
2313 |
|
movl %ebp,%esi |
2314 |
|
shldl $5,%ebp,%ebp |
2315 |
|
vpsrld $30,%xmm3,%xmm8 |
2316 |
|
vmovdqa %xmm9,32(%rsp) |
2375 |
– |
xorl %ebx,%edi |
2376 |
– |
addl %ebp,%edx |
2377 |
– |
shrdl $7,%eax,%eax |
2317 |
|
addl %edi,%edx |
2318 |
+ |
xorl %ebx,%esi |
2319 |
+ |
shrdl $7,%eax,%eax |
2320 |
+ |
addl %ebp,%edx |
2321 |
|
vpslld $2,%xmm3,%xmm3 |
2322 |
|
addl 56(%rsp),%ecx |
2323 |
< |
xorl %ebx,%esi |
2323 |
> |
xorl %eax,%esi |
2324 |
|
movl %edx,%edi |
2325 |
|
shldl $5,%edx,%edx |
2326 |
< |
xorl %eax,%esi |
2327 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2326 |
> |
addl %esi,%ecx |
2327 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2328 |
|
vmovups -32(%r15),%xmm14 |
2329 |
< |
addl %edx,%ecx |
2329 |
> |
xorl %eax,%edi |
2330 |
|
shrdl $7,%ebp,%ebp |
2331 |
< |
addl %esi,%ecx |
2331 |
> |
addl %edx,%ecx |
2332 |
|
vpor %xmm8,%xmm3,%xmm3 |
2333 |
|
addl 60(%rsp),%ebx |
2334 |
< |
xorl %eax,%edi |
2334 |
> |
xorl %ebp,%edi |
2335 |
|
movl %ecx,%esi |
2336 |
|
shldl $5,%ecx,%ecx |
2395 |
– |
xorl %ebp,%edi |
2396 |
– |
addl %ecx,%ebx |
2397 |
– |
shrdl $7,%edx,%edx |
2337 |
|
addl %edi,%ebx |
2399 |
– |
addl 0(%rsp),%eax |
2400 |
– |
vpaddd %xmm3,%xmm10,%xmm10 |
2338 |
|
xorl %ebp,%esi |
2339 |
+ |
shrdl $7,%edx,%edx |
2340 |
+ |
addl %ecx,%ebx |
2341 |
+ |
addl 0(%rsp),%eax |
2342 |
+ |
vpaddd %xmm3,%xmm10,%xmm9 |
2343 |
+ |
xorl %edx,%esi |
2344 |
|
movl %ebx,%edi |
2345 |
|
shldl $5,%ebx,%ebx |
2404 |
– |
xorl %edx,%esi |
2405 |
– |
movdqa %xmm10,48(%rsp) |
2406 |
– |
addl %ebx,%eax |
2407 |
– |
shrdl $7,%ecx,%ecx |
2346 |
|
addl %esi,%eax |
2347 |
+ |
vmovdqa %xmm9,48(%rsp) |
2348 |
+ |
xorl %edx,%edi |
2349 |
+ |
shrdl $7,%ecx,%ecx |
2350 |
+ |
addl %ebx,%eax |
2351 |
|
addl 4(%rsp),%ebp |
2352 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2352 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2353 |
|
vmovups -16(%r15),%xmm15 |
2354 |
< |
xorl %edx,%edi |
2354 |
> |
xorl %ecx,%edi |
2355 |
|
movl %eax,%esi |
2356 |
|
shldl $5,%eax,%eax |
2415 |
– |
xorl %ecx,%edi |
2416 |
– |
addl %eax,%ebp |
2417 |
– |
shrdl $7,%ebx,%ebx |
2357 |
|
addl %edi,%ebp |
2419 |
– |
addl 8(%rsp),%edx |
2358 |
|
xorl %ecx,%esi |
2359 |
+ |
shrdl $7,%ebx,%ebx |
2360 |
+ |
addl %eax,%ebp |
2361 |
+ |
addl 8(%rsp),%edx |
2362 |
+ |
xorl %ebx,%esi |
2363 |
|
movl %ebp,%edi |
2364 |
|
shldl $5,%ebp,%ebp |
2423 |
– |
xorl %ebx,%esi |
2424 |
– |
addl %ebp,%edx |
2425 |
– |
shrdl $7,%eax,%eax |
2365 |
|
addl %esi,%edx |
2427 |
– |
addl 12(%rsp),%ecx |
2366 |
|
xorl %ebx,%edi |
2367 |
+ |
shrdl $7,%eax,%eax |
2368 |
+ |
addl %ebp,%edx |
2369 |
+ |
addl 12(%rsp),%ecx |
2370 |
+ |
xorl %eax,%edi |
2371 |
|
movl %edx,%esi |
2372 |
|
shldl $5,%edx,%edx |
2373 |
< |
xorl %eax,%edi |
2374 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2373 |
> |
addl %edi,%ecx |
2374 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2375 |
|
vmovups 0(%r15),%xmm14 |
2376 |
< |
addl %edx,%ecx |
2376 |
> |
xorl %eax,%esi |
2377 |
|
shrdl $7,%ebp,%ebp |
2378 |
< |
addl %edi,%ecx |
2378 |
> |
addl %edx,%ecx |
2379 |
|
cmpq %r14,%r10 |
2380 |
|
je .Ldone_avx |
2381 |
< |
vmovdqa 64(%r11),%xmm6 |
2382 |
< |
vmovdqa 0(%r11),%xmm9 |
2381 |
> |
vmovdqa 64(%r11),%xmm9 |
2382 |
> |
vmovdqa 0(%r11),%xmm10 |
2383 |
|
vmovdqu 0(%r10),%xmm0 |
2384 |
|
vmovdqu 16(%r10),%xmm1 |
2385 |
|
vmovdqu 32(%r10),%xmm2 |
2386 |
|
vmovdqu 48(%r10),%xmm3 |
2387 |
< |
vpshufb %xmm6,%xmm0,%xmm0 |
2387 |
> |
vpshufb %xmm9,%xmm0,%xmm0 |
2388 |
|
addq $64,%r10 |
2389 |
|
addl 16(%rsp),%ebx |
2390 |
< |
xorl %eax,%esi |
2391 |
< |
vpshufb %xmm6,%xmm1,%xmm1 |
2390 |
> |
xorl %ebp,%esi |
2391 |
> |
vpshufb %xmm9,%xmm1,%xmm1 |
2392 |
|
movl %ecx,%edi |
2393 |
|
shldl $5,%ecx,%ecx |
2394 |
< |
vpaddd %xmm9,%xmm0,%xmm4 |
2453 |
< |
xorl %ebp,%esi |
2454 |
< |
addl %ecx,%ebx |
2455 |
< |
shrdl $7,%edx,%edx |
2394 |
> |
vpaddd %xmm10,%xmm0,%xmm8 |
2395 |
|
addl %esi,%ebx |
2457 |
– |
vmovdqa %xmm4,0(%rsp) |
2458 |
– |
addl 20(%rsp),%eax |
2396 |
|
xorl %ebp,%edi |
2397 |
+ |
shrdl $7,%edx,%edx |
2398 |
+ |
addl %ecx,%ebx |
2399 |
+ |
vmovdqa %xmm8,0(%rsp) |
2400 |
+ |
addl 20(%rsp),%eax |
2401 |
+ |
xorl %edx,%edi |
2402 |
|
movl %ebx,%esi |
2403 |
|
shldl $5,%ebx,%ebx |
2462 |
– |
xorl %edx,%edi |
2463 |
– |
addl %ebx,%eax |
2464 |
– |
shrdl $7,%ecx,%ecx |
2404 |
|
addl %edi,%eax |
2405 |
+ |
xorl %edx,%esi |
2406 |
+ |
shrdl $7,%ecx,%ecx |
2407 |
+ |
addl %ebx,%eax |
2408 |
|
addl 24(%rsp),%ebp |
2409 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2409 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2410 |
|
vmovups 16(%r15),%xmm15 |
2411 |
< |
xorl %edx,%esi |
2411 |
> |
xorl %ecx,%esi |
2412 |
|
movl %eax,%edi |
2413 |
|
shldl $5,%eax,%eax |
2472 |
– |
xorl %ecx,%esi |
2473 |
– |
addl %eax,%ebp |
2474 |
– |
shrdl $7,%ebx,%ebx |
2414 |
|
addl %esi,%ebp |
2476 |
– |
addl 28(%rsp),%edx |
2415 |
|
xorl %ecx,%edi |
2416 |
+ |
shrdl $7,%ebx,%ebx |
2417 |
+ |
addl %eax,%ebp |
2418 |
+ |
addl 28(%rsp),%edx |
2419 |
+ |
xorl %ebx,%edi |
2420 |
|
movl %ebp,%esi |
2421 |
|
shldl $5,%ebp,%ebp |
2480 |
– |
xorl %ebx,%edi |
2481 |
– |
addl %ebp,%edx |
2482 |
– |
shrdl $7,%eax,%eax |
2422 |
|
addl %edi,%edx |
2484 |
– |
addl 32(%rsp),%ecx |
2423 |
|
xorl %ebx,%esi |
2424 |
< |
vpshufb %xmm6,%xmm2,%xmm2 |
2424 |
> |
shrdl $7,%eax,%eax |
2425 |
> |
addl %ebp,%edx |
2426 |
> |
addl 32(%rsp),%ecx |
2427 |
> |
xorl %eax,%esi |
2428 |
> |
vpshufb %xmm9,%xmm2,%xmm2 |
2429 |
|
movl %edx,%edi |
2430 |
|
shldl $5,%edx,%edx |
2431 |
< |
vpaddd %xmm9,%xmm1,%xmm5 |
2432 |
< |
xorl %eax,%esi |
2433 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2431 |
> |
vpaddd %xmm10,%xmm1,%xmm8 |
2432 |
> |
addl %esi,%ecx |
2433 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2434 |
|
vmovups 32(%r15),%xmm14 |
2435 |
< |
addl %edx,%ecx |
2435 |
> |
xorl %eax,%edi |
2436 |
|
shrdl $7,%ebp,%ebp |
2437 |
< |
addl %esi,%ecx |
2438 |
< |
vmovdqa %xmm5,16(%rsp) |
2437 |
> |
addl %edx,%ecx |
2438 |
> |
vmovdqa %xmm8,16(%rsp) |
2439 |
|
addl 36(%rsp),%ebx |
2440 |
< |
xorl %eax,%edi |
2440 |
> |
xorl %ebp,%edi |
2441 |
|
movl %ecx,%esi |
2442 |
|
shldl $5,%ecx,%ecx |
2501 |
– |
xorl %ebp,%edi |
2502 |
– |
addl %ecx,%ebx |
2503 |
– |
shrdl $7,%edx,%edx |
2443 |
|
addl %edi,%ebx |
2505 |
– |
addl 40(%rsp),%eax |
2444 |
|
xorl %ebp,%esi |
2445 |
+ |
shrdl $7,%edx,%edx |
2446 |
+ |
addl %ecx,%ebx |
2447 |
+ |
addl 40(%rsp),%eax |
2448 |
+ |
xorl %edx,%esi |
2449 |
|
movl %ebx,%edi |
2450 |
|
shldl $5,%ebx,%ebx |
2509 |
– |
xorl %edx,%esi |
2510 |
– |
addl %ebx,%eax |
2511 |
– |
shrdl $7,%ecx,%ecx |
2451 |
|
addl %esi,%eax |
2452 |
+ |
xorl %edx,%edi |
2453 |
+ |
shrdl $7,%ecx,%ecx |
2454 |
+ |
addl %ebx,%eax |
2455 |
|
addl 44(%rsp),%ebp |
2456 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2456 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2457 |
|
vmovups 48(%r15),%xmm15 |
2458 |
< |
xorl %edx,%edi |
2458 |
> |
xorl %ecx,%edi |
2459 |
|
movl %eax,%esi |
2460 |
|
shldl $5,%eax,%eax |
2519 |
– |
xorl %ecx,%edi |
2520 |
– |
addl %eax,%ebp |
2521 |
– |
shrdl $7,%ebx,%ebx |
2461 |
|
addl %edi,%ebp |
2523 |
– |
addl 48(%rsp),%edx |
2462 |
|
xorl %ecx,%esi |
2463 |
< |
vpshufb %xmm6,%xmm3,%xmm3 |
2463 |
> |
shrdl $7,%ebx,%ebx |
2464 |
> |
addl %eax,%ebp |
2465 |
> |
addl 48(%rsp),%edx |
2466 |
> |
xorl %ebx,%esi |
2467 |
> |
vpshufb %xmm9,%xmm3,%xmm3 |
2468 |
|
movl %ebp,%edi |
2469 |
|
shldl $5,%ebp,%ebp |
2470 |
< |
vpaddd %xmm9,%xmm2,%xmm6 |
2529 |
< |
xorl %ebx,%esi |
2530 |
< |
addl %ebp,%edx |
2531 |
< |
shrdl $7,%eax,%eax |
2470 |
> |
vpaddd %xmm10,%xmm2,%xmm8 |
2471 |
|
addl %esi,%edx |
2533 |
– |
vmovdqa %xmm6,32(%rsp) |
2534 |
– |
addl 52(%rsp),%ecx |
2472 |
|
xorl %ebx,%edi |
2473 |
+ |
shrdl $7,%eax,%eax |
2474 |
+ |
addl %ebp,%edx |
2475 |
+ |
vmovdqa %xmm8,32(%rsp) |
2476 |
+ |
addl 52(%rsp),%ecx |
2477 |
+ |
xorl %eax,%edi |
2478 |
|
movl %edx,%esi |
2479 |
|
shldl $5,%edx,%edx |
2480 |
< |
xorl %eax,%edi |
2480 |
> |
addl %edi,%ecx |
2481 |
|
cmpl $11,%r8d |
2482 |
< |
jb .Lvaesenclast4 |
2483 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2482 |
> |
jb .Lvaesenclast9 |
2483 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2484 |
|
vmovups 64(%r15),%xmm14 |
2485 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2485 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2486 |
|
vmovups 80(%r15),%xmm15 |
2487 |
< |
je .Lvaesenclast4 |
2488 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2487 |
> |
je .Lvaesenclast9 |
2488 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2489 |
|
vmovups 96(%r15),%xmm14 |
2490 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2490 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2491 |
|
vmovups 112(%r15),%xmm15 |
2492 |
< |
.Lvaesenclast4: |
2493 |
< |
vaesenclast %xmm15,%xmm11,%xmm11 |
2492 |
> |
.Lvaesenclast9: |
2493 |
> |
vaesenclast %xmm15,%xmm12,%xmm12 |
2494 |
> |
vmovups -112(%r15),%xmm15 |
2495 |
|
vmovups 16-112(%r15),%xmm14 |
2496 |
< |
addl %edx,%ecx |
2496 |
> |
xorl %eax,%esi |
2497 |
|
shrdl $7,%ebp,%ebp |
2498 |
< |
addl %edi,%ecx |
2498 |
> |
addl %edx,%ecx |
2499 |
|
addl 56(%rsp),%ebx |
2500 |
< |
xorl %eax,%esi |
2500 |
> |
xorl %ebp,%esi |
2501 |
|
movl %ecx,%edi |
2502 |
|
shldl $5,%ecx,%ecx |
2560 |
– |
xorl %ebp,%esi |
2561 |
– |
addl %ecx,%ebx |
2562 |
– |
shrdl $7,%edx,%edx |
2503 |
|
addl %esi,%ebx |
2564 |
– |
addl 60(%rsp),%eax |
2504 |
|
xorl %ebp,%edi |
2505 |
+ |
shrdl $7,%edx,%edx |
2506 |
+ |
addl %ecx,%ebx |
2507 |
+ |
addl 60(%rsp),%eax |
2508 |
+ |
xorl %edx,%edi |
2509 |
|
movl %ebx,%esi |
2510 |
|
shldl $5,%ebx,%ebx |
2568 |
– |
xorl %edx,%edi |
2569 |
– |
addl %ebx,%eax |
2570 |
– |
shrdl $7,%ecx,%ecx |
2511 |
|
addl %edi,%eax |
2512 |
< |
vmovups %xmm11,48(%r13,%r12,1) |
2512 |
> |
shrdl $7,%ecx,%ecx |
2513 |
> |
addl %ebx,%eax |
2514 |
> |
vmovups %xmm12,48(%r13,%r12,1) |
2515 |
|
leaq 64(%r12),%r12 |
2516 |
|
|
2517 |
|
addl 0(%r9),%eax |
2523 |
|
movl %esi,4(%r9) |
2524 |
|
movl %esi,%ebx |
2525 |
|
movl %ecx,8(%r9) |
2526 |
+ |
movl %ecx,%edi |
2527 |
|
movl %edx,12(%r9) |
2528 |
+ |
xorl %edx,%edi |
2529 |
|
movl %ebp,16(%r9) |
2530 |
+ |
andl %edi,%esi |
2531 |
|
jmp .Loop_avx |
2532 |
|
|
2588 |
– |
.align 16 |
2533 |
|
.Ldone_avx: |
2534 |
|
addl 16(%rsp),%ebx |
2535 |
< |
xorl %eax,%esi |
2535 |
> |
xorl %ebp,%esi |
2536 |
|
movl %ecx,%edi |
2537 |
|
shldl $5,%ecx,%ecx |
2594 |
– |
xorl %ebp,%esi |
2595 |
– |
addl %ecx,%ebx |
2596 |
– |
shrdl $7,%edx,%edx |
2538 |
|
addl %esi,%ebx |
2598 |
– |
addl 20(%rsp),%eax |
2539 |
|
xorl %ebp,%edi |
2540 |
+ |
shrdl $7,%edx,%edx |
2541 |
+ |
addl %ecx,%ebx |
2542 |
+ |
addl 20(%rsp),%eax |
2543 |
+ |
xorl %edx,%edi |
2544 |
|
movl %ebx,%esi |
2545 |
|
shldl $5,%ebx,%ebx |
2602 |
– |
xorl %edx,%edi |
2603 |
– |
addl %ebx,%eax |
2604 |
– |
shrdl $7,%ecx,%ecx |
2546 |
|
addl %edi,%eax |
2547 |
+ |
xorl %edx,%esi |
2548 |
+ |
shrdl $7,%ecx,%ecx |
2549 |
+ |
addl %ebx,%eax |
2550 |
|
addl 24(%rsp),%ebp |
2551 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2551 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2552 |
|
vmovups 16(%r15),%xmm15 |
2553 |
< |
xorl %edx,%esi |
2553 |
> |
xorl %ecx,%esi |
2554 |
|
movl %eax,%edi |
2555 |
|
shldl $5,%eax,%eax |
2612 |
– |
xorl %ecx,%esi |
2613 |
– |
addl %eax,%ebp |
2614 |
– |
shrdl $7,%ebx,%ebx |
2556 |
|
addl %esi,%ebp |
2616 |
– |
addl 28(%rsp),%edx |
2557 |
|
xorl %ecx,%edi |
2558 |
+ |
shrdl $7,%ebx,%ebx |
2559 |
+ |
addl %eax,%ebp |
2560 |
+ |
addl 28(%rsp),%edx |
2561 |
+ |
xorl %ebx,%edi |
2562 |
|
movl %ebp,%esi |
2563 |
|
shldl $5,%ebp,%ebp |
2620 |
– |
xorl %ebx,%edi |
2621 |
– |
addl %ebp,%edx |
2622 |
– |
shrdl $7,%eax,%eax |
2564 |
|
addl %edi,%edx |
2624 |
– |
addl 32(%rsp),%ecx |
2565 |
|
xorl %ebx,%esi |
2566 |
+ |
shrdl $7,%eax,%eax |
2567 |
+ |
addl %ebp,%edx |
2568 |
+ |
addl 32(%rsp),%ecx |
2569 |
+ |
xorl %eax,%esi |
2570 |
|
movl %edx,%edi |
2571 |
|
shldl $5,%edx,%edx |
2572 |
< |
xorl %eax,%esi |
2573 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2572 |
> |
addl %esi,%ecx |
2573 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2574 |
|
vmovups 32(%r15),%xmm14 |
2575 |
< |
addl %edx,%ecx |
2575 |
> |
xorl %eax,%edi |
2576 |
|
shrdl $7,%ebp,%ebp |
2577 |
< |
addl %esi,%ecx |
2577 |
> |
addl %edx,%ecx |
2578 |
|
addl 36(%rsp),%ebx |
2579 |
< |
xorl %eax,%edi |
2579 |
> |
xorl %ebp,%edi |
2580 |
|
movl %ecx,%esi |
2581 |
|
shldl $5,%ecx,%ecx |
2638 |
– |
xorl %ebp,%edi |
2639 |
– |
addl %ecx,%ebx |
2640 |
– |
shrdl $7,%edx,%edx |
2582 |
|
addl %edi,%ebx |
2642 |
– |
addl 40(%rsp),%eax |
2583 |
|
xorl %ebp,%esi |
2584 |
+ |
shrdl $7,%edx,%edx |
2585 |
+ |
addl %ecx,%ebx |
2586 |
+ |
addl 40(%rsp),%eax |
2587 |
+ |
xorl %edx,%esi |
2588 |
|
movl %ebx,%edi |
2589 |
|
shldl $5,%ebx,%ebx |
2646 |
– |
xorl %edx,%esi |
2647 |
– |
addl %ebx,%eax |
2648 |
– |
shrdl $7,%ecx,%ecx |
2590 |
|
addl %esi,%eax |
2591 |
+ |
xorl %edx,%edi |
2592 |
+ |
shrdl $7,%ecx,%ecx |
2593 |
+ |
addl %ebx,%eax |
2594 |
|
addl 44(%rsp),%ebp |
2595 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2595 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2596 |
|
vmovups 48(%r15),%xmm15 |
2597 |
< |
xorl %edx,%edi |
2597 |
> |
xorl %ecx,%edi |
2598 |
|
movl %eax,%esi |
2599 |
|
shldl $5,%eax,%eax |
2656 |
– |
xorl %ecx,%edi |
2657 |
– |
addl %eax,%ebp |
2658 |
– |
shrdl $7,%ebx,%ebx |
2600 |
|
addl %edi,%ebp |
2660 |
– |
addl 48(%rsp),%edx |
2601 |
|
xorl %ecx,%esi |
2602 |
+ |
shrdl $7,%ebx,%ebx |
2603 |
+ |
addl %eax,%ebp |
2604 |
+ |
addl 48(%rsp),%edx |
2605 |
+ |
xorl %ebx,%esi |
2606 |
|
movl %ebp,%edi |
2607 |
|
shldl $5,%ebp,%ebp |
2664 |
– |
xorl %ebx,%esi |
2665 |
– |
addl %ebp,%edx |
2666 |
– |
shrdl $7,%eax,%eax |
2608 |
|
addl %esi,%edx |
2668 |
– |
addl 52(%rsp),%ecx |
2609 |
|
xorl %ebx,%edi |
2610 |
+ |
shrdl $7,%eax,%eax |
2611 |
+ |
addl %ebp,%edx |
2612 |
+ |
addl 52(%rsp),%ecx |
2613 |
+ |
xorl %eax,%edi |
2614 |
|
movl %edx,%esi |
2615 |
|
shldl $5,%edx,%edx |
2616 |
< |
xorl %eax,%edi |
2616 |
> |
addl %edi,%ecx |
2617 |
|
cmpl $11,%r8d |
2618 |
< |
jb .Lvaesenclast5 |
2619 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2618 |
> |
jb .Lvaesenclast10 |
2619 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2620 |
|
vmovups 64(%r15),%xmm14 |
2621 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2621 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2622 |
|
vmovups 80(%r15),%xmm15 |
2623 |
< |
je .Lvaesenclast5 |
2624 |
< |
vaesenc %xmm15,%xmm11,%xmm11 |
2623 |
> |
je .Lvaesenclast10 |
2624 |
> |
vaesenc %xmm15,%xmm12,%xmm12 |
2625 |
|
vmovups 96(%r15),%xmm14 |
2626 |
< |
vaesenc %xmm14,%xmm11,%xmm11 |
2626 |
> |
vaesenc %xmm14,%xmm12,%xmm12 |
2627 |
|
vmovups 112(%r15),%xmm15 |
2628 |
< |
.Lvaesenclast5: |
2629 |
< |
vaesenclast %xmm15,%xmm11,%xmm11 |
2628 |
> |
.Lvaesenclast10: |
2629 |
> |
vaesenclast %xmm15,%xmm12,%xmm12 |
2630 |
> |
vmovups -112(%r15),%xmm15 |
2631 |
|
vmovups 16-112(%r15),%xmm14 |
2632 |
< |
addl %edx,%ecx |
2632 |
> |
xorl %eax,%esi |
2633 |
|
shrdl $7,%ebp,%ebp |
2634 |
< |
addl %edi,%ecx |
2634 |
> |
addl %edx,%ecx |
2635 |
|
addl 56(%rsp),%ebx |
2636 |
< |
xorl %eax,%esi |
2636 |
> |
xorl %ebp,%esi |
2637 |
|
movl %ecx,%edi |
2638 |
|
shldl $5,%ecx,%ecx |
2694 |
– |
xorl %ebp,%esi |
2695 |
– |
addl %ecx,%ebx |
2696 |
– |
shrdl $7,%edx,%edx |
2639 |
|
addl %esi,%ebx |
2698 |
– |
addl 60(%rsp),%eax |
2640 |
|
xorl %ebp,%edi |
2641 |
+ |
shrdl $7,%edx,%edx |
2642 |
+ |
addl %ecx,%ebx |
2643 |
+ |
addl 60(%rsp),%eax |
2644 |
+ |
xorl %edx,%edi |
2645 |
|
movl %ebx,%esi |
2646 |
|
shldl $5,%ebx,%ebx |
2702 |
– |
xorl %edx,%edi |
2703 |
– |
addl %ebx,%eax |
2704 |
– |
shrdl $7,%ecx,%ecx |
2647 |
|
addl %edi,%eax |
2648 |
< |
vmovups %xmm11,48(%r13,%r12,1) |
2648 |
> |
shrdl $7,%ecx,%ecx |
2649 |
> |
addl %ebx,%eax |
2650 |
> |
vmovups %xmm12,48(%r13,%r12,1) |
2651 |
|
movq 88(%rsp),%r8 |
2652 |
|
|
2653 |
|
addl 0(%r9),%eax |
2660 |
|
movl %ecx,8(%r9) |
2661 |
|
movl %edx,12(%r9) |
2662 |
|
movl %ebp,16(%r9) |
2663 |
< |
vmovups %xmm11,(%r8) |
2663 |
> |
vmovups %xmm12,(%r8) |
2664 |
|
vzeroall |
2665 |
|
leaq 104(%rsp),%rsi |
2666 |
|
movq 0(%rsi),%r15 |
2675 |
|
.size aesni_cbc_sha1_enc_avx,.-aesni_cbc_sha1_enc_avx |
2676 |
|
.align 64 |
2677 |
|
K_XX_XX: |
2678 |
< |
.long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 |
2679 |
< |
.long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 |
2680 |
< |
.long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc |
2681 |
< |
.long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 |
2682 |
< |
.long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f |
2678 |
> |
.long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 |
2679 |
> |
.long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 |
2680 |
> |
.long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc |
2681 |
> |
.long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 |
2682 |
> |
.long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f |
2683 |
> |
.byte 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0 |
2684 |
|
|
2685 |
|
.byte 65,69,83,78,73,45,67,66,67,43,83,72,65,49,32,115,116,105,116,99,104,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0 |
2686 |
|
.align 64 |
2687 |
+ |
.type aesni_cbc_sha1_enc_shaext,@function |
2688 |
+ |
.align 32 |
2689 |
+ |
aesni_cbc_sha1_enc_shaext: |
2690 |
+ |
movq 8(%rsp),%r10 |
2691 |
+ |
movdqu (%r9),%xmm8 |
2692 |
+ |
movd 16(%r9),%xmm9 |
2693 |
+ |
movdqa K_XX_XX+80(%rip),%xmm7 |
2694 |
+ |
|
2695 |
+ |
movl 240(%rcx),%r11d |
2696 |
+ |
subq %rdi,%rsi |
2697 |
+ |
movups (%rcx),%xmm15 |
2698 |
+ |
movups (%r8),%xmm2 |
2699 |
+ |
movups 16(%rcx),%xmm0 |
2700 |
+ |
leaq 112(%rcx),%rcx |
2701 |
+ |
|
2702 |
+ |
pshufd $27,%xmm8,%xmm8 |
2703 |
+ |
pshufd $27,%xmm9,%xmm9 |
2704 |
+ |
jmp .Loop_shaext |
2705 |
+ |
|
2706 |
+ |
.align 16 |
2707 |
+ |
.Loop_shaext: |
2708 |
+ |
movups 0(%rdi),%xmm14 |
2709 |
+ |
xorps %xmm15,%xmm14 |
2710 |
+ |
xorps %xmm14,%xmm2 |
2711 |
+ |
movups -80(%rcx),%xmm1 |
2712 |
+ |
.byte 102,15,56,220,208 |
2713 |
+ |
movdqu (%r10),%xmm3 |
2714 |
+ |
movdqa %xmm9,%xmm12 |
2715 |
+ |
.byte 102,15,56,0,223 |
2716 |
+ |
movdqu 16(%r10),%xmm4 |
2717 |
+ |
movdqa %xmm8,%xmm11 |
2718 |
+ |
movups -64(%rcx),%xmm0 |
2719 |
+ |
.byte 102,15,56,220,209 |
2720 |
+ |
.byte 102,15,56,0,231 |
2721 |
+ |
|
2722 |
+ |
paddd %xmm3,%xmm9 |
2723 |
+ |
movdqu 32(%r10),%xmm5 |
2724 |
+ |
leaq 64(%r10),%r10 |
2725 |
+ |
pxor %xmm12,%xmm3 |
2726 |
+ |
movups -48(%rcx),%xmm1 |
2727 |
+ |
.byte 102,15,56,220,208 |
2728 |
+ |
pxor %xmm12,%xmm3 |
2729 |
+ |
movdqa %xmm8,%xmm10 |
2730 |
+ |
.byte 102,15,56,0,239 |
2731 |
+ |
.byte 69,15,58,204,193,0 |
2732 |
+ |
.byte 68,15,56,200,212 |
2733 |
+ |
movups -32(%rcx),%xmm0 |
2734 |
+ |
.byte 102,15,56,220,209 |
2735 |
+ |
.byte 15,56,201,220 |
2736 |
+ |
movdqu -16(%r10),%xmm6 |
2737 |
+ |
movdqa %xmm8,%xmm9 |
2738 |
+ |
.byte 102,15,56,0,247 |
2739 |
+ |
movups -16(%rcx),%xmm1 |
2740 |
+ |
.byte 102,15,56,220,208 |
2741 |
+ |
.byte 69,15,58,204,194,0 |
2742 |
+ |
.byte 68,15,56,200,205 |
2743 |
+ |
pxor %xmm5,%xmm3 |
2744 |
+ |
.byte 15,56,201,229 |
2745 |
+ |
movups 0(%rcx),%xmm0 |
2746 |
+ |
.byte 102,15,56,220,209 |
2747 |
+ |
movdqa %xmm8,%xmm10 |
2748 |
+ |
.byte 69,15,58,204,193,0 |
2749 |
+ |
.byte 68,15,56,200,214 |
2750 |
+ |
movups 16(%rcx),%xmm1 |
2751 |
+ |
.byte 102,15,56,220,208 |
2752 |
+ |
.byte 15,56,202,222 |
2753 |
+ |
pxor %xmm6,%xmm4 |
2754 |
+ |
.byte 15,56,201,238 |
2755 |
+ |
movups 32(%rcx),%xmm0 |
2756 |
+ |
.byte 102,15,56,220,209 |
2757 |
+ |
movdqa %xmm8,%xmm9 |
2758 |
+ |
.byte 69,15,58,204,194,0 |
2759 |
+ |
.byte 68,15,56,200,203 |
2760 |
+ |
movups 48(%rcx),%xmm1 |
2761 |
+ |
.byte 102,15,56,220,208 |
2762 |
+ |
.byte 15,56,202,227 |
2763 |
+ |
pxor %xmm3,%xmm5 |
2764 |
+ |
.byte 15,56,201,243 |
2765 |
+ |
cmpl $11,%r11d |
2766 |
+ |
jb .Laesenclast11 |
2767 |
+ |
movups 64(%rcx),%xmm0 |
2768 |
+ |
.byte 102,15,56,220,209 |
2769 |
+ |
movups 80(%rcx),%xmm1 |
2770 |
+ |
.byte 102,15,56,220,208 |
2771 |
+ |
je .Laesenclast11 |
2772 |
+ |
movups 96(%rcx),%xmm0 |
2773 |
+ |
.byte 102,15,56,220,209 |
2774 |
+ |
movups 112(%rcx),%xmm1 |
2775 |
+ |
.byte 102,15,56,220,208 |
2776 |
+ |
.Laesenclast11: |
2777 |
+ |
.byte 102,15,56,221,209 |
2778 |
+ |
movups 16-112(%rcx),%xmm0 |
2779 |
+ |
movdqa %xmm8,%xmm10 |
2780 |
+ |
.byte 69,15,58,204,193,0 |
2781 |
+ |
.byte 68,15,56,200,212 |
2782 |
+ |
movups 16(%rdi),%xmm14 |
2783 |
+ |
xorps %xmm15,%xmm14 |
2784 |
+ |
movups %xmm2,0(%rsi,%rdi,1) |
2785 |
+ |
xorps %xmm14,%xmm2 |
2786 |
+ |
movups -80(%rcx),%xmm1 |
2787 |
+ |
.byte 102,15,56,220,208 |
2788 |
+ |
.byte 15,56,202,236 |
2789 |
+ |
pxor %xmm4,%xmm6 |
2790 |
+ |
.byte 15,56,201,220 |
2791 |
+ |
movups -64(%rcx),%xmm0 |
2792 |
+ |
.byte 102,15,56,220,209 |
2793 |
+ |
movdqa %xmm8,%xmm9 |
2794 |
+ |
.byte 69,15,58,204,194,1 |
2795 |
+ |
.byte 68,15,56,200,205 |
2796 |
+ |
movups -48(%rcx),%xmm1 |
2797 |
+ |
.byte 102,15,56,220,208 |
2798 |
+ |
.byte 15,56,202,245 |
2799 |
+ |
pxor %xmm5,%xmm3 |
2800 |
+ |
.byte 15,56,201,229 |
2801 |
+ |
movups -32(%rcx),%xmm0 |
2802 |
+ |
.byte 102,15,56,220,209 |
2803 |
+ |
movdqa %xmm8,%xmm10 |
2804 |
+ |
.byte 69,15,58,204,193,1 |
2805 |
+ |
.byte 68,15,56,200,214 |
2806 |
+ |
movups -16(%rcx),%xmm1 |
2807 |
+ |
.byte 102,15,56,220,208 |
2808 |
+ |
.byte 15,56,202,222 |
2809 |
+ |
pxor %xmm6,%xmm4 |
2810 |
+ |
.byte 15,56,201,238 |
2811 |
+ |
movups 0(%rcx),%xmm0 |
2812 |
+ |
.byte 102,15,56,220,209 |
2813 |
+ |
movdqa %xmm8,%xmm9 |
2814 |
+ |
.byte 69,15,58,204,194,1 |
2815 |
+ |
.byte 68,15,56,200,203 |
2816 |
+ |
movups 16(%rcx),%xmm1 |
2817 |
+ |
.byte 102,15,56,220,208 |
2818 |
+ |
.byte 15,56,202,227 |
2819 |
+ |
pxor %xmm3,%xmm5 |
2820 |
+ |
.byte 15,56,201,243 |
2821 |
+ |
movups 32(%rcx),%xmm0 |
2822 |
+ |
.byte 102,15,56,220,209 |
2823 |
+ |
movdqa %xmm8,%xmm10 |
2824 |
+ |
.byte 69,15,58,204,193,1 |
2825 |
+ |
.byte 68,15,56,200,212 |
2826 |
+ |
movups 48(%rcx),%xmm1 |
2827 |
+ |
.byte 102,15,56,220,208 |
2828 |
+ |
.byte 15,56,202,236 |
2829 |
+ |
pxor %xmm4,%xmm6 |
2830 |
+ |
.byte 15,56,201,220 |
2831 |
+ |
cmpl $11,%r11d |
2832 |
+ |
jb .Laesenclast12 |
2833 |
+ |
movups 64(%rcx),%xmm0 |
2834 |
+ |
.byte 102,15,56,220,209 |
2835 |
+ |
movups 80(%rcx),%xmm1 |
2836 |
+ |
.byte 102,15,56,220,208 |
2837 |
+ |
je .Laesenclast12 |
2838 |
+ |
movups 96(%rcx),%xmm0 |
2839 |
+ |
.byte 102,15,56,220,209 |
2840 |
+ |
movups 112(%rcx),%xmm1 |
2841 |
+ |
.byte 102,15,56,220,208 |
2842 |
+ |
.Laesenclast12: |
2843 |
+ |
.byte 102,15,56,221,209 |
2844 |
+ |
movups 16-112(%rcx),%xmm0 |
2845 |
+ |
movdqa %xmm8,%xmm9 |
2846 |
+ |
.byte 69,15,58,204,194,1 |
2847 |
+ |
.byte 68,15,56,200,205 |
2848 |
+ |
movups 32(%rdi),%xmm14 |
2849 |
+ |
xorps %xmm15,%xmm14 |
2850 |
+ |
movups %xmm2,16(%rsi,%rdi,1) |
2851 |
+ |
xorps %xmm14,%xmm2 |
2852 |
+ |
movups -80(%rcx),%xmm1 |
2853 |
+ |
.byte 102,15,56,220,208 |
2854 |
+ |
.byte 15,56,202,245 |
2855 |
+ |
pxor %xmm5,%xmm3 |
2856 |
+ |
.byte 15,56,201,229 |
2857 |
+ |
movups -64(%rcx),%xmm0 |
2858 |
+ |
.byte 102,15,56,220,209 |
2859 |
+ |
movdqa %xmm8,%xmm10 |
2860 |
+ |
.byte 69,15,58,204,193,2 |
2861 |
+ |
.byte 68,15,56,200,214 |
2862 |
+ |
movups -48(%rcx),%xmm1 |
2863 |
+ |
.byte 102,15,56,220,208 |
2864 |
+ |
.byte 15,56,202,222 |
2865 |
+ |
pxor %xmm6,%xmm4 |
2866 |
+ |
.byte 15,56,201,238 |
2867 |
+ |
movups -32(%rcx),%xmm0 |
2868 |
+ |
.byte 102,15,56,220,209 |
2869 |
+ |
movdqa %xmm8,%xmm9 |
2870 |
+ |
.byte 69,15,58,204,194,2 |
2871 |
+ |
.byte 68,15,56,200,203 |
2872 |
+ |
movups -16(%rcx),%xmm1 |
2873 |
+ |
.byte 102,15,56,220,208 |
2874 |
+ |
.byte 15,56,202,227 |
2875 |
+ |
pxor %xmm3,%xmm5 |
2876 |
+ |
.byte 15,56,201,243 |
2877 |
+ |
movups 0(%rcx),%xmm0 |
2878 |
+ |
.byte 102,15,56,220,209 |
2879 |
+ |
movdqa %xmm8,%xmm10 |
2880 |
+ |
.byte 69,15,58,204,193,2 |
2881 |
+ |
.byte 68,15,56,200,212 |
2882 |
+ |
movups 16(%rcx),%xmm1 |
2883 |
+ |
.byte 102,15,56,220,208 |
2884 |
+ |
.byte 15,56,202,236 |
2885 |
+ |
pxor %xmm4,%xmm6 |
2886 |
+ |
.byte 15,56,201,220 |
2887 |
+ |
movups 32(%rcx),%xmm0 |
2888 |
+ |
.byte 102,15,56,220,209 |
2889 |
+ |
movdqa %xmm8,%xmm9 |
2890 |
+ |
.byte 69,15,58,204,194,2 |
2891 |
+ |
.byte 68,15,56,200,205 |
2892 |
+ |
movups 48(%rcx),%xmm1 |
2893 |
+ |
.byte 102,15,56,220,208 |
2894 |
+ |
.byte 15,56,202,245 |
2895 |
+ |
pxor %xmm5,%xmm3 |
2896 |
+ |
.byte 15,56,201,229 |
2897 |
+ |
cmpl $11,%r11d |
2898 |
+ |
jb .Laesenclast13 |
2899 |
+ |
movups 64(%rcx),%xmm0 |
2900 |
+ |
.byte 102,15,56,220,209 |
2901 |
+ |
movups 80(%rcx),%xmm1 |
2902 |
+ |
.byte 102,15,56,220,208 |
2903 |
+ |
je .Laesenclast13 |
2904 |
+ |
movups 96(%rcx),%xmm0 |
2905 |
+ |
.byte 102,15,56,220,209 |
2906 |
+ |
movups 112(%rcx),%xmm1 |
2907 |
+ |
.byte 102,15,56,220,208 |
2908 |
+ |
.Laesenclast13: |
2909 |
+ |
.byte 102,15,56,221,209 |
2910 |
+ |
movups 16-112(%rcx),%xmm0 |
2911 |
+ |
movdqa %xmm8,%xmm10 |
2912 |
+ |
.byte 69,15,58,204,193,2 |
2913 |
+ |
.byte 68,15,56,200,214 |
2914 |
+ |
movups 48(%rdi),%xmm14 |
2915 |
+ |
xorps %xmm15,%xmm14 |
2916 |
+ |
movups %xmm2,32(%rsi,%rdi,1) |
2917 |
+ |
xorps %xmm14,%xmm2 |
2918 |
+ |
movups -80(%rcx),%xmm1 |
2919 |
+ |
.byte 102,15,56,220,208 |
2920 |
+ |
.byte 15,56,202,222 |
2921 |
+ |
pxor %xmm6,%xmm4 |
2922 |
+ |
.byte 15,56,201,238 |
2923 |
+ |
movups -64(%rcx),%xmm0 |
2924 |
+ |
.byte 102,15,56,220,209 |
2925 |
+ |
movdqa %xmm8,%xmm9 |
2926 |
+ |
.byte 69,15,58,204,194,3 |
2927 |
+ |
.byte 68,15,56,200,203 |
2928 |
+ |
movups -48(%rcx),%xmm1 |
2929 |
+ |
.byte 102,15,56,220,208 |
2930 |
+ |
.byte 15,56,202,227 |
2931 |
+ |
pxor %xmm3,%xmm5 |
2932 |
+ |
.byte 15,56,201,243 |
2933 |
+ |
movups -32(%rcx),%xmm0 |
2934 |
+ |
.byte 102,15,56,220,209 |
2935 |
+ |
movdqa %xmm8,%xmm10 |
2936 |
+ |
.byte 69,15,58,204,193,3 |
2937 |
+ |
.byte 68,15,56,200,212 |
2938 |
+ |
.byte 15,56,202,236 |
2939 |
+ |
pxor %xmm4,%xmm6 |
2940 |
+ |
movups -16(%rcx),%xmm1 |
2941 |
+ |
.byte 102,15,56,220,208 |
2942 |
+ |
movdqa %xmm8,%xmm9 |
2943 |
+ |
.byte 69,15,58,204,194,3 |
2944 |
+ |
.byte 68,15,56,200,205 |
2945 |
+ |
.byte 15,56,202,245 |
2946 |
+ |
movups 0(%rcx),%xmm0 |
2947 |
+ |
.byte 102,15,56,220,209 |
2948 |
+ |
movdqa %xmm12,%xmm5 |
2949 |
+ |
movdqa %xmm8,%xmm10 |
2950 |
+ |
.byte 69,15,58,204,193,3 |
2951 |
+ |
.byte 68,15,56,200,214 |
2952 |
+ |
movups 16(%rcx),%xmm1 |
2953 |
+ |
.byte 102,15,56,220,208 |
2954 |
+ |
movdqa %xmm8,%xmm9 |
2955 |
+ |
.byte 69,15,58,204,194,3 |
2956 |
+ |
.byte 68,15,56,200,205 |
2957 |
+ |
movups 32(%rcx),%xmm0 |
2958 |
+ |
.byte 102,15,56,220,209 |
2959 |
+ |
movups 48(%rcx),%xmm1 |
2960 |
+ |
.byte 102,15,56,220,208 |
2961 |
+ |
cmpl $11,%r11d |
2962 |
+ |
jb .Laesenclast14 |
2963 |
+ |
movups 64(%rcx),%xmm0 |
2964 |
+ |
.byte 102,15,56,220,209 |
2965 |
+ |
movups 80(%rcx),%xmm1 |
2966 |
+ |
.byte 102,15,56,220,208 |
2967 |
+ |
je .Laesenclast14 |
2968 |
+ |
movups 96(%rcx),%xmm0 |
2969 |
+ |
.byte 102,15,56,220,209 |
2970 |
+ |
movups 112(%rcx),%xmm1 |
2971 |
+ |
.byte 102,15,56,220,208 |
2972 |
+ |
.Laesenclast14: |
2973 |
+ |
.byte 102,15,56,221,209 |
2974 |
+ |
movups 16-112(%rcx),%xmm0 |
2975 |
+ |
decq %rdx |
2976 |
+ |
|
2977 |
+ |
paddd %xmm11,%xmm8 |
2978 |
+ |
movups %xmm2,48(%rsi,%rdi,1) |
2979 |
+ |
leaq 64(%rdi),%rdi |
2980 |
+ |
jnz .Loop_shaext |
2981 |
+ |
|
2982 |
+ |
pshufd $27,%xmm8,%xmm8 |
2983 |
+ |
pshufd $27,%xmm9,%xmm9 |
2984 |
+ |
movups %xmm2,(%r8) |
2985 |
+ |
movdqu %xmm8,(%r9) |
2986 |
+ |
movd %xmm9,16(%r9) |
2987 |
+ |
.byte 0xf3,0xc3 |
2988 |
+ |
.size aesni_cbc_sha1_enc_shaext,.-aesni_cbc_sha1_enc_shaext |