Context Navigation

source: git/src-cryptopp/x64dll.asm

Last change on this file was e230cb0, checked in by David Stainton <dstainton415@…>, at 2016-10-12T13:27:29Z
Add cryptopp from tag CRYPTOPP_5_6_5
Property mode set to `100644`
File size: 46.2 KB

Line
1	include ksamd64.inc
2	EXTERNDEF ?Te@rdtable@CryptoPP@@3PA_KA:FAR
3	EXTERNDEF ?g_cacheLineSize@CryptoPP@@3IA:FAR
4	EXTERNDEF ?SHA256_K@CryptoPP@@3QBIB:FAR
5	.CODE
6
7	ALIGN 8
8	Baseline_Add PROC
9	lea rdx, [rdx+8*rcx]
10	lea r8, [r8+8*rcx]
11	lea r9, [r9+8*rcx]
12	neg rcx ; rcx is negative index
13	jz $1@Baseline_Add
14	mov rax,[r8+8*rcx]
15	add rax,[r9+8*rcx]
16	mov [rdx+8*rcx],rax
17	$0@Baseline_Add:
18	mov rax,[r8+8*rcx+8]
19	adc rax,[r9+8*rcx+8]
20	mov [rdx+8*rcx+8],rax
21	lea rcx,[rcx+2] ; advance index, avoid inc which causes slowdown on Intel Core 2
22	jrcxz $1@Baseline_Add ; loop until rcx overflows and becomes zero
23	mov rax,[r8+8*rcx]
24	adc rax,[r9+8*rcx]
25	mov [rdx+8*rcx],rax
26	jmp $0@Baseline_Add
27	$1@Baseline_Add:
28	mov rax, 0
29	adc rax, rax ; store carry into rax (return result register)
30	ret
31	Baseline_Add ENDP
32
33	ALIGN 8
34	Baseline_Sub PROC
35	lea rdx, [rdx+8*rcx]
36	lea r8, [r8+8*rcx]
37	lea r9, [r9+8*rcx]
38	neg rcx ; rcx is negative index
39	jz $1@Baseline_Sub
40	mov rax,[r8+8*rcx]
41	sub rax,[r9+8*rcx]
42	mov [rdx+8*rcx],rax
43	$0@Baseline_Sub:
44	mov rax,[r8+8*rcx+8]
45	sbb rax,[r9+8*rcx+8]
46	mov [rdx+8*rcx+8],rax
47	lea rcx,[rcx+2] ; advance index, avoid inc which causes slowdown on Intel Core 2
48	jrcxz $1@Baseline_Sub ; loop until rcx overflows and becomes zero
49	mov rax,[r8+8*rcx]
50	sbb rax,[r9+8*rcx]
51	mov [rdx+8*rcx],rax
52	jmp $0@Baseline_Sub
53	$1@Baseline_Sub:
54	mov rax, 0
55	adc rax, rax ; store carry into rax (return result register)
56
57	ret
58	Baseline_Sub ENDP
59
60	ALIGN 8
61	Rijndael_Enc_AdvancedProcessBlocks PROC FRAME
62	rex_push_reg rsi
63	push_reg rdi
64	push_reg rbx
65	push_reg r12
66	.endprolog
67	mov r8, rcx
68	mov r11, ?Te@rdtable@CryptoPP@@3PA_KA
69	mov edi, DWORD PTR [?g_cacheLineSize@CryptoPP@@3IA]
70	mov rsi, [(r8+16*19)]
71	mov rax, 16
72	and rax, rsi
73	movdqa xmm3, XMMWORD PTR [rdx+16+rax]
74	movdqa [(r8+16*12)], xmm3
75	lea rax, [rdx+rax+2*16]
76	sub rax, rsi
77	label0:
78	movdqa xmm0, [rax+rsi]
79	movdqa XMMWORD PTR [(r8+0)+rsi], xmm0
80	add rsi, 16
81	cmp rsi, 16*12
82	jl label0
83	movdqa xmm4, [rax+rsi]
84	movdqa xmm1, [rdx]
85	mov r12d, [rdx+4*4]
86	mov ebx, [rdx+5*4]
87	mov ecx, [rdx+6*4]
88	mov edx, [rdx+7*4]
89	xor rax, rax
90	label9:
91	mov esi, [r11+rax]
92	add rax, rdi
93	mov esi, [r11+rax]
94	add rax, rdi
95	mov esi, [r11+rax]
96	add rax, rdi
97	mov esi, [r11+rax]
98	add rax, rdi
99	cmp rax, 2048
100	jl label9
101	lfence
102	test DWORD PTR [(r8+16*18+8)], 1
103	jz label8
104	mov rsi, [(r8+16*14)]
105	movdqu xmm2, [rsi]
106	pxor xmm2, xmm1
107	psrldq xmm1, 14
108	movd eax, xmm1
109	mov al, BYTE PTR [rsi+15]
110	mov r10d, eax
111	movd eax, xmm2
112	psrldq xmm2, 4
113	movd edi, xmm2
114	psrldq xmm2, 4
115	movzx esi, al
116	xor r12d, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
117	movzx esi, ah
118	xor edx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
119	shr eax, 16
120	movzx esi, al
121	xor ecx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
122	movzx esi, ah
123	xor ebx, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
124	mov eax, edi
125	movd edi, xmm2
126	psrldq xmm2, 4
127	movzx esi, al
128	xor ebx, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
129	movzx esi, ah
130	xor r12d, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
131	shr eax, 16
132	movzx esi, al
133	xor edx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
134	movzx esi, ah
135	xor ecx, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
136	mov eax, edi
137	movd edi, xmm2
138	movzx esi, al
139	xor ecx, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
140	movzx esi, ah
141	xor ebx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
142	shr eax, 16
143	movzx esi, al
144	xor r12d, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
145	movzx esi, ah
146	xor edx, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
147	mov eax, edi
148	movzx esi, al
149	xor edx, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
150	movzx esi, ah
151	xor ecx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
152	shr eax, 16
153	movzx esi, al
154	xor ebx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
155	psrldq xmm2, 3
156	mov eax, [(r8+1612)+04]
157	mov edi, [(r8+1612)+24]
158	mov r9d, [(r8+1612)+34]
159	movzx esi, cl
160	xor r9d, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
161	movzx esi, bl
162	xor edi, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
163	movzx esi, bh
164	xor r9d, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
165	shr ebx, 16
166	movzx esi, bl
167	xor eax, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
168	movzx esi, bh
169	mov ebx, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
170	xor ebx, [(r8+1612)+14]
171	movzx esi, ch
172	xor eax, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
173	shr ecx, 16
174	movzx esi, dl
175	xor eax, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
176	movzx esi, dh
177	xor ebx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
178	shr edx, 16
179	movzx esi, ch
180	xor edi, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
181	movzx esi, cl
182	xor ebx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
183	movzx esi, dl
184	xor edi, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
185	movzx esi, dh
186	xor r9d, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
187	movd ecx, xmm2
188	mov edx, r12d
189	mov [(r8+0)+3*4], r9d
190	mov [(r8+0)+0*4], eax
191	mov [(r8+0)+1*4], ebx
192	mov [(r8+0)+2*4], edi
193	jmp label5
194	label3:
195	mov r12d, [(r8+1612)+04]
196	mov ebx, [(r8+1612)+14]
197	mov ecx, [(r8+1612)+24]
198	mov edx, [(r8+1612)+34]
199	label8:
200	mov rax, [(r8+16*14)]
201	movdqu xmm2, [rax]
202	mov rsi, [(r8+16*14)+8]
203	movdqu xmm5, [rsi]
204	pxor xmm2, xmm1
205	pxor xmm2, xmm5
206	movd eax, xmm2
207	psrldq xmm2, 4
208	movd edi, xmm2
209	psrldq xmm2, 4
210	movzx esi, al
211	xor r12d, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
212	movzx esi, ah
213	xor edx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
214	shr eax, 16
215	movzx esi, al
216	xor ecx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
217	movzx esi, ah
218	xor ebx, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
219	mov eax, edi
220	movd edi, xmm2
221	psrldq xmm2, 4
222	movzx esi, al
223	xor ebx, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
224	movzx esi, ah
225	xor r12d, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
226	shr eax, 16
227	movzx esi, al
228	xor edx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
229	movzx esi, ah
230	xor ecx, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
231	mov eax, edi
232	movd edi, xmm2
233	movzx esi, al
234	xor ecx, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
235	movzx esi, ah
236	xor ebx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
237	shr eax, 16
238	movzx esi, al
239	xor r12d, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
240	movzx esi, ah
241	xor edx, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
242	mov eax, edi
243	movzx esi, al
244	xor edx, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
245	movzx esi, ah
246	xor ecx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
247	shr eax, 16
248	movzx esi, al
249	xor ebx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
250	movzx esi, ah
251	xor r12d, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
252	mov eax, r12d
253	add r8, [(r8+16*19)]
254	add r8, 4*16
255	jmp label2
256	label1:
257	mov ecx, r10d
258	mov edx, r12d
259	mov eax, [(r8+0)+0*4]
260	mov ebx, [(r8+0)+1*4]
261	xor cl, ch
262	and rcx, 255
263	label5:
264	add r10d, 1
265	xor edx, DWORD PTR [r11+rcx*8+3]
266	movzx esi, dl
267	xor ebx, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
268	movzx esi, dh
269	mov ecx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
270	shr edx, 16
271	xor ecx, [(r8+0)+2*4]
272	movzx esi, dh
273	xor eax, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
274	movzx esi, dl
275	mov edx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
276	xor edx, [(r8+0)+3*4]
277	add r8, [(r8+16*19)]
278	add r8, 3*16
279	jmp label4
280	label2:
281	mov r9d, [(r8+0)-416+34]
282	mov edi, [(r8+0)-416+24]
283	movzx esi, cl
284	xor r9d, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
285	mov cl, al
286	movzx esi, ah
287	xor edi, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
288	shr eax, 16
289	movzx esi, bl
290	xor edi, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
291	movzx esi, bh
292	xor r9d, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
293	shr ebx, 16
294	movzx esi, al
295	xor r9d, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
296	movzx esi, ah
297	mov eax, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
298	movzx esi, bl
299	xor eax, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
300	movzx esi, bh
301	mov ebx, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
302	movzx esi, ch
303	xor eax, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
304	movzx esi, cl
305	xor ebx, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
306	shr ecx, 16
307	movzx esi, dl
308	xor eax, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
309	movzx esi, dh
310	xor ebx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
311	shr edx, 16
312	movzx esi, ch
313	xor edi, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
314	movzx esi, cl
315	xor ebx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
316	movzx esi, dl
317	xor edi, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
318	movzx esi, dh
319	xor r9d, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
320	mov ecx, edi
321	xor eax, [(r8+0)-416+04]
322	xor ebx, [(r8+0)-416+14]
323	mov edx, r9d
324	label4:
325	mov r9d, [(r8+0)-416+74]
326	mov edi, [(r8+0)-416+64]
327	movzx esi, cl
328	xor r9d, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
329	mov cl, al
330	movzx esi, ah
331	xor edi, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
332	shr eax, 16
333	movzx esi, bl
334	xor edi, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
335	movzx esi, bh
336	xor r9d, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
337	shr ebx, 16
338	movzx esi, al
339	xor r9d, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
340	movzx esi, ah
341	mov eax, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
342	movzx esi, bl
343	xor eax, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
344	movzx esi, bh
345	mov ebx, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
346	movzx esi, ch
347	xor eax, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
348	movzx esi, cl
349	xor ebx, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
350	shr ecx, 16
351	movzx esi, dl
352	xor eax, DWORD PTR [r11+8*rsi+(((3+3) MOD (4))+1)]
353	movzx esi, dh
354	xor ebx, DWORD PTR [r11+8*rsi+(((2+3) MOD (4))+1)]
355	shr edx, 16
356	movzx esi, ch
357	xor edi, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
358	movzx esi, cl
359	xor ebx, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
360	movzx esi, dl
361	xor edi, DWORD PTR [r11+8*rsi+(((1+3) MOD (4))+1)]
362	movzx esi, dh
363	xor r9d, DWORD PTR [r11+8*rsi+(((0+3) MOD (4))+1)]
364	mov ecx, edi
365	xor eax, [(r8+0)-416+44]
366	xor ebx, [(r8+0)-416+54]
367	mov edx, r9d
368	add r8, 32
369	test r8, 255
370	jnz label2
371	sub r8, 16*16
372	movzx esi, ch
373	movzx edi, BYTE PTR [r11+rsi*8+1]
374	movzx esi, dl
375	xor edi, DWORD PTR [r11+rsi*8+0]
376	mov WORD PTR [(r8+16*13)+2], di
377	movzx esi, dh
378	movzx edi, BYTE PTR [r11+rsi*8+1]
379	movzx esi, al
380	xor edi, DWORD PTR [r11+rsi*8+0]
381	mov WORD PTR [(r8+16*13)+6], di
382	shr edx, 16
383	movzx esi, ah
384	movzx edi, BYTE PTR [r11+rsi*8+1]
385	movzx esi, bl
386	xor edi, DWORD PTR [r11+rsi*8+0]
387	mov WORD PTR [(r8+16*13)+10], di
388	shr eax, 16
389	movzx esi, bh
390	movzx edi, BYTE PTR [r11+rsi*8+1]
391	movzx esi, cl
392	xor edi, DWORD PTR [r11+rsi*8+0]
393	mov WORD PTR [(r8+16*13)+14], di
394	shr ebx, 16
395	movzx esi, dh
396	movzx edi, BYTE PTR [r11+rsi*8+1]
397	movzx esi, al
398	xor edi, DWORD PTR [r11+rsi*8+0]
399	mov WORD PTR [(r8+16*13)+12], di
400	shr ecx, 16
401	movzx esi, ah
402	movzx edi, BYTE PTR [r11+rsi*8+1]
403	movzx esi, bl
404	xor edi, DWORD PTR [r11+rsi*8+0]
405	mov WORD PTR [(r8+16*13)+0], di
406	movzx esi, bh
407	movzx edi, BYTE PTR [r11+rsi*8+1]
408	movzx esi, cl
409	xor edi, DWORD PTR [r11+rsi*8+0]
410	mov WORD PTR [(r8+16*13)+4], di
411	movzx esi, ch
412	movzx edi, BYTE PTR [r11+rsi*8+1]
413	movzx esi, dl
414	xor edi, DWORD PTR [r11+rsi*8+0]
415	mov WORD PTR [(r8+16*13)+8], di
416	mov rax, [(r8+16*14)+16]
417	mov rbx, [(r8+16*14)+24]
418	mov rcx, [(r8+16*18+8)]
419	sub rcx, 16
420	movdqu xmm2, [rax]
421	pxor xmm2, xmm4
422	movdqa xmm0, [(r8+16*16)+16]
423	paddq xmm0, [(r8+16*14)+16]
424	movdqa [(r8+16*14)+16], xmm0
425	pxor xmm2, [(r8+16*13)]
426	movdqu [rbx], xmm2
427	jle label7
428	mov [(r8+16*18+8)], rcx
429	test rcx, 1
430	jnz label1
431	movdqa xmm0, [(r8+16*16)]
432	paddq xmm0, [(r8+16*14)]
433	movdqa [(r8+16*14)], xmm0
434	jmp label3
435	label7:
436	xorps xmm0, xmm0
437	lea rax, [(r8+0)+7*16]
438	movaps [rax-7*16], xmm0
439	movaps [rax-6*16], xmm0
440	movaps [rax-5*16], xmm0
441	movaps [rax-4*16], xmm0
442	movaps [rax-3*16], xmm0
443	movaps [rax-2*16], xmm0
444	movaps [rax-1*16], xmm0
445	movaps [rax+0*16], xmm0
446	movaps [rax+1*16], xmm0
447	movaps [rax+2*16], xmm0
448	movaps [rax+3*16], xmm0
449	movaps [rax+4*16], xmm0
450	movaps [rax+5*16], xmm0
451	movaps [rax+6*16], xmm0
452	pop r12
453	pop rbx
454	pop rdi
455	pop rsi
456	ret
457	Rijndael_Enc_AdvancedProcessBlocks ENDP
458
459	ALIGN 8
460	GCM_AuthenticateBlocks_2K PROC FRAME
461	rex_push_reg rsi
462	push_reg rdi
463	push_reg rbx
464	.endprolog
465	mov rsi, r8
466	mov r11, r9
467	movdqa xmm0, [rsi]
468	label0:
469	movdqu xmm4, [rcx]
470	pxor xmm0, xmm4
471	movd ebx, xmm0
472	mov eax, 0f0f0f0f0h
473	and eax, ebx
474	shl ebx, 4
475	and ebx, 0f0f0f0f0h
476	movzx edi, ah
477	movdqa xmm5, XMMWORD PTR [rsi + 32 + 1024 + rdi]
478	movzx edi, al
479	movdqa xmm4, XMMWORD PTR [rsi + 32 + 1024 + rdi]
480	shr eax, 16
481	movzx edi, ah
482	movdqa xmm3, XMMWORD PTR [rsi + 32 + 1024 + rdi]
483	movzx edi, al
484	movdqa xmm2, XMMWORD PTR [rsi + 32 + 1024 + rdi]
485	psrldq xmm0, 4
486	movd eax, xmm0
487	and eax, 0f0f0f0f0h
488	movzx edi, bh
489	pxor xmm5, XMMWORD PTR [rsi + 32 + (1-1)*256 + rdi]
490	movzx edi, bl
491	pxor xmm4, XMMWORD PTR [rsi + 32 + (1-1)*256 + rdi]
492	shr ebx, 16
493	movzx edi, bh
494	pxor xmm3, XMMWORD PTR [rsi + 32 + (1-1)*256 + rdi]
495	movzx edi, bl
496	pxor xmm2, XMMWORD PTR [rsi + 32 + (1-1)*256 + rdi]
497	movd ebx, xmm0
498	shl ebx, 4
499	and ebx, 0f0f0f0f0h
500	movzx edi, ah
501	pxor xmm5, XMMWORD PTR [rsi + 32 + 1024 + 1*256 + rdi]
502	movzx edi, al
503	pxor xmm4, XMMWORD PTR [rsi + 32 + 1024 + 1*256 + rdi]
504	shr eax, 16
505	movzx edi, ah
506	pxor xmm3, XMMWORD PTR [rsi + 32 + 1024 + 1*256 + rdi]
507	movzx edi, al
508	pxor xmm2, XMMWORD PTR [rsi + 32 + 1024 + 1*256 + rdi]
509	psrldq xmm0, 4
510	movd eax, xmm0
511	and eax, 0f0f0f0f0h
512	movzx edi, bh
513	pxor xmm5, XMMWORD PTR [rsi + 32 + (2-1)*256 + rdi]
514	movzx edi, bl
515	pxor xmm4, XMMWORD PTR [rsi + 32 + (2-1)*256 + rdi]
516	shr ebx, 16
517	movzx edi, bh
518	pxor xmm3, XMMWORD PTR [rsi + 32 + (2-1)*256 + rdi]
519	movzx edi, bl
520	pxor xmm2, XMMWORD PTR [rsi + 32 + (2-1)*256 + rdi]
521	movd ebx, xmm0
522	shl ebx, 4
523	and ebx, 0f0f0f0f0h
524	movzx edi, ah
525	pxor xmm5, XMMWORD PTR [rsi + 32 + 1024 + 2*256 + rdi]
526	movzx edi, al
527	pxor xmm4, XMMWORD PTR [rsi + 32 + 1024 + 2*256 + rdi]
528	shr eax, 16
529	movzx edi, ah
530	pxor xmm3, XMMWORD PTR [rsi + 32 + 1024 + 2*256 + rdi]
531	movzx edi, al
532	pxor xmm2, XMMWORD PTR [rsi + 32 + 1024 + 2*256 + rdi]
533	psrldq xmm0, 4
534	movd eax, xmm0
535	and eax, 0f0f0f0f0h
536	movzx edi, bh
537	pxor xmm5, XMMWORD PTR [rsi + 32 + (3-1)*256 + rdi]
538	movzx edi, bl
539	pxor xmm4, XMMWORD PTR [rsi + 32 + (3-1)*256 + rdi]
540	shr ebx, 16
541	movzx edi, bh
542	pxor xmm3, XMMWORD PTR [rsi + 32 + (3-1)*256 + rdi]
543	movzx edi, bl
544	pxor xmm2, XMMWORD PTR [rsi + 32 + (3-1)*256 + rdi]
545	movd ebx, xmm0
546	shl ebx, 4
547	and ebx, 0f0f0f0f0h
548	movzx edi, ah
549	pxor xmm5, XMMWORD PTR [rsi + 32 + 1024 + 3*256 + rdi]
550	movzx edi, al
551	pxor xmm4, XMMWORD PTR [rsi + 32 + 1024 + 3*256 + rdi]
552	shr eax, 16
553	movzx edi, ah
554	pxor xmm3, XMMWORD PTR [rsi + 32 + 1024 + 3*256 + rdi]
555	movzx edi, al
556	pxor xmm2, XMMWORD PTR [rsi + 32 + 1024 + 3*256 + rdi]
557	movzx edi, bh
558	pxor xmm5, XMMWORD PTR [rsi + 32 + 3*256 + rdi]
559	movzx edi, bl
560	pxor xmm4, XMMWORD PTR [rsi + 32 + 3*256 + rdi]
561	shr ebx, 16
562	movzx edi, bh
563	pxor xmm3, XMMWORD PTR [rsi + 32 + 3*256 + rdi]
564	movzx edi, bl
565	pxor xmm2, XMMWORD PTR [rsi + 32 + 3*256 + rdi]
566	movdqa xmm0, xmm3
567	pslldq xmm3, 1
568	pxor xmm2, xmm3
569	movdqa xmm1, xmm2
570	pslldq xmm2, 1
571	pxor xmm5, xmm2
572	psrldq xmm0, 15
573	movd rdi, xmm0
574	movzx eax, WORD PTR [r11 + rdi*2]
575	shl eax, 8
576	movdqa xmm0, xmm5
577	pslldq xmm5, 1
578	pxor xmm4, xmm5
579	psrldq xmm1, 15
580	movd rdi, xmm1
581	xor ax, WORD PTR [r11 + rdi*2]
582	shl eax, 8
583	psrldq xmm0, 15
584	movd rdi, xmm0
585	xor ax, WORD PTR [r11 + rdi*2]
586	movd xmm0, eax
587	pxor xmm0, xmm4
588	add rcx, 16
589	sub rdx, 1
590	jnz label0
591	movdqa [rsi], xmm0
592	pop rbx
593	pop rdi
594	pop rsi
595	ret
596	GCM_AuthenticateBlocks_2K ENDP
597
598	ALIGN 8
599	GCM_AuthenticateBlocks_64K PROC FRAME
600	rex_push_reg rsi
601	push_reg rdi
602	.endprolog
603	mov rsi, r8
604	movdqa xmm0, [rsi]
605	label1:
606	movdqu xmm1, [rcx]
607	pxor xmm1, xmm0
608	pxor xmm0, xmm0
609	movd eax, xmm1
610	psrldq xmm1, 4
611	movzx edi, al
612	add rdi, rdi
613	pxor xmm0, [rsi + 32 + (04+0)25616 + rdi8]
614	movzx edi, ah
615	add rdi, rdi
616	pxor xmm0, [rsi + 32 + (04+1)25616 + rdi8]
617	shr eax, 16
618	movzx edi, al
619	add rdi, rdi
620	pxor xmm0, [rsi + 32 + (04+2)25616 + rdi8]
621	movzx edi, ah
622	add rdi, rdi
623	pxor xmm0, [rsi + 32 + (04+3)25616 + rdi8]
624	movd eax, xmm1
625	psrldq xmm1, 4
626	movzx edi, al
627	add rdi, rdi
628	pxor xmm0, [rsi + 32 + (14+0)25616 + rdi8]
629	movzx edi, ah
630	add rdi, rdi
631	pxor xmm0, [rsi + 32 + (14+1)25616 + rdi8]
632	shr eax, 16
633	movzx edi, al
634	add rdi, rdi
635	pxor xmm0, [rsi + 32 + (14+2)25616 + rdi8]
636	movzx edi, ah
637	add rdi, rdi
638	pxor xmm0, [rsi + 32 + (14+3)25616 + rdi8]
639	movd eax, xmm1
640	psrldq xmm1, 4
641	movzx edi, al
642	add rdi, rdi
643	pxor xmm0, [rsi + 32 + (24+0)25616 + rdi8]
644	movzx edi, ah
645	add rdi, rdi
646	pxor xmm0, [rsi + 32 + (24+1)25616 + rdi8]
647	shr eax, 16
648	movzx edi, al
649	add rdi, rdi
650	pxor xmm0, [rsi + 32 + (24+2)25616 + rdi8]
651	movzx edi, ah
652	add rdi, rdi
653	pxor xmm0, [rsi + 32 + (24+3)25616 + rdi8]
654	movd eax, xmm1
655	psrldq xmm1, 4
656	movzx edi, al
657	add rdi, rdi
658	pxor xmm0, [rsi + 32 + (34+0)25616 + rdi8]
659	movzx edi, ah
660	add rdi, rdi
661	pxor xmm0, [rsi + 32 + (34+1)25616 + rdi8]
662	shr eax, 16
663	movzx edi, al
664	add rdi, rdi
665	pxor xmm0, [rsi + 32 + (34+2)25616 + rdi8]
666	movzx edi, ah
667	add rdi, rdi
668	pxor xmm0, [rsi + 32 + (34+3)25616 + rdi8]
669	add rcx, 16
670	sub rdx, 1
671	jnz label1
672	movdqa [rsi], xmm0
673	pop rdi
674	pop rsi
675	ret
676	GCM_AuthenticateBlocks_64K ENDP
677
678	ALIGN 8
679	X86_SHA256_HashBlocks PROC FRAME
680	rex_push_reg rsi
681	push_reg rdi
682	push_reg rbx
683	push_reg rbp
684	alloc_stack(84 + 164 + 4*8 + 8)
685	.endprolog
686	mov rdi, r8
687	lea rsi, [?SHA256_K@CryptoPP@@3QBIB + 48*4]
688	mov [rsp+84+164+1*8], rcx
689	mov [rsp+84+164+2*8], rdx
690	add rdi, rdx
691	mov [rsp+84+164+3*8], rdi
692	movdqa xmm0, XMMWORD PTR [rcx+0*16]
693	movdqa xmm1, XMMWORD PTR [rcx+1*16]
694	mov [rsp+84+164+0*8], rsi
695	label0:
696	sub rsi, 48*4
697	movdqa [rsp+((1024+7-(0+3)) MOD (8))*4], xmm1
698	movdqa [rsp+((1024+7-(0+7)) MOD (8))*4], xmm0
699	mov rbx, [rdx+0*8]
700	bswap rbx
701	mov [rsp+84+((1024+15-(0(1+1)+1)) MOD (16))*4], rbx
702	mov rbx, [rdx+1*8]
703	bswap rbx
704	mov [rsp+84+((1024+15-(1(1+1)+1)) MOD (16))*4], rbx
705	mov rbx, [rdx+2*8]
706	bswap rbx
707	mov [rsp+84+((1024+15-(2(1+1)+1)) MOD (16))*4], rbx
708	mov rbx, [rdx+3*8]
709	bswap rbx
710	mov [rsp+84+((1024+15-(3(1+1)+1)) MOD (16))*4], rbx
711	mov rbx, [rdx+4*8]
712	bswap rbx
713	mov [rsp+84+((1024+15-(4(1+1)+1)) MOD (16))*4], rbx
714	mov rbx, [rdx+5*8]
715	bswap rbx
716	mov [rsp+84+((1024+15-(5(1+1)+1)) MOD (16))*4], rbx
717	mov rbx, [rdx+6*8]
718	bswap rbx
719	mov [rsp+84+((1024+15-(6(1+1)+1)) MOD (16))*4], rbx
720	mov rbx, [rdx+7*8]
721	bswap rbx
722	mov [rsp+84+((1024+15-(7(1+1)+1)) MOD (16))*4], rbx
723	mov edi, [rsp+((1024+7-(0+3)) MOD (8))*4]
724	mov eax, [rsp+((1024+7-(0+6)) MOD (8))*4]
725	xor eax, [rsp+((1024+7-(0+5)) MOD (8))*4]
726	mov ecx, [rsp+((1024+7-(0+7)) MOD (8))*4]
727	mov edx, [rsp+((1024+7-(0+2)) MOD (8))*4]
728	xor edx, [rsp+((1024+7-(0+1)) MOD (8))*4]
729	and edx, edi
730	xor edx, [rsp+((1024+7-(0+1)) MOD (8))*4]
731	mov ebp, edi
732	ror edi, 6
733	ror ebp, 25
734	add edx, [rsi+(0)*4]
735	add edx, [rsp+84+((1024+15-(0)) MOD (16))4]
736	add edx, [rsp+((1024+7-(0)) MOD (8))*4]
737	xor ebp, edi
738	ror edi, 5
739	xor ebp, edi
740	add edx, ebp
741	mov ebx, ecx
742	xor ecx, [rsp+((1024+7-(0+6)) MOD (8))*4]
743	and eax, ecx
744	xor eax, [rsp+((1024+7-(0+6)) MOD (8))*4]
745	mov ebp, ebx
746	ror ebx, 2
747	add eax, edx
748	add edx, [rsp+((1024+7-(0+4)) MOD (8))*4]
749	mov [rsp+((1024+7-(0+4)) MOD (8))*4], edx
750	ror ebp, 22
751	xor ebp, ebx
752	ror ebx, 11
753	xor ebp, ebx
754	add eax, ebp
755	mov [rsp+((1024+7-(0)) MOD (8))*4], eax
756	mov edi, [rsp+((1024+7-(1+2)) MOD (8))*4]
757	xor edi, [rsp+((1024+7-(1+1)) MOD (8))*4]
758	and edi, edx
759	xor edi, [rsp+((1024+7-(1+1)) MOD (8))*4]
760	mov ebp, edx
761	ror edx, 6
762	ror ebp, 25
763	add edi, [rsi+(1)*4]
764	add edi, [rsp+84+((1024+15-(1)) MOD (16))4]
765	add edi, [rsp+((1024+7-(1)) MOD (8))*4]
766	xor ebp, edx
767	ror edx, 5
768	xor ebp, edx
769	add edi, ebp
770	mov ebx, eax
771	xor eax, [rsp+((1024+7-(1+6)) MOD (8))*4]
772	and ecx, eax
773	xor ecx, [rsp+((1024+7-(1+6)) MOD (8))*4]
774	mov ebp, ebx
775	ror ebx, 2
776	add ecx, edi
777	add edi, [rsp+((1024+7-(1+4)) MOD (8))*4]
778	mov [rsp+((1024+7-(1+4)) MOD (8))*4], edi
779	ror ebp, 22
780	xor ebp, ebx
781	ror ebx, 11
782	xor ebp, ebx
783	add ecx, ebp
784	mov [rsp+((1024+7-(1)) MOD (8))*4], ecx
785	mov edx, [rsp+((1024+7-(2+2)) MOD (8))*4]
786	xor edx, [rsp+((1024+7-(2+1)) MOD (8))*4]
787	and edx, edi
788	xor edx, [rsp+((1024+7-(2+1)) MOD (8))*4]
789	mov ebp, edi
790	ror edi, 6
791	ror ebp, 25
792	add edx, [rsi+(2)*4]
793	add edx, [rsp+84+((1024+15-(2)) MOD (16))4]
794	add edx, [rsp+((1024+7-(2)) MOD (8))*4]
795	xor ebp, edi
796	ror edi, 5
797	xor ebp, edi
798	add edx, ebp
799	mov ebx, ecx
800	xor ecx, [rsp+((1024+7-(2+6)) MOD (8))*4]
801	and eax, ecx
802	xor eax, [rsp+((1024+7-(2+6)) MOD (8))*4]
803	mov ebp, ebx
804	ror ebx, 2
805	add eax, edx
806	add edx, [rsp+((1024+7-(2+4)) MOD (8))*4]
807	mov [rsp+((1024+7-(2+4)) MOD (8))*4], edx
808	ror ebp, 22
809	xor ebp, ebx
810	ror ebx, 11
811	xor ebp, ebx
812	add eax, ebp
813	mov [rsp+((1024+7-(2)) MOD (8))*4], eax
814	mov edi, [rsp+((1024+7-(3+2)) MOD (8))*4]
815	xor edi, [rsp+((1024+7-(3+1)) MOD (8))*4]
816	and edi, edx
817	xor edi, [rsp+((1024+7-(3+1)) MOD (8))*4]
818	mov ebp, edx
819	ror edx, 6
820	ror ebp, 25
821	add edi, [rsi+(3)*4]
822	add edi, [rsp+84+((1024+15-(3)) MOD (16))4]
823	add edi, [rsp+((1024+7-(3)) MOD (8))*4]
824	xor ebp, edx
825	ror edx, 5
826	xor ebp, edx
827	add edi, ebp
828	mov ebx, eax
829	xor eax, [rsp+((1024+7-(3+6)) MOD (8))*4]
830	and ecx, eax
831	xor ecx, [rsp+((1024+7-(3+6)) MOD (8))*4]
832	mov ebp, ebx
833	ror ebx, 2
834	add ecx, edi
835	add edi, [rsp+((1024+7-(3+4)) MOD (8))*4]
836	mov [rsp+((1024+7-(3+4)) MOD (8))*4], edi
837	ror ebp, 22
838	xor ebp, ebx
839	ror ebx, 11
840	xor ebp, ebx
841	add ecx, ebp
842	mov [rsp+((1024+7-(3)) MOD (8))*4], ecx
843	mov edx, [rsp+((1024+7-(4+2)) MOD (8))*4]
844	xor edx, [rsp+((1024+7-(4+1)) MOD (8))*4]
845	and edx, edi
846	xor edx, [rsp+((1024+7-(4+1)) MOD (8))*4]
847	mov ebp, edi
848	ror edi, 6
849	ror ebp, 25
850	add edx, [rsi+(4)*4]
851	add edx, [rsp+84+((1024+15-(4)) MOD (16))4]
852	add edx, [rsp+((1024+7-(4)) MOD (8))*4]
853	xor ebp, edi
854	ror edi, 5
855	xor ebp, edi
856	add edx, ebp
857	mov ebx, ecx
858	xor ecx, [rsp+((1024+7-(4+6)) MOD (8))*4]
859	and eax, ecx
860	xor eax, [rsp+((1024+7-(4+6)) MOD (8))*4]
861	mov ebp, ebx
862	ror ebx, 2
863	add eax, edx
864	add edx, [rsp+((1024+7-(4+4)) MOD (8))*4]
865	mov [rsp+((1024+7-(4+4)) MOD (8))*4], edx
866	ror ebp, 22
867	xor ebp, ebx
868	ror ebx, 11
869	xor ebp, ebx
870	add eax, ebp
871	mov [rsp+((1024+7-(4)) MOD (8))*4], eax
872	mov edi, [rsp+((1024+7-(5+2)) MOD (8))*4]
873	xor edi, [rsp+((1024+7-(5+1)) MOD (8))*4]
874	and edi, edx
875	xor edi, [rsp+((1024+7-(5+1)) MOD (8))*4]
876	mov ebp, edx
877	ror edx, 6
878	ror ebp, 25
879	add edi, [rsi+(5)*4]
880	add edi, [rsp+84+((1024+15-(5)) MOD (16))4]
881	add edi, [rsp+((1024+7-(5)) MOD (8))*4]
882	xor ebp, edx
883	ror edx, 5
884	xor ebp, edx
885	add edi, ebp
886	mov ebx, eax
887	xor eax, [rsp+((1024+7-(5+6)) MOD (8))*4]
888	and ecx, eax
889	xor ecx, [rsp+((1024+7-(5+6)) MOD (8))*4]
890	mov ebp, ebx
891	ror ebx, 2
892	add ecx, edi
893	add edi, [rsp+((1024+7-(5+4)) MOD (8))*4]
894	mov [rsp+((1024+7-(5+4)) MOD (8))*4], edi
895	ror ebp, 22
896	xor ebp, ebx
897	ror ebx, 11
898	xor ebp, ebx
899	add ecx, ebp
900	mov [rsp+((1024+7-(5)) MOD (8))*4], ecx
901	mov edx, [rsp+((1024+7-(6+2)) MOD (8))*4]
902	xor edx, [rsp+((1024+7-(6+1)) MOD (8))*4]
903	and edx, edi
904	xor edx, [rsp+((1024+7-(6+1)) MOD (8))*4]
905	mov ebp, edi
906	ror edi, 6
907	ror ebp, 25
908	add edx, [rsi+(6)*4]
909	add edx, [rsp+84+((1024+15-(6)) MOD (16))4]
910	add edx, [rsp+((1024+7-(6)) MOD (8))*4]
911	xor ebp, edi
912	ror edi, 5
913	xor ebp, edi
914	add edx, ebp
915	mov ebx, ecx
916	xor ecx, [rsp+((1024+7-(6+6)) MOD (8))*4]
917	and eax, ecx
918	xor eax, [rsp+((1024+7-(6+6)) MOD (8))*4]
919	mov ebp, ebx
920	ror ebx, 2
921	add eax, edx
922	add edx, [rsp+((1024+7-(6+4)) MOD (8))*4]
923	mov [rsp+((1024+7-(6+4)) MOD (8))*4], edx
924	ror ebp, 22
925	xor ebp, ebx
926	ror ebx, 11
927	xor ebp, ebx
928	add eax, ebp
929	mov [rsp+((1024+7-(6)) MOD (8))*4], eax
930	mov edi, [rsp+((1024+7-(7+2)) MOD (8))*4]
931	xor edi, [rsp+((1024+7-(7+1)) MOD (8))*4]
932	and edi, edx
933	xor edi, [rsp+((1024+7-(7+1)) MOD (8))*4]
934	mov ebp, edx
935	ror edx, 6
936	ror ebp, 25
937	add edi, [rsi+(7)*4]
938	add edi, [rsp+84+((1024+15-(7)) MOD (16))4]
939	add edi, [rsp+((1024+7-(7)) MOD (8))*4]
940	xor ebp, edx
941	ror edx, 5
942	xor ebp, edx
943	add edi, ebp
944	mov ebx, eax
945	xor eax, [rsp+((1024+7-(7+6)) MOD (8))*4]
946	and ecx, eax
947	xor ecx, [rsp+((1024+7-(7+6)) MOD (8))*4]
948	mov ebp, ebx
949	ror ebx, 2
950	add ecx, edi
951	add edi, [rsp+((1024+7-(7+4)) MOD (8))*4]
952	mov [rsp+((1024+7-(7+4)) MOD (8))*4], edi
953	ror ebp, 22
954	xor ebp, ebx
955	ror ebx, 11
956	xor ebp, ebx
957	add ecx, ebp
958	mov [rsp+((1024+7-(7)) MOD (8))*4], ecx
959	mov edx, [rsp+((1024+7-(8+2)) MOD (8))*4]
960	xor edx, [rsp+((1024+7-(8+1)) MOD (8))*4]
961	and edx, edi
962	xor edx, [rsp+((1024+7-(8+1)) MOD (8))*4]
963	mov ebp, edi
964	ror edi, 6
965	ror ebp, 25
966	add edx, [rsi+(8)*4]
967	add edx, [rsp+84+((1024+15-(8)) MOD (16))4]
968	add edx, [rsp+((1024+7-(8)) MOD (8))*4]
969	xor ebp, edi
970	ror edi, 5
971	xor ebp, edi
972	add edx, ebp
973	mov ebx, ecx
974	xor ecx, [rsp+((1024+7-(8+6)) MOD (8))*4]
975	and eax, ecx
976	xor eax, [rsp+((1024+7-(8+6)) MOD (8))*4]
977	mov ebp, ebx
978	ror ebx, 2
979	add eax, edx
980	add edx, [rsp+((1024+7-(8+4)) MOD (8))*4]
981	mov [rsp+((1024+7-(8+4)) MOD (8))*4], edx
982	ror ebp, 22
983	xor ebp, ebx
984	ror ebx, 11
985	xor ebp, ebx
986	add eax, ebp
987	mov [rsp+((1024+7-(8)) MOD (8))*4], eax
988	mov edi, [rsp+((1024+7-(9+2)) MOD (8))*4]
989	xor edi, [rsp+((1024+7-(9+1)) MOD (8))*4]
990	and edi, edx
991	xor edi, [rsp+((1024+7-(9+1)) MOD (8))*4]
992	mov ebp, edx
993	ror edx, 6
994	ror ebp, 25
995	add edi, [rsi+(9)*4]
996	add edi, [rsp+84+((1024+15-(9)) MOD (16))4]
997	add edi, [rsp+((1024+7-(9)) MOD (8))*4]
998	xor ebp, edx
999	ror edx, 5
1000	xor ebp, edx
1001	add edi, ebp
1002	mov ebx, eax
1003	xor eax, [rsp+((1024+7-(9+6)) MOD (8))*4]
1004	and ecx, eax
1005	xor ecx, [rsp+((1024+7-(9+6)) MOD (8))*4]
1006	mov ebp, ebx
1007	ror ebx, 2
1008	add ecx, edi
1009	add edi, [rsp+((1024+7-(9+4)) MOD (8))*4]
1010	mov [rsp+((1024+7-(9+4)) MOD (8))*4], edi
1011	ror ebp, 22
1012	xor ebp, ebx
1013	ror ebx, 11
1014	xor ebp, ebx
1015	add ecx, ebp
1016	mov [rsp+((1024+7-(9)) MOD (8))*4], ecx
1017	mov edx, [rsp+((1024+7-(10+2)) MOD (8))*4]
1018	xor edx, [rsp+((1024+7-(10+1)) MOD (8))*4]
1019	and edx, edi
1020	xor edx, [rsp+((1024+7-(10+1)) MOD (8))*4]
1021	mov ebp, edi
1022	ror edi, 6
1023	ror ebp, 25
1024	add edx, [rsi+(10)*4]
1025	add edx, [rsp+84+((1024+15-(10)) MOD (16))4]
1026	add edx, [rsp+((1024+7-(10)) MOD (8))*4]
1027	xor ebp, edi
1028	ror edi, 5
1029	xor ebp, edi
1030	add edx, ebp
1031	mov ebx, ecx
1032	xor ecx, [rsp+((1024+7-(10+6)) MOD (8))*4]
1033	and eax, ecx
1034	xor eax, [rsp+((1024+7-(10+6)) MOD (8))*4]
1035	mov ebp, ebx
1036	ror ebx, 2
1037	add eax, edx
1038	add edx, [rsp+((1024+7-(10+4)) MOD (8))*4]
1039	mov [rsp+((1024+7-(10+4)) MOD (8))*4], edx
1040	ror ebp, 22
1041	xor ebp, ebx
1042	ror ebx, 11
1043	xor ebp, ebx
1044	add eax, ebp
1045	mov [rsp+((1024+7-(10)) MOD (8))*4], eax
1046	mov edi, [rsp+((1024+7-(11+2)) MOD (8))*4]
1047	xor edi, [rsp+((1024+7-(11+1)) MOD (8))*4]
1048	and edi, edx
1049	xor edi, [rsp+((1024+7-(11+1)) MOD (8))*4]
1050	mov ebp, edx
1051	ror edx, 6
1052	ror ebp, 25
1053	add edi, [rsi+(11)*4]
1054	add edi, [rsp+84+((1024+15-(11)) MOD (16))4]
1055	add edi, [rsp+((1024+7-(11)) MOD (8))*4]
1056	xor ebp, edx
1057	ror edx, 5
1058	xor ebp, edx
1059	add edi, ebp
1060	mov ebx, eax
1061	xor eax, [rsp+((1024+7-(11+6)) MOD (8))*4]
1062	and ecx, eax
1063	xor ecx, [rsp+((1024+7-(11+6)) MOD (8))*4]
1064	mov ebp, ebx
1065	ror ebx, 2
1066	add ecx, edi
1067	add edi, [rsp+((1024+7-(11+4)) MOD (8))*4]
1068	mov [rsp+((1024+7-(11+4)) MOD (8))*4], edi
1069	ror ebp, 22
1070	xor ebp, ebx
1071	ror ebx, 11
1072	xor ebp, ebx
1073	add ecx, ebp
1074	mov [rsp+((1024+7-(11)) MOD (8))*4], ecx
1075	mov edx, [rsp+((1024+7-(12+2)) MOD (8))*4]
1076	xor edx, [rsp+((1024+7-(12+1)) MOD (8))*4]
1077	and edx, edi
1078	xor edx, [rsp+((1024+7-(12+1)) MOD (8))*4]
1079	mov ebp, edi
1080	ror edi, 6
1081	ror ebp, 25
1082	add edx, [rsi+(12)*4]
1083	add edx, [rsp+84+((1024+15-(12)) MOD (16))4]
1084	add edx, [rsp+((1024+7-(12)) MOD (8))*4]
1085	xor ebp, edi
1086	ror edi, 5
1087	xor ebp, edi
1088	add edx, ebp
1089	mov ebx, ecx
1090	xor ecx, [rsp+((1024+7-(12+6)) MOD (8))*4]
1091	and eax, ecx
1092	xor eax, [rsp+((1024+7-(12+6)) MOD (8))*4]
1093	mov ebp, ebx
1094	ror ebx, 2
1095	add eax, edx
1096	add edx, [rsp+((1024+7-(12+4)) MOD (8))*4]
1097	mov [rsp+((1024+7-(12+4)) MOD (8))*4], edx
1098	ror ebp, 22
1099	xor ebp, ebx
1100	ror ebx, 11
1101	xor ebp, ebx
1102	add eax, ebp
1103	mov [rsp+((1024+7-(12)) MOD (8))*4], eax
1104	mov edi, [rsp+((1024+7-(13+2)) MOD (8))*4]
1105	xor edi, [rsp+((1024+7-(13+1)) MOD (8))*4]
1106	and edi, edx
1107	xor edi, [rsp+((1024+7-(13+1)) MOD (8))*4]
1108	mov ebp, edx
1109	ror edx, 6
1110	ror ebp, 25
1111	add edi, [rsi+(13)*4]
1112	add edi, [rsp+84+((1024+15-(13)) MOD (16))4]
1113	add edi, [rsp+((1024+7-(13)) MOD (8))*4]
1114	xor ebp, edx
1115	ror edx, 5
1116	xor ebp, edx
1117	add edi, ebp
1118	mov ebx, eax
1119	xor eax, [rsp+((1024+7-(13+6)) MOD (8))*4]
1120	and ecx, eax
1121	xor ecx, [rsp+((1024+7-(13+6)) MOD (8))*4]
1122	mov ebp, ebx
1123	ror ebx, 2
1124	add ecx, edi
1125	add edi, [rsp+((1024+7-(13+4)) MOD (8))*4]
1126	mov [rsp+((1024+7-(13+4)) MOD (8))*4], edi
1127	ror ebp, 22
1128	xor ebp, ebx
1129	ror ebx, 11
1130	xor ebp, ebx
1131	add ecx, ebp
1132	mov [rsp+((1024+7-(13)) MOD (8))*4], ecx
1133	mov edx, [rsp+((1024+7-(14+2)) MOD (8))*4]
1134	xor edx, [rsp+((1024+7-(14+1)) MOD (8))*4]
1135	and edx, edi
1136	xor edx, [rsp+((1024+7-(14+1)) MOD (8))*4]
1137	mov ebp, edi
1138	ror edi, 6
1139	ror ebp, 25
1140	add edx, [rsi+(14)*4]
1141	add edx, [rsp+84+((1024+15-(14)) MOD (16))4]
1142	add edx, [rsp+((1024+7-(14)) MOD (8))*4]
1143	xor ebp, edi
1144	ror edi, 5
1145	xor ebp, edi
1146	add edx, ebp
1147	mov ebx, ecx
1148	xor ecx, [rsp+((1024+7-(14+6)) MOD (8))*4]
1149	and eax, ecx
1150	xor eax, [rsp+((1024+7-(14+6)) MOD (8))*4]
1151	mov ebp, ebx
1152	ror ebx, 2
1153	add eax, edx
1154	add edx, [rsp+((1024+7-(14+4)) MOD (8))*4]
1155	mov [rsp+((1024+7-(14+4)) MOD (8))*4], edx
1156	ror ebp, 22
1157	xor ebp, ebx
1158	ror ebx, 11
1159	xor ebp, ebx
1160	add eax, ebp
1161	mov [rsp+((1024+7-(14)) MOD (8))*4], eax
1162	mov edi, [rsp+((1024+7-(15+2)) MOD (8))*4]
1163	xor edi, [rsp+((1024+7-(15+1)) MOD (8))*4]
1164	and edi, edx
1165	xor edi, [rsp+((1024+7-(15+1)) MOD (8))*4]
1166	mov ebp, edx
1167	ror edx, 6
1168	ror ebp, 25
1169	add edi, [rsi+(15)*4]
1170	add edi, [rsp+84+((1024+15-(15)) MOD (16))4]
1171	add edi, [rsp+((1024+7-(15)) MOD (8))*4]
1172	xor ebp, edx
1173	ror edx, 5
1174	xor ebp, edx
1175	add edi, ebp
1176	mov ebx, eax
1177	xor eax, [rsp+((1024+7-(15+6)) MOD (8))*4]
1178	and ecx, eax
1179	xor ecx, [rsp+((1024+7-(15+6)) MOD (8))*4]
1180	mov ebp, ebx
1181	ror ebx, 2
1182	add ecx, edi
1183	add edi, [rsp+((1024+7-(15+4)) MOD (8))*4]
1184	mov [rsp+((1024+7-(15+4)) MOD (8))*4], edi
1185	ror ebp, 22
1186	xor ebp, ebx
1187	ror ebx, 11
1188	xor ebp, ebx
1189	add ecx, ebp
1190	mov [rsp+((1024+7-(15)) MOD (8))*4], ecx
1191	label1:
1192	add rsi, 4*16
1193	mov edx, [rsp+((1024+7-(0+2)) MOD (8))*4]
1194	xor edx, [rsp+((1024+7-(0+1)) MOD (8))*4]
1195	and edx, edi
1196	xor edx, [rsp+((1024+7-(0+1)) MOD (8))*4]
1197	mov ebp, edi
1198	ror edi, 6
1199	ror ebp, 25
1200	xor ebp, edi
1201	ror edi, 5
1202	xor ebp, edi
1203	add edx, ebp
1204	mov ebp, [rsp+84+((1024+15-((0)-2)) MOD (16))4]
1205	mov edi, [rsp+84+((1024+15-((0)-15)) MOD (16))4]
1206	mov ebx, ebp
1207	shr ebp, 10
1208	ror ebx, 17
1209	xor ebp, ebx
1210	ror ebx, 2
1211	xor ebx, ebp
1212	add ebx, [rsp+84+((1024+15-((0)-7)) MOD (16))4]
1213	mov ebp, edi
1214	shr ebp, 3
1215	ror edi, 7
1216	add ebx, [rsp+84+((1024+15-(0)) MOD (16))4]
1217	xor ebp, edi
1218	add edx, [rsi+(0)*4]
1219	ror edi, 11
1220	add edx, [rsp+((1024+7-(0)) MOD (8))*4]
1221	xor ebp, edi
1222	add ebp, ebx
1223	mov [rsp+84+((1024+15-(0)) MOD (16))4], ebp
1224	add edx, ebp
1225	mov ebx, ecx
1226	xor ecx, [rsp+((1024+7-(0+6)) MOD (8))*4]
1227	and eax, ecx
1228	xor eax, [rsp+((1024+7-(0+6)) MOD (8))*4]
1229	mov ebp, ebx
1230	ror ebx, 2
1231	add eax, edx
1232	add edx, [rsp+((1024+7-(0+4)) MOD (8))*4]
1233	mov [rsp+((1024+7-(0+4)) MOD (8))*4], edx
1234	ror ebp, 22
1235	xor ebp, ebx
1236	ror ebx, 11
1237	xor ebp, ebx
1238	add eax, ebp
1239	mov [rsp+((1024+7-(0)) MOD (8))*4], eax
1240	mov edi, [rsp+((1024+7-(1+2)) MOD (8))*4]
1241	xor edi, [rsp+((1024+7-(1+1)) MOD (8))*4]
1242	and edi, edx
1243	xor edi, [rsp+((1024+7-(1+1)) MOD (8))*4]
1244	mov ebp, edx
1245	ror edx, 6
1246	ror ebp, 25
1247	xor ebp, edx
1248	ror edx, 5
1249	xor ebp, edx
1250	add edi, ebp
1251	mov ebp, [rsp+84+((1024+15-((1)-2)) MOD (16))4]
1252	mov edx, [rsp+84+((1024+15-((1)-15)) MOD (16))4]
1253	mov ebx, ebp
1254	shr ebp, 10
1255	ror ebx, 17
1256	xor ebp, ebx
1257	ror ebx, 2
1258	xor ebx, ebp
1259	add ebx, [rsp+84+((1024+15-((1)-7)) MOD (16))4]
1260	mov ebp, edx
1261	shr ebp, 3
1262	ror edx, 7
1263	add ebx, [rsp+84+((1024+15-(1)) MOD (16))4]
1264	xor ebp, edx
1265	add edi, [rsi+(1)*4]
1266	ror edx, 11
1267	add edi, [rsp+((1024+7-(1)) MOD (8))*4]
1268	xor ebp, edx
1269	add ebp, ebx
1270	mov [rsp+84+((1024+15-(1)) MOD (16))4], ebp
1271	add edi, ebp
1272	mov ebx, eax
1273	xor eax, [rsp+((1024+7-(1+6)) MOD (8))*4]
1274	and ecx, eax
1275	xor ecx, [rsp+((1024+7-(1+6)) MOD (8))*4]
1276	mov ebp, ebx
1277	ror ebx, 2
1278	add ecx, edi
1279	add edi, [rsp+((1024+7-(1+4)) MOD (8))*4]
1280	mov [rsp+((1024+7-(1+4)) MOD (8))*4], edi
1281	ror ebp, 22
1282	xor ebp, ebx
1283	ror ebx, 11
1284	xor ebp, ebx
1285	add ecx, ebp
1286	mov [rsp+((1024+7-(1)) MOD (8))*4], ecx
1287	mov edx, [rsp+((1024+7-(2+2)) MOD (8))*4]
1288	xor edx, [rsp+((1024+7-(2+1)) MOD (8))*4]
1289	and edx, edi
1290	xor edx, [rsp+((1024+7-(2+1)) MOD (8))*4]
1291	mov ebp, edi
1292	ror edi, 6
1293	ror ebp, 25
1294	xor ebp, edi
1295	ror edi, 5
1296	xor ebp, edi
1297	add edx, ebp
1298	mov ebp, [rsp+84+((1024+15-((2)-2)) MOD (16))4]
1299	mov edi, [rsp+84+((1024+15-((2)-15)) MOD (16))4]
1300	mov ebx, ebp
1301	shr ebp, 10
1302	ror ebx, 17
1303	xor ebp, ebx
1304	ror ebx, 2
1305	xor ebx, ebp
1306	add ebx, [rsp+84+((1024+15-((2)-7)) MOD (16))4]
1307	mov ebp, edi
1308	shr ebp, 3
1309	ror edi, 7
1310	add ebx, [rsp+84+((1024+15-(2)) MOD (16))4]
1311	xor ebp, edi
1312	add edx, [rsi+(2)*4]
1313	ror edi, 11
1314	add edx, [rsp+((1024+7-(2)) MOD (8))*4]
1315	xor ebp, edi
1316	add ebp, ebx
1317	mov [rsp+84+((1024+15-(2)) MOD (16))4], ebp
1318	add edx, ebp
1319	mov ebx, ecx
1320	xor ecx, [rsp+((1024+7-(2+6)) MOD (8))*4]
1321	and eax, ecx
1322	xor eax, [rsp+((1024+7-(2+6)) MOD (8))*4]
1323	mov ebp, ebx
1324	ror ebx, 2
1325	add eax, edx
1326	add edx, [rsp+((1024+7-(2+4)) MOD (8))*4]
1327	mov [rsp+((1024+7-(2+4)) MOD (8))*4], edx
1328	ror ebp, 22
1329	xor ebp, ebx
1330	ror ebx, 11
1331	xor ebp, ebx
1332	add eax, ebp
1333	mov [rsp+((1024+7-(2)) MOD (8))*4], eax
1334	mov edi, [rsp+((1024+7-(3+2)) MOD (8))*4]
1335	xor edi, [rsp+((1024+7-(3+1)) MOD (8))*4]
1336	and edi, edx
1337	xor edi, [rsp+((1024+7-(3+1)) MOD (8))*4]
1338	mov ebp, edx
1339	ror edx, 6
1340	ror ebp, 25
1341	xor ebp, edx
1342	ror edx, 5
1343	xor ebp, edx
1344	add edi, ebp
1345	mov ebp, [rsp+84+((1024+15-((3)-2)) MOD (16))4]
1346	mov edx, [rsp+84+((1024+15-((3)-15)) MOD (16))4]
1347	mov ebx, ebp
1348	shr ebp, 10
1349	ror ebx, 17
1350	xor ebp, ebx
1351	ror ebx, 2
1352	xor ebx, ebp
1353	add ebx, [rsp+84+((1024+15-((3)-7)) MOD (16))4]
1354	mov ebp, edx
1355	shr ebp, 3
1356	ror edx, 7
1357	add ebx, [rsp+84+((1024+15-(3)) MOD (16))4]
1358	xor ebp, edx
1359	add edi, [rsi+(3)*4]
1360	ror edx, 11
1361	add edi, [rsp+((1024+7-(3)) MOD (8))*4]
1362	xor ebp, edx
1363	add ebp, ebx
1364	mov [rsp+84+((1024+15-(3)) MOD (16))4], ebp
1365	add edi, ebp
1366	mov ebx, eax
1367	xor eax, [rsp+((1024+7-(3+6)) MOD (8))*4]
1368	and ecx, eax
1369	xor ecx, [rsp+((1024+7-(3+6)) MOD (8))*4]
1370	mov ebp, ebx
1371	ror ebx, 2
1372	add ecx, edi
1373	add edi, [rsp+((1024+7-(3+4)) MOD (8))*4]
1374	mov [rsp+((1024+7-(3+4)) MOD (8))*4], edi
1375	ror ebp, 22
1376	xor ebp, ebx
1377	ror ebx, 11
1378	xor ebp, ebx
1379	add ecx, ebp
1380	mov [rsp+((1024+7-(3)) MOD (8))*4], ecx
1381	mov edx, [rsp+((1024+7-(4+2)) MOD (8))*4]
1382	xor edx, [rsp+((1024+7-(4+1)) MOD (8))*4]
1383	and edx, edi
1384	xor edx, [rsp+((1024+7-(4+1)) MOD (8))*4]
1385	mov ebp, edi
1386	ror edi, 6
1387	ror ebp, 25
1388	xor ebp, edi
1389	ror edi, 5
1390	xor ebp, edi
1391	add edx, ebp
1392	mov ebp, [rsp+84+((1024+15-((4)-2)) MOD (16))4]
1393	mov edi, [rsp+84+((1024+15-((4)-15)) MOD (16))4]
1394	mov ebx, ebp
1395	shr ebp, 10
1396	ror ebx, 17
1397	xor ebp, ebx
1398	ror ebx, 2
1399	xor ebx, ebp
1400	add ebx, [rsp+84+((1024+15-((4)-7)) MOD (16))4]
1401	mov ebp, edi
1402	shr ebp, 3
1403	ror edi, 7
1404	add ebx, [rsp+84+((1024+15-(4)) MOD (16))4]
1405	xor ebp, edi
1406	add edx, [rsi+(4)*4]
1407	ror edi, 11
1408	add edx, [rsp+((1024+7-(4)) MOD (8))*4]
1409	xor ebp, edi
1410	add ebp, ebx
1411	mov [rsp+84+((1024+15-(4)) MOD (16))4], ebp
1412	add edx, ebp
1413	mov ebx, ecx
1414	xor ecx, [rsp+((1024+7-(4+6)) MOD (8))*4]
1415	and eax, ecx
1416	xor eax, [rsp+((1024+7-(4+6)) MOD (8))*4]
1417	mov ebp, ebx
1418	ror ebx, 2
1419	add eax, edx
1420	add edx, [rsp+((1024+7-(4+4)) MOD (8))*4]
1421	mov [rsp+((1024+7-(4+4)) MOD (8))*4], edx
1422	ror ebp, 22
1423	xor ebp, ebx
1424	ror ebx, 11
1425	xor ebp, ebx
1426	add eax, ebp
1427	mov [rsp+((1024+7-(4)) MOD (8))*4], eax
1428	mov edi, [rsp+((1024+7-(5+2)) MOD (8))*4]
1429	xor edi, [rsp+((1024+7-(5+1)) MOD (8))*4]
1430	and edi, edx
1431	xor edi, [rsp+((1024+7-(5+1)) MOD (8))*4]
1432	mov ebp, edx
1433	ror edx, 6
1434	ror ebp, 25
1435	xor ebp, edx
1436	ror edx, 5
1437	xor ebp, edx
1438	add edi, ebp
1439	mov ebp, [rsp+84+((1024+15-((5)-2)) MOD (16))4]
1440	mov edx, [rsp+84+((1024+15-((5)-15)) MOD (16))4]
1441	mov ebx, ebp
1442	shr ebp, 10
1443	ror ebx, 17
1444	xor ebp, ebx
1445	ror ebx, 2
1446	xor ebx, ebp
1447	add ebx, [rsp+84+((1024+15-((5)-7)) MOD (16))4]
1448	mov ebp, edx
1449	shr ebp, 3
1450	ror edx, 7
1451	add ebx, [rsp+84+((1024+15-(5)) MOD (16))4]
1452	xor ebp, edx
1453	add edi, [rsi+(5)*4]
1454	ror edx, 11
1455	add edi, [rsp+((1024+7-(5)) MOD (8))*4]
1456	xor ebp, edx
1457	add ebp, ebx
1458	mov [rsp+84+((1024+15-(5)) MOD (16))4], ebp
1459	add edi, ebp
1460	mov ebx, eax
1461	xor eax, [rsp+((1024+7-(5+6)) MOD (8))*4]
1462	and ecx, eax
1463	xor ecx, [rsp+((1024+7-(5+6)) MOD (8))*4]
1464	mov ebp, ebx
1465	ror ebx, 2
1466	add ecx, edi
1467	add edi, [rsp+((1024+7-(5+4)) MOD (8))*4]
1468	mov [rsp+((1024+7-(5+4)) MOD (8))*4], edi
1469	ror ebp, 22
1470	xor ebp, ebx
1471	ror ebx, 11
1472	xor ebp, ebx
1473	add ecx, ebp
1474	mov [rsp+((1024+7-(5)) MOD (8))*4], ecx
1475	mov edx, [rsp+((1024+7-(6+2)) MOD (8))*4]
1476	xor edx, [rsp+((1024+7-(6+1)) MOD (8))*4]
1477	and edx, edi
1478	xor edx, [rsp+((1024+7-(6+1)) MOD (8))*4]
1479	mov ebp, edi
1480	ror edi, 6
1481	ror ebp, 25
1482	xor ebp, edi
1483	ror edi, 5
1484	xor ebp, edi
1485	add edx, ebp
1486	mov ebp, [rsp+84+((1024+15-((6)-2)) MOD (16))4]
1487	mov edi, [rsp+84+((1024+15-((6)-15)) MOD (16))4]
1488	mov ebx, ebp
1489	shr ebp, 10
1490	ror ebx, 17
1491	xor ebp, ebx
1492	ror ebx, 2
1493	xor ebx, ebp
1494	add ebx, [rsp+84+((1024+15-((6)-7)) MOD (16))4]
1495	mov ebp, edi
1496	shr ebp, 3
1497	ror edi, 7
1498	add ebx, [rsp+84+((1024+15-(6)) MOD (16))4]
1499	xor ebp, edi
1500	add edx, [rsi+(6)*4]
1501	ror edi, 11
1502	add edx, [rsp+((1024+7-(6)) MOD (8))*4]
1503	xor ebp, edi
1504	add ebp, ebx
1505	mov [rsp+84+((1024+15-(6)) MOD (16))4], ebp
1506	add edx, ebp
1507	mov ebx, ecx
1508	xor ecx, [rsp+((1024+7-(6+6)) MOD (8))*4]
1509	and eax, ecx
1510	xor eax, [rsp+((1024+7-(6+6)) MOD (8))*4]
1511	mov ebp, ebx
1512	ror ebx, 2
1513	add eax, edx
1514	add edx, [rsp+((1024+7-(6+4)) MOD (8))*4]
1515	mov [rsp+((1024+7-(6+4)) MOD (8))*4], edx
1516	ror ebp, 22
1517	xor ebp, ebx
1518	ror ebx, 11
1519	xor ebp, ebx
1520	add eax, ebp
1521	mov [rsp+((1024+7-(6)) MOD (8))*4], eax
1522	mov edi, [rsp+((1024+7-(7+2)) MOD (8))*4]
1523	xor edi, [rsp+((1024+7-(7+1)) MOD (8))*4]
1524	and edi, edx
1525	xor edi, [rsp+((1024+7-(7+1)) MOD (8))*4]
1526	mov ebp, edx
1527	ror edx, 6
1528	ror ebp, 25
1529	xor ebp, edx
1530	ror edx, 5
1531	xor ebp, edx
1532	add edi, ebp
1533	mov ebp, [rsp+84+((1024+15-((7)-2)) MOD (16))4]
1534	mov edx, [rsp+84+((1024+15-((7)-15)) MOD (16))4]
1535	mov ebx, ebp
1536	shr ebp, 10
1537	ror ebx, 17
1538	xor ebp, ebx
1539	ror ebx, 2
1540	xor ebx, ebp
1541	add ebx, [rsp+84+((1024+15-((7)-7)) MOD (16))4]
1542	mov ebp, edx
1543	shr ebp, 3
1544	ror edx, 7
1545	add ebx, [rsp+84+((1024+15-(7)) MOD (16))4]
1546	xor ebp, edx
1547	add edi, [rsi+(7)*4]
1548	ror edx, 11
1549	add edi, [rsp+((1024+7-(7)) MOD (8))*4]
1550	xor ebp, edx
1551	add ebp, ebx
1552	mov [rsp+84+((1024+15-(7)) MOD (16))4], ebp
1553	add edi, ebp
1554	mov ebx, eax
1555	xor eax, [rsp+((1024+7-(7+6)) MOD (8))*4]
1556	and ecx, eax
1557	xor ecx, [rsp+((1024+7-(7+6)) MOD (8))*4]
1558	mov ebp, ebx
1559	ror ebx, 2
1560	add ecx, edi
1561	add edi, [rsp+((1024+7-(7+4)) MOD (8))*4]
1562	mov [rsp+((1024+7-(7+4)) MOD (8))*4], edi
1563	ror ebp, 22
1564	xor ebp, ebx
1565	ror ebx, 11
1566	xor ebp, ebx
1567	add ecx, ebp
1568	mov [rsp+((1024+7-(7)) MOD (8))*4], ecx
1569	mov edx, [rsp+((1024+7-(8+2)) MOD (8))*4]
1570	xor edx, [rsp+((1024+7-(8+1)) MOD (8))*4]
1571	and edx, edi
1572	xor edx, [rsp+((1024+7-(8+1)) MOD (8))*4]
1573	mov ebp, edi
1574	ror edi, 6
1575	ror ebp, 25
1576	xor ebp, edi
1577	ror edi, 5
1578	xor ebp, edi
1579	add edx, ebp
1580	mov ebp, [rsp+84+((1024+15-((8)-2)) MOD (16))4]
1581	mov edi, [rsp+84+((1024+15-((8)-15)) MOD (16))4]
1582	mov ebx, ebp
1583	shr ebp, 10
1584	ror ebx, 17
1585	xor ebp, ebx
1586	ror ebx, 2
1587	xor ebx, ebp
1588	add ebx, [rsp+84+((1024+15-((8)-7)) MOD (16))4]
1589	mov ebp, edi
1590	shr ebp, 3
1591	ror edi, 7
1592	add ebx, [rsp+84+((1024+15-(8)) MOD (16))4]
1593	xor ebp, edi
1594	add edx, [rsi+(8)*4]
1595	ror edi, 11
1596	add edx, [rsp+((1024+7-(8)) MOD (8))*4]
1597	xor ebp, edi
1598	add ebp, ebx
1599	mov [rsp+84+((1024+15-(8)) MOD (16))4], ebp
1600	add edx, ebp
1601	mov ebx, ecx
1602	xor ecx, [rsp+((1024+7-(8+6)) MOD (8))*4]
1603	and eax, ecx
1604	xor eax, [rsp+((1024+7-(8+6)) MOD (8))*4]
1605	mov ebp, ebx
1606	ror ebx, 2
1607	add eax, edx
1608	add edx, [rsp+((1024+7-(8+4)) MOD (8))*4]
1609	mov [rsp+((1024+7-(8+4)) MOD (8))*4], edx
1610	ror ebp, 22
1611	xor ebp, ebx
1612	ror ebx, 11
1613	xor ebp, ebx
1614	add eax, ebp
1615	mov [rsp+((1024+7-(8)) MOD (8))*4], eax
1616	mov edi, [rsp+((1024+7-(9+2)) MOD (8))*4]
1617	xor edi, [rsp+((1024+7-(9+1)) MOD (8))*4]
1618	and edi, edx
1619	xor edi, [rsp+((1024+7-(9+1)) MOD (8))*4]
1620	mov ebp, edx
1621	ror edx, 6
1622	ror ebp, 25
1623	xor ebp, edx
1624	ror edx, 5
1625	xor ebp, edx
1626	add edi, ebp
1627	mov ebp, [rsp+84+((1024+15-((9)-2)) MOD (16))4]
1628	mov edx, [rsp+84+((1024+15-((9)-15)) MOD (16))4]
1629	mov ebx, ebp
1630	shr ebp, 10
1631	ror ebx, 17
1632	xor ebp, ebx
1633	ror ebx, 2
1634	xor ebx, ebp
1635	add ebx, [rsp+84+((1024+15-((9)-7)) MOD (16))4]
1636	mov ebp, edx
1637	shr ebp, 3
1638	ror edx, 7
1639	add ebx, [rsp+84+((1024+15-(9)) MOD (16))4]
1640	xor ebp, edx
1641	add edi, [rsi+(9)*4]
1642	ror edx, 11
1643	add edi, [rsp+((1024+7-(9)) MOD (8))*4]
1644	xor ebp, edx
1645	add ebp, ebx
1646	mov [rsp+84+((1024+15-(9)) MOD (16))4], ebp
1647	add edi, ebp
1648	mov ebx, eax
1649	xor eax, [rsp+((1024+7-(9+6)) MOD (8))*4]
1650	and ecx, eax
1651	xor ecx, [rsp+((1024+7-(9+6)) MOD (8))*4]
1652	mov ebp, ebx
1653	ror ebx, 2
1654	add ecx, edi
1655	add edi, [rsp+((1024+7-(9+4)) MOD (8))*4]
1656	mov [rsp+((1024+7-(9+4)) MOD (8))*4], edi
1657	ror ebp, 22
1658	xor ebp, ebx
1659	ror ebx, 11
1660	xor ebp, ebx
1661	add ecx, ebp
1662	mov [rsp+((1024+7-(9)) MOD (8))*4], ecx
1663	mov edx, [rsp+((1024+7-(10+2)) MOD (8))*4]
1664	xor edx, [rsp+((1024+7-(10+1)) MOD (8))*4]
1665	and edx, edi
1666	xor edx, [rsp+((1024+7-(10+1)) MOD (8))*4]
1667	mov ebp, edi
1668	ror edi, 6
1669	ror ebp, 25
1670	xor ebp, edi
1671	ror edi, 5
1672	xor ebp, edi
1673	add edx, ebp
1674	mov ebp, [rsp+84+((1024+15-((10)-2)) MOD (16))4]
1675	mov edi, [rsp+84+((1024+15-((10)-15)) MOD (16))4]
1676	mov ebx, ebp
1677	shr ebp, 10
1678	ror ebx, 17
1679	xor ebp, ebx
1680	ror ebx, 2
1681	xor ebx, ebp
1682	add ebx, [rsp+84+((1024+15-((10)-7)) MOD (16))4]
1683	mov ebp, edi
1684	shr ebp, 3
1685	ror edi, 7
1686	add ebx, [rsp+84+((1024+15-(10)) MOD (16))4]
1687	xor ebp, edi
1688	add edx, [rsi+(10)*4]
1689	ror edi, 11
1690	add edx, [rsp+((1024+7-(10)) MOD (8))*4]
1691	xor ebp, edi
1692	add ebp, ebx
1693	mov [rsp+84+((1024+15-(10)) MOD (16))4], ebp
1694	add edx, ebp
1695	mov ebx, ecx
1696	xor ecx, [rsp+((1024+7-(10+6)) MOD (8))*4]
1697	and eax, ecx
1698	xor eax, [rsp+((1024+7-(10+6)) MOD (8))*4]
1699	mov ebp, ebx
1700	ror ebx, 2
1701	add eax, edx
1702	add edx, [rsp+((1024+7-(10+4)) MOD (8))*4]
1703	mov [rsp+((1024+7-(10+4)) MOD (8))*4], edx
1704	ror ebp, 22
1705	xor ebp, ebx
1706	ror ebx, 11
1707	xor ebp, ebx
1708	add eax, ebp
1709	mov [rsp+((1024+7-(10)) MOD (8))*4], eax
1710	mov edi, [rsp+((1024+7-(11+2)) MOD (8))*4]
1711	xor edi, [rsp+((1024+7-(11+1)) MOD (8))*4]
1712	and edi, edx
1713	xor edi, [rsp+((1024+7-(11+1)) MOD (8))*4]
1714	mov ebp, edx
1715	ror edx, 6
1716	ror ebp, 25
1717	xor ebp, edx
1718	ror edx, 5
1719	xor ebp, edx
1720	add edi, ebp
1721	mov ebp, [rsp+84+((1024+15-((11)-2)) MOD (16))4]
1722	mov edx, [rsp+84+((1024+15-((11)-15)) MOD (16))4]
1723	mov ebx, ebp
1724	shr ebp, 10
1725	ror ebx, 17
1726	xor ebp, ebx
1727	ror ebx, 2
1728	xor ebx, ebp
1729	add ebx, [rsp+84+((1024+15-((11)-7)) MOD (16))4]
1730	mov ebp, edx
1731	shr ebp, 3
1732	ror edx, 7
1733	add ebx, [rsp+84+((1024+15-(11)) MOD (16))4]
1734	xor ebp, edx
1735	add edi, [rsi+(11)*4]
1736	ror edx, 11
1737	add edi, [rsp+((1024+7-(11)) MOD (8))*4]
1738	xor ebp, edx
1739	add ebp, ebx
1740	mov [rsp+84+((1024+15-(11)) MOD (16))4], ebp
1741	add edi, ebp
1742	mov ebx, eax
1743	xor eax, [rsp+((1024+7-(11+6)) MOD (8))*4]
1744	and ecx, eax
1745	xor ecx, [rsp+((1024+7-(11+6)) MOD (8))*4]
1746	mov ebp, ebx
1747	ror ebx, 2
1748	add ecx, edi
1749	add edi, [rsp+((1024+7-(11+4)) MOD (8))*4]
1750	mov [rsp+((1024+7-(11+4)) MOD (8))*4], edi
1751	ror ebp, 22
1752	xor ebp, ebx
1753	ror ebx, 11
1754	xor ebp, ebx
1755	add ecx, ebp
1756	mov [rsp+((1024+7-(11)) MOD (8))*4], ecx
1757	mov edx, [rsp+((1024+7-(12+2)) MOD (8))*4]
1758	xor edx, [rsp+((1024+7-(12+1)) MOD (8))*4]
1759	and edx, edi
1760	xor edx, [rsp+((1024+7-(12+1)) MOD (8))*4]
1761	mov ebp, edi
1762	ror edi, 6
1763	ror ebp, 25
1764	xor ebp, edi
1765	ror edi, 5
1766	xor ebp, edi
1767	add edx, ebp
1768	mov ebp, [rsp+84+((1024+15-((12)-2)) MOD (16))4]
1769	mov edi, [rsp+84+((1024+15-((12)-15)) MOD (16))4]
1770	mov ebx, ebp
1771	shr ebp, 10
1772	ror ebx, 17
1773	xor ebp, ebx
1774	ror ebx, 2
1775	xor ebx, ebp
1776	add ebx, [rsp+84+((1024+15-((12)-7)) MOD (16))4]
1777	mov ebp, edi
1778	shr ebp, 3
1779	ror edi, 7
1780	add ebx, [rsp+84+((1024+15-(12)) MOD (16))4]
1781	xor ebp, edi
1782	add edx, [rsi+(12)*4]
1783	ror edi, 11
1784	add edx, [rsp+((1024+7-(12)) MOD (8))*4]
1785	xor ebp, edi
1786	add ebp, ebx
1787	mov [rsp+84+((1024+15-(12)) MOD (16))4], ebp
1788	add edx, ebp
1789	mov ebx, ecx
1790	xor ecx, [rsp+((1024+7-(12+6)) MOD (8))*4]
1791	and eax, ecx
1792	xor eax, [rsp+((1024+7-(12+6)) MOD (8))*4]
1793	mov ebp, ebx
1794	ror ebx, 2
1795	add eax, edx
1796	add edx, [rsp+((1024+7-(12+4)) MOD (8))*4]
1797	mov [rsp+((1024+7-(12+4)) MOD (8))*4], edx
1798	ror ebp, 22
1799	xor ebp, ebx
1800	ror ebx, 11
1801	xor ebp, ebx
1802	add eax, ebp
1803	mov [rsp+((1024+7-(12)) MOD (8))*4], eax
1804	mov edi, [rsp+((1024+7-(13+2)) MOD (8))*4]
1805	xor edi, [rsp+((1024+7-(13+1)) MOD (8))*4]
1806	and edi, edx
1807	xor edi, [rsp+((1024+7-(13+1)) MOD (8))*4]
1808	mov ebp, edx
1809	ror edx, 6
1810	ror ebp, 25
1811	xor ebp, edx
1812	ror edx, 5
1813	xor ebp, edx
1814	add edi, ebp
1815	mov ebp, [rsp+84+((1024+15-((13)-2)) MOD (16))4]
1816	mov edx, [rsp+84+((1024+15-((13)-15)) MOD (16))4]
1817	mov ebx, ebp
1818	shr ebp, 10
1819	ror ebx, 17
1820	xor ebp, ebx
1821	ror ebx, 2
1822	xor ebx, ebp
1823	add ebx, [rsp+84+((1024+15-((13)-7)) MOD (16))4]
1824	mov ebp, edx
1825	shr ebp, 3
1826	ror edx, 7
1827	add ebx, [rsp+84+((1024+15-(13)) MOD (16))4]
1828	xor ebp, edx
1829	add edi, [rsi+(13)*4]
1830	ror edx, 11
1831	add edi, [rsp+((1024+7-(13)) MOD (8))*4]
1832	xor ebp, edx
1833	add ebp, ebx
1834	mov [rsp+84+((1024+15-(13)) MOD (16))4], ebp
1835	add edi, ebp
1836	mov ebx, eax
1837	xor eax, [rsp+((1024+7-(13+6)) MOD (8))*4]
1838	and ecx, eax
1839	xor ecx, [rsp+((1024+7-(13+6)) MOD (8))*4]
1840	mov ebp, ebx
1841	ror ebx, 2
1842	add ecx, edi
1843	add edi, [rsp+((1024+7-(13+4)) MOD (8))*4]
1844	mov [rsp+((1024+7-(13+4)) MOD (8))*4], edi
1845	ror ebp, 22
1846	xor ebp, ebx
1847	ror ebx, 11
1848	xor ebp, ebx
1849	add ecx, ebp
1850	mov [rsp+((1024+7-(13)) MOD (8))*4], ecx
1851	mov edx, [rsp+((1024+7-(14+2)) MOD (8))*4]
1852	xor edx, [rsp+((1024+7-(14+1)) MOD (8))*4]
1853	and edx, edi
1854	xor edx, [rsp+((1024+7-(14+1)) MOD (8))*4]
1855	mov ebp, edi
1856	ror edi, 6
1857	ror ebp, 25
1858	xor ebp, edi
1859	ror edi, 5
1860	xor ebp, edi
1861	add edx, ebp
1862	mov ebp, [rsp+84+((1024+15-((14)-2)) MOD (16))4]
1863	mov edi, [rsp+84+((1024+15-((14)-15)) MOD (16))4]
1864	mov ebx, ebp
1865	shr ebp, 10
1866	ror ebx, 17
1867	xor ebp, ebx
1868	ror ebx, 2
1869	xor ebx, ebp
1870	add ebx, [rsp+84+((1024+15-((14)-7)) MOD (16))4]
1871	mov ebp, edi
1872	shr ebp, 3
1873	ror edi, 7
1874	add ebx, [rsp+84+((1024+15-(14)) MOD (16))4]
1875	xor ebp, edi
1876	add edx, [rsi+(14)*4]
1877	ror edi, 11
1878	add edx, [rsp+((1024+7-(14)) MOD (8))*4]
1879	xor ebp, edi
1880	add ebp, ebx
1881	mov [rsp+84+((1024+15-(14)) MOD (16))4], ebp
1882	add edx, ebp
1883	mov ebx, ecx
1884	xor ecx, [rsp+((1024+7-(14+6)) MOD (8))*4]
1885	and eax, ecx
1886	xor eax, [rsp+((1024+7-(14+6)) MOD (8))*4]
1887	mov ebp, ebx
1888	ror ebx, 2
1889	add eax, edx
1890	add edx, [rsp+((1024+7-(14+4)) MOD (8))*4]
1891	mov [rsp+((1024+7-(14+4)) MOD (8))*4], edx
1892	ror ebp, 22
1893	xor ebp, ebx
1894	ror ebx, 11
1895	xor ebp, ebx
1896	add eax, ebp
1897	mov [rsp+((1024+7-(14)) MOD (8))*4], eax
1898	mov edi, [rsp+((1024+7-(15+2)) MOD (8))*4]
1899	xor edi, [rsp+((1024+7-(15+1)) MOD (8))*4]
1900	and edi, edx
1901	xor edi, [rsp+((1024+7-(15+1)) MOD (8))*4]
1902	mov ebp, edx
1903	ror edx, 6
1904	ror ebp, 25
1905	xor ebp, edx
1906	ror edx, 5
1907	xor ebp, edx
1908	add edi, ebp
1909	mov ebp, [rsp+84+((1024+15-((15)-2)) MOD (16))4]
1910	mov edx, [rsp+84+((1024+15-((15)-15)) MOD (16))4]
1911	mov ebx, ebp
1912	shr ebp, 10
1913	ror ebx, 17
1914	xor ebp, ebx
1915	ror ebx, 2
1916	xor ebx, ebp
1917	add ebx, [rsp+84+((1024+15-((15)-7)) MOD (16))4]
1918	mov ebp, edx
1919	shr ebp, 3
1920	ror edx, 7
1921	add ebx, [rsp+84+((1024+15-(15)) MOD (16))4]
1922	xor ebp, edx
1923	add edi, [rsi+(15)*4]
1924	ror edx, 11
1925	add edi, [rsp+((1024+7-(15)) MOD (8))*4]
1926	xor ebp, edx
1927	add ebp, ebx
1928	mov [rsp+84+((1024+15-(15)) MOD (16))4], ebp
1929	add edi, ebp
1930	mov ebx, eax
1931	xor eax, [rsp+((1024+7-(15+6)) MOD (8))*4]
1932	and ecx, eax
1933	xor ecx, [rsp+((1024+7-(15+6)) MOD (8))*4]
1934	mov ebp, ebx
1935	ror ebx, 2
1936	add ecx, edi
1937	add edi, [rsp+((1024+7-(15+4)) MOD (8))*4]
1938	mov [rsp+((1024+7-(15+4)) MOD (8))*4], edi
1939	ror ebp, 22
1940	xor ebp, ebx
1941	ror ebx, 11
1942	xor ebp, ebx
1943	add ecx, ebp
1944	mov [rsp+((1024+7-(15)) MOD (8))*4], ecx
1945	cmp rsi, [rsp+84+164+0*8]
1946	jne label1
1947	mov rcx, [rsp+84+164+1*8]
1948	movdqa xmm1, XMMWORD PTR [rcx+1*16]
1949	movdqa xmm0, XMMWORD PTR [rcx+0*16]
1950	paddd xmm1, [rsp+((1024+7-(0+3)) MOD (8))*4]
1951	paddd xmm0, [rsp+((1024+7-(0+7)) MOD (8))*4]
1952	movdqa [rcx+1*16], xmm1
1953	movdqa [rcx+0*16], xmm0
1954	mov rdx, [rsp+84+164+2*8]
1955	add rdx, 64
1956	mov [rsp+84+164+2*8], rdx
1957	cmp rdx, [rsp+84+164+3*8]
1958	jne label0
1959	add rsp, 84 + 164 + 4*8 + 8
1960	pop rbp
1961	pop rbx
1962	pop rdi
1963	pop rsi
1964	ret
1965	X86_SHA256_HashBlocks ENDP
1966
1967	_TEXT ENDS
1968	END

Note: See TracBrowser for help on using the repository browser.

Download in other formats: