x86-mont.S@ 88461

Last change on this file since 88461 was 83531, checked in by vboxsync, 5 years ago
setting svn:sync-process=export for openssl-1.1.1f, all files except tests
File size: 6.4 KB

Line
1	%ifidn __OUTPUT_FORMAT__,obj
2	section code use32 class=code align=64
3	%elifidn __OUTPUT_FORMAT__,win32
4	section .text code align=64
5	%else
6	section .text code
7	%endif
8	global _bn_mul_mont
9	align 16
10	_bn_mul_mont:
11	L$_bn_mul_mont_begin:
12	push ebp
13	push ebx
14	push esi
15	push edi
16	xor eax,eax
17	mov edi,DWORD [40+esp]
18	cmp edi,4
19	jl NEAR L$000just_leave
20	lea esi,[20+esp]
21	lea edx,[24+esp]
22	add edi,2
23	neg edi
24	lea ebp,[edi*4+esp-32]
25	neg edi
26	mov eax,ebp
27	sub eax,edx
28	and eax,2047
29	sub ebp,eax
30	xor edx,ebp
31	and edx,2048
32	xor edx,2048
33	sub ebp,edx
34	and ebp,-64
35	mov eax,esp
36	sub eax,ebp
37	and eax,-4096
38	mov edx,esp
39	lea esp,[eax*1+ebp]
40	mov eax,DWORD [esp]
41	cmp esp,ebp
42	ja NEAR L$001page_walk
43	jmp NEAR L$002page_walk_done
44	align 16
45	L$001page_walk:
46	lea esp,[esp-4096]
47	mov eax,DWORD [esp]
48	cmp esp,ebp
49	ja NEAR L$001page_walk
50	L$002page_walk_done:
51	mov eax,DWORD [esi]
52	mov ebx,DWORD [4+esi]
53	mov ecx,DWORD [8+esi]
54	mov ebp,DWORD [12+esi]
55	mov esi,DWORD [16+esi]
56	mov esi,DWORD [esi]
57	mov DWORD [4+esp],eax
58	mov DWORD [8+esp],ebx
59	mov DWORD [12+esp],ecx
60	mov DWORD [16+esp],ebp
61	mov DWORD [20+esp],esi
62	lea ebx,[edi-3]
63	mov DWORD [24+esp],edx
64	mov esi,DWORD [8+esp]
65	lea ebp,[1+ebx]
66	mov edi,DWORD [12+esp]
67	xor ecx,ecx
68	mov edx,esi
69	and ebp,1
70	sub edx,edi
71	lea eax,[4+ebx*4+edi]
72	or ebp,edx
73	mov edi,DWORD [edi]
74	jz NEAR L$003bn_sqr_mont
75	mov DWORD [28+esp],eax
76	mov eax,DWORD [esi]
77	xor edx,edx
78	align 16
79	L$004mull:
80	mov ebp,edx
81	mul edi
82	add ebp,eax
83	lea ecx,[1+ecx]
84	adc edx,0
85	mov eax,DWORD [ecx*4+esi]
86	cmp ecx,ebx
87	mov DWORD [28+ecx*4+esp],ebp
88	jl NEAR L$004mull
89	mov ebp,edx
90	mul edi
91	mov edi,DWORD [20+esp]
92	add eax,ebp
93	mov esi,DWORD [16+esp]
94	adc edx,0
95	imul edi,DWORD [32+esp]
96	mov DWORD [32+ebx*4+esp],eax
97	xor ecx,ecx
98	mov DWORD [36+ebx*4+esp],edx
99	mov DWORD [40+ebx*4+esp],ecx
100	mov eax,DWORD [esi]
101	mul edi
102	add eax,DWORD [32+esp]
103	mov eax,DWORD [4+esi]
104	adc edx,0
105	inc ecx
106	jmp NEAR L$0052ndmadd
107	align 16
108	L$0061stmadd:
109	mov ebp,edx
110	mul edi
111	add ebp,DWORD [32+ecx*4+esp]
112	lea ecx,[1+ecx]
113	adc edx,0
114	add ebp,eax
115	mov eax,DWORD [ecx*4+esi]
116	adc edx,0
117	cmp ecx,ebx
118	mov DWORD [28+ecx*4+esp],ebp
119	jl NEAR L$0061stmadd
120	mov ebp,edx
121	mul edi
122	add eax,DWORD [32+ebx*4+esp]
123	mov edi,DWORD [20+esp]
124	adc edx,0
125	mov esi,DWORD [16+esp]
126	add ebp,eax
127	adc edx,0
128	imul edi,DWORD [32+esp]
129	xor ecx,ecx
130	add edx,DWORD [36+ebx*4+esp]
131	mov DWORD [32+ebx*4+esp],ebp
132	adc ecx,0
133	mov eax,DWORD [esi]
134	mov DWORD [36+ebx*4+esp],edx
135	mov DWORD [40+ebx*4+esp],ecx
136	mul edi
137	add eax,DWORD [32+esp]
138	mov eax,DWORD [4+esi]
139	adc edx,0
140	mov ecx,1
141	align 16
142	L$0052ndmadd:
143	mov ebp,edx
144	mul edi
145	add ebp,DWORD [32+ecx*4+esp]
146	lea ecx,[1+ecx]
147	adc edx,0
148	add ebp,eax
149	mov eax,DWORD [ecx*4+esi]
150	adc edx,0
151	cmp ecx,ebx
152	mov DWORD [24+ecx*4+esp],ebp
153	jl NEAR L$0052ndmadd
154	mov ebp,edx
155	mul edi
156	add ebp,DWORD [32+ebx*4+esp]
157	adc edx,0
158	add ebp,eax
159	adc edx,0
160	mov DWORD [28+ebx*4+esp],ebp
161	xor eax,eax
162	mov ecx,DWORD [12+esp]
163	add edx,DWORD [36+ebx*4+esp]
164	adc eax,DWORD [40+ebx*4+esp]
165	lea ecx,[4+ecx]
166	mov DWORD [32+ebx*4+esp],edx
167	cmp ecx,DWORD [28+esp]
168	mov DWORD [36+ebx*4+esp],eax
169	je NEAR L$007common_tail
170	mov edi,DWORD [ecx]
171	mov esi,DWORD [8+esp]
172	mov DWORD [12+esp],ecx
173	xor ecx,ecx
174	xor edx,edx
175	mov eax,DWORD [esi]
176	jmp NEAR L$0061stmadd
177	align 16
178	L$003bn_sqr_mont:
179	mov DWORD [esp],ebx
180	mov DWORD [12+esp],ecx
181	mov eax,edi
182	mul edi
183	mov DWORD [32+esp],eax
184	mov ebx,edx
185	shr edx,1
186	and ebx,1
187	inc ecx
188	align 16
189	L$008sqr:
190	mov eax,DWORD [ecx*4+esi]
191	mov ebp,edx
192	mul edi
193	add eax,ebp
194	lea ecx,[1+ecx]
195	adc edx,0
196	lea ebp,[eax*2+ebx]
197	shr eax,31
198	cmp ecx,DWORD [esp]
199	mov ebx,eax
200	mov DWORD [28+ecx*4+esp],ebp
201	jl NEAR L$008sqr
202	mov eax,DWORD [ecx*4+esi]
203	mov ebp,edx
204	mul edi
205	add eax,ebp
206	mov edi,DWORD [20+esp]
207	adc edx,0
208	mov esi,DWORD [16+esp]
209	lea ebp,[eax*2+ebx]
210	imul edi,DWORD [32+esp]
211	shr eax,31
212	mov DWORD [32+ecx*4+esp],ebp
213	lea ebp,[edx*2+eax]
214	mov eax,DWORD [esi]
215	shr edx,31
216	mov DWORD [36+ecx*4+esp],ebp
217	mov DWORD [40+ecx*4+esp],edx
218	mul edi
219	add eax,DWORD [32+esp]
220	mov ebx,ecx
221	adc edx,0
222	mov eax,DWORD [4+esi]
223	mov ecx,1
224	align 16
225	L$0093rdmadd:
226	mov ebp,edx
227	mul edi
228	add ebp,DWORD [32+ecx*4+esp]
229	adc edx,0
230	add ebp,eax
231	mov eax,DWORD [4+ecx*4+esi]
232	adc edx,0
233	mov DWORD [28+ecx*4+esp],ebp
234	mov ebp,edx
235	mul edi
236	add ebp,DWORD [36+ecx*4+esp]
237	lea ecx,[2+ecx]
238	adc edx,0
239	add ebp,eax
240	mov eax,DWORD [ecx*4+esi]
241	adc edx,0
242	cmp ecx,ebx
243	mov DWORD [24+ecx*4+esp],ebp
244	jl NEAR L$0093rdmadd
245	mov ebp,edx
246	mul edi
247	add ebp,DWORD [32+ebx*4+esp]
248	adc edx,0
249	add ebp,eax
250	adc edx,0
251	mov DWORD [28+ebx*4+esp],ebp
252	mov ecx,DWORD [12+esp]
253	xor eax,eax
254	mov esi,DWORD [8+esp]
255	add edx,DWORD [36+ebx*4+esp]
256	adc eax,DWORD [40+ebx*4+esp]
257	mov DWORD [32+ebx*4+esp],edx
258	cmp ecx,ebx
259	mov DWORD [36+ebx*4+esp],eax
260	je NEAR L$007common_tail
261	mov edi,DWORD [4+ecx*4+esi]
262	lea ecx,[1+ecx]
263	mov eax,edi
264	mov DWORD [12+esp],ecx
265	mul edi
266	add eax,DWORD [32+ecx*4+esp]
267	adc edx,0
268	mov DWORD [32+ecx*4+esp],eax
269	xor ebp,ebp
270	cmp ecx,ebx
271	lea ecx,[1+ecx]
272	je NEAR L$010sqrlast
273	mov ebx,edx
274	shr edx,1
275	and ebx,1
276	align 16
277	L$011sqradd:
278	mov eax,DWORD [ecx*4+esi]
279	mov ebp,edx
280	mul edi
281	add eax,ebp
282	lea ebp,[eax*1+eax]
283	adc edx,0
284	shr eax,31
285	add ebp,DWORD [32+ecx*4+esp]
286	lea ecx,[1+ecx]
287	adc eax,0
288	add ebp,ebx
289	adc eax,0
290	cmp ecx,DWORD [esp]
291	mov DWORD [28+ecx*4+esp],ebp
292	mov ebx,eax
293	jle NEAR L$011sqradd
294	mov ebp,edx
295	add edx,edx
296	shr ebp,31
297	add edx,ebx
298	adc ebp,0
299	L$010sqrlast:
300	mov edi,DWORD [20+esp]
301	mov esi,DWORD [16+esp]
302	imul edi,DWORD [32+esp]
303	add edx,DWORD [32+ecx*4+esp]
304	mov eax,DWORD [esi]
305	adc ebp,0
306	mov DWORD [32+ecx*4+esp],edx
307	mov DWORD [36+ecx*4+esp],ebp
308	mul edi
309	add eax,DWORD [32+esp]
310	lea ebx,[ecx-1]
311	adc edx,0
312	mov ecx,1
313	mov eax,DWORD [4+esi]
314	jmp NEAR L$0093rdmadd
315	align 16
316	L$007common_tail:
317	mov ebp,DWORD [16+esp]
318	mov edi,DWORD [4+esp]
319	lea esi,[32+esp]
320	mov eax,DWORD [esi]
321	mov ecx,ebx
322	xor edx,edx
323	align 16
324	L$012sub:
325	sbb eax,DWORD [edx*4+ebp]
326	mov DWORD [edx*4+edi],eax
327	dec ecx
328	mov eax,DWORD [4+edx*4+esi]
329	lea edx,[1+edx]
330	jge NEAR L$012sub
331	sbb eax,0
332	mov edx,-1
333	xor edx,eax
334	jmp NEAR L$013copy
335	align 16
336	L$013copy:
337	mov esi,DWORD [32+ebx*4+esp]
338	mov ebp,DWORD [ebx*4+edi]
339	mov DWORD [32+ebx*4+esp],ecx
340	and esi,eax
341	and ebp,edx
342	or ebp,esi
343	mov DWORD [ebx*4+edi],ebp
344	dec ebx
345	jge NEAR L$013copy
346	mov esp,DWORD [24+esp]
347	mov eax,1
348	L$000just_leave:
349	pop edi
350	pop esi
351	pop ebx
352	pop ebp
353	ret
354	db 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
355	db 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
356	db 54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
357	db 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
358	db 111,114,103,62,0

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

Download in other formats:

Original Format