x86-mont.S@ 83531

Last change on this file since 83531 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	.text
2	.globl bn_mul_mont
3	.type bn_mul_mont,@function
4	.align 16
5	bn_mul_mont:
6	.L_bn_mul_mont_begin:
7	pushl %ebp
8	pushl %ebx
9	pushl %esi
10	pushl %edi
11	xorl %eax,%eax
12	movl 40(%esp),%edi
13	cmpl $4,%edi
14	jl .L000just_leave
15	leal 20(%esp),%esi
16	leal 24(%esp),%edx
17	addl $2,%edi
18	negl %edi
19	leal -32(%esp,%edi,4),%ebp
20	negl %edi
21	movl %ebp,%eax
22	subl %edx,%eax
23	andl $2047,%eax
24	subl %eax,%ebp
25	xorl %ebp,%edx
26	andl $2048,%edx
27	xorl $2048,%edx
28	subl %edx,%ebp
29	andl $-64,%ebp
30	movl %esp,%eax
31	subl %ebp,%eax
32	andl $-4096,%eax
33	movl %esp,%edx
34	leal (%ebp,%eax,1),%esp
35	movl (%esp),%eax
36	cmpl %ebp,%esp
37	ja .L001page_walk
38	jmp .L002page_walk_done
39	.align 16
40	.L001page_walk:
41	leal -4096(%esp),%esp
42	movl (%esp),%eax
43	cmpl %ebp,%esp
44	ja .L001page_walk
45	.L002page_walk_done:
46	movl (%esi),%eax
47	movl 4(%esi),%ebx
48	movl 8(%esi),%ecx
49	movl 12(%esi),%ebp
50	movl 16(%esi),%esi
51	movl (%esi),%esi
52	movl %eax,4(%esp)
53	movl %ebx,8(%esp)
54	movl %ecx,12(%esp)
55	movl %ebp,16(%esp)
56	movl %esi,20(%esp)
57	leal -3(%edi),%ebx
58	movl %edx,24(%esp)
59	movl 8(%esp),%esi
60	leal 1(%ebx),%ebp
61	movl 12(%esp),%edi
62	xorl %ecx,%ecx
63	movl %esi,%edx
64	andl $1,%ebp
65	subl %edi,%edx
66	leal 4(%edi,%ebx,4),%eax
67	orl %edx,%ebp
68	movl (%edi),%edi
69	jz .L003bn_sqr_mont
70	movl %eax,28(%esp)
71	movl (%esi),%eax
72	xorl %edx,%edx
73	.align 16
74	.L004mull:
75	movl %edx,%ebp
76	mull %edi
77	addl %eax,%ebp
78	leal 1(%ecx),%ecx
79	adcl $0,%edx
80	movl (%esi,%ecx,4),%eax
81	cmpl %ebx,%ecx
82	movl %ebp,28(%esp,%ecx,4)
83	jl .L004mull
84	movl %edx,%ebp
85	mull %edi
86	movl 20(%esp),%edi
87	addl %ebp,%eax
88	movl 16(%esp),%esi
89	adcl $0,%edx
90	imull 32(%esp),%edi
91	movl %eax,32(%esp,%ebx,4)
92	xorl %ecx,%ecx
93	movl %edx,36(%esp,%ebx,4)
94	movl %ecx,40(%esp,%ebx,4)
95	movl (%esi),%eax
96	mull %edi
97	addl 32(%esp),%eax
98	movl 4(%esi),%eax
99	adcl $0,%edx
100	incl %ecx
101	jmp .L0052ndmadd
102	.align 16
103	.L0061stmadd:
104	movl %edx,%ebp
105	mull %edi
106	addl 32(%esp,%ecx,4),%ebp
107	leal 1(%ecx),%ecx
108	adcl $0,%edx
109	addl %eax,%ebp
110	movl (%esi,%ecx,4),%eax
111	adcl $0,%edx
112	cmpl %ebx,%ecx
113	movl %ebp,28(%esp,%ecx,4)
114	jl .L0061stmadd
115	movl %edx,%ebp
116	mull %edi
117	addl 32(%esp,%ebx,4),%eax
118	movl 20(%esp),%edi
119	adcl $0,%edx
120	movl 16(%esp),%esi
121	addl %eax,%ebp
122	adcl $0,%edx
123	imull 32(%esp),%edi
124	xorl %ecx,%ecx
125	addl 36(%esp,%ebx,4),%edx
126	movl %ebp,32(%esp,%ebx,4)
127	adcl $0,%ecx
128	movl (%esi),%eax
129	movl %edx,36(%esp,%ebx,4)
130	movl %ecx,40(%esp,%ebx,4)
131	mull %edi
132	addl 32(%esp),%eax
133	movl 4(%esi),%eax
134	adcl $0,%edx
135	movl $1,%ecx
136	.align 16
137	.L0052ndmadd:
138	movl %edx,%ebp
139	mull %edi
140	addl 32(%esp,%ecx,4),%ebp
141	leal 1(%ecx),%ecx
142	adcl $0,%edx
143	addl %eax,%ebp
144	movl (%esi,%ecx,4),%eax
145	adcl $0,%edx
146	cmpl %ebx,%ecx
147	movl %ebp,24(%esp,%ecx,4)
148	jl .L0052ndmadd
149	movl %edx,%ebp
150	mull %edi
151	addl 32(%esp,%ebx,4),%ebp
152	adcl $0,%edx
153	addl %eax,%ebp
154	adcl $0,%edx
155	movl %ebp,28(%esp,%ebx,4)
156	xorl %eax,%eax
157	movl 12(%esp),%ecx
158	addl 36(%esp,%ebx,4),%edx
159	adcl 40(%esp,%ebx,4),%eax
160	leal 4(%ecx),%ecx
161	movl %edx,32(%esp,%ebx,4)
162	cmpl 28(%esp),%ecx
163	movl %eax,36(%esp,%ebx,4)
164	je .L007common_tail
165	movl (%ecx),%edi
166	movl 8(%esp),%esi
167	movl %ecx,12(%esp)
168	xorl %ecx,%ecx
169	xorl %edx,%edx
170	movl (%esi),%eax
171	jmp .L0061stmadd
172	.align 16
173	.L003bn_sqr_mont:
174	movl %ebx,(%esp)
175	movl %ecx,12(%esp)
176	movl %edi,%eax
177	mull %edi
178	movl %eax,32(%esp)
179	movl %edx,%ebx
180	shrl $1,%edx
181	andl $1,%ebx
182	incl %ecx
183	.align 16
184	.L008sqr:
185	movl (%esi,%ecx,4),%eax
186	movl %edx,%ebp
187	mull %edi
188	addl %ebp,%eax
189	leal 1(%ecx),%ecx
190	adcl $0,%edx
191	leal (%ebx,%eax,2),%ebp
192	shrl $31,%eax
193	cmpl (%esp),%ecx
194	movl %eax,%ebx
195	movl %ebp,28(%esp,%ecx,4)
196	jl .L008sqr
197	movl (%esi,%ecx,4),%eax
198	movl %edx,%ebp
199	mull %edi
200	addl %ebp,%eax
201	movl 20(%esp),%edi
202	adcl $0,%edx
203	movl 16(%esp),%esi
204	leal (%ebx,%eax,2),%ebp
205	imull 32(%esp),%edi
206	shrl $31,%eax
207	movl %ebp,32(%esp,%ecx,4)
208	leal (%eax,%edx,2),%ebp
209	movl (%esi),%eax
210	shrl $31,%edx
211	movl %ebp,36(%esp,%ecx,4)
212	movl %edx,40(%esp,%ecx,4)
213	mull %edi
214	addl 32(%esp),%eax
215	movl %ecx,%ebx
216	adcl $0,%edx
217	movl 4(%esi),%eax
218	movl $1,%ecx
219	.align 16
220	.L0093rdmadd:
221	movl %edx,%ebp
222	mull %edi
223	addl 32(%esp,%ecx,4),%ebp
224	adcl $0,%edx
225	addl %eax,%ebp
226	movl 4(%esi,%ecx,4),%eax
227	adcl $0,%edx
228	movl %ebp,28(%esp,%ecx,4)
229	movl %edx,%ebp
230	mull %edi
231	addl 36(%esp,%ecx,4),%ebp
232	leal 2(%ecx),%ecx
233	adcl $0,%edx
234	addl %eax,%ebp
235	movl (%esi,%ecx,4),%eax
236	adcl $0,%edx
237	cmpl %ebx,%ecx
238	movl %ebp,24(%esp,%ecx,4)
239	jl .L0093rdmadd
240	movl %edx,%ebp
241	mull %edi
242	addl 32(%esp,%ebx,4),%ebp
243	adcl $0,%edx
244	addl %eax,%ebp
245	adcl $0,%edx
246	movl %ebp,28(%esp,%ebx,4)
247	movl 12(%esp),%ecx
248	xorl %eax,%eax
249	movl 8(%esp),%esi
250	addl 36(%esp,%ebx,4),%edx
251	adcl 40(%esp,%ebx,4),%eax
252	movl %edx,32(%esp,%ebx,4)
253	cmpl %ebx,%ecx
254	movl %eax,36(%esp,%ebx,4)
255	je .L007common_tail
256	movl 4(%esi,%ecx,4),%edi
257	leal 1(%ecx),%ecx
258	movl %edi,%eax
259	movl %ecx,12(%esp)
260	mull %edi
261	addl 32(%esp,%ecx,4),%eax
262	adcl $0,%edx
263	movl %eax,32(%esp,%ecx,4)
264	xorl %ebp,%ebp
265	cmpl %ebx,%ecx
266	leal 1(%ecx),%ecx
267	je .L010sqrlast
268	movl %edx,%ebx
269	shrl $1,%edx
270	andl $1,%ebx
271	.align 16
272	.L011sqradd:
273	movl (%esi,%ecx,4),%eax
274	movl %edx,%ebp
275	mull %edi
276	addl %ebp,%eax
277	leal (%eax,%eax,1),%ebp
278	adcl $0,%edx
279	shrl $31,%eax
280	addl 32(%esp,%ecx,4),%ebp
281	leal 1(%ecx),%ecx
282	adcl $0,%eax
283	addl %ebx,%ebp
284	adcl $0,%eax
285	cmpl (%esp),%ecx
286	movl %ebp,28(%esp,%ecx,4)
287	movl %eax,%ebx
288	jle .L011sqradd
289	movl %edx,%ebp
290	addl %edx,%edx
291	shrl $31,%ebp
292	addl %ebx,%edx
293	adcl $0,%ebp
294	.L010sqrlast:
295	movl 20(%esp),%edi
296	movl 16(%esp),%esi
297	imull 32(%esp),%edi
298	addl 32(%esp,%ecx,4),%edx
299	movl (%esi),%eax
300	adcl $0,%ebp
301	movl %edx,32(%esp,%ecx,4)
302	movl %ebp,36(%esp,%ecx,4)
303	mull %edi
304	addl 32(%esp),%eax
305	leal -1(%ecx),%ebx
306	adcl $0,%edx
307	movl $1,%ecx
308	movl 4(%esi),%eax
309	jmp .L0093rdmadd
310	.align 16
311	.L007common_tail:
312	movl 16(%esp),%ebp
313	movl 4(%esp),%edi
314	leal 32(%esp),%esi
315	movl (%esi),%eax
316	movl %ebx,%ecx
317	xorl %edx,%edx
318	.align 16
319	.L012sub:
320	sbbl (%ebp,%edx,4),%eax
321	movl %eax,(%edi,%edx,4)
322	decl %ecx
323	movl 4(%esi,%edx,4),%eax
324	leal 1(%edx),%edx
325	jge .L012sub
326	sbbl $0,%eax
327	movl $-1,%edx
328	xorl %eax,%edx
329	jmp .L013copy
330	.align 16
331	.L013copy:
332	movl 32(%esp,%ebx,4),%esi
333	movl (%edi,%ebx,4),%ebp
334	movl %ecx,32(%esp,%ebx,4)
335	andl %eax,%esi
336	andl %edx,%ebp
337	orl %esi,%ebp
338	movl %ebp,(%edi,%ebx,4)
339	decl %ebx
340	jge .L013copy
341	movl 24(%esp),%esp
342	movl $1,%eax
343	.L000just_leave:
344	popl %edi
345	popl %esi
346	popl %ebx
347	popl %ebp
348	ret
349	.size bn_mul_mont,.-.L_bn_mul_mont_begin
350	.byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
351	.byte 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
352	.byte 54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
353	.byte 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
354	.byte 111,114,103,62,0

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

Download in other formats:

Original Format