x86-mont.S@ 100942

Last change on this file since 100942 was 97372, checked in by vboxsync, 2 years ago
libs: Switch to openssl-3.0.7, bugref:10317
File size: 6.6 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	#ifdef __CET__
8
9	.byte 243,15,30,251
10	#endif
11
12	pushl %ebp
13	pushl %ebx
14	pushl %esi
15	pushl %edi
16	xorl %eax,%eax
17	movl 40(%esp),%edi
18	cmpl $4,%edi
19	jl .L000just_leave
20	leal 20(%esp),%esi
21	leal 24(%esp),%edx
22	addl $2,%edi
23	negl %edi
24	leal -32(%esp,%edi,4),%ebp
25	negl %edi
26	movl %ebp,%eax
27	subl %edx,%eax
28	andl $2047,%eax
29	subl %eax,%ebp
30	xorl %ebp,%edx
31	andl $2048,%edx
32	xorl $2048,%edx
33	subl %edx,%ebp
34	andl $-64,%ebp
35	movl %esp,%eax
36	subl %ebp,%eax
37	andl $-4096,%eax
38	movl %esp,%edx
39	leal (%ebp,%eax,1),%esp
40	movl (%esp),%eax
41	cmpl %ebp,%esp
42	ja .L001page_walk
43	jmp .L002page_walk_done
44	.align 16
45	.L001page_walk:
46	leal -4096(%esp),%esp
47	movl (%esp),%eax
48	cmpl %ebp,%esp
49	ja .L001page_walk
50	.L002page_walk_done:
51	movl (%esi),%eax
52	movl 4(%esi),%ebx
53	movl 8(%esi),%ecx
54	movl 12(%esi),%ebp
55	movl 16(%esi),%esi
56	movl (%esi),%esi
57	movl %eax,4(%esp)
58	movl %ebx,8(%esp)
59	movl %ecx,12(%esp)
60	movl %ebp,16(%esp)
61	movl %esi,20(%esp)
62	leal -3(%edi),%ebx
63	movl %edx,24(%esp)
64	movl 8(%esp),%esi
65	leal 1(%ebx),%ebp
66	movl 12(%esp),%edi
67	xorl %ecx,%ecx
68	movl %esi,%edx
69	andl $1,%ebp
70	subl %edi,%edx
71	leal 4(%edi,%ebx,4),%eax
72	orl %edx,%ebp
73	movl (%edi),%edi
74	jz .L003bn_sqr_mont
75	movl %eax,28(%esp)
76	movl (%esi),%eax
77	xorl %edx,%edx
78	.align 16
79	.L004mull:
80	movl %edx,%ebp
81	mull %edi
82	addl %eax,%ebp
83	leal 1(%ecx),%ecx
84	adcl $0,%edx
85	movl (%esi,%ecx,4),%eax
86	cmpl %ebx,%ecx
87	movl %ebp,28(%esp,%ecx,4)
88	jl .L004mull
89	movl %edx,%ebp
90	mull %edi
91	movl 20(%esp),%edi
92	addl %ebp,%eax
93	movl 16(%esp),%esi
94	adcl $0,%edx
95	imull 32(%esp),%edi
96	movl %eax,32(%esp,%ebx,4)
97	xorl %ecx,%ecx
98	movl %edx,36(%esp,%ebx,4)
99	movl %ecx,40(%esp,%ebx,4)
100	movl (%esi),%eax
101	mull %edi
102	addl 32(%esp),%eax
103	movl 4(%esi),%eax
104	adcl $0,%edx
105	incl %ecx
106	jmp .L0052ndmadd
107	.align 16
108	.L0061stmadd:
109	movl %edx,%ebp
110	mull %edi
111	addl 32(%esp,%ecx,4),%ebp
112	leal 1(%ecx),%ecx
113	adcl $0,%edx
114	addl %eax,%ebp
115	movl (%esi,%ecx,4),%eax
116	adcl $0,%edx
117	cmpl %ebx,%ecx
118	movl %ebp,28(%esp,%ecx,4)
119	jl .L0061stmadd
120	movl %edx,%ebp
121	mull %edi
122	addl 32(%esp,%ebx,4),%eax
123	movl 20(%esp),%edi
124	adcl $0,%edx
125	movl 16(%esp),%esi
126	addl %eax,%ebp
127	adcl $0,%edx
128	imull 32(%esp),%edi
129	xorl %ecx,%ecx
130	addl 36(%esp,%ebx,4),%edx
131	movl %ebp,32(%esp,%ebx,4)
132	adcl $0,%ecx
133	movl (%esi),%eax
134	movl %edx,36(%esp,%ebx,4)
135	movl %ecx,40(%esp,%ebx,4)
136	mull %edi
137	addl 32(%esp),%eax
138	movl 4(%esi),%eax
139	adcl $0,%edx
140	movl $1,%ecx
141	.align 16
142	.L0052ndmadd:
143	movl %edx,%ebp
144	mull %edi
145	addl 32(%esp,%ecx,4),%ebp
146	leal 1(%ecx),%ecx
147	adcl $0,%edx
148	addl %eax,%ebp
149	movl (%esi,%ecx,4),%eax
150	adcl $0,%edx
151	cmpl %ebx,%ecx
152	movl %ebp,24(%esp,%ecx,4)
153	jl .L0052ndmadd
154	movl %edx,%ebp
155	mull %edi
156	addl 32(%esp,%ebx,4),%ebp
157	adcl $0,%edx
158	addl %eax,%ebp
159	adcl $0,%edx
160	movl %ebp,28(%esp,%ebx,4)
161	xorl %eax,%eax
162	movl 12(%esp),%ecx
163	addl 36(%esp,%ebx,4),%edx
164	adcl 40(%esp,%ebx,4),%eax
165	leal 4(%ecx),%ecx
166	movl %edx,32(%esp,%ebx,4)
167	cmpl 28(%esp),%ecx
168	movl %eax,36(%esp,%ebx,4)
169	je .L007common_tail
170	movl (%ecx),%edi
171	movl 8(%esp),%esi
172	movl %ecx,12(%esp)
173	xorl %ecx,%ecx
174	xorl %edx,%edx
175	movl (%esi),%eax
176	jmp .L0061stmadd
177	.align 16
178	.L003bn_sqr_mont:
179	movl %ebx,(%esp)
180	movl %ecx,12(%esp)
181	movl %edi,%eax
182	mull %edi
183	movl %eax,32(%esp)
184	movl %edx,%ebx
185	shrl $1,%edx
186	andl $1,%ebx
187	incl %ecx
188	.align 16
189	.L008sqr:
190	movl (%esi,%ecx,4),%eax
191	movl %edx,%ebp
192	mull %edi
193	addl %ebp,%eax
194	leal 1(%ecx),%ecx
195	adcl $0,%edx
196	leal (%ebx,%eax,2),%ebp
197	shrl $31,%eax
198	cmpl (%esp),%ecx
199	movl %eax,%ebx
200	movl %ebp,28(%esp,%ecx,4)
201	jl .L008sqr
202	movl (%esi,%ecx,4),%eax
203	movl %edx,%ebp
204	mull %edi
205	addl %ebp,%eax
206	movl 20(%esp),%edi
207	adcl $0,%edx
208	movl 16(%esp),%esi
209	leal (%ebx,%eax,2),%ebp
210	imull 32(%esp),%edi
211	shrl $31,%eax
212	movl %ebp,32(%esp,%ecx,4)
213	leal (%eax,%edx,2),%ebp
214	movl (%esi),%eax
215	shrl $31,%edx
216	movl %ebp,36(%esp,%ecx,4)
217	movl %edx,40(%esp,%ecx,4)
218	mull %edi
219	addl 32(%esp),%eax
220	movl %ecx,%ebx
221	adcl $0,%edx
222	movl 4(%esi),%eax
223	movl $1,%ecx
224	.align 16
225	.L0093rdmadd:
226	movl %edx,%ebp
227	mull %edi
228	addl 32(%esp,%ecx,4),%ebp
229	adcl $0,%edx
230	addl %eax,%ebp
231	movl 4(%esi,%ecx,4),%eax
232	adcl $0,%edx
233	movl %ebp,28(%esp,%ecx,4)
234	movl %edx,%ebp
235	mull %edi
236	addl 36(%esp,%ecx,4),%ebp
237	leal 2(%ecx),%ecx
238	adcl $0,%edx
239	addl %eax,%ebp
240	movl (%esi,%ecx,4),%eax
241	adcl $0,%edx
242	cmpl %ebx,%ecx
243	movl %ebp,24(%esp,%ecx,4)
244	jl .L0093rdmadd
245	movl %edx,%ebp
246	mull %edi
247	addl 32(%esp,%ebx,4),%ebp
248	adcl $0,%edx
249	addl %eax,%ebp
250	adcl $0,%edx
251	movl %ebp,28(%esp,%ebx,4)
252	movl 12(%esp),%ecx
253	xorl %eax,%eax
254	movl 8(%esp),%esi
255	addl 36(%esp,%ebx,4),%edx
256	adcl 40(%esp,%ebx,4),%eax
257	movl %edx,32(%esp,%ebx,4)
258	cmpl %ebx,%ecx
259	movl %eax,36(%esp,%ebx,4)
260	je .L007common_tail
261	movl 4(%esi,%ecx,4),%edi
262	leal 1(%ecx),%ecx
263	movl %edi,%eax
264	movl %ecx,12(%esp)
265	mull %edi
266	addl 32(%esp,%ecx,4),%eax
267	adcl $0,%edx
268	movl %eax,32(%esp,%ecx,4)
269	xorl %ebp,%ebp
270	cmpl %ebx,%ecx
271	leal 1(%ecx),%ecx
272	je .L010sqrlast
273	movl %edx,%ebx
274	shrl $1,%edx
275	andl $1,%ebx
276	.align 16
277	.L011sqradd:
278	movl (%esi,%ecx,4),%eax
279	movl %edx,%ebp
280	mull %edi
281	addl %ebp,%eax
282	leal (%eax,%eax,1),%ebp
283	adcl $0,%edx
284	shrl $31,%eax
285	addl 32(%esp,%ecx,4),%ebp
286	leal 1(%ecx),%ecx
287	adcl $0,%eax
288	addl %ebx,%ebp
289	adcl $0,%eax
290	cmpl (%esp),%ecx
291	movl %ebp,28(%esp,%ecx,4)
292	movl %eax,%ebx
293	jle .L011sqradd
294	movl %edx,%ebp
295	addl %edx,%edx
296	shrl $31,%ebp
297	addl %ebx,%edx
298	adcl $0,%ebp
299	.L010sqrlast:
300	movl 20(%esp),%edi
301	movl 16(%esp),%esi
302	imull 32(%esp),%edi
303	addl 32(%esp,%ecx,4),%edx
304	movl (%esi),%eax
305	adcl $0,%ebp
306	movl %edx,32(%esp,%ecx,4)
307	movl %ebp,36(%esp,%ecx,4)
308	mull %edi
309	addl 32(%esp),%eax
310	leal -1(%ecx),%ebx
311	adcl $0,%edx
312	movl $1,%ecx
313	movl 4(%esi),%eax
314	jmp .L0093rdmadd
315	.align 16
316	.L007common_tail:
317	movl 16(%esp),%ebp
318	movl 4(%esp),%edi
319	leal 32(%esp),%esi
320	movl (%esi),%eax
321	movl %ebx,%ecx
322	xorl %edx,%edx
323	.align 16
324	.L012sub:
325	sbbl (%ebp,%edx,4),%eax
326	movl %eax,(%edi,%edx,4)
327	decl %ecx
328	movl 4(%esi,%edx,4),%eax
329	leal 1(%edx),%edx
330	jge .L012sub
331	sbbl $0,%eax
332	movl $-1,%edx
333	xorl %eax,%edx
334	jmp .L013copy
335	.align 16
336	.L013copy:
337	movl 32(%esp,%ebx,4),%esi
338	movl (%edi,%ebx,4),%ebp
339	movl %ecx,32(%esp,%ebx,4)
340	andl %eax,%esi
341	andl %edx,%ebp
342	orl %esi,%ebp
343	movl %ebp,(%edi,%ebx,4)
344	decl %ebx
345	jge .L013copy
346	movl 24(%esp),%esp
347	movl $1,%eax
348	.L000just_leave:
349	popl %edi
350	popl %esi
351	popl %ebx
352	popl %ebp
353	ret
354	.size bn_mul_mont,.-.L_bn_mul_mont_begin
355	.byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
356	.byte 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
357	.byte 54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
358	.byte 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
359	.byte 111,114,103,62,0
360
361	.section ".note.gnu.property", "a"
362	.p2align 2
363	.long 1f - 0f
364	.long 4f - 1f
365	.long 5
366	0:
367	.asciz "GNU"
368	1:
369	.p2align 2
370	.long 0xc0000002
371	.long 3f - 2f
372	2:
373	.long 3
374	3:
375	.p2align 2
376	4:

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

Download in other formats:

Original Format