x86-gf2m.S@ 98024

Last change on this file since 98024 was 97372, checked in by vboxsync, 2 years ago
libs: Switch to openssl-3.0.7, bugref:10317
File size: 5.8 KB

Line
1	.text
2	.align 4
3	__mul_1x1_mmx:
4	#ifdef __CET__
5
6	.byte 243,15,30,251
7	#endif
8
9	subl $36,%esp
10	movl %eax,%ecx
11	leal (%eax,%eax,1),%edx
12	andl $1073741823,%ecx
13	leal (%edx,%edx,1),%ebp
14	movl $0,(%esp)
15	andl $2147483647,%edx
16	movd %eax,%mm2
17	movd %ebx,%mm3
18	movl %ecx,4(%esp)
19	xorl %edx,%ecx
20	pxor %mm5,%mm5
21	pxor %mm4,%mm4
22	movl %edx,8(%esp)
23	xorl %ebp,%edx
24	movl %ecx,12(%esp)
25	pcmpgtd %mm2,%mm5
26	paddd %mm2,%mm2
27	xorl %edx,%ecx
28	movl %ebp,16(%esp)
29	xorl %edx,%ebp
30	pand %mm3,%mm5
31	pcmpgtd %mm2,%mm4
32	movl %ecx,20(%esp)
33	xorl %ecx,%ebp
34	psllq $31,%mm5
35	pand %mm3,%mm4
36	movl %edx,24(%esp)
37	movl $7,%esi
38	movl %ebp,28(%esp)
39	movl %esi,%ebp
40	andl %ebx,%esi
41	shrl $3,%ebx
42	movl %ebp,%edi
43	psllq $30,%mm4
44	andl %ebx,%edi
45	shrl $3,%ebx
46	movd (%esp,%esi,4),%mm0
47	movl %ebp,%esi
48	andl %ebx,%esi
49	shrl $3,%ebx
50	movd (%esp,%edi,4),%mm2
51	movl %ebp,%edi
52	psllq $3,%mm2
53	andl %ebx,%edi
54	shrl $3,%ebx
55	pxor %mm2,%mm0
56	movd (%esp,%esi,4),%mm1
57	movl %ebp,%esi
58	psllq $6,%mm1
59	andl %ebx,%esi
60	shrl $3,%ebx
61	pxor %mm1,%mm0
62	movd (%esp,%edi,4),%mm2
63	movl %ebp,%edi
64	psllq $9,%mm2
65	andl %ebx,%edi
66	shrl $3,%ebx
67	pxor %mm2,%mm0
68	movd (%esp,%esi,4),%mm1
69	movl %ebp,%esi
70	psllq $12,%mm1
71	andl %ebx,%esi
72	shrl $3,%ebx
73	pxor %mm1,%mm0
74	movd (%esp,%edi,4),%mm2
75	movl %ebp,%edi
76	psllq $15,%mm2
77	andl %ebx,%edi
78	shrl $3,%ebx
79	pxor %mm2,%mm0
80	movd (%esp,%esi,4),%mm1
81	movl %ebp,%esi
82	psllq $18,%mm1
83	andl %ebx,%esi
84	shrl $3,%ebx
85	pxor %mm1,%mm0
86	movd (%esp,%edi,4),%mm2
87	movl %ebp,%edi
88	psllq $21,%mm2
89	andl %ebx,%edi
90	shrl $3,%ebx
91	pxor %mm2,%mm0
92	movd (%esp,%esi,4),%mm1
93	movl %ebp,%esi
94	psllq $24,%mm1
95	andl %ebx,%esi
96	shrl $3,%ebx
97	pxor %mm1,%mm0
98	movd (%esp,%edi,4),%mm2
99	pxor %mm4,%mm0
100	psllq $27,%mm2
101	pxor %mm2,%mm0
102	movd (%esp,%esi,4),%mm1
103	pxor %mm5,%mm0
104	psllq $30,%mm1
105	addl $36,%esp
106	pxor %mm1,%mm0
107	ret
108	.align 4
109	__mul_1x1_ialu:
110	#ifdef __CET__
111
112	.byte 243,15,30,251
113	#endif
114
115	subl $36,%esp
116	movl %eax,%ecx
117	leal (%eax,%eax,1),%edx
118	leal (,%eax,4),%ebp
119	andl $1073741823,%ecx
120	leal (%eax,%eax,1),%edi
121	sarl $31,%eax
122	movl $0,(%esp)
123	andl $2147483647,%edx
124	movl %ecx,4(%esp)
125	xorl %edx,%ecx
126	movl %edx,8(%esp)
127	xorl %ebp,%edx
128	movl %ecx,12(%esp)
129	xorl %edx,%ecx
130	movl %ebp,16(%esp)
131	xorl %edx,%ebp
132	movl %ecx,20(%esp)
133	xorl %ecx,%ebp
134	sarl $31,%edi
135	andl %ebx,%eax
136	movl %edx,24(%esp)
137	andl %ebx,%edi
138	movl %ebp,28(%esp)
139	movl %eax,%edx
140	shll $31,%eax
141	movl %edi,%ecx
142	shrl $1,%edx
143	movl $7,%esi
144	shll $30,%edi
145	andl %ebx,%esi
146	shrl $2,%ecx
147	xorl %edi,%eax
148	shrl $3,%ebx
149	movl $7,%edi
150	andl %ebx,%edi
151	shrl $3,%ebx
152	xorl %ecx,%edx
153	xorl (%esp,%esi,4),%eax
154	movl $7,%esi
155	andl %ebx,%esi
156	shrl $3,%ebx
157	movl (%esp,%edi,4),%ebp
158	movl $7,%edi
159	movl %ebp,%ecx
160	shll $3,%ebp
161	andl %ebx,%edi
162	shrl $29,%ecx
163	xorl %ebp,%eax
164	shrl $3,%ebx
165	xorl %ecx,%edx
166	movl (%esp,%esi,4),%ecx
167	movl $7,%esi
168	movl %ecx,%ebp
169	shll $6,%ecx
170	andl %ebx,%esi
171	shrl $26,%ebp
172	xorl %ecx,%eax
173	shrl $3,%ebx
174	xorl %ebp,%edx
175	movl (%esp,%edi,4),%ebp
176	movl $7,%edi
177	movl %ebp,%ecx
178	shll $9,%ebp
179	andl %ebx,%edi
180	shrl $23,%ecx
181	xorl %ebp,%eax
182	shrl $3,%ebx
183	xorl %ecx,%edx
184	movl (%esp,%esi,4),%ecx
185	movl $7,%esi
186	movl %ecx,%ebp
187	shll $12,%ecx
188	andl %ebx,%esi
189	shrl $20,%ebp
190	xorl %ecx,%eax
191	shrl $3,%ebx
192	xorl %ebp,%edx
193	movl (%esp,%edi,4),%ebp
194	movl $7,%edi
195	movl %ebp,%ecx
196	shll $15,%ebp
197	andl %ebx,%edi
198	shrl $17,%ecx
199	xorl %ebp,%eax
200	shrl $3,%ebx
201	xorl %ecx,%edx
202	movl (%esp,%esi,4),%ecx
203	movl $7,%esi
204	movl %ecx,%ebp
205	shll $18,%ecx
206	andl %ebx,%esi
207	shrl $14,%ebp
208	xorl %ecx,%eax
209	shrl $3,%ebx
210	xorl %ebp,%edx
211	movl (%esp,%edi,4),%ebp
212	movl $7,%edi
213	movl %ebp,%ecx
214	shll $21,%ebp
215	andl %ebx,%edi
216	shrl $11,%ecx
217	xorl %ebp,%eax
218	shrl $3,%ebx
219	xorl %ecx,%edx
220	movl (%esp,%esi,4),%ecx
221	movl $7,%esi
222	movl %ecx,%ebp
223	shll $24,%ecx
224	andl %ebx,%esi
225	shrl $8,%ebp
226	xorl %ecx,%eax
227	shrl $3,%ebx
228	xorl %ebp,%edx
229	movl (%esp,%edi,4),%ebp
230	movl %ebp,%ecx
231	shll $27,%ebp
232	movl (%esp,%esi,4),%edi
233	shrl $5,%ecx
234	movl %edi,%esi
235	xorl %ebp,%eax
236	shll $30,%edi
237	xorl %ecx,%edx
238	shrl $2,%esi
239	xorl %edi,%eax
240	xorl %esi,%edx
241	addl $36,%esp
242	ret
243	.globl _bn_GF2m_mul_2x2
244	.align 4
245	_bn_GF2m_mul_2x2:
246	L_bn_GF2m_mul_2x2_begin:
247	#ifdef __CET__
248
249	.byte 243,15,30,251
250	#endif
251
252	call L000PIC_me_up
253	L000PIC_me_up:
254	popl %edx
255	movl L_OPENSSL_ia32cap_P$non_lazy_ptr-L000PIC_me_up(%edx),%edx
256	movl (%edx),%eax
257	movl 4(%edx),%edx
258	testl $8388608,%eax
259	jz L001ialu
260	pushl %ebp
261	pushl %ebx
262	pushl %esi
263	pushl %edi
264	movl 24(%esp),%eax
265	movl 32(%esp),%ebx
266	call __mul_1x1_mmx
267	movq %mm0,%mm7
268	movl 28(%esp),%eax
269	movl 36(%esp),%ebx
270	call __mul_1x1_mmx
271	movq %mm0,%mm6
272	movl 24(%esp),%eax
273	movl 32(%esp),%ebx
274	xorl 28(%esp),%eax
275	xorl 36(%esp),%ebx
276	call __mul_1x1_mmx
277	pxor %mm7,%mm0
278	movl 20(%esp),%eax
279	pxor %mm6,%mm0
280	movq %mm0,%mm2
281	psllq $32,%mm0
282	popl %edi
283	psrlq $32,%mm2
284	popl %esi
285	pxor %mm6,%mm0
286	popl %ebx
287	pxor %mm7,%mm2
288	movq %mm0,(%eax)
289	popl %ebp
290	movq %mm2,8(%eax)
291	emms
292	ret
293	.align 4,0x90
294	L001ialu:
295	pushl %ebp
296	pushl %ebx
297	pushl %esi
298	pushl %edi
299	subl $20,%esp
300	movl 44(%esp),%eax
301	movl 52(%esp),%ebx
302	call __mul_1x1_ialu
303	movl %eax,8(%esp)
304	movl %edx,12(%esp)
305	movl 48(%esp),%eax
306	movl 56(%esp),%ebx
307	call __mul_1x1_ialu
308	movl %eax,(%esp)
309	movl %edx,4(%esp)
310	movl 44(%esp),%eax
311	movl 52(%esp),%ebx
312	xorl 48(%esp),%eax
313	xorl 56(%esp),%ebx
314	call __mul_1x1_ialu
315	movl 40(%esp),%ebp
316	movl (%esp),%ebx
317	movl 4(%esp),%ecx
318	movl 8(%esp),%edi
319	movl 12(%esp),%esi
320	xorl %edx,%eax
321	xorl %ecx,%edx
322	xorl %ebx,%eax
323	movl %ebx,(%ebp)
324	xorl %edi,%edx
325	movl %esi,12(%ebp)
326	xorl %esi,%eax
327	addl $20,%esp
328	xorl %esi,%edx
329	popl %edi
330	xorl %edx,%eax
331	popl %esi
332	movl %edx,8(%ebp)
333	popl %ebx
334	movl %eax,4(%ebp)
335	popl %ebp
336	ret
337	.byte 71,70,40,50,94,109,41,32,77,117,108,116,105,112,108,105
338	.byte 99,97,116,105,111,110,32,102,111,114,32,120,56,54,44,32
339	.byte 67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
340	.byte 112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
341	.byte 62,0
342	.section __IMPORT,__pointers,non_lazy_symbol_pointers
343	L_OPENSSL_ia32cap_P$non_lazy_ptr:
344	.indirect_symbol _OPENSSL_ia32cap_P
345	.long 0
346	.comm _OPENSSL_ia32cap_P,16,2

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

Download in other formats:

Original Format