x86-gf2m.S@ 90293

Last change on this file since 90293 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: 5.7 KB

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

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

Download in other formats:

Original Format