summaryrefslogtreecommitdiff
path: root/src/lib/libcrypto/bn/asm
diff options
context:
space:
mode:
Diffstat (limited to 'src/lib/libcrypto/bn/asm')
-rw-r--r--src/lib/libcrypto/bn/asm/README30
-rw-r--r--src/lib/libcrypto/bn/asm/alpha.s344
-rw-r--r--src/lib/libcrypto/bn/asm/pa-risc.s710
-rw-r--r--src/lib/libcrypto/bn/asm/r3000.s646
4 files changed, 1730 insertions, 0 deletions
diff --git a/src/lib/libcrypto/bn/asm/README b/src/lib/libcrypto/bn/asm/README
new file mode 100644
index 0000000000..d93fbff77f
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/README
@@ -0,0 +1,30 @@
1All assember in this directory are just version of the file
2crypto/bn/bn_mulw.c.
3
4Quite a few of these files are just the assember output from gcc since on
5quite a few machines they are 2 times faster than the system compiler.
6
7For the x86, I have hand written assember because of the bad job all
8compilers seem to do on it. This normally gives a 2 time speed up in the RSA
9routines.
10
11For the DEC alpha, I also hand wrote the assember (except the division which
12is just the output from the C compiler pasted on the end of the file).
13On the 2 alpha C compilers I had access to, it was not possible to do
1464b x 64b -> 128b calculations (both long and the long long data types
15were 64 bits). So the hand assember gives access to the 128 bit result and
16a 2 times speedup :-).
17
18The x86xxxx.obj files are the assembled version of x86xxxx.asm files.
19I had such a hard time finding a macro assember for Microsoft, I decided to
20include the object file to save others the hassle :-).
21
22I have also included uu encoded versions of the .obj incase they get
23trashed.
24
25There are 2 versions of assember for the HP PA-RISC.
26pa-risc.s is the origional one which works fine.
27pa-risc2.s is a new version that often generates warnings but if the
28tests pass, it gives performance that is over 2 times faster than
29pa-risc.s.
30Both were generated using gcc :-)
diff --git a/src/lib/libcrypto/bn/asm/alpha.s b/src/lib/libcrypto/bn/asm/alpha.s
new file mode 100644
index 0000000000..1d17b1d619
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/alpha.s
@@ -0,0 +1,344 @@
1 # DEC Alpha assember
2 # The bn_div64 is actually gcc output but the other parts are hand done.
3 # Thanks to tzeruch@ceddec.com for sending me the gcc output for
4 # bn_div64.
5 .file 1 "bn_mulw.c"
6 .set noat
7gcc2_compiled.:
8__gnu_compiled_c:
9 .text
10 .align 3
11 .globl bn_mul_add_words
12 .ent bn_mul_add_words
13bn_mul_add_words:
14bn_mul_add_words..ng:
15 .frame $30,0,$26,0
16 .prologue 0
17 subq $18,2,$25 # num=-2
18 bis $31,$31,$0
19 blt $25,$42
20 .align 5
21$142:
22 subq $18,2,$18 # num-=2
23 subq $25,2,$25 # num-=2
24
25 ldq $1,0($17) # a[0]
26 ldq $2,8($17) # a[1]
27
28 mulq $19,$1,$3 # a[0]*w low part r3
29 umulh $19,$1,$1 # a[0]*w high part r1
30 mulq $19,$2,$4 # a[1]*w low part r4
31 umulh $19,$2,$2 # a[1]*w high part r2
32
33 ldq $22,0($16) # r[0] r22
34 ldq $23,8($16) # r[1] r23
35
36 addq $3,$22,$3 # a0 low part + r[0]
37 addq $4,$23,$4 # a1 low part + r[1]
38 cmpult $3,$22,$5 # overflow?
39 cmpult $4,$23,$6 # overflow?
40 addq $5,$1,$1 # high part + overflow
41 addq $6,$2,$2 # high part + overflow
42
43 addq $3,$0,$3 # add c
44 cmpult $3,$0,$5 # overflow?
45 stq $3,0($16)
46 addq $5,$1,$0 # c=high part + overflow
47
48 addq $4,$0,$4 # add c
49 cmpult $4,$0,$5 # overflow?
50 stq $4,8($16)
51 addq $5,$2,$0 # c=high part + overflow
52
53 ble $18,$43
54
55 addq $16,16,$16
56 addq $17,16,$17
57 blt $25,$42
58
59 br $31,$142
60$42:
61 ldq $1,0($17) # a[0]
62 umulh $19,$1,$3 # a[0]*w high part
63 mulq $19,$1,$1 # a[0]*w low part
64 ldq $2,0($16) # r[0]
65 addq $1,$2,$1 # low part + r[0]
66 cmpult $1,$2,$4 # overflow?
67 addq $4,$3,$3 # high part + overflow
68 addq $1,$0,$1 # add c
69 cmpult $1,$0,$4 # overflow?
70 addq $4,$3,$0 # c=high part + overflow
71 stq $1,0($16)
72
73 .align 4
74$43:
75 ret $31,($26),1
76 .end bn_mul_add_words
77 .align 3
78 .globl bn_mul_words
79 .ent bn_mul_words
80bn_mul_words:
81bn_mul_words..ng:
82 .frame $30,0,$26,0
83 .prologue 0
84 subq $18,2,$25 # num=-2
85 bis $31,$31,$0
86 blt $25,$242
87 .align 5
88$342:
89 subq $18,2,$18 # num-=2
90 subq $25,2,$25 # num-=2
91
92 ldq $1,0($17) # a[0]
93 ldq $2,8($17) # a[1]
94
95 mulq $19,$1,$3 # a[0]*w low part r3
96 umulh $19,$1,$1 # a[0]*w high part r1
97 mulq $19,$2,$4 # a[1]*w low part r4
98 umulh $19,$2,$2 # a[1]*w high part r2
99
100 addq $3,$0,$3 # add c
101 cmpult $3,$0,$5 # overflow?
102 stq $3,0($16)
103 addq $5,$1,$0 # c=high part + overflow
104
105 addq $4,$0,$4 # add c
106 cmpult $4,$0,$5 # overflow?
107 stq $4,8($16)
108 addq $5,$2,$0 # c=high part + overflow
109
110 ble $18,$243
111
112 addq $16,16,$16
113 addq $17,16,$17
114 blt $25,$242
115
116 br $31,$342
117$242:
118 ldq $1,0($17) # a[0]
119 umulh $19,$1,$3 # a[0]*w high part
120 mulq $19,$1,$1 # a[0]*w low part
121 addq $1,$0,$1 # add c
122 cmpult $1,$0,$4 # overflow?
123 addq $4,$3,$0 # c=high part + overflow
124 stq $1,0($16)
125$243:
126 ret $31,($26),1
127 .end bn_mul_words
128 .align 3
129 .globl bn_sqr_words
130 .ent bn_sqr_words
131bn_sqr_words:
132bn_sqr_words..ng:
133 .frame $30,0,$26,0
134 .prologue 0
135
136 subq $18,2,$25 # num=-2
137 blt $25,$442
138 .align 5
139$542:
140 subq $18,2,$18 # num-=2
141 subq $25,2,$25 # num-=2
142
143 ldq $1,0($17) # a[0]
144 ldq $4,8($17) # a[1]
145
146 mulq $1,$1,$2 # a[0]*w low part r2
147 umulh $1,$1,$3 # a[0]*w high part r3
148 mulq $4,$4,$5 # a[1]*w low part r5
149 umulh $4,$4,$6 # a[1]*w high part r6
150
151 stq $2,0($16) # r[0]
152 stq $3,8($16) # r[1]
153 stq $5,16($16) # r[3]
154 stq $6,24($16) # r[4]
155
156 ble $18,$443
157
158 addq $16,32,$16
159 addq $17,16,$17
160 blt $25,$442
161 br $31,$542
162
163$442:
164 ldq $1,0($17) # a[0]
165 mulq $1,$1,$2 # a[0]*w low part r2
166 umulh $1,$1,$3 # a[0]*w high part r3
167 stq $2,0($16) # r[0]
168 stq $3,8($16) # r[1]
169
170 .align 4
171$443:
172 ret $31,($26),1
173 .end bn_sqr_words
174
175 .align 3
176 .globl bn_add_words
177 .ent bn_add_words
178bn_add_words:
179bn_add_words..ng:
180 .frame $30,0,$26,0
181 .prologue 0
182
183 bis $31,$31,$8 # carry = 0
184 ble $19,$900
185$901:
186 ldq $0,0($17) # a[0]
187 ldq $1,0($18) # a[1]
188
189 addq $0,$1,$3 # c=a+b;
190 addq $17,8,$17 # a++
191
192 cmpult $3,$1,$7 # did we overflow?
193 addq $18,8,$18 # b++
194
195 addq $8,$3,$3 # c+=carry
196
197 cmpult $3,$8,$8 # did we overflow?
198 stq $3,($16) # r[0]=c
199
200 addq $7,$8,$8 # add into overflow
201 subq $19,1,$19 # loop--
202
203 addq $16,8,$16 # r++
204 bgt $19,$901
205$900:
206 bis $8,$8,$0 # return carry
207 ret $31,($26),1
208 .end bn_add_words
209
210 #
211 # What follows was taken directly from the C compiler with a few
212 # hacks to redo the lables.
213 #
214.text
215 .align 3
216 .globl bn_div64
217 .ent bn_div64
218bn_div64:
219 ldgp $29,0($27)
220bn_div64..ng:
221 lda $30,-48($30)
222 .frame $30,48,$26,0
223 stq $26,0($30)
224 stq $9,8($30)
225 stq $10,16($30)
226 stq $11,24($30)
227 stq $12,32($30)
228 stq $13,40($30)
229 .mask 0x4003e00,-48
230 .prologue 1
231 bis $16,$16,$9
232 bis $17,$17,$10
233 bis $18,$18,$11
234 bis $31,$31,$13
235 bis $31,2,$12
236 bne $11,$119
237 lda $0,-1
238 br $31,$136
239 .align 4
240$119:
241 bis $11,$11,$16
242 jsr $26,BN_num_bits_word
243 ldgp $29,0($26)
244 subq $0,64,$1
245 beq $1,$120
246 bis $31,1,$1
247 sll $1,$0,$1
248 cmpule $9,$1,$1
249 bne $1,$120
250 # lda $16,_IO_stderr_
251 # lda $17,$C32
252 # bis $0,$0,$18
253 # jsr $26,fprintf
254 # ldgp $29,0($26)
255 jsr $26,abort
256 ldgp $29,0($26)
257 .align 4
258$120:
259 bis $31,64,$3
260 cmpult $9,$11,$2
261 subq $3,$0,$1
262 addl $1,$31,$0
263 subq $9,$11,$1
264 cmoveq $2,$1,$9
265 beq $0,$122
266 zapnot $0,15,$2
267 subq $3,$0,$1
268 sll $11,$2,$11
269 sll $9,$2,$3
270 srl $10,$1,$1
271 sll $10,$2,$10
272 bis $3,$1,$9
273$122:
274 srl $11,32,$5
275 zapnot $11,15,$6
276 lda $7,-1
277 .align 5
278$123:
279 srl $9,32,$1
280 subq $1,$5,$1
281 bne $1,$126
282 zapnot $7,15,$27
283 br $31,$127
284 .align 4
285$126:
286 bis $9,$9,$24
287 bis $5,$5,$25
288 divqu $24,$25,$27
289$127:
290 srl $10,32,$4
291 .align 5
292$128:
293 mulq $27,$5,$1
294 subq $9,$1,$3
295 zapnot $3,240,$1
296 bne $1,$129
297 mulq $6,$27,$2
298 sll $3,32,$1
299 addq $1,$4,$1
300 cmpule $2,$1,$2
301 bne $2,$129
302 subq $27,1,$27
303 br $31,$128
304 .align 4
305$129:
306 mulq $27,$6,$1
307 mulq $27,$5,$4
308 srl $1,32,$3
309 sll $1,32,$1
310 addq $4,$3,$4
311 cmpult $10,$1,$2
312 subq $10,$1,$10
313 addq $2,$4,$2
314 cmpult $9,$2,$1
315 bis $2,$2,$4
316 beq $1,$134
317 addq $9,$11,$9
318 subq $27,1,$27
319$134:
320 subl $12,1,$12
321 subq $9,$4,$9
322 beq $12,$124
323 sll $27,32,$13
324 sll $9,32,$2
325 srl $10,32,$1
326 sll $10,32,$10
327 bis $2,$1,$9
328 br $31,$123
329 .align 4
330$124:
331 bis $13,$27,$0
332$136:
333 ldq $26,0($30)
334 ldq $9,8($30)
335 ldq $10,16($30)
336 ldq $11,24($30)
337 ldq $12,32($30)
338 ldq $13,40($30)
339 addq $30,48,$30
340 ret $31,($26),1
341 .end bn_div64
342 .ident "GCC: (GNU) 2.7.2.1"
343
344
diff --git a/src/lib/libcrypto/bn/asm/pa-risc.s b/src/lib/libcrypto/bn/asm/pa-risc.s
new file mode 100644
index 0000000000..775130a191
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/pa-risc.s
@@ -0,0 +1,710 @@
1 .SPACE $PRIVATE$
2 .SUBSPA $DATA$,QUAD=1,ALIGN=8,ACCESS=31
3 .SUBSPA $BSS$,QUAD=1,ALIGN=8,ACCESS=31,ZERO,SORT=82
4 .SPACE $TEXT$
5 .SUBSPA $LIT$,QUAD=0,ALIGN=8,ACCESS=44
6 .SUBSPA $CODE$,QUAD=0,ALIGN=8,ACCESS=44,CODE_ONLY
7 .IMPORT $global$,DATA
8 .IMPORT $$dyncall,MILLICODE
9; gcc_compiled.:
10 .SPACE $TEXT$
11 .SUBSPA $CODE$
12
13 .align 4
14 .EXPORT bn_mul_add_words,ENTRY,PRIV_LEV=3,ARGW0=GR,ARGW1=GR,ARGW2=GR,ARGW3=GR,RTNVAL=GR
15bn_mul_add_words
16 .PROC
17 .CALLINFO FRAME=0,CALLS,SAVE_RP
18 .ENTRY
19 stw %r2,-20(0,%r30)
20 ldi 0,%r28
21 extru %r23,31,16,%r2
22 stw %r2,-16(0,%r30)
23 extru %r23,15,16,%r23
24 ldil L'65536,%r31
25 fldws -16(0,%r30),%fr11R
26 stw %r23,-16(0,%r30)
27 ldo 12(%r25),%r29
28 ldo 12(%r26),%r23
29 fldws -16(0,%r30),%fr11L
30L$0002
31 ldw 0(0,%r25),%r19
32 extru %r19,31,16,%r20
33 stw %r20,-16(0,%r30)
34 extru %r19,15,16,%r19
35 fldws -16(0,%r30),%fr22L
36 stw %r19,-16(0,%r30)
37 xmpyu %fr22L,%fr11R,%fr8
38 fldws -16(0,%r30),%fr22L
39 fstws %fr8R,-16(0,%r30)
40 xmpyu %fr11R,%fr22L,%fr10
41 ldw -16(0,%r30),%r2
42 stw %r20,-16(0,%r30)
43 xmpyu %fr22L,%fr11L,%fr9
44 fldws -16(0,%r30),%fr22L
45 fstws %fr10R,-16(0,%r30)
46 copy %r2,%r22
47 ldw -16(0,%r30),%r2
48 fstws %fr9R,-16(0,%r30)
49 xmpyu %fr11L,%fr22L,%fr8
50 copy %r2,%r19
51 ldw -16(0,%r30),%r2
52 fstws %fr8R,-16(0,%r30)
53 copy %r2,%r20
54 ldw -16(0,%r30),%r2
55 addl %r2,%r19,%r21
56 comclr,<<= %r19,%r21,0
57 addl %r20,%r31,%r20
58L$0005
59 extru %r21,15,16,%r19
60 addl %r20,%r19,%r20
61 zdep %r21,15,16,%r19
62 addl %r22,%r19,%r22
63 comclr,<<= %r19,%r22,0
64 addi,tr 1,%r20,%r19
65 copy %r20,%r19
66 addl %r22,%r28,%r20
67 comclr,<<= %r28,%r20,0
68 addi 1,%r19,%r19
69 ldw 0(0,%r26),%r28
70 addl %r20,%r28,%r20
71 comclr,<<= %r28,%r20,0
72 addi,tr 1,%r19,%r28
73 copy %r19,%r28
74 addib,= -1,%r24,L$0003
75 stw %r20,0(0,%r26)
76 ldw -8(0,%r29),%r19
77 extru %r19,31,16,%r20
78 stw %r20,-16(0,%r30)
79 extru %r19,15,16,%r19
80 fldws -16(0,%r30),%fr22L
81 stw %r19,-16(0,%r30)
82 xmpyu %fr22L,%fr11R,%fr8
83 fldws -16(0,%r30),%fr22L
84 fstws %fr8R,-16(0,%r30)
85 xmpyu %fr11R,%fr22L,%fr10
86 ldw -16(0,%r30),%r2
87 stw %r20,-16(0,%r30)
88 xmpyu %fr22L,%fr11L,%fr9
89 fldws -16(0,%r30),%fr22L
90 fstws %fr10R,-16(0,%r30)
91 copy %r2,%r22
92 ldw -16(0,%r30),%r2
93 fstws %fr9R,-16(0,%r30)
94 xmpyu %fr11L,%fr22L,%fr8
95 copy %r2,%r19
96 ldw -16(0,%r30),%r2
97 fstws %fr8R,-16(0,%r30)
98 copy %r2,%r20
99 ldw -16(0,%r30),%r2
100 addl %r2,%r19,%r21
101 comclr,<<= %r19,%r21,0
102 addl %r20,%r31,%r20
103L$0010
104 extru %r21,15,16,%r19
105 addl %r20,%r19,%r20
106 zdep %r21,15,16,%r19
107 addl %r22,%r19,%r22
108 comclr,<<= %r19,%r22,0
109 addi,tr 1,%r20,%r19
110 copy %r20,%r19
111 addl %r22,%r28,%r20
112 comclr,<<= %r28,%r20,0
113 addi 1,%r19,%r19
114 ldw -8(0,%r23),%r28
115 addl %r20,%r28,%r20
116 comclr,<<= %r28,%r20,0
117 addi,tr 1,%r19,%r28
118 copy %r19,%r28
119 addib,= -1,%r24,L$0003
120 stw %r20,-8(0,%r23)
121 ldw -4(0,%r29),%r19
122 extru %r19,31,16,%r20
123 stw %r20,-16(0,%r30)
124 extru %r19,15,16,%r19
125 fldws -16(0,%r30),%fr22L
126 stw %r19,-16(0,%r30)
127 xmpyu %fr22L,%fr11R,%fr8
128 fldws -16(0,%r30),%fr22L
129 fstws %fr8R,-16(0,%r30)
130 xmpyu %fr11R,%fr22L,%fr10
131 ldw -16(0,%r30),%r2
132 stw %r20,-16(0,%r30)
133 xmpyu %fr22L,%fr11L,%fr9
134 fldws -16(0,%r30),%fr22L
135 fstws %fr10R,-16(0,%r30)
136 copy %r2,%r22
137 ldw -16(0,%r30),%r2
138 fstws %fr9R,-16(0,%r30)
139 xmpyu %fr11L,%fr22L,%fr8
140 copy %r2,%r19
141 ldw -16(0,%r30),%r2
142 fstws %fr8R,-16(0,%r30)
143 copy %r2,%r20
144 ldw -16(0,%r30),%r2
145 addl %r2,%r19,%r21
146 comclr,<<= %r19,%r21,0
147 addl %r20,%r31,%r20
148L$0015
149 extru %r21,15,16,%r19
150 addl %r20,%r19,%r20
151 zdep %r21,15,16,%r19
152 addl %r22,%r19,%r22
153 comclr,<<= %r19,%r22,0
154 addi,tr 1,%r20,%r19
155 copy %r20,%r19
156 addl %r22,%r28,%r20
157 comclr,<<= %r28,%r20,0
158 addi 1,%r19,%r19
159 ldw -4(0,%r23),%r28
160 addl %r20,%r28,%r20
161 comclr,<<= %r28,%r20,0
162 addi,tr 1,%r19,%r28
163 copy %r19,%r28
164 addib,= -1,%r24,L$0003
165 stw %r20,-4(0,%r23)
166 ldw 0(0,%r29),%r19
167 extru %r19,31,16,%r20
168 stw %r20,-16(0,%r30)
169 extru %r19,15,16,%r19
170 fldws -16(0,%r30),%fr22L
171 stw %r19,-16(0,%r30)
172 xmpyu %fr22L,%fr11R,%fr8
173 fldws -16(0,%r30),%fr22L
174 fstws %fr8R,-16(0,%r30)
175 xmpyu %fr11R,%fr22L,%fr10
176 ldw -16(0,%r30),%r2
177 stw %r20,-16(0,%r30)
178 xmpyu %fr22L,%fr11L,%fr9
179 fldws -16(0,%r30),%fr22L
180 fstws %fr10R,-16(0,%r30)
181 copy %r2,%r22
182 ldw -16(0,%r30),%r2
183 fstws %fr9R,-16(0,%r30)
184 xmpyu %fr11L,%fr22L,%fr8
185 copy %r2,%r19
186 ldw -16(0,%r30),%r2
187 fstws %fr8R,-16(0,%r30)
188 copy %r2,%r20
189 ldw -16(0,%r30),%r2
190 addl %r2,%r19,%r21
191 comclr,<<= %r19,%r21,0
192 addl %r20,%r31,%r20
193L$0020
194 extru %r21,15,16,%r19
195 addl %r20,%r19,%r20
196 zdep %r21,15,16,%r19
197 addl %r22,%r19,%r22
198 comclr,<<= %r19,%r22,0
199 addi,tr 1,%r20,%r19
200 copy %r20,%r19
201 addl %r22,%r28,%r20
202 comclr,<<= %r28,%r20,0
203 addi 1,%r19,%r19
204 ldw 0(0,%r23),%r28
205 addl %r20,%r28,%r20
206 comclr,<<= %r28,%r20,0
207 addi,tr 1,%r19,%r28
208 copy %r19,%r28
209 addib,= -1,%r24,L$0003
210 stw %r20,0(0,%r23)
211 ldo 16(%r29),%r29
212 ldo 16(%r25),%r25
213 ldo 16(%r23),%r23
214 bl L$0002,0
215 ldo 16(%r26),%r26
216L$0003
217 ldw -20(0,%r30),%r2
218 bv,n 0(%r2)
219 .EXIT
220 .PROCEND
221 .align 4
222 .EXPORT bn_mul_words,ENTRY,PRIV_LEV=3,ARGW0=GR,ARGW1=GR,ARGW2=GR,ARGW3=GR,RTNVAL=GR
223bn_mul_words
224 .PROC
225 .CALLINFO FRAME=0,CALLS,SAVE_RP
226 .ENTRY
227 stw %r2,-20(0,%r30)
228 ldi 0,%r28
229 extru %r23,31,16,%r2
230 stw %r2,-16(0,%r30)
231 extru %r23,15,16,%r23
232 ldil L'65536,%r31
233 fldws -16(0,%r30),%fr11R
234 stw %r23,-16(0,%r30)
235 ldo 12(%r26),%r29
236 ldo 12(%r25),%r23
237 fldws -16(0,%r30),%fr11L
238L$0026
239 ldw 0(0,%r25),%r19
240 extru %r19,31,16,%r20
241 stw %r20,-16(0,%r30)
242 extru %r19,15,16,%r19
243 fldws -16(0,%r30),%fr22L
244 stw %r19,-16(0,%r30)
245 xmpyu %fr22L,%fr11R,%fr8
246 fldws -16(0,%r30),%fr22L
247 fstws %fr8R,-16(0,%r30)
248 xmpyu %fr11R,%fr22L,%fr10
249 ldw -16(0,%r30),%r2
250 stw %r20,-16(0,%r30)
251 xmpyu %fr22L,%fr11L,%fr9
252 fldws -16(0,%r30),%fr22L
253 fstws %fr10R,-16(0,%r30)
254 copy %r2,%r22
255 ldw -16(0,%r30),%r2
256 fstws %fr9R,-16(0,%r30)
257 xmpyu %fr11L,%fr22L,%fr8
258 copy %r2,%r19
259 ldw -16(0,%r30),%r2
260 fstws %fr8R,-16(0,%r30)
261 copy %r2,%r20
262 ldw -16(0,%r30),%r2
263 addl %r2,%r19,%r21
264 comclr,<<= %r19,%r21,0
265 addl %r20,%r31,%r20
266L$0029
267 extru %r21,15,16,%r19
268 addl %r20,%r19,%r20
269 zdep %r21,15,16,%r19
270 addl %r22,%r19,%r22
271 comclr,<<= %r19,%r22,0
272 addi,tr 1,%r20,%r19
273 copy %r20,%r19
274 addl %r22,%r28,%r20
275 comclr,<<= %r28,%r20,0
276 addi,tr 1,%r19,%r28
277 copy %r19,%r28
278 addib,= -1,%r24,L$0027
279 stw %r20,0(0,%r26)
280 ldw -8(0,%r23),%r19
281 extru %r19,31,16,%r20
282 stw %r20,-16(0,%r30)
283 extru %r19,15,16,%r19
284 fldws -16(0,%r30),%fr22L
285 stw %r19,-16(0,%r30)
286 xmpyu %fr22L,%fr11R,%fr8
287 fldws -16(0,%r30),%fr22L
288 fstws %fr8R,-16(0,%r30)
289 xmpyu %fr11R,%fr22L,%fr10
290 ldw -16(0,%r30),%r2
291 stw %r20,-16(0,%r30)
292 xmpyu %fr22L,%fr11L,%fr9
293 fldws -16(0,%r30),%fr22L
294 fstws %fr10R,-16(0,%r30)
295 copy %r2,%r22
296 ldw -16(0,%r30),%r2
297 fstws %fr9R,-16(0,%r30)
298 xmpyu %fr11L,%fr22L,%fr8
299 copy %r2,%r19
300 ldw -16(0,%r30),%r2
301 fstws %fr8R,-16(0,%r30)
302 copy %r2,%r20
303 ldw -16(0,%r30),%r2
304 addl %r2,%r19,%r21
305 comclr,<<= %r19,%r21,0
306 addl %r20,%r31,%r20
307L$0033
308 extru %r21,15,16,%r19
309 addl %r20,%r19,%r20
310 zdep %r21,15,16,%r19
311 addl %r22,%r19,%r22
312 comclr,<<= %r19,%r22,0
313 addi,tr 1,%r20,%r19
314 copy %r20,%r19
315 addl %r22,%r28,%r20
316 comclr,<<= %r28,%r20,0
317 addi,tr 1,%r19,%r28
318 copy %r19,%r28
319 addib,= -1,%r24,L$0027
320 stw %r20,-8(0,%r29)
321 ldw -4(0,%r23),%r19
322 extru %r19,31,16,%r20
323 stw %r20,-16(0,%r30)
324 extru %r19,15,16,%r19
325 fldws -16(0,%r30),%fr22L
326 stw %r19,-16(0,%r30)
327 xmpyu %fr22L,%fr11R,%fr8
328 fldws -16(0,%r30),%fr22L
329 fstws %fr8R,-16(0,%r30)
330 xmpyu %fr11R,%fr22L,%fr10
331 ldw -16(0,%r30),%r2
332 stw %r20,-16(0,%r30)
333 xmpyu %fr22L,%fr11L,%fr9
334 fldws -16(0,%r30),%fr22L
335 fstws %fr10R,-16(0,%r30)
336 copy %r2,%r22
337 ldw -16(0,%r30),%r2
338 fstws %fr9R,-16(0,%r30)
339 xmpyu %fr11L,%fr22L,%fr8
340 copy %r2,%r19
341 ldw -16(0,%r30),%r2
342 fstws %fr8R,-16(0,%r30)
343 copy %r2,%r20
344 ldw -16(0,%r30),%r2
345 addl %r2,%r19,%r21
346 comclr,<<= %r19,%r21,0
347 addl %r20,%r31,%r20
348L$0037
349 extru %r21,15,16,%r19
350 addl %r20,%r19,%r20
351 zdep %r21,15,16,%r19
352 addl %r22,%r19,%r22
353 comclr,<<= %r19,%r22,0
354 addi,tr 1,%r20,%r19
355 copy %r20,%r19
356 addl %r22,%r28,%r20
357 comclr,<<= %r28,%r20,0
358 addi,tr 1,%r19,%r28
359 copy %r19,%r28
360 addib,= -1,%r24,L$0027
361 stw %r20,-4(0,%r29)
362 ldw 0(0,%r23),%r19
363 extru %r19,31,16,%r20
364 stw %r20,-16(0,%r30)
365 extru %r19,15,16,%r19
366 fldws -16(0,%r30),%fr22L
367 stw %r19,-16(0,%r30)
368 xmpyu %fr22L,%fr11R,%fr8
369 fldws -16(0,%r30),%fr22L
370 fstws %fr8R,-16(0,%r30)
371 xmpyu %fr11R,%fr22L,%fr10
372 ldw -16(0,%r30),%r2
373 stw %r20,-16(0,%r30)
374 xmpyu %fr22L,%fr11L,%fr9
375 fldws -16(0,%r30),%fr22L
376 fstws %fr10R,-16(0,%r30)
377 copy %r2,%r22
378 ldw -16(0,%r30),%r2
379 fstws %fr9R,-16(0,%r30)
380 xmpyu %fr11L,%fr22L,%fr8
381 copy %r2,%r19
382 ldw -16(0,%r30),%r2
383 fstws %fr8R,-16(0,%r30)
384 copy %r2,%r20
385 ldw -16(0,%r30),%r2
386 addl %r2,%r19,%r21
387 comclr,<<= %r19,%r21,0
388 addl %r20,%r31,%r20
389L$0041
390 extru %r21,15,16,%r19
391 addl %r20,%r19,%r20
392 zdep %r21,15,16,%r19
393 addl %r22,%r19,%r22
394 comclr,<<= %r19,%r22,0
395 addi,tr 1,%r20,%r19
396 copy %r20,%r19
397 addl %r22,%r28,%r20
398 comclr,<<= %r28,%r20,0
399 addi,tr 1,%r19,%r28
400 copy %r19,%r28
401 addib,= -1,%r24,L$0027
402 stw %r20,0(0,%r29)
403 ldo 16(%r23),%r23
404 ldo 16(%r25),%r25
405 ldo 16(%r29),%r29
406 bl L$0026,0
407 ldo 16(%r26),%r26
408L$0027
409 ldw -20(0,%r30),%r2
410 bv,n 0(%r2)
411 .EXIT
412 .PROCEND
413 .align 4
414 .EXPORT bn_sqr_words,ENTRY,PRIV_LEV=3,ARGW0=GR,ARGW1=GR,ARGW2=GR
415bn_sqr_words
416 .PROC
417 .CALLINFO FRAME=0,NO_CALLS
418 .ENTRY
419 ldo 28(%r26),%r23
420 ldo 12(%r25),%r28
421L$0046
422 ldw 0(0,%r25),%r21
423 extru %r21,31,16,%r22
424 stw %r22,-16(0,%r30)
425 extru %r21,15,16,%r21
426 fldws -16(0,%r30),%fr10L
427 stw %r21,-16(0,%r30)
428 fldws -16(0,%r30),%fr10R
429 xmpyu %fr10L,%fr10R,%fr8
430 fstws %fr8R,-16(0,%r30)
431 ldw -16(0,%r30),%r29
432 stw %r22,-16(0,%r30)
433 fldws -16(0,%r30),%fr10R
434 stw %r21,-16(0,%r30)
435 copy %r29,%r19
436 xmpyu %fr10L,%fr10R,%fr8
437 fldws -16(0,%r30),%fr10L
438 stw %r21,-16(0,%r30)
439 fldws -16(0,%r30),%fr10R
440 fstws %fr8R,-16(0,%r30)
441 extru %r19,16,17,%r20
442 zdep %r19,14,15,%r19
443 ldw -16(0,%r30),%r29
444 xmpyu %fr10L,%fr10R,%fr9
445 addl %r29,%r19,%r22
446 stw %r22,0(0,%r26)
447 fstws %fr9R,-16(0,%r30)
448 ldw -16(0,%r30),%r29
449 addl %r29,%r20,%r21
450 comclr,<<= %r19,%r22,0
451 addi 1,%r21,%r21
452 addib,= -1,%r24,L$0057
453 stw %r21,-24(0,%r23)
454 ldw -8(0,%r28),%r21
455 extru %r21,31,16,%r22
456 stw %r22,-16(0,%r30)
457 extru %r21,15,16,%r21
458 fldws -16(0,%r30),%fr10L
459 stw %r21,-16(0,%r30)
460 fldws -16(0,%r30),%fr10R
461 xmpyu %fr10L,%fr10R,%fr8
462 fstws %fr8R,-16(0,%r30)
463 ldw -16(0,%r30),%r29
464 stw %r22,-16(0,%r30)
465 fldws -16(0,%r30),%fr10R
466 stw %r21,-16(0,%r30)
467 copy %r29,%r19
468 xmpyu %fr10L,%fr10R,%fr8
469 fldws -16(0,%r30),%fr10L
470 stw %r21,-16(0,%r30)
471 fldws -16(0,%r30),%fr10R
472 fstws %fr8R,-16(0,%r30)
473 extru %r19,16,17,%r20
474 zdep %r19,14,15,%r19
475 ldw -16(0,%r30),%r29
476 xmpyu %fr10L,%fr10R,%fr9
477 addl %r29,%r19,%r22
478 stw %r22,-20(0,%r23)
479 fstws %fr9R,-16(0,%r30)
480 ldw -16(0,%r30),%r29
481 addl %r29,%r20,%r21
482 comclr,<<= %r19,%r22,0
483 addi 1,%r21,%r21
484 addib,= -1,%r24,L$0057
485 stw %r21,-16(0,%r23)
486 ldw -4(0,%r28),%r21
487 extru %r21,31,16,%r22
488 stw %r22,-16(0,%r30)
489 extru %r21,15,16,%r21
490 fldws -16(0,%r30),%fr10L
491 stw %r21,-16(0,%r30)
492 fldws -16(0,%r30),%fr10R
493 xmpyu %fr10L,%fr10R,%fr8
494 fstws %fr8R,-16(0,%r30)
495 ldw -16(0,%r30),%r29
496 stw %r22,-16(0,%r30)
497 fldws -16(0,%r30),%fr10R
498 stw %r21,-16(0,%r30)
499 copy %r29,%r19
500 xmpyu %fr10L,%fr10R,%fr8
501 fldws -16(0,%r30),%fr10L
502 stw %r21,-16(0,%r30)
503 fldws -16(0,%r30),%fr10R
504 fstws %fr8R,-16(0,%r30)
505 extru %r19,16,17,%r20
506 zdep %r19,14,15,%r19
507 ldw -16(0,%r30),%r29
508 xmpyu %fr10L,%fr10R,%fr9
509 addl %r29,%r19,%r22
510 stw %r22,-12(0,%r23)
511 fstws %fr9R,-16(0,%r30)
512 ldw -16(0,%r30),%r29
513 addl %r29,%r20,%r21
514 comclr,<<= %r19,%r22,0
515 addi 1,%r21,%r21
516 addib,= -1,%r24,L$0057
517 stw %r21,-8(0,%r23)
518 ldw 0(0,%r28),%r21
519 extru %r21,31,16,%r22
520 stw %r22,-16(0,%r30)
521 extru %r21,15,16,%r21
522 fldws -16(0,%r30),%fr10L
523 stw %r21,-16(0,%r30)
524 fldws -16(0,%r30),%fr10R
525 xmpyu %fr10L,%fr10R,%fr8
526 fstws %fr8R,-16(0,%r30)
527 ldw -16(0,%r30),%r29
528 stw %r22,-16(0,%r30)
529 fldws -16(0,%r30),%fr10R
530 stw %r21,-16(0,%r30)
531 copy %r29,%r19
532 xmpyu %fr10L,%fr10R,%fr8
533 fldws -16(0,%r30),%fr10L
534 stw %r21,-16(0,%r30)
535 fldws -16(0,%r30),%fr10R
536 fstws %fr8R,-16(0,%r30)
537 extru %r19,16,17,%r20
538 zdep %r19,14,15,%r19
539 ldw -16(0,%r30),%r29
540 xmpyu %fr10L,%fr10R,%fr9
541 addl %r29,%r19,%r22
542 stw %r22,-4(0,%r23)
543 fstws %fr9R,-16(0,%r30)
544 ldw -16(0,%r30),%r29
545 addl %r29,%r20,%r21
546 comclr,<<= %r19,%r22,0
547 addi 1,%r21,%r21
548 addib,= -1,%r24,L$0057
549 stw %r21,0(0,%r23)
550 ldo 16(%r28),%r28
551 ldo 16(%r25),%r25
552 ldo 32(%r23),%r23
553 bl L$0046,0
554 ldo 32(%r26),%r26
555L$0057
556 bv,n 0(%r2)
557 .EXIT
558 .PROCEND
559 .IMPORT BN_num_bits_word,CODE
560 .IMPORT fprintf,CODE
561 .IMPORT __iob,DATA
562 .SPACE $TEXT$
563 .SUBSPA $LIT$
564
565 .align 4
566L$C0000
567 .STRING "Division would overflow\x0a\x00"
568 .IMPORT abort,CODE
569 .SPACE $TEXT$
570 .SUBSPA $CODE$
571
572 .align 4
573 .EXPORT bn_div64,ENTRY,PRIV_LEV=3,ARGW0=GR,ARGW1=GR,ARGW2=GR,RTNVAL=GR
574bn_div64
575 .PROC
576 .CALLINFO FRAME=128,CALLS,SAVE_RP,ENTRY_GR=8
577 .ENTRY
578 stw %r2,-20(0,%r30)
579 stwm %r8,128(0,%r30)
580 stw %r7,-124(0,%r30)
581 stw %r4,-112(0,%r30)
582 stw %r3,-108(0,%r30)
583 copy %r26,%r3
584 copy %r25,%r4
585 stw %r6,-120(0,%r30)
586 ldi 0,%r7
587 stw %r5,-116(0,%r30)
588 movb,<> %r24,%r5,L$0059
589 ldi 2,%r6
590 bl L$0076,0
591 ldi -1,%r28
592L$0059
593 .CALL ARGW0=GR
594 bl BN_num_bits_word,%r2
595 copy %r5,%r26
596 ldi 32,%r19
597 comb,= %r19,%r28,L$0060
598 subi 31,%r28,%r19
599 mtsar %r19
600 zvdepi 1,32,%r19
601 comb,>>= %r19,%r3,L$0060
602 addil LR'__iob-$global$+32,%r27
603 ldo RR'__iob-$global$+32(%r1),%r26
604 ldil LR'L$C0000,%r25
605 .CALL ARGW0=GR,ARGW1=GR
606 bl fprintf,%r2
607 ldo RR'L$C0000(%r25),%r25
608 .CALL
609 bl abort,%r2
610 nop
611L$0060
612 comb,>> %r5,%r3,L$0061
613 subi 32,%r28,%r28
614 sub %r3,%r5,%r3
615L$0061
616 comib,= 0,%r28,L$0062
617 subi 31,%r28,%r19
618 mtsar %r19
619 zvdep %r5,32,%r5
620 zvdep %r3,32,%r21
621 subi 32,%r28,%r20
622 mtsar %r20
623 vshd 0,%r4,%r20
624 or %r21,%r20,%r3
625 mtsar %r19
626 zvdep %r4,32,%r4
627L$0062
628 extru %r5,15,16,%r23
629 extru %r5,31,16,%r28
630L$0063
631 extru %r3,15,16,%r19
632 comb,<> %r23,%r19,L$0066
633 copy %r3,%r26
634 bl L$0067,0
635 zdepi -1,31,16,%r29
636L$0066
637 .IMPORT $$divU,MILLICODE
638 bl $$divU,%r31
639 copy %r23,%r25
640L$0067
641 stw %r29,-16(0,%r30)
642 fldws -16(0,%r30),%fr10L
643 stw %r28,-16(0,%r30)
644 fldws -16(0,%r30),%fr10R
645 stw %r23,-16(0,%r30)
646 xmpyu %fr10L,%fr10R,%fr8
647 fldws -16(0,%r30),%fr10R
648 fstws %fr8R,-16(0,%r30)
649 xmpyu %fr10L,%fr10R,%fr9
650 ldw -16(0,%r30),%r8
651 fstws %fr9R,-16(0,%r30)
652 copy %r8,%r22
653 ldw -16(0,%r30),%r8
654 extru %r4,15,16,%r24
655 copy %r8,%r21
656L$0068
657 sub %r3,%r21,%r20
658 copy %r20,%r19
659 depi 0,31,16,%r19
660 comib,<> 0,%r19,L$0069
661 zdep %r20,15,16,%r19
662 addl %r19,%r24,%r19
663 comb,>>= %r19,%r22,L$0069
664 sub %r22,%r28,%r22
665 sub %r21,%r23,%r21
666 bl L$0068,0
667 ldo -1(%r29),%r29
668L$0069
669 stw %r29,-16(0,%r30)
670 fldws -16(0,%r30),%fr10L
671 stw %r28,-16(0,%r30)
672 fldws -16(0,%r30),%fr10R
673 xmpyu %fr10L,%fr10R,%fr8
674 fstws %fr8R,-16(0,%r30)
675 ldw -16(0,%r30),%r8
676 stw %r23,-16(0,%r30)
677 fldws -16(0,%r30),%fr10R
678 copy %r8,%r19
679 xmpyu %fr10L,%fr10R,%fr8
680 fstws %fr8R,-16(0,%r30)
681 extru %r19,15,16,%r20
682 ldw -16(0,%r30),%r8
683 zdep %r19,15,16,%r19
684 addl %r8,%r20,%r20
685 comclr,<<= %r19,%r4,0
686 addi 1,%r20,%r20
687 comb,<<= %r20,%r3,L$0074
688 sub %r4,%r19,%r4
689 addl %r3,%r5,%r3
690 ldo -1(%r29),%r29
691L$0074
692 addib,= -1,%r6,L$0064
693 sub %r3,%r20,%r3
694 zdep %r29,15,16,%r7
695 shd %r3,%r4,16,%r3
696 bl L$0063,0
697 zdep %r4,15,16,%r4
698L$0064
699 or %r7,%r29,%r28
700L$0076
701 ldw -148(0,%r30),%r2
702 ldw -124(0,%r30),%r7
703 ldw -120(0,%r30),%r6
704 ldw -116(0,%r30),%r5
705 ldw -112(0,%r30),%r4
706 ldw -108(0,%r30),%r3
707 bv 0(%r2)
708 ldwm -128(0,%r30),%r8
709 .EXIT
710 .PROCEND
diff --git a/src/lib/libcrypto/bn/asm/r3000.s b/src/lib/libcrypto/bn/asm/r3000.s
new file mode 100644
index 0000000000..e95269afa3
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/r3000.s
@@ -0,0 +1,646 @@
1 .file 1 "../bn_mulw.c"
2 .set nobopt
3 .option pic2
4
5 # GNU C 2.6.3 [AL 1.1, MM 40] SGI running IRIX 5.0 compiled by GNU C
6
7 # Cc1 defaults:
8 # -mabicalls
9
10 # Cc1 arguments (-G value = 0, Cpu = 3000, ISA = 1):
11 # -quiet -dumpbase -O2 -o
12
13gcc2_compiled.:
14__gnu_compiled_c:
15 .rdata
16
17 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
18 .byte 0x6e,0x3a,0x20,0x31,0x2e,0x34,0x39,0x20
19 .byte 0x24,0x0
20
21 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
22 .byte 0x6e,0x3a,0x20,0x31,0x2e,0x33,0x34,0x20
23 .byte 0x24,0x0
24
25 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
26 .byte 0x6e,0x3a,0x20,0x31,0x2e,0x35,0x20,0x24
27 .byte 0x0
28
29 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
30 .byte 0x6e,0x3a,0x20,0x31,0x2e,0x38,0x20,0x24
31 .byte 0x0
32
33 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
34 .byte 0x6e,0x3a,0x20,0x31,0x2e,0x32,0x33,0x20
35 .byte 0x24,0x0
36
37 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
38 .byte 0x6e,0x3a,0x20,0x31,0x2e,0x37,0x38,0x20
39 .byte 0x24,0x0
40
41 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
42 .byte 0x6e,0x3a,0x20,0x33,0x2e,0x37,0x30,0x20
43 .byte 0x24,0x0
44
45 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
46 .byte 0x6e,0x3a,0x20,0x31,0x2e,0x32,0x20,0x24
47 .byte 0x0
48
49 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
50 .byte 0x6e,0x3a,0x20,0x31,0x2e,0x34,0x20,0x24
51 .byte 0x0
52
53 .byte 0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f
54 .byte 0x6e,0x3a,0x20,0x31,0x2e,0x38,0x20,0x24
55 .byte 0x0
56 .text
57 .align 2
58 .globl bn_mul_add_words
59 .ent bn_mul_add_words
60bn_mul_add_words:
61 .frame $sp,0,$31 # vars= 0, regs= 0/0, args= 0, extra= 0
62 .mask 0x00000000,0
63 .fmask 0x00000000,0
64 .set noreorder
65 .cpload $25
66 .set reorder
67 move $12,$4
68 move $14,$5
69 move $9,$6
70 move $13,$7
71 move $8,$0
72 addu $10,$12,12
73 addu $11,$14,12
74$L2:
75 lw $6,0($14)
76 #nop
77 multu $13,$6
78 mfhi $6
79 mflo $7
80 #nop
81 move $5,$8
82 move $4,$0
83 lw $3,0($12)
84 addu $9,$9,-1
85 move $2,$0
86 addu $7,$7,$3
87 sltu $8,$7,$3
88 addu $6,$6,$2
89 addu $6,$6,$8
90 addu $7,$7,$5
91 sltu $2,$7,$5
92 addu $6,$6,$4
93 addu $6,$6,$2
94 srl $3,$6,0
95 move $2,$0
96 move $8,$3
97 .set noreorder
98 .set nomacro
99 beq $9,$0,$L3
100 sw $7,0($12)
101 .set macro
102 .set reorder
103
104 lw $6,-8($11)
105 #nop
106 multu $13,$6
107 mfhi $6
108 mflo $7
109 #nop
110 move $5,$8
111 move $4,$0
112 lw $3,-8($10)
113 addu $9,$9,-1
114 move $2,$0
115 addu $7,$7,$3
116 sltu $8,$7,$3
117 addu $6,$6,$2
118 addu $6,$6,$8
119 addu $7,$7,$5
120 sltu $2,$7,$5
121 addu $6,$6,$4
122 addu $6,$6,$2
123 srl $3,$6,0
124 move $2,$0
125 move $8,$3
126 .set noreorder
127 .set nomacro
128 beq $9,$0,$L3
129 sw $7,-8($10)
130 .set macro
131 .set reorder
132
133 lw $6,-4($11)
134 #nop
135 multu $13,$6
136 mfhi $6
137 mflo $7
138 #nop
139 move $5,$8
140 move $4,$0
141 lw $3,-4($10)
142 addu $9,$9,-1
143 move $2,$0
144 addu $7,$7,$3
145 sltu $8,$7,$3
146 addu $6,$6,$2
147 addu $6,$6,$8
148 addu $7,$7,$5
149 sltu $2,$7,$5
150 addu $6,$6,$4
151 addu $6,$6,$2
152 srl $3,$6,0
153 move $2,$0
154 move $8,$3
155 .set noreorder
156 .set nomacro
157 beq $9,$0,$L3
158 sw $7,-4($10)
159 .set macro
160 .set reorder
161
162 lw $6,0($11)
163 #nop
164 multu $13,$6
165 mfhi $6
166 mflo $7
167 #nop
168 move $5,$8
169 move $4,$0
170 lw $3,0($10)
171 addu $9,$9,-1
172 move $2,$0
173 addu $7,$7,$3
174 sltu $8,$7,$3
175 addu $6,$6,$2
176 addu $6,$6,$8
177 addu $7,$7,$5
178 sltu $2,$7,$5
179 addu $6,$6,$4
180 addu $6,$6,$2
181 srl $3,$6,0
182 move $2,$0
183 move $8,$3
184 .set noreorder
185 .set nomacro
186 beq $9,$0,$L3
187 sw $7,0($10)
188 .set macro
189 .set reorder
190
191 addu $11,$11,16
192 addu $14,$14,16
193 addu $10,$10,16
194 .set noreorder
195 .set nomacro
196 j $L2
197 addu $12,$12,16
198 .set macro
199 .set reorder
200
201$L3:
202 .set noreorder
203 .set nomacro
204 j $31
205 move $2,$8
206 .set macro
207 .set reorder
208
209 .end bn_mul_add_words
210 .align 2
211 .globl bn_mul_words
212 .ent bn_mul_words
213bn_mul_words:
214 .frame $sp,0,$31 # vars= 0, regs= 0/0, args= 0, extra= 0
215 .mask 0x00000000,0
216 .fmask 0x00000000,0
217 .set noreorder
218 .cpload $25
219 .set reorder
220 move $11,$4
221 move $12,$5
222 move $8,$6
223 move $6,$0
224 addu $10,$11,12
225 addu $9,$12,12
226$L10:
227 lw $4,0($12)
228 #nop
229 multu $7,$4
230 mfhi $4
231 mflo $5
232 #nop
233 move $3,$6
234 move $2,$0
235 addu $8,$8,-1
236 addu $5,$5,$3
237 sltu $6,$5,$3
238 addu $4,$4,$2
239 addu $4,$4,$6
240 srl $3,$4,0
241 move $2,$0
242 move $6,$3
243 .set noreorder
244 .set nomacro
245 beq $8,$0,$L11
246 sw $5,0($11)
247 .set macro
248 .set reorder
249
250 lw $4,-8($9)
251 #nop
252 multu $7,$4
253 mfhi $4
254 mflo $5
255 #nop
256 move $3,$6
257 move $2,$0
258 addu $8,$8,-1
259 addu $5,$5,$3
260 sltu $6,$5,$3
261 addu $4,$4,$2
262 addu $4,$4,$6
263 srl $3,$4,0
264 move $2,$0
265 move $6,$3
266 .set noreorder
267 .set nomacro
268 beq $8,$0,$L11
269 sw $5,-8($10)
270 .set macro
271 .set reorder
272
273 lw $4,-4($9)
274 #nop
275 multu $7,$4
276 mfhi $4
277 mflo $5
278 #nop
279 move $3,$6
280 move $2,$0
281 addu $8,$8,-1
282 addu $5,$5,$3
283 sltu $6,$5,$3
284 addu $4,$4,$2
285 addu $4,$4,$6
286 srl $3,$4,0
287 move $2,$0
288 move $6,$3
289 .set noreorder
290 .set nomacro
291 beq $8,$0,$L11
292 sw $5,-4($10)
293 .set macro
294 .set reorder
295
296 lw $4,0($9)
297 #nop
298 multu $7,$4
299 mfhi $4
300 mflo $5
301 #nop
302 move $3,$6
303 move $2,$0
304 addu $8,$8,-1
305 addu $5,$5,$3
306 sltu $6,$5,$3
307 addu $4,$4,$2
308 addu $4,$4,$6
309 srl $3,$4,0
310 move $2,$0
311 move $6,$3
312 .set noreorder
313 .set nomacro
314 beq $8,$0,$L11
315 sw $5,0($10)
316 .set macro
317 .set reorder
318
319 addu $9,$9,16
320 addu $12,$12,16
321 addu $10,$10,16
322 .set noreorder
323 .set nomacro
324 j $L10
325 addu $11,$11,16
326 .set macro
327 .set reorder
328
329$L11:
330 .set noreorder
331 .set nomacro
332 j $31
333 move $2,$6
334 .set macro
335 .set reorder
336
337 .end bn_mul_words
338 .align 2
339 .globl bn_sqr_words
340 .ent bn_sqr_words
341bn_sqr_words:
342 .frame $sp,0,$31 # vars= 0, regs= 0/0, args= 0, extra= 0
343 .mask 0x00000000,0
344 .fmask 0x00000000,0
345 .set noreorder
346 .cpload $25
347 .set reorder
348 move $9,$4
349 addu $7,$9,28
350 addu $8,$5,12
351$L18:
352 lw $2,0($5)
353 #nop
354 multu $2,$2
355 mfhi $2
356 mflo $3
357 #nop
358 addu $6,$6,-1
359 sw $3,0($9)
360 srl $3,$2,0
361 move $2,$0
362 .set noreorder
363 .set nomacro
364 beq $6,$0,$L19
365 sw $3,-24($7)
366 .set macro
367 .set reorder
368
369 lw $2,-8($8)
370 #nop
371 multu $2,$2
372 mfhi $2
373 mflo $3
374 #nop
375 addu $6,$6,-1
376 sw $3,-20($7)
377 srl $3,$2,0
378 move $2,$0
379 .set noreorder
380 .set nomacro
381 beq $6,$0,$L19
382 sw $3,-16($7)
383 .set macro
384 .set reorder
385
386 lw $2,-4($8)
387 #nop
388 multu $2,$2
389 mfhi $2
390 mflo $3
391 #nop
392 addu $6,$6,-1
393 sw $3,-12($7)
394 srl $3,$2,0
395 move $2,$0
396 .set noreorder
397 .set nomacro
398 beq $6,$0,$L19
399 sw $3,-8($7)
400 .set macro
401 .set reorder
402
403 lw $2,0($8)
404 #nop
405 multu $2,$2
406 mfhi $2
407 mflo $3
408 #nop
409 addu $6,$6,-1
410 sw $3,-4($7)
411 srl $3,$2,0
412 move $2,$0
413 .set noreorder
414 .set nomacro
415 beq $6,$0,$L19
416 sw $3,0($7)
417 .set macro
418 .set reorder
419
420 addu $8,$8,16
421 addu $5,$5,16
422 addu $7,$7,32
423 .set noreorder
424 .set nomacro
425 j $L18
426 addu $9,$9,32
427 .set macro
428 .set reorder
429
430$L19:
431 j $31
432 .end bn_sqr_words
433 .rdata
434 .align 2
435$LC0:
436
437 .byte 0x44,0x69,0x76,0x69,0x73,0x69,0x6f,0x6e
438 .byte 0x20,0x77,0x6f,0x75,0x6c,0x64,0x20,0x6f
439 .byte 0x76,0x65,0x72,0x66,0x6c,0x6f,0x77,0xa
440 .byte 0x0
441 .text
442 .align 2
443 .globl bn_div64
444 .ent bn_div64
445bn_div64:
446 .frame $sp,56,$31 # vars= 0, regs= 7/0, args= 16, extra= 8
447 .mask 0x901f0000,-8
448 .fmask 0x00000000,0
449 .set noreorder
450 .cpload $25
451 .set reorder
452 subu $sp,$sp,56
453 .cprestore 16
454 sw $16,24($sp)
455 move $16,$4
456 sw $17,28($sp)
457 move $17,$5
458 sw $18,32($sp)
459 move $18,$6
460 sw $20,40($sp)
461 move $20,$0
462 sw $19,36($sp)
463 li $19,0x00000002 # 2
464 sw $31,48($sp)
465 .set noreorder
466 .set nomacro
467 bne $18,$0,$L26
468 sw $28,44($sp)
469 .set macro
470 .set reorder
471
472 .set noreorder
473 .set nomacro
474 j $L43
475 li $2,-1 # 0xffffffff
476 .set macro
477 .set reorder
478
479$L26:
480 move $4,$18
481 jal BN_num_bits_word
482 move $4,$2
483 li $2,0x00000020 # 32
484 .set noreorder
485 .set nomacro
486 beq $4,$2,$L27
487 li $2,0x00000001 # 1
488 .set macro
489 .set reorder
490
491 sll $2,$2,$4
492 sltu $2,$2,$16
493 .set noreorder
494 .set nomacro
495 beq $2,$0,$L44
496 li $5,0x00000020 # 32
497 .set macro
498 .set reorder
499
500 la $4,__iob+32
501 la $5,$LC0
502 jal fprintf
503 jal abort
504$L27:
505 li $5,0x00000020 # 32
506$L44:
507 sltu $2,$16,$18
508 .set noreorder
509 .set nomacro
510 bne $2,$0,$L28
511 subu $4,$5,$4
512 .set macro
513 .set reorder
514
515 subu $16,$16,$18
516$L28:
517 .set noreorder
518 .set nomacro
519 beq $4,$0,$L29
520 li $10,-65536 # 0xffff0000
521 .set macro
522 .set reorder
523
524 sll $18,$18,$4
525 sll $3,$16,$4
526 subu $2,$5,$4
527 srl $2,$17,$2
528 or $16,$3,$2
529 sll $17,$17,$4
530$L29:
531 srl $7,$18,16
532 andi $9,$18,0xffff
533$L30:
534 srl $2,$16,16
535 .set noreorder
536 .set nomacro
537 beq $2,$7,$L34
538 li $6,0x0000ffff # 65535
539 .set macro
540 .set reorder
541
542 divu $6,$16,$7
543$L34:
544 mult $6,$9
545 mflo $5
546 #nop
547 #nop
548 mult $6,$7
549 and $2,$17,$10
550 srl $8,$2,16
551 mflo $4
552$L35:
553 subu $3,$16,$4
554 and $2,$3,$10
555 .set noreorder
556 .set nomacro
557 bne $2,$0,$L36
558 sll $2,$3,16
559 .set macro
560 .set reorder
561
562 addu $2,$2,$8
563 sltu $2,$2,$5
564 .set noreorder
565 .set nomacro
566 beq $2,$0,$L36
567 subu $5,$5,$9
568 .set macro
569 .set reorder
570
571 subu $4,$4,$7
572 .set noreorder
573 .set nomacro
574 j $L35
575 addu $6,$6,-1
576 .set macro
577 .set reorder
578
579$L36:
580 mult $6,$7
581 mflo $5
582 #nop
583 #nop
584 mult $6,$9
585 mflo $4
586 #nop
587 #nop
588 srl $3,$4,16
589 sll $2,$4,16
590 and $4,$2,$10
591 sltu $2,$17,$4
592 .set noreorder
593 .set nomacro
594 beq $2,$0,$L40
595 addu $5,$5,$3
596 .set macro
597 .set reorder
598
599 addu $5,$5,1
600$L40:
601 sltu $2,$16,$5
602 .set noreorder
603 .set nomacro
604 beq $2,$0,$L41
605 subu $17,$17,$4
606 .set macro
607 .set reorder
608
609 addu $16,$16,$18
610 addu $6,$6,-1
611$L41:
612 addu $19,$19,-1
613 .set noreorder
614 .set nomacro
615 beq $19,$0,$L31
616 subu $16,$16,$5
617 .set macro
618 .set reorder
619
620 sll $20,$6,16
621 sll $3,$16,16
622 srl $2,$17,16
623 or $16,$3,$2
624 .set noreorder
625 .set nomacro
626 j $L30
627 sll $17,$17,16
628 .set macro
629 .set reorder
630
631$L31:
632 or $2,$20,$6
633$L43:
634 lw $31,48($sp)
635 lw $20,40($sp)
636 lw $19,36($sp)
637 lw $18,32($sp)
638 lw $17,28($sp)
639 lw $16,24($sp)
640 addu $sp,$sp,56
641 j $31
642 .end bn_div64
643
644 .globl abort .text
645 .globl fprintf .text
646 .globl BN_num_bits_word .text
generated by cgit v1.2.3-55-g6feb (git 2.39.1) at 2025-08-04 19:32:05 +0000