1 1.1 christos #include "sparc_arch.h" 2 1.1 christos 3 1.1 christos #ifdef __arch64__ 4 1.1 christos .register %g2,#scratch 5 1.1 christos .register %g3,#scratch 6 1.1 christos #endif 7 1.1 christos 8 1.1 christos .section ".text",#alloc,#execinstr 9 1.1 christos 10 1.1 christos #ifdef __PIC__ 11 1.1 christos SPARC_PIC_THUNK(%g1) 12 1.1 christos #endif 13 1.1 christos .globl bn_mul_mont_t4_8 14 1.1 christos .align 32 15 1.1 christos bn_mul_mont_t4_8: 16 1.1 christos #ifdef __arch64__ 17 1.1 christos mov 0,%g5 18 1.1 christos mov -128,%g4 19 1.1 christos #elif defined(SPARCV9_64BIT_STACK) 20 1.1 christos SPARC_LOAD_ADDRESS_LEAF(OPENSSL_sparcv9cap_P,%g1,%g5) 21 1.1 christos ld [%g1+0],%g1 ! OPENSSL_sparcv9_P[0] 22 1.1 christos mov -2047,%g4 23 1.1 christos and %g1,SPARCV9_64BIT_STACK,%g1 24 1.1 christos movrz %g1,0,%g4 25 1.1 christos mov -1,%g5 26 1.1 christos add %g4,-128,%g4 27 1.1 christos #else 28 1.1 christos mov -1,%g5 29 1.1 christos mov -128,%g4 30 1.1 christos #endif 31 1.1 christos sllx %g5,32,%g5 32 1.1 christos save %sp,%g4,%sp 33 1.1 christos #ifndef __arch64__ 34 1.1 christos save %sp,-128,%sp ! warm it up 35 1.1 christos save %sp,-128,%sp 36 1.1 christos save %sp,-128,%sp 37 1.1 christos save %sp,-128,%sp 38 1.1 christos save %sp,-128,%sp 39 1.1 christos save %sp,-128,%sp 40 1.1 christos restore 41 1.1 christos restore 42 1.1 christos restore 43 1.1 christos restore 44 1.1 christos restore 45 1.1 christos restore 46 1.1 christos #endif 47 1.1 christos and %sp,1,%g4 48 1.1 christos or %g5,%fp,%fp 49 1.1 christos or %g4,%g5,%g5 50 1.1 christos 51 1.1 christos ! copy arguments to global registers 52 1.1 christos mov %i0,%g1 53 1.1 christos mov %i1,%g2 54 1.1 christos mov %i2,%g3 55 1.1 christos mov %i3,%g4 56 1.1 christos ld [%i4+0],%f1 ! load *n0 57 1.1 christos ld [%i4+4],%f0 58 1.1 christos .word 0xbbb00f00 !fsrc2 %f0,%f0,%f60 59 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 60 1.1 christos ld [%g2+0*8+0],%l1 61 1.1 christos ld [%g2+0*8+4],%l0 62 1.1 christos sllx %l0,32,%l0 63 1.1 christos or %l1,%l0,%l0 64 1.1 christos ld [%g2+1*8+0],%l2 65 1.1 christos ld [%g2+1*8+4],%l1 66 1.1 christos sllx %l1,32,%l1 67 1.1 christos or %l2,%l1,%l1 68 1.1 christos ld [%g2+2*8+0],%l3 69 1.1 christos ld [%g2+2*8+4],%l2 70 1.1 christos sllx %l2,32,%l2 71 1.1 christos or %l3,%l2,%l2 72 1.1 christos ld [%g2+3*8+0],%l4 73 1.1 christos ld [%g2+3*8+4],%l3 74 1.1 christos sllx %l3,32,%l3 75 1.1 christos or %l4,%l3,%l3 76 1.1 christos ld [%g2+4*8+0],%l5 77 1.1 christos ld [%g2+4*8+4],%l4 78 1.1 christos sllx %l4,32,%l4 79 1.1 christos or %l5,%l4,%l4 80 1.1 christos ld [%g2+5*8+0],%l6 81 1.1 christos ld [%g2+5*8+4],%l5 82 1.1 christos sllx %l5,32,%l5 83 1.1 christos or %l6,%l5,%l5 84 1.1 christos ld [%g2+6*8+0],%l7 85 1.1 christos ld [%g2+6*8+4],%l6 86 1.1 christos sllx %l6,32,%l6 87 1.1 christos or %l7,%l6,%l6 88 1.1 christos ld [%g2+7*8+0],%o0 89 1.1 christos ld [%g2+7*8+4],%l7 90 1.1 christos sllx %l7,32,%l7 91 1.1 christos or %o0,%l7,%l7 92 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 93 1.1 christos ld [%g4+0*8+0],%l1 94 1.1 christos ld [%g4+0*8+4],%l0 95 1.1 christos sllx %l0,32,%l0 96 1.1 christos or %l1,%l0,%l0 97 1.1 christos ld [%g4+1*8+0],%l2 98 1.1 christos ld [%g4+1*8+4],%l1 99 1.1 christos sllx %l1,32,%l1 100 1.1 christos or %l2,%l1,%l1 101 1.1 christos ld [%g4+2*8+0],%l3 102 1.1 christos ld [%g4+2*8+4],%l2 103 1.1 christos sllx %l2,32,%l2 104 1.1 christos or %l3,%l2,%l2 105 1.1 christos ld [%g4+3*8+0],%l4 106 1.1 christos ld [%g4+3*8+4],%l3 107 1.1 christos sllx %l3,32,%l3 108 1.1 christos or %l4,%l3,%l3 109 1.1 christos ld [%g4+4*8+0],%l5 110 1.1 christos ld [%g4+4*8+4],%l4 111 1.1 christos sllx %l4,32,%l4 112 1.1 christos or %l5,%l4,%l4 113 1.1 christos ld [%g4+5*8+0],%l6 114 1.1 christos ld [%g4+5*8+4],%l5 115 1.1 christos sllx %l5,32,%l5 116 1.1 christos or %l6,%l5,%l5 117 1.1 christos ld [%g4+6*8+0],%l7 118 1.1 christos ld [%g4+6*8+4],%l6 119 1.1 christos sllx %l6,32,%l6 120 1.1 christos or %l7,%l6,%l6 121 1.1 christos ld [%g4+7*8+0],%o0 122 1.1 christos ld [%g4+7*8+4],%l7 123 1.1 christos sllx %l7,32,%l7 124 1.1 christos or %o0,%l7,%l7 125 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 126 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 127 1.1 christos cmp %g2,%g3 128 1.1 christos be SIZE_T_CC,.Lmsquare_8 129 1.1 christos nop 130 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 131 1.1 christos ld [%g3+0*8+0],%i1 132 1.1 christos ld [%g3+0*8+4],%i0 133 1.1 christos sllx %i0,32,%i0 134 1.1 christos or %i1,%i0,%i0 135 1.1 christos ld [%g3+1*8+0],%i2 136 1.1 christos ld [%g3+1*8+4],%i1 137 1.1 christos sllx %i1,32,%i1 138 1.1 christos or %i2,%i1,%i1 139 1.1 christos ld [%g3+2*8+0],%i3 140 1.1 christos ld [%g3+2*8+4],%i2 141 1.1 christos sllx %i2,32,%i2 142 1.1 christos or %i3,%i2,%i2 143 1.1 christos ld [%g3+3*8+0],%i4 144 1.1 christos ld [%g3+3*8+4],%i3 145 1.1 christos sllx %i3,32,%i3 146 1.1 christos or %i4,%i3,%i3 147 1.1 christos ld [%g3+4*8+0],%i5 148 1.1 christos ld [%g3+4*8+4],%i4 149 1.1 christos sllx %i4,32,%i4 150 1.1 christos or %i5,%i4,%i4 151 1.1 christos ld [%g3+5*8+0],%l0 152 1.1 christos ld [%g3+5*8+4],%i5 153 1.1 christos sllx %i5,32,%i5 154 1.1 christos or %l0,%i5,%i5 155 1.1 christos ld [%g3+6*8+0],%l1 156 1.1 christos ld [%g3+6*8+4],%l0 157 1.1 christos sllx %l0,32,%l0 158 1.1 christos or %l1,%l0,%l0 159 1.1 christos ld [%g3+7*8+0],%l2 160 1.1 christos ld [%g3+7*8+4],%l1 161 1.1 christos sllx %l1,32,%l1 162 1.1 christos or %l2,%l1,%l1 163 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 164 1.1 christos .word 0x81b02920+8-1 ! montmul 8-1 165 1.1 christos .Lmresume_8: 166 1.1 christos fbu,pn %fcc3,.Lmabort_8 167 1.1 christos #ifndef __arch64__ 168 1.1 christos and %fp,%g5,%g5 169 1.1 christos brz,pn %g5,.Lmabort_8 170 1.1 christos #endif 171 1.1 christos nop 172 1.1 christos #ifdef __arch64__ 173 1.1 christos restore 174 1.1 christos restore 175 1.1 christos restore 176 1.1 christos restore 177 1.1 christos restore 178 1.1 christos #else 179 1.1 christos restore; and %fp,%g5,%g5 180 1.1 christos restore; and %fp,%g5,%g5 181 1.1 christos restore; and %fp,%g5,%g5 182 1.1 christos restore; and %fp,%g5,%g5 183 1.1 christos brz,pn %g5,.Lmabort1_8 184 1.1 christos restore 185 1.1 christos #endif 186 1.1 christos .word 0x81b02310 !movxtod %l0,%f0 187 1.1 christos .word 0x85b02311 !movxtod %l1,%f2 188 1.1 christos .word 0x89b02312 !movxtod %l2,%f4 189 1.1 christos .word 0x8db02313 !movxtod %l3,%f6 190 1.1 christos .word 0x91b02314 !movxtod %l4,%f8 191 1.1 christos .word 0x95b02315 !movxtod %l5,%f10 192 1.1 christos .word 0x99b02316 !movxtod %l6,%f12 193 1.1 christos .word 0x9db02317 !movxtod %l7,%f14 194 1.1 christos #ifdef __arch64__ 195 1.1 christos restore 196 1.1 christos #else 197 1.1 christos and %fp,%g5,%g5 198 1.1 christos restore 199 1.1 christos and %g5,1,%o7 200 1.1 christos and %fp,%g5,%g5 201 1.1 christos srl %fp,0,%fp ! just in case? 202 1.1 christos or %o7,%g5,%g5 203 1.1 christos brz,a,pn %g5,.Lmdone_8 204 1.1 christos mov 0,%i0 ! return failure 205 1.1 christos #endif 206 1.1 christos st %f1,[%g1+0*8+0] 207 1.1 christos st %f0,[%g1+0*8+4] 208 1.1 christos st %f3,[%g1+1*8+0] 209 1.1 christos st %f2,[%g1+1*8+4] 210 1.1 christos st %f5,[%g1+2*8+0] 211 1.1 christos st %f4,[%g1+2*8+4] 212 1.1 christos st %f7,[%g1+3*8+0] 213 1.1 christos st %f6,[%g1+3*8+4] 214 1.1 christos st %f9,[%g1+4*8+0] 215 1.1 christos st %f8,[%g1+4*8+4] 216 1.1 christos st %f11,[%g1+5*8+0] 217 1.1 christos st %f10,[%g1+5*8+4] 218 1.1 christos st %f13,[%g1+6*8+0] 219 1.1 christos st %f12,[%g1+6*8+4] 220 1.1 christos st %f15,[%g1+7*8+0] 221 1.1 christos st %f14,[%g1+7*8+4] 222 1.1 christos mov 1,%i0 ! return success 223 1.1 christos .Lmdone_8: 224 1.1 christos ret 225 1.1 christos restore 226 1.1 christos 227 1.1 christos .Lmabort_8: 228 1.1 christos restore 229 1.1 christos restore 230 1.1 christos restore 231 1.1 christos restore 232 1.1 christos restore 233 1.1 christos .Lmabort1_8: 234 1.1 christos restore 235 1.1 christos 236 1.1 christos mov 0,%i0 ! return failure 237 1.1 christos ret 238 1.1 christos restore 239 1.1 christos 240 1.1 christos .align 32 241 1.1 christos .Lmsquare_8: 242 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 243 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 244 1.1 christos .word 0x81b02940+8-1 ! montsqr 8-1 245 1.1 christos ba .Lmresume_8 246 1.1 christos nop 247 1.1 christos .type bn_mul_mont_t4_8, #function 248 1.1 christos .size bn_mul_mont_t4_8, .-bn_mul_mont_t4_8 249 1.1 christos .globl bn_mul_mont_t4_16 250 1.1 christos .align 32 251 1.1 christos bn_mul_mont_t4_16: 252 1.1 christos #ifdef __arch64__ 253 1.1 christos mov 0,%g5 254 1.1 christos mov -128,%g4 255 1.1 christos #elif defined(SPARCV9_64BIT_STACK) 256 1.1 christos SPARC_LOAD_ADDRESS_LEAF(OPENSSL_sparcv9cap_P,%g1,%g5) 257 1.1 christos ld [%g1+0],%g1 ! OPENSSL_sparcv9_P[0] 258 1.1 christos mov -2047,%g4 259 1.1 christos and %g1,SPARCV9_64BIT_STACK,%g1 260 1.1 christos movrz %g1,0,%g4 261 1.1 christos mov -1,%g5 262 1.1 christos add %g4,-128,%g4 263 1.1 christos #else 264 1.1 christos mov -1,%g5 265 1.1 christos mov -128,%g4 266 1.1 christos #endif 267 1.1 christos sllx %g5,32,%g5 268 1.1 christos save %sp,%g4,%sp 269 1.1 christos #ifndef __arch64__ 270 1.1 christos save %sp,-128,%sp ! warm it up 271 1.1 christos save %sp,-128,%sp 272 1.1 christos save %sp,-128,%sp 273 1.1 christos save %sp,-128,%sp 274 1.1 christos save %sp,-128,%sp 275 1.1 christos save %sp,-128,%sp 276 1.1 christos restore 277 1.1 christos restore 278 1.1 christos restore 279 1.1 christos restore 280 1.1 christos restore 281 1.1 christos restore 282 1.1 christos #endif 283 1.1 christos and %sp,1,%g4 284 1.1 christos or %g5,%fp,%fp 285 1.1 christos or %g4,%g5,%g5 286 1.1 christos 287 1.1 christos ! copy arguments to global registers 288 1.1 christos mov %i0,%g1 289 1.1 christos mov %i1,%g2 290 1.1 christos mov %i2,%g3 291 1.1 christos mov %i3,%g4 292 1.1 christos ld [%i4+0],%f1 ! load *n0 293 1.1 christos ld [%i4+4],%f0 294 1.1 christos .word 0xbbb00f00 !fsrc2 %f0,%f0,%f60 295 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 296 1.1 christos ld [%g2+0*8+0],%l1 297 1.1 christos ld [%g2+0*8+4],%l0 298 1.1 christos sllx %l0,32,%l0 299 1.1 christos or %l1,%l0,%l0 300 1.1 christos ld [%g2+1*8+0],%l2 301 1.1 christos ld [%g2+1*8+4],%l1 302 1.1 christos sllx %l1,32,%l1 303 1.1 christos or %l2,%l1,%l1 304 1.1 christos ld [%g2+2*8+0],%l3 305 1.1 christos ld [%g2+2*8+4],%l2 306 1.1 christos sllx %l2,32,%l2 307 1.1 christos or %l3,%l2,%l2 308 1.1 christos ld [%g2+3*8+0],%l4 309 1.1 christos ld [%g2+3*8+4],%l3 310 1.1 christos sllx %l3,32,%l3 311 1.1 christos or %l4,%l3,%l3 312 1.1 christos ld [%g2+4*8+0],%l5 313 1.1 christos ld [%g2+4*8+4],%l4 314 1.1 christos sllx %l4,32,%l4 315 1.1 christos or %l5,%l4,%l4 316 1.1 christos ld [%g2+5*8+0],%l6 317 1.1 christos ld [%g2+5*8+4],%l5 318 1.1 christos sllx %l5,32,%l5 319 1.1 christos or %l6,%l5,%l5 320 1.1 christos ld [%g2+6*8+0],%l7 321 1.1 christos ld [%g2+6*8+4],%l6 322 1.1 christos sllx %l6,32,%l6 323 1.1 christos or %l7,%l6,%l6 324 1.1 christos ld [%g2+7*8+0],%o0 325 1.1 christos ld [%g2+7*8+4],%l7 326 1.1 christos sllx %l7,32,%l7 327 1.1 christos or %o0,%l7,%l7 328 1.1 christos ld [%g2+8*8+0],%o1 329 1.1 christos ld [%g2+8*8+4],%o0 330 1.1 christos sllx %o0,32,%o0 331 1.1 christos or %o1,%o0,%o0 332 1.1 christos ld [%g2+9*8+0],%o2 333 1.1 christos ld [%g2+9*8+4],%o1 334 1.1 christos sllx %o1,32,%o1 335 1.1 christos or %o2,%o1,%o1 336 1.1 christos ld [%g2+10*8+0],%o3 337 1.1 christos ld [%g2+10*8+4],%o2 338 1.1 christos sllx %o2,32,%o2 339 1.1 christos or %o3,%o2,%o2 340 1.1 christos ld [%g2+11*8+0],%o4 341 1.1 christos ld [%g2+11*8+4],%o3 342 1.1 christos sllx %o3,32,%o3 343 1.1 christos or %o4,%o3,%o3 344 1.1 christos ld [%g2+12*8+0],%o5 345 1.1 christos ld [%g2+12*8+4],%o4 346 1.1 christos sllx %o4,32,%o4 347 1.1 christos or %o5,%o4,%o4 348 1.1 christos ld [%g2+13*8+0],%o7 349 1.1 christos ld [%g2+13*8+4],%o5 350 1.1 christos sllx %o5,32,%o5 351 1.1 christos or %o7,%o5,%o5 352 1.1 christos ld [%g2+14*8+0],%f5 353 1.1 christos ld [%g2+14*8+4],%f4 354 1.1 christos .word 0xb1b00f04 !fsrc2 %f0,%f4,%f24 355 1.1 christos ld [%g2+15*8+0],%f7 356 1.1 christos ld [%g2+15*8+4],%f6 357 1.1 christos .word 0xb5b00f06 !fsrc2 %f0,%f6,%f26 358 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 359 1.1 christos ld [%g4+0*8+0],%l1 360 1.1 christos ld [%g4+0*8+4],%l0 361 1.1 christos sllx %l0,32,%l0 362 1.1 christos or %l1,%l0,%l0 363 1.1 christos ld [%g4+1*8+0],%l2 364 1.1 christos ld [%g4+1*8+4],%l1 365 1.1 christos sllx %l1,32,%l1 366 1.1 christos or %l2,%l1,%l1 367 1.1 christos ld [%g4+2*8+0],%l3 368 1.1 christos ld [%g4+2*8+4],%l2 369 1.1 christos sllx %l2,32,%l2 370 1.1 christos or %l3,%l2,%l2 371 1.1 christos ld [%g4+3*8+0],%l4 372 1.1 christos ld [%g4+3*8+4],%l3 373 1.1 christos sllx %l3,32,%l3 374 1.1 christos or %l4,%l3,%l3 375 1.1 christos ld [%g4+4*8+0],%l5 376 1.1 christos ld [%g4+4*8+4],%l4 377 1.1 christos sllx %l4,32,%l4 378 1.1 christos or %l5,%l4,%l4 379 1.1 christos ld [%g4+5*8+0],%l6 380 1.1 christos ld [%g4+5*8+4],%l5 381 1.1 christos sllx %l5,32,%l5 382 1.1 christos or %l6,%l5,%l5 383 1.1 christos ld [%g4+6*8+0],%l7 384 1.1 christos ld [%g4+6*8+4],%l6 385 1.1 christos sllx %l6,32,%l6 386 1.1 christos or %l7,%l6,%l6 387 1.1 christos ld [%g4+7*8+0],%o0 388 1.1 christos ld [%g4+7*8+4],%l7 389 1.1 christos sllx %l7,32,%l7 390 1.1 christos or %o0,%l7,%l7 391 1.1 christos ld [%g4+8*8+0],%o1 392 1.1 christos ld [%g4+8*8+4],%o0 393 1.1 christos sllx %o0,32,%o0 394 1.1 christos or %o1,%o0,%o0 395 1.1 christos ld [%g4+9*8+0],%o2 396 1.1 christos ld [%g4+9*8+4],%o1 397 1.1 christos sllx %o1,32,%o1 398 1.1 christos or %o2,%o1,%o1 399 1.1 christos ld [%g4+10*8+0],%o3 400 1.1 christos ld [%g4+10*8+4],%o2 401 1.1 christos sllx %o2,32,%o2 402 1.1 christos or %o3,%o2,%o2 403 1.1 christos ld [%g4+11*8+0],%o4 404 1.1 christos ld [%g4+11*8+4],%o3 405 1.1 christos sllx %o3,32,%o3 406 1.1 christos or %o4,%o3,%o3 407 1.1 christos ld [%g4+12*8+0],%o5 408 1.1 christos ld [%g4+12*8+4],%o4 409 1.1 christos sllx %o4,32,%o4 410 1.1 christos or %o5,%o4,%o4 411 1.1 christos ld [%g4+13*8+0],%o7 412 1.1 christos ld [%g4+13*8+4],%o5 413 1.1 christos sllx %o5,32,%o5 414 1.1 christos or %o7,%o5,%o5 415 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 416 1.1 christos ld [%g4+14*8+0],%l1 417 1.1 christos ld [%g4+14*8+4],%l0 418 1.1 christos sllx %l0,32,%l0 419 1.1 christos or %l1,%l0,%l0 420 1.1 christos ld [%g4+15*8+0],%l2 421 1.1 christos ld [%g4+15*8+4],%l1 422 1.1 christos sllx %l1,32,%l1 423 1.1 christos or %l2,%l1,%l1 424 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 425 1.1 christos cmp %g2,%g3 426 1.1 christos be SIZE_T_CC,.Lmsquare_16 427 1.1 christos nop 428 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 429 1.1 christos ld [%g3+0*8+0],%i1 430 1.1 christos ld [%g3+0*8+4],%i0 431 1.1 christos sllx %i0,32,%i0 432 1.1 christos or %i1,%i0,%i0 433 1.1 christos ld [%g3+1*8+0],%i2 434 1.1 christos ld [%g3+1*8+4],%i1 435 1.1 christos sllx %i1,32,%i1 436 1.1 christos or %i2,%i1,%i1 437 1.1 christos ld [%g3+2*8+0],%i3 438 1.1 christos ld [%g3+2*8+4],%i2 439 1.1 christos sllx %i2,32,%i2 440 1.1 christos or %i3,%i2,%i2 441 1.1 christos ld [%g3+3*8+0],%i4 442 1.1 christos ld [%g3+3*8+4],%i3 443 1.1 christos sllx %i3,32,%i3 444 1.1 christos or %i4,%i3,%i3 445 1.1 christos ld [%g3+4*8+0],%i5 446 1.1 christos ld [%g3+4*8+4],%i4 447 1.1 christos sllx %i4,32,%i4 448 1.1 christos or %i5,%i4,%i4 449 1.1 christos ld [%g3+5*8+0],%l0 450 1.1 christos ld [%g3+5*8+4],%i5 451 1.1 christos sllx %i5,32,%i5 452 1.1 christos or %l0,%i5,%i5 453 1.1 christos ld [%g3+6*8+0],%l1 454 1.1 christos ld [%g3+6*8+4],%l0 455 1.1 christos sllx %l0,32,%l0 456 1.1 christos or %l1,%l0,%l0 457 1.1 christos ld [%g3+7*8+0],%l2 458 1.1 christos ld [%g3+7*8+4],%l1 459 1.1 christos sllx %l1,32,%l1 460 1.1 christos or %l2,%l1,%l1 461 1.1 christos ld [%g3+8*8+0],%l3 462 1.1 christos ld [%g3+8*8+4],%l2 463 1.1 christos sllx %l2,32,%l2 464 1.1 christos or %l3,%l2,%l2 465 1.1 christos ld [%g3+9*8+0],%l4 466 1.1 christos ld [%g3+9*8+4],%l3 467 1.1 christos sllx %l3,32,%l3 468 1.1 christos or %l4,%l3,%l3 469 1.1 christos ld [%g3+10*8+0],%l5 470 1.1 christos ld [%g3+10*8+4],%l4 471 1.1 christos sllx %l4,32,%l4 472 1.1 christos or %l5,%l4,%l4 473 1.1 christos ld [%g3+11*8+0],%l6 474 1.1 christos ld [%g3+11*8+4],%l5 475 1.1 christos sllx %l5,32,%l5 476 1.1 christos or %l6,%l5,%l5 477 1.1 christos ld [%g3+12*8+0],%l7 478 1.1 christos ld [%g3+12*8+4],%l6 479 1.1 christos sllx %l6,32,%l6 480 1.1 christos or %l7,%l6,%l6 481 1.1 christos ld [%g3+13*8+0],%o7 482 1.1 christos ld [%g3+13*8+4],%l7 483 1.1 christos sllx %l7,32,%l7 484 1.1 christos or %o7,%l7,%l7 485 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 486 1.1 christos ld [%g3+14*8+0],%i1 487 1.1 christos ld [%g3+14*8+4],%i0 488 1.1 christos sllx %i0,32,%i0 489 1.1 christos or %i1,%i0,%i0 490 1.1 christos ld [%g3+15*8+0],%o7 491 1.1 christos ld [%g3+15*8+4],%i1 492 1.1 christos sllx %i1,32,%i1 493 1.1 christos or %o7,%i1,%i1 494 1.1 christos .word 0x81b02920+16-1 ! montmul 16-1 495 1.1 christos .Lmresume_16: 496 1.1 christos fbu,pn %fcc3,.Lmabort_16 497 1.1 christos #ifndef __arch64__ 498 1.1 christos and %fp,%g5,%g5 499 1.1 christos brz,pn %g5,.Lmabort_16 500 1.1 christos #endif 501 1.1 christos nop 502 1.1 christos #ifdef __arch64__ 503 1.1 christos restore 504 1.1 christos restore 505 1.1 christos restore 506 1.1 christos restore 507 1.1 christos restore 508 1.1 christos #else 509 1.1 christos restore; and %fp,%g5,%g5 510 1.1 christos restore; and %fp,%g5,%g5 511 1.1 christos restore; and %fp,%g5,%g5 512 1.1 christos restore; and %fp,%g5,%g5 513 1.1 christos brz,pn %g5,.Lmabort1_16 514 1.1 christos restore 515 1.1 christos #endif 516 1.1 christos .word 0x81b02310 !movxtod %l0,%f0 517 1.1 christos .word 0x85b02311 !movxtod %l1,%f2 518 1.1 christos .word 0x89b02312 !movxtod %l2,%f4 519 1.1 christos .word 0x8db02313 !movxtod %l3,%f6 520 1.1 christos .word 0x91b02314 !movxtod %l4,%f8 521 1.1 christos .word 0x95b02315 !movxtod %l5,%f10 522 1.1 christos .word 0x99b02316 !movxtod %l6,%f12 523 1.1 christos .word 0x9db02317 !movxtod %l7,%f14 524 1.1 christos .word 0xa1b02308 !movxtod %o0,%f16 525 1.1 christos .word 0xa5b02309 !movxtod %o1,%f18 526 1.1 christos .word 0xa9b0230a !movxtod %o2,%f20 527 1.1 christos .word 0xadb0230b !movxtod %o3,%f22 528 1.1 christos .word 0xbbb0230c !movxtod %o4,%f60 529 1.1 christos .word 0xbfb0230d !movxtod %o5,%f62 530 1.1 christos #ifdef __arch64__ 531 1.1 christos restore 532 1.1 christos #else 533 1.1 christos and %fp,%g5,%g5 534 1.1 christos restore 535 1.1 christos and %g5,1,%o7 536 1.1 christos and %fp,%g5,%g5 537 1.1 christos srl %fp,0,%fp ! just in case? 538 1.1 christos or %o7,%g5,%g5 539 1.1 christos brz,a,pn %g5,.Lmdone_16 540 1.1 christos mov 0,%i0 ! return failure 541 1.1 christos #endif 542 1.1 christos st %f1,[%g1+0*8+0] 543 1.1 christos st %f0,[%g1+0*8+4] 544 1.1 christos st %f3,[%g1+1*8+0] 545 1.1 christos st %f2,[%g1+1*8+4] 546 1.1 christos st %f5,[%g1+2*8+0] 547 1.1 christos st %f4,[%g1+2*8+4] 548 1.1 christos st %f7,[%g1+3*8+0] 549 1.1 christos st %f6,[%g1+3*8+4] 550 1.1 christos st %f9,[%g1+4*8+0] 551 1.1 christos st %f8,[%g1+4*8+4] 552 1.1 christos st %f11,[%g1+5*8+0] 553 1.1 christos st %f10,[%g1+5*8+4] 554 1.1 christos st %f13,[%g1+6*8+0] 555 1.1 christos st %f12,[%g1+6*8+4] 556 1.1 christos st %f15,[%g1+7*8+0] 557 1.1 christos st %f14,[%g1+7*8+4] 558 1.1 christos st %f17,[%g1+8*8+0] 559 1.1 christos st %f16,[%g1+8*8+4] 560 1.1 christos st %f19,[%g1+9*8+0] 561 1.1 christos st %f18,[%g1+9*8+4] 562 1.1 christos st %f21,[%g1+10*8+0] 563 1.1 christos st %f20,[%g1+10*8+4] 564 1.1 christos st %f23,[%g1+11*8+0] 565 1.1 christos st %f22,[%g1+11*8+4] 566 1.1 christos .word 0x81b00f1d !fsrc2 %f0,%f60,%f0 567 1.1 christos st %f1,[%g1+12*8+0] 568 1.1 christos st %f0,[%g1+12*8+4] 569 1.1 christos .word 0x85b00f1f !fsrc2 %f0,%f62,%f2 570 1.1 christos st %f3,[%g1+13*8+0] 571 1.1 christos st %f2,[%g1+13*8+4] 572 1.1 christos .word 0x89b00f18 !fsrc2 %f0,%f24,%f4 573 1.1 christos st %f5,[%g1+14*8+0] 574 1.1 christos st %f4,[%g1+14*8+4] 575 1.1 christos .word 0x8db00f1a !fsrc2 %f0,%f26,%f6 576 1.1 christos st %f7,[%g1+15*8+0] 577 1.1 christos st %f6,[%g1+15*8+4] 578 1.1 christos mov 1,%i0 ! return success 579 1.1 christos .Lmdone_16: 580 1.1 christos ret 581 1.1 christos restore 582 1.1 christos 583 1.1 christos .Lmabort_16: 584 1.1 christos restore 585 1.1 christos restore 586 1.1 christos restore 587 1.1 christos restore 588 1.1 christos restore 589 1.1 christos .Lmabort1_16: 590 1.1 christos restore 591 1.1 christos 592 1.1 christos mov 0,%i0 ! return failure 593 1.1 christos ret 594 1.1 christos restore 595 1.1 christos 596 1.1 christos .align 32 597 1.1 christos .Lmsquare_16: 598 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 599 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 600 1.1 christos .word 0x81b02940+16-1 ! montsqr 16-1 601 1.1 christos ba .Lmresume_16 602 1.1 christos nop 603 1.1 christos .type bn_mul_mont_t4_16, #function 604 1.1 christos .size bn_mul_mont_t4_16, .-bn_mul_mont_t4_16 605 1.1 christos .globl bn_mul_mont_t4_24 606 1.1 christos .align 32 607 1.1 christos bn_mul_mont_t4_24: 608 1.1 christos #ifdef __arch64__ 609 1.1 christos mov 0,%g5 610 1.1 christos mov -128,%g4 611 1.1 christos #elif defined(SPARCV9_64BIT_STACK) 612 1.1 christos SPARC_LOAD_ADDRESS_LEAF(OPENSSL_sparcv9cap_P,%g1,%g5) 613 1.1 christos ld [%g1+0],%g1 ! OPENSSL_sparcv9_P[0] 614 1.1 christos mov -2047,%g4 615 1.1 christos and %g1,SPARCV9_64BIT_STACK,%g1 616 1.1 christos movrz %g1,0,%g4 617 1.1 christos mov -1,%g5 618 1.1 christos add %g4,-128,%g4 619 1.1 christos #else 620 1.1 christos mov -1,%g5 621 1.1 christos mov -128,%g4 622 1.1 christos #endif 623 1.1 christos sllx %g5,32,%g5 624 1.1 christos save %sp,%g4,%sp 625 1.1 christos #ifndef __arch64__ 626 1.1 christos save %sp,-128,%sp ! warm it up 627 1.1 christos save %sp,-128,%sp 628 1.1 christos save %sp,-128,%sp 629 1.1 christos save %sp,-128,%sp 630 1.1 christos save %sp,-128,%sp 631 1.1 christos save %sp,-128,%sp 632 1.1 christos restore 633 1.1 christos restore 634 1.1 christos restore 635 1.1 christos restore 636 1.1 christos restore 637 1.1 christos restore 638 1.1 christos #endif 639 1.1 christos and %sp,1,%g4 640 1.1 christos or %g5,%fp,%fp 641 1.1 christos or %g4,%g5,%g5 642 1.1 christos 643 1.1 christos ! copy arguments to global registers 644 1.1 christos mov %i0,%g1 645 1.1 christos mov %i1,%g2 646 1.1 christos mov %i2,%g3 647 1.1 christos mov %i3,%g4 648 1.1 christos ld [%i4+0],%f1 ! load *n0 649 1.1 christos ld [%i4+4],%f0 650 1.1 christos .word 0xbbb00f00 !fsrc2 %f0,%f0,%f60 651 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 652 1.1 christos ld [%g2+0*8+0],%l1 653 1.1 christos ld [%g2+0*8+4],%l0 654 1.1 christos sllx %l0,32,%l0 655 1.1 christos or %l1,%l0,%l0 656 1.1 christos ld [%g2+1*8+0],%l2 657 1.1 christos ld [%g2+1*8+4],%l1 658 1.1 christos sllx %l1,32,%l1 659 1.1 christos or %l2,%l1,%l1 660 1.1 christos ld [%g2+2*8+0],%l3 661 1.1 christos ld [%g2+2*8+4],%l2 662 1.1 christos sllx %l2,32,%l2 663 1.1 christos or %l3,%l2,%l2 664 1.1 christos ld [%g2+3*8+0],%l4 665 1.1 christos ld [%g2+3*8+4],%l3 666 1.1 christos sllx %l3,32,%l3 667 1.1 christos or %l4,%l3,%l3 668 1.1 christos ld [%g2+4*8+0],%l5 669 1.1 christos ld [%g2+4*8+4],%l4 670 1.1 christos sllx %l4,32,%l4 671 1.1 christos or %l5,%l4,%l4 672 1.1 christos ld [%g2+5*8+0],%l6 673 1.1 christos ld [%g2+5*8+4],%l5 674 1.1 christos sllx %l5,32,%l5 675 1.1 christos or %l6,%l5,%l5 676 1.1 christos ld [%g2+6*8+0],%l7 677 1.1 christos ld [%g2+6*8+4],%l6 678 1.1 christos sllx %l6,32,%l6 679 1.1 christos or %l7,%l6,%l6 680 1.1 christos ld [%g2+7*8+0],%o0 681 1.1 christos ld [%g2+7*8+4],%l7 682 1.1 christos sllx %l7,32,%l7 683 1.1 christos or %o0,%l7,%l7 684 1.1 christos ld [%g2+8*8+0],%o1 685 1.1 christos ld [%g2+8*8+4],%o0 686 1.1 christos sllx %o0,32,%o0 687 1.1 christos or %o1,%o0,%o0 688 1.1 christos ld [%g2+9*8+0],%o2 689 1.1 christos ld [%g2+9*8+4],%o1 690 1.1 christos sllx %o1,32,%o1 691 1.1 christos or %o2,%o1,%o1 692 1.1 christos ld [%g2+10*8+0],%o3 693 1.1 christos ld [%g2+10*8+4],%o2 694 1.1 christos sllx %o2,32,%o2 695 1.1 christos or %o3,%o2,%o2 696 1.1 christos ld [%g2+11*8+0],%o4 697 1.1 christos ld [%g2+11*8+4],%o3 698 1.1 christos sllx %o3,32,%o3 699 1.1 christos or %o4,%o3,%o3 700 1.1 christos ld [%g2+12*8+0],%o5 701 1.1 christos ld [%g2+12*8+4],%o4 702 1.1 christos sllx %o4,32,%o4 703 1.1 christos or %o5,%o4,%o4 704 1.1 christos ld [%g2+13*8+0],%o7 705 1.1 christos ld [%g2+13*8+4],%o5 706 1.1 christos sllx %o5,32,%o5 707 1.1 christos or %o7,%o5,%o5 708 1.1 christos ld [%g2+14*8+0],%f5 709 1.1 christos ld [%g2+14*8+4],%f4 710 1.1 christos .word 0xb1b00f04 !fsrc2 %f0,%f4,%f24 711 1.1 christos ld [%g2+15*8+0],%f7 712 1.1 christos ld [%g2+15*8+4],%f6 713 1.1 christos .word 0xb5b00f06 !fsrc2 %f0,%f6,%f26 714 1.1 christos ld [%g2+16*8+0],%f1 715 1.1 christos ld [%g2+16*8+4],%f0 716 1.1 christos .word 0xb9b00f00 !fsrc2 %f0,%f0,%f28 717 1.1 christos ld [%g2+17*8+0],%f3 718 1.1 christos ld [%g2+17*8+4],%f2 719 1.1 christos .word 0xbdb00f02 !fsrc2 %f0,%f2,%f30 720 1.1 christos ld [%g2+18*8+0],%f5 721 1.1 christos ld [%g2+18*8+4],%f4 722 1.1 christos .word 0x83b00f04 !fsrc2 %f0,%f4,%f32 723 1.1 christos ld [%g2+19*8+0],%f7 724 1.1 christos ld [%g2+19*8+4],%f6 725 1.1 christos .word 0x87b00f06 !fsrc2 %f0,%f6,%f34 726 1.1 christos ld [%g2+20*8+0],%f1 727 1.1 christos ld [%g2+20*8+4],%f0 728 1.1 christos .word 0x8bb00f00 !fsrc2 %f0,%f0,%f36 729 1.1 christos ld [%g2+21*8+0],%f3 730 1.1 christos ld [%g2+21*8+4],%f2 731 1.1 christos .word 0x8fb00f02 !fsrc2 %f0,%f2,%f38 732 1.1 christos ld [%g2+22*8+0],%f5 733 1.1 christos ld [%g2+22*8+4],%f4 734 1.1 christos .word 0x93b00f04 !fsrc2 %f0,%f4,%f40 735 1.1 christos ld [%g2+23*8+0],%f7 736 1.1 christos ld [%g2+23*8+4],%f6 737 1.1 christos .word 0x97b00f06 !fsrc2 %f0,%f6,%f42 738 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 739 1.1 christos ld [%g4+0*8+0],%l1 740 1.1 christos ld [%g4+0*8+4],%l0 741 1.1 christos sllx %l0,32,%l0 742 1.1 christos or %l1,%l0,%l0 743 1.1 christos ld [%g4+1*8+0],%l2 744 1.1 christos ld [%g4+1*8+4],%l1 745 1.1 christos sllx %l1,32,%l1 746 1.1 christos or %l2,%l1,%l1 747 1.1 christos ld [%g4+2*8+0],%l3 748 1.1 christos ld [%g4+2*8+4],%l2 749 1.1 christos sllx %l2,32,%l2 750 1.1 christos or %l3,%l2,%l2 751 1.1 christos ld [%g4+3*8+0],%l4 752 1.1 christos ld [%g4+3*8+4],%l3 753 1.1 christos sllx %l3,32,%l3 754 1.1 christos or %l4,%l3,%l3 755 1.1 christos ld [%g4+4*8+0],%l5 756 1.1 christos ld [%g4+4*8+4],%l4 757 1.1 christos sllx %l4,32,%l4 758 1.1 christos or %l5,%l4,%l4 759 1.1 christos ld [%g4+5*8+0],%l6 760 1.1 christos ld [%g4+5*8+4],%l5 761 1.1 christos sllx %l5,32,%l5 762 1.1 christos or %l6,%l5,%l5 763 1.1 christos ld [%g4+6*8+0],%l7 764 1.1 christos ld [%g4+6*8+4],%l6 765 1.1 christos sllx %l6,32,%l6 766 1.1 christos or %l7,%l6,%l6 767 1.1 christos ld [%g4+7*8+0],%o0 768 1.1 christos ld [%g4+7*8+4],%l7 769 1.1 christos sllx %l7,32,%l7 770 1.1 christos or %o0,%l7,%l7 771 1.1 christos ld [%g4+8*8+0],%o1 772 1.1 christos ld [%g4+8*8+4],%o0 773 1.1 christos sllx %o0,32,%o0 774 1.1 christos or %o1,%o0,%o0 775 1.1 christos ld [%g4+9*8+0],%o2 776 1.1 christos ld [%g4+9*8+4],%o1 777 1.1 christos sllx %o1,32,%o1 778 1.1 christos or %o2,%o1,%o1 779 1.1 christos ld [%g4+10*8+0],%o3 780 1.1 christos ld [%g4+10*8+4],%o2 781 1.1 christos sllx %o2,32,%o2 782 1.1 christos or %o3,%o2,%o2 783 1.1 christos ld [%g4+11*8+0],%o4 784 1.1 christos ld [%g4+11*8+4],%o3 785 1.1 christos sllx %o3,32,%o3 786 1.1 christos or %o4,%o3,%o3 787 1.1 christos ld [%g4+12*8+0],%o5 788 1.1 christos ld [%g4+12*8+4],%o4 789 1.1 christos sllx %o4,32,%o4 790 1.1 christos or %o5,%o4,%o4 791 1.1 christos ld [%g4+13*8+0],%o7 792 1.1 christos ld [%g4+13*8+4],%o5 793 1.1 christos sllx %o5,32,%o5 794 1.1 christos or %o7,%o5,%o5 795 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 796 1.1 christos ld [%g4+14*8+0],%l1 797 1.1 christos ld [%g4+14*8+4],%l0 798 1.1 christos sllx %l0,32,%l0 799 1.1 christos or %l1,%l0,%l0 800 1.1 christos ld [%g4+15*8+0],%l2 801 1.1 christos ld [%g4+15*8+4],%l1 802 1.1 christos sllx %l1,32,%l1 803 1.1 christos or %l2,%l1,%l1 804 1.1 christos ld [%g4+16*8+0],%l3 805 1.1 christos ld [%g4+16*8+4],%l2 806 1.1 christos sllx %l2,32,%l2 807 1.1 christos or %l3,%l2,%l2 808 1.1 christos ld [%g4+17*8+0],%l4 809 1.1 christos ld [%g4+17*8+4],%l3 810 1.1 christos sllx %l3,32,%l3 811 1.1 christos or %l4,%l3,%l3 812 1.1 christos ld [%g4+18*8+0],%l5 813 1.1 christos ld [%g4+18*8+4],%l4 814 1.1 christos sllx %l4,32,%l4 815 1.1 christos or %l5,%l4,%l4 816 1.1 christos ld [%g4+19*8+0],%l6 817 1.1 christos ld [%g4+19*8+4],%l5 818 1.1 christos sllx %l5,32,%l5 819 1.1 christos or %l6,%l5,%l5 820 1.1 christos ld [%g4+20*8+0],%l7 821 1.1 christos ld [%g4+20*8+4],%l6 822 1.1 christos sllx %l6,32,%l6 823 1.1 christos or %l7,%l6,%l6 824 1.1 christos ld [%g4+21*8+0],%o0 825 1.1 christos ld [%g4+21*8+4],%l7 826 1.1 christos sllx %l7,32,%l7 827 1.1 christos or %o0,%l7,%l7 828 1.1 christos ld [%g4+22*8+0],%o1 829 1.1 christos ld [%g4+22*8+4],%o0 830 1.1 christos sllx %o0,32,%o0 831 1.1 christos or %o1,%o0,%o0 832 1.1 christos ld [%g4+23*8+0],%o2 833 1.1 christos ld [%g4+23*8+4],%o1 834 1.1 christos sllx %o1,32,%o1 835 1.1 christos or %o2,%o1,%o1 836 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 837 1.1 christos cmp %g2,%g3 838 1.1 christos be SIZE_T_CC,.Lmsquare_24 839 1.1 christos nop 840 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 841 1.1 christos ld [%g3+0*8+0],%i1 842 1.1 christos ld [%g3+0*8+4],%i0 843 1.1 christos sllx %i0,32,%i0 844 1.1 christos or %i1,%i0,%i0 845 1.1 christos ld [%g3+1*8+0],%i2 846 1.1 christos ld [%g3+1*8+4],%i1 847 1.1 christos sllx %i1,32,%i1 848 1.1 christos or %i2,%i1,%i1 849 1.1 christos ld [%g3+2*8+0],%i3 850 1.1 christos ld [%g3+2*8+4],%i2 851 1.1 christos sllx %i2,32,%i2 852 1.1 christos or %i3,%i2,%i2 853 1.1 christos ld [%g3+3*8+0],%i4 854 1.1 christos ld [%g3+3*8+4],%i3 855 1.1 christos sllx %i3,32,%i3 856 1.1 christos or %i4,%i3,%i3 857 1.1 christos ld [%g3+4*8+0],%i5 858 1.1 christos ld [%g3+4*8+4],%i4 859 1.1 christos sllx %i4,32,%i4 860 1.1 christos or %i5,%i4,%i4 861 1.1 christos ld [%g3+5*8+0],%l0 862 1.1 christos ld [%g3+5*8+4],%i5 863 1.1 christos sllx %i5,32,%i5 864 1.1 christos or %l0,%i5,%i5 865 1.1 christos ld [%g3+6*8+0],%l1 866 1.1 christos ld [%g3+6*8+4],%l0 867 1.1 christos sllx %l0,32,%l0 868 1.1 christos or %l1,%l0,%l0 869 1.1 christos ld [%g3+7*8+0],%l2 870 1.1 christos ld [%g3+7*8+4],%l1 871 1.1 christos sllx %l1,32,%l1 872 1.1 christos or %l2,%l1,%l1 873 1.1 christos ld [%g3+8*8+0],%l3 874 1.1 christos ld [%g3+8*8+4],%l2 875 1.1 christos sllx %l2,32,%l2 876 1.1 christos or %l3,%l2,%l2 877 1.1 christos ld [%g3+9*8+0],%l4 878 1.1 christos ld [%g3+9*8+4],%l3 879 1.1 christos sllx %l3,32,%l3 880 1.1 christos or %l4,%l3,%l3 881 1.1 christos ld [%g3+10*8+0],%l5 882 1.1 christos ld [%g3+10*8+4],%l4 883 1.1 christos sllx %l4,32,%l4 884 1.1 christos or %l5,%l4,%l4 885 1.1 christos ld [%g3+11*8+0],%l6 886 1.1 christos ld [%g3+11*8+4],%l5 887 1.1 christos sllx %l5,32,%l5 888 1.1 christos or %l6,%l5,%l5 889 1.1 christos ld [%g3+12*8+0],%l7 890 1.1 christos ld [%g3+12*8+4],%l6 891 1.1 christos sllx %l6,32,%l6 892 1.1 christos or %l7,%l6,%l6 893 1.1 christos ld [%g3+13*8+0],%o7 894 1.1 christos ld [%g3+13*8+4],%l7 895 1.1 christos sllx %l7,32,%l7 896 1.1 christos or %o7,%l7,%l7 897 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 898 1.1 christos ld [%g3+14*8+0],%i1 899 1.1 christos ld [%g3+14*8+4],%i0 900 1.1 christos sllx %i0,32,%i0 901 1.1 christos or %i1,%i0,%i0 902 1.1 christos ld [%g3+15*8+0],%i2 903 1.1 christos ld [%g3+15*8+4],%i1 904 1.1 christos sllx %i1,32,%i1 905 1.1 christos or %i2,%i1,%i1 906 1.1 christos ld [%g3+16*8+0],%i3 907 1.1 christos ld [%g3+16*8+4],%i2 908 1.1 christos sllx %i2,32,%i2 909 1.1 christos or %i3,%i2,%i2 910 1.1 christos ld [%g3+17*8+0],%i4 911 1.1 christos ld [%g3+17*8+4],%i3 912 1.1 christos sllx %i3,32,%i3 913 1.1 christos or %i4,%i3,%i3 914 1.1 christos ld [%g3+18*8+0],%i5 915 1.1 christos ld [%g3+18*8+4],%i4 916 1.1 christos sllx %i4,32,%i4 917 1.1 christos or %i5,%i4,%i4 918 1.1 christos ld [%g3+19*8+0],%l0 919 1.1 christos ld [%g3+19*8+4],%i5 920 1.1 christos sllx %i5,32,%i5 921 1.1 christos or %l0,%i5,%i5 922 1.1 christos ld [%g3+20*8+0],%l1 923 1.1 christos ld [%g3+20*8+4],%l0 924 1.1 christos sllx %l0,32,%l0 925 1.1 christos or %l1,%l0,%l0 926 1.1 christos ld [%g3+21*8+0],%l2 927 1.1 christos ld [%g3+21*8+4],%l1 928 1.1 christos sllx %l1,32,%l1 929 1.1 christos or %l2,%l1,%l1 930 1.1 christos ld [%g3+22*8+0],%l3 931 1.1 christos ld [%g3+22*8+4],%l2 932 1.1 christos sllx %l2,32,%l2 933 1.1 christos or %l3,%l2,%l2 934 1.1 christos ld [%g3+23*8+0],%o7 935 1.1 christos ld [%g3+23*8+4],%l3 936 1.1 christos sllx %l3,32,%l3 937 1.1 christos or %o7,%l3,%l3 938 1.1 christos .word 0x81b02920+24-1 ! montmul 24-1 939 1.1 christos .Lmresume_24: 940 1.1 christos fbu,pn %fcc3,.Lmabort_24 941 1.1 christos #ifndef __arch64__ 942 1.1 christos and %fp,%g5,%g5 943 1.1 christos brz,pn %g5,.Lmabort_24 944 1.1 christos #endif 945 1.1 christos nop 946 1.1 christos #ifdef __arch64__ 947 1.1 christos restore 948 1.1 christos restore 949 1.1 christos restore 950 1.1 christos restore 951 1.1 christos restore 952 1.1 christos #else 953 1.1 christos restore; and %fp,%g5,%g5 954 1.1 christos restore; and %fp,%g5,%g5 955 1.1 christos restore; and %fp,%g5,%g5 956 1.1 christos restore; and %fp,%g5,%g5 957 1.1 christos brz,pn %g5,.Lmabort1_24 958 1.1 christos restore 959 1.1 christos #endif 960 1.1 christos .word 0x81b02310 !movxtod %l0,%f0 961 1.1 christos .word 0x85b02311 !movxtod %l1,%f2 962 1.1 christos .word 0x89b02312 !movxtod %l2,%f4 963 1.1 christos .word 0x8db02313 !movxtod %l3,%f6 964 1.1 christos .word 0x91b02314 !movxtod %l4,%f8 965 1.1 christos .word 0x95b02315 !movxtod %l5,%f10 966 1.1 christos .word 0x99b02316 !movxtod %l6,%f12 967 1.1 christos .word 0x9db02317 !movxtod %l7,%f14 968 1.1 christos .word 0xa1b02308 !movxtod %o0,%f16 969 1.1 christos .word 0xa5b02309 !movxtod %o1,%f18 970 1.1 christos .word 0xa9b0230a !movxtod %o2,%f20 971 1.1 christos .word 0xadb0230b !movxtod %o3,%f22 972 1.1 christos .word 0xbbb0230c !movxtod %o4,%f60 973 1.1 christos .word 0xbfb0230d !movxtod %o5,%f62 974 1.1 christos #ifdef __arch64__ 975 1.1 christos restore 976 1.1 christos #else 977 1.1 christos and %fp,%g5,%g5 978 1.1 christos restore 979 1.1 christos and %g5,1,%o7 980 1.1 christos and %fp,%g5,%g5 981 1.1 christos srl %fp,0,%fp ! just in case? 982 1.1 christos or %o7,%g5,%g5 983 1.1 christos brz,a,pn %g5,.Lmdone_24 984 1.1 christos mov 0,%i0 ! return failure 985 1.1 christos #endif 986 1.1 christos st %f1,[%g1+0*8+0] 987 1.1 christos st %f0,[%g1+0*8+4] 988 1.1 christos st %f3,[%g1+1*8+0] 989 1.1 christos st %f2,[%g1+1*8+4] 990 1.1 christos st %f5,[%g1+2*8+0] 991 1.1 christos st %f4,[%g1+2*8+4] 992 1.1 christos st %f7,[%g1+3*8+0] 993 1.1 christos st %f6,[%g1+3*8+4] 994 1.1 christos st %f9,[%g1+4*8+0] 995 1.1 christos st %f8,[%g1+4*8+4] 996 1.1 christos st %f11,[%g1+5*8+0] 997 1.1 christos st %f10,[%g1+5*8+4] 998 1.1 christos st %f13,[%g1+6*8+0] 999 1.1 christos st %f12,[%g1+6*8+4] 1000 1.1 christos st %f15,[%g1+7*8+0] 1001 1.1 christos st %f14,[%g1+7*8+4] 1002 1.1 christos st %f17,[%g1+8*8+0] 1003 1.1 christos st %f16,[%g1+8*8+4] 1004 1.1 christos st %f19,[%g1+9*8+0] 1005 1.1 christos st %f18,[%g1+9*8+4] 1006 1.1 christos st %f21,[%g1+10*8+0] 1007 1.1 christos st %f20,[%g1+10*8+4] 1008 1.1 christos st %f23,[%g1+11*8+0] 1009 1.1 christos st %f22,[%g1+11*8+4] 1010 1.1 christos .word 0x81b00f1d !fsrc2 %f0,%f60,%f0 1011 1.1 christos st %f1,[%g1+12*8+0] 1012 1.1 christos st %f0,[%g1+12*8+4] 1013 1.1 christos .word 0x85b00f1f !fsrc2 %f0,%f62,%f2 1014 1.1 christos st %f3,[%g1+13*8+0] 1015 1.1 christos st %f2,[%g1+13*8+4] 1016 1.1 christos .word 0x89b00f18 !fsrc2 %f0,%f24,%f4 1017 1.1 christos st %f5,[%g1+14*8+0] 1018 1.1 christos st %f4,[%g1+14*8+4] 1019 1.1 christos .word 0x8db00f1a !fsrc2 %f0,%f26,%f6 1020 1.1 christos st %f7,[%g1+15*8+0] 1021 1.1 christos st %f6,[%g1+15*8+4] 1022 1.1 christos .word 0x81b00f1c !fsrc2 %f0,%f28,%f0 1023 1.1 christos st %f1,[%g1+16*8+0] 1024 1.1 christos st %f0,[%g1+16*8+4] 1025 1.1 christos .word 0x85b00f1e !fsrc2 %f0,%f30,%f2 1026 1.1 christos st %f3,[%g1+17*8+0] 1027 1.1 christos st %f2,[%g1+17*8+4] 1028 1.1 christos .word 0x89b00f01 !fsrc2 %f0,%f32,%f4 1029 1.1 christos st %f5,[%g1+18*8+0] 1030 1.1 christos st %f4,[%g1+18*8+4] 1031 1.1 christos .word 0x8db00f03 !fsrc2 %f0,%f34,%f6 1032 1.1 christos st %f7,[%g1+19*8+0] 1033 1.1 christos st %f6,[%g1+19*8+4] 1034 1.1 christos .word 0x81b00f05 !fsrc2 %f0,%f36,%f0 1035 1.1 christos st %f1,[%g1+20*8+0] 1036 1.1 christos st %f0,[%g1+20*8+4] 1037 1.1 christos .word 0x85b00f07 !fsrc2 %f0,%f38,%f2 1038 1.1 christos st %f3,[%g1+21*8+0] 1039 1.1 christos st %f2,[%g1+21*8+4] 1040 1.1 christos .word 0x89b00f09 !fsrc2 %f0,%f40,%f4 1041 1.1 christos st %f5,[%g1+22*8+0] 1042 1.1 christos st %f4,[%g1+22*8+4] 1043 1.1 christos .word 0x8db00f0b !fsrc2 %f0,%f42,%f6 1044 1.1 christos st %f7,[%g1+23*8+0] 1045 1.1 christos st %f6,[%g1+23*8+4] 1046 1.1 christos mov 1,%i0 ! return success 1047 1.1 christos .Lmdone_24: 1048 1.1 christos ret 1049 1.1 christos restore 1050 1.1 christos 1051 1.1 christos .Lmabort_24: 1052 1.1 christos restore 1053 1.1 christos restore 1054 1.1 christos restore 1055 1.1 christos restore 1056 1.1 christos restore 1057 1.1 christos .Lmabort1_24: 1058 1.1 christos restore 1059 1.1 christos 1060 1.1 christos mov 0,%i0 ! return failure 1061 1.1 christos ret 1062 1.1 christos restore 1063 1.1 christos 1064 1.1 christos .align 32 1065 1.1 christos .Lmsquare_24: 1066 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1067 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1068 1.1 christos .word 0x81b02940+24-1 ! montsqr 24-1 1069 1.1 christos ba .Lmresume_24 1070 1.1 christos nop 1071 1.1 christos .type bn_mul_mont_t4_24, #function 1072 1.1 christos .size bn_mul_mont_t4_24, .-bn_mul_mont_t4_24 1073 1.1 christos .globl bn_mul_mont_t4_32 1074 1.1 christos .align 32 1075 1.1 christos bn_mul_mont_t4_32: 1076 1.1 christos #ifdef __arch64__ 1077 1.1 christos mov 0,%g5 1078 1.1 christos mov -128,%g4 1079 1.1 christos #elif defined(SPARCV9_64BIT_STACK) 1080 1.1 christos SPARC_LOAD_ADDRESS_LEAF(OPENSSL_sparcv9cap_P,%g1,%g5) 1081 1.1 christos ld [%g1+0],%g1 ! OPENSSL_sparcv9_P[0] 1082 1.1 christos mov -2047,%g4 1083 1.1 christos and %g1,SPARCV9_64BIT_STACK,%g1 1084 1.1 christos movrz %g1,0,%g4 1085 1.1 christos mov -1,%g5 1086 1.1 christos add %g4,-128,%g4 1087 1.1 christos #else 1088 1.1 christos mov -1,%g5 1089 1.1 christos mov -128,%g4 1090 1.1 christos #endif 1091 1.1 christos sllx %g5,32,%g5 1092 1.1 christos save %sp,%g4,%sp 1093 1.1 christos #ifndef __arch64__ 1094 1.1 christos save %sp,-128,%sp ! warm it up 1095 1.1 christos save %sp,-128,%sp 1096 1.1 christos save %sp,-128,%sp 1097 1.1 christos save %sp,-128,%sp 1098 1.1 christos save %sp,-128,%sp 1099 1.1 christos save %sp,-128,%sp 1100 1.1 christos restore 1101 1.1 christos restore 1102 1.1 christos restore 1103 1.1 christos restore 1104 1.1 christos restore 1105 1.1 christos restore 1106 1.1 christos #endif 1107 1.1 christos and %sp,1,%g4 1108 1.1 christos or %g5,%fp,%fp 1109 1.1 christos or %g4,%g5,%g5 1110 1.1 christos 1111 1.1 christos ! copy arguments to global registers 1112 1.1 christos mov %i0,%g1 1113 1.1 christos mov %i1,%g2 1114 1.1 christos mov %i2,%g3 1115 1.1 christos mov %i3,%g4 1116 1.1 christos ld [%i4+0],%f1 ! load *n0 1117 1.1 christos ld [%i4+4],%f0 1118 1.1 christos .word 0xbbb00f00 !fsrc2 %f0,%f0,%f60 1119 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1120 1.1 christos ld [%g2+0*8+0],%l1 1121 1.1 christos ld [%g2+0*8+4],%l0 1122 1.1 christos sllx %l0,32,%l0 1123 1.1 christos or %l1,%l0,%l0 1124 1.1 christos ld [%g2+1*8+0],%l2 1125 1.1 christos ld [%g2+1*8+4],%l1 1126 1.1 christos sllx %l1,32,%l1 1127 1.1 christos or %l2,%l1,%l1 1128 1.1 christos ld [%g2+2*8+0],%l3 1129 1.1 christos ld [%g2+2*8+4],%l2 1130 1.1 christos sllx %l2,32,%l2 1131 1.1 christos or %l3,%l2,%l2 1132 1.1 christos ld [%g2+3*8+0],%l4 1133 1.1 christos ld [%g2+3*8+4],%l3 1134 1.1 christos sllx %l3,32,%l3 1135 1.1 christos or %l4,%l3,%l3 1136 1.1 christos ld [%g2+4*8+0],%l5 1137 1.1 christos ld [%g2+4*8+4],%l4 1138 1.1 christos sllx %l4,32,%l4 1139 1.1 christos or %l5,%l4,%l4 1140 1.1 christos ld [%g2+5*8+0],%l6 1141 1.1 christos ld [%g2+5*8+4],%l5 1142 1.1 christos sllx %l5,32,%l5 1143 1.1 christos or %l6,%l5,%l5 1144 1.1 christos ld [%g2+6*8+0],%l7 1145 1.1 christos ld [%g2+6*8+4],%l6 1146 1.1 christos sllx %l6,32,%l6 1147 1.1 christos or %l7,%l6,%l6 1148 1.1 christos ld [%g2+7*8+0],%o0 1149 1.1 christos ld [%g2+7*8+4],%l7 1150 1.1 christos sllx %l7,32,%l7 1151 1.1 christos or %o0,%l7,%l7 1152 1.1 christos ld [%g2+8*8+0],%o1 1153 1.1 christos ld [%g2+8*8+4],%o0 1154 1.1 christos sllx %o0,32,%o0 1155 1.1 christos or %o1,%o0,%o0 1156 1.1 christos ld [%g2+9*8+0],%o2 1157 1.1 christos ld [%g2+9*8+4],%o1 1158 1.1 christos sllx %o1,32,%o1 1159 1.1 christos or %o2,%o1,%o1 1160 1.1 christos ld [%g2+10*8+0],%o3 1161 1.1 christos ld [%g2+10*8+4],%o2 1162 1.1 christos sllx %o2,32,%o2 1163 1.1 christos or %o3,%o2,%o2 1164 1.1 christos ld [%g2+11*8+0],%o4 1165 1.1 christos ld [%g2+11*8+4],%o3 1166 1.1 christos sllx %o3,32,%o3 1167 1.1 christos or %o4,%o3,%o3 1168 1.1 christos ld [%g2+12*8+0],%o5 1169 1.1 christos ld [%g2+12*8+4],%o4 1170 1.1 christos sllx %o4,32,%o4 1171 1.1 christos or %o5,%o4,%o4 1172 1.1 christos ld [%g2+13*8+0],%o7 1173 1.1 christos ld [%g2+13*8+4],%o5 1174 1.1 christos sllx %o5,32,%o5 1175 1.1 christos or %o7,%o5,%o5 1176 1.1 christos ld [%g2+14*8+0],%f5 1177 1.1 christos ld [%g2+14*8+4],%f4 1178 1.1 christos .word 0xb1b00f04 !fsrc2 %f0,%f4,%f24 1179 1.1 christos ld [%g2+15*8+0],%f7 1180 1.1 christos ld [%g2+15*8+4],%f6 1181 1.1 christos .word 0xb5b00f06 !fsrc2 %f0,%f6,%f26 1182 1.1 christos ld [%g2+16*8+0],%f1 1183 1.1 christos ld [%g2+16*8+4],%f0 1184 1.1 christos .word 0xb9b00f00 !fsrc2 %f0,%f0,%f28 1185 1.1 christos ld [%g2+17*8+0],%f3 1186 1.1 christos ld [%g2+17*8+4],%f2 1187 1.1 christos .word 0xbdb00f02 !fsrc2 %f0,%f2,%f30 1188 1.1 christos ld [%g2+18*8+0],%f5 1189 1.1 christos ld [%g2+18*8+4],%f4 1190 1.1 christos .word 0x83b00f04 !fsrc2 %f0,%f4,%f32 1191 1.1 christos ld [%g2+19*8+0],%f7 1192 1.1 christos ld [%g2+19*8+4],%f6 1193 1.1 christos .word 0x87b00f06 !fsrc2 %f0,%f6,%f34 1194 1.1 christos ld [%g2+20*8+0],%f1 1195 1.1 christos ld [%g2+20*8+4],%f0 1196 1.1 christos .word 0x8bb00f00 !fsrc2 %f0,%f0,%f36 1197 1.1 christos ld [%g2+21*8+0],%f3 1198 1.1 christos ld [%g2+21*8+4],%f2 1199 1.1 christos .word 0x8fb00f02 !fsrc2 %f0,%f2,%f38 1200 1.1 christos ld [%g2+22*8+0],%f5 1201 1.1 christos ld [%g2+22*8+4],%f4 1202 1.1 christos .word 0x93b00f04 !fsrc2 %f0,%f4,%f40 1203 1.1 christos ld [%g2+23*8+0],%f7 1204 1.1 christos ld [%g2+23*8+4],%f6 1205 1.1 christos .word 0x97b00f06 !fsrc2 %f0,%f6,%f42 1206 1.1 christos ld [%g2+24*8+0],%f1 1207 1.1 christos ld [%g2+24*8+4],%f0 1208 1.1 christos .word 0x9bb00f00 !fsrc2 %f0,%f0,%f44 1209 1.1 christos ld [%g2+25*8+0],%f3 1210 1.1 christos ld [%g2+25*8+4],%f2 1211 1.1 christos .word 0x9fb00f02 !fsrc2 %f0,%f2,%f46 1212 1.1 christos ld [%g2+26*8+0],%f5 1213 1.1 christos ld [%g2+26*8+4],%f4 1214 1.1 christos .word 0xa3b00f04 !fsrc2 %f0,%f4,%f48 1215 1.1 christos ld [%g2+27*8+0],%f7 1216 1.1 christos ld [%g2+27*8+4],%f6 1217 1.1 christos .word 0xa7b00f06 !fsrc2 %f0,%f6,%f50 1218 1.1 christos ld [%g2+28*8+0],%f1 1219 1.1 christos ld [%g2+28*8+4],%f0 1220 1.1 christos .word 0xabb00f00 !fsrc2 %f0,%f0,%f52 1221 1.1 christos ld [%g2+29*8+0],%f3 1222 1.1 christos ld [%g2+29*8+4],%f2 1223 1.1 christos .word 0xafb00f02 !fsrc2 %f0,%f2,%f54 1224 1.1 christos ld [%g2+30*8+0],%f5 1225 1.1 christos ld [%g2+30*8+4],%f4 1226 1.1 christos .word 0xb3b00f04 !fsrc2 %f0,%f4,%f56 1227 1.1 christos ld [%g2+31*8+0],%f7 1228 1.1 christos ld [%g2+31*8+4],%f6 1229 1.1 christos .word 0xb7b00f06 !fsrc2 %f0,%f6,%f58 1230 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1231 1.1 christos ld [%g4+0*8+0],%l1 1232 1.1 christos ld [%g4+0*8+4],%l0 1233 1.1 christos sllx %l0,32,%l0 1234 1.1 christos or %l1,%l0,%l0 1235 1.1 christos ld [%g4+1*8+0],%l2 1236 1.1 christos ld [%g4+1*8+4],%l1 1237 1.1 christos sllx %l1,32,%l1 1238 1.1 christos or %l2,%l1,%l1 1239 1.1 christos ld [%g4+2*8+0],%l3 1240 1.1 christos ld [%g4+2*8+4],%l2 1241 1.1 christos sllx %l2,32,%l2 1242 1.1 christos or %l3,%l2,%l2 1243 1.1 christos ld [%g4+3*8+0],%l4 1244 1.1 christos ld [%g4+3*8+4],%l3 1245 1.1 christos sllx %l3,32,%l3 1246 1.1 christos or %l4,%l3,%l3 1247 1.1 christos ld [%g4+4*8+0],%l5 1248 1.1 christos ld [%g4+4*8+4],%l4 1249 1.1 christos sllx %l4,32,%l4 1250 1.1 christos or %l5,%l4,%l4 1251 1.1 christos ld [%g4+5*8+0],%l6 1252 1.1 christos ld [%g4+5*8+4],%l5 1253 1.1 christos sllx %l5,32,%l5 1254 1.1 christos or %l6,%l5,%l5 1255 1.1 christos ld [%g4+6*8+0],%l7 1256 1.1 christos ld [%g4+6*8+4],%l6 1257 1.1 christos sllx %l6,32,%l6 1258 1.1 christos or %l7,%l6,%l6 1259 1.1 christos ld [%g4+7*8+0],%o0 1260 1.1 christos ld [%g4+7*8+4],%l7 1261 1.1 christos sllx %l7,32,%l7 1262 1.1 christos or %o0,%l7,%l7 1263 1.1 christos ld [%g4+8*8+0],%o1 1264 1.1 christos ld [%g4+8*8+4],%o0 1265 1.1 christos sllx %o0,32,%o0 1266 1.1 christos or %o1,%o0,%o0 1267 1.1 christos ld [%g4+9*8+0],%o2 1268 1.1 christos ld [%g4+9*8+4],%o1 1269 1.1 christos sllx %o1,32,%o1 1270 1.1 christos or %o2,%o1,%o1 1271 1.1 christos ld [%g4+10*8+0],%o3 1272 1.1 christos ld [%g4+10*8+4],%o2 1273 1.1 christos sllx %o2,32,%o2 1274 1.1 christos or %o3,%o2,%o2 1275 1.1 christos ld [%g4+11*8+0],%o4 1276 1.1 christos ld [%g4+11*8+4],%o3 1277 1.1 christos sllx %o3,32,%o3 1278 1.1 christos or %o4,%o3,%o3 1279 1.1 christos ld [%g4+12*8+0],%o5 1280 1.1 christos ld [%g4+12*8+4],%o4 1281 1.1 christos sllx %o4,32,%o4 1282 1.1 christos or %o5,%o4,%o4 1283 1.1 christos ld [%g4+13*8+0],%o7 1284 1.1 christos ld [%g4+13*8+4],%o5 1285 1.1 christos sllx %o5,32,%o5 1286 1.1 christos or %o7,%o5,%o5 1287 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1288 1.1 christos ld [%g4+14*8+0],%l1 1289 1.1 christos ld [%g4+14*8+4],%l0 1290 1.1 christos sllx %l0,32,%l0 1291 1.1 christos or %l1,%l0,%l0 1292 1.1 christos ld [%g4+15*8+0],%l2 1293 1.1 christos ld [%g4+15*8+4],%l1 1294 1.1 christos sllx %l1,32,%l1 1295 1.1 christos or %l2,%l1,%l1 1296 1.1 christos ld [%g4+16*8+0],%l3 1297 1.1 christos ld [%g4+16*8+4],%l2 1298 1.1 christos sllx %l2,32,%l2 1299 1.1 christos or %l3,%l2,%l2 1300 1.1 christos ld [%g4+17*8+0],%l4 1301 1.1 christos ld [%g4+17*8+4],%l3 1302 1.1 christos sllx %l3,32,%l3 1303 1.1 christos or %l4,%l3,%l3 1304 1.1 christos ld [%g4+18*8+0],%l5 1305 1.1 christos ld [%g4+18*8+4],%l4 1306 1.1 christos sllx %l4,32,%l4 1307 1.1 christos or %l5,%l4,%l4 1308 1.1 christos ld [%g4+19*8+0],%l6 1309 1.1 christos ld [%g4+19*8+4],%l5 1310 1.1 christos sllx %l5,32,%l5 1311 1.1 christos or %l6,%l5,%l5 1312 1.1 christos ld [%g4+20*8+0],%l7 1313 1.1 christos ld [%g4+20*8+4],%l6 1314 1.1 christos sllx %l6,32,%l6 1315 1.1 christos or %l7,%l6,%l6 1316 1.1 christos ld [%g4+21*8+0],%o0 1317 1.1 christos ld [%g4+21*8+4],%l7 1318 1.1 christos sllx %l7,32,%l7 1319 1.1 christos or %o0,%l7,%l7 1320 1.1 christos ld [%g4+22*8+0],%o1 1321 1.1 christos ld [%g4+22*8+4],%o0 1322 1.1 christos sllx %o0,32,%o0 1323 1.1 christos or %o1,%o0,%o0 1324 1.1 christos ld [%g4+23*8+0],%o2 1325 1.1 christos ld [%g4+23*8+4],%o1 1326 1.1 christos sllx %o1,32,%o1 1327 1.1 christos or %o2,%o1,%o1 1328 1.1 christos ld [%g4+24*8+0],%o3 1329 1.1 christos ld [%g4+24*8+4],%o2 1330 1.1 christos sllx %o2,32,%o2 1331 1.1 christos or %o3,%o2,%o2 1332 1.1 christos ld [%g4+25*8+0],%o4 1333 1.1 christos ld [%g4+25*8+4],%o3 1334 1.1 christos sllx %o3,32,%o3 1335 1.1 christos or %o4,%o3,%o3 1336 1.1 christos ld [%g4+26*8+0],%o5 1337 1.1 christos ld [%g4+26*8+4],%o4 1338 1.1 christos sllx %o4,32,%o4 1339 1.1 christos or %o5,%o4,%o4 1340 1.1 christos ld [%g4+27*8+0],%o7 1341 1.1 christos ld [%g4+27*8+4],%o5 1342 1.1 christos sllx %o5,32,%o5 1343 1.1 christos or %o7,%o5,%o5 1344 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1345 1.1 christos ld [%g4+28*8+0],%l1 1346 1.1 christos ld [%g4+28*8+4],%l0 1347 1.1 christos sllx %l0,32,%l0 1348 1.1 christos or %l1,%l0,%l0 1349 1.1 christos ld [%g4+29*8+0],%l2 1350 1.1 christos ld [%g4+29*8+4],%l1 1351 1.1 christos sllx %l1,32,%l1 1352 1.1 christos or %l2,%l1,%l1 1353 1.1 christos ld [%g4+30*8+0],%l3 1354 1.1 christos ld [%g4+30*8+4],%l2 1355 1.1 christos sllx %l2,32,%l2 1356 1.1 christos or %l3,%l2,%l2 1357 1.1 christos ld [%g4+31*8+0],%o7 1358 1.1 christos ld [%g4+31*8+4],%l3 1359 1.1 christos sllx %l3,32,%l3 1360 1.1 christos or %o7,%l3,%l3 1361 1.1 christos cmp %g2,%g3 1362 1.1 christos be SIZE_T_CC,.Lmsquare_32 1363 1.1 christos nop 1364 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1365 1.1 christos ld [%g3+0*8+0],%i1 1366 1.1 christos ld [%g3+0*8+4],%i0 1367 1.1 christos sllx %i0,32,%i0 1368 1.1 christos or %i1,%i0,%i0 1369 1.1 christos ld [%g3+1*8+0],%i2 1370 1.1 christos ld [%g3+1*8+4],%i1 1371 1.1 christos sllx %i1,32,%i1 1372 1.1 christos or %i2,%i1,%i1 1373 1.1 christos ld [%g3+2*8+0],%i3 1374 1.1 christos ld [%g3+2*8+4],%i2 1375 1.1 christos sllx %i2,32,%i2 1376 1.1 christos or %i3,%i2,%i2 1377 1.1 christos ld [%g3+3*8+0],%i4 1378 1.1 christos ld [%g3+3*8+4],%i3 1379 1.1 christos sllx %i3,32,%i3 1380 1.1 christos or %i4,%i3,%i3 1381 1.1 christos ld [%g3+4*8+0],%i5 1382 1.1 christos ld [%g3+4*8+4],%i4 1383 1.1 christos sllx %i4,32,%i4 1384 1.1 christos or %i5,%i4,%i4 1385 1.1 christos ld [%g3+5*8+0],%l0 1386 1.1 christos ld [%g3+5*8+4],%i5 1387 1.1 christos sllx %i5,32,%i5 1388 1.1 christos or %l0,%i5,%i5 1389 1.1 christos ld [%g3+6*8+0],%l1 1390 1.1 christos ld [%g3+6*8+4],%l0 1391 1.1 christos sllx %l0,32,%l0 1392 1.1 christos or %l1,%l0,%l0 1393 1.1 christos ld [%g3+7*8+0],%l2 1394 1.1 christos ld [%g3+7*8+4],%l1 1395 1.1 christos sllx %l1,32,%l1 1396 1.1 christos or %l2,%l1,%l1 1397 1.1 christos ld [%g3+8*8+0],%l3 1398 1.1 christos ld [%g3+8*8+4],%l2 1399 1.1 christos sllx %l2,32,%l2 1400 1.1 christos or %l3,%l2,%l2 1401 1.1 christos ld [%g3+9*8+0],%l4 1402 1.1 christos ld [%g3+9*8+4],%l3 1403 1.1 christos sllx %l3,32,%l3 1404 1.1 christos or %l4,%l3,%l3 1405 1.1 christos ld [%g3+10*8+0],%l5 1406 1.1 christos ld [%g3+10*8+4],%l4 1407 1.1 christos sllx %l4,32,%l4 1408 1.1 christos or %l5,%l4,%l4 1409 1.1 christos ld [%g3+11*8+0],%l6 1410 1.1 christos ld [%g3+11*8+4],%l5 1411 1.1 christos sllx %l5,32,%l5 1412 1.1 christos or %l6,%l5,%l5 1413 1.1 christos ld [%g3+12*8+0],%l7 1414 1.1 christos ld [%g3+12*8+4],%l6 1415 1.1 christos sllx %l6,32,%l6 1416 1.1 christos or %l7,%l6,%l6 1417 1.1 christos ld [%g3+13*8+0],%o7 1418 1.1 christos ld [%g3+13*8+4],%l7 1419 1.1 christos sllx %l7,32,%l7 1420 1.1 christos or %o7,%l7,%l7 1421 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1422 1.1 christos ld [%g3+14*8+0],%i1 1423 1.1 christos ld [%g3+14*8+4],%i0 1424 1.1 christos sllx %i0,32,%i0 1425 1.1 christos or %i1,%i0,%i0 1426 1.1 christos ld [%g3+15*8+0],%i2 1427 1.1 christos ld [%g3+15*8+4],%i1 1428 1.1 christos sllx %i1,32,%i1 1429 1.1 christos or %i2,%i1,%i1 1430 1.1 christos ld [%g3+16*8+0],%i3 1431 1.1 christos ld [%g3+16*8+4],%i2 1432 1.1 christos sllx %i2,32,%i2 1433 1.1 christos or %i3,%i2,%i2 1434 1.1 christos ld [%g3+17*8+0],%i4 1435 1.1 christos ld [%g3+17*8+4],%i3 1436 1.1 christos sllx %i3,32,%i3 1437 1.1 christos or %i4,%i3,%i3 1438 1.1 christos ld [%g3+18*8+0],%i5 1439 1.1 christos ld [%g3+18*8+4],%i4 1440 1.1 christos sllx %i4,32,%i4 1441 1.1 christos or %i5,%i4,%i4 1442 1.1 christos ld [%g3+19*8+0],%l0 1443 1.1 christos ld [%g3+19*8+4],%i5 1444 1.1 christos sllx %i5,32,%i5 1445 1.1 christos or %l0,%i5,%i5 1446 1.1 christos ld [%g3+20*8+0],%l1 1447 1.1 christos ld [%g3+20*8+4],%l0 1448 1.1 christos sllx %l0,32,%l0 1449 1.1 christos or %l1,%l0,%l0 1450 1.1 christos ld [%g3+21*8+0],%l2 1451 1.1 christos ld [%g3+21*8+4],%l1 1452 1.1 christos sllx %l1,32,%l1 1453 1.1 christos or %l2,%l1,%l1 1454 1.1 christos ld [%g3+22*8+0],%l3 1455 1.1 christos ld [%g3+22*8+4],%l2 1456 1.1 christos sllx %l2,32,%l2 1457 1.1 christos or %l3,%l2,%l2 1458 1.1 christos ld [%g3+23*8+0],%l4 1459 1.1 christos ld [%g3+23*8+4],%l3 1460 1.1 christos sllx %l3,32,%l3 1461 1.1 christos or %l4,%l3,%l3 1462 1.1 christos ld [%g3+24*8+0],%l5 1463 1.1 christos ld [%g3+24*8+4],%l4 1464 1.1 christos sllx %l4,32,%l4 1465 1.1 christos or %l5,%l4,%l4 1466 1.1 christos ld [%g3+25*8+0],%l6 1467 1.1 christos ld [%g3+25*8+4],%l5 1468 1.1 christos sllx %l5,32,%l5 1469 1.1 christos or %l6,%l5,%l5 1470 1.1 christos ld [%g3+26*8+0],%l7 1471 1.1 christos ld [%g3+26*8+4],%l6 1472 1.1 christos sllx %l6,32,%l6 1473 1.1 christos or %l7,%l6,%l6 1474 1.1 christos ld [%g3+27*8+0],%o0 1475 1.1 christos ld [%g3+27*8+4],%l7 1476 1.1 christos sllx %l7,32,%l7 1477 1.1 christos or %o0,%l7,%l7 1478 1.1 christos ld [%g3+28*8+0],%o1 1479 1.1 christos ld [%g3+28*8+4],%o0 1480 1.1 christos sllx %o0,32,%o0 1481 1.1 christos or %o1,%o0,%o0 1482 1.1 christos ld [%g3+29*8+0],%o2 1483 1.1 christos ld [%g3+29*8+4],%o1 1484 1.1 christos sllx %o1,32,%o1 1485 1.1 christos or %o2,%o1,%o1 1486 1.1 christos ld [%g3+30*8+0],%o3 1487 1.1 christos ld [%g3+30*8+4],%o2 1488 1.1 christos sllx %o2,32,%o2 1489 1.1 christos or %o3,%o2,%o2 1490 1.1 christos ld [%g3+31*8+0],%o7 1491 1.1 christos ld [%g3+31*8+4],%o3 1492 1.1 christos sllx %o3,32,%o3 1493 1.1 christos or %o7,%o3,%o3 1494 1.1 christos .word 0x81b02920+32-1 ! montmul 32-1 1495 1.1 christos .Lmresume_32: 1496 1.1 christos fbu,pn %fcc3,.Lmabort_32 1497 1.1 christos #ifndef __arch64__ 1498 1.1 christos and %fp,%g5,%g5 1499 1.1 christos brz,pn %g5,.Lmabort_32 1500 1.1 christos #endif 1501 1.1 christos nop 1502 1.1 christos #ifdef __arch64__ 1503 1.1 christos restore 1504 1.1 christos restore 1505 1.1 christos restore 1506 1.1 christos restore 1507 1.1 christos restore 1508 1.1 christos #else 1509 1.1 christos restore; and %fp,%g5,%g5 1510 1.1 christos restore; and %fp,%g5,%g5 1511 1.1 christos restore; and %fp,%g5,%g5 1512 1.1 christos restore; and %fp,%g5,%g5 1513 1.1 christos brz,pn %g5,.Lmabort1_32 1514 1.1 christos restore 1515 1.1 christos #endif 1516 1.1 christos .word 0x81b02310 !movxtod %l0,%f0 1517 1.1 christos .word 0x85b02311 !movxtod %l1,%f2 1518 1.1 christos .word 0x89b02312 !movxtod %l2,%f4 1519 1.1 christos .word 0x8db02313 !movxtod %l3,%f6 1520 1.1 christos .word 0x91b02314 !movxtod %l4,%f8 1521 1.1 christos .word 0x95b02315 !movxtod %l5,%f10 1522 1.1 christos .word 0x99b02316 !movxtod %l6,%f12 1523 1.1 christos .word 0x9db02317 !movxtod %l7,%f14 1524 1.1 christos .word 0xa1b02308 !movxtod %o0,%f16 1525 1.1 christos .word 0xa5b02309 !movxtod %o1,%f18 1526 1.1 christos .word 0xa9b0230a !movxtod %o2,%f20 1527 1.1 christos .word 0xadb0230b !movxtod %o3,%f22 1528 1.1 christos .word 0xbbb0230c !movxtod %o4,%f60 1529 1.1 christos .word 0xbfb0230d !movxtod %o5,%f62 1530 1.1 christos #ifdef __arch64__ 1531 1.1 christos restore 1532 1.1 christos #else 1533 1.1 christos and %fp,%g5,%g5 1534 1.1 christos restore 1535 1.1 christos and %g5,1,%o7 1536 1.1 christos and %fp,%g5,%g5 1537 1.1 christos srl %fp,0,%fp ! just in case? 1538 1.1 christos or %o7,%g5,%g5 1539 1.1 christos brz,a,pn %g5,.Lmdone_32 1540 1.1 christos mov 0,%i0 ! return failure 1541 1.1 christos #endif 1542 1.1 christos st %f1,[%g1+0*8+0] 1543 1.1 christos st %f0,[%g1+0*8+4] 1544 1.1 christos st %f3,[%g1+1*8+0] 1545 1.1 christos st %f2,[%g1+1*8+4] 1546 1.1 christos st %f5,[%g1+2*8+0] 1547 1.1 christos st %f4,[%g1+2*8+4] 1548 1.1 christos st %f7,[%g1+3*8+0] 1549 1.1 christos st %f6,[%g1+3*8+4] 1550 1.1 christos st %f9,[%g1+4*8+0] 1551 1.1 christos st %f8,[%g1+4*8+4] 1552 1.1 christos st %f11,[%g1+5*8+0] 1553 1.1 christos st %f10,[%g1+5*8+4] 1554 1.1 christos st %f13,[%g1+6*8+0] 1555 1.1 christos st %f12,[%g1+6*8+4] 1556 1.1 christos st %f15,[%g1+7*8+0] 1557 1.1 christos st %f14,[%g1+7*8+4] 1558 1.1 christos st %f17,[%g1+8*8+0] 1559 1.1 christos st %f16,[%g1+8*8+4] 1560 1.1 christos st %f19,[%g1+9*8+0] 1561 1.1 christos st %f18,[%g1+9*8+4] 1562 1.1 christos st %f21,[%g1+10*8+0] 1563 1.1 christos st %f20,[%g1+10*8+4] 1564 1.1 christos st %f23,[%g1+11*8+0] 1565 1.1 christos st %f22,[%g1+11*8+4] 1566 1.1 christos .word 0x81b00f1d !fsrc2 %f0,%f60,%f0 1567 1.1 christos st %f1,[%g1+12*8+0] 1568 1.1 christos st %f0,[%g1+12*8+4] 1569 1.1 christos .word 0x85b00f1f !fsrc2 %f0,%f62,%f2 1570 1.1 christos st %f3,[%g1+13*8+0] 1571 1.1 christos st %f2,[%g1+13*8+4] 1572 1.1 christos .word 0x89b00f18 !fsrc2 %f0,%f24,%f4 1573 1.1 christos st %f5,[%g1+14*8+0] 1574 1.1 christos st %f4,[%g1+14*8+4] 1575 1.1 christos .word 0x8db00f1a !fsrc2 %f0,%f26,%f6 1576 1.1 christos st %f7,[%g1+15*8+0] 1577 1.1 christos st %f6,[%g1+15*8+4] 1578 1.1 christos .word 0x81b00f1c !fsrc2 %f0,%f28,%f0 1579 1.1 christos st %f1,[%g1+16*8+0] 1580 1.1 christos st %f0,[%g1+16*8+4] 1581 1.1 christos .word 0x85b00f1e !fsrc2 %f0,%f30,%f2 1582 1.1 christos st %f3,[%g1+17*8+0] 1583 1.1 christos st %f2,[%g1+17*8+4] 1584 1.1 christos .word 0x89b00f01 !fsrc2 %f0,%f32,%f4 1585 1.1 christos st %f5,[%g1+18*8+0] 1586 1.1 christos st %f4,[%g1+18*8+4] 1587 1.1 christos .word 0x8db00f03 !fsrc2 %f0,%f34,%f6 1588 1.1 christos st %f7,[%g1+19*8+0] 1589 1.1 christos st %f6,[%g1+19*8+4] 1590 1.1 christos .word 0x81b00f05 !fsrc2 %f0,%f36,%f0 1591 1.1 christos st %f1,[%g1+20*8+0] 1592 1.1 christos st %f0,[%g1+20*8+4] 1593 1.1 christos .word 0x85b00f07 !fsrc2 %f0,%f38,%f2 1594 1.1 christos st %f3,[%g1+21*8+0] 1595 1.1 christos st %f2,[%g1+21*8+4] 1596 1.1 christos .word 0x89b00f09 !fsrc2 %f0,%f40,%f4 1597 1.1 christos st %f5,[%g1+22*8+0] 1598 1.1 christos st %f4,[%g1+22*8+4] 1599 1.1 christos .word 0x8db00f0b !fsrc2 %f0,%f42,%f6 1600 1.1 christos st %f7,[%g1+23*8+0] 1601 1.1 christos st %f6,[%g1+23*8+4] 1602 1.1 christos .word 0x81b00f0d !fsrc2 %f0,%f44,%f0 1603 1.1 christos st %f1,[%g1+24*8+0] 1604 1.1 christos st %f0,[%g1+24*8+4] 1605 1.1 christos .word 0x85b00f0f !fsrc2 %f0,%f46,%f2 1606 1.1 christos st %f3,[%g1+25*8+0] 1607 1.1 christos st %f2,[%g1+25*8+4] 1608 1.1 christos .word 0x89b00f11 !fsrc2 %f0,%f48,%f4 1609 1.1 christos st %f5,[%g1+26*8+0] 1610 1.1 christos st %f4,[%g1+26*8+4] 1611 1.1 christos .word 0x8db00f13 !fsrc2 %f0,%f50,%f6 1612 1.1 christos st %f7,[%g1+27*8+0] 1613 1.1 christos st %f6,[%g1+27*8+4] 1614 1.1 christos .word 0x81b00f15 !fsrc2 %f0,%f52,%f0 1615 1.1 christos st %f1,[%g1+28*8+0] 1616 1.1 christos st %f0,[%g1+28*8+4] 1617 1.1 christos .word 0x85b00f17 !fsrc2 %f0,%f54,%f2 1618 1.1 christos st %f3,[%g1+29*8+0] 1619 1.1 christos st %f2,[%g1+29*8+4] 1620 1.1 christos .word 0x89b00f19 !fsrc2 %f0,%f56,%f4 1621 1.1 christos st %f5,[%g1+30*8+0] 1622 1.1 christos st %f4,[%g1+30*8+4] 1623 1.1 christos .word 0x8db00f1b !fsrc2 %f0,%f58,%f6 1624 1.1 christos st %f7,[%g1+31*8+0] 1625 1.1 christos st %f6,[%g1+31*8+4] 1626 1.1 christos mov 1,%i0 ! return success 1627 1.1 christos .Lmdone_32: 1628 1.1 christos ret 1629 1.1 christos restore 1630 1.1 christos 1631 1.1 christos .Lmabort_32: 1632 1.1 christos restore 1633 1.1 christos restore 1634 1.1 christos restore 1635 1.1 christos restore 1636 1.1 christos restore 1637 1.1 christos .Lmabort1_32: 1638 1.1 christos restore 1639 1.1 christos 1640 1.1 christos mov 0,%i0 ! return failure 1641 1.1 christos ret 1642 1.1 christos restore 1643 1.1 christos 1644 1.1 christos .align 32 1645 1.1 christos .Lmsquare_32: 1646 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1647 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1648 1.1 christos .word 0x81b02940+32-1 ! montsqr 32-1 1649 1.1 christos ba .Lmresume_32 1650 1.1 christos nop 1651 1.1 christos .type bn_mul_mont_t4_32, #function 1652 1.1 christos .size bn_mul_mont_t4_32, .-bn_mul_mont_t4_32 1653 1.1 christos .globl bn_pwr5_mont_t4_8 1654 1.1 christos .align 32 1655 1.1 christos bn_pwr5_mont_t4_8: 1656 1.1 christos #ifdef __arch64__ 1657 1.1 christos mov 0,%g5 1658 1.1 christos mov -128,%g4 1659 1.1 christos #elif defined(SPARCV9_64BIT_STACK) 1660 1.1 christos SPARC_LOAD_ADDRESS_LEAF(OPENSSL_sparcv9cap_P,%g1,%g5) 1661 1.1 christos ld [%g1+0],%g1 ! OPENSSL_sparcv9_P[0] 1662 1.1 christos mov -2047,%g4 1663 1.1 christos and %g1,SPARCV9_64BIT_STACK,%g1 1664 1.1 christos movrz %g1,0,%g4 1665 1.1 christos mov -1,%g5 1666 1.1 christos add %g4,-128,%g4 1667 1.1 christos #else 1668 1.1 christos mov -1,%g5 1669 1.1 christos mov -128,%g4 1670 1.1 christos #endif 1671 1.1 christos sllx %g5,32,%g5 1672 1.1 christos save %sp,%g4,%sp 1673 1.1 christos #ifndef __arch64__ 1674 1.1 christos save %sp,-128,%sp ! warm it up 1675 1.1 christos save %sp,-128,%sp 1676 1.1 christos save %sp,-128,%sp 1677 1.1 christos save %sp,-128,%sp 1678 1.1 christos save %sp,-128,%sp 1679 1.1 christos save %sp,-128,%sp 1680 1.1 christos restore 1681 1.1 christos restore 1682 1.1 christos restore 1683 1.1 christos restore 1684 1.1 christos restore 1685 1.1 christos restore 1686 1.1 christos #endif 1687 1.1 christos and %sp,1,%g4 1688 1.1 christos or %g5,%fp,%fp 1689 1.1 christos or %g4,%g5,%g5 1690 1.1 christos 1691 1.1 christos ! copy arguments to global registers 1692 1.1 christos mov %i0,%g1 1693 1.1 christos mov %i1,%g2 1694 1.1 christos ld [%i2+0],%f1 ! load *n0 1695 1.1 christos ld [%i2+4],%f0 1696 1.1 christos mov %i3,%g3 1697 1.1 christos srl %i4,%g0,%i4 ! pack last arguments 1698 1.1 christos sllx %i5,32,%g4 1699 1.1 christos or %i4,%g4,%g4 1700 1.1 christos .word 0xbbb00f00 !fsrc2 %f0,%f0,%f60 1701 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1702 1.1 christos ldx [%g1+0*8],%l0 1703 1.1 christos ldx [%g1+1*8],%l1 1704 1.1 christos ldx [%g1+2*8],%l2 1705 1.1 christos ldx [%g1+3*8],%l3 1706 1.1 christos ldx [%g1+4*8],%l4 1707 1.1 christos ldx [%g1+5*8],%l5 1708 1.1 christos ldx [%g1+6*8],%l6 1709 1.1 christos ldx [%g1+7*8],%l7 1710 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1711 1.1 christos ldx [%g2+0*8],%l0 1712 1.1 christos ldx [%g2+1*8],%l1 1713 1.1 christos ldx [%g2+2*8],%l2 1714 1.1 christos ldx [%g2+3*8],%l3 1715 1.1 christos ldx [%g2+4*8],%l4 1716 1.1 christos ldx [%g2+5*8],%l5 1717 1.1 christos ldx [%g2+6*8],%l6 1718 1.1 christos ldx [%g2+7*8],%l7 1719 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1720 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1721 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1722 1.1 christos 1723 1.1 christos srlx %g4, 32, %o4 ! unpack %g4 1724 1.1 christos srl %g4, %g0, %o5 1725 1.1 christos sub %o4, 5, %o4 1726 1.1 christos mov %g3, %o7 1727 1.1 christos sllx %o4, 32, %g4 ! re-pack %g4 1728 1.1 christos or %o5, %g4, %g4 1729 1.1 christos srl %o5, %o4, %o5 1730 1.1 christos srl %o5, 2, %o4 1731 1.1 christos and %o5, 3, %o5 1732 1.1 christos and %o4, 7, %o4 1733 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 1734 1.1 christos add %o5, %o7, %o7 ! of the pwrtbl 1735 1.1 christos or %g0, 1, %o5 1736 1.1 christos sll %o5, %o4, %o4 1737 1.1 christos wr %o4, %g0, %ccr 1738 1.1 christos b .Lstride_8 1739 1.1 christos nop 1740 1.1 christos .align 16 1741 1.1 christos .Lstride_8: 1742 1.1 christos ldx [%o7+0*32], %i0 1743 1.1 christos ldx [%o7+8*32], %i1 1744 1.1 christos ldx [%o7+1*32], %o4 1745 1.1 christos ldx [%o7+9*32], %o5 1746 1.1 christos movvs %icc, %o4, %i0 1747 1.1 christos ldx [%o7+2*32], %o4 1748 1.1 christos movvs %icc, %o5, %i1 1749 1.1 christos ldx [%o7+10*32],%o5 1750 1.1 christos move %icc, %o4, %i0 1751 1.1 christos ldx [%o7+3*32], %o4 1752 1.1 christos move %icc, %o5, %i1 1753 1.1 christos ldx [%o7+11*32],%o5 1754 1.1 christos movneg %icc, %o4, %i0 1755 1.1 christos ldx [%o7+4*32], %o4 1756 1.1 christos movneg %icc, %o5, %i1 1757 1.1 christos ldx [%o7+12*32],%o5 1758 1.1 christos movcs %xcc, %o4, %i0 1759 1.1 christos ldx [%o7+5*32],%o4 1760 1.1 christos movcs %xcc, %o5, %i1 1761 1.1 christos ldx [%o7+13*32],%o5 1762 1.1 christos movvs %xcc, %o4, %i0 1763 1.1 christos ldx [%o7+6*32], %o4 1764 1.1 christos movvs %xcc, %o5, %i1 1765 1.1 christos ldx [%o7+14*32],%o5 1766 1.1 christos move %xcc, %o4, %i0 1767 1.1 christos ldx [%o7+7*32], %o4 1768 1.1 christos move %xcc, %o5, %i1 1769 1.1 christos ldx [%o7+15*32],%o5 1770 1.1 christos movneg %xcc, %o4, %i0 1771 1.1 christos add %o7,16*32, %o7 1772 1.1 christos movneg %xcc, %o5, %i1 1773 1.1 christos ldx [%o7+0*32], %i2 1774 1.1 christos ldx [%o7+8*32], %i3 1775 1.1 christos ldx [%o7+1*32], %o4 1776 1.1 christos ldx [%o7+9*32], %o5 1777 1.1 christos movvs %icc, %o4, %i2 1778 1.1 christos ldx [%o7+2*32], %o4 1779 1.1 christos movvs %icc, %o5, %i3 1780 1.1 christos ldx [%o7+10*32],%o5 1781 1.1 christos move %icc, %o4, %i2 1782 1.1 christos ldx [%o7+3*32], %o4 1783 1.1 christos move %icc, %o5, %i3 1784 1.1 christos ldx [%o7+11*32],%o5 1785 1.1 christos movneg %icc, %o4, %i2 1786 1.1 christos ldx [%o7+4*32], %o4 1787 1.1 christos movneg %icc, %o5, %i3 1788 1.1 christos ldx [%o7+12*32],%o5 1789 1.1 christos movcs %xcc, %o4, %i2 1790 1.1 christos ldx [%o7+5*32],%o4 1791 1.1 christos movcs %xcc, %o5, %i3 1792 1.1 christos ldx [%o7+13*32],%o5 1793 1.1 christos movvs %xcc, %o4, %i2 1794 1.1 christos ldx [%o7+6*32], %o4 1795 1.1 christos movvs %xcc, %o5, %i3 1796 1.1 christos ldx [%o7+14*32],%o5 1797 1.1 christos move %xcc, %o4, %i2 1798 1.1 christos ldx [%o7+7*32], %o4 1799 1.1 christos move %xcc, %o5, %i3 1800 1.1 christos ldx [%o7+15*32],%o5 1801 1.1 christos movneg %xcc, %o4, %i2 1802 1.1 christos add %o7,16*32, %o7 1803 1.1 christos movneg %xcc, %o5, %i3 1804 1.1 christos ldx [%o7+0*32], %i4 1805 1.1 christos ldx [%o7+8*32], %i5 1806 1.1 christos ldx [%o7+1*32], %o4 1807 1.1 christos ldx [%o7+9*32], %o5 1808 1.1 christos movvs %icc, %o4, %i4 1809 1.1 christos ldx [%o7+2*32], %o4 1810 1.1 christos movvs %icc, %o5, %i5 1811 1.1 christos ldx [%o7+10*32],%o5 1812 1.1 christos move %icc, %o4, %i4 1813 1.1 christos ldx [%o7+3*32], %o4 1814 1.1 christos move %icc, %o5, %i5 1815 1.1 christos ldx [%o7+11*32],%o5 1816 1.1 christos movneg %icc, %o4, %i4 1817 1.1 christos ldx [%o7+4*32], %o4 1818 1.1 christos movneg %icc, %o5, %i5 1819 1.1 christos ldx [%o7+12*32],%o5 1820 1.1 christos movcs %xcc, %o4, %i4 1821 1.1 christos ldx [%o7+5*32],%o4 1822 1.1 christos movcs %xcc, %o5, %i5 1823 1.1 christos ldx [%o7+13*32],%o5 1824 1.1 christos movvs %xcc, %o4, %i4 1825 1.1 christos ldx [%o7+6*32], %o4 1826 1.1 christos movvs %xcc, %o5, %i5 1827 1.1 christos ldx [%o7+14*32],%o5 1828 1.1 christos move %xcc, %o4, %i4 1829 1.1 christos ldx [%o7+7*32], %o4 1830 1.1 christos move %xcc, %o5, %i5 1831 1.1 christos ldx [%o7+15*32],%o5 1832 1.1 christos movneg %xcc, %o4, %i4 1833 1.1 christos add %o7,16*32, %o7 1834 1.1 christos movneg %xcc, %o5, %i5 1835 1.1 christos ldx [%o7+0*32], %l0 1836 1.1 christos ldx [%o7+8*32], %l1 1837 1.1 christos ldx [%o7+1*32], %o4 1838 1.1 christos ldx [%o7+9*32], %o5 1839 1.1 christos movvs %icc, %o4, %l0 1840 1.1 christos ldx [%o7+2*32], %o4 1841 1.1 christos movvs %icc, %o5, %l1 1842 1.1 christos ldx [%o7+10*32],%o5 1843 1.1 christos move %icc, %o4, %l0 1844 1.1 christos ldx [%o7+3*32], %o4 1845 1.1 christos move %icc, %o5, %l1 1846 1.1 christos ldx [%o7+11*32],%o5 1847 1.1 christos movneg %icc, %o4, %l0 1848 1.1 christos ldx [%o7+4*32], %o4 1849 1.1 christos movneg %icc, %o5, %l1 1850 1.1 christos ldx [%o7+12*32],%o5 1851 1.1 christos movcs %xcc, %o4, %l0 1852 1.1 christos ldx [%o7+5*32],%o4 1853 1.1 christos movcs %xcc, %o5, %l1 1854 1.1 christos ldx [%o7+13*32],%o5 1855 1.1 christos movvs %xcc, %o4, %l0 1856 1.1 christos ldx [%o7+6*32], %o4 1857 1.1 christos movvs %xcc, %o5, %l1 1858 1.1 christos ldx [%o7+14*32],%o5 1859 1.1 christos move %xcc, %o4, %l0 1860 1.1 christos ldx [%o7+7*32], %o4 1861 1.1 christos move %xcc, %o5, %l1 1862 1.1 christos ldx [%o7+15*32],%o5 1863 1.1 christos movneg %xcc, %o4, %l0 1864 1.1 christos add %o7,16*32, %o7 1865 1.1 christos movneg %xcc, %o5, %l1 1866 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 1867 1.1 christos srax %g4, 32, %o4 ! unpack %g4 1868 1.1 christos srl %g4, %g0, %o5 1869 1.1 christos sub %o4, 5, %o4 1870 1.1 christos mov %g3, %i7 1871 1.1 christos sllx %o4, 32, %g4 ! re-pack %g4 1872 1.1 christos or %o5, %g4, %g4 1873 1.1 christos srl %o5, %o4, %o5 1874 1.1 christos srl %o5, 2, %o4 1875 1.1 christos and %o5, 3, %o5 1876 1.1 christos and %o4, 7, %o4 1877 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 1878 1.1 christos add %o5, %i7, %i7 ! of the pwrtbl 1879 1.1 christos or %g0, 1, %o5 1880 1.1 christos sll %o5, %o4, %o4 1881 1.1 christos .word 0x81b02940+8-1 ! montsqr 8-1 1882 1.1 christos fbu,pn %fcc3,.Labort_8 1883 1.1 christos #ifndef __arch64__ 1884 1.1 christos and %fp,%g5,%g5 1885 1.1 christos brz,pn %g5,.Labort_8 1886 1.1 christos #endif 1887 1.1 christos nop 1888 1.1 christos .word 0x81b02940+8-1 ! montsqr 8-1 1889 1.1 christos fbu,pn %fcc3,.Labort_8 1890 1.1 christos #ifndef __arch64__ 1891 1.1 christos and %fp,%g5,%g5 1892 1.1 christos brz,pn %g5,.Labort_8 1893 1.1 christos #endif 1894 1.1 christos nop 1895 1.1 christos .word 0x81b02940+8-1 ! montsqr 8-1 1896 1.1 christos fbu,pn %fcc3,.Labort_8 1897 1.1 christos #ifndef __arch64__ 1898 1.1 christos and %fp,%g5,%g5 1899 1.1 christos brz,pn %g5,.Labort_8 1900 1.1 christos #endif 1901 1.1 christos nop 1902 1.1 christos .word 0x81b02940+8-1 ! montsqr 8-1 1903 1.1 christos fbu,pn %fcc3,.Labort_8 1904 1.1 christos #ifndef __arch64__ 1905 1.1 christos and %fp,%g5,%g5 1906 1.1 christos brz,pn %g5,.Labort_8 1907 1.1 christos #endif 1908 1.1 christos nop 1909 1.1 christos .word 0x81b02940+8-1 ! montsqr 8-1 1910 1.1 christos fbu,pn %fcc3,.Labort_8 1911 1.1 christos #ifndef __arch64__ 1912 1.1 christos and %fp,%g5,%g5 1913 1.1 christos brz,pn %g5,.Labort_8 1914 1.1 christos #endif 1915 1.1 christos nop 1916 1.1 christos wr %o4, %g0, %ccr 1917 1.1 christos .word 0x81b02920+8-1 ! montmul 8-1 1918 1.1 christos fbu,pn %fcc3,.Labort_8 1919 1.1 christos #ifndef __arch64__ 1920 1.1 christos and %fp,%g5,%g5 1921 1.1 christos brz,pn %g5,.Labort_8 1922 1.1 christos #endif 1923 1.1 christos 1924 1.1 christos srax %g4, 32, %o4 1925 1.1 christos #ifdef __arch64__ 1926 1.1 christos brgez %o4,.Lstride_8 1927 1.1 christos restore 1928 1.1 christos restore 1929 1.1 christos restore 1930 1.1 christos restore 1931 1.1 christos restore 1932 1.1 christos #else 1933 1.1 christos brgez %o4,.Lstride_8 1934 1.1 christos restore; and %fp,%g5,%g5 1935 1.1 christos restore; and %fp,%g5,%g5 1936 1.1 christos restore; and %fp,%g5,%g5 1937 1.1 christos restore; and %fp,%g5,%g5 1938 1.1 christos brz,pn %g5,.Labort1_8 1939 1.1 christos restore 1940 1.1 christos #endif 1941 1.1 christos .word 0x81b02310 !movxtod %l0,%f0 1942 1.1 christos .word 0x85b02311 !movxtod %l1,%f2 1943 1.1 christos .word 0x89b02312 !movxtod %l2,%f4 1944 1.1 christos .word 0x8db02313 !movxtod %l3,%f6 1945 1.1 christos .word 0x91b02314 !movxtod %l4,%f8 1946 1.1 christos .word 0x95b02315 !movxtod %l5,%f10 1947 1.1 christos .word 0x99b02316 !movxtod %l6,%f12 1948 1.1 christos .word 0x9db02317 !movxtod %l7,%f14 1949 1.1 christos #ifdef __arch64__ 1950 1.1 christos restore 1951 1.1 christos #else 1952 1.1 christos and %fp,%g5,%g5 1953 1.1 christos restore 1954 1.1 christos and %g5,1,%o7 1955 1.1 christos and %fp,%g5,%g5 1956 1.1 christos srl %fp,0,%fp ! just in case? 1957 1.1 christos or %o7,%g5,%g5 1958 1.1 christos brz,a,pn %g5,.Ldone_8 1959 1.1 christos mov 0,%i0 ! return failure 1960 1.1 christos #endif 1961 1.1 christos std %f0,[%g1+0*8] 1962 1.1 christos std %f2,[%g1+1*8] 1963 1.1 christos std %f4,[%g1+2*8] 1964 1.1 christos std %f6,[%g1+3*8] 1965 1.1 christos std %f8,[%g1+4*8] 1966 1.1 christos std %f10,[%g1+5*8] 1967 1.1 christos std %f12,[%g1+6*8] 1968 1.1 christos std %f14,[%g1+7*8] 1969 1.1 christos mov 1,%i0 ! return success 1970 1.1 christos .Ldone_8: 1971 1.1 christos ret 1972 1.1 christos restore 1973 1.1 christos 1974 1.1 christos .Labort_8: 1975 1.1 christos restore 1976 1.1 christos restore 1977 1.1 christos restore 1978 1.1 christos restore 1979 1.1 christos restore 1980 1.1 christos .Labort1_8: 1981 1.1 christos restore 1982 1.1 christos 1983 1.1 christos mov 0,%i0 ! return failure 1984 1.1 christos ret 1985 1.1 christos restore 1986 1.1 christos .type bn_pwr5_mont_t4_8, #function 1987 1.1 christos .size bn_pwr5_mont_t4_8, .-bn_pwr5_mont_t4_8 1988 1.1 christos .globl bn_pwr5_mont_t4_16 1989 1.1 christos .align 32 1990 1.1 christos bn_pwr5_mont_t4_16: 1991 1.1 christos #ifdef __arch64__ 1992 1.1 christos mov 0,%g5 1993 1.1 christos mov -128,%g4 1994 1.1 christos #elif defined(SPARCV9_64BIT_STACK) 1995 1.1 christos SPARC_LOAD_ADDRESS_LEAF(OPENSSL_sparcv9cap_P,%g1,%g5) 1996 1.1 christos ld [%g1+0],%g1 ! OPENSSL_sparcv9_P[0] 1997 1.1 christos mov -2047,%g4 1998 1.1 christos and %g1,SPARCV9_64BIT_STACK,%g1 1999 1.1 christos movrz %g1,0,%g4 2000 1.1 christos mov -1,%g5 2001 1.1 christos add %g4,-128,%g4 2002 1.1 christos #else 2003 1.1 christos mov -1,%g5 2004 1.1 christos mov -128,%g4 2005 1.1 christos #endif 2006 1.1 christos sllx %g5,32,%g5 2007 1.1 christos save %sp,%g4,%sp 2008 1.1 christos #ifndef __arch64__ 2009 1.1 christos save %sp,-128,%sp ! warm it up 2010 1.1 christos save %sp,-128,%sp 2011 1.1 christos save %sp,-128,%sp 2012 1.1 christos save %sp,-128,%sp 2013 1.1 christos save %sp,-128,%sp 2014 1.1 christos save %sp,-128,%sp 2015 1.1 christos restore 2016 1.1 christos restore 2017 1.1 christos restore 2018 1.1 christos restore 2019 1.1 christos restore 2020 1.1 christos restore 2021 1.1 christos #endif 2022 1.1 christos and %sp,1,%g4 2023 1.1 christos or %g5,%fp,%fp 2024 1.1 christos or %g4,%g5,%g5 2025 1.1 christos 2026 1.1 christos ! copy arguments to global registers 2027 1.1 christos mov %i0,%g1 2028 1.1 christos mov %i1,%g2 2029 1.1 christos ld [%i2+0],%f1 ! load *n0 2030 1.1 christos ld [%i2+4],%f0 2031 1.1 christos mov %i3,%g3 2032 1.1 christos srl %i4,%g0,%i4 ! pack last arguments 2033 1.1 christos sllx %i5,32,%g4 2034 1.1 christos or %i4,%g4,%g4 2035 1.1 christos .word 0xbbb00f00 !fsrc2 %f0,%f0,%f60 2036 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2037 1.1 christos ldx [%g1+0*8],%l0 2038 1.1 christos ldx [%g1+1*8],%l1 2039 1.1 christos ldx [%g1+2*8],%l2 2040 1.1 christos ldx [%g1+3*8],%l3 2041 1.1 christos ldx [%g1+4*8],%l4 2042 1.1 christos ldx [%g1+5*8],%l5 2043 1.1 christos ldx [%g1+6*8],%l6 2044 1.1 christos ldx [%g1+7*8],%l7 2045 1.1 christos ldx [%g1+8*8],%o0 2046 1.1 christos ldx [%g1+9*8],%o1 2047 1.1 christos ldx [%g1+10*8],%o2 2048 1.1 christos ldx [%g1+11*8],%o3 2049 1.1 christos ldx [%g1+12*8],%o4 2050 1.1 christos ldx [%g1+13*8],%o5 2051 1.1 christos ldd [%g1+14*8],%f24 2052 1.1 christos ldd [%g1+15*8],%f26 2053 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2054 1.1 christos ldx [%g2+0*8],%l0 2055 1.1 christos ldx [%g2+1*8],%l1 2056 1.1 christos ldx [%g2+2*8],%l2 2057 1.1 christos ldx [%g2+3*8],%l3 2058 1.1 christos ldx [%g2+4*8],%l4 2059 1.1 christos ldx [%g2+5*8],%l5 2060 1.1 christos ldx [%g2+6*8],%l6 2061 1.1 christos ldx [%g2+7*8],%l7 2062 1.1 christos ldx [%g2+8*8],%o0 2063 1.1 christos ldx [%g2+9*8],%o1 2064 1.1 christos ldx [%g2+10*8],%o2 2065 1.1 christos ldx [%g2+11*8],%o3 2066 1.1 christos ldx [%g2+12*8],%o4 2067 1.1 christos ldx [%g2+13*8],%o5 2068 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2069 1.1 christos ldx [%g2+14*8],%l0 2070 1.1 christos ldx [%g2+15*8],%l1 2071 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2072 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2073 1.1 christos 2074 1.1 christos srlx %g4, 32, %o4 ! unpack %g4 2075 1.1 christos srl %g4, %g0, %o5 2076 1.1 christos sub %o4, 5, %o4 2077 1.1 christos mov %g3, %o7 2078 1.1 christos sllx %o4, 32, %g4 ! re-pack %g4 2079 1.1 christos or %o5, %g4, %g4 2080 1.1 christos srl %o5, %o4, %o5 2081 1.1 christos srl %o5, 2, %o4 2082 1.1 christos and %o5, 3, %o5 2083 1.1 christos and %o4, 7, %o4 2084 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 2085 1.1 christos add %o5, %o7, %o7 ! of the pwrtbl 2086 1.1 christos or %g0, 1, %o5 2087 1.1 christos sll %o5, %o4, %o4 2088 1.1 christos wr %o4, %g0, %ccr 2089 1.1 christos b .Lstride_16 2090 1.1 christos nop 2091 1.1 christos .align 16 2092 1.1 christos .Lstride_16: 2093 1.1 christos ldx [%o7+0*32], %i0 2094 1.1 christos ldx [%o7+8*32], %i1 2095 1.1 christos ldx [%o7+1*32], %o4 2096 1.1 christos ldx [%o7+9*32], %o5 2097 1.1 christos movvs %icc, %o4, %i0 2098 1.1 christos ldx [%o7+2*32], %o4 2099 1.1 christos movvs %icc, %o5, %i1 2100 1.1 christos ldx [%o7+10*32],%o5 2101 1.1 christos move %icc, %o4, %i0 2102 1.1 christos ldx [%o7+3*32], %o4 2103 1.1 christos move %icc, %o5, %i1 2104 1.1 christos ldx [%o7+11*32],%o5 2105 1.1 christos movneg %icc, %o4, %i0 2106 1.1 christos ldx [%o7+4*32], %o4 2107 1.1 christos movneg %icc, %o5, %i1 2108 1.1 christos ldx [%o7+12*32],%o5 2109 1.1 christos movcs %xcc, %o4, %i0 2110 1.1 christos ldx [%o7+5*32],%o4 2111 1.1 christos movcs %xcc, %o5, %i1 2112 1.1 christos ldx [%o7+13*32],%o5 2113 1.1 christos movvs %xcc, %o4, %i0 2114 1.1 christos ldx [%o7+6*32], %o4 2115 1.1 christos movvs %xcc, %o5, %i1 2116 1.1 christos ldx [%o7+14*32],%o5 2117 1.1 christos move %xcc, %o4, %i0 2118 1.1 christos ldx [%o7+7*32], %o4 2119 1.1 christos move %xcc, %o5, %i1 2120 1.1 christos ldx [%o7+15*32],%o5 2121 1.1 christos movneg %xcc, %o4, %i0 2122 1.1 christos add %o7,16*32, %o7 2123 1.1 christos movneg %xcc, %o5, %i1 2124 1.1 christos ldx [%o7+0*32], %i2 2125 1.1 christos ldx [%o7+8*32], %i3 2126 1.1 christos ldx [%o7+1*32], %o4 2127 1.1 christos ldx [%o7+9*32], %o5 2128 1.1 christos movvs %icc, %o4, %i2 2129 1.1 christos ldx [%o7+2*32], %o4 2130 1.1 christos movvs %icc, %o5, %i3 2131 1.1 christos ldx [%o7+10*32],%o5 2132 1.1 christos move %icc, %o4, %i2 2133 1.1 christos ldx [%o7+3*32], %o4 2134 1.1 christos move %icc, %o5, %i3 2135 1.1 christos ldx [%o7+11*32],%o5 2136 1.1 christos movneg %icc, %o4, %i2 2137 1.1 christos ldx [%o7+4*32], %o4 2138 1.1 christos movneg %icc, %o5, %i3 2139 1.1 christos ldx [%o7+12*32],%o5 2140 1.1 christos movcs %xcc, %o4, %i2 2141 1.1 christos ldx [%o7+5*32],%o4 2142 1.1 christos movcs %xcc, %o5, %i3 2143 1.1 christos ldx [%o7+13*32],%o5 2144 1.1 christos movvs %xcc, %o4, %i2 2145 1.1 christos ldx [%o7+6*32], %o4 2146 1.1 christos movvs %xcc, %o5, %i3 2147 1.1 christos ldx [%o7+14*32],%o5 2148 1.1 christos move %xcc, %o4, %i2 2149 1.1 christos ldx [%o7+7*32], %o4 2150 1.1 christos move %xcc, %o5, %i3 2151 1.1 christos ldx [%o7+15*32],%o5 2152 1.1 christos movneg %xcc, %o4, %i2 2153 1.1 christos add %o7,16*32, %o7 2154 1.1 christos movneg %xcc, %o5, %i3 2155 1.1 christos ldx [%o7+0*32], %i4 2156 1.1 christos ldx [%o7+8*32], %i5 2157 1.1 christos ldx [%o7+1*32], %o4 2158 1.1 christos ldx [%o7+9*32], %o5 2159 1.1 christos movvs %icc, %o4, %i4 2160 1.1 christos ldx [%o7+2*32], %o4 2161 1.1 christos movvs %icc, %o5, %i5 2162 1.1 christos ldx [%o7+10*32],%o5 2163 1.1 christos move %icc, %o4, %i4 2164 1.1 christos ldx [%o7+3*32], %o4 2165 1.1 christos move %icc, %o5, %i5 2166 1.1 christos ldx [%o7+11*32],%o5 2167 1.1 christos movneg %icc, %o4, %i4 2168 1.1 christos ldx [%o7+4*32], %o4 2169 1.1 christos movneg %icc, %o5, %i5 2170 1.1 christos ldx [%o7+12*32],%o5 2171 1.1 christos movcs %xcc, %o4, %i4 2172 1.1 christos ldx [%o7+5*32],%o4 2173 1.1 christos movcs %xcc, %o5, %i5 2174 1.1 christos ldx [%o7+13*32],%o5 2175 1.1 christos movvs %xcc, %o4, %i4 2176 1.1 christos ldx [%o7+6*32], %o4 2177 1.1 christos movvs %xcc, %o5, %i5 2178 1.1 christos ldx [%o7+14*32],%o5 2179 1.1 christos move %xcc, %o4, %i4 2180 1.1 christos ldx [%o7+7*32], %o4 2181 1.1 christos move %xcc, %o5, %i5 2182 1.1 christos ldx [%o7+15*32],%o5 2183 1.1 christos movneg %xcc, %o4, %i4 2184 1.1 christos add %o7,16*32, %o7 2185 1.1 christos movneg %xcc, %o5, %i5 2186 1.1 christos ldx [%o7+0*32], %l0 2187 1.1 christos ldx [%o7+8*32], %l1 2188 1.1 christos ldx [%o7+1*32], %o4 2189 1.1 christos ldx [%o7+9*32], %o5 2190 1.1 christos movvs %icc, %o4, %l0 2191 1.1 christos ldx [%o7+2*32], %o4 2192 1.1 christos movvs %icc, %o5, %l1 2193 1.1 christos ldx [%o7+10*32],%o5 2194 1.1 christos move %icc, %o4, %l0 2195 1.1 christos ldx [%o7+3*32], %o4 2196 1.1 christos move %icc, %o5, %l1 2197 1.1 christos ldx [%o7+11*32],%o5 2198 1.1 christos movneg %icc, %o4, %l0 2199 1.1 christos ldx [%o7+4*32], %o4 2200 1.1 christos movneg %icc, %o5, %l1 2201 1.1 christos ldx [%o7+12*32],%o5 2202 1.1 christos movcs %xcc, %o4, %l0 2203 1.1 christos ldx [%o7+5*32],%o4 2204 1.1 christos movcs %xcc, %o5, %l1 2205 1.1 christos ldx [%o7+13*32],%o5 2206 1.1 christos movvs %xcc, %o4, %l0 2207 1.1 christos ldx [%o7+6*32], %o4 2208 1.1 christos movvs %xcc, %o5, %l1 2209 1.1 christos ldx [%o7+14*32],%o5 2210 1.1 christos move %xcc, %o4, %l0 2211 1.1 christos ldx [%o7+7*32], %o4 2212 1.1 christos move %xcc, %o5, %l1 2213 1.1 christos ldx [%o7+15*32],%o5 2214 1.1 christos movneg %xcc, %o4, %l0 2215 1.1 christos add %o7,16*32, %o7 2216 1.1 christos movneg %xcc, %o5, %l1 2217 1.1 christos ldx [%o7+0*32], %l2 2218 1.1 christos ldx [%o7+8*32], %l3 2219 1.1 christos ldx [%o7+1*32], %o4 2220 1.1 christos ldx [%o7+9*32], %o5 2221 1.1 christos movvs %icc, %o4, %l2 2222 1.1 christos ldx [%o7+2*32], %o4 2223 1.1 christos movvs %icc, %o5, %l3 2224 1.1 christos ldx [%o7+10*32],%o5 2225 1.1 christos move %icc, %o4, %l2 2226 1.1 christos ldx [%o7+3*32], %o4 2227 1.1 christos move %icc, %o5, %l3 2228 1.1 christos ldx [%o7+11*32],%o5 2229 1.1 christos movneg %icc, %o4, %l2 2230 1.1 christos ldx [%o7+4*32], %o4 2231 1.1 christos movneg %icc, %o5, %l3 2232 1.1 christos ldx [%o7+12*32],%o5 2233 1.1 christos movcs %xcc, %o4, %l2 2234 1.1 christos ldx [%o7+5*32],%o4 2235 1.1 christos movcs %xcc, %o5, %l3 2236 1.1 christos ldx [%o7+13*32],%o5 2237 1.1 christos movvs %xcc, %o4, %l2 2238 1.1 christos ldx [%o7+6*32], %o4 2239 1.1 christos movvs %xcc, %o5, %l3 2240 1.1 christos ldx [%o7+14*32],%o5 2241 1.1 christos move %xcc, %o4, %l2 2242 1.1 christos ldx [%o7+7*32], %o4 2243 1.1 christos move %xcc, %o5, %l3 2244 1.1 christos ldx [%o7+15*32],%o5 2245 1.1 christos movneg %xcc, %o4, %l2 2246 1.1 christos add %o7,16*32, %o7 2247 1.1 christos movneg %xcc, %o5, %l3 2248 1.1 christos ldx [%o7+0*32], %l4 2249 1.1 christos ldx [%o7+8*32], %l5 2250 1.1 christos ldx [%o7+1*32], %o4 2251 1.1 christos ldx [%o7+9*32], %o5 2252 1.1 christos movvs %icc, %o4, %l4 2253 1.1 christos ldx [%o7+2*32], %o4 2254 1.1 christos movvs %icc, %o5, %l5 2255 1.1 christos ldx [%o7+10*32],%o5 2256 1.1 christos move %icc, %o4, %l4 2257 1.1 christos ldx [%o7+3*32], %o4 2258 1.1 christos move %icc, %o5, %l5 2259 1.1 christos ldx [%o7+11*32],%o5 2260 1.1 christos movneg %icc, %o4, %l4 2261 1.1 christos ldx [%o7+4*32], %o4 2262 1.1 christos movneg %icc, %o5, %l5 2263 1.1 christos ldx [%o7+12*32],%o5 2264 1.1 christos movcs %xcc, %o4, %l4 2265 1.1 christos ldx [%o7+5*32],%o4 2266 1.1 christos movcs %xcc, %o5, %l5 2267 1.1 christos ldx [%o7+13*32],%o5 2268 1.1 christos movvs %xcc, %o4, %l4 2269 1.1 christos ldx [%o7+6*32], %o4 2270 1.1 christos movvs %xcc, %o5, %l5 2271 1.1 christos ldx [%o7+14*32],%o5 2272 1.1 christos move %xcc, %o4, %l4 2273 1.1 christos ldx [%o7+7*32], %o4 2274 1.1 christos move %xcc, %o5, %l5 2275 1.1 christos ldx [%o7+15*32],%o5 2276 1.1 christos movneg %xcc, %o4, %l4 2277 1.1 christos add %o7,16*32, %o7 2278 1.1 christos movneg %xcc, %o5, %l5 2279 1.1 christos ldx [%o7+0*32], %l6 2280 1.1 christos ldx [%o7+8*32], %l7 2281 1.1 christos ldx [%o7+1*32], %o4 2282 1.1 christos ldx [%o7+9*32], %o5 2283 1.1 christos movvs %icc, %o4, %l6 2284 1.1 christos ldx [%o7+2*32], %o4 2285 1.1 christos movvs %icc, %o5, %l7 2286 1.1 christos ldx [%o7+10*32],%o5 2287 1.1 christos move %icc, %o4, %l6 2288 1.1 christos ldx [%o7+3*32], %o4 2289 1.1 christos move %icc, %o5, %l7 2290 1.1 christos ldx [%o7+11*32],%o5 2291 1.1 christos movneg %icc, %o4, %l6 2292 1.1 christos ldx [%o7+4*32], %o4 2293 1.1 christos movneg %icc, %o5, %l7 2294 1.1 christos ldx [%o7+12*32],%o5 2295 1.1 christos movcs %xcc, %o4, %l6 2296 1.1 christos ldx [%o7+5*32],%o4 2297 1.1 christos movcs %xcc, %o5, %l7 2298 1.1 christos ldx [%o7+13*32],%o5 2299 1.1 christos movvs %xcc, %o4, %l6 2300 1.1 christos ldx [%o7+6*32], %o4 2301 1.1 christos movvs %xcc, %o5, %l7 2302 1.1 christos ldx [%o7+14*32],%o5 2303 1.1 christos move %xcc, %o4, %l6 2304 1.1 christos ldx [%o7+7*32], %o4 2305 1.1 christos move %xcc, %o5, %l7 2306 1.1 christos ldx [%o7+15*32],%o5 2307 1.1 christos movneg %xcc, %o4, %l6 2308 1.1 christos add %o7,16*32, %o7 2309 1.1 christos movneg %xcc, %o5, %l7 2310 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2311 1.1 christos ldx [%i7+0*32], %i0 2312 1.1 christos ldx [%i7+8*32], %i1 2313 1.1 christos ldx [%i7+1*32], %o4 2314 1.1 christos ldx [%i7+9*32], %o5 2315 1.1 christos movvs %icc, %o4, %i0 2316 1.1 christos ldx [%i7+2*32], %o4 2317 1.1 christos movvs %icc, %o5, %i1 2318 1.1 christos ldx [%i7+10*32],%o5 2319 1.1 christos move %icc, %o4, %i0 2320 1.1 christos ldx [%i7+3*32], %o4 2321 1.1 christos move %icc, %o5, %i1 2322 1.1 christos ldx [%i7+11*32],%o5 2323 1.1 christos movneg %icc, %o4, %i0 2324 1.1 christos ldx [%i7+4*32], %o4 2325 1.1 christos movneg %icc, %o5, %i1 2326 1.1 christos ldx [%i7+12*32],%o5 2327 1.1 christos movcs %xcc, %o4, %i0 2328 1.1 christos ldx [%i7+5*32],%o4 2329 1.1 christos movcs %xcc, %o5, %i1 2330 1.1 christos ldx [%i7+13*32],%o5 2331 1.1 christos movvs %xcc, %o4, %i0 2332 1.1 christos ldx [%i7+6*32], %o4 2333 1.1 christos movvs %xcc, %o5, %i1 2334 1.1 christos ldx [%i7+14*32],%o5 2335 1.1 christos move %xcc, %o4, %i0 2336 1.1 christos ldx [%i7+7*32], %o4 2337 1.1 christos move %xcc, %o5, %i1 2338 1.1 christos ldx [%i7+15*32],%o5 2339 1.1 christos movneg %xcc, %o4, %i0 2340 1.1 christos add %i7,16*32, %i7 2341 1.1 christos movneg %xcc, %o5, %i1 2342 1.1 christos srax %g4, 32, %o4 ! unpack %g4 2343 1.1 christos srl %g4, %g0, %o5 2344 1.1 christos sub %o4, 5, %o4 2345 1.1 christos mov %g3, %i7 2346 1.1 christos sllx %o4, 32, %g4 ! re-pack %g4 2347 1.1 christos or %o5, %g4, %g4 2348 1.1 christos srl %o5, %o4, %o5 2349 1.1 christos srl %o5, 2, %o4 2350 1.1 christos and %o5, 3, %o5 2351 1.1 christos and %o4, 7, %o4 2352 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 2353 1.1 christos add %o5, %i7, %i7 ! of the pwrtbl 2354 1.1 christos or %g0, 1, %o5 2355 1.1 christos sll %o5, %o4, %o4 2356 1.1 christos .word 0x81b02940+16-1 ! montsqr 16-1 2357 1.1 christos fbu,pn %fcc3,.Labort_16 2358 1.1 christos #ifndef __arch64__ 2359 1.1 christos and %fp,%g5,%g5 2360 1.1 christos brz,pn %g5,.Labort_16 2361 1.1 christos #endif 2362 1.1 christos nop 2363 1.1 christos .word 0x81b02940+16-1 ! montsqr 16-1 2364 1.1 christos fbu,pn %fcc3,.Labort_16 2365 1.1 christos #ifndef __arch64__ 2366 1.1 christos and %fp,%g5,%g5 2367 1.1 christos brz,pn %g5,.Labort_16 2368 1.1 christos #endif 2369 1.1 christos nop 2370 1.1 christos .word 0x81b02940+16-1 ! montsqr 16-1 2371 1.1 christos fbu,pn %fcc3,.Labort_16 2372 1.1 christos #ifndef __arch64__ 2373 1.1 christos and %fp,%g5,%g5 2374 1.1 christos brz,pn %g5,.Labort_16 2375 1.1 christos #endif 2376 1.1 christos nop 2377 1.1 christos .word 0x81b02940+16-1 ! montsqr 16-1 2378 1.1 christos fbu,pn %fcc3,.Labort_16 2379 1.1 christos #ifndef __arch64__ 2380 1.1 christos and %fp,%g5,%g5 2381 1.1 christos brz,pn %g5,.Labort_16 2382 1.1 christos #endif 2383 1.1 christos nop 2384 1.1 christos .word 0x81b02940+16-1 ! montsqr 16-1 2385 1.1 christos fbu,pn %fcc3,.Labort_16 2386 1.1 christos #ifndef __arch64__ 2387 1.1 christos and %fp,%g5,%g5 2388 1.1 christos brz,pn %g5,.Labort_16 2389 1.1 christos #endif 2390 1.1 christos nop 2391 1.1 christos wr %o4, %g0, %ccr 2392 1.1 christos .word 0x81b02920+16-1 ! montmul 16-1 2393 1.1 christos fbu,pn %fcc3,.Labort_16 2394 1.1 christos #ifndef __arch64__ 2395 1.1 christos and %fp,%g5,%g5 2396 1.1 christos brz,pn %g5,.Labort_16 2397 1.1 christos #endif 2398 1.1 christos 2399 1.1 christos srax %g4, 32, %o4 2400 1.1 christos #ifdef __arch64__ 2401 1.1 christos brgez %o4,.Lstride_16 2402 1.1 christos restore 2403 1.1 christos restore 2404 1.1 christos restore 2405 1.1 christos restore 2406 1.1 christos restore 2407 1.1 christos #else 2408 1.1 christos brgez %o4,.Lstride_16 2409 1.1 christos restore; and %fp,%g5,%g5 2410 1.1 christos restore; and %fp,%g5,%g5 2411 1.1 christos restore; and %fp,%g5,%g5 2412 1.1 christos restore; and %fp,%g5,%g5 2413 1.1 christos brz,pn %g5,.Labort1_16 2414 1.1 christos restore 2415 1.1 christos #endif 2416 1.1 christos .word 0x81b02310 !movxtod %l0,%f0 2417 1.1 christos .word 0x85b02311 !movxtod %l1,%f2 2418 1.1 christos .word 0x89b02312 !movxtod %l2,%f4 2419 1.1 christos .word 0x8db02313 !movxtod %l3,%f6 2420 1.1 christos .word 0x91b02314 !movxtod %l4,%f8 2421 1.1 christos .word 0x95b02315 !movxtod %l5,%f10 2422 1.1 christos .word 0x99b02316 !movxtod %l6,%f12 2423 1.1 christos .word 0x9db02317 !movxtod %l7,%f14 2424 1.1 christos .word 0xa1b02308 !movxtod %o0,%f16 2425 1.1 christos .word 0xa5b02309 !movxtod %o1,%f18 2426 1.1 christos .word 0xa9b0230a !movxtod %o2,%f20 2427 1.1 christos .word 0xadb0230b !movxtod %o3,%f22 2428 1.1 christos .word 0xbbb0230c !movxtod %o4,%f60 2429 1.1 christos .word 0xbfb0230d !movxtod %o5,%f62 2430 1.1 christos #ifdef __arch64__ 2431 1.1 christos restore 2432 1.1 christos #else 2433 1.1 christos and %fp,%g5,%g5 2434 1.1 christos restore 2435 1.1 christos and %g5,1,%o7 2436 1.1 christos and %fp,%g5,%g5 2437 1.1 christos srl %fp,0,%fp ! just in case? 2438 1.1 christos or %o7,%g5,%g5 2439 1.1 christos brz,a,pn %g5,.Ldone_16 2440 1.1 christos mov 0,%i0 ! return failure 2441 1.1 christos #endif 2442 1.1 christos std %f0,[%g1+0*8] 2443 1.1 christos std %f2,[%g1+1*8] 2444 1.1 christos std %f4,[%g1+2*8] 2445 1.1 christos std %f6,[%g1+3*8] 2446 1.1 christos std %f8,[%g1+4*8] 2447 1.1 christos std %f10,[%g1+5*8] 2448 1.1 christos std %f12,[%g1+6*8] 2449 1.1 christos std %f14,[%g1+7*8] 2450 1.1 christos std %f16,[%g1+8*8] 2451 1.1 christos std %f18,[%g1+9*8] 2452 1.1 christos std %f20,[%g1+10*8] 2453 1.1 christos std %f22,[%g1+11*8] 2454 1.1 christos std %f60,[%g1+12*8] 2455 1.1 christos std %f62,[%g1+13*8] 2456 1.1 christos std %f24,[%g1+14*8] 2457 1.1 christos std %f26,[%g1+15*8] 2458 1.1 christos mov 1,%i0 ! return success 2459 1.1 christos .Ldone_16: 2460 1.1 christos ret 2461 1.1 christos restore 2462 1.1 christos 2463 1.1 christos .Labort_16: 2464 1.1 christos restore 2465 1.1 christos restore 2466 1.1 christos restore 2467 1.1 christos restore 2468 1.1 christos restore 2469 1.1 christos .Labort1_16: 2470 1.1 christos restore 2471 1.1 christos 2472 1.1 christos mov 0,%i0 ! return failure 2473 1.1 christos ret 2474 1.1 christos restore 2475 1.1 christos .type bn_pwr5_mont_t4_16, #function 2476 1.1 christos .size bn_pwr5_mont_t4_16, .-bn_pwr5_mont_t4_16 2477 1.1 christos .globl bn_pwr5_mont_t4_24 2478 1.1 christos .align 32 2479 1.1 christos bn_pwr5_mont_t4_24: 2480 1.1 christos #ifdef __arch64__ 2481 1.1 christos mov 0,%g5 2482 1.1 christos mov -128,%g4 2483 1.1 christos #elif defined(SPARCV9_64BIT_STACK) 2484 1.1 christos SPARC_LOAD_ADDRESS_LEAF(OPENSSL_sparcv9cap_P,%g1,%g5) 2485 1.1 christos ld [%g1+0],%g1 ! OPENSSL_sparcv9_P[0] 2486 1.1 christos mov -2047,%g4 2487 1.1 christos and %g1,SPARCV9_64BIT_STACK,%g1 2488 1.1 christos movrz %g1,0,%g4 2489 1.1 christos mov -1,%g5 2490 1.1 christos add %g4,-128,%g4 2491 1.1 christos #else 2492 1.1 christos mov -1,%g5 2493 1.1 christos mov -128,%g4 2494 1.1 christos #endif 2495 1.1 christos sllx %g5,32,%g5 2496 1.1 christos save %sp,%g4,%sp 2497 1.1 christos #ifndef __arch64__ 2498 1.1 christos save %sp,-128,%sp ! warm it up 2499 1.1 christos save %sp,-128,%sp 2500 1.1 christos save %sp,-128,%sp 2501 1.1 christos save %sp,-128,%sp 2502 1.1 christos save %sp,-128,%sp 2503 1.1 christos save %sp,-128,%sp 2504 1.1 christos restore 2505 1.1 christos restore 2506 1.1 christos restore 2507 1.1 christos restore 2508 1.1 christos restore 2509 1.1 christos restore 2510 1.1 christos #endif 2511 1.1 christos and %sp,1,%g4 2512 1.1 christos or %g5,%fp,%fp 2513 1.1 christos or %g4,%g5,%g5 2514 1.1 christos 2515 1.1 christos ! copy arguments to global registers 2516 1.1 christos mov %i0,%g1 2517 1.1 christos mov %i1,%g2 2518 1.1 christos ld [%i2+0],%f1 ! load *n0 2519 1.1 christos ld [%i2+4],%f0 2520 1.1 christos mov %i3,%g3 2521 1.1 christos srl %i4,%g0,%i4 ! pack last arguments 2522 1.1 christos sllx %i5,32,%g4 2523 1.1 christos or %i4,%g4,%g4 2524 1.1 christos .word 0xbbb00f00 !fsrc2 %f0,%f0,%f60 2525 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2526 1.1 christos ldx [%g1+0*8],%l0 2527 1.1 christos ldx [%g1+1*8],%l1 2528 1.1 christos ldx [%g1+2*8],%l2 2529 1.1 christos ldx [%g1+3*8],%l3 2530 1.1 christos ldx [%g1+4*8],%l4 2531 1.1 christos ldx [%g1+5*8],%l5 2532 1.1 christos ldx [%g1+6*8],%l6 2533 1.1 christos ldx [%g1+7*8],%l7 2534 1.1 christos ldx [%g1+8*8],%o0 2535 1.1 christos ldx [%g1+9*8],%o1 2536 1.1 christos ldx [%g1+10*8],%o2 2537 1.1 christos ldx [%g1+11*8],%o3 2538 1.1 christos ldx [%g1+12*8],%o4 2539 1.1 christos ldx [%g1+13*8],%o5 2540 1.1 christos ldd [%g1+14*8],%f24 2541 1.1 christos ldd [%g1+15*8],%f26 2542 1.1 christos ldd [%g1+16*8],%f28 2543 1.1 christos ldd [%g1+17*8],%f30 2544 1.1 christos ldd [%g1+18*8],%f32 2545 1.1 christos ldd [%g1+19*8],%f34 2546 1.1 christos ldd [%g1+20*8],%f36 2547 1.1 christos ldd [%g1+21*8],%f38 2548 1.1 christos ldd [%g1+22*8],%f40 2549 1.1 christos ldd [%g1+23*8],%f42 2550 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2551 1.1 christos ldx [%g2+0*8],%l0 2552 1.1 christos ldx [%g2+1*8],%l1 2553 1.1 christos ldx [%g2+2*8],%l2 2554 1.1 christos ldx [%g2+3*8],%l3 2555 1.1 christos ldx [%g2+4*8],%l4 2556 1.1 christos ldx [%g2+5*8],%l5 2557 1.1 christos ldx [%g2+6*8],%l6 2558 1.1 christos ldx [%g2+7*8],%l7 2559 1.1 christos ldx [%g2+8*8],%o0 2560 1.1 christos ldx [%g2+9*8],%o1 2561 1.1 christos ldx [%g2+10*8],%o2 2562 1.1 christos ldx [%g2+11*8],%o3 2563 1.1 christos ldx [%g2+12*8],%o4 2564 1.1 christos ldx [%g2+13*8],%o5 2565 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2566 1.1 christos ldx [%g2+14*8],%l0 2567 1.1 christos ldx [%g2+15*8],%l1 2568 1.1 christos ldx [%g2+16*8],%l2 2569 1.1 christos ldx [%g2+17*8],%l3 2570 1.1 christos ldx [%g2+18*8],%l4 2571 1.1 christos ldx [%g2+19*8],%l5 2572 1.1 christos ldx [%g2+20*8],%l6 2573 1.1 christos ldx [%g2+21*8],%l7 2574 1.1 christos ldx [%g2+22*8],%o0 2575 1.1 christos ldx [%g2+23*8],%o1 2576 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2577 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2578 1.1 christos 2579 1.1 christos srlx %g4, 32, %o4 ! unpack %g4 2580 1.1 christos srl %g4, %g0, %o5 2581 1.1 christos sub %o4, 5, %o4 2582 1.1 christos mov %g3, %o7 2583 1.1 christos sllx %o4, 32, %g4 ! re-pack %g4 2584 1.1 christos or %o5, %g4, %g4 2585 1.1 christos srl %o5, %o4, %o5 2586 1.1 christos srl %o5, 2, %o4 2587 1.1 christos and %o5, 3, %o5 2588 1.1 christos and %o4, 7, %o4 2589 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 2590 1.1 christos add %o5, %o7, %o7 ! of the pwrtbl 2591 1.1 christos or %g0, 1, %o5 2592 1.1 christos sll %o5, %o4, %o4 2593 1.1 christos wr %o4, %g0, %ccr 2594 1.1 christos b .Lstride_24 2595 1.1 christos nop 2596 1.1 christos .align 16 2597 1.1 christos .Lstride_24: 2598 1.1 christos ldx [%o7+0*32], %i0 2599 1.1 christos ldx [%o7+8*32], %i1 2600 1.1 christos ldx [%o7+1*32], %o4 2601 1.1 christos ldx [%o7+9*32], %o5 2602 1.1 christos movvs %icc, %o4, %i0 2603 1.1 christos ldx [%o7+2*32], %o4 2604 1.1 christos movvs %icc, %o5, %i1 2605 1.1 christos ldx [%o7+10*32],%o5 2606 1.1 christos move %icc, %o4, %i0 2607 1.1 christos ldx [%o7+3*32], %o4 2608 1.1 christos move %icc, %o5, %i1 2609 1.1 christos ldx [%o7+11*32],%o5 2610 1.1 christos movneg %icc, %o4, %i0 2611 1.1 christos ldx [%o7+4*32], %o4 2612 1.1 christos movneg %icc, %o5, %i1 2613 1.1 christos ldx [%o7+12*32],%o5 2614 1.1 christos movcs %xcc, %o4, %i0 2615 1.1 christos ldx [%o7+5*32],%o4 2616 1.1 christos movcs %xcc, %o5, %i1 2617 1.1 christos ldx [%o7+13*32],%o5 2618 1.1 christos movvs %xcc, %o4, %i0 2619 1.1 christos ldx [%o7+6*32], %o4 2620 1.1 christos movvs %xcc, %o5, %i1 2621 1.1 christos ldx [%o7+14*32],%o5 2622 1.1 christos move %xcc, %o4, %i0 2623 1.1 christos ldx [%o7+7*32], %o4 2624 1.1 christos move %xcc, %o5, %i1 2625 1.1 christos ldx [%o7+15*32],%o5 2626 1.1 christos movneg %xcc, %o4, %i0 2627 1.1 christos add %o7,16*32, %o7 2628 1.1 christos movneg %xcc, %o5, %i1 2629 1.1 christos ldx [%o7+0*32], %i2 2630 1.1 christos ldx [%o7+8*32], %i3 2631 1.1 christos ldx [%o7+1*32], %o4 2632 1.1 christos ldx [%o7+9*32], %o5 2633 1.1 christos movvs %icc, %o4, %i2 2634 1.1 christos ldx [%o7+2*32], %o4 2635 1.1 christos movvs %icc, %o5, %i3 2636 1.1 christos ldx [%o7+10*32],%o5 2637 1.1 christos move %icc, %o4, %i2 2638 1.1 christos ldx [%o7+3*32], %o4 2639 1.1 christos move %icc, %o5, %i3 2640 1.1 christos ldx [%o7+11*32],%o5 2641 1.1 christos movneg %icc, %o4, %i2 2642 1.1 christos ldx [%o7+4*32], %o4 2643 1.1 christos movneg %icc, %o5, %i3 2644 1.1 christos ldx [%o7+12*32],%o5 2645 1.1 christos movcs %xcc, %o4, %i2 2646 1.1 christos ldx [%o7+5*32],%o4 2647 1.1 christos movcs %xcc, %o5, %i3 2648 1.1 christos ldx [%o7+13*32],%o5 2649 1.1 christos movvs %xcc, %o4, %i2 2650 1.1 christos ldx [%o7+6*32], %o4 2651 1.1 christos movvs %xcc, %o5, %i3 2652 1.1 christos ldx [%o7+14*32],%o5 2653 1.1 christos move %xcc, %o4, %i2 2654 1.1 christos ldx [%o7+7*32], %o4 2655 1.1 christos move %xcc, %o5, %i3 2656 1.1 christos ldx [%o7+15*32],%o5 2657 1.1 christos movneg %xcc, %o4, %i2 2658 1.1 christos add %o7,16*32, %o7 2659 1.1 christos movneg %xcc, %o5, %i3 2660 1.1 christos ldx [%o7+0*32], %i4 2661 1.1 christos ldx [%o7+8*32], %i5 2662 1.1 christos ldx [%o7+1*32], %o4 2663 1.1 christos ldx [%o7+9*32], %o5 2664 1.1 christos movvs %icc, %o4, %i4 2665 1.1 christos ldx [%o7+2*32], %o4 2666 1.1 christos movvs %icc, %o5, %i5 2667 1.1 christos ldx [%o7+10*32],%o5 2668 1.1 christos move %icc, %o4, %i4 2669 1.1 christos ldx [%o7+3*32], %o4 2670 1.1 christos move %icc, %o5, %i5 2671 1.1 christos ldx [%o7+11*32],%o5 2672 1.1 christos movneg %icc, %o4, %i4 2673 1.1 christos ldx [%o7+4*32], %o4 2674 1.1 christos movneg %icc, %o5, %i5 2675 1.1 christos ldx [%o7+12*32],%o5 2676 1.1 christos movcs %xcc, %o4, %i4 2677 1.1 christos ldx [%o7+5*32],%o4 2678 1.1 christos movcs %xcc, %o5, %i5 2679 1.1 christos ldx [%o7+13*32],%o5 2680 1.1 christos movvs %xcc, %o4, %i4 2681 1.1 christos ldx [%o7+6*32], %o4 2682 1.1 christos movvs %xcc, %o5, %i5 2683 1.1 christos ldx [%o7+14*32],%o5 2684 1.1 christos move %xcc, %o4, %i4 2685 1.1 christos ldx [%o7+7*32], %o4 2686 1.1 christos move %xcc, %o5, %i5 2687 1.1 christos ldx [%o7+15*32],%o5 2688 1.1 christos movneg %xcc, %o4, %i4 2689 1.1 christos add %o7,16*32, %o7 2690 1.1 christos movneg %xcc, %o5, %i5 2691 1.1 christos ldx [%o7+0*32], %l0 2692 1.1 christos ldx [%o7+8*32], %l1 2693 1.1 christos ldx [%o7+1*32], %o4 2694 1.1 christos ldx [%o7+9*32], %o5 2695 1.1 christos movvs %icc, %o4, %l0 2696 1.1 christos ldx [%o7+2*32], %o4 2697 1.1 christos movvs %icc, %o5, %l1 2698 1.1 christos ldx [%o7+10*32],%o5 2699 1.1 christos move %icc, %o4, %l0 2700 1.1 christos ldx [%o7+3*32], %o4 2701 1.1 christos move %icc, %o5, %l1 2702 1.1 christos ldx [%o7+11*32],%o5 2703 1.1 christos movneg %icc, %o4, %l0 2704 1.1 christos ldx [%o7+4*32], %o4 2705 1.1 christos movneg %icc, %o5, %l1 2706 1.1 christos ldx [%o7+12*32],%o5 2707 1.1 christos movcs %xcc, %o4, %l0 2708 1.1 christos ldx [%o7+5*32],%o4 2709 1.1 christos movcs %xcc, %o5, %l1 2710 1.1 christos ldx [%o7+13*32],%o5 2711 1.1 christos movvs %xcc, %o4, %l0 2712 1.1 christos ldx [%o7+6*32], %o4 2713 1.1 christos movvs %xcc, %o5, %l1 2714 1.1 christos ldx [%o7+14*32],%o5 2715 1.1 christos move %xcc, %o4, %l0 2716 1.1 christos ldx [%o7+7*32], %o4 2717 1.1 christos move %xcc, %o5, %l1 2718 1.1 christos ldx [%o7+15*32],%o5 2719 1.1 christos movneg %xcc, %o4, %l0 2720 1.1 christos add %o7,16*32, %o7 2721 1.1 christos movneg %xcc, %o5, %l1 2722 1.1 christos ldx [%o7+0*32], %l2 2723 1.1 christos ldx [%o7+8*32], %l3 2724 1.1 christos ldx [%o7+1*32], %o4 2725 1.1 christos ldx [%o7+9*32], %o5 2726 1.1 christos movvs %icc, %o4, %l2 2727 1.1 christos ldx [%o7+2*32], %o4 2728 1.1 christos movvs %icc, %o5, %l3 2729 1.1 christos ldx [%o7+10*32],%o5 2730 1.1 christos move %icc, %o4, %l2 2731 1.1 christos ldx [%o7+3*32], %o4 2732 1.1 christos move %icc, %o5, %l3 2733 1.1 christos ldx [%o7+11*32],%o5 2734 1.1 christos movneg %icc, %o4, %l2 2735 1.1 christos ldx [%o7+4*32], %o4 2736 1.1 christos movneg %icc, %o5, %l3 2737 1.1 christos ldx [%o7+12*32],%o5 2738 1.1 christos movcs %xcc, %o4, %l2 2739 1.1 christos ldx [%o7+5*32],%o4 2740 1.1 christos movcs %xcc, %o5, %l3 2741 1.1 christos ldx [%o7+13*32],%o5 2742 1.1 christos movvs %xcc, %o4, %l2 2743 1.1 christos ldx [%o7+6*32], %o4 2744 1.1 christos movvs %xcc, %o5, %l3 2745 1.1 christos ldx [%o7+14*32],%o5 2746 1.1 christos move %xcc, %o4, %l2 2747 1.1 christos ldx [%o7+7*32], %o4 2748 1.1 christos move %xcc, %o5, %l3 2749 1.1 christos ldx [%o7+15*32],%o5 2750 1.1 christos movneg %xcc, %o4, %l2 2751 1.1 christos add %o7,16*32, %o7 2752 1.1 christos movneg %xcc, %o5, %l3 2753 1.1 christos ldx [%o7+0*32], %l4 2754 1.1 christos ldx [%o7+8*32], %l5 2755 1.1 christos ldx [%o7+1*32], %o4 2756 1.1 christos ldx [%o7+9*32], %o5 2757 1.1 christos movvs %icc, %o4, %l4 2758 1.1 christos ldx [%o7+2*32], %o4 2759 1.1 christos movvs %icc, %o5, %l5 2760 1.1 christos ldx [%o7+10*32],%o5 2761 1.1 christos move %icc, %o4, %l4 2762 1.1 christos ldx [%o7+3*32], %o4 2763 1.1 christos move %icc, %o5, %l5 2764 1.1 christos ldx [%o7+11*32],%o5 2765 1.1 christos movneg %icc, %o4, %l4 2766 1.1 christos ldx [%o7+4*32], %o4 2767 1.1 christos movneg %icc, %o5, %l5 2768 1.1 christos ldx [%o7+12*32],%o5 2769 1.1 christos movcs %xcc, %o4, %l4 2770 1.1 christos ldx [%o7+5*32],%o4 2771 1.1 christos movcs %xcc, %o5, %l5 2772 1.1 christos ldx [%o7+13*32],%o5 2773 1.1 christos movvs %xcc, %o4, %l4 2774 1.1 christos ldx [%o7+6*32], %o4 2775 1.1 christos movvs %xcc, %o5, %l5 2776 1.1 christos ldx [%o7+14*32],%o5 2777 1.1 christos move %xcc, %o4, %l4 2778 1.1 christos ldx [%o7+7*32], %o4 2779 1.1 christos move %xcc, %o5, %l5 2780 1.1 christos ldx [%o7+15*32],%o5 2781 1.1 christos movneg %xcc, %o4, %l4 2782 1.1 christos add %o7,16*32, %o7 2783 1.1 christos movneg %xcc, %o5, %l5 2784 1.1 christos ldx [%o7+0*32], %l6 2785 1.1 christos ldx [%o7+8*32], %l7 2786 1.1 christos ldx [%o7+1*32], %o4 2787 1.1 christos ldx [%o7+9*32], %o5 2788 1.1 christos movvs %icc, %o4, %l6 2789 1.1 christos ldx [%o7+2*32], %o4 2790 1.1 christos movvs %icc, %o5, %l7 2791 1.1 christos ldx [%o7+10*32],%o5 2792 1.1 christos move %icc, %o4, %l6 2793 1.1 christos ldx [%o7+3*32], %o4 2794 1.1 christos move %icc, %o5, %l7 2795 1.1 christos ldx [%o7+11*32],%o5 2796 1.1 christos movneg %icc, %o4, %l6 2797 1.1 christos ldx [%o7+4*32], %o4 2798 1.1 christos movneg %icc, %o5, %l7 2799 1.1 christos ldx [%o7+12*32],%o5 2800 1.1 christos movcs %xcc, %o4, %l6 2801 1.1 christos ldx [%o7+5*32],%o4 2802 1.1 christos movcs %xcc, %o5, %l7 2803 1.1 christos ldx [%o7+13*32],%o5 2804 1.1 christos movvs %xcc, %o4, %l6 2805 1.1 christos ldx [%o7+6*32], %o4 2806 1.1 christos movvs %xcc, %o5, %l7 2807 1.1 christos ldx [%o7+14*32],%o5 2808 1.1 christos move %xcc, %o4, %l6 2809 1.1 christos ldx [%o7+7*32], %o4 2810 1.1 christos move %xcc, %o5, %l7 2811 1.1 christos ldx [%o7+15*32],%o5 2812 1.1 christos movneg %xcc, %o4, %l6 2813 1.1 christos add %o7,16*32, %o7 2814 1.1 christos movneg %xcc, %o5, %l7 2815 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 2816 1.1 christos ldx [%i7+0*32], %i0 2817 1.1 christos ldx [%i7+8*32], %i1 2818 1.1 christos ldx [%i7+1*32], %o4 2819 1.1 christos ldx [%i7+9*32], %o5 2820 1.1 christos movvs %icc, %o4, %i0 2821 1.1 christos ldx [%i7+2*32], %o4 2822 1.1 christos movvs %icc, %o5, %i1 2823 1.1 christos ldx [%i7+10*32],%o5 2824 1.1 christos move %icc, %o4, %i0 2825 1.1 christos ldx [%i7+3*32], %o4 2826 1.1 christos move %icc, %o5, %i1 2827 1.1 christos ldx [%i7+11*32],%o5 2828 1.1 christos movneg %icc, %o4, %i0 2829 1.1 christos ldx [%i7+4*32], %o4 2830 1.1 christos movneg %icc, %o5, %i1 2831 1.1 christos ldx [%i7+12*32],%o5 2832 1.1 christos movcs %xcc, %o4, %i0 2833 1.1 christos ldx [%i7+5*32],%o4 2834 1.1 christos movcs %xcc, %o5, %i1 2835 1.1 christos ldx [%i7+13*32],%o5 2836 1.1 christos movvs %xcc, %o4, %i0 2837 1.1 christos ldx [%i7+6*32], %o4 2838 1.1 christos movvs %xcc, %o5, %i1 2839 1.1 christos ldx [%i7+14*32],%o5 2840 1.1 christos move %xcc, %o4, %i0 2841 1.1 christos ldx [%i7+7*32], %o4 2842 1.1 christos move %xcc, %o5, %i1 2843 1.1 christos ldx [%i7+15*32],%o5 2844 1.1 christos movneg %xcc, %o4, %i0 2845 1.1 christos add %i7,16*32, %i7 2846 1.1 christos movneg %xcc, %o5, %i1 2847 1.1 christos ldx [%i7+0*32], %i2 2848 1.1 christos ldx [%i7+8*32], %i3 2849 1.1 christos ldx [%i7+1*32], %o4 2850 1.1 christos ldx [%i7+9*32], %o5 2851 1.1 christos movvs %icc, %o4, %i2 2852 1.1 christos ldx [%i7+2*32], %o4 2853 1.1 christos movvs %icc, %o5, %i3 2854 1.1 christos ldx [%i7+10*32],%o5 2855 1.1 christos move %icc, %o4, %i2 2856 1.1 christos ldx [%i7+3*32], %o4 2857 1.1 christos move %icc, %o5, %i3 2858 1.1 christos ldx [%i7+11*32],%o5 2859 1.1 christos movneg %icc, %o4, %i2 2860 1.1 christos ldx [%i7+4*32], %o4 2861 1.1 christos movneg %icc, %o5, %i3 2862 1.1 christos ldx [%i7+12*32],%o5 2863 1.1 christos movcs %xcc, %o4, %i2 2864 1.1 christos ldx [%i7+5*32],%o4 2865 1.1 christos movcs %xcc, %o5, %i3 2866 1.1 christos ldx [%i7+13*32],%o5 2867 1.1 christos movvs %xcc, %o4, %i2 2868 1.1 christos ldx [%i7+6*32], %o4 2869 1.1 christos movvs %xcc, %o5, %i3 2870 1.1 christos ldx [%i7+14*32],%o5 2871 1.1 christos move %xcc, %o4, %i2 2872 1.1 christos ldx [%i7+7*32], %o4 2873 1.1 christos move %xcc, %o5, %i3 2874 1.1 christos ldx [%i7+15*32],%o5 2875 1.1 christos movneg %xcc, %o4, %i2 2876 1.1 christos add %i7,16*32, %i7 2877 1.1 christos movneg %xcc, %o5, %i3 2878 1.1 christos ldx [%i7+0*32], %i4 2879 1.1 christos ldx [%i7+8*32], %i5 2880 1.1 christos ldx [%i7+1*32], %o4 2881 1.1 christos ldx [%i7+9*32], %o5 2882 1.1 christos movvs %icc, %o4, %i4 2883 1.1 christos ldx [%i7+2*32], %o4 2884 1.1 christos movvs %icc, %o5, %i5 2885 1.1 christos ldx [%i7+10*32],%o5 2886 1.1 christos move %icc, %o4, %i4 2887 1.1 christos ldx [%i7+3*32], %o4 2888 1.1 christos move %icc, %o5, %i5 2889 1.1 christos ldx [%i7+11*32],%o5 2890 1.1 christos movneg %icc, %o4, %i4 2891 1.1 christos ldx [%i7+4*32], %o4 2892 1.1 christos movneg %icc, %o5, %i5 2893 1.1 christos ldx [%i7+12*32],%o5 2894 1.1 christos movcs %xcc, %o4, %i4 2895 1.1 christos ldx [%i7+5*32],%o4 2896 1.1 christos movcs %xcc, %o5, %i5 2897 1.1 christos ldx [%i7+13*32],%o5 2898 1.1 christos movvs %xcc, %o4, %i4 2899 1.1 christos ldx [%i7+6*32], %o4 2900 1.1 christos movvs %xcc, %o5, %i5 2901 1.1 christos ldx [%i7+14*32],%o5 2902 1.1 christos move %xcc, %o4, %i4 2903 1.1 christos ldx [%i7+7*32], %o4 2904 1.1 christos move %xcc, %o5, %i5 2905 1.1 christos ldx [%i7+15*32],%o5 2906 1.1 christos movneg %xcc, %o4, %i4 2907 1.1 christos add %i7,16*32, %i7 2908 1.1 christos movneg %xcc, %o5, %i5 2909 1.1 christos ldx [%i7+0*32], %l0 2910 1.1 christos ldx [%i7+8*32], %l1 2911 1.1 christos ldx [%i7+1*32], %o4 2912 1.1 christos ldx [%i7+9*32], %o5 2913 1.1 christos movvs %icc, %o4, %l0 2914 1.1 christos ldx [%i7+2*32], %o4 2915 1.1 christos movvs %icc, %o5, %l1 2916 1.1 christos ldx [%i7+10*32],%o5 2917 1.1 christos move %icc, %o4, %l0 2918 1.1 christos ldx [%i7+3*32], %o4 2919 1.1 christos move %icc, %o5, %l1 2920 1.1 christos ldx [%i7+11*32],%o5 2921 1.1 christos movneg %icc, %o4, %l0 2922 1.1 christos ldx [%i7+4*32], %o4 2923 1.1 christos movneg %icc, %o5, %l1 2924 1.1 christos ldx [%i7+12*32],%o5 2925 1.1 christos movcs %xcc, %o4, %l0 2926 1.1 christos ldx [%i7+5*32],%o4 2927 1.1 christos movcs %xcc, %o5, %l1 2928 1.1 christos ldx [%i7+13*32],%o5 2929 1.1 christos movvs %xcc, %o4, %l0 2930 1.1 christos ldx [%i7+6*32], %o4 2931 1.1 christos movvs %xcc, %o5, %l1 2932 1.1 christos ldx [%i7+14*32],%o5 2933 1.1 christos move %xcc, %o4, %l0 2934 1.1 christos ldx [%i7+7*32], %o4 2935 1.1 christos move %xcc, %o5, %l1 2936 1.1 christos ldx [%i7+15*32],%o5 2937 1.1 christos movneg %xcc, %o4, %l0 2938 1.1 christos add %i7,16*32, %i7 2939 1.1 christos movneg %xcc, %o5, %l1 2940 1.1 christos ldx [%i7+0*32], %l2 2941 1.1 christos ldx [%i7+8*32], %l3 2942 1.1 christos ldx [%i7+1*32], %o4 2943 1.1 christos ldx [%i7+9*32], %o5 2944 1.1 christos movvs %icc, %o4, %l2 2945 1.1 christos ldx [%i7+2*32], %o4 2946 1.1 christos movvs %icc, %o5, %l3 2947 1.1 christos ldx [%i7+10*32],%o5 2948 1.1 christos move %icc, %o4, %l2 2949 1.1 christos ldx [%i7+3*32], %o4 2950 1.1 christos move %icc, %o5, %l3 2951 1.1 christos ldx [%i7+11*32],%o5 2952 1.1 christos movneg %icc, %o4, %l2 2953 1.1 christos ldx [%i7+4*32], %o4 2954 1.1 christos movneg %icc, %o5, %l3 2955 1.1 christos ldx [%i7+12*32],%o5 2956 1.1 christos movcs %xcc, %o4, %l2 2957 1.1 christos ldx [%i7+5*32],%o4 2958 1.1 christos movcs %xcc, %o5, %l3 2959 1.1 christos ldx [%i7+13*32],%o5 2960 1.1 christos movvs %xcc, %o4, %l2 2961 1.1 christos ldx [%i7+6*32], %o4 2962 1.1 christos movvs %xcc, %o5, %l3 2963 1.1 christos ldx [%i7+14*32],%o5 2964 1.1 christos move %xcc, %o4, %l2 2965 1.1 christos ldx [%i7+7*32], %o4 2966 1.1 christos move %xcc, %o5, %l3 2967 1.1 christos ldx [%i7+15*32],%o5 2968 1.1 christos movneg %xcc, %o4, %l2 2969 1.1 christos add %i7,16*32, %i7 2970 1.1 christos movneg %xcc, %o5, %l3 2971 1.1 christos srax %g4, 32, %o4 ! unpack %g4 2972 1.1 christos srl %g4, %g0, %o5 2973 1.1 christos sub %o4, 5, %o4 2974 1.1 christos mov %g3, %i7 2975 1.1 christos sllx %o4, 32, %g4 ! re-pack %g4 2976 1.1 christos or %o5, %g4, %g4 2977 1.1 christos srl %o5, %o4, %o5 2978 1.1 christos srl %o5, 2, %o4 2979 1.1 christos and %o5, 3, %o5 2980 1.1 christos and %o4, 7, %o4 2981 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 2982 1.1 christos add %o5, %i7, %i7 ! of the pwrtbl 2983 1.1 christos or %g0, 1, %o5 2984 1.1 christos sll %o5, %o4, %o4 2985 1.1 christos .word 0x81b02940+24-1 ! montsqr 24-1 2986 1.1 christos fbu,pn %fcc3,.Labort_24 2987 1.1 christos #ifndef __arch64__ 2988 1.1 christos and %fp,%g5,%g5 2989 1.1 christos brz,pn %g5,.Labort_24 2990 1.1 christos #endif 2991 1.1 christos nop 2992 1.1 christos .word 0x81b02940+24-1 ! montsqr 24-1 2993 1.1 christos fbu,pn %fcc3,.Labort_24 2994 1.1 christos #ifndef __arch64__ 2995 1.1 christos and %fp,%g5,%g5 2996 1.1 christos brz,pn %g5,.Labort_24 2997 1.1 christos #endif 2998 1.1 christos nop 2999 1.1 christos .word 0x81b02940+24-1 ! montsqr 24-1 3000 1.1 christos fbu,pn %fcc3,.Labort_24 3001 1.1 christos #ifndef __arch64__ 3002 1.1 christos and %fp,%g5,%g5 3003 1.1 christos brz,pn %g5,.Labort_24 3004 1.1 christos #endif 3005 1.1 christos nop 3006 1.1 christos .word 0x81b02940+24-1 ! montsqr 24-1 3007 1.1 christos fbu,pn %fcc3,.Labort_24 3008 1.1 christos #ifndef __arch64__ 3009 1.1 christos and %fp,%g5,%g5 3010 1.1 christos brz,pn %g5,.Labort_24 3011 1.1 christos #endif 3012 1.1 christos nop 3013 1.1 christos .word 0x81b02940+24-1 ! montsqr 24-1 3014 1.1 christos fbu,pn %fcc3,.Labort_24 3015 1.1 christos #ifndef __arch64__ 3016 1.1 christos and %fp,%g5,%g5 3017 1.1 christos brz,pn %g5,.Labort_24 3018 1.1 christos #endif 3019 1.1 christos nop 3020 1.1 christos wr %o4, %g0, %ccr 3021 1.1 christos .word 0x81b02920+24-1 ! montmul 24-1 3022 1.1 christos fbu,pn %fcc3,.Labort_24 3023 1.1 christos #ifndef __arch64__ 3024 1.1 christos and %fp,%g5,%g5 3025 1.1 christos brz,pn %g5,.Labort_24 3026 1.1 christos #endif 3027 1.1 christos 3028 1.1 christos srax %g4, 32, %o4 3029 1.1 christos #ifdef __arch64__ 3030 1.1 christos brgez %o4,.Lstride_24 3031 1.1 christos restore 3032 1.1 christos restore 3033 1.1 christos restore 3034 1.1 christos restore 3035 1.1 christos restore 3036 1.1 christos #else 3037 1.1 christos brgez %o4,.Lstride_24 3038 1.1 christos restore; and %fp,%g5,%g5 3039 1.1 christos restore; and %fp,%g5,%g5 3040 1.1 christos restore; and %fp,%g5,%g5 3041 1.1 christos restore; and %fp,%g5,%g5 3042 1.1 christos brz,pn %g5,.Labort1_24 3043 1.1 christos restore 3044 1.1 christos #endif 3045 1.1 christos .word 0x81b02310 !movxtod %l0,%f0 3046 1.1 christos .word 0x85b02311 !movxtod %l1,%f2 3047 1.1 christos .word 0x89b02312 !movxtod %l2,%f4 3048 1.1 christos .word 0x8db02313 !movxtod %l3,%f6 3049 1.1 christos .word 0x91b02314 !movxtod %l4,%f8 3050 1.1 christos .word 0x95b02315 !movxtod %l5,%f10 3051 1.1 christos .word 0x99b02316 !movxtod %l6,%f12 3052 1.1 christos .word 0x9db02317 !movxtod %l7,%f14 3053 1.1 christos .word 0xa1b02308 !movxtod %o0,%f16 3054 1.1 christos .word 0xa5b02309 !movxtod %o1,%f18 3055 1.1 christos .word 0xa9b0230a !movxtod %o2,%f20 3056 1.1 christos .word 0xadb0230b !movxtod %o3,%f22 3057 1.1 christos .word 0xbbb0230c !movxtod %o4,%f60 3058 1.1 christos .word 0xbfb0230d !movxtod %o5,%f62 3059 1.1 christos #ifdef __arch64__ 3060 1.1 christos restore 3061 1.1 christos #else 3062 1.1 christos and %fp,%g5,%g5 3063 1.1 christos restore 3064 1.1 christos and %g5,1,%o7 3065 1.1 christos and %fp,%g5,%g5 3066 1.1 christos srl %fp,0,%fp ! just in case? 3067 1.1 christos or %o7,%g5,%g5 3068 1.1 christos brz,a,pn %g5,.Ldone_24 3069 1.1 christos mov 0,%i0 ! return failure 3070 1.1 christos #endif 3071 1.1 christos std %f0,[%g1+0*8] 3072 1.1 christos std %f2,[%g1+1*8] 3073 1.1 christos std %f4,[%g1+2*8] 3074 1.1 christos std %f6,[%g1+3*8] 3075 1.1 christos std %f8,[%g1+4*8] 3076 1.1 christos std %f10,[%g1+5*8] 3077 1.1 christos std %f12,[%g1+6*8] 3078 1.1 christos std %f14,[%g1+7*8] 3079 1.1 christos std %f16,[%g1+8*8] 3080 1.1 christos std %f18,[%g1+9*8] 3081 1.1 christos std %f20,[%g1+10*8] 3082 1.1 christos std %f22,[%g1+11*8] 3083 1.1 christos std %f60,[%g1+12*8] 3084 1.1 christos std %f62,[%g1+13*8] 3085 1.1 christos std %f24,[%g1+14*8] 3086 1.1 christos std %f26,[%g1+15*8] 3087 1.1 christos std %f28,[%g1+16*8] 3088 1.1 christos std %f30,[%g1+17*8] 3089 1.1 christos std %f32,[%g1+18*8] 3090 1.1 christos std %f34,[%g1+19*8] 3091 1.1 christos std %f36,[%g1+20*8] 3092 1.1 christos std %f38,[%g1+21*8] 3093 1.1 christos std %f40,[%g1+22*8] 3094 1.1 christos std %f42,[%g1+23*8] 3095 1.1 christos mov 1,%i0 ! return success 3096 1.1 christos .Ldone_24: 3097 1.1 christos ret 3098 1.1 christos restore 3099 1.1 christos 3100 1.1 christos .Labort_24: 3101 1.1 christos restore 3102 1.1 christos restore 3103 1.1 christos restore 3104 1.1 christos restore 3105 1.1 christos restore 3106 1.1 christos .Labort1_24: 3107 1.1 christos restore 3108 1.1 christos 3109 1.1 christos mov 0,%i0 ! return failure 3110 1.1 christos ret 3111 1.1 christos restore 3112 1.1 christos .type bn_pwr5_mont_t4_24, #function 3113 1.1 christos .size bn_pwr5_mont_t4_24, .-bn_pwr5_mont_t4_24 3114 1.1 christos .globl bn_pwr5_mont_t4_32 3115 1.1 christos .align 32 3116 1.1 christos bn_pwr5_mont_t4_32: 3117 1.1 christos #ifdef __arch64__ 3118 1.1 christos mov 0,%g5 3119 1.1 christos mov -128,%g4 3120 1.1 christos #elif defined(SPARCV9_64BIT_STACK) 3121 1.1 christos SPARC_LOAD_ADDRESS_LEAF(OPENSSL_sparcv9cap_P,%g1,%g5) 3122 1.1 christos ld [%g1+0],%g1 ! OPENSSL_sparcv9_P[0] 3123 1.1 christos mov -2047,%g4 3124 1.1 christos and %g1,SPARCV9_64BIT_STACK,%g1 3125 1.1 christos movrz %g1,0,%g4 3126 1.1 christos mov -1,%g5 3127 1.1 christos add %g4,-128,%g4 3128 1.1 christos #else 3129 1.1 christos mov -1,%g5 3130 1.1 christos mov -128,%g4 3131 1.1 christos #endif 3132 1.1 christos sllx %g5,32,%g5 3133 1.1 christos save %sp,%g4,%sp 3134 1.1 christos #ifndef __arch64__ 3135 1.1 christos save %sp,-128,%sp ! warm it up 3136 1.1 christos save %sp,-128,%sp 3137 1.1 christos save %sp,-128,%sp 3138 1.1 christos save %sp,-128,%sp 3139 1.1 christos save %sp,-128,%sp 3140 1.1 christos save %sp,-128,%sp 3141 1.1 christos restore 3142 1.1 christos restore 3143 1.1 christos restore 3144 1.1 christos restore 3145 1.1 christos restore 3146 1.1 christos restore 3147 1.1 christos #endif 3148 1.1 christos and %sp,1,%g4 3149 1.1 christos or %g5,%fp,%fp 3150 1.1 christos or %g4,%g5,%g5 3151 1.1 christos 3152 1.1 christos ! copy arguments to global registers 3153 1.1 christos mov %i0,%g1 3154 1.1 christos mov %i1,%g2 3155 1.1 christos ld [%i2+0],%f1 ! load *n0 3156 1.1 christos ld [%i2+4],%f0 3157 1.1 christos mov %i3,%g3 3158 1.1 christos srl %i4,%g0,%i4 ! pack last arguments 3159 1.1 christos sllx %i5,32,%g4 3160 1.1 christos or %i4,%g4,%g4 3161 1.1 christos .word 0xbbb00f00 !fsrc2 %f0,%f0,%f60 3162 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 3163 1.1 christos ldx [%g1+0*8],%l0 3164 1.1 christos ldx [%g1+1*8],%l1 3165 1.1 christos ldx [%g1+2*8],%l2 3166 1.1 christos ldx [%g1+3*8],%l3 3167 1.1 christos ldx [%g1+4*8],%l4 3168 1.1 christos ldx [%g1+5*8],%l5 3169 1.1 christos ldx [%g1+6*8],%l6 3170 1.1 christos ldx [%g1+7*8],%l7 3171 1.1 christos ldx [%g1+8*8],%o0 3172 1.1 christos ldx [%g1+9*8],%o1 3173 1.1 christos ldx [%g1+10*8],%o2 3174 1.1 christos ldx [%g1+11*8],%o3 3175 1.1 christos ldx [%g1+12*8],%o4 3176 1.1 christos ldx [%g1+13*8],%o5 3177 1.1 christos ldd [%g1+14*8],%f24 3178 1.1 christos ldd [%g1+15*8],%f26 3179 1.1 christos ldd [%g1+16*8],%f28 3180 1.1 christos ldd [%g1+17*8],%f30 3181 1.1 christos ldd [%g1+18*8],%f32 3182 1.1 christos ldd [%g1+19*8],%f34 3183 1.1 christos ldd [%g1+20*8],%f36 3184 1.1 christos ldd [%g1+21*8],%f38 3185 1.1 christos ldd [%g1+22*8],%f40 3186 1.1 christos ldd [%g1+23*8],%f42 3187 1.1 christos ldd [%g1+24*8],%f44 3188 1.1 christos ldd [%g1+25*8],%f46 3189 1.1 christos ldd [%g1+26*8],%f48 3190 1.1 christos ldd [%g1+27*8],%f50 3191 1.1 christos ldd [%g1+28*8],%f52 3192 1.1 christos ldd [%g1+29*8],%f54 3193 1.1 christos ldd [%g1+30*8],%f56 3194 1.1 christos ldd [%g1+31*8],%f58 3195 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 3196 1.1 christos ldx [%g2+0*8],%l0 3197 1.1 christos ldx [%g2+1*8],%l1 3198 1.1 christos ldx [%g2+2*8],%l2 3199 1.1 christos ldx [%g2+3*8],%l3 3200 1.1 christos ldx [%g2+4*8],%l4 3201 1.1 christos ldx [%g2+5*8],%l5 3202 1.1 christos ldx [%g2+6*8],%l6 3203 1.1 christos ldx [%g2+7*8],%l7 3204 1.1 christos ldx [%g2+8*8],%o0 3205 1.1 christos ldx [%g2+9*8],%o1 3206 1.1 christos ldx [%g2+10*8],%o2 3207 1.1 christos ldx [%g2+11*8],%o3 3208 1.1 christos ldx [%g2+12*8],%o4 3209 1.1 christos ldx [%g2+13*8],%o5 3210 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 3211 1.1 christos ldx [%g2+14*8],%l0 3212 1.1 christos ldx [%g2+15*8],%l1 3213 1.1 christos ldx [%g2+16*8],%l2 3214 1.1 christos ldx [%g2+17*8],%l3 3215 1.1 christos ldx [%g2+18*8],%l4 3216 1.1 christos ldx [%g2+19*8],%l5 3217 1.1 christos ldx [%g2+20*8],%l6 3218 1.1 christos ldx [%g2+21*8],%l7 3219 1.1 christos ldx [%g2+22*8],%o0 3220 1.1 christos ldx [%g2+23*8],%o1 3221 1.1 christos ldx [%g2+24*8],%o2 3222 1.1 christos ldx [%g2+25*8],%o3 3223 1.1 christos ldx [%g2+26*8],%o4 3224 1.1 christos ldx [%g2+27*8],%o5 3225 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 3226 1.1 christos ldx [%g2+28*8],%l0 3227 1.1 christos ldx [%g2+29*8],%l1 3228 1.1 christos ldx [%g2+30*8],%l2 3229 1.1 christos ldx [%g2+31*8],%l3 3230 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 3231 1.1 christos 3232 1.1 christos srlx %g4, 32, %o4 ! unpack %g4 3233 1.1 christos srl %g4, %g0, %o5 3234 1.1 christos sub %o4, 5, %o4 3235 1.1 christos mov %g3, %o7 3236 1.1 christos sllx %o4, 32, %g4 ! re-pack %g4 3237 1.1 christos or %o5, %g4, %g4 3238 1.1 christos srl %o5, %o4, %o5 3239 1.1 christos srl %o5, 2, %o4 3240 1.1 christos and %o5, 3, %o5 3241 1.1 christos and %o4, 7, %o4 3242 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 3243 1.1 christos add %o5, %o7, %o7 ! of the pwrtbl 3244 1.1 christos or %g0, 1, %o5 3245 1.1 christos sll %o5, %o4, %o4 3246 1.1 christos wr %o4, %g0, %ccr 3247 1.1 christos b .Lstride_32 3248 1.1 christos nop 3249 1.1 christos .align 16 3250 1.1 christos .Lstride_32: 3251 1.1 christos ldx [%o7+0*32], %i0 3252 1.1 christos ldx [%o7+8*32], %i1 3253 1.1 christos ldx [%o7+1*32], %o4 3254 1.1 christos ldx [%o7+9*32], %o5 3255 1.1 christos movvs %icc, %o4, %i0 3256 1.1 christos ldx [%o7+2*32], %o4 3257 1.1 christos movvs %icc, %o5, %i1 3258 1.1 christos ldx [%o7+10*32],%o5 3259 1.1 christos move %icc, %o4, %i0 3260 1.1 christos ldx [%o7+3*32], %o4 3261 1.1 christos move %icc, %o5, %i1 3262 1.1 christos ldx [%o7+11*32],%o5 3263 1.1 christos movneg %icc, %o4, %i0 3264 1.1 christos ldx [%o7+4*32], %o4 3265 1.1 christos movneg %icc, %o5, %i1 3266 1.1 christos ldx [%o7+12*32],%o5 3267 1.1 christos movcs %xcc, %o4, %i0 3268 1.1 christos ldx [%o7+5*32],%o4 3269 1.1 christos movcs %xcc, %o5, %i1 3270 1.1 christos ldx [%o7+13*32],%o5 3271 1.1 christos movvs %xcc, %o4, %i0 3272 1.1 christos ldx [%o7+6*32], %o4 3273 1.1 christos movvs %xcc, %o5, %i1 3274 1.1 christos ldx [%o7+14*32],%o5 3275 1.1 christos move %xcc, %o4, %i0 3276 1.1 christos ldx [%o7+7*32], %o4 3277 1.1 christos move %xcc, %o5, %i1 3278 1.1 christos ldx [%o7+15*32],%o5 3279 1.1 christos movneg %xcc, %o4, %i0 3280 1.1 christos add %o7,16*32, %o7 3281 1.1 christos movneg %xcc, %o5, %i1 3282 1.1 christos ldx [%o7+0*32], %i2 3283 1.1 christos ldx [%o7+8*32], %i3 3284 1.1 christos ldx [%o7+1*32], %o4 3285 1.1 christos ldx [%o7+9*32], %o5 3286 1.1 christos movvs %icc, %o4, %i2 3287 1.1 christos ldx [%o7+2*32], %o4 3288 1.1 christos movvs %icc, %o5, %i3 3289 1.1 christos ldx [%o7+10*32],%o5 3290 1.1 christos move %icc, %o4, %i2 3291 1.1 christos ldx [%o7+3*32], %o4 3292 1.1 christos move %icc, %o5, %i3 3293 1.1 christos ldx [%o7+11*32],%o5 3294 1.1 christos movneg %icc, %o4, %i2 3295 1.1 christos ldx [%o7+4*32], %o4 3296 1.1 christos movneg %icc, %o5, %i3 3297 1.1 christos ldx [%o7+12*32],%o5 3298 1.1 christos movcs %xcc, %o4, %i2 3299 1.1 christos ldx [%o7+5*32],%o4 3300 1.1 christos movcs %xcc, %o5, %i3 3301 1.1 christos ldx [%o7+13*32],%o5 3302 1.1 christos movvs %xcc, %o4, %i2 3303 1.1 christos ldx [%o7+6*32], %o4 3304 1.1 christos movvs %xcc, %o5, %i3 3305 1.1 christos ldx [%o7+14*32],%o5 3306 1.1 christos move %xcc, %o4, %i2 3307 1.1 christos ldx [%o7+7*32], %o4 3308 1.1 christos move %xcc, %o5, %i3 3309 1.1 christos ldx [%o7+15*32],%o5 3310 1.1 christos movneg %xcc, %o4, %i2 3311 1.1 christos add %o7,16*32, %o7 3312 1.1 christos movneg %xcc, %o5, %i3 3313 1.1 christos ldx [%o7+0*32], %i4 3314 1.1 christos ldx [%o7+8*32], %i5 3315 1.1 christos ldx [%o7+1*32], %o4 3316 1.1 christos ldx [%o7+9*32], %o5 3317 1.1 christos movvs %icc, %o4, %i4 3318 1.1 christos ldx [%o7+2*32], %o4 3319 1.1 christos movvs %icc, %o5, %i5 3320 1.1 christos ldx [%o7+10*32],%o5 3321 1.1 christos move %icc, %o4, %i4 3322 1.1 christos ldx [%o7+3*32], %o4 3323 1.1 christos move %icc, %o5, %i5 3324 1.1 christos ldx [%o7+11*32],%o5 3325 1.1 christos movneg %icc, %o4, %i4 3326 1.1 christos ldx [%o7+4*32], %o4 3327 1.1 christos movneg %icc, %o5, %i5 3328 1.1 christos ldx [%o7+12*32],%o5 3329 1.1 christos movcs %xcc, %o4, %i4 3330 1.1 christos ldx [%o7+5*32],%o4 3331 1.1 christos movcs %xcc, %o5, %i5 3332 1.1 christos ldx [%o7+13*32],%o5 3333 1.1 christos movvs %xcc, %o4, %i4 3334 1.1 christos ldx [%o7+6*32], %o4 3335 1.1 christos movvs %xcc, %o5, %i5 3336 1.1 christos ldx [%o7+14*32],%o5 3337 1.1 christos move %xcc, %o4, %i4 3338 1.1 christos ldx [%o7+7*32], %o4 3339 1.1 christos move %xcc, %o5, %i5 3340 1.1 christos ldx [%o7+15*32],%o5 3341 1.1 christos movneg %xcc, %o4, %i4 3342 1.1 christos add %o7,16*32, %o7 3343 1.1 christos movneg %xcc, %o5, %i5 3344 1.1 christos ldx [%o7+0*32], %l0 3345 1.1 christos ldx [%o7+8*32], %l1 3346 1.1 christos ldx [%o7+1*32], %o4 3347 1.1 christos ldx [%o7+9*32], %o5 3348 1.1 christos movvs %icc, %o4, %l0 3349 1.1 christos ldx [%o7+2*32], %o4 3350 1.1 christos movvs %icc, %o5, %l1 3351 1.1 christos ldx [%o7+10*32],%o5 3352 1.1 christos move %icc, %o4, %l0 3353 1.1 christos ldx [%o7+3*32], %o4 3354 1.1 christos move %icc, %o5, %l1 3355 1.1 christos ldx [%o7+11*32],%o5 3356 1.1 christos movneg %icc, %o4, %l0 3357 1.1 christos ldx [%o7+4*32], %o4 3358 1.1 christos movneg %icc, %o5, %l1 3359 1.1 christos ldx [%o7+12*32],%o5 3360 1.1 christos movcs %xcc, %o4, %l0 3361 1.1 christos ldx [%o7+5*32],%o4 3362 1.1 christos movcs %xcc, %o5, %l1 3363 1.1 christos ldx [%o7+13*32],%o5 3364 1.1 christos movvs %xcc, %o4, %l0 3365 1.1 christos ldx [%o7+6*32], %o4 3366 1.1 christos movvs %xcc, %o5, %l1 3367 1.1 christos ldx [%o7+14*32],%o5 3368 1.1 christos move %xcc, %o4, %l0 3369 1.1 christos ldx [%o7+7*32], %o4 3370 1.1 christos move %xcc, %o5, %l1 3371 1.1 christos ldx [%o7+15*32],%o5 3372 1.1 christos movneg %xcc, %o4, %l0 3373 1.1 christos add %o7,16*32, %o7 3374 1.1 christos movneg %xcc, %o5, %l1 3375 1.1 christos ldx [%o7+0*32], %l2 3376 1.1 christos ldx [%o7+8*32], %l3 3377 1.1 christos ldx [%o7+1*32], %o4 3378 1.1 christos ldx [%o7+9*32], %o5 3379 1.1 christos movvs %icc, %o4, %l2 3380 1.1 christos ldx [%o7+2*32], %o4 3381 1.1 christos movvs %icc, %o5, %l3 3382 1.1 christos ldx [%o7+10*32],%o5 3383 1.1 christos move %icc, %o4, %l2 3384 1.1 christos ldx [%o7+3*32], %o4 3385 1.1 christos move %icc, %o5, %l3 3386 1.1 christos ldx [%o7+11*32],%o5 3387 1.1 christos movneg %icc, %o4, %l2 3388 1.1 christos ldx [%o7+4*32], %o4 3389 1.1 christos movneg %icc, %o5, %l3 3390 1.1 christos ldx [%o7+12*32],%o5 3391 1.1 christos movcs %xcc, %o4, %l2 3392 1.1 christos ldx [%o7+5*32],%o4 3393 1.1 christos movcs %xcc, %o5, %l3 3394 1.1 christos ldx [%o7+13*32],%o5 3395 1.1 christos movvs %xcc, %o4, %l2 3396 1.1 christos ldx [%o7+6*32], %o4 3397 1.1 christos movvs %xcc, %o5, %l3 3398 1.1 christos ldx [%o7+14*32],%o5 3399 1.1 christos move %xcc, %o4, %l2 3400 1.1 christos ldx [%o7+7*32], %o4 3401 1.1 christos move %xcc, %o5, %l3 3402 1.1 christos ldx [%o7+15*32],%o5 3403 1.1 christos movneg %xcc, %o4, %l2 3404 1.1 christos add %o7,16*32, %o7 3405 1.1 christos movneg %xcc, %o5, %l3 3406 1.1 christos ldx [%o7+0*32], %l4 3407 1.1 christos ldx [%o7+8*32], %l5 3408 1.1 christos ldx [%o7+1*32], %o4 3409 1.1 christos ldx [%o7+9*32], %o5 3410 1.1 christos movvs %icc, %o4, %l4 3411 1.1 christos ldx [%o7+2*32], %o4 3412 1.1 christos movvs %icc, %o5, %l5 3413 1.1 christos ldx [%o7+10*32],%o5 3414 1.1 christos move %icc, %o4, %l4 3415 1.1 christos ldx [%o7+3*32], %o4 3416 1.1 christos move %icc, %o5, %l5 3417 1.1 christos ldx [%o7+11*32],%o5 3418 1.1 christos movneg %icc, %o4, %l4 3419 1.1 christos ldx [%o7+4*32], %o4 3420 1.1 christos movneg %icc, %o5, %l5 3421 1.1 christos ldx [%o7+12*32],%o5 3422 1.1 christos movcs %xcc, %o4, %l4 3423 1.1 christos ldx [%o7+5*32],%o4 3424 1.1 christos movcs %xcc, %o5, %l5 3425 1.1 christos ldx [%o7+13*32],%o5 3426 1.1 christos movvs %xcc, %o4, %l4 3427 1.1 christos ldx [%o7+6*32], %o4 3428 1.1 christos movvs %xcc, %o5, %l5 3429 1.1 christos ldx [%o7+14*32],%o5 3430 1.1 christos move %xcc, %o4, %l4 3431 1.1 christos ldx [%o7+7*32], %o4 3432 1.1 christos move %xcc, %o5, %l5 3433 1.1 christos ldx [%o7+15*32],%o5 3434 1.1 christos movneg %xcc, %o4, %l4 3435 1.1 christos add %o7,16*32, %o7 3436 1.1 christos movneg %xcc, %o5, %l5 3437 1.1 christos ldx [%o7+0*32], %l6 3438 1.1 christos ldx [%o7+8*32], %l7 3439 1.1 christos ldx [%o7+1*32], %o4 3440 1.1 christos ldx [%o7+9*32], %o5 3441 1.1 christos movvs %icc, %o4, %l6 3442 1.1 christos ldx [%o7+2*32], %o4 3443 1.1 christos movvs %icc, %o5, %l7 3444 1.1 christos ldx [%o7+10*32],%o5 3445 1.1 christos move %icc, %o4, %l6 3446 1.1 christos ldx [%o7+3*32], %o4 3447 1.1 christos move %icc, %o5, %l7 3448 1.1 christos ldx [%o7+11*32],%o5 3449 1.1 christos movneg %icc, %o4, %l6 3450 1.1 christos ldx [%o7+4*32], %o4 3451 1.1 christos movneg %icc, %o5, %l7 3452 1.1 christos ldx [%o7+12*32],%o5 3453 1.1 christos movcs %xcc, %o4, %l6 3454 1.1 christos ldx [%o7+5*32],%o4 3455 1.1 christos movcs %xcc, %o5, %l7 3456 1.1 christos ldx [%o7+13*32],%o5 3457 1.1 christos movvs %xcc, %o4, %l6 3458 1.1 christos ldx [%o7+6*32], %o4 3459 1.1 christos movvs %xcc, %o5, %l7 3460 1.1 christos ldx [%o7+14*32],%o5 3461 1.1 christos move %xcc, %o4, %l6 3462 1.1 christos ldx [%o7+7*32], %o4 3463 1.1 christos move %xcc, %o5, %l7 3464 1.1 christos ldx [%o7+15*32],%o5 3465 1.1 christos movneg %xcc, %o4, %l6 3466 1.1 christos add %o7,16*32, %o7 3467 1.1 christos movneg %xcc, %o5, %l7 3468 1.1 christos save %sp,-128,%sp; or %g5,%fp,%fp 3469 1.1 christos ldx [%i7+0*32], %i0 3470 1.1 christos ldx [%i7+8*32], %i1 3471 1.1 christos ldx [%i7+1*32], %o4 3472 1.1 christos ldx [%i7+9*32], %o5 3473 1.1 christos movvs %icc, %o4, %i0 3474 1.1 christos ldx [%i7+2*32], %o4 3475 1.1 christos movvs %icc, %o5, %i1 3476 1.1 christos ldx [%i7+10*32],%o5 3477 1.1 christos move %icc, %o4, %i0 3478 1.1 christos ldx [%i7+3*32], %o4 3479 1.1 christos move %icc, %o5, %i1 3480 1.1 christos ldx [%i7+11*32],%o5 3481 1.1 christos movneg %icc, %o4, %i0 3482 1.1 christos ldx [%i7+4*32], %o4 3483 1.1 christos movneg %icc, %o5, %i1 3484 1.1 christos ldx [%i7+12*32],%o5 3485 1.1 christos movcs %xcc, %o4, %i0 3486 1.1 christos ldx [%i7+5*32],%o4 3487 1.1 christos movcs %xcc, %o5, %i1 3488 1.1 christos ldx [%i7+13*32],%o5 3489 1.1 christos movvs %xcc, %o4, %i0 3490 1.1 christos ldx [%i7+6*32], %o4 3491 1.1 christos movvs %xcc, %o5, %i1 3492 1.1 christos ldx [%i7+14*32],%o5 3493 1.1 christos move %xcc, %o4, %i0 3494 1.1 christos ldx [%i7+7*32], %o4 3495 1.1 christos move %xcc, %o5, %i1 3496 1.1 christos ldx [%i7+15*32],%o5 3497 1.1 christos movneg %xcc, %o4, %i0 3498 1.1 christos add %i7,16*32, %i7 3499 1.1 christos movneg %xcc, %o5, %i1 3500 1.1 christos ldx [%i7+0*32], %i2 3501 1.1 christos ldx [%i7+8*32], %i3 3502 1.1 christos ldx [%i7+1*32], %o4 3503 1.1 christos ldx [%i7+9*32], %o5 3504 1.1 christos movvs %icc, %o4, %i2 3505 1.1 christos ldx [%i7+2*32], %o4 3506 1.1 christos movvs %icc, %o5, %i3 3507 1.1 christos ldx [%i7+10*32],%o5 3508 1.1 christos move %icc, %o4, %i2 3509 1.1 christos ldx [%i7+3*32], %o4 3510 1.1 christos move %icc, %o5, %i3 3511 1.1 christos ldx [%i7+11*32],%o5 3512 1.1 christos movneg %icc, %o4, %i2 3513 1.1 christos ldx [%i7+4*32], %o4 3514 1.1 christos movneg %icc, %o5, %i3 3515 1.1 christos ldx [%i7+12*32],%o5 3516 1.1 christos movcs %xcc, %o4, %i2 3517 1.1 christos ldx [%i7+5*32],%o4 3518 1.1 christos movcs %xcc, %o5, %i3 3519 1.1 christos ldx [%i7+13*32],%o5 3520 1.1 christos movvs %xcc, %o4, %i2 3521 1.1 christos ldx [%i7+6*32], %o4 3522 1.1 christos movvs %xcc, %o5, %i3 3523 1.1 christos ldx [%i7+14*32],%o5 3524 1.1 christos move %xcc, %o4, %i2 3525 1.1 christos ldx [%i7+7*32], %o4 3526 1.1 christos move %xcc, %o5, %i3 3527 1.1 christos ldx [%i7+15*32],%o5 3528 1.1 christos movneg %xcc, %o4, %i2 3529 1.1 christos add %i7,16*32, %i7 3530 1.1 christos movneg %xcc, %o5, %i3 3531 1.1 christos ldx [%i7+0*32], %i4 3532 1.1 christos ldx [%i7+8*32], %i5 3533 1.1 christos ldx [%i7+1*32], %o4 3534 1.1 christos ldx [%i7+9*32], %o5 3535 1.1 christos movvs %icc, %o4, %i4 3536 1.1 christos ldx [%i7+2*32], %o4 3537 1.1 christos movvs %icc, %o5, %i5 3538 1.1 christos ldx [%i7+10*32],%o5 3539 1.1 christos move %icc, %o4, %i4 3540 1.1 christos ldx [%i7+3*32], %o4 3541 1.1 christos move %icc, %o5, %i5 3542 1.1 christos ldx [%i7+11*32],%o5 3543 1.1 christos movneg %icc, %o4, %i4 3544 1.1 christos ldx [%i7+4*32], %o4 3545 1.1 christos movneg %icc, %o5, %i5 3546 1.1 christos ldx [%i7+12*32],%o5 3547 1.1 christos movcs %xcc, %o4, %i4 3548 1.1 christos ldx [%i7+5*32],%o4 3549 1.1 christos movcs %xcc, %o5, %i5 3550 1.1 christos ldx [%i7+13*32],%o5 3551 1.1 christos movvs %xcc, %o4, %i4 3552 1.1 christos ldx [%i7+6*32], %o4 3553 1.1 christos movvs %xcc, %o5, %i5 3554 1.1 christos ldx [%i7+14*32],%o5 3555 1.1 christos move %xcc, %o4, %i4 3556 1.1 christos ldx [%i7+7*32], %o4 3557 1.1 christos move %xcc, %o5, %i5 3558 1.1 christos ldx [%i7+15*32],%o5 3559 1.1 christos movneg %xcc, %o4, %i4 3560 1.1 christos add %i7,16*32, %i7 3561 1.1 christos movneg %xcc, %o5, %i5 3562 1.1 christos ldx [%i7+0*32], %l0 3563 1.1 christos ldx [%i7+8*32], %l1 3564 1.1 christos ldx [%i7+1*32], %o4 3565 1.1 christos ldx [%i7+9*32], %o5 3566 1.1 christos movvs %icc, %o4, %l0 3567 1.1 christos ldx [%i7+2*32], %o4 3568 1.1 christos movvs %icc, %o5, %l1 3569 1.1 christos ldx [%i7+10*32],%o5 3570 1.1 christos move %icc, %o4, %l0 3571 1.1 christos ldx [%i7+3*32], %o4 3572 1.1 christos move %icc, %o5, %l1 3573 1.1 christos ldx [%i7+11*32],%o5 3574 1.1 christos movneg %icc, %o4, %l0 3575 1.1 christos ldx [%i7+4*32], %o4 3576 1.1 christos movneg %icc, %o5, %l1 3577 1.1 christos ldx [%i7+12*32],%o5 3578 1.1 christos movcs %xcc, %o4, %l0 3579 1.1 christos ldx [%i7+5*32],%o4 3580 1.1 christos movcs %xcc, %o5, %l1 3581 1.1 christos ldx [%i7+13*32],%o5 3582 1.1 christos movvs %xcc, %o4, %l0 3583 1.1 christos ldx [%i7+6*32], %o4 3584 1.1 christos movvs %xcc, %o5, %l1 3585 1.1 christos ldx [%i7+14*32],%o5 3586 1.1 christos move %xcc, %o4, %l0 3587 1.1 christos ldx [%i7+7*32], %o4 3588 1.1 christos move %xcc, %o5, %l1 3589 1.1 christos ldx [%i7+15*32],%o5 3590 1.1 christos movneg %xcc, %o4, %l0 3591 1.1 christos add %i7,16*32, %i7 3592 1.1 christos movneg %xcc, %o5, %l1 3593 1.1 christos ldx [%i7+0*32], %l2 3594 1.1 christos ldx [%i7+8*32], %l3 3595 1.1 christos ldx [%i7+1*32], %o4 3596 1.1 christos ldx [%i7+9*32], %o5 3597 1.1 christos movvs %icc, %o4, %l2 3598 1.1 christos ldx [%i7+2*32], %o4 3599 1.1 christos movvs %icc, %o5, %l3 3600 1.1 christos ldx [%i7+10*32],%o5 3601 1.1 christos move %icc, %o4, %l2 3602 1.1 christos ldx [%i7+3*32], %o4 3603 1.1 christos move %icc, %o5, %l3 3604 1.1 christos ldx [%i7+11*32],%o5 3605 1.1 christos movneg %icc, %o4, %l2 3606 1.1 christos ldx [%i7+4*32], %o4 3607 1.1 christos movneg %icc, %o5, %l3 3608 1.1 christos ldx [%i7+12*32],%o5 3609 1.1 christos movcs %xcc, %o4, %l2 3610 1.1 christos ldx [%i7+5*32],%o4 3611 1.1 christos movcs %xcc, %o5, %l3 3612 1.1 christos ldx [%i7+13*32],%o5 3613 1.1 christos movvs %xcc, %o4, %l2 3614 1.1 christos ldx [%i7+6*32], %o4 3615 1.1 christos movvs %xcc, %o5, %l3 3616 1.1 christos ldx [%i7+14*32],%o5 3617 1.1 christos move %xcc, %o4, %l2 3618 1.1 christos ldx [%i7+7*32], %o4 3619 1.1 christos move %xcc, %o5, %l3 3620 1.1 christos ldx [%i7+15*32],%o5 3621 1.1 christos movneg %xcc, %o4, %l2 3622 1.1 christos add %i7,16*32, %i7 3623 1.1 christos movneg %xcc, %o5, %l3 3624 1.1 christos ldx [%i7+0*32], %l4 3625 1.1 christos ldx [%i7+8*32], %l5 3626 1.1 christos ldx [%i7+1*32], %o4 3627 1.1 christos ldx [%i7+9*32], %o5 3628 1.1 christos movvs %icc, %o4, %l4 3629 1.1 christos ldx [%i7+2*32], %o4 3630 1.1 christos movvs %icc, %o5, %l5 3631 1.1 christos ldx [%i7+10*32],%o5 3632 1.1 christos move %icc, %o4, %l4 3633 1.1 christos ldx [%i7+3*32], %o4 3634 1.1 christos move %icc, %o5, %l5 3635 1.1 christos ldx [%i7+11*32],%o5 3636 1.1 christos movneg %icc, %o4, %l4 3637 1.1 christos ldx [%i7+4*32], %o4 3638 1.1 christos movneg %icc, %o5, %l5 3639 1.1 christos ldx [%i7+12*32],%o5 3640 1.1 christos movcs %xcc, %o4, %l4 3641 1.1 christos ldx [%i7+5*32],%o4 3642 1.1 christos movcs %xcc, %o5, %l5 3643 1.1 christos ldx [%i7+13*32],%o5 3644 1.1 christos movvs %xcc, %o4, %l4 3645 1.1 christos ldx [%i7+6*32], %o4 3646 1.1 christos movvs %xcc, %o5, %l5 3647 1.1 christos ldx [%i7+14*32],%o5 3648 1.1 christos move %xcc, %o4, %l4 3649 1.1 christos ldx [%i7+7*32], %o4 3650 1.1 christos move %xcc, %o5, %l5 3651 1.1 christos ldx [%i7+15*32],%o5 3652 1.1 christos movneg %xcc, %o4, %l4 3653 1.1 christos add %i7,16*32, %i7 3654 1.1 christos movneg %xcc, %o5, %l5 3655 1.1 christos ldx [%i7+0*32], %l6 3656 1.1 christos ldx [%i7+8*32], %l7 3657 1.1 christos ldx [%i7+1*32], %o4 3658 1.1 christos ldx [%i7+9*32], %o5 3659 1.1 christos movvs %icc, %o4, %l6 3660 1.1 christos ldx [%i7+2*32], %o4 3661 1.1 christos movvs %icc, %o5, %l7 3662 1.1 christos ldx [%i7+10*32],%o5 3663 1.1 christos move %icc, %o4, %l6 3664 1.1 christos ldx [%i7+3*32], %o4 3665 1.1 christos move %icc, %o5, %l7 3666 1.1 christos ldx [%i7+11*32],%o5 3667 1.1 christos movneg %icc, %o4, %l6 3668 1.1 christos ldx [%i7+4*32], %o4 3669 1.1 christos movneg %icc, %o5, %l7 3670 1.1 christos ldx [%i7+12*32],%o5 3671 1.1 christos movcs %xcc, %o4, %l6 3672 1.1 christos ldx [%i7+5*32],%o4 3673 1.1 christos movcs %xcc, %o5, %l7 3674 1.1 christos ldx [%i7+13*32],%o5 3675 1.1 christos movvs %xcc, %o4, %l6 3676 1.1 christos ldx [%i7+6*32], %o4 3677 1.1 christos movvs %xcc, %o5, %l7 3678 1.1 christos ldx [%i7+14*32],%o5 3679 1.1 christos move %xcc, %o4, %l6 3680 1.1 christos ldx [%i7+7*32], %o4 3681 1.1 christos move %xcc, %o5, %l7 3682 1.1 christos ldx [%i7+15*32],%o5 3683 1.1 christos movneg %xcc, %o4, %l6 3684 1.1 christos add %i7,16*32, %i7 3685 1.1 christos movneg %xcc, %o5, %l7 3686 1.1 christos ldx [%i7+0*32], %o0 3687 1.1 christos ldx [%i7+8*32], %o1 3688 1.1 christos ldx [%i7+1*32], %o4 3689 1.1 christos ldx [%i7+9*32], %o5 3690 1.1 christos movvs %icc, %o4, %o0 3691 1.1 christos ldx [%i7+2*32], %o4 3692 1.1 christos movvs %icc, %o5, %o1 3693 1.1 christos ldx [%i7+10*32],%o5 3694 1.1 christos move %icc, %o4, %o0 3695 1.1 christos ldx [%i7+3*32], %o4 3696 1.1 christos move %icc, %o5, %o1 3697 1.1 christos ldx [%i7+11*32],%o5 3698 1.1 christos movneg %icc, %o4, %o0 3699 1.1 christos ldx [%i7+4*32], %o4 3700 1.1 christos movneg %icc, %o5, %o1 3701 1.1 christos ldx [%i7+12*32],%o5 3702 1.1 christos movcs %xcc, %o4, %o0 3703 1.1 christos ldx [%i7+5*32],%o4 3704 1.1 christos movcs %xcc, %o5, %o1 3705 1.1 christos ldx [%i7+13*32],%o5 3706 1.1 christos movvs %xcc, %o4, %o0 3707 1.1 christos ldx [%i7+6*32], %o4 3708 1.1 christos movvs %xcc, %o5, %o1 3709 1.1 christos ldx [%i7+14*32],%o5 3710 1.1 christos move %xcc, %o4, %o0 3711 1.1 christos ldx [%i7+7*32], %o4 3712 1.1 christos move %xcc, %o5, %o1 3713 1.1 christos ldx [%i7+15*32],%o5 3714 1.1 christos movneg %xcc, %o4, %o0 3715 1.1 christos add %i7,16*32, %i7 3716 1.1 christos movneg %xcc, %o5, %o1 3717 1.1 christos ldx [%i7+0*32], %o2 3718 1.1 christos ldx [%i7+8*32], %o3 3719 1.1 christos ldx [%i7+1*32], %o4 3720 1.1 christos ldx [%i7+9*32], %o5 3721 1.1 christos movvs %icc, %o4, %o2 3722 1.1 christos ldx [%i7+2*32], %o4 3723 1.1 christos movvs %icc, %o5, %o3 3724 1.1 christos ldx [%i7+10*32],%o5 3725 1.1 christos move %icc, %o4, %o2 3726 1.1 christos ldx [%i7+3*32], %o4 3727 1.1 christos move %icc, %o5, %o3 3728 1.1 christos ldx [%i7+11*32],%o5 3729 1.1 christos movneg %icc, %o4, %o2 3730 1.1 christos ldx [%i7+4*32], %o4 3731 1.1 christos movneg %icc, %o5, %o3 3732 1.1 christos ldx [%i7+12*32],%o5 3733 1.1 christos movcs %xcc, %o4, %o2 3734 1.1 christos ldx [%i7+5*32],%o4 3735 1.1 christos movcs %xcc, %o5, %o3 3736 1.1 christos ldx [%i7+13*32],%o5 3737 1.1 christos movvs %xcc, %o4, %o2 3738 1.1 christos ldx [%i7+6*32], %o4 3739 1.1 christos movvs %xcc, %o5, %o3 3740 1.1 christos ldx [%i7+14*32],%o5 3741 1.1 christos move %xcc, %o4, %o2 3742 1.1 christos ldx [%i7+7*32], %o4 3743 1.1 christos move %xcc, %o5, %o3 3744 1.1 christos ldx [%i7+15*32],%o5 3745 1.1 christos movneg %xcc, %o4, %o2 3746 1.1 christos add %i7,16*32, %i7 3747 1.1 christos movneg %xcc, %o5, %o3 3748 1.1 christos srax %g4, 32, %o4 ! unpack %g4 3749 1.1 christos srl %g4, %g0, %o5 3750 1.1 christos sub %o4, 5, %o4 3751 1.1 christos mov %g3, %i7 3752 1.1 christos sllx %o4, 32, %g4 ! re-pack %g4 3753 1.1 christos or %o5, %g4, %g4 3754 1.1 christos srl %o5, %o4, %o5 3755 1.1 christos srl %o5, 2, %o4 3756 1.1 christos and %o5, 3, %o5 3757 1.1 christos and %o4, 7, %o4 3758 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 3759 1.1 christos add %o5, %i7, %i7 ! of the pwrtbl 3760 1.1 christos or %g0, 1, %o5 3761 1.1 christos sll %o5, %o4, %o4 3762 1.1 christos .word 0x81b02940+32-1 ! montsqr 32-1 3763 1.1 christos fbu,pn %fcc3,.Labort_32 3764 1.1 christos #ifndef __arch64__ 3765 1.1 christos and %fp,%g5,%g5 3766 1.1 christos brz,pn %g5,.Labort_32 3767 1.1 christos #endif 3768 1.1 christos nop 3769 1.1 christos .word 0x81b02940+32-1 ! montsqr 32-1 3770 1.1 christos fbu,pn %fcc3,.Labort_32 3771 1.1 christos #ifndef __arch64__ 3772 1.1 christos and %fp,%g5,%g5 3773 1.1 christos brz,pn %g5,.Labort_32 3774 1.1 christos #endif 3775 1.1 christos nop 3776 1.1 christos .word 0x81b02940+32-1 ! montsqr 32-1 3777 1.1 christos fbu,pn %fcc3,.Labort_32 3778 1.1 christos #ifndef __arch64__ 3779 1.1 christos and %fp,%g5,%g5 3780 1.1 christos brz,pn %g5,.Labort_32 3781 1.1 christos #endif 3782 1.1 christos nop 3783 1.1 christos .word 0x81b02940+32-1 ! montsqr 32-1 3784 1.1 christos fbu,pn %fcc3,.Labort_32 3785 1.1 christos #ifndef __arch64__ 3786 1.1 christos and %fp,%g5,%g5 3787 1.1 christos brz,pn %g5,.Labort_32 3788 1.1 christos #endif 3789 1.1 christos nop 3790 1.1 christos .word 0x81b02940+32-1 ! montsqr 32-1 3791 1.1 christos fbu,pn %fcc3,.Labort_32 3792 1.1 christos #ifndef __arch64__ 3793 1.1 christos and %fp,%g5,%g5 3794 1.1 christos brz,pn %g5,.Labort_32 3795 1.1 christos #endif 3796 1.1 christos nop 3797 1.1 christos wr %o4, %g0, %ccr 3798 1.1 christos .word 0x81b02920+32-1 ! montmul 32-1 3799 1.1 christos fbu,pn %fcc3,.Labort_32 3800 1.1 christos #ifndef __arch64__ 3801 1.1 christos and %fp,%g5,%g5 3802 1.1 christos brz,pn %g5,.Labort_32 3803 1.1 christos #endif 3804 1.1 christos 3805 1.1 christos srax %g4, 32, %o4 3806 1.1 christos #ifdef __arch64__ 3807 1.1 christos brgez %o4,.Lstride_32 3808 1.1 christos restore 3809 1.1 christos restore 3810 1.1 christos restore 3811 1.1 christos restore 3812 1.1 christos restore 3813 1.1 christos #else 3814 1.1 christos brgez %o4,.Lstride_32 3815 1.1 christos restore; and %fp,%g5,%g5 3816 1.1 christos restore; and %fp,%g5,%g5 3817 1.1 christos restore; and %fp,%g5,%g5 3818 1.1 christos restore; and %fp,%g5,%g5 3819 1.1 christos brz,pn %g5,.Labort1_32 3820 1.1 christos restore 3821 1.1 christos #endif 3822 1.1 christos .word 0x81b02310 !movxtod %l0,%f0 3823 1.1 christos .word 0x85b02311 !movxtod %l1,%f2 3824 1.1 christos .word 0x89b02312 !movxtod %l2,%f4 3825 1.1 christos .word 0x8db02313 !movxtod %l3,%f6 3826 1.1 christos .word 0x91b02314 !movxtod %l4,%f8 3827 1.1 christos .word 0x95b02315 !movxtod %l5,%f10 3828 1.1 christos .word 0x99b02316 !movxtod %l6,%f12 3829 1.1 christos .word 0x9db02317 !movxtod %l7,%f14 3830 1.1 christos .word 0xa1b02308 !movxtod %o0,%f16 3831 1.1 christos .word 0xa5b02309 !movxtod %o1,%f18 3832 1.1 christos .word 0xa9b0230a !movxtod %o2,%f20 3833 1.1 christos .word 0xadb0230b !movxtod %o3,%f22 3834 1.1 christos .word 0xbbb0230c !movxtod %o4,%f60 3835 1.1 christos .word 0xbfb0230d !movxtod %o5,%f62 3836 1.1 christos #ifdef __arch64__ 3837 1.1 christos restore 3838 1.1 christos #else 3839 1.1 christos and %fp,%g5,%g5 3840 1.1 christos restore 3841 1.1 christos and %g5,1,%o7 3842 1.1 christos and %fp,%g5,%g5 3843 1.1 christos srl %fp,0,%fp ! just in case? 3844 1.1 christos or %o7,%g5,%g5 3845 1.1 christos brz,a,pn %g5,.Ldone_32 3846 1.1 christos mov 0,%i0 ! return failure 3847 1.1 christos #endif 3848 1.1 christos std %f0,[%g1+0*8] 3849 1.1 christos std %f2,[%g1+1*8] 3850 1.1 christos std %f4,[%g1+2*8] 3851 1.1 christos std %f6,[%g1+3*8] 3852 1.1 christos std %f8,[%g1+4*8] 3853 1.1 christos std %f10,[%g1+5*8] 3854 1.1 christos std %f12,[%g1+6*8] 3855 1.1 christos std %f14,[%g1+7*8] 3856 1.1 christos std %f16,[%g1+8*8] 3857 1.1 christos std %f18,[%g1+9*8] 3858 1.1 christos std %f20,[%g1+10*8] 3859 1.1 christos std %f22,[%g1+11*8] 3860 1.1 christos std %f60,[%g1+12*8] 3861 1.1 christos std %f62,[%g1+13*8] 3862 1.1 christos std %f24,[%g1+14*8] 3863 1.1 christos std %f26,[%g1+15*8] 3864 1.1 christos std %f28,[%g1+16*8] 3865 1.1 christos std %f30,[%g1+17*8] 3866 1.1 christos std %f32,[%g1+18*8] 3867 1.1 christos std %f34,[%g1+19*8] 3868 1.1 christos std %f36,[%g1+20*8] 3869 1.1 christos std %f38,[%g1+21*8] 3870 1.1 christos std %f40,[%g1+22*8] 3871 1.1 christos std %f42,[%g1+23*8] 3872 1.1 christos std %f44,[%g1+24*8] 3873 1.1 christos std %f46,[%g1+25*8] 3874 1.1 christos std %f48,[%g1+26*8] 3875 1.1 christos std %f50,[%g1+27*8] 3876 1.1 christos std %f52,[%g1+28*8] 3877 1.1 christos std %f54,[%g1+29*8] 3878 1.1 christos std %f56,[%g1+30*8] 3879 1.1 christos std %f58,[%g1+31*8] 3880 1.1 christos mov 1,%i0 ! return success 3881 1.1 christos .Ldone_32: 3882 1.1 christos ret 3883 1.1 christos restore 3884 1.1 christos 3885 1.1 christos .Labort_32: 3886 1.1 christos restore 3887 1.1 christos restore 3888 1.1 christos restore 3889 1.1 christos restore 3890 1.1 christos restore 3891 1.1 christos .Labort1_32: 3892 1.1 christos restore 3893 1.1 christos 3894 1.1 christos mov 0,%i0 ! return failure 3895 1.1 christos ret 3896 1.1 christos restore 3897 1.1 christos .type bn_pwr5_mont_t4_32, #function 3898 1.1 christos .size bn_pwr5_mont_t4_32, .-bn_pwr5_mont_t4_32 3899 1.1 christos .globl bn_mul_mont_t4 3900 1.1 christos .align 32 3901 1.1 christos bn_mul_mont_t4: 3902 1.1 christos add %sp, STACK_BIAS, %g4 ! real top of stack 3903 1.1 christos sll %o5, 3, %o5 ! size in bytes 3904 1.1 christos add %o5, 63, %g1 3905 1.1 christos andn %g1, 63, %g1 ! buffer size rounded up to 64 bytes 3906 1.1 christos sub %g4, %g1, %g1 3907 1.1 christos andn %g1, 63, %g1 ! align at 64 byte 3908 1.1 christos sub %g1, STACK_FRAME, %g1 ! new top of stack 3909 1.1 christos sub %g1, %g4, %g1 3910 1.1 christos 3911 1.1 christos save %sp, %g1, %sp 3912 1.1 christos ld [%i4+0], %l0 ! pull n0[0..1] value 3913 1.1 christos ld [%i4+4], %l1 3914 1.1 christos add %sp, STACK_BIAS+STACK_FRAME, %l5 3915 1.1 christos ldx [%i2+0], %g2 ! m0=bp[0] 3916 1.1 christos sllx %l1, 32, %g1 3917 1.1 christos add %i2, 8, %i2 3918 1.1 christos or %l0, %g1, %g1 3919 1.1 christos 3920 1.1 christos ldx [%i1+0], %o2 ! ap[0] 3922 1.1 christos 3923 1.1 christos mulx %o2, %g2, %g4 ! ap[0]*bp[0] 3924 1.1 christos .word 0x8bb282c2 !umulxhi %o2,%g2,%g5 3925 1.1 christos 3926 1.1 christos ldx [%i1+8], %o2 ! ap[1] 3927 1.1 christos add %i1, 16, %i1 3928 1.1 christos ldx [%i3+0], %o4 ! np[0] 3929 1.1 christos 3930 1.1 christos mulx %g4, %g1, %g3 ! "tp[0]"*n0 3931 1.1 christos 3932 1.1 christos mulx %o2, %g2, %o3 ! ap[1]*bp[0] 3933 1.1 christos .word 0x95b282c2 !umulxhi %o2,%g2,%o2 ! ahi=aj 3934 1.1 christos 3935 1.1 christos mulx %o4, %g3, %o0 ! np[0]*m1 3936 1.1 christos .word 0x93b302c3 !umulxhi %o4,%g3,%o1 3937 1.1 christos 3938 1.1 christos ldx [%i3+8], %o4 ! np[1] 3939 1.1 christos 3940 1.1 christos addcc %g4, %o0, %o0 3941 1.1 christos add %i3, 16, %i3 3942 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 3943 1.1 christos 3944 1.1 christos mulx %o4, %g3, %o5 ! np[1]*m1 3945 1.1 christos .word 0x99b302c3 !umulxhi %o4,%g3,%o4 ! nhi=nj 3946 1.1 christos 3947 1.1 christos ba .L1st 3949 1.1 christos sub %i5, 24, %l4 ! cnt=num-3 3950 1.1 christos 3951 1.1 christos .align 16 3952 1.1 christos .L1st: 3953 1.1 christos addcc %o3, %g5, %g4 3954 1.1 christos .word 0x8bb28220 !addxc %o2,%g0,%g5 3955 1.1 christos 3956 1.1 christos ldx [%i1+0], %o2 ! ap[j] 3957 1.1 christos addcc %o5, %o1, %o0 3958 1.1 christos add %i1, 8, %i1 3959 1.1 christos .word 0x93b30220 !addxc %o4,%g0,%o1 ! nhi=nj 3960 1.1 christos 3961 1.1 christos ldx [%i3+0], %o4 ! np[j] 3962 1.1 christos mulx %o2, %g2, %o3 ! ap[j]*bp[0] 3963 1.1 christos add %i3, 8, %i3 3964 1.1 christos .word 0x95b282c2 !umulxhi %o2,%g2,%o2 ! ahi=aj 3965 1.1 christos 3966 1.1 christos mulx %o4, %g3, %o5 ! np[j]*m1 3967 1.1 christos addcc %g4, %o0, %o0 ! np[j]*m1+ap[j]*bp[0] 3968 1.1 christos .word 0x99b302c3 !umulxhi %o4,%g3,%o4 ! nhi=nj 3969 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 3970 1.1 christos stxa %o0, [%l5]0xe2 ! tp[j-1] 3971 1.1 christos add %l5, 8, %l5 ! tp++ 3972 1.1 christos 3973 1.1 christos brnz,pt %l4, .L1st 3974 1.1 christos sub %l4, 8, %l4 ! j-- 3975 1.1 christos !.L1st 3976 1.1 christos addcc %o3, %g5, %g4 3977 1.1 christos .word 0x8bb28220 !addxc %o2,%g0,%g5 ! ahi=aj 3978 1.1 christos 3979 1.1 christos addcc %o5, %o1, %o0 3980 1.1 christos .word 0x93b30220 !addxc %o4,%g0,%o1 3981 1.1 christos addcc %g4, %o0, %o0 ! np[j]*m1+ap[j]*bp[0] 3982 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 3983 1.1 christos stxa %o0, [%l5]0xe2 ! tp[j-1] 3984 1.1 christos add %l5, 8, %l5 3985 1.1 christos 3986 1.1 christos addcc %g5, %o1, %o1 3987 1.1 christos .word 0xa1b00220 !addxc %g0,%g0,%l0 ! upmost overflow bit 3988 1.1 christos stxa %o1, [%l5]0xe2 3989 1.1 christos add %l5, 8, %l5 3990 1.1 christos 3991 1.1 christos ba .Louter 3993 1.1 christos sub %i5, 16, %l1 ! i=num-2 3994 1.1 christos 3995 1.1 christos .align 16 3996 1.1 christos .Louter: 3997 1.1 christos ldx [%i2+0], %g2 ! m0=bp[i] 3998 1.1 christos add %i2, 8, %i2 3999 1.1 christos 4000 1.1 christos sub %i1, %i5, %i1 ! rewind 4001 1.1 christos sub %i3, %i5, %i3 4002 1.1 christos sub %l5, %i5, %l5 4003 1.1 christos 4004 1.1 christos ldx [%i1+0], %o2 ! ap[0] 4005 1.1 christos ldx [%i3+0], %o4 ! np[0] 4006 1.1 christos 4007 1.1 christos mulx %o2, %g2, %g4 ! ap[0]*bp[i] 4008 1.1 christos ldx [%l5], %o7 ! tp[0] 4009 1.1 christos .word 0x8bb282c2 !umulxhi %o2,%g2,%g5 4010 1.1 christos ldx [%i1+8], %o2 ! ap[1] 4011 1.1 christos addcc %g4, %o7, %g4 ! ap[0]*bp[i]+tp[0] 4012 1.1 christos mulx %o2, %g2, %o3 ! ap[1]*bp[i] 4013 1.1 christos .word 0x8bb00225 !addxc %g0,%g5,%g5 4014 1.1 christos mulx %g4, %g1, %g3 ! tp[0]*n0 4015 1.1 christos .word 0x95b282c2 !umulxhi %o2,%g2,%o2 ! ahi=aj 4016 1.1 christos mulx %o4, %g3, %o0 ! np[0]*m1 4017 1.1 christos add %i1, 16, %i1 4018 1.1 christos .word 0x93b302c3 !umulxhi %o4,%g3,%o1 4019 1.1 christos ldx [%i3+8], %o4 ! np[1] 4020 1.1 christos add %i3, 16, %i3 4021 1.1 christos addcc %o0, %g4, %o0 4022 1.1 christos mulx %o4, %g3, %o5 ! np[1]*m1 4023 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 4024 1.1 christos .word 0x99b302c3 !umulxhi %o4,%g3,%o4 ! nhi=nj 4025 1.1 christos 4026 1.1 christos ba .Linner 4028 1.1 christos sub %i5, 24, %l4 ! cnt=num-3 4029 1.1 christos .align 16 4030 1.1 christos .Linner: 4031 1.1 christos addcc %o3, %g5, %g4 4032 1.1 christos ldx [%l5+8], %o7 ! tp[j] 4033 1.1 christos .word 0x8bb28220 !addxc %o2,%g0,%g5 ! ahi=aj 4034 1.1 christos ldx [%i1+0], %o2 ! ap[j] 4035 1.1 christos add %i1, 8, %i1 4036 1.1 christos addcc %o5, %o1, %o0 4037 1.1 christos mulx %o2, %g2, %o3 ! ap[j]*bp[i] 4038 1.1 christos .word 0x93b30220 !addxc %o4,%g0,%o1 ! nhi=nj 4039 1.1 christos ldx [%i3+0], %o4 ! np[j] 4040 1.1 christos add %i3, 8, %i3 4041 1.1 christos .word 0x95b282c2 !umulxhi %o2,%g2,%o2 ! ahi=aj 4042 1.1 christos addcc %g4, %o7, %g4 ! ap[j]*bp[i]+tp[j] 4043 1.1 christos mulx %o4, %g3, %o5 ! np[j]*m1 4044 1.1 christos .word 0x8bb00225 !addxc %g0,%g5,%g5 4045 1.1 christos .word 0x99b302c3 !umulxhi %o4,%g3,%o4 ! nhi=nj 4046 1.1 christos addcc %o0, %g4, %o0 ! np[j]*m1+ap[j]*bp[i]+tp[j] 4047 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 4048 1.1 christos stx %o0, [%l5] ! tp[j-1] 4049 1.1 christos add %l5, 8, %l5 4050 1.1 christos brnz,pt %l4, .Linner 4051 1.1 christos sub %l4, 8, %l4 4052 1.1 christos !.Linner 4053 1.1 christos ldx [%l5+8], %o7 ! tp[j] 4054 1.1 christos addcc %o3, %g5, %g4 4055 1.1 christos .word 0x8bb28220 !addxc %o2,%g0,%g5 ! ahi=aj 4056 1.1 christos addcc %g4, %o7, %g4 ! ap[j]*bp[i]+tp[j] 4057 1.1 christos .word 0x8bb00225 !addxc %g0,%g5,%g5 4058 1.1 christos 4059 1.1 christos addcc %o5, %o1, %o0 4060 1.1 christos .word 0x93b30220 !addxc %o4,%g0,%o1 ! nhi=nj 4061 1.1 christos addcc %o0, %g4, %o0 ! np[j]*m1+ap[j]*bp[i]+tp[j] 4062 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 4063 1.1 christos stx %o0, [%l5] ! tp[j-1] 4064 1.1 christos 4065 1.1 christos subcc %g0, %l0, %g0 ! move upmost overflow to CCR.xcc 4066 1.1 christos .word 0x93b24265 !addxccc %o1,%g5,%o1 4067 1.1 christos .word 0xa1b00220 !addxc %g0,%g0,%l0 4068 1.1 christos stx %o1, [%l5+8] 4069 1.1 christos add %l5, 16, %l5 4070 1.1 christos 4071 1.1 christos brnz,pt %l1, .Louter 4072 1.1 christos sub %l1, 8, %l1 4073 1.1 christos 4074 1.1 christos sub %i1, %i5, %i1 ! rewind 4076 1.1 christos sub %i3, %i5, %i3 4077 1.1 christos sub %l5, %i5, %l5 4078 1.1 christos ba .Lsub 4079 1.1 christos subcc %i5, 8, %l4 ! cnt=num-1 and clear CCR.xcc 4080 1.1 christos 4081 1.1 christos .align 16 4082 1.1 christos .Lsub: 4083 1.1 christos ldx [%l5], %o7 4084 1.1 christos add %l5, 8, %l5 4085 1.1 christos ldx [%i3+0], %o4 4086 1.1 christos add %i3, 8, %i3 4087 1.1 christos subccc %o7, %o4, %l2 ! tp[j]-np[j] 4088 1.1 christos srlx %o7, 32, %o7 4089 1.1 christos srlx %o4, 32, %o4 4090 1.1 christos subccc %o7, %o4, %l3 4091 1.1 christos add %i0, 8, %i0 4092 1.1 christos st %l2, [%i0-4] ! reverse order 4093 1.1 christos st %l3, [%i0-8] 4094 1.1 christos brnz,pt %l4, .Lsub 4095 1.1 christos sub %l4, 8, %l4 4096 1.1.1.2 christos 4097 1.1 christos sub %i3, %i5, %i3 ! rewind 4098 1.1 christos sub %l5, %i5, %l5 4099 1.1 christos sub %i0, %i5, %i0 4100 1.1 christos 4101 1.1.1.2 christos subccc %l0, %g0, %l0 ! handle upmost overflow bit 4102 1.1.1.2 christos ba .Lcopy 4103 1.1.1.2 christos sub %i5, 8, %l4 4104 1.1 christos 4105 1.1 christos .align 16 4106 1.1.1.2 christos .Lcopy: ! conditional copy 4107 1.1 christos ldx [%l5], %o7 4108 1.1 christos ldx [%i0+0], %l2 4109 1.1 christos stx %g0, [%l5] ! zap 4110 1.1 christos add %l5, 8, %l5 4111 1.1 christos movcs %icc, %o7, %l2 4112 1.1 christos stx %l2, [%i0+0] 4113 1.1 christos add %i0, 8, %i0 4114 1.1 christos brnz %l4, .Lcopy 4115 1.1 christos sub %l4, 8, %l4 4116 1.1 christos 4117 1.1 christos mov 1, %o0 4118 1.1 christos ret 4119 1.1 christos restore 4120 1.1 christos .type bn_mul_mont_t4, #function 4121 1.1 christos .size bn_mul_mont_t4, .-bn_mul_mont_t4 4122 1.1 christos .globl bn_mul_mont_gather5_t4 4123 1.1 christos .align 32 4124 1.1 christos bn_mul_mont_gather5_t4: 4125 1.1 christos add %sp, STACK_BIAS, %g4 ! real top of stack 4126 1.1 christos sll %o5, 3, %o5 ! size in bytes 4127 1.1 christos add %o5, 63, %g1 4128 1.1 christos andn %g1, 63, %g1 ! buffer size rounded up to 64 bytes 4129 1.1 christos sub %g4, %g1, %g1 4130 1.1 christos andn %g1, 63, %g1 ! align at 64 byte 4131 1.1 christos sub %g1, STACK_FRAME, %g1 ! new top of stack 4132 1.1 christos sub %g1, %g4, %g1 4133 1.1 christos LDPTR [%sp+STACK_7thARG], %g4 ! load power, 7th argument 4134 1.1 christos 4135 1.1 christos save %sp, %g1, %sp 4136 1.1 christos srl %g4, 2, %o4 4137 1.1 christos and %g4, 3, %o5 4138 1.1 christos and %o4, 7, %o4 4139 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 4140 1.1 christos add %o5, %i2, %i2 ! of the pwrtbl 4141 1.1 christos or %g0, 1, %o5 4142 1.1 christos sll %o5, %o4, %l7 4143 1.1 christos wr %l7, %g0, %ccr 4144 1.1 christos ldx [%i2+0*32], %g2 4145 1.1 christos ldx [%i2+1*32], %o4 4146 1.1 christos ldx [%i2+2*32], %o5 4147 1.1 christos movvs %icc, %o4, %g2 4148 1.1 christos ldx [%i2+3*32], %o4 4149 1.1 christos move %icc, %o5, %g2 4150 1.1 christos ldx [%i2+4*32], %o5 4151 1.1 christos movneg %icc, %o4, %g2 4152 1.1 christos ldx [%i2+5*32], %o4 4153 1.1 christos movcs %xcc, %o5, %g2 4154 1.1 christos ldx [%i2+6*32], %o5 4155 1.1 christos movvs %xcc, %o4, %g2 4156 1.1 christos ldx [%i2+7*32], %o4 4157 1.1 christos move %xcc, %o5, %g2 4158 1.1 christos add %i2,8*32, %i2 4159 1.1 christos movneg %xcc, %o4, %g2 4160 1.1 christos ld [%i4+0], %l0 ! pull n0[0..1] value 4161 1.1 christos ld [%i4+4], %l1 4162 1.1 christos add %sp, STACK_BIAS+STACK_FRAME, %l5 4163 1.1 christos sllx %l1, 32, %g1 4164 1.1 christos or %l0, %g1, %g1 4165 1.1 christos 4166 1.1 christos ldx [%i1+0], %o2 ! ap[0] 4168 1.1 christos 4169 1.1 christos mulx %o2, %g2, %g4 ! ap[0]*bp[0] 4170 1.1 christos .word 0x8bb282c2 !umulxhi %o2,%g2,%g5 4171 1.1 christos 4172 1.1 christos ldx [%i1+8], %o2 ! ap[1] 4173 1.1 christos add %i1, 16, %i1 4174 1.1 christos ldx [%i3+0], %o4 ! np[0] 4175 1.1 christos 4176 1.1 christos mulx %g4, %g1, %g3 ! "tp[0]"*n0 4177 1.1 christos 4178 1.1 christos mulx %o2, %g2, %o3 ! ap[1]*bp[0] 4179 1.1 christos .word 0x95b282c2 !umulxhi %o2,%g2,%o2 ! ahi=aj 4180 1.1 christos 4181 1.1 christos mulx %o4, %g3, %o0 ! np[0]*m1 4182 1.1 christos .word 0x93b302c3 !umulxhi %o4,%g3,%o1 4183 1.1 christos 4184 1.1 christos ldx [%i3+8], %o4 ! np[1] 4185 1.1 christos 4186 1.1 christos addcc %g4, %o0, %o0 4187 1.1 christos add %i3, 16, %i3 4188 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 4189 1.1 christos 4190 1.1 christos mulx %o4, %g3, %o5 ! np[1]*m1 4191 1.1 christos .word 0x99b302c3 !umulxhi %o4,%g3,%o4 ! nhi=nj 4192 1.1 christos 4193 1.1 christos ba .L1st_g5 4195 1.1 christos sub %i5, 24, %l4 ! cnt=num-3 4196 1.1 christos 4197 1.1 christos .align 16 4198 1.1 christos .L1st_g5: 4199 1.1 christos addcc %o3, %g5, %g4 4200 1.1 christos .word 0x8bb28220 !addxc %o2,%g0,%g5 4201 1.1 christos 4202 1.1 christos ldx [%i1+0], %o2 ! ap[j] 4203 1.1 christos addcc %o5, %o1, %o0 4204 1.1 christos add %i1, 8, %i1 4205 1.1 christos .word 0x93b30220 !addxc %o4,%g0,%o1 ! nhi=nj 4206 1.1 christos 4207 1.1 christos ldx [%i3+0], %o4 ! np[j] 4208 1.1 christos mulx %o2, %g2, %o3 ! ap[j]*bp[0] 4209 1.1 christos add %i3, 8, %i3 4210 1.1 christos .word 0x95b282c2 !umulxhi %o2,%g2,%o2 ! ahi=aj 4211 1.1 christos 4212 1.1 christos mulx %o4, %g3, %o5 ! np[j]*m1 4213 1.1 christos addcc %g4, %o0, %o0 ! np[j]*m1+ap[j]*bp[0] 4214 1.1 christos .word 0x99b302c3 !umulxhi %o4,%g3,%o4 ! nhi=nj 4215 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 4216 1.1 christos stxa %o0, [%l5]0xe2 ! tp[j-1] 4217 1.1 christos add %l5, 8, %l5 ! tp++ 4218 1.1 christos 4219 1.1 christos brnz,pt %l4, .L1st_g5 4220 1.1 christos sub %l4, 8, %l4 ! j-- 4221 1.1 christos !.L1st_g5 4222 1.1 christos addcc %o3, %g5, %g4 4223 1.1 christos .word 0x8bb28220 !addxc %o2,%g0,%g5 ! ahi=aj 4224 1.1 christos 4225 1.1 christos addcc %o5, %o1, %o0 4226 1.1 christos .word 0x93b30220 !addxc %o4,%g0,%o1 4227 1.1 christos addcc %g4, %o0, %o0 ! np[j]*m1+ap[j]*bp[0] 4228 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 4229 1.1 christos stxa %o0, [%l5]0xe2 ! tp[j-1] 4230 1.1 christos add %l5, 8, %l5 4231 1.1 christos 4232 1.1 christos addcc %g5, %o1, %o1 4233 1.1 christos .word 0xa1b00220 !addxc %g0,%g0,%l0 ! upmost overflow bit 4234 1.1 christos stxa %o1, [%l5]0xe2 4235 1.1 christos add %l5, 8, %l5 4236 1.1 christos 4237 1.1 christos ba .Louter_g5 4239 1.1 christos sub %i5, 16, %l1 ! i=num-2 4240 1.1 christos 4241 1.1 christos .align 16 4242 1.1 christos .Louter_g5: 4243 1.1 christos wr %l7, %g0, %ccr 4244 1.1 christos ldx [%i2+0*32], %g2 4245 1.1 christos ldx [%i2+1*32], %o4 4246 1.1 christos ldx [%i2+2*32], %o5 4247 1.1 christos movvs %icc, %o4, %g2 4248 1.1 christos ldx [%i2+3*32], %o4 4249 1.1 christos move %icc, %o5, %g2 4250 1.1 christos ldx [%i2+4*32], %o5 4251 1.1 christos movneg %icc, %o4, %g2 4252 1.1 christos ldx [%i2+5*32], %o4 4253 1.1 christos movcs %xcc, %o5, %g2 4254 1.1 christos ldx [%i2+6*32], %o5 4255 1.1 christos movvs %xcc, %o4, %g2 4256 1.1 christos ldx [%i2+7*32], %o4 4257 1.1 christos move %xcc, %o5, %g2 4258 1.1 christos add %i2,8*32, %i2 4259 1.1 christos movneg %xcc, %o4, %g2 4260 1.1 christos sub %i1, %i5, %i1 ! rewind 4261 1.1 christos sub %i3, %i5, %i3 4262 1.1 christos sub %l5, %i5, %l5 4263 1.1 christos 4264 1.1 christos ldx [%i1+0], %o2 ! ap[0] 4265 1.1 christos ldx [%i3+0], %o4 ! np[0] 4266 1.1 christos 4267 1.1 christos mulx %o2, %g2, %g4 ! ap[0]*bp[i] 4268 1.1 christos ldx [%l5], %o7 ! tp[0] 4269 1.1 christos .word 0x8bb282c2 !umulxhi %o2,%g2,%g5 4270 1.1 christos ldx [%i1+8], %o2 ! ap[1] 4271 1.1 christos addcc %g4, %o7, %g4 ! ap[0]*bp[i]+tp[0] 4272 1.1 christos mulx %o2, %g2, %o3 ! ap[1]*bp[i] 4273 1.1 christos .word 0x8bb00225 !addxc %g0,%g5,%g5 4274 1.1 christos mulx %g4, %g1, %g3 ! tp[0]*n0 4275 1.1 christos .word 0x95b282c2 !umulxhi %o2,%g2,%o2 ! ahi=aj 4276 1.1 christos mulx %o4, %g3, %o0 ! np[0]*m1 4277 1.1 christos add %i1, 16, %i1 4278 1.1 christos .word 0x93b302c3 !umulxhi %o4,%g3,%o1 4279 1.1 christos ldx [%i3+8], %o4 ! np[1] 4280 1.1 christos add %i3, 16, %i3 4281 1.1 christos addcc %o0, %g4, %o0 4282 1.1 christos mulx %o4, %g3, %o5 ! np[1]*m1 4283 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 4284 1.1 christos .word 0x99b302c3 !umulxhi %o4,%g3,%o4 ! nhi=nj 4285 1.1 christos 4286 1.1 christos ba .Linner_g5 4288 1.1 christos sub %i5, 24, %l4 ! cnt=num-3 4289 1.1 christos .align 16 4290 1.1 christos .Linner_g5: 4291 1.1 christos addcc %o3, %g5, %g4 4292 1.1 christos ldx [%l5+8], %o7 ! tp[j] 4293 1.1 christos .word 0x8bb28220 !addxc %o2,%g0,%g5 ! ahi=aj 4294 1.1 christos ldx [%i1+0], %o2 ! ap[j] 4295 1.1 christos add %i1, 8, %i1 4296 1.1 christos addcc %o5, %o1, %o0 4297 1.1 christos mulx %o2, %g2, %o3 ! ap[j]*bp[i] 4298 1.1 christos .word 0x93b30220 !addxc %o4,%g0,%o1 ! nhi=nj 4299 1.1 christos ldx [%i3+0], %o4 ! np[j] 4300 1.1 christos add %i3, 8, %i3 4301 1.1 christos .word 0x95b282c2 !umulxhi %o2,%g2,%o2 ! ahi=aj 4302 1.1 christos addcc %g4, %o7, %g4 ! ap[j]*bp[i]+tp[j] 4303 1.1 christos mulx %o4, %g3, %o5 ! np[j]*m1 4304 1.1 christos .word 0x8bb00225 !addxc %g0,%g5,%g5 4305 1.1 christos .word 0x99b302c3 !umulxhi %o4,%g3,%o4 ! nhi=nj 4306 1.1 christos addcc %o0, %g4, %o0 ! np[j]*m1+ap[j]*bp[i]+tp[j] 4307 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 4308 1.1 christos stx %o0, [%l5] ! tp[j-1] 4309 1.1 christos add %l5, 8, %l5 4310 1.1 christos brnz,pt %l4, .Linner_g5 4311 1.1 christos sub %l4, 8, %l4 4312 1.1 christos !.Linner_g5 4313 1.1 christos ldx [%l5+8], %o7 ! tp[j] 4314 1.1 christos addcc %o3, %g5, %g4 4315 1.1 christos .word 0x8bb28220 !addxc %o2,%g0,%g5 ! ahi=aj 4316 1.1 christos addcc %g4, %o7, %g4 ! ap[j]*bp[i]+tp[j] 4317 1.1 christos .word 0x8bb00225 !addxc %g0,%g5,%g5 4318 1.1 christos 4319 1.1 christos addcc %o5, %o1, %o0 4320 1.1 christos .word 0x93b30220 !addxc %o4,%g0,%o1 ! nhi=nj 4321 1.1 christos addcc %o0, %g4, %o0 ! np[j]*m1+ap[j]*bp[i]+tp[j] 4322 1.1 christos .word 0x93b00229 !addxc %g0,%o1,%o1 4323 1.1 christos stx %o0, [%l5] ! tp[j-1] 4324 1.1 christos 4325 1.1 christos subcc %g0, %l0, %g0 ! move upmost overflow to CCR.xcc 4326 1.1 christos .word 0x93b24265 !addxccc %o1,%g5,%o1 4327 1.1 christos .word 0xa1b00220 !addxc %g0,%g0,%l0 4328 1.1 christos stx %o1, [%l5+8] 4329 1.1 christos add %l5, 16, %l5 4330 1.1 christos 4331 1.1 christos brnz,pt %l1, .Louter_g5 4332 1.1 christos sub %l1, 8, %l1 4333 1.1 christos 4334 1.1 christos sub %i1, %i5, %i1 ! rewind 4336 1.1 christos sub %i3, %i5, %i3 4337 1.1 christos sub %l5, %i5, %l5 4338 1.1 christos ba .Lsub_g5 4339 1.1 christos subcc %i5, 8, %l4 ! cnt=num-1 and clear CCR.xcc 4340 1.1 christos 4341 1.1 christos .align 16 4342 1.1 christos .Lsub_g5: 4343 1.1 christos ldx [%l5], %o7 4344 1.1 christos add %l5, 8, %l5 4345 1.1 christos ldx [%i3+0], %o4 4346 1.1 christos add %i3, 8, %i3 4347 1.1 christos subccc %o7, %o4, %l2 ! tp[j]-np[j] 4348 1.1 christos srlx %o7, 32, %o7 4349 1.1 christos srlx %o4, 32, %o4 4350 1.1 christos subccc %o7, %o4, %l3 4351 1.1.1.2 christos add %i0, 8, %i0 4352 1.1 christos st %l2, [%i0-4] ! reverse order 4353 1.1 christos st %l3, [%i0-8] 4354 1.1 christos brnz,pt %l4, .Lsub_g5 4355 1.1 christos sub %l4, 8, %l4 4356 1.1.1.2 christos 4357 1.1.1.2 christos sub %i3, %i5, %i3 ! rewind 4358 1.1.1.2 christos sub %l5, %i5, %l5 4359 1.1 christos sub %i0, %i5, %i0 4360 1.1 christos 4361 1.1.1.2 christos subccc %l0, %g0, %l0 ! handle upmost overflow bit 4362 1.1 christos ba .Lcopy_g5 4363 1.1 christos sub %i5, 8, %l4 4364 1.1 christos 4365 1.1 christos .align 16 4366 1.1 christos .Lcopy_g5: ! conditional copy 4367 1.1 christos ldx [%l5], %o7 4368 1.1 christos ldx [%i0+0], %l2 4369 1.1 christos stx %g0, [%l5] ! zap 4370 1.1 christos add %l5, 8, %l5 4371 1.1 christos movcs %icc, %o7, %l2 4372 1.1 christos stx %l2, [%i0+0] 4373 1.1 christos add %i0, 8, %i0 4374 1.1 christos brnz %l4, .Lcopy_g5 4375 1.1 christos sub %l4, 8, %l4 4376 1.1 christos 4377 1.1 christos mov 1, %o0 4378 1.1 christos ret 4379 1.1 christos restore 4380 1.1 christos .type bn_mul_mont_gather5_t4, #function 4381 1.1 christos .size bn_mul_mont_gather5_t4, .-bn_mul_mont_gather5_t4 4382 1.1 christos .globl bn_flip_t4 4383 1.1 christos .align 32 4384 1.1 christos bn_flip_t4: 4385 1.1 christos .Loop_flip: 4386 1.1 christos ld [%o1+0], %o4 4387 1.1 christos sub %o2, 1, %o2 4388 1.1 christos ld [%o1+4], %o5 4389 1.1 christos add %o1, 8, %o1 4390 1.1 christos st %o5, [%o0+0] 4391 1.1 christos st %o4, [%o0+4] 4392 1.1 christos brnz %o2, .Loop_flip 4393 1.1 christos add %o0, 8, %o0 4394 1.1 christos retl 4395 1.1 christos nop 4396 1.1 christos .type bn_flip_t4, #function 4397 1.1 christos .size bn_flip_t4, .-bn_flip_t4 4398 1.1 christos 4399 1.1 christos .globl bn_flip_n_scatter5_t4 4400 1.1 christos .align 32 4401 1.1 christos bn_flip_n_scatter5_t4: 4402 1.1 christos sll %o3, 3, %o3 4403 1.1 christos srl %o1, 1, %o1 4404 1.1 christos add %o3, %o2, %o2 ! &pwrtbl[pwr] 4405 1.1 christos sub %o1, 1, %o1 4406 1.1 christos .Loop_flip_n_scatter5: 4407 1.1 christos ld [%o0+0], %o4 ! inp[i] 4408 1.1 christos ld [%o0+4], %o5 4409 1.1 christos add %o0, 8, %o0 4410 1.1 christos sllx %o5, 32, %o5 4411 1.1 christos or %o4, %o5, %o5 4412 1.1 christos stx %o5, [%o2] 4413 1.1 christos add %o2, 32*8, %o2 4414 1.1 christos brnz %o1, .Loop_flip_n_scatter5 4415 1.1 christos sub %o1, 1, %o1 4416 1.1 christos retl 4417 1.1 christos nop 4418 1.1 christos .type bn_flip_n_scatter5_t4, #function 4419 1.1 christos .size bn_flip_n_scatter5_t4, .-bn_flip_n_scatter5_t4 4420 1.1 christos 4421 1.1 christos .globl bn_gather5_t4 4422 1.1 christos .align 32 4423 1.1 christos bn_gather5_t4: 4424 1.1 christos srl %o3, 2, %o4 4425 1.1 christos and %o3, 3, %o5 4426 1.1 christos and %o4, 7, %o4 4427 1.1 christos sll %o5, 3, %o5 ! offset within first cache line 4428 1.1 christos add %o5, %o2, %o2 ! of the pwrtbl 4429 1.1 christos or %g0, 1, %o5 4430 1.1 christos sll %o5, %o4, %g1 4431 1.1 christos wr %g1, %g0, %ccr 4432 1.1 christos sub %o1, 1, %o1 4433 1.1 christos .Loop_gather5: 4434 1.1 christos ldx [%o2+0*32], %g1 4435 1.1 christos ldx [%o2+1*32], %o4 4436 1.1 christos ldx [%o2+2*32], %o5 4437 1.1 christos movvs %icc, %o4, %g1 4438 1.1 christos ldx [%o2+3*32], %o4 4439 1.1 christos move %icc, %o5, %g1 4440 1.1 christos ldx [%o2+4*32], %o5 4441 1.1 christos movneg %icc, %o4, %g1 4442 1.1 christos ldx [%o2+5*32], %o4 4443 1.1 christos movcs %xcc, %o5, %g1 4444 1.1 christos ldx [%o2+6*32], %o5 4445 1.1 christos movvs %xcc, %o4, %g1 4446 1.1 christos ldx [%o2+7*32], %o4 4447 1.1 christos move %xcc, %o5, %g1 4448 1.1 christos add %o2,8*32, %o2 4449 1.1 christos movneg %xcc, %o4, %g1 4450 1.1 christos stx %g1, [%o0] 4451 1.1 christos add %o0, 8, %o0 4452 brnz %o1, .Loop_gather5 4453 sub %o1, 1, %o1 4454 4455 retl 4456 nop 4457 .type bn_gather5_t4, #function 4458 .size bn_gather5_t4, .-bn_gather5_t4 4459 4460 .asciz "Montgomery Multiplication for SPARC T4, David S. Miller, Andy Polyakov" 4461 .align 4 4462