| ... | ... | @@ -0,0 +1,1788 @@ |
| 1 | const builtin = @import("builtin"); |
| 2 | const common = @import("./common.zig"); |
| 3 | |
| 4 | fn __hexagon_divsi3() callconv(.naked) noreturn { |
| 5 | asm volatile ( |
| 6 | \\ { |
| 7 | \\ p0 = cmp.ge(r0,#0) |
| 8 | \\ p1 = cmp.ge(r1,#0) |
| 9 | \\ r1 = abs(r0) |
| 10 | \\ r2 = abs(r1) |
| 11 | \\ } |
| 12 | \\ { |
| 13 | \\ r3 = cl0(r1) |
| 14 | \\ r4 = cl0(r2) |
| 15 | \\ r5 = sub(r1,r2) |
| 16 | \\ p2 = cmp.gtu(r2,r1) |
| 17 | \\ } |
| 18 | \\ { |
| 19 | \\ r0 = #0 |
| 20 | \\ p1 = xor(p0,p1) |
| 21 | \\ p0 = cmp.gtu(r2,r5) |
| 22 | \\ if (p2) jumpr r31 |
| 23 | \\ } |
| 24 | \\ |
| 25 | \\ { |
| 26 | \\ r0 = mux(p1,#-1,#1) |
| 27 | \\ if (p0) jumpr r31 |
| 28 | \\ r4 = sub(r4,r3) |
| 29 | \\ r3 = #1 |
| 30 | \\ } |
| 31 | \\ { |
| 32 | \\ r0 = #0 |
| 33 | \\ r3:2 = vlslw(r3:2,r4) |
| 34 | \\ loop0(1f,r4) |
| 35 | \\ } |
| 36 | \\ .falign |
| 37 | \\ 1: |
| 38 | \\ { |
| 39 | \\ p0 = cmp.gtu(r2,r1) |
| 40 | \\ if (!p0.new) r1 = sub(r1,r2) |
| 41 | \\ if (!p0.new) r0 = add(r0,r3) |
| 42 | \\ r3:2 = vlsrw(r3:2,#1) |
| 43 | \\ }:endloop0 |
| 44 | \\ { |
| 45 | \\ p0 = cmp.gtu(r2,r1) |
| 46 | \\ if (!p0.new) r0 = add(r0,r3) |
| 47 | \\ if (!p1) jumpr r31 |
| 48 | \\ } |
| 49 | \\ { |
| 50 | \\ r0 = neg(r0) |
| 51 | \\ jumpr r31 |
| 52 | \\ } |
| 53 | ); |
| 54 | } |
| 55 | |
| 56 | fn __hexagon_umodsi3() callconv(.naked) noreturn { |
| 57 | asm volatile ( |
| 58 | \\ { |
| 59 | \\ r2 = cl0(r0) |
| 60 | \\ r3 = cl0(r1) |
| 61 | \\ p0 = cmp.gtu(r1,r0) |
| 62 | \\ } |
| 63 | \\ { |
| 64 | \\ r2 = sub(r3,r2) |
| 65 | \\ if (p0) jumpr r31 |
| 66 | \\ } |
| 67 | \\ { |
| 68 | \\ loop0(1f,r2) |
| 69 | \\ p1 = cmp.eq(r2,#0) |
| 70 | \\ r2 = lsl(r1,r2) |
| 71 | \\ } |
| 72 | \\ .falign |
| 73 | \\ 1: |
| 74 | \\ { |
| 75 | \\ p0 = cmp.gtu(r2,r0) |
| 76 | \\ if (!p0.new) r0 = sub(r0,r2) |
| 77 | \\ r2 = lsr(r2,#1) |
| 78 | \\ if (p1) r1 = #0 |
| 79 | \\ }:endloop0 |
| 80 | \\ { |
| 81 | \\ p0 = cmp.gtu(r2,r0) |
| 82 | \\ if (!p0.new) r0 = sub(r0,r1) |
| 83 | \\ jumpr r31 |
| 84 | \\ } |
| 85 | ); |
| 86 | } |
| 87 | |
| 88 | fn __hexagon_sqrtf() callconv(.naked) noreturn { |
| 89 | asm volatile ( |
| 90 | \\ { |
| 91 | \\ r3,p0 = sfinvsqrta(r0) |
| 92 | \\ r5 = sffixupr(r0) |
| 93 | \\ r4 = ##0x3f000000 |
| 94 | \\ r1:0 = combine(#0,#0) |
| 95 | \\ } |
| 96 | \\ { |
| 97 | \\ r0 += sfmpy(r3,r5):lib |
| 98 | \\ r1 += sfmpy(r3,r4):lib |
| 99 | \\ r2 = r4 |
| 100 | \\ r3 = r5 |
| 101 | \\ } |
| 102 | \\ { |
| 103 | \\ r2 -= sfmpy(r0,r1):lib |
| 104 | \\ p1 = sfclass(r5,#1) |
| 105 | \\ |
| 106 | \\ } |
| 107 | \\ { |
| 108 | \\ r0 += sfmpy(r0,r2):lib |
| 109 | \\ r1 += sfmpy(r1,r2):lib |
| 110 | \\ r2 = r4 |
| 111 | \\ r3 = r5 |
| 112 | \\ } |
| 113 | \\ { |
| 114 | \\ r2 -= sfmpy(r0,r1):lib |
| 115 | \\ r3 -= sfmpy(r0,r0):lib |
| 116 | \\ } |
| 117 | \\ { |
| 118 | \\ r0 += sfmpy(r1,r3):lib |
| 119 | \\ r1 += sfmpy(r1,r2):lib |
| 120 | \\ r2 = r4 |
| 121 | \\ r3 = r5 |
| 122 | \\ } |
| 123 | \\ { |
| 124 | \\ |
| 125 | \\ r3 -= sfmpy(r0,r0):lib |
| 126 | \\ if (p1) r0 = or(r0,r5) |
| 127 | \\ } |
| 128 | \\ { |
| 129 | \\ r0 += sfmpy(r1,r3,p0):scale |
| 130 | \\ jumpr r31 |
| 131 | \\ } |
| 132 | ); |
| 133 | } |
| 134 | |
| 135 | fn __hexagon_moddi3() callconv(.naked) noreturn { |
| 136 | asm volatile ( |
| 137 | \\ { |
| 138 | \\ p3 = tstbit(r1,#31) |
| 139 | \\ } |
| 140 | \\ { |
| 141 | \\ r1:0 = abs(r1:0) |
| 142 | \\ r3:2 = abs(r3:2) |
| 143 | \\ } |
| 144 | \\ { |
| 145 | \\ r6 = cl0(r1:0) |
| 146 | \\ r7 = cl0(r3:2) |
| 147 | \\ r5:4 = r3:2 |
| 148 | \\ r3:2 = r1:0 |
| 149 | \\ } |
| 150 | \\ { |
| 151 | \\ r10 = sub(r7,r6) |
| 152 | \\ r1:0 = #0 |
| 153 | \\ r15:14 = #1 |
| 154 | \\ } |
| 155 | \\ { |
| 156 | \\ r11 = add(r10,#1) |
| 157 | \\ r13:12 = lsl(r5:4,r10) |
| 158 | \\ r15:14 = lsl(r15:14,r10) |
| 159 | \\ } |
| 160 | \\ { |
| 161 | \\ p0 = cmp.gtu(r5:4,r3:2) |
| 162 | \\ loop0(1f,r11) |
| 163 | \\ } |
| 164 | \\ { |
| 165 | \\ if (p0) jump .hexagon_moddi3_return |
| 166 | \\ } |
| 167 | \\ .falign |
| 168 | \\ 1: |
| 169 | \\ { |
| 170 | \\ p0 = cmp.gtu(r13:12,r3:2) |
| 171 | \\ } |
| 172 | \\ { |
| 173 | \\ r7:6 = sub(r3:2, r13:12) |
| 174 | \\ r9:8 = add(r1:0, r15:14) |
| 175 | \\ } |
| 176 | \\ { |
| 177 | \\ r1:0 = vmux(p0, r1:0, r9:8) |
| 178 | \\ r3:2 = vmux(p0, r3:2, r7:6) |
| 179 | \\ } |
| 180 | \\ { |
| 181 | \\ r15:14 = lsr(r15:14, #1) |
| 182 | \\ r13:12 = lsr(r13:12, #1) |
| 183 | \\ }:endloop0 |
| 184 | \\ |
| 185 | \\ .hexagon_moddi3_return: |
| 186 | \\ { |
| 187 | \\ r1:0 = neg(r3:2) |
| 188 | \\ } |
| 189 | \\ { |
| 190 | \\ r1:0 = vmux(p3,r1:0,r3:2) |
| 191 | \\ jumpr r31 |
| 192 | \\ } |
| 193 | ); |
| 194 | } |
| 195 | |
| 196 | fn __hexagon_divdi3() callconv(.naked) noreturn { |
| 197 | asm volatile ( |
| 198 | \\ { |
| 199 | \\ p2 = tstbit(r1,#31) |
| 200 | \\ p3 = tstbit(r3,#31) |
| 201 | \\ } |
| 202 | \\ { |
| 203 | \\ r1:0 = abs(r1:0) |
| 204 | \\ r3:2 = abs(r3:2) |
| 205 | \\ } |
| 206 | \\ { |
| 207 | \\ r6 = cl0(r1:0) |
| 208 | \\ r7 = cl0(r3:2) |
| 209 | \\ r5:4 = r3:2 |
| 210 | \\ r3:2 = r1:0 |
| 211 | \\ } |
| 212 | \\ { |
| 213 | \\ p3 = xor(p2,p3) |
| 214 | \\ r10 = sub(r7,r6) |
| 215 | \\ r1:0 = #0 |
| 216 | \\ r15:14 = #1 |
| 217 | \\ } |
| 218 | \\ { |
| 219 | \\ r11 = add(r10,#1) |
| 220 | \\ r13:12 = lsl(r5:4,r10) |
| 221 | \\ r15:14 = lsl(r15:14,r10) |
| 222 | \\ } |
| 223 | \\ { |
| 224 | \\ p0 = cmp.gtu(r5:4,r3:2) |
| 225 | \\ loop0(1f,r11) |
| 226 | \\ } |
| 227 | \\ { |
| 228 | \\ if (p0) jump .hexagon_divdi3_return |
| 229 | \\ } |
| 230 | \\ .falign |
| 231 | \\ 1: |
| 232 | \\ { |
| 233 | \\ p0 = cmp.gtu(r13:12,r3:2) |
| 234 | \\ } |
| 235 | \\ { |
| 236 | \\ r7:6 = sub(r3:2, r13:12) |
| 237 | \\ r9:8 = add(r1:0, r15:14) |
| 238 | \\ } |
| 239 | \\ { |
| 240 | \\ r1:0 = vmux(p0, r1:0, r9:8) |
| 241 | \\ r3:2 = vmux(p0, r3:2, r7:6) |
| 242 | \\ } |
| 243 | \\ { |
| 244 | \\ r15:14 = lsr(r15:14, #1) |
| 245 | \\ r13:12 = lsr(r13:12, #1) |
| 246 | \\ }:endloop0 |
| 247 | \\ |
| 248 | \\ .hexagon_divdi3_return: |
| 249 | \\ { |
| 250 | \\ r3:2 = neg(r1:0) |
| 251 | \\ } |
| 252 | \\ { |
| 253 | \\ r1:0 = vmux(p3,r3:2,r1:0) |
| 254 | \\ jumpr r31 |
| 255 | \\ } |
| 256 | ); |
| 257 | } |
| 258 | |
| 259 | fn __hexagon_divsf3() callconv(.naked) noreturn { |
| 260 | asm volatile ( |
| 261 | \\ { |
| 262 | \\ r2,p0 = sfrecipa(r0,r1) |
| 263 | \\ r4 = sffixupd(r0,r1) |
| 264 | \\ r3 = ##0x3f800000 |
| 265 | \\ } |
| 266 | \\ { |
| 267 | \\ r5 = sffixupn(r0,r1) |
| 268 | \\ r3 -= sfmpy(r4,r2):lib |
| 269 | \\ r6 = ##0x80000000 |
| 270 | \\ r7 = r3 |
| 271 | \\ } |
| 272 | \\ { |
| 273 | \\ r2 += sfmpy(r3,r2):lib |
| 274 | \\ r3 = r7 |
| 275 | \\ r6 = r5 |
| 276 | \\ r0 = and(r6,r5) |
| 277 | \\ } |
| 278 | \\ { |
| 279 | \\ r3 -= sfmpy(r4,r2):lib |
| 280 | \\ r0 += sfmpy(r5,r2):lib |
| 281 | \\ } |
| 282 | \\ { |
| 283 | \\ r2 += sfmpy(r3,r2):lib |
| 284 | \\ r6 -= sfmpy(r0,r4):lib |
| 285 | \\ } |
| 286 | \\ { |
| 287 | \\ r0 += sfmpy(r6,r2):lib |
| 288 | \\ } |
| 289 | \\ { |
| 290 | \\ r5 -= sfmpy(r0,r4):lib |
| 291 | \\ } |
| 292 | \\ { |
| 293 | \\ r0 += sfmpy(r5,r2,p0):scale |
| 294 | \\ jumpr r31 |
| 295 | \\ } |
| 296 | ); |
| 297 | } |
| 298 | |
| 299 | fn __hexagon_udivdi3() callconv(.naked) noreturn { |
| 300 | asm volatile ( |
| 301 | \\ { |
| 302 | \\ r6 = cl0(r1:0) |
| 303 | \\ r7 = cl0(r3:2) |
| 304 | \\ r5:4 = r3:2 |
| 305 | \\ r3:2 = r1:0 |
| 306 | \\ } |
| 307 | \\ { |
| 308 | \\ r10 = sub(r7,r6) |
| 309 | \\ r1:0 = #0 |
| 310 | \\ r15:14 = #1 |
| 311 | \\ } |
| 312 | \\ { |
| 313 | \\ r11 = add(r10,#1) |
| 314 | \\ r13:12 = lsl(r5:4,r10) |
| 315 | \\ r15:14 = lsl(r15:14,r10) |
| 316 | \\ } |
| 317 | \\ { |
| 318 | \\ p0 = cmp.gtu(r5:4,r3:2) |
| 319 | \\ loop0(1f,r11) |
| 320 | \\ } |
| 321 | \\ { |
| 322 | \\ if (p0) jumpr r31 |
| 323 | \\ } |
| 324 | \\ .falign |
| 325 | \\ 1: |
| 326 | \\ { |
| 327 | \\ p0 = cmp.gtu(r13:12,r3:2) |
| 328 | \\ } |
| 329 | \\ { |
| 330 | \\ r7:6 = sub(r3:2, r13:12) |
| 331 | \\ r9:8 = add(r1:0, r15:14) |
| 332 | \\ } |
| 333 | \\ { |
| 334 | \\ r1:0 = vmux(p0, r1:0, r9:8) |
| 335 | \\ r3:2 = vmux(p0, r3:2, r7:6) |
| 336 | \\ } |
| 337 | \\ { |
| 338 | \\ r15:14 = lsr(r15:14, #1) |
| 339 | \\ r13:12 = lsr(r13:12, #1) |
| 340 | \\ }:endloop0 |
| 341 | \\ { |
| 342 | \\ jumpr r31 |
| 343 | \\ } |
| 344 | ); |
| 345 | } |
| 346 | |
| 347 | fn __hexagon_umoddi3() callconv(.naked) noreturn { |
| 348 | asm volatile ( |
| 349 | \\ { |
| 350 | \\ r6 = cl0(r1:0) |
| 351 | \\ r7 = cl0(r3:2) |
| 352 | \\ r5:4 = r3:2 |
| 353 | \\ r3:2 = r1:0 |
| 354 | \\ } |
| 355 | \\ { |
| 356 | \\ r10 = sub(r7,r6) |
| 357 | \\ r1:0 = #0 |
| 358 | \\ r15:14 = #1 |
| 359 | \\ } |
| 360 | \\ { |
| 361 | \\ r11 = add(r10,#1) |
| 362 | \\ r13:12 = lsl(r5:4,r10) |
| 363 | \\ r15:14 = lsl(r15:14,r10) |
| 364 | \\ } |
| 365 | \\ { |
| 366 | \\ p0 = cmp.gtu(r5:4,r3:2) |
| 367 | \\ loop0(1f,r11) |
| 368 | \\ } |
| 369 | \\ { |
| 370 | \\ if (p0) jump .hexagon_umoddi3_return |
| 371 | \\ } |
| 372 | \\ .falign |
| 373 | \\ 1: |
| 374 | \\ { |
| 375 | \\ p0 = cmp.gtu(r13:12,r3:2) |
| 376 | \\ } |
| 377 | \\ { |
| 378 | \\ r7:6 = sub(r3:2, r13:12) |
| 379 | \\ r9:8 = add(r1:0, r15:14) |
| 380 | \\ } |
| 381 | \\ { |
| 382 | \\ r1:0 = vmux(p0, r1:0, r9:8) |
| 383 | \\ r3:2 = vmux(p0, r3:2, r7:6) |
| 384 | \\ } |
| 385 | \\ { |
| 386 | \\ r15:14 = lsr(r15:14, #1) |
| 387 | \\ r13:12 = lsr(r13:12, #1) |
| 388 | \\ }:endloop0 |
| 389 | \\ |
| 390 | \\ .hexagon_umoddi3_return: |
| 391 | \\ { |
| 392 | \\ r1:0 = r3:2 |
| 393 | \\ jumpr r31 |
| 394 | \\ } |
| 395 | ); |
| 396 | } |
| 397 | |
| 398 | fn __hexagon_modsi3() callconv(.naked) noreturn { |
| 399 | asm volatile ( |
| 400 | \\ { |
| 401 | \\ p2 = cmp.ge(r0,#0) |
| 402 | \\ r2 = abs(r0) |
| 403 | \\ r1 = abs(r1) |
| 404 | \\ } |
| 405 | \\ { |
| 406 | \\ r3 = cl0(r2) |
| 407 | \\ r4 = cl0(r1) |
| 408 | \\ p0 = cmp.gtu(r1,r2) |
| 409 | \\ } |
| 410 | \\ { |
| 411 | \\ r3 = sub(r4,r3) |
| 412 | \\ if (p0) jumpr r31 |
| 413 | \\ } |
| 414 | \\ { |
| 415 | \\ p1 = cmp.eq(r3,#0) |
| 416 | \\ loop0(1f,r3) |
| 417 | \\ r0 = r2 |
| 418 | \\ r2 = lsl(r1,r3) |
| 419 | \\ } |
| 420 | \\ .falign |
| 421 | \\ 1: |
| 422 | \\ { |
| 423 | \\ p0 = cmp.gtu(r2,r0) |
| 424 | \\ if (!p0.new) r0 = sub(r0,r2) |
| 425 | \\ r2 = lsr(r2,#1) |
| 426 | \\ if (p1) r1 = #0 |
| 427 | \\ }:endloop0 |
| 428 | \\ { |
| 429 | \\ p0 = cmp.gtu(r2,r0) |
| 430 | \\ if (!p0.new) r0 = sub(r0,r1) |
| 431 | \\ if (p2) jumpr r31 |
| 432 | \\ } |
| 433 | \\ { |
| 434 | \\ r0 = neg(r0) |
| 435 | \\ jumpr r31 |
| 436 | \\ } |
| 437 | ); |
| 438 | } |
| 439 | |
| 440 | fn __hexagon_memcpy_likely_aligned_min32bytes_mult8bytes() callconv(.naked) noreturn { |
| 441 | asm volatile ( |
| 442 | \\ { |
| 443 | \\ p0 = bitsclr(r1,#7) |
| 444 | \\ p0 = bitsclr(r0,#7) |
| 445 | \\ if (p0.new) r5:4 = memd(r1) |
| 446 | \\ r3 = #-3 |
| 447 | \\ } |
| 448 | \\ { |
| 449 | \\ if (!p0) jump .Lmemcpy_call |
| 450 | \\ if (p0) memd(r0++#8) = r5:4 |
| 451 | \\ if (p0) r5:4 = memd(r1+#8) |
| 452 | \\ r3 += lsr(r2,#3) |
| 453 | \\ } |
| 454 | \\ { |
| 455 | \\ memd(r0++#8) = r5:4 |
| 456 | \\ r5:4 = memd(r1+#16) |
| 457 | \\ r1 = add(r1,#24) |
| 458 | \\ loop0(1f,r3) |
| 459 | \\ } |
| 460 | \\ .falign |
| 461 | \\ 1: |
| 462 | \\ { |
| 463 | \\ memd(r0++#8) = r5:4 |
| 464 | \\ r5:4 = memd(r1++#8) |
| 465 | \\ }:endloop0 |
| 466 | \\ { |
| 467 | \\ memd(r0) = r5:4 |
| 468 | \\ r0 -= add(r2,#-8) |
| 469 | \\ jumpr r31 |
| 470 | \\ } |
| 471 | \\ .Lmemcpy_call: |
| 472 | \\ jump memcpy@PLT |
| 473 | ); |
| 474 | } |
| 475 | |
| 476 | fn __hexagon_udivsi3() callconv(.naked) noreturn { |
| 477 | asm volatile ( |
| 478 | \\ { |
| 479 | \\ r2 = cl0(r0) |
| 480 | \\ r3 = cl0(r1) |
| 481 | \\ r5:4 = combine(#1,#0) |
| 482 | \\ p0 = cmp.gtu(r1,r0) |
| 483 | \\ } |
| 484 | \\ { |
| 485 | \\ r6 = sub(r3,r2) |
| 486 | \\ r4 = r1 |
| 487 | \\ r1:0 = combine(r0,r4) |
| 488 | \\ if (p0) jumpr r31 |
| 489 | \\ } |
| 490 | \\ { |
| 491 | \\ r3:2 = vlslw(r5:4,r6) |
| 492 | \\ loop0(1f,r6) |
| 493 | \\ } |
| 494 | \\ .falign |
| 495 | \\ 1: |
| 496 | \\ { |
| 497 | \\ p0 = cmp.gtu(r2,r1) |
| 498 | \\ if (!p0.new) r1 = sub(r1,r2) |
| 499 | \\ if (!p0.new) r0 = add(r0,r3) |
| 500 | \\ r3:2 = vlsrw(r3:2,#1) |
| 501 | \\ }:endloop0 |
| 502 | \\ { |
| 503 | \\ p0 = cmp.gtu(r2,r1) |
| 504 | \\ if (!p0.new) r0 = add(r0,r3) |
| 505 | \\ jumpr r31 |
| 506 | \\ } |
| 507 | ); |
| 508 | } |
| 509 | |
| 510 | fn __hexagon_adddf3() align(32) callconv(.naked) noreturn { |
| 511 | asm volatile ( |
| 512 | \\ { |
| 513 | \\ r4 = extractu(r1,#11,#20) |
| 514 | \\ r5 = extractu(r3,#11,#20) |
| 515 | \\ r13:12 = combine(##0x20000000,#0) |
| 516 | \\ } |
| 517 | \\ { |
| 518 | \\ p3 = dfclass(r1:0,#2) |
| 519 | \\ p3 = dfclass(r3:2,#2) |
| 520 | \\ r9:8 = r13:12 |
| 521 | \\ p2 = cmp.gtu(r5,r4) |
| 522 | \\ } |
| 523 | \\ { |
| 524 | \\ if (!p3) jump .Ladd_abnormal |
| 525 | \\ if (p2) r1:0 = r3:2 |
| 526 | \\ if (p2) r3:2 = r1:0 |
| 527 | \\ if (p2) r5:4 = combine(r4,r5) |
| 528 | \\ } |
| 529 | \\ { |
| 530 | \\ r13:12 = insert(r1:0,#52,#11 -2) |
| 531 | \\ r9:8 = insert(r3:2,#52,#11 -2) |
| 532 | \\ r15 = sub(r4,r5) |
| 533 | \\ r7:6 = combine(#62,#1) |
| 534 | \\ } |
| 535 | \\ |
| 536 | \\ |
| 537 | \\ |
| 538 | \\ |
| 539 | \\ |
| 540 | \\ .Ladd_continue: |
| 541 | \\ { |
| 542 | \\ r15 = min(r15,r7) |
| 543 | \\ |
| 544 | \\ r11:10 = neg(r13:12) |
| 545 | \\ p2 = cmp.gt(r1,#-1) |
| 546 | \\ r14 = #0 |
| 547 | \\ } |
| 548 | \\ { |
| 549 | \\ if (!p2) r13:12 = r11:10 |
| 550 | \\ r11:10 = extractu(r9:8,r15:14) |
| 551 | \\ r9:8 = ASR(r9:8,r15) |
| 552 | \\ |
| 553 | \\ |
| 554 | \\ |
| 555 | \\ |
| 556 | \\ r15:14 = #0 |
| 557 | \\ } |
| 558 | \\ { |
| 559 | \\ p1 = cmp.eq(r11:10,r15:14) |
| 560 | \\ if (!p1.new) r8 = or(r8,r6) |
| 561 | \\ r5 = add(r4,#-1024 -60) |
| 562 | \\ p3 = cmp.gt(r3,#-1) |
| 563 | \\ } |
| 564 | \\ { |
| 565 | \\ r13:12 = add(r13:12,r9:8) |
| 566 | \\ r11:10 = sub(r13:12,r9:8) |
| 567 | \\ r7:6 = combine(#54,##2045) |
| 568 | \\ } |
| 569 | \\ { |
| 570 | \\ p0 = cmp.gtu(r4,r7) |
| 571 | \\ p0 = !cmp.gtu(r4,r6) |
| 572 | \\ if (!p0.new) jump:nt .Ladd_ovf_unf |
| 573 | \\ if (!p3) r13:12 = r11:10 |
| 574 | \\ } |
| 575 | \\ { |
| 576 | \\ r1:0 = convert_d2df(r13:12) |
| 577 | \\ p0 = cmp.eq(r13,#0) |
| 578 | \\ p0 = cmp.eq(r12,#0) |
| 579 | \\ if (p0.new) jump:nt .Ladd_zero |
| 580 | \\ } |
| 581 | \\ { |
| 582 | \\ r1 += asl(r5,#20) |
| 583 | \\ jumpr r31 |
| 584 | \\ } |
| 585 | \\ |
| 586 | \\ .falign |
| 587 | \\ .Ladd_zero: |
| 588 | \\ |
| 589 | \\ |
| 590 | \\ { |
| 591 | \\ r28 = USR |
| 592 | \\ r1:0 = #0 |
| 593 | \\ r3 = #1 |
| 594 | \\ } |
| 595 | \\ { |
| 596 | \\ r28 = extractu(r28,#2,#22) |
| 597 | \\ r3 = asl(r3,#31) |
| 598 | \\ } |
| 599 | \\ { |
| 600 | \\ p0 = cmp.eq(r28,#2) |
| 601 | \\ if (p0.new) r1 = xor(r1,r3) |
| 602 | \\ jumpr r31 |
| 603 | \\ } |
| 604 | \\ .falign |
| 605 | \\ .Ladd_ovf_unf: |
| 606 | \\ { |
| 607 | \\ r1:0 = convert_d2df(r13:12) |
| 608 | \\ p0 = cmp.eq(r13,#0) |
| 609 | \\ p0 = cmp.eq(r12,#0) |
| 610 | \\ if (p0.new) jump:nt .Ladd_zero |
| 611 | \\ } |
| 612 | \\ { |
| 613 | \\ r28 = extractu(r1,#11,#20) |
| 614 | \\ r1 += asl(r5,#20) |
| 615 | \\ } |
| 616 | \\ { |
| 617 | \\ r5 = add(r5,r28) |
| 618 | \\ r3:2 = combine(##0x00100000,#0) |
| 619 | \\ } |
| 620 | \\ { |
| 621 | \\ p0 = cmp.gt(r5,##1024 +1024 -2) |
| 622 | \\ if (p0.new) jump:nt .Ladd_ovf |
| 623 | \\ } |
| 624 | \\ { |
| 625 | \\ p0 = cmp.gt(r5,#0) |
| 626 | \\ if (p0.new) jumpr:t r31 |
| 627 | \\ r28 = sub(#1,r5) |
| 628 | \\ } |
| 629 | \\ { |
| 630 | \\ r3:2 = insert(r1:0,#52,#0) |
| 631 | \\ r1:0 = r13:12 |
| 632 | \\ } |
| 633 | \\ { |
| 634 | \\ r3:2 = lsr(r3:2,r28) |
| 635 | \\ } |
| 636 | \\ { |
| 637 | \\ r1:0 = insert(r3:2,#63,#0) |
| 638 | \\ jumpr r31 |
| 639 | \\ } |
| 640 | \\ .falign |
| 641 | \\ .Ladd_ovf: |
| 642 | \\ |
| 643 | \\ { |
| 644 | \\ r1:0 = r13:12 |
| 645 | \\ r28 = USR |
| 646 | \\ r13:12 = combine(##0x7fefffff,#-1) |
| 647 | \\ } |
| 648 | \\ { |
| 649 | \\ r5 = extractu(r28,#2,#22) |
| 650 | \\ r28 = or(r28,#0x28) |
| 651 | \\ r9:8 = combine(##0x7ff00000,#0) |
| 652 | \\ } |
| 653 | \\ { |
| 654 | \\ USR = r28 |
| 655 | \\ r5 ^= lsr(r1,#31) |
| 656 | \\ r28 = r5 |
| 657 | \\ } |
| 658 | \\ { |
| 659 | \\ p0 = !cmp.eq(r28,#1) |
| 660 | \\ p0 = !cmp.eq(r5,#2) |
| 661 | \\ if (p0.new) r13:12 = r9:8 |
| 662 | \\ } |
| 663 | \\ { |
| 664 | \\ r1:0 = insert(r13:12,#63,#0) |
| 665 | \\ } |
| 666 | \\ { |
| 667 | \\ p0 = dfcmp.eq(r1:0,r1:0) |
| 668 | \\ jumpr r31 |
| 669 | \\ } |
| 670 | \\ |
| 671 | \\ .Ladd_abnormal: |
| 672 | \\ { |
| 673 | \\ r13:12 = extractu(r1:0,#63,#0) |
| 674 | \\ r9:8 = extractu(r3:2,#63,#0) |
| 675 | \\ } |
| 676 | \\ { |
| 677 | \\ p3 = cmp.gtu(r13:12,r9:8) |
| 678 | \\ if (!p3.new) r1:0 = r3:2 |
| 679 | \\ if (!p3.new) r3:2 = r1:0 |
| 680 | \\ } |
| 681 | \\ { |
| 682 | \\ |
| 683 | \\ p0 = dfclass(r1:0,#0x0f) |
| 684 | \\ if (!p0.new) jump:nt .Linvalid_nan_add |
| 685 | \\ if (!p3) r13:12 = r9:8 |
| 686 | \\ if (!p3) r9:8 = r13:12 |
| 687 | \\ } |
| 688 | \\ { |
| 689 | \\ |
| 690 | \\ |
| 691 | \\ p1 = dfclass(r1:0,#0x08) |
| 692 | \\ if (p1.new) jump:nt .Linf_add |
| 693 | \\ } |
| 694 | \\ { |
| 695 | \\ p2 = dfclass(r3:2,#0x01) |
| 696 | \\ if (p2.new) jump:nt .LB_zero |
| 697 | \\ r13:12 = #0 |
| 698 | \\ } |
| 699 | \\ |
| 700 | \\ { |
| 701 | \\ p0 = dfclass(r1:0,#4) |
| 702 | \\ if (p0.new) jump:nt .Ladd_two_subnormal |
| 703 | \\ r13:12 = combine(##0x20000000,#0) |
| 704 | \\ } |
| 705 | \\ { |
| 706 | \\ r4 = extractu(r1,#11,#20) |
| 707 | \\ r5 = #1 |
| 708 | \\ |
| 709 | \\ r9:8 = asl(r9:8,#11 -2) |
| 710 | \\ } |
| 711 | \\ |
| 712 | \\ |
| 713 | \\ |
| 714 | \\ { |
| 715 | \\ r13:12 = insert(r1:0,#52,#11 -2) |
| 716 | \\ r15 = sub(r4,r5) |
| 717 | \\ r7:6 = combine(#62,#1) |
| 718 | \\ jump .Ladd_continue |
| 719 | \\ } |
| 720 | \\ |
| 721 | \\ .Ladd_two_subnormal: |
| 722 | \\ { |
| 723 | \\ r13:12 = extractu(r1:0,#63,#0) |
| 724 | \\ r9:8 = extractu(r3:2,#63,#0) |
| 725 | \\ } |
| 726 | \\ { |
| 727 | \\ r13:12 = neg(r13:12) |
| 728 | \\ r9:8 = neg(r9:8) |
| 729 | \\ p0 = cmp.gt(r1,#-1) |
| 730 | \\ p1 = cmp.gt(r3,#-1) |
| 731 | \\ } |
| 732 | \\ { |
| 733 | \\ if (p0) r13:12 = r1:0 |
| 734 | \\ if (p1) r9:8 = r3:2 |
| 735 | \\ } |
| 736 | \\ { |
| 737 | \\ r13:12 = add(r13:12,r9:8) |
| 738 | \\ } |
| 739 | \\ { |
| 740 | \\ r9:8 = neg(r13:12) |
| 741 | \\ p0 = cmp.gt(r13,#-1) |
| 742 | \\ r3:2 = #0 |
| 743 | \\ } |
| 744 | \\ { |
| 745 | \\ if (!p0) r1:0 = r9:8 |
| 746 | \\ if (p0) r1:0 = r13:12 |
| 747 | \\ r3 = ##0x80000000 |
| 748 | \\ } |
| 749 | \\ { |
| 750 | \\ if (!p0) r1 = or(r1,r3) |
| 751 | \\ p0 = dfcmp.eq(r1:0,r3:2) |
| 752 | \\ if (p0.new) jump:nt .Lzero_plus_zero |
| 753 | \\ } |
| 754 | \\ { |
| 755 | \\ jumpr r31 |
| 756 | \\ } |
| 757 | \\ |
| 758 | \\ .Linvalid_nan_add: |
| 759 | \\ { |
| 760 | \\ r28 = convert_df2sf(r1:0) |
| 761 | \\ p0 = dfclass(r3:2,#0x0f) |
| 762 | \\ if (p0.new) r3:2 = r1:0 |
| 763 | \\ } |
| 764 | \\ { |
| 765 | \\ r2 = convert_df2sf(r3:2) |
| 766 | \\ r1:0 = #-1 |
| 767 | \\ jumpr r31 |
| 768 | \\ } |
| 769 | \\ .falign |
| 770 | \\ .LB_zero: |
| 771 | \\ { |
| 772 | \\ p0 = dfcmp.eq(r13:12,r1:0) |
| 773 | \\ if (!p0.new) jumpr:t r31 |
| 774 | \\ } |
| 775 | \\ |
| 776 | \\ |
| 777 | \\ |
| 778 | \\ |
| 779 | \\ .Lzero_plus_zero: |
| 780 | \\ { |
| 781 | \\ p0 = cmp.eq(r1:0,r3:2) |
| 782 | \\ if (p0.new) jumpr:t r31 |
| 783 | \\ } |
| 784 | \\ { |
| 785 | \\ r28 = USR |
| 786 | \\ } |
| 787 | \\ { |
| 788 | \\ r28 = extractu(r28,#2,#22) |
| 789 | \\ r1:0 = #0 |
| 790 | \\ } |
| 791 | \\ { |
| 792 | \\ p0 = cmp.eq(r28,#2) |
| 793 | \\ if (p0.new) r1 = ##0x80000000 |
| 794 | \\ jumpr r31 |
| 795 | \\ } |
| 796 | \\ .Linf_add: |
| 797 | \\ |
| 798 | \\ { |
| 799 | \\ p0 = !cmp.eq(r1,r3) |
| 800 | \\ p0 = dfclass(r3:2,#8) |
| 801 | \\ if (!p0.new) jumpr:t r31 |
| 802 | \\ } |
| 803 | \\ { |
| 804 | \\ r2 = ##0x7f800001 |
| 805 | \\ } |
| 806 | \\ { |
| 807 | \\ r1:0 = convert_sf2df(r2) |
| 808 | \\ jumpr r31 |
| 809 | \\ } |
| 810 | ); |
| 811 | } |
| 812 | |
| 813 | fn __hexagon_subdf3() align(32) callconv(.naked) noreturn { |
| 814 | asm volatile ( |
| 815 | \\ { |
| 816 | \\ r3 = togglebit(r3,#31) |
| 817 | \\ jump ##__hexagon_adddf3 |
| 818 | \\ } |
| 819 | ); |
| 820 | } |
| 821 | |
| 822 | fn __hexagon_divdf3() align(32) callconv(.naked) noreturn { |
| 823 | asm volatile ( |
| 824 | \\ { |
| 825 | \\ p2 = dfclass(r1:0,#0x02) |
| 826 | \\ p2 = dfclass(r3:2,#0x02) |
| 827 | \\ r13:12 = combine(r3,r1) |
| 828 | \\ r28 = xor(r1,r3) |
| 829 | \\ } |
| 830 | \\ { |
| 831 | \\ if (!p2) jump .Ldiv_abnormal |
| 832 | \\ r7:6 = extractu(r3:2,#23,#52 -23) |
| 833 | \\ r8 = ##0x3f800001 |
| 834 | \\ } |
| 835 | \\ { |
| 836 | \\ r9 = or(r8,r6) |
| 837 | \\ r13 = extractu(r13,#11,#52 -32) |
| 838 | \\ r12 = extractu(r12,#11,#52 -32) |
| 839 | \\ p3 = cmp.gt(r28,#-1) |
| 840 | \\ } |
| 841 | \\ |
| 842 | \\ |
| 843 | \\ .Ldenorm_continue: |
| 844 | \\ { |
| 845 | \\ r11,p0 = sfrecipa(r8,r9) |
| 846 | \\ r10 = and(r8,#-2) |
| 847 | \\ r28 = #1 |
| 848 | \\ r12 = sub(r12,r13) |
| 849 | \\ } |
| 850 | \\ |
| 851 | \\ |
| 852 | \\ { |
| 853 | \\ r10 -= sfmpy(r11,r9):lib |
| 854 | \\ r1 = insert(r28,#11 +1,#52 -32) |
| 855 | \\ r13 = ##0x00800000 << 3 |
| 856 | \\ } |
| 857 | \\ { |
| 858 | \\ r11 += sfmpy(r11,r10):lib |
| 859 | \\ r3 = insert(r28,#11 +1,#52 -32) |
| 860 | \\ r10 = and(r8,#-2) |
| 861 | \\ } |
| 862 | \\ { |
| 863 | \\ r10 -= sfmpy(r11,r9):lib |
| 864 | \\ r5 = #-0x3ff +1 |
| 865 | \\ r4 = #0x3ff -1 |
| 866 | \\ } |
| 867 | \\ { |
| 868 | \\ r11 += sfmpy(r11,r10):lib |
| 869 | \\ p1 = cmp.gt(r12,r5) |
| 870 | \\ p1 = !cmp.gt(r12,r4) |
| 871 | \\ } |
| 872 | \\ { |
| 873 | \\ r13 = insert(r11,#23,#3) |
| 874 | \\ r5:4 = #0 |
| 875 | \\ r12 = add(r12,#-61) |
| 876 | \\ } |
| 877 | \\ |
| 878 | \\ |
| 879 | \\ |
| 880 | \\ |
| 881 | \\ { |
| 882 | \\ r13 = add(r13,#((-3) << 3)) |
| 883 | \\ } |
| 884 | \\ { r7:6 = mpyu(r13,r1); r1:0 = asl(r1:0,# ( 15 )); }; { r6 = # 0; r1:0 -= mpyu(r7,r2); r15:14 = mpyu(r7,r3); }; { r5:4 += ASL(r7:6, # ( 14 )); r1:0 -= asl(r15:14, # 32); } |
| 885 | \\ { r7:6 = mpyu(r13,r1); r1:0 = asl(r1:0,# ( 15 )); }; { r6 = # 0; r1:0 -= mpyu(r7,r2); r15:14 = mpyu(r7,r3); }; { r5:4 += ASR(r7:6, # ( 1 )); r1:0 -= asl(r15:14, # 32); } |
| 886 | \\ { r7:6 = mpyu(r13,r1); r1:0 = asl(r1:0,# ( 15 )); }; { r6 = # 0; r1:0 -= mpyu(r7,r2); r15:14 = mpyu(r7,r3); }; { r5:4 += ASR(r7:6, # ( 16 )); r1:0 -= asl(r15:14, # 32); } |
| 887 | \\ { r7:6 = mpyu(r13,r1); r1:0 = asl(r1:0,# ( 15 )); }; { r6 = # 0; r1:0 -= mpyu(r7,r2); r15:14 = mpyu(r7,r3); }; { r5:4 += ASR(r7:6, # ( 31 )); r1:0 -= asl(r15:14, # 32); r7:6=# ( 0 ); } |
| 888 | \\ |
| 889 | \\ |
| 890 | \\ |
| 891 | \\ |
| 892 | \\ |
| 893 | \\ |
| 894 | \\ |
| 895 | \\ { |
| 896 | \\ |
| 897 | \\ r15:14 = sub(r1:0,r3:2) |
| 898 | \\ p0 = cmp.gtu(r3:2,r1:0) |
| 899 | \\ |
| 900 | \\ if (!p0.new) r6 = #2 |
| 901 | \\ } |
| 902 | \\ { |
| 903 | \\ r5:4 = add(r5:4,r7:6) |
| 904 | \\ if (!p0) r1:0 = r15:14 |
| 905 | \\ r15:14 = #0 |
| 906 | \\ } |
| 907 | \\ { |
| 908 | \\ p0 = cmp.eq(r1:0,r15:14) |
| 909 | \\ if (!p0.new) r4 = or(r4,r28) |
| 910 | \\ } |
| 911 | \\ { |
| 912 | \\ r7:6 = neg(r5:4) |
| 913 | \\ } |
| 914 | \\ { |
| 915 | \\ if (!p3) r5:4 = r7:6 |
| 916 | \\ } |
| 917 | \\ { |
| 918 | \\ r1:0 = convert_d2df(r5:4) |
| 919 | \\ if (!p1) jump .Ldiv_ovf_unf |
| 920 | \\ } |
| 921 | \\ { |
| 922 | \\ r1 += asl(r12,#52 -32) |
| 923 | \\ jumpr r31 |
| 924 | \\ } |
| 925 | \\ |
| 926 | \\ .Ldiv_ovf_unf: |
| 927 | \\ { |
| 928 | \\ r1 += asl(r12,#52 -32) |
| 929 | \\ r13 = extractu(r1,#11,#52 -32) |
| 930 | \\ } |
| 931 | \\ { |
| 932 | \\ r7:6 = abs(r5:4) |
| 933 | \\ r12 = add(r12,r13) |
| 934 | \\ } |
| 935 | \\ { |
| 936 | \\ p0 = cmp.gt(r12,##0x3ff +0x3ff) |
| 937 | \\ if (p0.new) jump:nt .Ldiv_ovf |
| 938 | \\ } |
| 939 | \\ { |
| 940 | \\ p0 = cmp.gt(r12,#0) |
| 941 | \\ if (p0.new) jump:nt .Lpossible_unf |
| 942 | \\ } |
| 943 | \\ { |
| 944 | \\ r13 = add(clb(r7:6),#-1) |
| 945 | \\ r12 = sub(#7,r12) |
| 946 | \\ r10 = USR |
| 947 | \\ r11 = #63 |
| 948 | \\ } |
| 949 | \\ { |
| 950 | \\ r13 = min(r12,r11) |
| 951 | \\ r11 = or(r10,#0x030) |
| 952 | \\ r7:6 = asl(r7:6,r13) |
| 953 | \\ r12 = #0 |
| 954 | \\ } |
| 955 | \\ { |
| 956 | \\ r15:14 = extractu(r7:6,r13:12) |
| 957 | \\ r7:6 = lsr(r7:6,r13) |
| 958 | \\ r3:2 = #1 |
| 959 | \\ } |
| 960 | \\ { |
| 961 | \\ p0 = cmp.gtu(r3:2,r15:14) |
| 962 | \\ if (!p0.new) r6 = or(r2,r6) |
| 963 | \\ r7 = setbit(r7,#52 -32+4) |
| 964 | \\ } |
| 965 | \\ { |
| 966 | \\ r5:4 = neg(r7:6) |
| 967 | \\ p0 = bitsclr(r6,#(1<<4)-1) |
| 968 | \\ if (!p0.new) r10 = r11 |
| 969 | \\ } |
| 970 | \\ { |
| 971 | \\ USR = r10 |
| 972 | \\ if (p3) r5:4 = r7:6 |
| 973 | \\ r10 = #-0x3ff -(52 +4) |
| 974 | \\ } |
| 975 | \\ { |
| 976 | \\ r1:0 = convert_d2df(r5:4) |
| 977 | \\ } |
| 978 | \\ { |
| 979 | \\ r1 += asl(r10,#52 -32) |
| 980 | \\ jumpr r31 |
| 981 | \\ } |
| 982 | \\ |
| 983 | \\ |
| 984 | \\ .Lpossible_unf: |
| 985 | \\ |
| 986 | \\ |
| 987 | \\ { |
| 988 | \\ r3:2 = extractu(r1:0,#63,#0) |
| 989 | \\ r15:14 = combine(##0x00100000,#0) |
| 990 | \\ r10 = #0x7FFF |
| 991 | \\ } |
| 992 | \\ { |
| 993 | \\ p0 = dfcmp.eq(r15:14,r3:2) |
| 994 | \\ p0 = bitsset(r7,r10) |
| 995 | \\ } |
| 996 | \\ |
| 997 | \\ |
| 998 | \\ |
| 999 | \\ |
| 1000 | \\ |
| 1001 | \\ |
| 1002 | \\ { |
| 1003 | \\ if (!p0) jumpr r31 |
| 1004 | \\ r10 = USR |
| 1005 | \\ } |
| 1006 | \\ |
| 1007 | \\ { |
| 1008 | \\ r10 = or(r10,#0x30) |
| 1009 | \\ } |
| 1010 | \\ { |
| 1011 | \\ USR = r10 |
| 1012 | \\ } |
| 1013 | \\ { |
| 1014 | \\ p0 = dfcmp.eq(r1:0,r1:0) |
| 1015 | \\ jumpr r31 |
| 1016 | \\ } |
| 1017 | \\ |
| 1018 | \\ .Ldiv_ovf: |
| 1019 | \\ |
| 1020 | \\ |
| 1021 | \\ |
| 1022 | \\ { |
| 1023 | \\ r10 = USR |
| 1024 | \\ r3:2 = combine(##0x7fefffff,#-1) |
| 1025 | \\ r1 = mux(p3,#0,#-1) |
| 1026 | \\ } |
| 1027 | \\ { |
| 1028 | \\ r7:6 = combine(##0x7ff00000,#0) |
| 1029 | \\ r5 = extractu(r10,#2,#22) |
| 1030 | \\ r10 = or(r10,#0x28) |
| 1031 | \\ } |
| 1032 | \\ { |
| 1033 | \\ USR = r10 |
| 1034 | \\ r5 ^= lsr(r1,#31) |
| 1035 | \\ r4 = r5 |
| 1036 | \\ } |
| 1037 | \\ { |
| 1038 | \\ p0 = !cmp.eq(r4,#1) |
| 1039 | \\ p0 = !cmp.eq(r5,#2) |
| 1040 | \\ if (p0.new) r3:2 = r7:6 |
| 1041 | \\ p0 = dfcmp.eq(r3:2,r3:2) |
| 1042 | \\ } |
| 1043 | \\ { |
| 1044 | \\ r1:0 = insert(r3:2,#63,#0) |
| 1045 | \\ jumpr r31 |
| 1046 | \\ } |
| 1047 | \\ |
| 1048 | \\ |
| 1049 | \\ |
| 1050 | \\ |
| 1051 | \\ |
| 1052 | \\ |
| 1053 | \\ |
| 1054 | \\ .Ldiv_abnormal: |
| 1055 | \\ { |
| 1056 | \\ p0 = dfclass(r1:0,#0x0F) |
| 1057 | \\ p0 = dfclass(r3:2,#0x0F) |
| 1058 | \\ p3 = cmp.gt(r28,#-1) |
| 1059 | \\ } |
| 1060 | \\ { |
| 1061 | \\ p1 = dfclass(r1:0,#0x08) |
| 1062 | \\ p1 = dfclass(r3:2,#0x08) |
| 1063 | \\ } |
| 1064 | \\ { |
| 1065 | \\ p2 = dfclass(r1:0,#0x01) |
| 1066 | \\ p2 = dfclass(r3:2,#0x01) |
| 1067 | \\ } |
| 1068 | \\ { |
| 1069 | \\ if (!p0) jump .Ldiv_nan |
| 1070 | \\ if (p1) jump .Ldiv_invalid |
| 1071 | \\ } |
| 1072 | \\ { |
| 1073 | \\ if (p2) jump .Ldiv_invalid |
| 1074 | \\ } |
| 1075 | \\ { |
| 1076 | \\ p2 = dfclass(r1:0,#(0x0F ^ 0x01)) |
| 1077 | \\ p2 = dfclass(r3:2,#(0x0F ^ 0x08)) |
| 1078 | \\ } |
| 1079 | \\ { |
| 1080 | \\ p1 = dfclass(r1:0,#(0x0F ^ 0x08)) |
| 1081 | \\ p1 = dfclass(r3:2,#(0x0F ^ 0x01)) |
| 1082 | \\ } |
| 1083 | \\ { |
| 1084 | \\ if (!p2) jump .Ldiv_zero_result |
| 1085 | \\ if (!p1) jump .Ldiv_inf_result |
| 1086 | \\ } |
| 1087 | \\ |
| 1088 | \\ |
| 1089 | \\ |
| 1090 | \\ |
| 1091 | \\ |
| 1092 | \\ { |
| 1093 | \\ p0 = dfclass(r1:0,#0x02) |
| 1094 | \\ p1 = dfclass(r3:2,#0x02) |
| 1095 | \\ r10 = ##0x00100000 |
| 1096 | \\ } |
| 1097 | \\ { |
| 1098 | \\ r13:12 = combine(r3,r1) |
| 1099 | \\ r1 = insert(r10,#11 +1,#52 -32) |
| 1100 | \\ r3 = insert(r10,#11 +1,#52 -32) |
| 1101 | \\ } |
| 1102 | \\ { |
| 1103 | \\ if (p0) r1 = or(r1,r10) |
| 1104 | \\ if (p1) r3 = or(r3,r10) |
| 1105 | \\ } |
| 1106 | \\ { |
| 1107 | \\ r5 = add(clb(r1:0),#-11) |
| 1108 | \\ r4 = add(clb(r3:2),#-11) |
| 1109 | \\ r10 = #1 |
| 1110 | \\ } |
| 1111 | \\ { |
| 1112 | \\ r12 = extractu(r12,#11,#52 -32) |
| 1113 | \\ r13 = extractu(r13,#11,#52 -32) |
| 1114 | \\ } |
| 1115 | \\ { |
| 1116 | \\ r1:0 = asl(r1:0,r5) |
| 1117 | \\ r3:2 = asl(r3:2,r4) |
| 1118 | \\ if (!p0) r12 = sub(r10,r5) |
| 1119 | \\ if (!p1) r13 = sub(r10,r4) |
| 1120 | \\ } |
| 1121 | \\ { |
| 1122 | \\ r7:6 = extractu(r3:2,#23,#52 -23) |
| 1123 | \\ } |
| 1124 | \\ { |
| 1125 | \\ r9 = or(r8,r6) |
| 1126 | \\ jump .Ldenorm_continue |
| 1127 | \\ } |
| 1128 | \\ |
| 1129 | \\ .Ldiv_zero_result: |
| 1130 | \\ { |
| 1131 | \\ r1 = xor(r1,r3) |
| 1132 | \\ r3:2 = #0 |
| 1133 | \\ } |
| 1134 | \\ { |
| 1135 | \\ r1:0 = insert(r3:2,#63,#0) |
| 1136 | \\ jumpr r31 |
| 1137 | \\ } |
| 1138 | \\ .Ldiv_inf_result: |
| 1139 | \\ { |
| 1140 | \\ p2 = dfclass(r3:2,#0x01) |
| 1141 | \\ p2 = dfclass(r1:0,#(0x0F ^ 0x08)) |
| 1142 | \\ } |
| 1143 | \\ { |
| 1144 | \\ r10 = USR |
| 1145 | \\ if (!p2) jump 1f |
| 1146 | \\ r1 = xor(r1,r3) |
| 1147 | \\ } |
| 1148 | \\ { |
| 1149 | \\ r10 = or(r10,#0x04) |
| 1150 | \\ } |
| 1151 | \\ { |
| 1152 | \\ USR = r10 |
| 1153 | \\ } |
| 1154 | \\ 1: |
| 1155 | \\ { |
| 1156 | \\ r3:2 = combine(##0x7ff00000,#0) |
| 1157 | \\ p0 = dfcmp.uo(r3:2,r3:2) |
| 1158 | \\ } |
| 1159 | \\ { |
| 1160 | \\ r1:0 = insert(r3:2,#63,#0) |
| 1161 | \\ jumpr r31 |
| 1162 | \\ } |
| 1163 | \\ .Ldiv_nan: |
| 1164 | \\ { |
| 1165 | \\ p0 = dfclass(r1:0,#0x10) |
| 1166 | \\ p1 = dfclass(r3:2,#0x10) |
| 1167 | \\ if (!p0.new) r1:0 = r3:2 |
| 1168 | \\ if (!p1.new) r3:2 = r1:0 |
| 1169 | \\ } |
| 1170 | \\ { |
| 1171 | \\ r5 = convert_df2sf(r1:0) |
| 1172 | \\ r4 = convert_df2sf(r3:2) |
| 1173 | \\ } |
| 1174 | \\ { |
| 1175 | \\ r1:0 = #-1 |
| 1176 | \\ jumpr r31 |
| 1177 | \\ } |
| 1178 | \\ |
| 1179 | \\ .Ldiv_invalid: |
| 1180 | \\ { |
| 1181 | \\ r10 = ##0x7f800001 |
| 1182 | \\ } |
| 1183 | \\ { |
| 1184 | \\ r1:0 = convert_sf2df(r10) |
| 1185 | \\ jumpr r31 |
| 1186 | \\ } |
| 1187 | ); |
| 1188 | } |
| 1189 | |
| 1190 | fn __hexagon_muldf3() align(32) callconv(.naked) noreturn { |
| 1191 | asm volatile ( |
| 1192 | \\ { |
| 1193 | \\ p0 = dfclass(r1:0,#2) |
| 1194 | \\ p0 = dfclass(r3:2,#2) |
| 1195 | \\ r13:12 = combine(##0x40000000,#0) |
| 1196 | \\ } |
| 1197 | \\ { |
| 1198 | \\ r13:12 = insert(r1:0,#52,#11 -1) |
| 1199 | \\ r5:4 = asl(r3:2,#11 -1) |
| 1200 | \\ r28 = #-1024 |
| 1201 | \\ r9:8 = #1 |
| 1202 | \\ } |
| 1203 | \\ { |
| 1204 | \\ r7:6 = mpyu(r4,r13) |
| 1205 | \\ r5:4 = insert(r9:8,#2,#62) |
| 1206 | \\ } |
| 1207 | \\ |
| 1208 | \\ |
| 1209 | \\ |
| 1210 | \\ |
| 1211 | \\ { |
| 1212 | \\ r15:14 = mpyu(r12,r4) |
| 1213 | \\ r7:6 += mpyu(r12,r5) |
| 1214 | \\ } |
| 1215 | \\ { |
| 1216 | \\ r7:6 += lsr(r15:14,#32) |
| 1217 | \\ r11:10 = mpyu(r13,r5) |
| 1218 | \\ r5:4 = combine(##1024 +1024 -4,#0) |
| 1219 | \\ } |
| 1220 | \\ { |
| 1221 | \\ r11:10 += lsr(r7:6,#32) |
| 1222 | \\ if (!p0) jump .Lmul_abnormal |
| 1223 | \\ p1 = cmp.eq(r14,#0) |
| 1224 | \\ p1 = cmp.eq(r6,#0) |
| 1225 | \\ } |
| 1226 | \\ { |
| 1227 | \\ if (!p1) r10 = or(r10,r8) |
| 1228 | \\ r6 = extractu(r1,#11,#20) |
| 1229 | \\ r7 = extractu(r3,#11,#20) |
| 1230 | \\ } |
| 1231 | \\ { |
| 1232 | \\ r15:14 = neg(r11:10) |
| 1233 | \\ r6 += add(r28,r7) |
| 1234 | \\ r28 = xor(r1,r3) |
| 1235 | \\ } |
| 1236 | \\ { |
| 1237 | \\ if (!p2.new) r11:10 = r15:14 |
| 1238 | \\ p2 = cmp.gt(r28,#-1) |
| 1239 | \\ p0 = !cmp.gt(r6,r5) |
| 1240 | \\ p0 = cmp.gt(r6,r4) |
| 1241 | \\ if (!p0.new) jump:nt .Lmul_ovf_unf |
| 1242 | \\ } |
| 1243 | \\ { |
| 1244 | \\ r1:0 = convert_d2df(r11:10) |
| 1245 | \\ r6 = add(r6,#-1024 -58) |
| 1246 | \\ } |
| 1247 | \\ { |
| 1248 | \\ r1 += asl(r6,#20) |
| 1249 | \\ jumpr r31 |
| 1250 | \\ } |
| 1251 | \\ |
| 1252 | \\ .falign |
| 1253 | \\ .Lpossible_unf: |
| 1254 | \\ { |
| 1255 | \\ p0 = cmp.eq(r0,#0) |
| 1256 | \\ p0 = bitsclr(r1,r4) |
| 1257 | \\ if (!p0.new) jumpr:t r31 |
| 1258 | \\ r5 = #0x7fff |
| 1259 | \\ } |
| 1260 | \\ { |
| 1261 | \\ p0 = bitsset(r13,r5) |
| 1262 | \\ r4 = USR |
| 1263 | \\ r5 = #0x030 |
| 1264 | \\ } |
| 1265 | \\ { |
| 1266 | \\ if (p0) r4 = or(r4,r5) |
| 1267 | \\ } |
| 1268 | \\ { |
| 1269 | \\ USR = r4 |
| 1270 | \\ } |
| 1271 | \\ { |
| 1272 | \\ p0 = dfcmp.eq(r1:0,r1:0) |
| 1273 | \\ jumpr r31 |
| 1274 | \\ } |
| 1275 | \\ .falign |
| 1276 | \\ .Lmul_ovf_unf: |
| 1277 | \\ { |
| 1278 | \\ r1:0 = convert_d2df(r11:10) |
| 1279 | \\ r13:12 = abs(r11:10) |
| 1280 | \\ r7 = add(r6,#-1024 -58) |
| 1281 | \\ } |
| 1282 | \\ { |
| 1283 | \\ r1 += asl(r7,#20) |
| 1284 | \\ r7 = extractu(r1,#11,#20) |
| 1285 | \\ r4 = ##0x7FEFFFFF |
| 1286 | \\ } |
| 1287 | \\ { |
| 1288 | \\ r7 += add(r6,##-1024 -58) |
| 1289 | \\ |
| 1290 | \\ r5 = #0 |
| 1291 | \\ } |
| 1292 | \\ { |
| 1293 | \\ p0 = cmp.gt(r7,##1024 +1024 -2) |
| 1294 | \\ if (p0.new) jump:nt .Lmul_ovf |
| 1295 | \\ } |
| 1296 | \\ { |
| 1297 | \\ p0 = cmp.gt(r7,#0) |
| 1298 | \\ if (p0.new) jump:nt .Lpossible_unf |
| 1299 | \\ r5 = sub(r6,r5) |
| 1300 | \\ r28 = #63 |
| 1301 | \\ } |
| 1302 | \\ { |
| 1303 | \\ r4 = #0 |
| 1304 | \\ r5 = sub(#5,r5) |
| 1305 | \\ } |
| 1306 | \\ { |
| 1307 | \\ p3 = cmp.gt(r11,#-1) |
| 1308 | \\ r5 = min(r5,r28) |
| 1309 | \\ r11:10 = r13:12 |
| 1310 | \\ } |
| 1311 | \\ { |
| 1312 | \\ r28 = USR |
| 1313 | \\ r15:14 = extractu(r11:10,r5:4) |
| 1314 | \\ } |
| 1315 | \\ { |
| 1316 | \\ r11:10 = asr(r11:10,r5) |
| 1317 | \\ r4 = #0x0030 |
| 1318 | \\ r1 = insert(r9,#11,#20) |
| 1319 | \\ } |
| 1320 | \\ { |
| 1321 | \\ p0 = cmp.gtu(r9:8,r15:14) |
| 1322 | \\ if (!p0.new) r10 = or(r10,r8) |
| 1323 | \\ r11 = setbit(r11,#20 +3) |
| 1324 | \\ } |
| 1325 | \\ { |
| 1326 | \\ r15:14 = neg(r11:10) |
| 1327 | \\ p1 = bitsclr(r10,#0x7) |
| 1328 | \\ if (!p1.new) r28 = or(r4,r28) |
| 1329 | \\ } |
| 1330 | \\ { |
| 1331 | \\ if (!p3) r11:10 = r15:14 |
| 1332 | \\ USR = r28 |
| 1333 | \\ } |
| 1334 | \\ { |
| 1335 | \\ r1:0 = convert_d2df(r11:10) |
| 1336 | \\ p0 = dfcmp.eq(r1:0,r1:0) |
| 1337 | \\ } |
| 1338 | \\ { |
| 1339 | \\ r1 = insert(r9,#11 -1,#20 +1) |
| 1340 | \\ jumpr r31 |
| 1341 | \\ } |
| 1342 | \\ .falign |
| 1343 | \\ .Lmul_ovf: |
| 1344 | \\ |
| 1345 | \\ { |
| 1346 | \\ r28 = USR |
| 1347 | \\ r13:12 = combine(##0x7fefffff,#-1) |
| 1348 | \\ r1:0 = r11:10 |
| 1349 | \\ } |
| 1350 | \\ { |
| 1351 | \\ r14 = extractu(r28,#2,#22) |
| 1352 | \\ r28 = or(r28,#0x28) |
| 1353 | \\ r5:4 = combine(##0x7ff00000,#0) |
| 1354 | \\ } |
| 1355 | \\ { |
| 1356 | \\ USR = r28 |
| 1357 | \\ r14 ^= lsr(r1,#31) |
| 1358 | \\ r28 = r14 |
| 1359 | \\ } |
| 1360 | \\ { |
| 1361 | \\ p0 = !cmp.eq(r28,#1) |
| 1362 | \\ p0 = !cmp.eq(r14,#2) |
| 1363 | \\ if (p0.new) r13:12 = r5:4 |
| 1364 | \\ p0 = dfcmp.eq(r1:0,r1:0) |
| 1365 | \\ } |
| 1366 | \\ { |
| 1367 | \\ r1:0 = insert(r13:12,#63,#0) |
| 1368 | \\ jumpr r31 |
| 1369 | \\ } |
| 1370 | \\ |
| 1371 | \\ .Lmul_abnormal: |
| 1372 | \\ { |
| 1373 | \\ r13:12 = extractu(r1:0,#63,#0) |
| 1374 | \\ r5:4 = extractu(r3:2,#63,#0) |
| 1375 | \\ } |
| 1376 | \\ { |
| 1377 | \\ p3 = cmp.gtu(r13:12,r5:4) |
| 1378 | \\ if (!p3.new) r1:0 = r3:2 |
| 1379 | \\ if (!p3.new) r3:2 = r1:0 |
| 1380 | \\ } |
| 1381 | \\ { |
| 1382 | \\ |
| 1383 | \\ p0 = dfclass(r1:0,#0x0f) |
| 1384 | \\ if (!p0.new) jump:nt .Linvalid_nan |
| 1385 | \\ if (!p3) r13:12 = r5:4 |
| 1386 | \\ if (!p3) r5:4 = r13:12 |
| 1387 | \\ } |
| 1388 | \\ { |
| 1389 | \\ |
| 1390 | \\ p1 = dfclass(r1:0,#0x08) |
| 1391 | \\ p1 = dfclass(r3:2,#0x0e) |
| 1392 | \\ } |
| 1393 | \\ { |
| 1394 | \\ |
| 1395 | \\ |
| 1396 | \\ p0 = dfclass(r1:0,#0x08) |
| 1397 | \\ p0 = dfclass(r3:2,#0x01) |
| 1398 | \\ } |
| 1399 | \\ { |
| 1400 | \\ if (p1) jump .Ltrue_inf |
| 1401 | \\ p2 = dfclass(r3:2,#0x01) |
| 1402 | \\ } |
| 1403 | \\ { |
| 1404 | \\ if (p0) jump .Linvalid_zeroinf |
| 1405 | \\ if (p2) jump .Ltrue_zero |
| 1406 | \\ r28 = ##0x7c000000 |
| 1407 | \\ } |
| 1408 | \\ |
| 1409 | \\ |
| 1410 | \\ |
| 1411 | \\ |
| 1412 | \\ |
| 1413 | \\ { |
| 1414 | \\ p0 = bitsclr(r1,r28) |
| 1415 | \\ if (p0.new) jump:nt .Lmul_tiny |
| 1416 | \\ } |
| 1417 | \\ { |
| 1418 | \\ r28 = cl0(r5:4) |
| 1419 | \\ } |
| 1420 | \\ { |
| 1421 | \\ r28 = add(r28,#-11) |
| 1422 | \\ } |
| 1423 | \\ { |
| 1424 | \\ r5:4 = asl(r5:4,r28) |
| 1425 | \\ } |
| 1426 | \\ { |
| 1427 | \\ r3:2 = insert(r5:4,#63,#0) |
| 1428 | \\ r1 -= asl(r28,#20) |
| 1429 | \\ } |
| 1430 | \\ jump __hexagon_muldf3 |
| 1431 | \\ .Lmul_tiny: |
| 1432 | \\ { |
| 1433 | \\ r28 = USR |
| 1434 | \\ r1:0 = xor(r1:0,r3:2) |
| 1435 | \\ } |
| 1436 | \\ { |
| 1437 | \\ r28 = or(r28,#0x30) |
| 1438 | \\ r1:0 = insert(r9:8,#63,#0) |
| 1439 | \\ r5 = extractu(r28,#2,#22) |
| 1440 | \\ } |
| 1441 | \\ { |
| 1442 | \\ USR = r28 |
| 1443 | \\ p0 = cmp.gt(r5,#1) |
| 1444 | \\ if (!p0.new) r0 = #0 |
| 1445 | \\ r5 ^= lsr(r1,#31) |
| 1446 | \\ } |
| 1447 | \\ { |
| 1448 | \\ p0 = cmp.eq(r5,#3) |
| 1449 | \\ if (!p0.new) r0 = #0 |
| 1450 | \\ jumpr r31 |
| 1451 | \\ } |
| 1452 | \\ .Linvalid_zeroinf: |
| 1453 | \\ { |
| 1454 | \\ r28 = USR |
| 1455 | \\ } |
| 1456 | \\ { |
| 1457 | \\ r1:0 = #-1 |
| 1458 | \\ r28 = or(r28,#2) |
| 1459 | \\ } |
| 1460 | \\ { |
| 1461 | \\ USR = r28 |
| 1462 | \\ } |
| 1463 | \\ { |
| 1464 | \\ p0 = dfcmp.uo(r1:0,r1:0) |
| 1465 | \\ jumpr r31 |
| 1466 | \\ } |
| 1467 | \\ .Linvalid_nan: |
| 1468 | \\ { |
| 1469 | \\ p0 = dfclass(r3:2,#0x0f) |
| 1470 | \\ r28 = convert_df2sf(r1:0) |
| 1471 | \\ if (p0.new) r3:2 = r1:0 |
| 1472 | \\ } |
| 1473 | \\ { |
| 1474 | \\ r2 = convert_df2sf(r3:2) |
| 1475 | \\ r1:0 = #-1 |
| 1476 | \\ jumpr r31 |
| 1477 | \\ } |
| 1478 | \\ .falign |
| 1479 | \\ .Ltrue_zero: |
| 1480 | \\ { |
| 1481 | \\ r1:0 = r3:2 |
| 1482 | \\ r3:2 = r1:0 |
| 1483 | \\ } |
| 1484 | \\ .Ltrue_inf: |
| 1485 | \\ { |
| 1486 | \\ r3 = extract(r3,#1,#31) |
| 1487 | \\ } |
| 1488 | \\ { |
| 1489 | \\ r1 ^= asl(r3,#31) |
| 1490 | \\ jumpr r31 |
| 1491 | \\ } |
| 1492 | ); |
| 1493 | } |
| 1494 | |
| 1495 | fn __hexagon_sqrtdf2() align(32) callconv(.naked) noreturn { |
| 1496 | asm volatile ( |
| 1497 | \\ { |
| 1498 | \\ r15:14 = extractu(r1:0,#23 +1,#52 -23) |
| 1499 | \\ r28 = extractu(r1,#11,#52 -32) |
| 1500 | \\ r5:4 = combine(##0x3f000004,#1) |
| 1501 | \\ } |
| 1502 | \\ { |
| 1503 | \\ p2 = dfclass(r1:0,#0x02) |
| 1504 | \\ p2 = cmp.gt(r1,#-1) |
| 1505 | \\ if (!p2.new) jump:nt .Lsqrt_abnormal |
| 1506 | \\ r9 = or(r5,r14) |
| 1507 | \\ } |
| 1508 | \\ |
| 1509 | \\ .Ldenormal_restart: |
| 1510 | \\ { |
| 1511 | \\ r11:10 = r1:0 |
| 1512 | \\ r7,p0 = sfinvsqrta(r9) |
| 1513 | \\ r5 = and(r5,#-16) |
| 1514 | \\ r3:2 = #0 |
| 1515 | \\ } |
| 1516 | \\ { |
| 1517 | \\ r3 += sfmpy(r7,r9):lib |
| 1518 | \\ r2 += sfmpy(r7,r5):lib |
| 1519 | \\ r6 = r5 |
| 1520 | \\ |
| 1521 | \\ |
| 1522 | \\ r9 = and(r28,#1) |
| 1523 | \\ } |
| 1524 | \\ { |
| 1525 | \\ r6 -= sfmpy(r3,r2):lib |
| 1526 | \\ r11 = insert(r4,#11 +1,#52 -32) |
| 1527 | \\ p1 = cmp.gtu(r9,#0) |
| 1528 | \\ } |
| 1529 | \\ { |
| 1530 | \\ r3 += sfmpy(r3,r6):lib |
| 1531 | \\ r2 += sfmpy(r2,r6):lib |
| 1532 | \\ r6 = r5 |
| 1533 | \\ r9 = mux(p1,#8,#9) |
| 1534 | \\ } |
| 1535 | \\ { |
| 1536 | \\ r6 -= sfmpy(r3,r2):lib |
| 1537 | \\ r11:10 = asl(r11:10,r9) |
| 1538 | \\ r9 = mux(p1,#3,#2) |
| 1539 | \\ } |
| 1540 | \\ { |
| 1541 | \\ r2 += sfmpy(r2,r6):lib |
| 1542 | \\ |
| 1543 | \\ r15:14 = asl(r11:10,r9) |
| 1544 | \\ } |
| 1545 | \\ { |
| 1546 | \\ r2 = and(r2,##0x007fffff) |
| 1547 | \\ } |
| 1548 | \\ { |
| 1549 | \\ r2 = add(r2,##0x00800000 - 3) |
| 1550 | \\ r9 = mux(p1,#7,#8) |
| 1551 | \\ } |
| 1552 | \\ { |
| 1553 | \\ r8 = asl(r2,r9) |
| 1554 | \\ r9 = mux(p1,#15-(1+1),#15-(1+0)) |
| 1555 | \\ } |
| 1556 | \\ { |
| 1557 | \\ r13:12 = mpyu(r8,r15) |
| 1558 | \\ } |
| 1559 | \\ { |
| 1560 | \\ r1:0 = asl(r11:10,#15) |
| 1561 | \\ r15:14 = mpyu(r13,r13) |
| 1562 | \\ p1 = cmp.eq(r0,r0) |
| 1563 | \\ } |
| 1564 | \\ { |
| 1565 | \\ r1:0 -= asl(r15:14,#15) |
| 1566 | \\ r15:14 = mpyu(r13,r12) |
| 1567 | \\ p2 = cmp.eq(r0,r0) |
| 1568 | \\ } |
| 1569 | \\ { |
| 1570 | \\ r1:0 -= lsr(r15:14,#16) |
| 1571 | \\ p3 = cmp.eq(r0,r0) |
| 1572 | \\ } |
| 1573 | \\ { |
| 1574 | \\ r1:0 = mpyu(r1,r8) |
| 1575 | \\ } |
| 1576 | \\ { |
| 1577 | \\ r13:12 += lsr(r1:0,r9) |
| 1578 | \\ r9 = add(r9,#16) |
| 1579 | \\ r1:0 = asl(r11:10,#31) |
| 1580 | \\ } |
| 1581 | \\ |
| 1582 | \\ { |
| 1583 | \\ r15:14 = mpyu(r13,r13) |
| 1584 | \\ r1:0 -= mpyu(r13,r12) |
| 1585 | \\ } |
| 1586 | \\ { |
| 1587 | \\ r1:0 -= asl(r15:14,#31) |
| 1588 | \\ r15:14 = mpyu(r12,r12) |
| 1589 | \\ } |
| 1590 | \\ { |
| 1591 | \\ r1:0 -= lsr(r15:14,#33) |
| 1592 | \\ } |
| 1593 | \\ { |
| 1594 | \\ r1:0 = mpyu(r1,r8) |
| 1595 | \\ } |
| 1596 | \\ { |
| 1597 | \\ r13:12 += lsr(r1:0,r9) |
| 1598 | \\ r9 = add(r9,#16) |
| 1599 | \\ r1:0 = asl(r11:10,#47) |
| 1600 | \\ } |
| 1601 | \\ |
| 1602 | \\ { |
| 1603 | \\ r15:14 = mpyu(r13,r13) |
| 1604 | \\ } |
| 1605 | \\ { |
| 1606 | \\ r1:0 -= asl(r15:14,#47) |
| 1607 | \\ r15:14 = mpyu(r13,r12) |
| 1608 | \\ } |
| 1609 | \\ { |
| 1610 | \\ r1:0 -= asl(r15:14,#16) |
| 1611 | \\ r15:14 = mpyu(r12,r12) |
| 1612 | \\ } |
| 1613 | \\ { |
| 1614 | \\ r1:0 -= lsr(r15:14,#17) |
| 1615 | \\ } |
| 1616 | \\ { |
| 1617 | \\ r1:0 = mpyu(r1,r8) |
| 1618 | \\ } |
| 1619 | \\ { |
| 1620 | \\ r13:12 += lsr(r1:0,r9) |
| 1621 | \\ } |
| 1622 | \\ { |
| 1623 | \\ r3:2 = mpyu(r13,r12) |
| 1624 | \\ r5:4 = mpyu(r12,r12) |
| 1625 | \\ r15:14 = #0 |
| 1626 | \\ r1:0 = #0 |
| 1627 | \\ } |
| 1628 | \\ { |
| 1629 | \\ r3:2 += lsr(r5:4,#33) |
| 1630 | \\ r5:4 += asl(r3:2,#33) |
| 1631 | \\ p1 = cmp.eq(r0,r0) |
| 1632 | \\ } |
| 1633 | \\ { |
| 1634 | \\ r7:6 = mpyu(r13,r13) |
| 1635 | \\ r1:0 = sub(r1:0,r5:4,p1):carry |
| 1636 | \\ r9:8 = #1 |
| 1637 | \\ } |
| 1638 | \\ { |
| 1639 | \\ r7:6 += lsr(r3:2,#31) |
| 1640 | \\ r9:8 += asl(r13:12,#1) |
| 1641 | \\ } |
| 1642 | \\ |
| 1643 | \\ |
| 1644 | \\ |
| 1645 | \\ |
| 1646 | \\ |
| 1647 | \\ { |
| 1648 | \\ r15:14 = sub(r11:10,r7:6,p1):carry |
| 1649 | \\ r5:4 = sub(r1:0,r9:8,p2):carry |
| 1650 | \\ |
| 1651 | \\ |
| 1652 | \\ |
| 1653 | \\ |
| 1654 | \\ r7:6 = #1 |
| 1655 | \\ r11:10 = #0 |
| 1656 | \\ } |
| 1657 | \\ { |
| 1658 | \\ r3:2 = sub(r15:14,r11:10,p2):carry |
| 1659 | \\ r7:6 = add(r13:12,r7:6) |
| 1660 | \\ r28 = add(r28,#-0x3ff) |
| 1661 | \\ } |
| 1662 | \\ { |
| 1663 | \\ |
| 1664 | \\ if (p2) r13:12 = r7:6 |
| 1665 | \\ if (p2) r1:0 = r5:4 |
| 1666 | \\ if (p2) r15:14 = r3:2 |
| 1667 | \\ } |
| 1668 | \\ { |
| 1669 | \\ r5:4 = sub(r1:0,r9:8,p3):carry |
| 1670 | \\ r7:6 = #1 |
| 1671 | \\ r28 = asr(r28,#1) |
| 1672 | \\ } |
| 1673 | \\ { |
| 1674 | \\ r3:2 = sub(r15:14,r11:10,p3):carry |
| 1675 | \\ r7:6 = add(r13:12,r7:6) |
| 1676 | \\ } |
| 1677 | \\ { |
| 1678 | \\ if (p3) r13:12 = r7:6 |
| 1679 | \\ if (p3) r1:0 = r5:4 |
| 1680 | \\ |
| 1681 | \\ |
| 1682 | \\ |
| 1683 | \\ |
| 1684 | \\ |
| 1685 | \\ r2 = #1 |
| 1686 | \\ } |
| 1687 | \\ { |
| 1688 | \\ p0 = cmp.eq(r1:0,r11:10) |
| 1689 | \\ if (!p0.new) r12 = or(r12,r2) |
| 1690 | \\ r3 = cl0(r13:12) |
| 1691 | \\ r28 = add(r28,#-63) |
| 1692 | \\ } |
| 1693 | \\ |
| 1694 | \\ |
| 1695 | \\ |
| 1696 | \\ { |
| 1697 | \\ r1:0 = convert_ud2df(r13:12) |
| 1698 | \\ r28 = add(r28,r3) |
| 1699 | \\ } |
| 1700 | \\ { |
| 1701 | \\ r1 += asl(r28,#52 -32) |
| 1702 | \\ jumpr r31 |
| 1703 | \\ } |
| 1704 | \\ .Lsqrt_abnormal: |
| 1705 | \\ { |
| 1706 | \\ p0 = dfclass(r1:0,#0x01) |
| 1707 | \\ if (p0.new) jumpr:t r31 |
| 1708 | \\ } |
| 1709 | \\ { |
| 1710 | \\ p0 = dfclass(r1:0,#0x10) |
| 1711 | \\ if (p0.new) jump:nt .Lsqrt_nan |
| 1712 | \\ } |
| 1713 | \\ { |
| 1714 | \\ p0 = cmp.gt(r1,#-1) |
| 1715 | \\ if (!p0.new) jump:nt .Lsqrt_invalid_neg |
| 1716 | \\ if (!p0.new) r28 = ##0x7F800001 |
| 1717 | \\ } |
| 1718 | \\ { |
| 1719 | \\ p0 = dfclass(r1:0,#0x08) |
| 1720 | \\ if (p0.new) jumpr:nt r31 |
| 1721 | \\ } |
| 1722 | \\ |
| 1723 | \\ |
| 1724 | \\ { |
| 1725 | \\ r1:0 = extractu(r1:0,#52,#0) |
| 1726 | \\ } |
| 1727 | \\ { |
| 1728 | \\ r28 = add(clb(r1:0),#-11) |
| 1729 | \\ } |
| 1730 | \\ { |
| 1731 | \\ r1:0 = asl(r1:0,r28) |
| 1732 | \\ r28 = sub(#1,r28) |
| 1733 | \\ } |
| 1734 | \\ { |
| 1735 | \\ r1 = insert(r28,#1,#52 -32) |
| 1736 | \\ } |
| 1737 | \\ { |
| 1738 | \\ r3:2 = extractu(r1:0,#23 +1,#52 -23) |
| 1739 | \\ r5 = ##0x3f000004 |
| 1740 | \\ } |
| 1741 | \\ { |
| 1742 | \\ r9 = or(r5,r2) |
| 1743 | \\ r5 = and(r5,#-16) |
| 1744 | \\ jump .Ldenormal_restart |
| 1745 | \\ } |
| 1746 | \\ .Lsqrt_nan: |
| 1747 | \\ { |
| 1748 | \\ r28 = convert_df2sf(r1:0) |
| 1749 | \\ r1:0 = #-1 |
| 1750 | \\ jumpr r31 |
| 1751 | \\ } |
| 1752 | \\ .Lsqrt_invalid_neg: |
| 1753 | \\ { |
| 1754 | \\ r1:0 = convert_sf2df(r28) |
| 1755 | \\ jumpr r31 |
| 1756 | \\ } |
| 1757 | ); |
| 1758 | } |
| 1759 | |
| 1760 | comptime { |
| 1761 | if (builtin.cpu.arch == .hexagon) { |
| 1762 | @export(__hexagon_adddf3, .{ .name = "__hexagon_adddf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1763 | @export(__hexagon_adddf3, .{ .name = "__hexagon_fast_adddf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1764 | @export(__hexagon_subdf3, .{ .name = "__hexagon_subdf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1765 | @export(__hexagon_subdf3, .{ .name = "__hexagon_fast_subdf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1766 | @export(__hexagon_divdf3, .{ .name = "__hexagon_divdf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1767 | @export(__hexagon_divdf3, .{ .name = "__hexagon_fast_divdf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1768 | @export(__hexagon_muldf3, .{ .name = "__hexagon_muldf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1769 | @export(__hexagon_muldf3, .{ .name = "__hexagon_fast_muldf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1770 | @export(__hexagon_sqrtdf2, .{ .name = "__hexagon_sqrtdf2", .linkage = common.linkage, .visibility = common.visibility }); |
| 1771 | @export(__hexagon_sqrtdf2, .{ .name = "__hexagon_fast2_sqrtdf2", .linkage = common.linkage, .visibility = common.visibility }); |
| 1772 | @export(__hexagon_sqrtdf2, .{ .name = "__hexagon_sqrt", .linkage = common.linkage, .visibility = common.visibility }); |
| 1773 | @export(__hexagon_divsf3, .{ .name = "__hexagon_divsf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1774 | @export(__hexagon_divsf3, .{ .name = "__hexagon_fast_divsf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1775 | @export(__hexagon_divsi3, .{ .name = "__hexagon_divsi3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1776 | @export(__hexagon_umodsi3, .{ .name = "__hexagon_umodsi3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1777 | @export(__hexagon_sqrtf, .{ .name = "__hexagon_sqrtf", .linkage = common.linkage, .visibility = common.visibility }); |
| 1778 | @export(__hexagon_sqrtf, .{ .name = "__hexagon_fast2_sqrtf", .linkage = common.linkage, .visibility = common.visibility }); |
| 1779 | @export(__hexagon_moddi3, .{ .name = "__hexagon_moddi3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1780 | @export(__hexagon_divdi3, .{ .name = "__hexagon_divdi3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1781 | @export(__hexagon_divsf3, .{ .name = "__hexagon_divsf3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1782 | @export(__hexagon_udivdi3, .{ .name = "__hexagon_udivdi3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1783 | @export(__hexagon_umoddi3, .{ .name = "__hexagon_umoddi3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1784 | @export(__hexagon_modsi3, .{ .name = "__hexagon_modsi3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1785 | @export(__hexagon_memcpy_likely_aligned_min32bytes_mult8bytes, .{ .name = "__hexagon_memcpy_likely_aligned_min32bytes_mult8bytes", .linkage = common.linkage, .visibility = common.visibility }); |
| 1786 | @export(__hexagon_udivsi3, .{ .name = "__hexagon_udivsi3", .linkage = common.linkage, .visibility = common.visibility }); |
| 1787 | } |
| 1788 | } |
Brian Cain <brian.cain@oss.qualcomm.com> 2024-11-23 19:17:21-06:00
GitHub <noreply@github.com> 2024-11-24 01:17:21+00:00
Signed by PGP key B5690EEEBB952194
lib