| 1 |
;; Copyright (C) 2019-2025 Free Software Foundation, Inc. |
| 2 |
;; |
| 3 |
;; This file is part of LIBF7, which is part of GCC. |
| 4 |
;; |
| 5 |
;; GCC is free software; you can redistribute it and/or modify it under |
| 6 |
;; the terms of the GNU General Public License as published by the Free |
| 7 |
;; Software Foundation; either version 3, or (at your option) any later |
| 8 |
;; version. |
| 9 |
;; |
| 10 |
;; GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
| 11 |
;; WARRANTY; without even the implied warranty of MERCHANTABILITY or |
| 12 |
;; FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| 13 |
;; for more details. |
| 14 |
;; |
| 15 |
;; Under Section 7 of GPL version 3, you are granted additional |
| 16 |
;; permissions described in the GCC Runtime Library Exception, version |
| 17 |
;; 3.1, as published by the Free Software Foundation. |
| 18 |
;; |
| 19 |
;; You should have received a copy of the GNU General Public License and |
| 20 |
;; a copy of the GCC Runtime Library Exception along with this program; |
| 21 |
;; see the files COPYING3 and COPYING.RUNTIME respectively. If not, see |
| 22 |
;; <http://www.gnu.org/licenses/>. */ |
| 23 |
|
| 24 |
#ifndef __AVR_TINY__ |
| 25 |
|
| 26 |
#define ASM_DEFS_HAVE_DEFUN |
| 27 |
|
| 28 |
#include "asm-defs.h" |
| 29 |
#include "libf7.h" |
| 30 |
|
| 31 |
#define ZERO __zero_reg__ |
| 32 |
#define TMP __tmp_reg__ |
| 33 |
|
| 34 |
#define F7(name) F7_(name##_asm) |
| 35 |
|
| 36 |
.macro F7call name |
| 37 |
.global F7(\name\()) |
| 38 |
XCALL F7(\name\()) |
| 39 |
.endm |
| 40 |
|
| 41 |
.macro F7jmp name |
| 42 |
.global F7(\name\()) |
| 43 |
XJMP F7(\name\()) |
| 44 |
.endm |
| 45 |
|
| 46 |
;; Just for visibility in disassembly. |
| 47 |
.macro LLL name |
| 48 |
.global LLL.\name |
| 49 |
LLL.\name: |
| 50 |
nop |
| 51 |
.endm |
| 52 |
|
| 53 |
.macro DEFUN name |
| 54 |
.section .text.libf7.asm.\name, "ax", @progbits |
| 55 |
.global F7(\name\()) |
| 56 |
.func F7(\name\()) |
| 57 |
F7(\name\()) : |
| 58 |
.endm |
| 59 |
|
| 60 |
.macro ENDF name |
| 61 |
.size F7(\name\()), . - F7(\name\()) |
| 62 |
.endfunc |
| 63 |
.endm |
| 64 |
|
| 65 |
.macro LABEL name |
| 66 |
.global F7(\name\()) |
| 67 |
F7(\name\()) : |
| 68 |
.endm |
| 69 |
|
| 70 |
.macro _DEFUN name |
| 71 |
.section .text.libf7.asm.\name, "ax", @progbits |
| 72 |
.weak \name |
| 73 |
.type \name, @function |
| 74 |
\name : |
| 75 |
.endm |
| 76 |
|
| 77 |
.macro _ENDF name |
| 78 |
.size \name, . - \name |
| 79 |
.endm |
| 80 |
|
| 81 |
.macro _LABEL name |
| 82 |
.weak \name |
| 83 |
.type \name, @function |
| 84 |
\name : |
| 85 |
.endm |
| 86 |
|
| 87 |
#define F7_NAME(X) F7_(X) |
| 88 |
|
| 89 |
;; Make a weak alias. |
| 90 |
.macro ALIAS sym |
| 91 |
.weak \sym |
| 92 |
.type \sym, @function |
| 93 |
\sym: |
| 94 |
.endm |
| 95 |
|
| 96 |
;; Make a weak alias if double is 64 bits wide. |
| 97 |
.macro DALIAS sym |
| 98 |
#if defined (WITH_LIBF7_MATH_SYMBOLS) && __SIZEOF_DOUBLE__ == 8 |
| 99 |
ALIAS \sym |
| 100 |
#endif |
| 101 |
.endm |
| 102 |
|
| 103 |
;; Make a weak alias if long double is 64 bits wide. |
| 104 |
.macro LALIAS sym |
| 105 |
#if defined (WITH_LIBF7_MATH_SYMBOLS) && __SIZEOF_LONG_DOUBLE__ == 8 |
| 106 |
ALIAS \sym |
| 107 |
#endif |
| 108 |
.endm |
| 109 |
|
| 110 |
#define Off 1 |
| 111 |
#define Expo (Off + F7_MANT_BYTES) |
| 112 |
|
| 113 |
#ifdef F7MOD_classify_ |
| 114 |
;; r24 = classify (*Z) |
| 115 |
;; NaN -> F7_FLAG_nan |
| 116 |
;; INF -> F7_FLAG_inf [ | F7_FLAG_sign ] |
| 117 |
;; ==0 -> F7_FLAG_zero |
| 118 |
;; ... -> 0 [ | F7_FLAG_sign ] |
| 119 |
|
| 120 |
;; Clobbers: None (no TMP, no T). |
| 121 |
DEFUN classify |
| 122 |
|
| 123 |
ld r24, Z |
| 124 |
lsr r24 |
| 125 |
brne .Lnan_or_inf |
| 126 |
|
| 127 |
ldd r24, Z+6+Off |
| 128 |
tst r24 |
| 129 |
brpl 0f |
| 130 |
sbc r24, r24 |
| 131 |
andi r24, F7_FLAG_sign |
| 132 |
ret |
| 133 |
|
| 134 |
0: ldi r24, F7_FLAG_zero |
| 135 |
ret |
| 136 |
|
| 137 |
.Lnan_or_inf: |
| 138 |
rol r24 |
| 139 |
ret |
| 140 |
|
| 141 |
ENDF classify |
| 142 |
#endif /* F7MOD_classify_ */ |
| 143 |
|
| 144 |
#ifdef F7MOD_clr_ |
| 145 |
DEFUN clr |
| 146 |
std Z+0, ZERO |
| 147 |
std Z+0+Off, ZERO |
| 148 |
std Z+1+Off, ZERO |
| 149 |
std Z+2+Off, ZERO |
| 150 |
std Z+3+Off, ZERO |
| 151 |
std Z+4+Off, ZERO |
| 152 |
std Z+5+Off, ZERO |
| 153 |
std Z+6+Off, ZERO |
| 154 |
std Z+0+Expo, ZERO |
| 155 |
std Z+1+Expo, ZERO |
| 156 |
ret |
| 157 |
ENDF clr |
| 158 |
|
| 159 |
#endif /* F7MOD_clr_ */ |
| 160 |
|
| 161 |
#ifdef F7MOD_clz_ |
| 162 |
;; The libcc CLZ implementations like __clzsi2 aka. __builtin_clzl are |
| 163 |
;; not very well suited for out purpose, so implement our own. |
| 164 |
|
| 165 |
#define ZBITS r26 |
| 166 |
.macro .test.byte reg |
| 167 |
or ZERO, \reg |
| 168 |
brne .Loop_bit |
| 169 |
subi ZBITS, -8 |
| 170 |
.endm |
| 171 |
|
| 172 |
;; R26 = CLZ (uint64_t R18); CLZ (0) = 64. |
| 173 |
;; Unchanged: T |
| 174 |
DEFUN clzdi2 |
| 175 |
clr ZBITS |
| 176 |
;; Catch the common case of normalized .mant for speed-up. |
| 177 |
tst r25 |
| 178 |
brmi 9f |
| 179 |
.test.byte r25 |
| 180 |
.test.byte r24 |
| 181 |
.test.byte r23 |
| 182 |
.test.byte r22 |
| 183 |
.test.byte r21 |
| 184 |
.test.byte r20 |
| 185 |
.test.byte r19 |
| 186 |
.test.byte r18 |
| 187 |
.Ldone: |
| 188 |
clr ZERO |
| 189 |
9: ret |
| 190 |
|
| 191 |
.Loop_bit: |
| 192 |
lsl ZERO |
| 193 |
brcs .Ldone |
| 194 |
inc ZBITS |
| 195 |
rjmp .Loop_bit |
| 196 |
|
| 197 |
ENDF clzdi2 |
| 198 |
#undef ZBITS |
| 199 |
#endif /* F7MOD_clz_ */ |
| 200 |
|
| 201 |
#ifdef F7MOD_cmp_mant_ |
| 202 |
DEFUN cmp_mant |
| 203 |
|
| 204 |
adiw X, 6 + Off |
| 205 |
ld r24, X $ ldd TMP, Z+6+Off $ SUB r24, TMP |
| 206 |
brne .Lunequal |
| 207 |
|
| 208 |
sbiw X, 6 |
| 209 |
ld r24, X+ $ ldd TMP, Z+0+Off $ SUB r24, TMP |
| 210 |
ld r24, X+ $ ldd TMP, Z+1+Off $ sbc r24, TMP |
| 211 |
ld r24, X+ $ ldd TMP, Z+2+Off $ sbc r24, TMP |
| 212 |
ld r24, X+ $ ldd TMP, Z+3+Off $ sbc r24, TMP |
| 213 |
ld r24, X+ $ ldd TMP, Z+4+Off $ sbc r24, TMP |
| 214 |
ld r24, X+ $ ldd TMP, Z+5+Off $ sbc r24, TMP |
| 215 |
;; MSBs are already known to be equal |
| 216 |
breq 9f |
| 217 |
.Lunequal: |
| 218 |
sbc r24, r24 |
| 219 |
sbci r24, -1 |
| 220 |
9: sbiw X, 6 + Off |
| 221 |
ret |
| 222 |
ENDF cmp_mant |
| 223 |
#endif /* F7MOD_cmp_mant_ */ |
| 224 |
|
| 225 |
#define CA 18 |
| 226 |
#define C0 CA+1 |
| 227 |
#define C1 C0+1 |
| 228 |
#define C2 C0+2 |
| 229 |
#define C3 C0+3 |
| 230 |
#define C4 C0+4 |
| 231 |
#define C5 C0+5 |
| 232 |
#define C6 C0+6 |
| 233 |
#define Carry r16 |
| 234 |
#define Flags 18 |
| 235 |
|
| 236 |
#ifdef F7MOD_store_ |
| 237 |
;; Z->flags = CA. |
| 238 |
;; Z->mant = C[7]. |
| 239 |
DEFUN store_mant.with_flags |
| 240 |
st Z, CA |
| 241 |
|
| 242 |
;; Z->mant = C[7]. |
| 243 |
LABEL store_mant |
| 244 |
std Z+0+Off, C0 |
| 245 |
std Z+1+Off, C1 |
| 246 |
std Z+2+Off, C2 |
| 247 |
std Z+3+Off, C3 |
| 248 |
std Z+4+Off, C4 |
| 249 |
std Z+5+Off, C5 |
| 250 |
std Z+6+Off, C6 |
| 251 |
ret |
| 252 |
ENDF store_mant.with_flags |
| 253 |
#endif /* F7MOD_store_ */ |
| 254 |
|
| 255 |
#ifdef F7MOD_load_ |
| 256 |
;; CA = Z->flags |
| 257 |
;; C[7] = Z->mant |
| 258 |
DEFUN load_mant.with_flags |
| 259 |
ld CA, Z |
| 260 |
skipnext |
| 261 |
|
| 262 |
;; CA = 0 |
| 263 |
;; C[7] = Z->mant |
| 264 |
LABEL load_mant.clr_CA |
| 265 |
LABEL load_mant.clr_flags |
| 266 |
clr CA ; May be skipped |
| 267 |
|
| 268 |
;; C[7] = Z->mant |
| 269 |
LABEL load_mant |
| 270 |
ldd C0, Z+0+Off |
| 271 |
ldd C1, Z+1+Off |
| 272 |
ldd C2, Z+2+Off |
| 273 |
ldd C3, Z+3+Off |
| 274 |
ldd C4, Z+4+Off |
| 275 |
ldd C5, Z+5+Off |
| 276 |
ldd C6, Z+6+Off |
| 277 |
ret |
| 278 |
ENDF load_mant.with_flags |
| 279 |
#endif /* F7MOD_load_ */ |
| 280 |
|
| 281 |
#ifdef F7MOD_copy_ |
| 282 |
DEFUN copy |
| 283 |
cp XL, ZL |
| 284 |
cpc XH, ZH |
| 285 |
breq 9f |
| 286 |
adiw XL, F7_SIZEOF |
| 287 |
adiw ZL, F7_SIZEOF |
| 288 |
set |
| 289 |
bld ZERO, 1 |
| 290 |
bld ZERO, 3 ; ZERO = 0b1010 = 10. |
| 291 |
.Loop: |
| 292 |
ld TMP, -X |
| 293 |
st -Z, TMP |
| 294 |
dec ZERO |
| 295 |
brne .Loop |
| 296 |
9: ret |
| 297 |
ENDF copy |
| 298 |
#endif /* F7MOD_copy_ */ |
| 299 |
|
| 300 |
#ifdef F7MOD_copy_P_ |
| 301 |
DEFUN copy_P |
| 302 |
set |
| 303 |
bld ZERO, 1 |
| 304 |
bld ZERO, 3 ; ZERO = 0b1010 = 10. |
| 305 |
.Loop: |
| 306 |
#ifdef __AVR_HAVE_LPMX__ |
| 307 |
lpm TMP, Z+ |
| 308 |
#else |
| 309 |
lpm |
| 310 |
adiw Z, 1 |
| 311 |
#endif /* Have LPMx */ |
| 312 |
st X+, TMP |
| 313 |
dec ZERO |
| 314 |
brne .Loop |
| 315 |
sbiw X, F7_SIZEOF |
| 316 |
sbiw Z, F7_SIZEOF |
| 317 |
ret |
| 318 |
ENDF copy_P |
| 319 |
#endif /* F7MOD_copy_P_ */ |
| 320 |
|
| 321 |
#ifdef F7MOD_copy_mant_ |
| 322 |
DEFUN copy_mant |
| 323 |
cp XL, ZL |
| 324 |
cpc XH, ZH |
| 325 |
breq 9f |
| 326 |
adiw XL, 1 |
| 327 |
adiw ZL, 1 |
| 328 |
set |
| 329 |
bld ZERO, 3 |
| 330 |
dec ZERO ; ZERO = 7 |
| 331 |
.Loop: |
| 332 |
ld TMP, X+ |
| 333 |
st Z+, TMP |
| 334 |
dec ZERO |
| 335 |
brne .Loop |
| 336 |
sbiw XL, 8 |
| 337 |
sbiw ZL, 8 |
| 338 |
9: ret |
| 339 |
ENDF copy_mant |
| 340 |
#endif /* F7MOD_copy_mant_ */ |
| 341 |
|
| 342 |
|
| 343 |
#ifdef F7MOD_clr_mant_lsbs_ |
| 344 |
DEFUN clr_mant_lsbs |
| 345 |
push r16 |
| 346 |
mov r16, r20 |
| 347 |
wmov XL, r24 |
| 348 |
|
| 349 |
wmov ZL, r22 |
| 350 |
F7call load_mant |
| 351 |
|
| 352 |
F7call lshrdi3 |
| 353 |
|
| 354 |
clr CA |
| 355 |
|
| 356 |
F7call ashldi3 |
| 357 |
|
| 358 |
pop r16 |
| 359 |
|
| 360 |
wmov ZL, XL |
| 361 |
F7jmp store_mant |
| 362 |
|
| 363 |
ENDF clr_mant_lsbs |
| 364 |
#endif /* F7MOD_clr_mant_lsbs_ */ |
| 365 |
|
| 366 |
|
| 367 |
#ifdef F7MOD_normalize_with_carry_ |
| 368 |
;; Z = &f7_t |
| 369 |
;; C[] = .mant may be not normalized |
| 370 |
;; Carry === r16 = Addend to Z->expo in [-64, 128). |
| 371 |
;; Normalize C[], set Flags, and adjust Z->expo. |
| 372 |
;; Return CA (after normalization) in TMP. |
| 373 |
;; Unchanged: T |
| 374 |
#define Addend r17 |
| 375 |
#define Zbits r26 |
| 376 |
#define expL r26 |
| 377 |
#define expH r27 |
| 378 |
DEFUN normalize_with_carry |
| 379 |
mov Addend, Carry |
| 380 |
tst C6 |
| 381 |
brmi .Lshift.0 |
| 382 |
;; r26 = CLZ (uint64_t R18) |
| 383 |
F7call clzdi2 |
| 384 |
cpi Zbits, 64 |
| 385 |
breq .Lclr |
| 386 |
sub Addend, Zbits |
| 387 |
mov r16, Zbits |
| 388 |
|
| 389 |
F7call ashldi3 |
| 390 |
;; Assert (R25.7 == 1) |
| 391 |
.Lshift.0: |
| 392 |
mov TMP, CA |
| 393 |
ld Flags, Z |
| 394 |
|
| 395 |
;; .expo += Addend |
| 396 |
ldd expL, Z+0+Expo |
| 397 |
ldd expH, Z+1+Expo |
| 398 |
;; Sign-extend Addend |
| 399 |
clr r16 |
| 400 |
sbrc Addend, 7 |
| 401 |
com r16 |
| 402 |
|
| 403 |
;; exp += (int8_t) Addend, i.e. sign-extend Addend. |
| 404 |
add expL, Addend |
| 405 |
adc expH, r16 |
| 406 |
brvc .Lnormal |
| 407 |
tst r16 |
| 408 |
brmi .Lclr |
| 409 |
;; Overflow |
| 410 |
#if F7_HAVE_Inf == 1 |
| 411 |
ori Flags, F7_FLAG_inf |
| 412 |
#else |
| 413 |
ldi Flags, F7_FLAG_nan |
| 414 |
#endif /* Have Inf */ |
| 415 |
ret |
| 416 |
|
| 417 |
.Lnormal: |
| 418 |
std Z+0+Expo, expL |
| 419 |
std Z+1+Expo, expH |
| 420 |
ret |
| 421 |
|
| 422 |
.Lclr: |
| 423 |
;; Underflow or Zero. |
| 424 |
clr TMP |
| 425 |
.global __clr_8 |
| 426 |
XJMP __clr_8 |
| 427 |
|
| 428 |
LABEL normalize.store_with_flags |
| 429 |
;; no rounding |
| 430 |
set |
| 431 |
skipnext |
| 432 |
LABEL normalize.round.store_with_flags |
| 433 |
;; with rounding |
| 434 |
clt ; skipped ? |
| 435 |
LABEL normalize.maybe_round.store_with_flags |
| 436 |
F7call normalize_with_carry |
| 437 |
;; We have: |
| 438 |
;; Z = &f7_t |
| 439 |
;; X = .expo |
| 440 |
;; C[] = .mant |
| 441 |
;; R18 = .flags |
| 442 |
;; TMP = byte below .mant after normalization |
| 443 |
;; T = 1 => no rounding. |
| 444 |
brts .Lstore |
| 445 |
lsl TMP |
| 446 |
adc C0, ZERO |
| 447 |
brcc .Lstore |
| 448 |
adc C1, ZERO |
| 449 |
adc C2, ZERO |
| 450 |
adc C3, ZERO |
| 451 |
adc C4, ZERO |
| 452 |
adc C5, ZERO |
| 453 |
adc C6, ZERO |
| 454 |
brcc .Lstore |
| 455 |
;; We only come here if C6 overflowed, i.e. C[] is 0 now. |
| 456 |
;; .mant = 1.0 by restoring the MSbit. |
| 457 |
ror C6 |
| 458 |
;; .expo += 1 and override the .expo stored during normalize. |
| 459 |
adiw expL, 1 |
| 460 |
std Z+0+Expo, expL |
| 461 |
std Z+1+Expo, expH |
| 462 |
|
| 463 |
.Lstore: |
| 464 |
F7call store_mant.with_flags |
| 465 |
|
| 466 |
;; Return the byte below .mant after normalization. |
| 467 |
;; This is only useful without rounding; the caller will know. |
| 468 |
mov R24, TMP |
| 469 |
ret |
| 470 |
ENDF normalize_with_carry |
| 471 |
#endif /* F7MOD_normalize_with_carry_ */ |
| 472 |
|
| 473 |
|
| 474 |
#ifdef F7MOD_normalize_ |
| 475 |
;; Using above functionality from C. |
| 476 |
;; f7_t* normalize (f7_t *cc) |
| 477 |
;; Adjusts cc->expo |
| 478 |
;; Clears cc->flags |
| 479 |
DEFUN normalize |
| 480 |
push r17 |
| 481 |
push r16 |
| 482 |
wmov ZL, r24 |
| 483 |
F7call load_mant.clr_CA |
| 484 |
clr Carry |
| 485 |
st Z, ZERO |
| 486 |
F7call normalize.store_with_flags |
| 487 |
wmov r24, Z |
| 488 |
pop r16 |
| 489 |
pop r17 |
| 490 |
ret |
| 491 |
ENDF normalize |
| 492 |
#endif /* F7MOD_normalize_ */ |
| 493 |
|
| 494 |
|
| 495 |
#ifdef F7MOD_store_expo_ |
| 496 |
#define Done r24 |
| 497 |
#define expLO r24 |
| 498 |
#define expHI r25 |
| 499 |
;; expo == INT16_MAX => *Z = Inf, return Done = true. |
| 500 |
;; expo == INT16_MIN => *Z = 0x0, return Done = true. |
| 501 |
;; else => Z->expo = expo, return Done = false. |
| 502 |
DEFUN store_expo |
| 503 |
cpi expHI, 0x80 |
| 504 |
cpc expLO, ZERO |
| 505 |
breq .Ltiny |
| 506 |
adiw expLO, 1 |
| 507 |
brvs .Lhuge |
| 508 |
sbiw expLO, 1 |
| 509 |
std Z+0+Expo, expLO |
| 510 |
std Z+1+Expo, expHI |
| 511 |
ldi Done, 0 |
| 512 |
ret |
| 513 |
|
| 514 |
.Lhuge: |
| 515 |
#if F7_HAVE_Inf == 1 |
| 516 |
ld Done, Z |
| 517 |
andi Done, F7_FLAG_sign |
| 518 |
ori Done, F7_FLAG_inf |
| 519 |
#else |
| 520 |
ldi Done, F7_FLAG_nan |
| 521 |
#endif /* Have Inf */ |
| 522 |
st Z, Done |
| 523 |
ldi Done, 1 |
| 524 |
ret |
| 525 |
|
| 526 |
.Ltiny: |
| 527 |
ldi Done, 1 |
| 528 |
F7jmp clr |
| 529 |
ENDF store_expo |
| 530 |
#endif /* F7MOD_store_expo_ */ |
| 531 |
|
| 532 |
|
| 533 |
#ifdef F7MOD_set_u64_ |
| 534 |
DEFUN set_s64 |
| 535 |
set |
| 536 |
skipnext |
| 537 |
;; ... |
| 538 |
LABEL set_u64 |
| 539 |
clt ; Skipped? |
| 540 |
wmov Zl, r16 |
| 541 |
;; TMP holds .flags. |
| 542 |
clr TMP |
| 543 |
brtc .Lnot.negative |
| 544 |
|
| 545 |
bst C6, 7 |
| 546 |
brtc .Lnot.negative |
| 547 |
bld TMP, F7_FLAGNO_sign |
| 548 |
.global __negdi2 |
| 549 |
XCALL __negdi2 |
| 550 |
|
| 551 |
.Lnot.negative: |
| 552 |
st Z, TMP |
| 553 |
std Z+0+Expo, ZERO |
| 554 |
std Z+1+Expo, ZERO |
| 555 |
ldi Carry, 63 |
| 556 |
F7call normalize.round.store_with_flags |
| 557 |
wmov r24, Z |
| 558 |
wmov r16, Z ; Unclobber r16. |
| 559 |
ret |
| 560 |
ENDF set_s64 |
| 561 |
#endif /* F7MOD_set_u64_ */ |
| 562 |
|
| 563 |
|
| 564 |
#ifdef F7MOD_to_integer_ |
| 565 |
#define Mask r26 |
| 566 |
DEFUN to_integer |
| 567 |
wmov ZL, r24 |
| 568 |
mov Mask, r22 |
| 569 |
|
| 570 |
F7call load_mant.with_flags |
| 571 |
|
| 572 |
sbrc Flags, F7_FLAGNO_nan |
| 573 |
rjmp .Lset_0x8000 |
| 574 |
|
| 575 |
sbrc Flags, F7_FLAGNO_inf |
| 576 |
rjmp .Lsaturate |
| 577 |
|
| 578 |
sbrs C6, 7 |
| 579 |
rjmp .Lset_0x0000 |
| 580 |
|
| 581 |
bst Flags, F7_FLAGNO_sign |
| 582 |
ldd r27, Z+0+Expo |
| 583 |
;; Does .expo have bits outside Mask? ... |
| 584 |
mov TMP, Mask |
| 585 |
com TMP |
| 586 |
and TMP, r27 |
| 587 |
ldd r27, Z+1+Expo |
| 588 |
tst r27 |
| 589 |
brmi .Lset_0x0000 ; ...yes: .expo is < 0 => return 0 |
| 590 |
or TMP, r27 |
| 591 |
brne .Lsaturate.T ; ...yes: .expo > Mask => saturate |
| 592 |
|
| 593 |
;; ...no: Shift right to meet .expo = 0. |
| 594 |
PUSH r16 |
| 595 |
ldd r16, Z+0+Expo |
| 596 |
eor r16, Mask |
| 597 |
and r16, Mask |
| 598 |
clr CA |
| 599 |
F7call lshrdi3 |
| 600 |
POP r16 |
| 601 |
tst C6 |
| 602 |
brmi .Lsaturate.T ; > INTxx_MAX => saturate |
| 603 |
|
| 604 |
brtc 9f ; >= 0 => return |
| 605 |
sbrc Mask, 5 |
| 606 |
.global __negdi2 |
| 607 |
XJMP __negdi2 |
| 608 |
sbrc Mask, 4 |
| 609 |
.global __negsi2 |
| 610 |
XJMP __negsi2 |
| 611 |
neg C6 |
| 612 |
neg C5 |
| 613 |
sbci C6, 0 |
| 614 |
9: ret |
| 615 |
|
| 616 |
.Lsaturate: |
| 617 |
bst Flags, F7_FLAGNO_sign |
| 618 |
.Lsaturate.T: |
| 619 |
|
| 620 |
#if F7_HAVE_Inf |
| 621 |
brts .Lset_0x8000 |
| 622 |
|
| 623 |
.Lset_0x7fff: |
| 624 |
;; +Inf => return INTxx_MAX |
| 625 |
sec |
| 626 |
.global __sbc_8 |
| 627 |
XCALL __sbc_8 |
| 628 |
ldi C6, 0x7f |
| 629 |
ret |
| 630 |
#endif /* F7_HAVE_Inf */ |
| 631 |
|
| 632 |
.Lset_0x8000: |
| 633 |
;; NaN or -Inf => return INTxx_MIN |
| 634 |
.global __clr_8 |
| 635 |
XCALL __clr_8 |
| 636 |
ldi C6, 0x80 |
| 637 |
ret |
| 638 |
|
| 639 |
.Lset_0x0000: |
| 640 |
;; Small value => return 0x0 |
| 641 |
.global __clr_8 |
| 642 |
XJMP __clr_8 |
| 643 |
|
| 644 |
ENDF to_integer |
| 645 |
#endif /* F7MOD_to_integer_ */ |
| 646 |
|
| 647 |
|
| 648 |
#ifdef F7MOD_to_unsigned_ |
| 649 |
#define Mask r26 |
| 650 |
DEFUN to_unsigned |
| 651 |
wmov ZL, r24 |
| 652 |
mov Mask, r22 |
| 653 |
|
| 654 |
F7call load_mant.with_flags |
| 655 |
|
| 656 |
sbrc Flags, F7_FLAGNO_nan |
| 657 |
rjmp .Lset_0xffff |
| 658 |
|
| 659 |
sbrc Flags, F7_FLAGNO_sign |
| 660 |
rjmp .Lset_0x0000 |
| 661 |
|
| 662 |
sbrc Flags, F7_FLAGNO_inf |
| 663 |
rjmp .Lset_0xffff |
| 664 |
|
| 665 |
sbrs C6, 7 |
| 666 |
rjmp .Lset_0x0000 |
| 667 |
|
| 668 |
ldd r27, Z+0+Expo |
| 669 |
;; Does .expo have bits outside Mask? ... |
| 670 |
mov TMP, Mask |
| 671 |
com TMP |
| 672 |
and TMP, r27 |
| 673 |
ldd r27, Z+1+Expo |
| 674 |
tst r27 |
| 675 |
brmi .Lset_0x0000 ; ...yes: .expo is < 0 => return 0 |
| 676 |
or TMP, r27 |
| 677 |
brne .Lset_0xffff ; ...yes: .expo > Mask => saturate |
| 678 |
|
| 679 |
;; ...no: Shift right to meet .expo = 0. |
| 680 |
PUSH r16 |
| 681 |
ldd r16, Z+0+Expo |
| 682 |
eor r16, Mask |
| 683 |
and r16, Mask |
| 684 |
clr CA |
| 685 |
F7call lshrdi3 |
| 686 |
POP r16 |
| 687 |
ret |
| 688 |
|
| 689 |
.Lset_0xffff: |
| 690 |
;; return UINTxx_MAX |
| 691 |
sec |
| 692 |
.global __sbc_8 |
| 693 |
XJMP __sbc_8 |
| 694 |
|
| 695 |
.Lset_0x0000: |
| 696 |
;; Small value => return 0x0 |
| 697 |
.global __clr_8 |
| 698 |
XJMP __clr_8 |
| 699 |
|
| 700 |
ENDF to_unsigned |
| 701 |
#endif /* F7MOD_to_unsigned_ */ |
| 702 |
|
| 703 |
|
| 704 |
#ifdef F7MOD_addsub_mant_scaled_ |
| 705 |
;; int8_t f7_addsub_mant_scaled_asm (f7_t *r24, const f7_t *r22, const f7_t 20*, |
| 706 |
;; uint8_t r18); |
| 707 |
;; R18.0 = 1 : ADD |
| 708 |
;; R18.0 = 0 : SUB |
| 709 |
;; R18[7..1] : Scale |
| 710 |
;; Compute *R24 = *R22 + *R20 >> R18[7..1]. |
| 711 |
|
| 712 |
#define BA 10 |
| 713 |
#define B0 BA+1 |
| 714 |
#define B1 B0+1 |
| 715 |
#define B2 B0+2 |
| 716 |
#define B3 B0+3 |
| 717 |
#define B4 B0+4 |
| 718 |
#define B5 B0+5 |
| 719 |
#define B6 B0+6 |
| 720 |
|
| 721 |
DEFUN addsub_mant_scaled |
| 722 |
do_prologue_saves 10 |
| 723 |
|
| 724 |
bst r18, 0 ;; ADD ? |
| 725 |
lsr r18 |
| 726 |
mov r16, r18 |
| 727 |
|
| 728 |
wmov ZL, r20 |
| 729 |
wmov YL, r22 |
| 730 |
;; C[] = bb >> shift |
| 731 |
wmov XL, r24 |
| 732 |
|
| 733 |
F7call load_mant.clr_CA |
| 734 |
F7call lshrdi3 |
| 735 |
|
| 736 |
wmov BA, CA |
| 737 |
wmov B1, C1 |
| 738 |
wmov B3, C3 |
| 739 |
wmov B5, C5 |
| 740 |
wmov ZL, YL |
| 741 |
F7call load_mant.clr_CA |
| 742 |
|
| 743 |
wmov ZL, XL |
| 744 |
|
| 745 |
brts .Ladd |
| 746 |
|
| 747 |
.global __subdi3 |
| 748 |
XCALL __subdi3 |
| 749 |
|
| 750 |
breq .Lzero |
| 751 |
brcc .Lround |
| 752 |
;; C = 1: Can underflow happen at all ? |
| 753 |
.Lzero: |
| 754 |
F7call clr |
| 755 |
rjmp .Lepilogue |
| 756 |
|
| 757 |
.Ladd: |
| 758 |
.global __adddi3 |
| 759 |
XCALL __adddi3 |
| 760 |
brcc .Lround |
| 761 |
ldi Carry, 1 |
| 762 |
.global __lshrdi3 |
| 763 |
XCALL __lshrdi3 |
| 764 |
ori C6, 1 << 7 |
| 765 |
skipnext |
| 766 |
.Lround: |
| 767 |
clr Carry ; skipped? |
| 768 |
F7call normalize.round.store_with_flags |
| 769 |
|
| 770 |
.Lepilogue: |
| 771 |
do_epilogue_restores 10 |
| 772 |
|
| 773 |
ENDF addsub_mant_scaled |
| 774 |
|
| 775 |
#if !defined (__AVR_HAVE_MOVW__) || !defined (__AVR_HAVE_JMP_CALL__) |
| 776 |
DEFUN lshrdi3 |
| 777 |
.global __lshrdi3 |
| 778 |
XJMP __lshrdi3 |
| 779 |
ENDF lshrdi3 |
| 780 |
DEFUN ashldi3 |
| 781 |
.global __ashldi3 |
| 782 |
XJMP __ashldi3 |
| 783 |
ENDF ashldi3 |
| 784 |
#else |
| 785 |
|
| 786 |
# Basically just a wrapper around libgcc's __lshrdi3. |
| 787 |
DEFUN lshrdi3 |
| 788 |
;; Handle bit 5 of shift offset. |
| 789 |
sbrs r16, 5 |
| 790 |
rjmp 4f |
| 791 |
wmov CA, C3 |
| 792 |
wmov C1, C5 |
| 793 |
clr C6 $ clr C5 $ wmov C3, C5 |
| 794 |
4: |
| 795 |
;; Handle bit 4 of shift offset. |
| 796 |
sbrs r16, 4 |
| 797 |
rjmp 3f |
| 798 |
wmov CA, C1 |
| 799 |
wmov C1, C3 |
| 800 |
wmov C3, C5 |
| 801 |
clr C6 $ clr C5 |
| 802 |
3: |
| 803 |
;; Handle bits 3...0 of shift offset. |
| 804 |
push r16 |
| 805 |
andi r16, 0xf |
| 806 |
breq 0f |
| 807 |
|
| 808 |
.global __lshrdi3 |
| 809 |
XCALL __lshrdi3 |
| 810 |
0: |
| 811 |
pop r16 |
| 812 |
ret |
| 813 |
ENDF lshrdi3 |
| 814 |
|
| 815 |
# Basically just a wrapper around libgcc's __ashldi3. |
| 816 |
DEFUN ashldi3 |
| 817 |
;; Handle bit 5 of shift offset. |
| 818 |
sbrs r16, 5 |
| 819 |
rjmp 4f |
| 820 |
wmov C5, C1 |
| 821 |
wmov C3, CA |
| 822 |
clr C2 $ clr C1 $ wmov CA, C1 |
| 823 |
4: |
| 824 |
;; Handle bit 4 of shift offset. |
| 825 |
sbrs r16, 4 |
| 826 |
rjmp 3f |
| 827 |
wmov C5, C3 |
| 828 |
wmov C3, C1 |
| 829 |
wmov C1, CA |
| 830 |
clr CA $ clr C0 |
| 831 |
3: |
| 832 |
;; Handle bits 3...0 of shift offset. |
| 833 |
push r16 |
| 834 |
andi r16, 0xf |
| 835 |
breq 0f |
| 836 |
|
| 837 |
.global __ashldi3 |
| 838 |
XCALL __ashldi3 |
| 839 |
0: |
| 840 |
pop r16 |
| 841 |
ret |
| 842 |
ENDF ashldi3 |
| 843 |
#endif /* Small device */ |
| 844 |
|
| 845 |
#endif /* F7MOD_addsub_mant_scaled_ */ |
| 846 |
|
| 847 |
#if defined F7MOD_mul_mant_ && defined (__AVR_HAVE_MUL__) |
| 848 |
#define A0 11 |
| 849 |
#define A1 A0+1 |
| 850 |
#define A2 A0+2 |
| 851 |
#define A3 A0+3 |
| 852 |
#define A4 A0+4 |
| 853 |
#define A5 A0+5 |
| 854 |
#define A6 A0+6 |
| 855 |
|
| 856 |
#define TT0 26 |
| 857 |
#define TT1 TT0+1 |
| 858 |
#define TT2 28 |
| 859 |
#define TT3 TT2+1 |
| 860 |
|
| 861 |
#define BB 10 |
| 862 |
|
| 863 |
;; R18.0 = 1: No rounding. |
| 864 |
|
| 865 |
DEFUN mul_mant |
| 866 |
;; 10 = Y, R17...R10 |
| 867 |
do_prologue_saves 10 |
| 868 |
;; T = R18.0: Skip rounding? |
| 869 |
bst r18, 0 |
| 870 |
;; Save result address for later. |
| 871 |
push r25 |
| 872 |
push r24 |
| 873 |
;; Load A's mantissa. |
| 874 |
movw ZL, r22 |
| 875 |
LDD A0, Z+0+Off |
| 876 |
LDD A1, Z+1+Off |
| 877 |
LDD A2, Z+2+Off |
| 878 |
LDD A3, Z+3+Off |
| 879 |
LDD A4, Z+4+Off |
| 880 |
LDD A5, Z+5+Off |
| 881 |
LDD A6, Z+6+Off |
| 882 |
movw ZL, r20 |
| 883 |
|
| 884 |
;; 6 * 6 -> 6:5 |
| 885 |
;; 4 * 6 -> 4:3 |
| 886 |
;; 2 * 6 -> 2:1 |
| 887 |
;; 0 * 6 -> 0:a |
| 888 |
ldd BB, Z+6+Off |
| 889 |
mul A6, BB $ movw C5, r0 |
| 890 |
mul A4, BB $ movw C3, r0 |
| 891 |
mul A2, BB $ movw C1, r0 |
| 892 |
mul A0, BB $ movw CA, r0 |
| 893 |
|
| 894 |
;; 5 * 6 -> 5:4 |
| 895 |
;; 3 * 6 -> 3:2 |
| 896 |
;; 1 * 6 -> 1:0 |
| 897 |
mul A5, BB $ movw TT2, r0 |
| 898 |
mul A3, BB $ movw TT0, r0 |
| 899 |
mul A1, BB |
| 900 |
ADD C0, r0 $ adc C1, r1 |
| 901 |
adc C2, TT0 $ adc C3, TT1 |
| 902 |
adc C4, TT2 $ adc C5, TT3 $ clr ZERO |
| 903 |
adc C6, ZERO |
| 904 |
;; Done B6 |
| 905 |
|
| 906 |
;; 6 * 5 -> 5:4 |
| 907 |
;; 4 * 5 -> 3:2 |
| 908 |
;; 2 * 5 -> 1:0 |
| 909 |
;; 0 * 5 -> a:- |
| 910 |
ldd BB, Z+5+Off |
| 911 |
mul A0, BB |
| 912 |
;; Done A0 |
| 913 |
#define Atmp A0 |
| 914 |
#define Null A0 |
| 915 |
|
| 916 |
mov Atmp, r1 |
| 917 |
mul A6, BB $ movw TT2, r0 |
| 918 |
mul A4, BB $ movw TT0, r0 |
| 919 |
mul A2, BB |
| 920 |
|
| 921 |
ADD CA, Atmp |
| 922 |
adc C0, r0 $ adc C1, r1 |
| 923 |
adc C2, TT0 $ adc C3, TT1 |
| 924 |
adc C4, TT2 $ adc C5, TT3 $ clr Null |
| 925 |
adc C6, Null |
| 926 |
|
| 927 |
;; 1 * 5 -> 0:a |
| 928 |
;; 3 * 5 -> 2:1 |
| 929 |
;; 5 * 5 -> 4:3 |
| 930 |
mul A1, BB $ movw TT0, r0 |
| 931 |
mul A3, BB $ movw TT2, r0 |
| 932 |
mul A5, BB |
| 933 |
|
| 934 |
ADD CA, TT0 $ adc C0, TT1 |
| 935 |
adc C1, TT2 $ adc C2, TT3 |
| 936 |
adc C3, r0 $ adc C4, r1 |
| 937 |
adc C5, Null $ adc C6, Null |
| 938 |
;; Done B5 |
| 939 |
|
| 940 |
;; 2 * 4 -> 0:a |
| 941 |
;; 4 * 4 -> 2:1 |
| 942 |
;; 6 * 4 -> 4:3 |
| 943 |
ldd BB, Z+4+Off |
| 944 |
mul A2, BB $ movw TT0, r0 |
| 945 |
mul A4, BB $ movw TT2, r0 |
| 946 |
mul A6, BB |
| 947 |
|
| 948 |
ADD CA, TT0 $ adc C0, TT1 |
| 949 |
adc C1, TT2 $ adc C2, TT3 |
| 950 |
adc C3, r0 $ adc C4, r1 |
| 951 |
adc C5, Null $ adc C6, Null |
| 952 |
|
| 953 |
;; 1 * 4 -> a:- |
| 954 |
;; 3 * 4 -> 1:0 |
| 955 |
;; 5 * 4 -> 3:2 |
| 956 |
mul A1, BB $ mov TT1, r1 |
| 957 |
mul A3, BB $ movw TT2, r0 |
| 958 |
mul A5, BB |
| 959 |
;; Done A1 |
| 960 |
;; Done B4 |
| 961 |
ADD CA, TT1 |
| 962 |
adc C0, TT2 $ adc C1, TT3 |
| 963 |
adc C2, r0 $ adc C3, r1 |
| 964 |
;; Accumulate carry for C3 in TT1. |
| 965 |
;; Accumulate carry for C4 in A1. |
| 966 |
#define Cry3 TT1 |
| 967 |
#define Cry4 A1 |
| 968 |
clr Cry3 |
| 969 |
clr Cry4 |
| 970 |
rol Cry4 |
| 971 |
|
| 972 |
;; 6 * 2 -> 2:1 |
| 973 |
;; 6 * 3 -> 3:2 |
| 974 |
;; 5 * 3 -> 2:1 |
| 975 |
ldd BB, Z+2+Off |
| 976 |
mul A6, BB |
| 977 |
add C1, r0 |
| 978 |
adc C2, r1 |
| 979 |
adc Cry3, Null |
| 980 |
|
| 981 |
ldd BB, Z+3+Off |
| 982 |
mul A6, BB |
| 983 |
add C2, r0 |
| 984 |
adc C3, r1 |
| 985 |
adc Cry4, Null |
| 986 |
|
| 987 |
mul A5, BB |
| 988 |
add C1, r0 |
| 989 |
adc C2, r1 |
| 990 |
adc Cry3, Null |
| 991 |
|
| 992 |
;; Perform the remaining 11 multiplications in 4 loopings: |
| 993 |
;; 4 * 3 -> 1:0 |
| 994 |
;; 3 * 3 -> 0:a |
| 995 |
;; 2 * 3 -> a:- |
| 996 |
;; |
| 997 |
;; 5 * 2 -> 1:0 |
| 998 |
;; 4 * 2 -> 0:a |
| 999 |
;; 3 * 2 -> a:- |
| 1000 |
;; |
| 1001 |
;; 6 * 1 -> 1:0 |
| 1002 |
;; 5 * 1 -> 0:a |
| 1003 |
;; 4 * 1 -> a:- |
| 1004 |
;; |
| 1005 |
;; . * 0 -> 1:0 (=0) |
| 1006 |
;; 6 * 0 -> 0:a |
| 1007 |
;; 5 * 0 -> a:- |
| 1008 |
|
| 1009 |
;; BB already contains B3, hence let Z point one past B2 so that |
| 1010 |
;; the LD *, -Z below will pick up B2, B1, B0. |
| 1011 |
adiw r30, 1 + Off+2 |
| 1012 |
|
| 1013 |
;; Accumulate carry for C2 in TT2. |
| 1014 |
#define Cry2 TT2 |
| 1015 |
clr Cry2 |
| 1016 |
|
| 1017 |
;; TT3 is the loop counter, iterate over B3...B0. |
| 1018 |
ldi TT3, 4 |
| 1019 |
rjmp .Loop_start |
| 1020 |
|
| 1021 |
.Loop: |
| 1022 |
;; We use A2...A4 below; so shift bytes of A into place. |
| 1023 |
mov A2, A3 |
| 1024 |
mov A3, A4 |
| 1025 |
mov A4, A5 |
| 1026 |
mov A5, A6 |
| 1027 |
clr A6 |
| 1028 |
ld BB, -Z |
| 1029 |
.Loop_start: |
| 1030 |
mul A3, BB |
| 1031 |
ADD CA, r0 $ adc C0, r1 $ adc C1, Null $ adc Cry2, Null |
| 1032 |
MUL A2, BB |
| 1033 |
mov TT0, r1 |
| 1034 |
MUL A4, BB |
| 1035 |
ADD CA, TT0 $ adc C0, r0 $ adc C1, r1 $ adc Cry2, Null |
| 1036 |
dec TT3 |
| 1037 |
brne .Loop |
| 1038 |
|
| 1039 |
clr ZERO |
| 1040 |
ADD C2, Cry2 |
| 1041 |
adc C3, Cry3 |
| 1042 |
adc C4, Cry4 |
| 1043 |
adc C5, ZERO |
| 1044 |
adc C6, ZERO |
| 1045 |
|
| 1046 |
;; Finally... |
| 1047 |
|
| 1048 |
pop ZL |
| 1049 |
pop ZH |
| 1050 |
;; The high byte is at least 0x40 and at most 0xfe. |
| 1051 |
;; The result has to be left-shifted by one in order to scale it |
| 1052 |
;; correctly. |
| 1053 |
|
| 1054 |
ldi Carry, 1 |
| 1055 |
F7call normalize.maybe_round.store_with_flags |
| 1056 |
|
| 1057 |
do_epilogue_restores 10 |
| 1058 |
|
| 1059 |
ENDF mul_mant |
| 1060 |
#endif /* F7MOD_mul_mant_ && MUL */ |
| 1061 |
|
| 1062 |
#if defined F7MOD_mul_mant_ && ! defined (__AVR_HAVE_MUL__) |
| 1063 |
#define AA TMP |
| 1064 |
#define A0 13 |
| 1065 |
#define A1 A0+1 |
| 1066 |
#define A2 A0+2 |
| 1067 |
#define A3 A0+3 |
| 1068 |
#define A4 A0+4 |
| 1069 |
#define A5 r26 |
| 1070 |
#define A6 r27 |
| 1071 |
#define BB ZERO |
| 1072 |
#define Bits r29 |
| 1073 |
#define Bytes r28 |
| 1074 |
|
| 1075 |
DEFUN mul_mant |
| 1076 |
do_prologue_saves 7 |
| 1077 |
bst r18, 0 ; T = 1: Don't round. |
| 1078 |
;; Save result address for later. |
| 1079 |
push r25 |
| 1080 |
push r24 |
| 1081 |
;; Load 1st operand mantissa. |
| 1082 |
wmov r30, r22 |
| 1083 |
clr AA |
| 1084 |
LDD A0, Z+0+Off |
| 1085 |
LDD A1, Z+1+Off |
| 1086 |
LDD A2, Z+2+Off |
| 1087 |
LDD A3, Z+3+Off |
| 1088 |
LDD A4, Z+4+Off |
| 1089 |
LDD A5, Z+5+Off |
| 1090 |
LDD A6, Z+6+Off |
| 1091 |
;; Let Z point one past .mant of the 2nd input operand. |
| 1092 |
wmov r30, r20 |
| 1093 |
adiw r30, Expo |
| 1094 |
|
| 1095 |
;; Clear the result mantissa. |
| 1096 |
.global __clr_8 |
| 1097 |
XCALL __clr_8 |
| 1098 |
|
| 1099 |
;; Loop over the bytes of B's mantissa from highest to lowest. |
| 1100 |
;; "+1" because we jump into the loop. |
| 1101 |
ldi Bytes, 1 + F7_MANT_BYTES |
| 1102 |
|
| 1103 |
;; Divide one operand by 2 so that the result mantissa won't overflow. |
| 1104 |
;; This is accounted for by "Carry = 1" below. |
| 1105 |
ldi Bits, 1 |
| 1106 |
rjmp .Loop_entry |
| 1107 |
|
| 1108 |
.Loop_bytes: |
| 1109 |
ld BB, -Z |
| 1110 |
;; Loop over the bits of B's mantissa from highest to lowest. |
| 1111 |
ldi Bits, 8 |
| 1112 |
.Loop_bits: |
| 1113 |
lsl BB |
| 1114 |
brcc .Lnext_bit |
| 1115 |
|
| 1116 |
ADD CA, AA |
| 1117 |
adc C0, A0 |
| 1118 |
adc C1, A1 |
| 1119 |
adc C2, A2 |
| 1120 |
adc C3, A3 |
| 1121 |
adc C4, A4 |
| 1122 |
adc C5, A5 |
| 1123 |
adc C6, A6 |
| 1124 |
|
| 1125 |
.Lnext_bit: |
| 1126 |
.Loop_entry: |
| 1127 |
LSR A6 |
| 1128 |
ror A5 |
| 1129 |
ror A4 |
| 1130 |
ror A3 |
| 1131 |
ror A2 |
| 1132 |
ror A1 |
| 1133 |
ror A0 |
| 1134 |
ror AA |
| 1135 |
|
| 1136 |
dec Bits |
| 1137 |
brne .Loop_bits |
| 1138 |
|
| 1139 |
dec Bytes |
| 1140 |
brne .Loop_bytes |
| 1141 |
|
| 1142 |
;; Finally... |
| 1143 |
|
| 1144 |
pop ZL |
| 1145 |
pop ZH |
| 1146 |
|
| 1147 |
;; The result has to be left-shifted by one (multiplied by 2) in order |
| 1148 |
;; to undo the division by 2 of the 1st operand. |
| 1149 |
ldi Carry, 1 |
| 1150 |
F7call normalize.maybe_round.store_with_flags |
| 1151 |
|
| 1152 |
do_epilogue_restores 7 |
| 1153 |
ENDF mul_mant |
| 1154 |
#endif /* F7MOD_mul_mant_ && ! MUL */ |
| 1155 |
|
| 1156 |
|
| 1157 |
#if defined (F7MOD_div_) |
| 1158 |
|
| 1159 |
;; Dividend is C[] |
| 1160 |
|
| 1161 |
;; Divisor |
| 1162 |
#define A0 9 |
| 1163 |
#define A1 10 |
| 1164 |
#define A2 11 |
| 1165 |
#define A3 12 |
| 1166 |
#define A4 13 |
| 1167 |
#define A5 14 |
| 1168 |
#define A6 15 |
| 1169 |
|
| 1170 |
;; Quotient |
| 1171 |
#define Q0 0 /* === TMP */ |
| 1172 |
#define Q1 Q0+1 /* === ZERO */ |
| 1173 |
#define Q2 26 |
| 1174 |
#define Q3 Q2+1 |
| 1175 |
#define Q4 28 |
| 1176 |
#define Q5 Q4+1 |
| 1177 |
#define Q6 16 |
| 1178 |
#define Q7 Q6+1 |
| 1179 |
|
| 1180 |
#define Cnt CA |
| 1181 |
#define QBits r8 |
| 1182 |
|
| 1183 |
DEFUN div |
| 1184 |
do_prologue_saves 12 |
| 1185 |
|
| 1186 |
;; Number of bits requested for the quotient. |
| 1187 |
;; This is usually 2 + F7_MANT_BITS. |
| 1188 |
mov QBits, r20 |
| 1189 |
wmov ZL, r22 |
| 1190 |
LDD A0, Z+0+Off |
| 1191 |
LDD A1, Z+1+Off |
| 1192 |
LDD A2, Z+2+Off |
| 1193 |
LDD A3, Z+3+Off |
| 1194 |
LDD A4, Z+4+Off |
| 1195 |
LDD A5, Z+5+Off |
| 1196 |
LDD A6, Z+6+Off |
| 1197 |
wmov ZL, r24 |
| 1198 |
F7call load_mant |
| 1199 |
|
| 1200 |
;; Clear quotient Q[]. |
| 1201 |
clr Q0 ; === TMP |
| 1202 |
;clr Q1 ; === ZERO |
| 1203 |
wmov Q2, Q0 |
| 1204 |
wmov Q4, Q0 |
| 1205 |
wmov Q6, Q0 |
| 1206 |
|
| 1207 |
;; C[] and A[] are valid mantissae, i.e. their MSBit is set. Therefore, |
| 1208 |
;; quotient Q[] will be in [0x0.ff..., 0x0.40...] and to adjust Q[] we |
| 1209 |
;; need at most 1 left-shift. Compute F7_MANT_BITS + 2 bits of the |
| 1210 |
;; quotient: One bit is used for rounding, and one bit might be consumed |
| 1211 |
;; by the mentioned left-shift. |
| 1212 |
mov Cnt, QBits |
| 1213 |
rjmp .Loop_start |
| 1214 |
|
| 1215 |
.Loop: |
| 1216 |
;; Shift dividend. |
| 1217 |
LSL C0 |
| 1218 |
rol C1 |
| 1219 |
rol C2 |
| 1220 |
rol C3 |
| 1221 |
rol C4 |
| 1222 |
rol C5 |
| 1223 |
rol C6 |
| 1224 |
brcs .Lfits |
| 1225 |
;; Compare dividend against divisor. |
| 1226 |
.Loop_start: |
| 1227 |
CP C0, A0 |
| 1228 |
cpc C1, A1 |
| 1229 |
cpc C2, A2 |
| 1230 |
cpc C3, A3 |
| 1231 |
cpc C4, A4 |
| 1232 |
cpc C5, A5 |
| 1233 |
cpc C6, A6 |
| 1234 |
;; Shift 0 into quotient. |
| 1235 |
brlo 1f |
| 1236 |
.Lfits: |
| 1237 |
;; Divisor fits into dividend. |
| 1238 |
SUB C0, A0 |
| 1239 |
sbc C1, A1 |
| 1240 |
sbc C2, A2 |
| 1241 |
sbc C3, A3 |
| 1242 |
sbc C4, A4 |
| 1243 |
sbc C5, A5 |
| 1244 |
sbc C6, A6 |
| 1245 |
;; Shift 1 into quotient. |
| 1246 |
sec |
| 1247 |
rol Q0 |
| 1248 |
skipnext |
| 1249 |
1: lsl Q0 |
| 1250 |
rol Q1 |
| 1251 |
rol Q2 |
| 1252 |
rol Q3 |
| 1253 |
rol Q4 |
| 1254 |
rol Q5 |
| 1255 |
rol Q6 |
| 1256 |
rol Q7 |
| 1257 |
dec Cnt |
| 1258 |
brne .Loop |
| 1259 |
|
| 1260 |
wmov CA, Q0 |
| 1261 |
wmov C1, Q2 |
| 1262 |
wmov C3, Q4 |
| 1263 |
wmov C5, Q6 |
| 1264 |
clr ZERO |
| 1265 |
|
| 1266 |
ldi Carry, 64 |
| 1267 |
sub Carry, QBits |
| 1268 |
F7call normalize.round.store_with_flags |
| 1269 |
|
| 1270 |
do_epilogue_restores 12 |
| 1271 |
ENDF div |
| 1272 |
|
| 1273 |
#endif /* F7MOD_div_ */ |
| 1274 |
|
| 1275 |
|
| 1276 |
#ifdef F7MOD_sqrt_approx_ |
| 1277 |
;; ReMainder |
| 1278 |
#define MX 16 |
| 1279 |
#define M0 17 |
| 1280 |
#define M1 26 |
| 1281 |
#define M2 27 |
| 1282 |
#define M3 14 |
| 1283 |
#define M4 15 |
| 1284 |
#define M5 TMP |
| 1285 |
#define M6 r29 |
| 1286 |
|
| 1287 |
#define AA ZERO |
| 1288 |
#define One r13 |
| 1289 |
#define Bits r28 |
| 1290 |
#define Bytes M6 |
| 1291 |
|
| 1292 |
;; Extend C[] by 0b01 at the low end. |
| 1293 |
#define CX (0b01 << 6) |
| 1294 |
|
| 1295 |
;;; Compute square-root of const f7_t *R22 for a positive number. |
| 1296 |
DEFUN sqrt_approx |
| 1297 |
;; 7 = Y, R17...R13 |
| 1298 |
do_prologue_saves 7 |
| 1299 |
|
| 1300 |
wmov ZL, r22 ; Input const f7_t* |
| 1301 |
wmov YL, r24 ; Output f7_t* |
| 1302 |
F7call load_mant |
| 1303 |
ldi CA, 0xff |
| 1304 |
|
| 1305 |
;; The paper-pencil method for the mantissa consumes bits in pairs and |
| 1306 |
;; expects the input as Q-format 2.*, but mant is in 1.*. This means |
| 1307 |
;; we have to shift one to the right. If expo is odd, then we shift |
| 1308 |
;; one to the left and subtract one from expo in order to compensate |
| 1309 |
;; and to get an even exponent. |
| 1310 |
|
| 1311 |
;; Divide expo by 2 because we are doing sqrt. |
| 1312 |
ldd XH, Z+Expo+1 |
| 1313 |
ldd XL, Z+Expo+0 |
| 1314 |
asr XH |
| 1315 |
ror XL |
| 1316 |
;; Store expo to result. |
| 1317 |
wmov ZL, YL |
| 1318 |
std Z+Expo+0, XL |
| 1319 |
std Z+Expo+1, XH |
| 1320 |
|
| 1321 |
brcs 1f |
| 1322 |
;; Expo was even. Do >>=1 in order to get Q2.* as explained above. |
| 1323 |
LSR C6 $ ror C5 $ ror C4 $ ror C3 |
| 1324 |
ror C2 $ ror C1 $ ror C0 $ ror CA |
| 1325 |
1: |
| 1326 |
;; For odd expo, >>=1 to get Q2.* and <<=1 to get an even expo cancel out. |
| 1327 |
;; And the right-shift of the exponent implicitly subtracted 1 from it |
| 1328 |
;; as needed. |
| 1329 |
F7call store_mant.with_flags |
| 1330 |
|
| 1331 |
;; Let Z point one past the mantissa's MSB. |
| 1332 |
adiw ZL, Off + F7_MANT_BYTES |
| 1333 |
|
| 1334 |
;; Clear the result mantissa. |
| 1335 |
.global __clr_8 |
| 1336 |
XCALL __clr_8 |
| 1337 |
;; Clear ReMainder. M6 === Bytes will be zero when Bytes is down to 0. |
| 1338 |
clr M5 |
| 1339 |
wmov M3, C3 |
| 1340 |
wmov M1, C1 |
| 1341 |
wmov MX, CA |
| 1342 |
|
| 1343 |
clr One |
| 1344 |
inc One |
| 1345 |
|
| 1346 |
;; "+1" because .flags extends the mantissa at the low end. |
| 1347 |
ldi Bytes, 1 + F7_MANT_BYTES |
| 1348 |
.Loop_bytes: |
| 1349 |
ld AA, -Z |
| 1350 |
ldi Bits, 8 |
| 1351 |
.Loop_bits: |
| 1352 |
;; Shift top 2 bits of MX into M[]. |
| 1353 |
LSL MX $ rol M0 $ rol M1 $ rol M2 $ rol M3 |
| 1354 |
LSL MX $ rol M0 $ rol M1 $ rol M2 $ rol M3 |
| 1355 |
|
| 1356 |
;; "Take down" 2 bits from A[] to MX.7 and MX.6 |
| 1357 |
mov MX, AA |
| 1358 |
andi MX, 0xc0 |
| 1359 |
lsl AA |
| 1360 |
lsl AA |
| 1361 |
|
| 1362 |
;; Compare remainder against current result extended by 0b01. |
| 1363 |
CPI MX, CX |
| 1364 |
cpc M0, C0 |
| 1365 |
cpc M1, C1 |
| 1366 |
cpc M2, C2 |
| 1367 |
cpc M3, C3 |
| 1368 |
brcs 1f |
| 1369 |
;; If the extended result fits, subtract it from M and set the |
| 1370 |
;; next result bit to 1. |
| 1371 |
SUBI MX, CX |
| 1372 |
sbc M0, C0 |
| 1373 |
sbc M1, C1 |
| 1374 |
sbc M2, C2 |
| 1375 |
sbc M3, C3 |
| 1376 |
1: |
| 1377 |
;; If it doesn't fit, set the next result bit to 0 (and don't subtract). |
| 1378 |
rol C0 |
| 1379 |
eor C0, One |
| 1380 |
rol C1 |
| 1381 |
rol C2 |
| 1382 |
rol C3 |
| 1383 |
|
| 1384 |
subi Bits, 2 |
| 1385 |
brne .Loop_bits |
| 1386 |
;; AA (=== ZERO) is zero again. |
| 1387 |
|
| 1388 |
dec Bytes |
| 1389 |
brne .Loop_bytes |
| 1390 |
;; B6 (=== Bytes) is zero now. |
| 1391 |
|
| 1392 |
;; Now we consumed all the 64 bits of the extended mantissa, but we |
| 1393 |
;; only expanded 64 / 2 = 32 bits of the result, which is currently |
| 1394 |
;; held in C3 ... C0. Do the same like above, but on all bytes. |
| 1395 |
;; Shift in 00's because the mantissa is exhausted. |
| 1396 |
|
| 1397 |
;; "-1" because flags is part of the mantissa, which is already consumed. |
| 1398 |
ldi Bits, 8 * (F7_MANT_BYTES - 1) |
| 1399 |
.Loop2_bits: |
| 1400 |
;; Shift top 2 bits of MX into M[]. |
| 1401 |
.Ltwice: |
| 1402 |
LSL MX |
| 1403 |
rol M0 |
| 1404 |
rol M1 |
| 1405 |
rol M2 |
| 1406 |
rol M3 |
| 1407 |
rol M4 |
| 1408 |
rol M5 |
| 1409 |
rol M6 |
| 1410 |
subi Bits, 0x80 |
| 1411 |
brmi .Ltwice |
| 1412 |
|
| 1413 |
;; "Take down" two 0's to MX.7 and MX.6 |
| 1414 |
; clr MX ;; MX is already zero. |
| 1415 |
|
| 1416 |
;; Compare remainder against current result extended by 0b01. |
| 1417 |
CPI MX, CX |
| 1418 |
cpc M0, C0 |
| 1419 |
cpc M1, C1 |
| 1420 |
cpc M2, C2 |
| 1421 |
cpc M3, C3 |
| 1422 |
cpc M4, C4 |
| 1423 |
cpc M5, C5 |
| 1424 |
cpc M6, C6 |
| 1425 |
brcs 1f |
| 1426 |
;; If the extended result fits, subtract it from M and set the |
| 1427 |
;; next result bit to 1. |
| 1428 |
SUBI MX, CX |
| 1429 |
sbc M0, C0 |
| 1430 |
sbc M1, C1 |
| 1431 |
sbc M2, C2 |
| 1432 |
sbc M3, C3 |
| 1433 |
sbc M4, C4 |
| 1434 |
sbc M5, C5 |
| 1435 |
sbc M6, C6 |
| 1436 |
1: |
| 1437 |
;; If it doesn't fit, set the next result bit to 0 (and don't subtract). |
| 1438 |
rol C0 |
| 1439 |
eor C0, One |
| 1440 |
rol C1 |
| 1441 |
rol C2 |
| 1442 |
rol C3 |
| 1443 |
rol C4 |
| 1444 |
rol C5 |
| 1445 |
rol C6 |
| 1446 |
|
| 1447 |
subi Bits, 2 |
| 1448 |
brne .Loop2_bits |
| 1449 |
|
| 1450 |
;; Set flags. |
| 1451 |
st Z, ZERO |
| 1452 |
F7call store_mant |
| 1453 |
|
| 1454 |
do_epilogue_restores 7 |
| 1455 |
ENDF sqrt_approx |
| 1456 |
#endif /* F7MOD_sqrt_approx_ */ |
| 1457 |
|
| 1458 |
|
| 1459 |
#if defined (F7MOD_sqrt16_) && defined (__AVR_HAVE_MUL__) |
| 1460 |
|
| 1461 |
#define Mask C6 |
| 1462 |
#define Q0 C3 /* = R22 */ |
| 1463 |
#define Q1 C4 /* = R23 */ |
| 1464 |
|
| 1465 |
;; uint16_t R24 = sqrt16_XXX (uint16_t R24); |
| 1466 |
;; Clobbers: R22, R23, TMP. |
| 1467 |
;; |
| 1468 |
;; XXX = floor: Return integral part of square-root of R25:R24 with R25 = 0. |
| 1469 |
;; Error is in [0, -1 LSB). |
| 1470 |
;; XXX = round: Return quare-root of R25:R24 rounded to nearest integer. |
| 1471 |
;; R25 = (Q[] >= 65281) = (Q > 0xff00), i.e. if Q[] is not |
| 1472 |
;; bigger than 0xff00, then the result fits in 8 bits. |
| 1473 |
;; Return C = 0 if the result is the same as for XXX = floor, |
| 1474 |
;; error in [0, -1/2 LSB) |
| 1475 |
;; Return C = 1 if the result is one higher than for XXX = floor, |
| 1476 |
;; error in [1/2 LSB, 0). |
| 1477 |
DEFUN sqrt16_round |
| 1478 |
set |
| 1479 |
skipnext |
| 1480 |
;; ... |
| 1481 |
LABEL sqrt16_floor |
| 1482 |
clt ; Skipped? |
| 1483 |
movw Q0, r24 |
| 1484 |
clr C5 |
| 1485 |
ldi Mask, 1 << 7 |
| 1486 |
|
| 1487 |
.Loop_mask: |
| 1488 |
add C5, Mask |
| 1489 |
mul C5, C5 |
| 1490 |
cp Q0, R0 |
| 1491 |
cpc Q1, R1 |
| 1492 |
brsh 1f |
| 1493 |
sub C5, Mask |
| 1494 |
1: lsr Mask |
| 1495 |
brne .Loop_mask |
| 1496 |
|
| 1497 |
brtc .Ldone ; No rounding => C6 will be 0. |
| 1498 |
|
| 1499 |
;; Rounding: (X + 1/2)^2 = X^2 + X + 1/4, thus probing |
| 1500 |
;; for bit -1 is testing Q[] against C5^2 + C5. |
| 1501 |
mul C5, C5 |
| 1502 |
add R0, C5 |
| 1503 |
adc R1, C6 ; Exploit C6 === Mask = 0. |
| 1504 |
cp R0, Q0 |
| 1505 |
cpc R1, Q1 |
| 1506 |
brcc .Ldone |
| 1507 |
;; If C5^2 + C5 + 1/4 fits into Q[], then round up and C = 1. |
| 1508 |
adiw C5, 1 ; Exploit C6 === Mask = 0. |
| 1509 |
sec |
| 1510 |
|
| 1511 |
.Ldone: |
| 1512 |
clr __zero_reg__ |
| 1513 |
ret |
| 1514 |
ENDF sqrt16_round |
| 1515 |
#undef Mask |
| 1516 |
#undef Q0 |
| 1517 |
#undef Q1 |
| 1518 |
#endif /* F7MOD_sqrt16_ && MUL */ |
| 1519 |
|
| 1520 |
|
| 1521 |
#undef CA |
| 1522 |
#undef C0 |
| 1523 |
#undef C1 |
| 1524 |
#undef C2 |
| 1525 |
#undef C3 |
| 1526 |
#undef C4 |
| 1527 |
#undef C5 |
| 1528 |
#undef C6 |
| 1529 |
#undef Carry |
| 1530 |
|
| 1531 |
|
| 1532 |
#ifdef F7MOD_D_fma_ |
| 1533 |
_DEFUN __fma |
| 1534 |
DALIAS fma |
| 1535 |
LALIAS fmal |
| 1536 |
|
| 1537 |
#define n_pushed 4 |
| 1538 |
#define n_frame (2 * F7_SIZEOF) |
| 1539 |
|
| 1540 |
do_prologue_saves n_pushed, n_frame |
| 1541 |
;; Y = FramePointer + 1 |
| 1542 |
adiw Y, 1 |
| 1543 |
|
| 1544 |
;; FP + 1 = (f7_t) arg1 |
| 1545 |
wmov r16, Y |
| 1546 |
;; The double argument arg1 is already in R18[]. |
| 1547 |
XCALL F7_NAME (set_double_impl) |
| 1548 |
|
| 1549 |
;; The double argument arg2 is in R10[]. Move it to R18[]. |
| 1550 |
wmov r18, r10 |
| 1551 |
wmov r20, r12 |
| 1552 |
wmov r22, r14 |
| 1553 |
;; R16, R17 are clobbered. Fetch them from where prologue_saves put them. |
| 1554 |
ldd r24, Y + n_frame + 3 ; Saved R16 |
| 1555 |
ldd r25, Y + n_frame + 2 ; Saved R17 |
| 1556 |
;; FP + 1 + 10 = (f7_t) arg2 |
| 1557 |
subi r16, lo8 (-F7_SIZEOF) |
| 1558 |
sbci r17, hi8 (-F7_SIZEOF) |
| 1559 |
XCALL F7_NAME (set_double_impl) |
| 1560 |
|
| 1561 |
wmov r24, Y ; &arg1 |
| 1562 |
wmov r22, r16 ; &arg2 |
| 1563 |
XCALL F7_NAME (Imul) ; arg1 *= arg2 |
| 1564 |
|
| 1565 |
;; The 3rd double argument arg3 was passed on the stack. Move it to R18[], |
| 1566 |
;; Don't use f7_set_pdouble() because that function is unused (for now). |
| 1567 |
.irp n, 0, 1, 2, 3, 4, 5, 6, 7 |
| 1568 |
ldd 18+\n, Y + n_frame + n_pushed + PC_SIZE + \n |
| 1569 |
.endr |
| 1570 |
XCALL F7_NAME (set_double_impl) |
| 1571 |
|
| 1572 |
wmov r24, Y ; &arg1 |
| 1573 |
wmov r22, r16 ; &arg2 |
| 1574 |
XCALL F7_NAME (Iadd) ; arg1 += arg2 |
| 1575 |
|
| 1576 |
wmov r24, Y ; &arg1 |
| 1577 |
XCALL F7_NAME (get_double) |
| 1578 |
|
| 1579 |
do_epilogue_restores n_pushed, n_frame |
| 1580 |
_ENDF __fma |
| 1581 |
#endif /* F7MOD_D_fma_ */ |
| 1582 |
|
| 1583 |
|
| 1584 |
#ifdef F7MOD_D_fabs_ |
| 1585 |
_DEFUN __fabs |
| 1586 |
DALIAS fabs |
| 1587 |
LALIAS fabsl |
| 1588 |
andi R25, 0b01111111 |
| 1589 |
ret |
| 1590 |
_ENDF __fabs |
| 1591 |
#endif /* F7MOD_D_fabs_ */ |
| 1592 |
|
| 1593 |
|
| 1594 |
#ifdef F7MOD_D_neg_ |
| 1595 |
_DEFUN __neg |
| 1596 |
_LABEL __negdf2 |
| 1597 |
subi R25, 0b10000000 |
| 1598 |
ret |
| 1599 |
_ENDF __neg |
| 1600 |
#endif /* F7MOD_D_neg_ */ |
| 1601 |
|
| 1602 |
|
| 1603 |
#ifdef F7MOD_D_signbit_ |
| 1604 |
_DEFUN __signbit |
| 1605 |
DALIAS signbit |
| 1606 |
LALIAS signbitl |
| 1607 |
bst R25, 7 |
| 1608 |
clr R25 |
| 1609 |
clr R24 |
| 1610 |
bld R24, 0 |
| 1611 |
ret |
| 1612 |
_ENDF __signbit |
| 1613 |
#endif /* F7MOD_D_signbit_ */ |
| 1614 |
|
| 1615 |
|
| 1616 |
#ifdef F7MOD_D_copysign_ |
| 1617 |
_DEFUN __copysign |
| 1618 |
DALIAS copysign |
| 1619 |
LALIAS copysignl |
| 1620 |
bst R17, 7 |
| 1621 |
bld R25, 7 |
| 1622 |
ret |
| 1623 |
_ENDF __copysign |
| 1624 |
#endif /* F7MOD_D_copysign_ */ |
| 1625 |
|
| 1626 |
|
| 1627 |
#ifdef F7MOD_D_isinf_ |
| 1628 |
;;; +Inf -> +1 |
| 1629 |
;;; -Inf -> -1 |
| 1630 |
_DEFUN __isinf |
| 1631 |
DALIAS isinf |
| 1632 |
LALIAS isinfl |
| 1633 |
;; Save sign for later |
| 1634 |
push R25 |
| 1635 |
F7call class_D |
| 1636 |
pop TMP |
| 1637 |
ldi R24, 0 |
| 1638 |
ldi R25, 0 |
| 1639 |
;; Inf: T = Z = 1. |
| 1640 |
brtc 0f ; ordinary number |
| 1641 |
brne 0f ; Nan |
| 1642 |
ldi R24, 1 |
| 1643 |
sbrc TMP, 7 |
| 1644 |
sbiw R24, 2 |
| 1645 |
0: ret |
| 1646 |
_ENDF __isinf |
| 1647 |
#endif /* F7MOD_D_isinf_ */ |
| 1648 |
|
| 1649 |
|
| 1650 |
#ifdef F7MOD_D_isnan_ |
| 1651 |
_DEFUN __isnan |
| 1652 |
DALIAS isnan |
| 1653 |
LALIAS isnanl |
| 1654 |
F7call class_D |
| 1655 |
;; NaN: T = 1, Z = 0. |
| 1656 |
brtc 0f |
| 1657 |
ldi R24, 1 |
| 1658 |
brne 1f |
| 1659 |
0: |
| 1660 |
clr R24 |
| 1661 |
1: |
| 1662 |
clr R25 |
| 1663 |
ret |
| 1664 |
_ENDF __isnan |
| 1665 |
#endif /* F7MOD_D_isnan_ */ |
| 1666 |
|
| 1667 |
|
| 1668 |
#ifdef F7MOD_D_isfinite_ |
| 1669 |
_DEFUN __isfinite |
| 1670 |
DALIAS isfinite |
| 1671 |
LALIAS isfinitel |
| 1672 |
F7call class_D |
| 1673 |
;; Number <=> T = 0. |
| 1674 |
bld R24, 0 |
| 1675 |
com R24 |
| 1676 |
andi R24, 1 |
| 1677 |
clr R25 |
| 1678 |
ret |
| 1679 |
_ENDF __isfinite |
| 1680 |
#endif /* F7MOD_D_isfinite_ */ |
| 1681 |
|
| 1682 |
|
| 1683 |
#ifdef F7MOD_D_class_ |
| 1684 |
;; The encoded exponent has 11 Bits. |
| 1685 |
#define MAX_BIASED_EXPO 0b0111111111110000 |
| 1686 |
|
| 1687 |
;; Classify a double in R18[] |
| 1688 |
;; Number: T-Flag = 0. |
| 1689 |
;; +-Inf : T-Flag = 1, Z-Flag = 1. |
| 1690 |
;; NaN : T-Flag = 1, Z-Flag = 0. |
| 1691 |
DEFUN class_D |
| 1692 |
wmov R26, R24 |
| 1693 |
andi R26, lo8 (MAX_BIASED_EXPO) |
| 1694 |
andi R27, hi8 (MAX_BIASED_EXPO) |
| 1695 |
subi R26, lo8 (MAX_BIASED_EXPO) |
| 1696 |
sbci R27, hi8 (MAX_BIASED_EXPO) |
| 1697 |
clt |
| 1698 |
brne .L.number |
| 1699 |
set |
| 1700 |
;; Set sign and expo to 0. |
| 1701 |
clr R25 |
| 1702 |
andi R24, lo8 (~MAX_BIASED_EXPO) |
| 1703 |
;; What remains is the mantissa. |
| 1704 |
;; Mantissa == 0 => +/-Inf. |
| 1705 |
;; Mantissa != 0 => NaN. |
| 1706 |
;; Compare R18[] against sign_extend(R26) with R26 = 0. |
| 1707 |
.global __cmpdi2_s8 |
| 1708 |
XJMP __cmpdi2_s8 |
| 1709 |
.L.number: |
| 1710 |
ret |
| 1711 |
|
| 1712 |
ENDF class_D |
| 1713 |
#endif /* F7MOD_D_class_ */ |
| 1714 |
|
| 1715 |
|
| 1716 |
#ifdef F7MOD_D_cmp_ |
| 1717 |
|
| 1718 |
#define A0 18 |
| 1719 |
#define A1 A0 + 1 |
| 1720 |
#define A2 A0 + 2 |
| 1721 |
#define A3 A0 + 3 |
| 1722 |
#define A4 A0 + 4 |
| 1723 |
#define A5 A0 + 5 |
| 1724 |
#define A6 A0 + 6 |
| 1725 |
#define A7 A0 + 7 |
| 1726 |
|
| 1727 |
#define B0 10 |
| 1728 |
#define B1 B0 + 1 |
| 1729 |
#define B2 B0 + 2 |
| 1730 |
#define B3 B0 + 3 |
| 1731 |
#define B4 B0 + 4 |
| 1732 |
#define B5 B0 + 5 |
| 1733 |
#define B6 B0 + 6 |
| 1734 |
#define B7 B0 + 7 |
| 1735 |
|
| 1736 |
#define AA5 XH |
| 1737 |
#define AA6 ZL |
| 1738 |
#define AA7 ZH |
| 1739 |
|
| 1740 |
#define BB0 A0 |
| 1741 |
#define BB1 A1 |
| 1742 |
#define BB2 A2 |
| 1743 |
#define BB3 A3 |
| 1744 |
#define BB4 A4 |
| 1745 |
#define BB5 A5 |
| 1746 |
#define BB6 A6 |
| 1747 |
#define BB7 A7 |
| 1748 |
|
| 1749 |
;;; Helper for __<cmp>df2 and __unorddf2. |
| 1750 |
;;; T = 1: Comparison is unordered. |
| 1751 |
;;; T = 0: Comparison is ordered, and Z, N, C, S flags are set according |
| 1752 |
;;; to compare (double A, double B) as if set by a signed int comparison. |
| 1753 |
;;; Note that f(+0) = f(-0) = 0. |
| 1754 |
;;; In any case: |
| 1755 |
;;; - return R24 = 1. |
| 1756 |
;;; - return R25.0 = isNaN (A) |
| 1757 |
;;; - return R25.1 = isNaN (B) |
| 1758 |
DEFUN D_cmp |
| 1759 |
rcall D_cmp.map_i64 |
| 1760 |
bld __tmp_reg__, 0 |
| 1761 |
push __tmp_reg__ |
| 1762 |
;; Save A somewhere else... |
| 1763 |
wmov AA6, A6 |
| 1764 |
mov AA5, A5 |
| 1765 |
push A4 |
| 1766 |
push A3 |
| 1767 |
push A2 |
| 1768 |
push A1 |
| 1769 |
mov r0, A0 |
| 1770 |
;; ... so that we can use D_cmp.map_i64 on B. |
| 1771 |
wmov BB6, B6 |
| 1772 |
wmov BB4, B4 |
| 1773 |
wmov BB2, B2 |
| 1774 |
wmov BB0, B0 |
| 1775 |
rcall D_cmp.map_i64 |
| 1776 |
;; Run the following code even when B is NaN (T=1) so as to pop the regs. |
| 1777 |
;; In the non-NaN case, AA and BB can be compared like int64_t for the |
| 1778 |
;; sake of comparing A and B as double. |
| 1779 |
CP r0, BB0 $ pop r0 |
| 1780 |
cpc r0, BB1 $ pop r0 |
| 1781 |
cpc r0, BB2 $ pop r0 |
| 1782 |
cpc r0, BB3 $ pop r0 |
| 1783 |
cpc r0, BB4 |
| 1784 |
cpc AA5, BB5 |
| 1785 |
cpc AA6, BB6 |
| 1786 |
cpc AA7, BB7 |
| 1787 |
pop r25 |
| 1788 |
;; R25.0 <=> A is NaN |
| 1789 |
;; R25.1 <=> B is NaN |
| 1790 |
;; T <=> comparison is unordered |
| 1791 |
bld r25, 1 |
| 1792 |
sbrc r25, 0 |
| 1793 |
set |
| 1794 |
ldi r24, 1 |
| 1795 |
ret |
| 1796 |
|
| 1797 |
;;; A is NaN: Set T=1. |
| 1798 |
;;; A is not a NaN: Set T=0, and map double A to int64_t such that |
| 1799 |
;;; f(A) <cmp> f(B) iff A <cmp> B, i.e. we can treat the result |
| 1800 |
;;; as int64_t for the matter of double comparison. |
| 1801 |
;;; Clobbers: XL. |
| 1802 |
D_cmp.map_i64: |
| 1803 |
bst A7, 7 |
| 1804 |
cbr A7, 0x80 |
| 1805 |
;; If Inf < |A|, then we have a NaN. |
| 1806 |
CP __zero_reg__, A0 |
| 1807 |
cpc __zero_reg__, A1 |
| 1808 |
cpc __zero_reg__, A2 |
| 1809 |
cpc __zero_reg__, A3 |
| 1810 |
cpc __zero_reg__, A4 |
| 1811 |
cpc __zero_reg__, A5 |
| 1812 |
ldi XL, lo8(0x7ff0) $ cpc XL, A6 |
| 1813 |
ldi XL, hi8(0x7ff0) $ cpc XL, A7 |
| 1814 |
brlo .Lunord |
| 1815 |
brtc 9f |
| 1816 |
clt |
| 1817 |
.global __negdi2 |
| 1818 |
XJMP __negdi2 |
| 1819 |
.Lunord: |
| 1820 |
set |
| 1821 |
9: ret |
| 1822 |
|
| 1823 |
ENDF D_cmp |
| 1824 |
#endif /* F7MOD_D_cmp_ */ |
| 1825 |
|
| 1826 |
|
| 1827 |
;; bool __ledf2 (double, double); |
| 1828 |
#ifdef F7MOD_D_le_ |
| 1829 |
_DEFUN __ledf2 |
| 1830 |
F7call D_cmp |
| 1831 |
brts 0f |
| 1832 |
breq 1f |
| 1833 |
brlt 1f |
| 1834 |
0: ldi r24, 0 |
| 1835 |
1: ret |
| 1836 |
_ENDF __ledf2 |
| 1837 |
#endif /* F7MOD_D_le_ */ |
| 1838 |
|
| 1839 |
;; bool __ltdf2 (double, double); |
| 1840 |
#ifdef F7MOD_D_lt_ |
| 1841 |
_DEFUN __ltdf2 |
| 1842 |
F7call D_cmp |
| 1843 |
brts 0f |
| 1844 |
brlt 1f |
| 1845 |
0: ldi r24, 0 |
| 1846 |
1: ret |
| 1847 |
_ENDF __ltdf2 |
| 1848 |
#endif /* F7MOD_D_lt_ */ |
| 1849 |
|
| 1850 |
;; bool __gedf2 (double, double); |
| 1851 |
#ifdef F7MOD_D_ge_ |
| 1852 |
_DEFUN __gedf2 |
| 1853 |
F7call D_cmp |
| 1854 |
brts 0f |
| 1855 |
brge 1f |
| 1856 |
0: ldi r24, 0 |
| 1857 |
1: ret |
| 1858 |
_ENDF __gedf2 |
| 1859 |
#endif /* F7MOD_D_ge_ */ |
| 1860 |
|
| 1861 |
;; bool __gtdf2 (double, double); |
| 1862 |
#ifdef F7MOD_D_gt_ |
| 1863 |
_DEFUN __gtdf2 |
| 1864 |
F7call D_cmp |
| 1865 |
brts 0f |
| 1866 |
breq 0f |
| 1867 |
brge 1f |
| 1868 |
0: ldi r24, 0 |
| 1869 |
1: ret |
| 1870 |
_ENDF __gtdf2 |
| 1871 |
#endif /* F7MOD_D_gt_ */ |
| 1872 |
|
| 1873 |
;; bool __nedf2 (double, double); |
| 1874 |
#ifdef F7MOD_D_ne_ |
| 1875 |
_DEFUN __nedf2 |
| 1876 |
F7call D_cmp |
| 1877 |
brts 0f |
| 1878 |
brne 1f |
| 1879 |
0: ldi r24, 0 |
| 1880 |
1: ret |
| 1881 |
_ENDF __nedf2 |
| 1882 |
#endif /* F7MOD_D_ne_ */ |
| 1883 |
|
| 1884 |
;; bool __eqdf2 (double, double); |
| 1885 |
#ifdef F7MOD_D_eq_ |
| 1886 |
_DEFUN __eqdf2 |
| 1887 |
F7call D_cmp |
| 1888 |
brts 0f |
| 1889 |
breq 1f |
| 1890 |
0: ldi r24, 0 |
| 1891 |
1: ret |
| 1892 |
_ENDF __eqdf2 |
| 1893 |
#endif /* F7MOD_D_eq_ */ |
| 1894 |
|
| 1895 |
;; bool __unorddf2 (double, double); |
| 1896 |
#ifdef F7MOD_D_unord_ |
| 1897 |
_DEFUN __unorddf2 |
| 1898 |
F7call D_cmp |
| 1899 |
bld r24, 0 |
| 1900 |
ret |
| 1901 |
_ENDF __unorddf2 |
| 1902 |
#endif /* F7MOD_D_unord_ */ |
| 1903 |
|
| 1904 |
#ifdef F7MOD_D_fminfmax_ |
| 1905 |
_DEFUN __fmin |
| 1906 |
DALIAS fmin |
| 1907 |
LALIAS fminl |
| 1908 |
inc __zero_reg__ |
| 1909 |
|
| 1910 |
_LABEL __fmax |
| 1911 |
DALIAS fmax |
| 1912 |
LALIAS fmaxl |
| 1913 |
;; Push A[]. |
| 1914 |
push r25 |
| 1915 |
push r24 |
| 1916 |
push r23 |
| 1917 |
push r22 |
| 1918 |
push r21 |
| 1919 |
push r20 |
| 1920 |
push r19 |
| 1921 |
push r18 |
| 1922 |
;; fmin or fmax |
| 1923 |
push __zero_reg__ |
| 1924 |
clr __zero_reg__ |
| 1925 |
|
| 1926 |
XCALL __gedf2 |
| 1927 |
|
| 1928 |
pop __tmp_reg__ |
| 1929 |
andi r25, 0x3 ; NaNs? |
| 1930 |
brne .Lnan |
| 1931 |
;; No NaNs involved. |
| 1932 |
eor __tmp_reg__, r24 ; (f == fmin) ^ (A >= B) |
| 1933 |
brne 1f |
| 1934 |
2: |
| 1935 |
;; Return B since the cases are: |
| 1936 |
;; fmax && A < B |
| 1937 |
;; fmin && A >= B |
| 1938 |
#ifdef __AVR_XMEGA__ |
| 1939 |
in XL, __SP_L__ |
| 1940 |
in XH, __SP_H__ |
| 1941 |
adiw XL, 8 |
| 1942 |
out __SP_L__, XL |
| 1943 |
out __SP_H__, XH |
| 1944 |
#else |
| 1945 |
pop r0 $ pop r0 $ pop r0 $ pop r0 |
| 1946 |
pop r0 $ pop r0 $ pop r0 $ pop r0 |
| 1947 |
#endif |
| 1948 |
wmov r24, r16 |
| 1949 |
wmov r22, r14 |
| 1950 |
wmov r20, r12 |
| 1951 |
wmov r18, r10 |
| 1952 |
ret |
| 1953 |
1: |
| 1954 |
;; Return A since the cases are: |
| 1955 |
;; fmax && A >= B |
| 1956 |
;; fmin && A < B |
| 1957 |
pop r18 |
| 1958 |
pop r19 |
| 1959 |
pop r20 |
| 1960 |
pop r21 |
| 1961 |
pop r22 |
| 1962 |
pop r23 |
| 1963 |
pop r24 |
| 1964 |
pop r25 |
| 1965 |
ret |
| 1966 |
|
| 1967 |
.Lnan: |
| 1968 |
;; There are NaNs. |
| 1969 |
;; When only the 1st argument is a NaN, then return the 2nd argument |
| 1970 |
cpi r25, 0x1 |
| 1971 |
breq 2b |
| 1972 |
;; When the 2nd argument is a NaN, then return the 1st argument. |
| 1973 |
;; When both arguments are NaNs, then return NaN (e.g. the 1st argument). |
| 1974 |
rjmp 1b |
| 1975 |
_ENDF __fmax |
| 1976 |
#endif /* F7MOD_D_fminfmax_ */ |
| 1977 |
|
| 1978 |
|
| 1979 |
#ifdef F7MOD_D_sincos_ |
| 1980 |
;;; void sincos (double R18, double *R16, double *R14); |
| 1981 |
_DEFUN __sincos |
| 1982 |
DALIAS sincos |
| 1983 |
LALIAS sincosl |
| 1984 |
|
| 1985 |
#define n_pushed 4 |
| 1986 |
#define n_frame (2 * F7_SIZEOF) |
| 1987 |
do_prologue_saves n_pushed, n_frame |
| 1988 |
;; Y = FramePointer + 1 |
| 1989 |
adiw Y, 1 |
| 1990 |
;; R16 = frame-arg 1 |
| 1991 |
wmov r16, Y |
| 1992 |
;; The double argument is in R18[]. |
| 1993 |
XCALL F7_NAME (set_double_impl) |
| 1994 |
;; void f7_sincos (f7_t *ss, f7_t *cc, const f7_t *aa) |
| 1995 |
;; Note that aa may equal ss or cc. |
| 1996 |
wmov r20, r16 ; aa |
| 1997 |
wmov r24, r16 ; ss = FP + 1 |
| 1998 |
subi r16, lo8(-F7_SIZEOF) |
| 1999 |
sbci r17, hi8(-F7_SIZEOF) |
| 2000 |
wmov r22, r16 ; cc = FP + 1 + F7_SIZEOF |
| 2001 |
XCALL F7_NAME (sincos) |
| 2002 |
|
| 2003 |
;; double R18 = get_double (cc) |
| 2004 |
wmov r24, r16 |
| 2005 |
XCALL F7_NAME (get_double) |
| 2006 |
wmov XL, r14 ; double *pcos |
| 2007 |
rcall store.r18.X ; *pcos = R18 |
| 2008 |
|
| 2009 |
;; double R18 = get_double (ss) |
| 2010 |
wmov r24, Y |
| 2011 |
XCALL F7_NAME (get_double) |
| 2012 |
ldd XL, Y + n_frame + 3 ; Saved R16 |
| 2013 |
ldd XH, Y + n_frame + 2 ; Saved R17 |
| 2014 |
rcall store.r18.X ; *psin = R18 |
| 2015 |
|
| 2016 |
do_epilogue_restores n_pushed, n_frame |
| 2017 |
|
| 2018 |
store.r18.X: |
| 2019 |
st X+, r18 |
| 2020 |
st X+, r19 |
| 2021 |
st X+, r20 |
| 2022 |
st X+, r21 |
| 2023 |
st X+, r22 |
| 2024 |
st X+, r23 |
| 2025 |
st X+, r24 |
| 2026 |
st X+, r25 |
| 2027 |
ret |
| 2028 |
_ENDF __sincos |
| 2029 |
#endif /* F7MOD_D_sincos_ */ |
| 2030 |
|
| 2031 |
#ifdef F7MOD_call_dd_ |
| 2032 |
|
| 2033 |
;; Provide double wrappers for functions that operate on f7_t and get f7_t*. |
| 2034 |
;; |
| 2035 |
;; We set up a frame of sizeof(f7_t), convert the input double in R18[] to |
| 2036 |
;; f7_t in that frame location, then call *Z and finally convert the result f7_t |
| 2037 |
;; to double R18[] if that's requested. |
| 2038 |
;; |
| 2039 |
;; call_dd: double func (double A) |
| 2040 |
;; void (*Z) (f7_t *aa, const f7_t *aa) |
| 2041 |
;; |
| 2042 |
;; call_dx: double func (type_t A) , sizeof(type_t) <= 4 |
| 2043 |
;; void (*Z) (f7_t *aa, type_t) |
| 2044 |
;; |
| 2045 |
;; call_xd: type_t func (double A) |
| 2046 |
;; type_t (*Z) (const f7_t *aa) |
| 2047 |
;; |
| 2048 |
;; call_ddx: double func (double A, word_t) , sizeof (word_t) <= 2 |
| 2049 |
;; void (*Z) (f7_t *aa, const f7_t *aa, word_t) |
| 2050 |
|
| 2051 |
#define WHAT R13 |
| 2052 |
|
| 2053 |
DEFUN call_dd ; WHAT = R13 = 3 |
| 2054 |
inc ZERO |
| 2055 |
LABEL call_xd ; WHAT = R13 = 2 |
| 2056 |
inc ZERO |
| 2057 |
LABEL call_ddx ; WHAT = R13 = 1 |
| 2058 |
inc ZERO |
| 2059 |
LABEL call_dx ; WHAT = R13 = 0 |
| 2060 |
push WHAT |
| 2061 |
mov WHAT, ZERO |
| 2062 |
clr ZERO |
| 2063 |
;; R14/R15 hold Z, the address of the f7_worker function, until we need it. |
| 2064 |
push r14 |
| 2065 |
push r15 |
| 2066 |
wmov r14, Z |
| 2067 |
|
| 2068 |
#define n_pushed 4 |
| 2069 |
#define n_frame F7_SIZEOF |
| 2070 |
|
| 2071 |
do_prologue_saves n_pushed, n_frame |
| 2072 |
;; Y = FramePointer + 1 |
| 2073 |
adiw Y, 1 |
| 2074 |
dec WHAT |
| 2075 |
brmi .Ldx ; WHAT was initially 0. |
| 2076 |
;; FP + 1 = (f7_t) arg1 |
| 2077 |
wmov r16, Y |
| 2078 |
;; The double argument is in R18[]. |
| 2079 |
XCALL F7_NAME (set_double_impl) |
| 2080 |
tst WHAT |
| 2081 |
brne .Lno.ddx ; WHAT was initially != 1. |
| 2082 |
;; call_ddx: Set R20/21 to the 2-byte scalar / pointer argument. |
| 2083 |
;; Fetch it from where prologue_saves put it. |
| 2084 |
ldd r20, Y + n_frame + 3 ; Saved R16 |
| 2085 |
ldd r21, Y + n_frame + 2 ; Saved R17 |
| 2086 |
.Lno.ddx: |
| 2087 |
wmov r22, Y ; &arg1 (input) |
| 2088 |
.Ldo.dx: |
| 2089 |
wmov r24, Y ; &arg1 (output) |
| 2090 |
wmov Z, r14 |
| 2091 |
XICALL |
| 2092 |
dec WHAT |
| 2093 |
breq .Lepilogue ; WHAT was initially 2: Return non-double. |
| 2094 |
wmov r24, Y ; &arg1 |
| 2095 |
XCALL F7_NAME (get_double) |
| 2096 |
.Lepilogue: |
| 2097 |
;; + 3 to account for R13...R15 pushed prior to do_prologue_saves. |
| 2098 |
do_epilogue_restores n_pushed + 3, n_frame |
| 2099 |
|
| 2100 |
.Ldx: |
| 2101 |
;; call_dx: Copy the 4-byte input scalar from R22[4] to R20[4]. |
| 2102 |
wmov r20, r22 |
| 2103 |
wmov r22, r24 |
| 2104 |
rjmp .Ldo.dx |
| 2105 |
|
| 2106 |
ENDF call_dd |
| 2107 |
#endif /* F7MOD_call_dd_ */ |
| 2108 |
|
| 2109 |
|
| 2110 |
#ifdef F7MOD_call_ddd_ |
| 2111 |
|
| 2112 |
;; Provide double wrappers for functions that operate on f7_t and get f7_t*. |
| 2113 |
;; |
| 2114 |
;; We set up a frame of 2 * sizeof(f7_t), convert the input doubles in R18[] |
| 2115 |
;; and R10[] to f7_t in these frame locations, then call *Z and finally |
| 2116 |
;; convert the result f7_t to double R18[] if that's requested. |
| 2117 |
;; |
| 2118 |
;; call_ddd: double func (double A, double B) |
| 2119 |
;; void (*Z) (f7_t *aa, const f7_t *aa, const f7_t *bb) |
| 2120 |
;; |
| 2121 |
;; call_xdd: type_t func (double A, double B) |
| 2122 |
;; type_t (*Z) (const f7_t *aa, const f7_t *bb) |
| 2123 |
|
| 2124 |
DEFUN call_ddd |
| 2125 |
inc ZERO |
| 2126 |
LABEL call_xdd |
| 2127 |
;; R8/R9 hold Z, the address of the f7_worker function, until we need it. |
| 2128 |
push r9 |
| 2129 |
push r8 |
| 2130 |
wmov r8, Z |
| 2131 |
;; This is an argument to call.2 and will be accessed by the arg pointer. |
| 2132 |
push ZERO |
| 2133 |
clr ZERO |
| 2134 |
rcall call.2 |
| 2135 |
pop TMP |
| 2136 |
pop r8 |
| 2137 |
pop r9 |
| 2138 |
ret |
| 2139 |
|
| 2140 |
#define n_pushed 4 |
| 2141 |
#define n_frame (2 * F7_SIZEOF) |
| 2142 |
|
| 2143 |
call.2: |
| 2144 |
do_prologue_saves n_pushed, n_frame |
| 2145 |
;; Y = FramePointer + 1 |
| 2146 |
adiw Y, 1 |
| 2147 |
;; FP + 1 = (f7_t) arg1 |
| 2148 |
wmov r16, Y |
| 2149 |
;; First double argument is already in R18[]. |
| 2150 |
XCALL F7_NAME (set_double_impl) |
| 2151 |
;; FP + 11 = (f7_t) arg2 |
| 2152 |
subi r16, lo8 (-F7_SIZEOF) |
| 2153 |
sbci r17, hi8 (-F7_SIZEOF) |
| 2154 |
;; Move second double argument to R18[]. |
| 2155 |
wmov r18, r10 |
| 2156 |
wmov r20, r12 |
| 2157 |
wmov r22, r14 |
| 2158 |
;; Get high word of arg2 from where prologue_saves put it. |
| 2159 |
ldd r24, Y + n_frame + 3 ; Saved R16 |
| 2160 |
ldd r25, Y + n_frame + 2 ; Saved R17 |
| 2161 |
XCALL F7_NAME (set_double_impl) |
| 2162 |
;; Z (f7_t *arg1, const f7_t *arg1, const f7_t *arg2) |
| 2163 |
wmov Z, r8 |
| 2164 |
wmov r24, Y ; &arg1 |
| 2165 |
;; WHAT == 0 => call_xdd |
| 2166 |
;; WHAT != 0 => call_ddd |
| 2167 |
ldd TMP, Y + n_frame + n_pushed + PC_SIZE |
| 2168 |
tst TMP |
| 2169 |
breq .Lxdd |
| 2170 |
wmov r22, Y ; &arg1 |
| 2171 |
wmov r20, r16 ; &arg2 |
| 2172 |
XICALL |
| 2173 |
wmov r24, Y ; &arg1 |
| 2174 |
XCALL F7_NAME (get_double) |
| 2175 |
.Lepilogue: |
| 2176 |
do_epilogue_restores n_pushed, n_frame |
| 2177 |
.Lxdd: |
| 2178 |
wmov r22, r16 ; &arg2 |
| 2179 |
XICALL |
| 2180 |
rjmp .Lepilogue |
| 2181 |
ENDF call_ddd |
| 2182 |
#endif /* F7MOD_call_ddd_ */ |
| 2183 |
|
| 2184 |
#include "f7-wraps.h" |
| 2185 |
|
| 2186 |
;;; Some additional, singular wraps that don't match any pattern. |
| 2187 |
|
| 2188 |
;; double __powidf2 (double, int) ; __builtin_powi |
| 2189 |
#ifdef F7MOD_D_powi_ |
| 2190 |
_DEFUN __powidf2 |
| 2191 |
.global F7_NAME(powi) |
| 2192 |
ldi ZH, hi8(gs(F7_NAME(powi))) |
| 2193 |
ldi ZL, lo8(gs(F7_NAME(powi))) |
| 2194 |
F7jmp call_ddx |
| 2195 |
_ENDF __powidf2 |
| 2196 |
#endif /* F7MOD_D_powi_ */ |
| 2197 |
|
| 2198 |
|
| 2199 |
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
| 2200 |
;;; Fixed-point -> double conversions. |
| 2201 |
|
| 2202 |
;;; The double exponent starts at bit 52 since the encoded mantissa has 52 bits. |
| 2203 |
;;; Note that when X is a multiple of 16, then dex_lo(x) evaluates to 0. |
| 2204 |
#define DEX16(x) (x) << (52 - 48) |
| 2205 |
#define dex_lo(x) lo8 (DEX16 (x)) |
| 2206 |
#define dex_hi(x) hi8 (DEX16 (x)) |
| 2207 |
|
| 2208 |
#ifdef F7MOD_usa2D_ |
| 2209 |
_DEFUN __fractusadf |
| 2210 |
;; Convert USI to DF. |
| 2211 |
XCALL __floatunsidf |
| 2212 |
;; The MSB indicates a value of 0. |
| 2213 |
cpse r25, __zero_reg__ |
| 2214 |
;; Divide non-zero values by 2^16 in order to adjust for FBIT = 16. |
| 2215 |
subi r25, dex_hi (16) |
| 2216 |
ret |
| 2217 |
_ENDF __fractusadf |
| 2218 |
#endif /* F7MOD_usa2D_ */ |
| 2219 |
|
| 2220 |
#ifdef F7MOD_sa2D_ |
| 2221 |
_DEFUN __fractsadf |
| 2222 |
;; Convert SI to DF. |
| 2223 |
XCALL __floatsidf |
| 2224 |
;; The MSB indicates a value of 0. |
| 2225 |
tst r25 |
| 2226 |
breq 0f |
| 2227 |
;; Divide non-zero values by 2^15 in order to adjust for FBIT = 15. |
| 2228 |
subi r24, dex_lo (15) |
| 2229 |
sbci r25, dex_hi (15) |
| 2230 |
0: ret |
| 2231 |
_ENDF __fractsadf |
| 2232 |
#endif /* F7MOD_sa2D_ */ |
| 2233 |
|
| 2234 |
#ifdef F7MOD_uha2D_ |
| 2235 |
_DEFUN __fractuhadf |
| 2236 |
;; Extend UHA to USA. |
| 2237 |
clr r22 |
| 2238 |
mov r23, r24 |
| 2239 |
mov r24, r25 |
| 2240 |
clr r25 |
| 2241 |
XJMP __fractusadf |
| 2242 |
_ENDF __fractuhadf |
| 2243 |
#endif /* F7MOD_uha2D_ */ |
| 2244 |
|
| 2245 |
#ifdef F7MOD_ha2D_ |
| 2246 |
_DEFUN __fracthadf |
| 2247 |
;; Extend HA to SA. |
| 2248 |
clr r22 |
| 2249 |
mov r23, r24 |
| 2250 |
mov r24, r25 |
| 2251 |
lsl r25 |
| 2252 |
sbc r25, r25 |
| 2253 |
XJMP __fractsadf |
| 2254 |
_ENDF __fracthadf |
| 2255 |
#endif /* F7MOD_ha2D_ */ |
| 2256 |
|
| 2257 |
|
| 2258 |
#ifdef F7MOD_usq2D_ |
| 2259 |
_DEFUN __fractusqdf |
| 2260 |
;; Convert USI to DF. |
| 2261 |
XCALL __floatunsidf |
| 2262 |
;; The MSB indicates a value of 0. |
| 2263 |
cpse r25, __zero_reg__ |
| 2264 |
;; Divide non-zero values by 2^32 in order to adjust for FBIT = 32. |
| 2265 |
subi r25, dex_hi (32) |
| 2266 |
ret |
| 2267 |
_ENDF __fractusqdf |
| 2268 |
#endif /* F7MOD_usq2D_ */ |
| 2269 |
|
| 2270 |
#ifdef F7MOD_sq2D_ |
| 2271 |
_DEFUN __fractsqdf |
| 2272 |
;; Convert SI to DF. |
| 2273 |
XCALL __floatsidf |
| 2274 |
;; The MSB indicates a value of 0. |
| 2275 |
tst r25 |
| 2276 |
breq 0f |
| 2277 |
;; Divide non-zero values by 2^31 in order to adjust for FBIT = 31. |
| 2278 |
subi r24, dex_lo (31) |
| 2279 |
sbci r25, dex_hi (31) |
| 2280 |
0: ret |
| 2281 |
_ENDF __fractsqdf |
| 2282 |
#endif /* F7MOD_sq2D_ */ |
| 2283 |
|
| 2284 |
#ifdef F7MOD_uqq2D_ |
| 2285 |
_DEFUN __fractuqqdf |
| 2286 |
;; Extend UQQ to UHQ. |
| 2287 |
mov r25, r24 |
| 2288 |
clr r24 |
| 2289 |
_LABEL __fractuhqdf |
| 2290 |
;; Extend UHQ to USQ. |
| 2291 |
clr r23 |
| 2292 |
clr r22 |
| 2293 |
XJMP __fractusqdf |
| 2294 |
_ENDF __fractuqqdf |
| 2295 |
#endif /* F7MOD_uqq2D_ */ |
| 2296 |
|
| 2297 |
#ifdef F7MOD_qq2D_ |
| 2298 |
_DEFUN __fractqqdf |
| 2299 |
;; Extend QQ to HQ. |
| 2300 |
mov r25, r24 |
| 2301 |
clr r24 |
| 2302 |
_LABEL __fracthqdf |
| 2303 |
;; Extend HQ to SQ. |
| 2304 |
clr r23 |
| 2305 |
clr r22 |
| 2306 |
XJMP __fractsqdf |
| 2307 |
_ENDF __fractqqdf |
| 2308 |
#endif /* F7MOD_qq2D_ */ |
| 2309 |
|
| 2310 |
|
| 2311 |
#ifdef F7MOD_fx2D_ |
| 2312 |
;;; Convert an [un]signed 64-bit fixed point value R18[] to double. |
| 2313 |
;;; double ufx64_to_d (ufx64_t R18, uint8_t fbit R30); |
| 2314 |
;;; double sfx64_to_d (sfx64_t R18, uint8_t fbit R30); |
| 2315 |
;;; FBIT accounts for the fractional bits of the fixed point type, |
| 2316 |
;;; e.g. pass __DA_FBIT__ for long accum etc. |
| 2317 |
DEFUN ufx64_to_d |
| 2318 |
push r30 |
| 2319 |
XCALL __floatundidf |
| 2320 |
rjmp 1f |
| 2321 |
|
| 2322 |
LABEL sfx64_to_d |
| 2323 |
push r30 |
| 2324 |
XCALL __floatdidf |
| 2325 |
1: pop r30 |
| 2326 |
|
| 2327 |
;; The MSB indicates a value of 0. Notice that at this point, |
| 2328 |
;; we have a normal floating point number with a biased exponent |
| 2329 |
;; that won't underflow below. |
| 2330 |
tst r25 |
| 2331 |
breq 0f |
| 2332 |
|
| 2333 |
;; Subtract fbit from the double exponent. R30[] = fbit << 4. |
| 2334 |
swap r30 |
| 2335 |
mov r31, r30 |
| 2336 |
andi r30, lo8 (0x0ff0) |
| 2337 |
andi r31, hi8 (0x0ff0) |
| 2338 |
sub r24, r30 |
| 2339 |
sbc r25, r31 |
| 2340 |
|
| 2341 |
0: ret |
| 2342 |
ENDF ufx64_to_d |
| 2343 |
#endif /* F7MOD_fx2D_ */ |
| 2344 |
|
| 2345 |
#ifdef F7MOD_da2D_ |
| 2346 |
_DEFUN __fractdadf |
| 2347 |
ldi r30, __DA_FBIT__ |
| 2348 |
F7jmp sfx64_to_d |
| 2349 |
_ENDF __fractdadf |
| 2350 |
#endif /* F7MOD_da2D_ */ |
| 2351 |
|
| 2352 |
#ifdef F7MOD_uda2D_ |
| 2353 |
_DEFUN __fractudadf |
| 2354 |
ldi r30, __UDA_FBIT__ |
| 2355 |
F7jmp ufx64_to_d |
| 2356 |
_ENDF __fractudadf |
| 2357 |
#endif /* F7MOD_uda2D_ */ |
| 2358 |
|
| 2359 |
#ifdef F7MOD_ta2D_ |
| 2360 |
_DEFUN __fracttadf |
| 2361 |
ldi r30, __TA_FBIT__ |
| 2362 |
F7jmp sfx64_to_d |
| 2363 |
_ENDF __fracttadf |
| 2364 |
#endif /* F7MOD_ta2D_ */ |
| 2365 |
|
| 2366 |
#ifdef F7MOD_uta2D_ |
| 2367 |
_DEFUN __fractutadf |
| 2368 |
ldi r30, __UTA_FBIT__ |
| 2369 |
F7jmp ufx64_to_d |
| 2370 |
_ENDF __fractutadf |
| 2371 |
#endif /* F7MOD_uta2D_ */ |
| 2372 |
|
| 2373 |
#ifdef F7MOD_dq2D_ |
| 2374 |
_DEFUN __fractdqdf |
| 2375 |
ldi r30, __DQ_FBIT__ |
| 2376 |
F7jmp sfx64_to_d |
| 2377 |
_ENDF __fractdqdf |
| 2378 |
#endif /* F7MOD_dq2D_ */ |
| 2379 |
|
| 2380 |
#ifdef F7MOD_udq2D_ |
| 2381 |
_DEFUN __fractudqdf |
| 2382 |
ldi r30, __UDQ_FBIT__ |
| 2383 |
F7jmp ufx64_to_d |
| 2384 |
_ENDF __fractudqdf |
| 2385 |
#endif /* F7MOD_udq2D_ */ |
| 2386 |
|
| 2387 |
|
| 2388 |
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
| 2389 |
;;; Double -> fixed-point conversions. |
| 2390 |
|
| 2391 |
#ifdef F7MOD_D2qq_ |
| 2392 |
_DEFUN __fractdfqq |
| 2393 |
;; Multiply with 2^{24+7} to get a QQ result in r25. |
| 2394 |
subi r24, dex_lo (-31) |
| 2395 |
sbci r25, dex_hi (-31) |
| 2396 |
XCALL __fixdfsi |
| 2397 |
mov r24, r25 |
| 2398 |
ret |
| 2399 |
_ENDF __fractdfqq |
| 2400 |
#endif /* F7MOD_D2qq_ */ |
| 2401 |
|
| 2402 |
#ifdef F7MOD_D2uqq_ |
| 2403 |
_DEFUN __fractdfuqq |
| 2404 |
;; Multiply with 2^{24+8} to get a UQQ result in r25. |
| 2405 |
subi r25, dex_hi (-32) |
| 2406 |
XCALL __fixunsdfsi |
| 2407 |
mov r24, r25 |
| 2408 |
ret |
| 2409 |
_ENDF __fractdfuqq |
| 2410 |
#endif /* F7MOD_D2uqq_ */ |
| 2411 |
|
| 2412 |
#ifdef F7MOD_D2ha_ |
| 2413 |
_DEFUN __fractdfha |
| 2414 |
;; Multiply with 2^{16+7} to get a HA result in r25:r24 |
| 2415 |
subi r24, dex_lo (-23) |
| 2416 |
sbci r25, dex_hi (-23) |
| 2417 |
XJMP __fixdfsi |
| 2418 |
_ENDF __fractdfha |
| 2419 |
#endif /* F7MOD_D2ha_ */ |
| 2420 |
|
| 2421 |
#ifdef F7MOD_D2uha_ |
| 2422 |
_DEFUN __fractdfuha |
| 2423 |
;; Multiply with 2^24 to get a UHA result in r25:r24. |
| 2424 |
subi r24, dex_lo (-24) |
| 2425 |
sbci r25, dex_hi (-24) |
| 2426 |
XJMP __fixunsdfsi |
| 2427 |
_ENDF __fractdfuha |
| 2428 |
#endif /* F7MOD_D2uha_ */ |
| 2429 |
|
| 2430 |
#ifdef F7MOD_D2hq_ |
| 2431 |
_DEFUN __fractdfhq |
| 2432 |
;; Multiply with 2^{16+15} to get a HQ result in r25:r24 |
| 2433 |
;; resp. with 2^31 to get a SQ result in r25:r22. |
| 2434 |
_LABEL __fractdfsq |
| 2435 |
subi r24, dex_lo (-31) |
| 2436 |
sbci r25, dex_hi (-31) |
| 2437 |
XJMP __fixdfsi |
| 2438 |
_ENDF __fractdfhq |
| 2439 |
#endif /* F7MOD_D2hq_ */ |
| 2440 |
|
| 2441 |
#ifdef F7MOD_D2uhq_ |
| 2442 |
_DEFUN __fractdfuhq |
| 2443 |
;; Multiply with 2^{16+16} to get a UHQ result in r25:r24 |
| 2444 |
;; resp. with 2^32 to get a USQ result in r25:r22. |
| 2445 |
_LABEL __fractdfusq |
| 2446 |
subi r25, dex_hi (-32) |
| 2447 |
XJMP __fixunsdfsi |
| 2448 |
_ENDF __fractdfuhq |
| 2449 |
#endif /* F7MOD_D2uhq_ */ |
| 2450 |
|
| 2451 |
#ifdef F7MOD_D2sa_ |
| 2452 |
_DEFUN __fractdfsa |
| 2453 |
;; Multiply with 2^15 to get a SA result in r25:r22. |
| 2454 |
subi r24, dex_lo (-15) |
| 2455 |
sbci r25, dex_hi (-15) |
| 2456 |
XJMP __fixdfsi |
| 2457 |
_ENDF __fractdfsa |
| 2458 |
#endif /* F7MOD_D2sa_ */ |
| 2459 |
|
| 2460 |
#ifdef F7MOD_D2usa_ |
| 2461 |
_DEFUN __fractdfusa |
| 2462 |
;; Multiply with 2^16 to get a USA result in r25:r22. |
| 2463 |
subi r25, dex_hi (-16) |
| 2464 |
XJMP __fixunsdfsi |
| 2465 |
_ENDF __fractdfusa |
| 2466 |
#endif /* F7MOD_D2usa_ */ |
| 2467 |
|
| 2468 |
|
| 2469 |
#ifdef F7MOD_D2fx_ |
| 2470 |
;;; Convert double R18 to a 64-bit fixed point value. |
| 2471 |
;;; R18 d_to_fx64 (double R18, int8_t fbit R16); |
| 2472 |
;;; |FBIT| accounts for the extra fbits compared to [u]int64_t. |
| 2473 |
;;; FBIT < 0 indicates a signed conversion. |
| 2474 |
;;; FBIT > 0 indicates an unsigned conversion. |
| 2475 |
DEFUN d_to_fx64 |
| 2476 |
do_prologue_saves 4, F7_SIZEOF |
| 2477 |
|
| 2478 |
;; Y = FramePointer + 1 |
| 2479 |
adiw Y, 1 |
| 2480 |
|
| 2481 |
;; FP + 1 = (f7_t) arg |
| 2482 |
wmov r16, Y |
| 2483 |
|
| 2484 |
;; Convert double R18 to f7_t *R16. |
| 2485 |
XCALL F7_NAME (set_double_impl) |
| 2486 |
|
| 2487 |
;; Prepare for *R24 = ldexp (*R24, R22). ldexp() accounts for the |
| 2488 |
;; fractional bits of the fixed point value, i.e. we have to <<= FBIT. |
| 2489 |
wmov r24, r16 |
| 2490 |
;; Zero-extend FBIT. |
| 2491 |
clr r23 |
| 2492 |
;; R16 has been clobbered. Fetch it from where prologue_saves put it. |
| 2493 |
ldd r22, Y + F7_SIZEOF + 3 ; Saved R16 |
| 2494 |
tst r22 |
| 2495 |
brmi 1f |
| 2496 |
|
| 2497 |
;; Positive FBIT indicates an unsigned conversion. |
| 2498 |
XCALL F7_NAME (Ildexp) |
| 2499 |
XCALL F7_NAME (get_u64) |
| 2500 |
rjmp 9f |
| 2501 |
|
| 2502 |
1: ;; Negative FBIT indicates a signed conversion. |
| 2503 |
neg r22 ; FBIT := |FBIT| |
| 2504 |
XCALL F7_NAME (Ildexp) |
| 2505 |
XCALL F7_NAME (get_s64) |
| 2506 |
|
| 2507 |
9: do_epilogue_restores 4, F7_SIZEOF |
| 2508 |
ENDF d_to_fx64 |
| 2509 |
#endif /* F7MOD_D2fx_ */ |
| 2510 |
|
| 2511 |
|
| 2512 |
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
| 2513 |
;;; [u]int32_t -> double conversions. |
| 2514 |
|
| 2515 |
;; double __floatsidf (int32_t); |
| 2516 |
#ifdef F7MOD_D_floatsidf_ |
| 2517 |
_DEFUN __floatsidf |
| 2518 |
bst r25, 7 |
| 2519 |
brtc 0f |
| 2520 |
XCALL __negsi2 |
| 2521 |
0: XJMP __floatunsidf.ge0 |
| 2522 |
_ENDF __floatsidf |
| 2523 |
#endif /* F7MOD_D_floatsidf_ */ |
| 2524 |
|
| 2525 |
;; double __floatunsidf (uint32_t); |
| 2526 |
#ifdef F7MOD_D_floatunsidf_ |
| 2527 |
_DEFUN __floatunsidf |
| 2528 |
clt |
| 2529 |
_LABEL __floatunsidf.ge0 |
| 2530 |
;; Zero-extend SI at the low end. |
| 2531 |
clr r18 |
| 2532 |
clr r19 |
| 2533 |
wmov r20, r18 |
| 2534 |
;; Input is zero? |
| 2535 |
sbiw r24, 0 |
| 2536 |
sbci r23, 0 |
| 2537 |
sbci r22, 0 |
| 2538 |
breq 9f |
| 2539 |
;; No: The double exponent of 0x80000000 is 31 plus a bias of 1023. |
| 2540 |
;; Align the SI value such that the MSBit is as R25.4. |
| 2541 |
;; For each << we have to subtract 1 from the exponent, and for |
| 2542 |
;; each >> we have to add 1. Since we want the MSB in R25.4 and |
| 2543 |
;; not in R25.7, the initial exponent must be reduced by 3. |
| 2544 |
ldi Xl, dex_lo (31 + 1023 - 3) |
| 2545 |
ldi Xh, dex_hi (31 + 1023 - 3) |
| 2546 |
;; Move the MSByte to R25. |
| 2547 |
1: tst r25 |
| 2548 |
brne 2f |
| 2549 |
subi Xl, dex_lo (8) |
| 2550 |
sbci Xh, dex_hi (8) |
| 2551 |
mov r25, r24 |
| 2552 |
mov r24, r23 |
| 2553 |
mov r23, r22 |
| 2554 |
clr r22 |
| 2555 |
rjmp 1b |
| 2556 |
2: ;; Now we have R25 != 0. |
| 2557 |
cpi r25, 0x20 |
| 2558 |
brlo 3f |
| 2559 |
adiw Xl, DEX16 (1) |
| 2560 |
lsr r25 |
| 2561 |
ror r24 |
| 2562 |
ror r23 |
| 2563 |
ror r22 |
| 2564 |
ror r21 |
| 2565 |
rjmp 2b |
| 2566 |
3: cpi r25, 0x10 |
| 2567 |
brsh 4f |
| 2568 |
sbiw Xl, DEX16 (1) |
| 2569 |
lsl r22 |
| 2570 |
rol r23 |
| 2571 |
rol r24 |
| 2572 |
rol r25 |
| 2573 |
rjmp 3b |
| 2574 |
4: ;; Move the mantissa into place and clear the redundant leading 1. |
| 2575 |
cbr r25, 0x10 |
| 2576 |
mov r20, r21 |
| 2577 |
mov r21, r22 |
| 2578 |
mov r22, r23 |
| 2579 |
mov r23, r24 |
| 2580 |
mov r24, r25 |
| 2581 |
;; Insert the biased exponent. |
| 2582 |
or r24, Xl |
| 2583 |
mov r25, Xh |
| 2584 |
;; Insert the sign. |
| 2585 |
bld r25, 7 |
| 2586 |
9: ret |
| 2587 |
_ENDF __floatunsidf |
| 2588 |
#endif /* F7MOD_D_floatunsidf_ */ |
| 2589 |
|
| 2590 |
|
| 2591 |
#endif /* !AVR_TINY */ |