ieee754-sf.S 39.5 KB
Newer Older
<
bwilson's avatar
bwilson committed
1
/* IEEE-754 single-precision functions for Xtensa
jakub's avatar
jakub committed
2
   Copyright (C) 2006-2016 Free Software Foundation, Inc.
bwilson's avatar
bwilson committed
3 4 5 6 7 8
   Contributed by Bob Wilson (bwilson@tensilica.com) at Tensilica.

   This file is part of GCC.

   GCC is free software; you can redistribute it and/or modify it
   under the terms of the GNU General Public License as published by
9
   the Free Software Foundation; either version 3, or (at your option)
bwilson's avatar
bwilson committed
10 11 12 13 14 15 16
   any later version.

   GCC is distributed in the hope that it will be useful, but WITHOUT
   ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
   or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public
   License for more details.

17 18 19 20 21 22 23 24
   Under Section 7 of GPL version 3, you are granted additional
   permissions described in the GCC Runtime Library Exception, version
   3.1, as published by the Free Software Foundation.

   You should have received a copy of the GNU General Public License and
   a copy of the GCC Runtime Library Exception along with this program;
   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
   <http://www.gnu.org/licenses/>.  */
bwilson's avatar
bwilson committed
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48

#ifdef __XTENSA_EB__
#define xh a2
#define xl a3
#define yh a4
#define yl a5
#else
#define xh a3
#define xl a2
#define yh a5
#define yl a4
#endif

/*  Warning!  The branch displacements for some Xtensa branch instructions
    are quite small, and this code has been carefully laid out to keep
    branch targets in range.  If you change anything, be sure to check that
    the assembler is not relaxing anything to branch over a jump.  */

#ifdef L_negsf2

	.align	4
	.global	__negsf2
	.type	__negsf2, @function
__negsf2:
49
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
50 51
	movi	a4, 0x80000000
	xor	a2, a2, a4
52
	leaf_return
bwilson's avatar
bwilson committed
53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

#endif /* L_negsf2 */

#ifdef L_addsubsf3

	/* Addition */
__addsf3_aux:

	/* Handle NaNs and Infinities.  (This code is placed before the
	   start of the function just to keep it in range of the limited
	   branch displacements.)  */

.Ladd_xnan_or_inf:
	/* If y is neither Infinity nor NaN, return x.  */
	bnall	a3, a6, 1f
	/* If x is a NaN, return it.  Otherwise, return y.  */
	slli	a7, a2, 9
	beqz	a7, .Ladd_ynan_or_inf
71
1:	leaf_return
bwilson's avatar
bwilson committed
72 73 74 75

.Ladd_ynan_or_inf:
	/* Return y.  */
	mov	a2, a3
76
	leaf_return
bwilson's avatar
bwilson committed
77 78 79 80 81 82 83 84 85 86 87

.Ladd_opposite_signs:
	/* Operand signs differ.  Do a subtraction.  */
	slli	a7, a6, 8
	xor	a3, a3, a7
	j	.Lsub_same_sign

	.align	4
	.global	__addsf3
	.type	__addsf3, @function
__addsf3:
88
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155
	movi	a6, 0x7f800000

	/* Check if the two operands have the same sign.  */
	xor	a7, a2, a3
	bltz	a7, .Ladd_opposite_signs

.Ladd_same_sign:	
	/* Check if either exponent == 0x7f8 (i.e., NaN or Infinity).  */
	ball	a2, a6, .Ladd_xnan_or_inf
	ball	a3, a6, .Ladd_ynan_or_inf

	/* Compare the exponents.  The smaller operand will be shifted
	   right by the exponent difference and added to the larger
	   one.  */
	extui	a7, a2, 23, 9
	extui	a8, a3, 23, 9
	bltu	a7, a8, .Ladd_shiftx

.Ladd_shifty:
	/* Check if the smaller (or equal) exponent is zero.  */
	bnone	a3, a6, .Ladd_yexpzero

	/* Replace y sign/exponent with 0x008.  */
	or	a3, a3, a6
	slli	a3, a3, 8
	srli	a3, a3, 8

.Ladd_yexpdiff:
	/* Compute the exponent difference.  */
	sub	a10, a7, a8

	/* Exponent difference > 32 -- just return the bigger value.  */
	bgeui	a10, 32, 1f
	
	/* Shift y right by the exponent difference.  Any bits that are
	   shifted out of y are saved in a9 for rounding the result.  */
	ssr	a10
	movi	a9, 0
	src	a9, a3, a9
	srl	a3, a3

	/* Do the addition.  */
	add	a2, a2, a3

	/* Check if the add overflowed into the exponent.  */
	extui	a10, a2, 23, 9
	beq	a10, a7, .Ladd_round
	mov	a8, a7
	j	.Ladd_carry

.Ladd_yexpzero:
	/* y is a subnormal value.  Replace its sign/exponent with zero,
	   i.e., no implicit "1.0", and increment the apparent exponent
	   because subnormals behave as if they had the minimum (nonzero)
	   exponent.  Test for the case when both exponents are zero.  */
	slli	a3, a3, 9
	srli	a3, a3, 9
	bnone	a2, a6, .Ladd_bothexpzero
	addi	a8, a8, 1
	j	.Ladd_yexpdiff

.Ladd_bothexpzero:
	/* Both exponents are zero.  Handle this as a special case.  There
	   is no need to shift or round, and the normal code for handling
	   a carry into the exponent field will not work because it
	   assumes there is an implicit "1.0" that needs to be added.  */
	add	a2, a2, a3
156
1:	leaf_return
bwilson's avatar
bwilson committed
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197

.Ladd_xexpzero:
	/* Same as "yexpzero" except skip handling the case when both
	   exponents are zero.  */
	slli	a2, a2, 9
	srli	a2, a2, 9
	addi	a7, a7, 1
	j	.Ladd_xexpdiff

.Ladd_shiftx:
	/* Same thing as the "shifty" code, but with x and y swapped.  Also,
	   because the exponent difference is always nonzero in this version,
	   the shift sequence can use SLL and skip loading a constant zero.  */
	bnone	a2, a6, .Ladd_xexpzero

	or	a2, a2, a6
	slli	a2, a2, 8
	srli	a2, a2, 8

.Ladd_xexpdiff:
	sub	a10, a8, a7
	bgeui	a10, 32, .Ladd_returny
	
	ssr	a10
	sll	a9, a2
	srl	a2, a2

	add	a2, a2, a3

	/* Check if the add overflowed into the exponent.  */
	extui	a10, a2, 23, 9
	bne	a10, a8, .Ladd_carry

.Ladd_round:
	/* Round up if the leftover fraction is >= 1/2.  */
	bgez	a9, 1f
	addi	a2, a2, 1

	/* Check if the leftover fraction is exactly 1/2.  */
	slli	a9, a9, 1
	beqz	a9, .Ladd_exactlyhalf
198
1:	leaf_return
bwilson's avatar
bwilson committed
199 200 201

.Ladd_returny:
	mov	a2, a3
202
	leaf_return
bwilson's avatar
bwilson committed
203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239

.Ladd_carry:	
	/* The addition has overflowed into the exponent field, so the
	   value needs to be renormalized.  The mantissa of the result
	   can be recovered by subtracting the original exponent and
	   adding 0x800000 (which is the explicit "1.0" for the
	   mantissa of the non-shifted operand -- the "1.0" for the
	   shifted operand was already added).  The mantissa can then
	   be shifted right by one bit.  The explicit "1.0" of the
	   shifted mantissa then needs to be replaced by the exponent,
	   incremented by one to account for the normalizing shift.
	   It is faster to combine these operations: do the shift first
	   and combine the additions and subtractions.  If x is the
	   original exponent, the result is:
	       shifted mantissa - (x << 22) + (1 << 22) + (x << 23)
	   or:
	       shifted mantissa + ((x + 1) << 22)
	   Note that the exponent is incremented here by leaving the
	   explicit "1.0" of the mantissa in the exponent field.  */

	/* Shift x right by one bit.  Save the lsb.  */
	mov	a10, a2
	srli	a2, a2, 1

	/* See explanation above.  The original exponent is in a8.  */
	addi	a8, a8, 1
	slli	a8, a8, 22
	add	a2, a2, a8

	/* Return an Infinity if the exponent overflowed.  */
	ball	a2, a6, .Ladd_infinity
	
	/* Same thing as the "round" code except the msb of the leftover
	   fraction is bit 0 of a10, with the rest of the fraction in a9.  */
	bbci.l	a10, 0, 1f
	addi	a2, a2, 1
	beqz	a9, .Ladd_exactlyhalf
240
1:	leaf_return
bwilson's avatar
bwilson committed
241 242 243 244 245 246 247 248 249

.Ladd_infinity:
	/* Clear the mantissa.  */
	srli	a2, a2, 23
	slli	a2, a2, 23

	/* The sign bit may have been lost in a carry-out.  Put it back.  */
	slli	a8, a8, 1
	or	a2, a2, a8
250
	leaf_return
bwilson's avatar
bwilson committed
251 252 253 254 255

.Ladd_exactlyhalf:
	/* Round down to the nearest even value.  */
	srli	a2, a2, 1
	slli	a2, a2, 1
256
	leaf_return
bwilson's avatar
bwilson committed
257 258 259 260 261 262 263 264 265 266 267 268 269 270 271


	/* Subtraction */
__subsf3_aux:
	
	/* Handle NaNs and Infinities.  (This code is placed before the
	   start of the function just to keep it in range of the limited
	   branch displacements.)  */

.Lsub_xnan_or_inf:
	/* If y is neither Infinity nor NaN, return x.  */
	bnall	a3, a6, 1f
	/* Both x and y are either NaN or Inf, so the result is NaN.  */
	movi	a4, 0x400000	/* make it a quiet NaN */
	or	a2, a2, a4
272
1:	leaf_return
bwilson's avatar
bwilson committed
273 274 275 276 277

.Lsub_ynan_or_inf:
	/* Negate y and return it.  */
	slli	a7, a6, 8
	xor	a2, a3, a7
278
	leaf_return
bwilson's avatar
bwilson committed
279 280 281 282 283 284 285 286 287 288 289

.Lsub_opposite_signs:
	/* Operand signs differ.  Do an addition.  */
	slli	a7, a6, 8
	xor	a3, a3, a7
	j	.Ladd_same_sign

	.align	4
	.global	__subsf3
	.type	__subsf3, @function
__subsf3:
290
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363
	movi	a6, 0x7f800000

	/* Check if the two operands have the same sign.  */
	xor	a7, a2, a3
	bltz	a7, .Lsub_opposite_signs

.Lsub_same_sign:	
	/* Check if either exponent == 0x7f8 (i.e., NaN or Infinity).  */
	ball	a2, a6, .Lsub_xnan_or_inf
	ball	a3, a6, .Lsub_ynan_or_inf

	/* Compare the operands.  In contrast to addition, the entire
	   value matters here.  */
	extui	a7, a2, 23, 8
	extui	a8, a3, 23, 8
	bltu	a2, a3, .Lsub_xsmaller

.Lsub_ysmaller:
	/* Check if the smaller (or equal) exponent is zero.  */
	bnone	a3, a6, .Lsub_yexpzero

	/* Replace y sign/exponent with 0x008.  */
	or	a3, a3, a6
	slli	a3, a3, 8
	srli	a3, a3, 8

.Lsub_yexpdiff:
	/* Compute the exponent difference.  */
	sub	a10, a7, a8

	/* Exponent difference > 32 -- just return the bigger value.  */
	bgeui	a10, 32, 1f
	
	/* Shift y right by the exponent difference.  Any bits that are
	   shifted out of y are saved in a9 for rounding the result.  */
	ssr	a10
	movi	a9, 0
	src	a9, a3, a9
	srl	a3, a3

	sub	a2, a2, a3

	/* Subtract the leftover bits in a9 from zero and propagate any
	   borrow from a2.  */
	neg	a9, a9
	addi	a10, a2, -1
	movnez	a2, a10, a9

	/* Check if the subtract underflowed into the exponent.  */
	extui	a10, a2, 23, 8
	beq	a10, a7, .Lsub_round
	j	.Lsub_borrow

.Lsub_yexpzero:
	/* Return zero if the inputs are equal.  (For the non-subnormal
	   case, subtracting the "1.0" will cause a borrow from the exponent
	   and this case can be detected when handling the borrow.)  */
	beq	a2, a3, .Lsub_return_zero

	/* y is a subnormal value.  Replace its sign/exponent with zero,
	   i.e., no implicit "1.0".  Unless x is also a subnormal, increment
	   y's apparent exponent because subnormals behave as if they had
	   the minimum (nonzero) exponent.  */
	slli	a3, a3, 9
	srli	a3, a3, 9
	bnone	a2, a6, .Lsub_yexpdiff
	addi	a8, a8, 1
	j	.Lsub_yexpdiff

.Lsub_returny:
	/* Negate and return y.  */
	slli	a7, a6, 8
	xor	a2, a3, a7
364
1:	leaf_return
bwilson's avatar
bwilson committed
365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405

.Lsub_xsmaller:
	/* Same thing as the "ysmaller" code, but with x and y swapped and
	   with y negated.  */
	bnone	a2, a6, .Lsub_xexpzero

	or	a2, a2, a6
	slli	a2, a2, 8
	srli	a2, a2, 8

.Lsub_xexpdiff:
	sub	a10, a8, a7
	bgeui	a10, 32, .Lsub_returny
	
	ssr	a10
	movi	a9, 0
	src	a9, a2, a9
	srl	a2, a2

	/* Negate y.  */
	slli	a11, a6, 8
	xor	a3, a3, a11

	sub	a2, a3, a2

	neg	a9, a9
	addi	a10, a2, -1
	movnez	a2, a10, a9

	/* Check if the subtract underflowed into the exponent.  */
	extui	a10, a2, 23, 8
	bne	a10, a8, .Lsub_borrow

.Lsub_round:
	/* Round up if the leftover fraction is >= 1/2.  */
	bgez	a9, 1f
	addi	a2, a2, 1

	/* Check if the leftover fraction is exactly 1/2.  */
	slli	a9, a9, 1
	beqz	a9, .Lsub_exactlyhalf
406
1:	leaf_return
bwilson's avatar
bwilson committed
407 408 409 410 411 412 413 414 415 416 417 418

.Lsub_xexpzero:
	/* Same as "yexpzero".  */
	beq	a2, a3, .Lsub_return_zero
	slli	a2, a2, 9
	srli	a2, a2, 9
	bnone	a3, a6, .Lsub_xexpdiff
	addi	a7, a7, 1
	j	.Lsub_xexpdiff

.Lsub_return_zero:
	movi	a2, 0
419
	leaf_return
bwilson's avatar
bwilson committed
420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454

.Lsub_borrow:	
	/* The subtraction has underflowed into the exponent field, so the
	   value needs to be renormalized.  Shift the mantissa left as
	   needed to remove any leading zeros and adjust the exponent
	   accordingly.  If the exponent is not large enough to remove
	   all the leading zeros, the result will be a subnormal value.  */

	slli	a8, a2, 9
	beqz	a8, .Lsub_xzero
	do_nsau	a6, a8, a7, a11
	srli	a8, a8, 9
	bge	a6, a10, .Lsub_subnormal
	addi	a6, a6, 1

.Lsub_normalize_shift:
	/* Shift the mantissa (a8/a9) left by a6.  */
	ssl	a6
	src	a8, a8, a9
	sll	a9, a9

	/* Combine the shifted mantissa with the sign and exponent,
	   decrementing the exponent by a6.  (The exponent has already
	   been decremented by one due to the borrow from the subtraction,
	   but adding the mantissa will increment the exponent by one.)  */
	srli	a2, a2, 23
	sub	a2, a2, a6
	slli	a2, a2, 23
	add	a2, a2, a8
	j	.Lsub_round

.Lsub_exactlyhalf:
	/* Round down to the nearest even value.  */
	srli	a2, a2, 1
	slli	a2, a2, 1
455
	leaf_return
bwilson's avatar
bwilson committed
456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485

.Lsub_xzero:
	/* If there was a borrow from the exponent, and the mantissa and
	   guard digits are all zero, then the inputs were equal and the
	   result should be zero.  */
	beqz	a9, .Lsub_return_zero

	/* Only the guard digit is nonzero.  Shift by min(24, a10).  */
	addi	a11, a10, -24
	movi	a6, 24
	movltz	a6, a10, a11
	j	.Lsub_normalize_shift

.Lsub_subnormal:
	/* The exponent is too small to shift away all the leading zeros.
	   Set a6 to the current exponent (which has already been
	   decremented by the borrow) so that the exponent of the result
	   will be zero.  Do not add 1 to a6 in this case, because: (1)
	   adding the mantissa will not increment the exponent, so there is
	   no need to subtract anything extra from the exponent to
	   compensate, and (2) the effective exponent of a subnormal is 1
	   not 0 so the shift amount must be 1 smaller than normal. */
	mov	a6, a10
	j	.Lsub_normalize_shift

#endif /* L_addsubsf3 */

#ifdef L_mulsf3

	/* Multiplication */
486 487 488 489
#if !XCHAL_HAVE_MUL16 && !XCHAL_HAVE_MUL32 && !XCHAL_HAVE_MAC16
#define XCHAL_NO_MUL 1
#endif

490
	.literal_position
bwilson's avatar
bwilson committed
491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573
__mulsf3_aux:

	/* Handle unusual cases (zeros, subnormals, NaNs and Infinities).
	   (This code is placed before the start of the function just to
	   keep it in range of the limited branch displacements.)  */

.Lmul_xexpzero:
	/* Clear the sign bit of x.  */
	slli	a2, a2, 1
	srli	a2, a2, 1

	/* If x is zero, return zero.  */
	beqz	a2, .Lmul_return_zero

	/* Normalize x.  Adjust the exponent in a8.  */
	do_nsau	a10, a2, a11, a12
	addi	a10, a10, -8
	ssl	a10
	sll	a2, a2 
	movi	a8, 1
	sub	a8, a8, a10
	j	.Lmul_xnormalized	
	
.Lmul_yexpzero:
	/* Clear the sign bit of y.  */
	slli	a3, a3, 1
	srli	a3, a3, 1

	/* If y is zero, return zero.  */
	beqz	a3, .Lmul_return_zero

	/* Normalize y.  Adjust the exponent in a9.  */
	do_nsau	a10, a3, a11, a12
	addi	a10, a10, -8
	ssl	a10
	sll	a3, a3
	movi	a9, 1
	sub	a9, a9, a10
	j	.Lmul_ynormalized	

.Lmul_return_zero:
	/* Return zero with the appropriate sign bit.  */
	srli	a2, a7, 31
	slli	a2, a2, 31
	j	.Lmul_done

.Lmul_xnan_or_inf:
	/* If y is zero, return NaN.  */
	slli	a8, a3, 1
	bnez	a8, 1f
	movi	a4, 0x400000	/* make it a quiet NaN */
	or	a2, a2, a4
	j	.Lmul_done
1:
	/* If y is NaN, return y.  */
	bnall	a3, a6, .Lmul_returnx
	slli	a8, a3, 9
	beqz	a8, .Lmul_returnx

.Lmul_returny:
	mov	a2, a3

.Lmul_returnx:
	/* Set the sign bit and return.  */
	extui	a7, a7, 31, 1
	slli	a2, a2, 1
	ssai	1
	src	a2, a7, a2
	j	.Lmul_done

.Lmul_ynan_or_inf:
	/* If x is zero, return NaN.  */
	slli	a8, a2, 1
	bnez	a8, .Lmul_returny
	movi	a7, 0x400000	/* make it a quiet NaN */
	or	a2, a3, a7
	j	.Lmul_done

	.align	4
	.global	__mulsf3
	.type	__mulsf3, @function
__mulsf3:
#if __XTENSA_CALL0_ABI__
574
	leaf_entry sp, 32
bwilson's avatar
bwilson committed
575 576 577 578 579
	addi	sp, sp, -32
	s32i	a12, sp, 16
	s32i	a13, sp, 20
	s32i	a14, sp, 24
	s32i	a15, sp, 28
580 581 582 583 584 585
#elif XCHAL_NO_MUL
	/* This is not really a leaf function; allocate enough stack space
	   to allow CALL12s to a helper function.  */
	leaf_entry sp, 64
#else
	leaf_entry sp, 32
bwilson's avatar
bwilson committed
586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641
#endif
	movi	a6, 0x7f800000

	/* Get the sign of the result.  */
	xor	a7, a2, a3

	/* Check for NaN and infinity.  */
	ball	a2, a6, .Lmul_xnan_or_inf
	ball	a3, a6, .Lmul_ynan_or_inf

	/* Extract the exponents.  */
	extui	a8, a2, 23, 8
	extui	a9, a3, 23, 8

	beqz	a8, .Lmul_xexpzero
.Lmul_xnormalized:	
	beqz	a9, .Lmul_yexpzero
.Lmul_ynormalized:	

	/* Add the exponents.  */
	add	a8, a8, a9

	/* Replace sign/exponent fields with explicit "1.0".  */
	movi	a10, 0xffffff
	or	a2, a2, a6
	and	a2, a2, a10
	or	a3, a3, a6
	and	a3, a3, a10

	/* Multiply 32x32 to 64 bits.  The result ends up in a2/a6.  */

#if XCHAL_HAVE_MUL32_HIGH

	mull	a6, a2, a3
	muluh	a2, a2, a3

#else

	/* Break the inputs into 16-bit chunks and compute 4 32-bit partial
	   products.  These partial products are:

		0 xl * yl

		1 xl * yh
		2 xh * yl

		3 xh * yh

	   If using the Mul16 or Mul32 multiplier options, these input
	   chunks must be stored in separate registers.  For Mac16, the
	   UMUL.AA.* opcodes can specify that the inputs come from either
	   half of the registers, so there is no need to shift them out
	   ahead of time.  If there is no multiply hardware, the 16-bit
	   chunks can be extracted when setting up the arguments to the
	   separate multiply function.  */

642
#if __XTENSA_CALL0_ABI__ && XCHAL_NO_MUL
bwilson's avatar
bwilson committed
643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701
	/* Calling a separate multiply function will clobber a0 and requires
	   use of a8 as a temporary, so save those values now.  (The function
	   uses a custom ABI so nothing else needs to be saved.)  */
	s32i	a0, sp, 0
	s32i	a8, sp, 4
#endif

#if XCHAL_HAVE_MUL16 || XCHAL_HAVE_MUL32

#define a2h a4
#define a3h a5

	/* Get the high halves of the inputs into registers.  */
	srli	a2h, a2, 16
	srli	a3h, a3, 16

#define a2l a2
#define a3l a3

#if XCHAL_HAVE_MUL32 && !XCHAL_HAVE_MUL16
	/* Clear the high halves of the inputs.  This does not matter
	   for MUL16 because the high bits are ignored.  */
	extui	a2, a2, 0, 16
	extui	a3, a3, 0, 16
#endif
#endif /* MUL16 || MUL32 */


#if XCHAL_HAVE_MUL16

#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
	mul16u	dst, xreg ## xhalf, yreg ## yhalf

#elif XCHAL_HAVE_MUL32

#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
	mull	dst, xreg ## xhalf, yreg ## yhalf

#elif XCHAL_HAVE_MAC16

/* The preprocessor insists on inserting a space when concatenating after
   a period in the definition of do_mul below.  These macros are a workaround
   using underscores instead of periods when doing the concatenation.  */
#define umul_aa_ll umul.aa.ll
#define umul_aa_lh umul.aa.lh
#define umul_aa_hl umul.aa.hl
#define umul_aa_hh umul.aa.hh

#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
	umul_aa_ ## xhalf ## yhalf	xreg, yreg; \
	rsr	dst, ACCLO

#else /* no multiply hardware */
	
#define set_arg_l(dst, src) \
	extui	dst, src, 0, 16
#define set_arg_h(dst, src) \
	srli	dst, src, 16

702
#if __XTENSA_CALL0_ABI__
bwilson's avatar
bwilson committed
703 704 705 706 707
#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
	set_arg_ ## xhalf (a13, xreg); \
	set_arg_ ## yhalf (a14, yreg); \
	call0	.Lmul_mulsi3; \
	mov	dst, a12
708 709 710 711 712 713 714 715 716
#else
#define do_mul(dst, xreg, xhalf, yreg, yhalf) \
	set_arg_ ## xhalf (a14, xreg); \
	set_arg_ ## yhalf (a15, yreg); \
	call12	.Lmul_mulsi3; \
	mov	dst, a14
#endif /* __XTENSA_CALL0_ABI__ */

#endif /* no multiply hardware */
bwilson's avatar
bwilson committed
717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741

	/* Add pp1 and pp2 into a6 with carry-out in a9.  */
	do_mul(a6, a2, l, a3, h)	/* pp 1 */
	do_mul(a11, a2, h, a3, l)	/* pp 2 */
	movi	a9, 0
	add	a6, a6, a11
	bgeu	a6, a11, 1f
	addi	a9, a9, 1
1:
	/* Shift the high half of a9/a6 into position in a9.  Note that
	   this value can be safely incremented without any carry-outs.  */
	ssai	16
	src	a9, a9, a6

	/* Compute the low word into a6.  */
	do_mul(a11, a2, l, a3, l)	/* pp 0 */
	sll	a6, a6
	add	a6, a6, a11
	bgeu	a6, a11, 1f
	addi	a9, a9, 1
1:
	/* Compute the high word into a2.  */
	do_mul(a2, a2, h, a3, h)	/* pp 3 */
	add	a2, a2, a9
	
742
#if __XTENSA_CALL0_ABI__ && XCHAL_NO_MUL
bwilson's avatar
bwilson committed
743 744 745 746
	/* Restore values saved on the stack during the multiplication.  */
	l32i	a0, sp, 0
	l32i	a8, sp, 4
#endif
747
#endif /* ! XCHAL_HAVE_MUL32_HIGH */
bwilson's avatar
bwilson committed
748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797

	/* Shift left by 9 bits, unless there was a carry-out from the
	   multiply, in which case, shift by 8 bits and increment the
	   exponent.  */
	movi	a4, 9
	srli	a5, a2, 24 - 9
	beqz	a5, 1f
	addi	a4, a4, -1
	addi	a8, a8, 1
1:	ssl	a4
	src	a2, a2, a6
	sll	a6, a6

	/* Subtract the extra bias from the exponent sum (plus one to account
	   for the explicit "1.0" of the mantissa that will be added to the
	   exponent in the final result).  */
	movi	a4, 0x80
	sub	a8, a8, a4
	
	/* Check for over/underflow.  The value in a8 is one less than the
	   final exponent, so values in the range 0..fd are OK here.  */
	movi	a4, 0xfe
	bgeu	a8, a4, .Lmul_overflow
	
.Lmul_round:
	/* Round.  */
	bgez	a6, .Lmul_rounded
	addi	a2, a2, 1
	slli	a6, a6, 1
	beqz	a6, .Lmul_exactlyhalf

.Lmul_rounded:
	/* Add the exponent to the mantissa.  */
	slli	a8, a8, 23
	add	a2, a2, a8

.Lmul_addsign:
	/* Add the sign bit.  */
	srli	a7, a7, 31
	slli	a7, a7, 31
	or	a2, a2, a7

.Lmul_done:
#if __XTENSA_CALL0_ABI__
	l32i	a12, sp, 16
	l32i	a13, sp, 20
	l32i	a14, sp, 24
	l32i	a15, sp, 28
	addi	sp, sp, 32
#endif
798
	leaf_return
bwilson's avatar
bwilson committed
799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842

.Lmul_exactlyhalf:
	/* Round down to the nearest even value.  */
	srli	a2, a2, 1
	slli	a2, a2, 1
	j	.Lmul_rounded

.Lmul_overflow:
	bltz	a8, .Lmul_underflow
	/* Return +/- Infinity.  */
	movi	a8, 0xff
	slli	a2, a8, 23
	j	.Lmul_addsign

.Lmul_underflow:
	/* Create a subnormal value, where the exponent field contains zero,
	   but the effective exponent is 1.  The value of a8 is one less than
	   the actual exponent, so just negate it to get the shift amount.  */
	neg	a8, a8
	mov	a9, a6
	ssr	a8
	bgeui	a8, 32, .Lmul_flush_to_zero
	
	/* Shift a2 right.  Any bits that are shifted out of a2 are saved
	   in a6 (combined with the shifted-out bits currently in a6) for
	   rounding the result.  */
	sll	a6, a2
	srl	a2, a2

	/* Set the exponent to zero.  */
	movi	a8, 0

	/* Pack any nonzero bits shifted out into a6.  */
	beqz	a9, .Lmul_round
	movi	a9, 1
	or	a6, a6, a9
	j	.Lmul_round
	
.Lmul_flush_to_zero:
	/* Return zero with the appropriate sign bit.  */
	srli	a2, a7, 31
	slli	a2, a2, 31
	j	.Lmul_done

843
#if XCHAL_NO_MUL
bwilson's avatar
bwilson committed
844 845 846
	
	/* For Xtensa processors with no multiply hardware, this simplified
	   version of _mulsi3 is used for multiplying 16-bit chunks of
847 848 849
	   the floating-point mantissas.  When using CALL0, this function
	   uses a custom ABI: the inputs are passed in a13 and a14, the
	   result is returned in a12, and a8 and a15 are clobbered.  */
bwilson's avatar
bwilson committed
850 851
	.align	4
.Lmul_mulsi3:
852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883
	leaf_entry sp, 16
	.macro mul_mulsi3_body dst, src1, src2, tmp1, tmp2
	movi	\dst, 0
1:	add	\tmp1, \src2, \dst
	extui	\tmp2, \src1, 0, 1
	movnez	\dst, \tmp1, \tmp2

	do_addx2 \tmp1, \src2, \dst, \tmp1
	extui	\tmp2, \src1, 1, 1
	movnez	\dst, \tmp1, \tmp2

	do_addx4 \tmp1, \src2, \dst, \tmp1
	extui	\tmp2, \src1, 2, 1
	movnez	\dst, \tmp1, \tmp2

	do_addx8 \tmp1, \src2, \dst, \tmp1
	extui	\tmp2, \src1, 3, 1
	movnez	\dst, \tmp1, \tmp2

	srli	\src1, \src1, 4
	slli	\src2, \src2, 4
	bnez	\src1, 1b
	.endm
#if __XTENSA_CALL0_ABI__
	mul_mulsi3_body a12, a13, a14, a15, a8
#else
	/* The result will be written into a2, so save that argument in a4.  */
	mov	a4, a2
	mul_mulsi3_body a2, a4, a3, a5, a6
#endif
	leaf_return
#endif /* XCHAL_NO_MUL */
bwilson's avatar
bwilson committed
884 885 886 887 888
#endif /* L_mulsf3 */

#ifdef L_divsf3

	/* Division */
889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933

#if XCHAL_HAVE_FP_DIV

	.align	4
	.global	__divsf3
	.type	__divsf3, @function
__divsf3:
	leaf_entry	sp, 16

	wfr		f1, a2	/* dividend */
	wfr		f2, a3	/* divisor */

	div0.s		f3, f2
	nexp01.s	f4, f2
	const.s		f5, 1
	maddn.s		f5, f4, f3
	mov.s		f6, f3
	mov.s		f7, f2
	nexp01.s	f2, f1
	maddn.s		f6, f5, f6
	const.s		f5, 1
	const.s		f0, 0
	neg.s		f8, f2
	maddn.s		f5, f4, f6
	maddn.s		f0, f8, f3
	mkdadj.s	f7, f1
	maddn.s		f6, f5, f6
	maddn.s		f8, f4, f0
	const.s		f3, 1
	maddn.s		f3, f4, f6
	maddn.s		f0, f8, f6
	neg.s		f2, f2
	maddn.s		f6, f3, f6
	maddn.s		f2, f4, f0
	addexpm.s	f0, f7
	addexp.s	f6, f7
	divn.s		f0, f2, f6

	rfr		a2, f0

	leaf_return

#else

	.literal_position
bwilson's avatar
bwilson committed
934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966
__divsf3_aux:

	/* Handle unusual cases (zeros, subnormals, NaNs and Infinities).
	   (This code is placed before the start of the function just to
	   keep it in range of the limited branch displacements.)  */

.Ldiv_yexpzero:
	/* Clear the sign bit of y.  */
	slli	a3, a3, 1
	srli	a3, a3, 1

	/* Check for division by zero.  */
	beqz	a3, .Ldiv_yzero

	/* Normalize y.  Adjust the exponent in a9.  */
	do_nsau	a10, a3, a4, a5
	addi	a10, a10, -8
	ssl	a10
	sll	a3, a3
	movi	a9, 1
	sub	a9, a9, a10
	j	.Ldiv_ynormalized	

.Ldiv_yzero:
	/* y is zero.  Return NaN if x is also zero; otherwise, infinity.  */
	slli	a4, a2, 1
	srli	a4, a4, 1
	srli	a2, a7, 31
	slli	a2, a2, 31
	or	a2, a2, a6
	bnez	a4, 1f
	movi	a4, 0x400000	/* make it a quiet NaN */
	or	a2, a2, a4
967
1:	leaf_return
bwilson's avatar
bwilson committed
968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989

.Ldiv_xexpzero:
	/* Clear the sign bit of x.  */
	slli	a2, a2, 1
	srli	a2, a2, 1

	/* If x is zero, return zero.  */
	beqz	a2, .Ldiv_return_zero

	/* Normalize x.  Adjust the exponent in a8.  */
	do_nsau	a10, a2, a4, a5
	addi	a10, a10, -8
	ssl	a10
	sll	a2, a2
	movi	a8, 1
	sub	a8, a8, a10
	j	.Ldiv_xnormalized	
	
.Ldiv_return_zero:
	/* Return zero with the appropriate sign bit.  */
	srli	a2, a7, 31
	slli	a2, a2, 31
990
	leaf_return
bwilson's avatar
bwilson committed
991 992 993 994 995 996 997 998 999 1000

.Ldiv_xnan_or_inf:
	/* Set the sign bit of the result.  */
	srli	a7, a3, 31
	slli	a7, a7, 31
	xor	a2, a2, a7
	/* If y is NaN or Inf, return NaN.  */
	bnall	a3, a6, 1f
	movi	a4, 0x400000	/* make it a quiet NaN */
	or	a2, a2, a4
1001
1:	leaf_return
bwilson's avatar
bwilson committed
1002 1003 1004 1005 1006 1007 1008

.Ldiv_ynan_or_inf:
	/* If y is Infinity, return zero.  */
	slli	a8, a3, 9
	beqz	a8, .Ldiv_return_zero
	/* y is NaN; return it.  */
	mov	a2, a3
1009
	leaf_return
bwilson's avatar
bwilson committed
1010 1011 1012 1013 1014

	.align	4
	.global	__divsf3
	.type	__divsf3, @function
__divsf3:
1015
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107
	movi	a6, 0x7f800000

	/* Get the sign of the result.  */
	xor	a7, a2, a3

	/* Check for NaN and infinity.  */
	ball	a2, a6, .Ldiv_xnan_or_inf
	ball	a3, a6, .Ldiv_ynan_or_inf

	/* Extract the exponents.  */
	extui	a8, a2, 23, 8
	extui	a9, a3, 23, 8

	beqz	a9, .Ldiv_yexpzero
.Ldiv_ynormalized:	
	beqz	a8, .Ldiv_xexpzero
.Ldiv_xnormalized:	

	/* Subtract the exponents.  */
	sub	a8, a8, a9

	/* Replace sign/exponent fields with explicit "1.0".  */
	movi	a10, 0xffffff
	or	a2, a2, a6
	and	a2, a2, a10
	or	a3, a3, a6
	and	a3, a3, a10

	/* The first digit of the mantissa division must be a one.
	   Shift x (and adjust the exponent) as needed to make this true.  */
	bltu	a3, a2, 1f
	slli	a2, a2, 1
	addi	a8, a8, -1
1:
	/* Do the first subtraction and shift.  */
	sub	a2, a2, a3
	slli	a2, a2, 1

	/* Put the quotient into a10.  */
	movi	a10, 1

	/* Divide one bit at a time for 23 bits.  */
	movi	a9, 23
#if XCHAL_HAVE_LOOPS
	loop	a9, .Ldiv_loopend
#endif
.Ldiv_loop:
	/* Shift the quotient << 1.  */
	slli	a10, a10, 1

	/* Is this digit a 0 or 1?  */
	bltu	a2, a3, 1f

	/* Output a 1 and subtract.  */
	addi	a10, a10, 1
	sub	a2, a2, a3

	/* Shift the dividend << 1.  */
1:	slli	a2, a2, 1

#if !XCHAL_HAVE_LOOPS
	addi	a9, a9, -1
	bnez	a9, .Ldiv_loop
#endif
.Ldiv_loopend:

	/* Add the exponent bias (less one to account for the explicit "1.0"
	   of the mantissa that will be added to the exponent in the final
	   result).  */
	addi	a8, a8, 0x7e
	
	/* Check for over/underflow.  The value in a8 is one less than the
	   final exponent, so values in the range 0..fd are OK here.  */
	movi	a4, 0xfe
	bgeu	a8, a4, .Ldiv_overflow
	
.Ldiv_round:
	/* Round.  The remainder (<< 1) is in a2.  */
	bltu	a2, a3, .Ldiv_rounded
	addi	a10, a10, 1
	beq	a2, a3, .Ldiv_exactlyhalf

.Ldiv_rounded:
	/* Add the exponent to the mantissa.  */
	slli	a8, a8, 23
	add	a2, a10, a8

.Ldiv_addsign:
	/* Add the sign bit.  */
	srli	a7, a7, 31
	slli	a7, a7, 31
	or	a2, a2, a7
1108
	leaf_return
bwilson's avatar
bwilson committed
1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156

.Ldiv_overflow:
	bltz	a8, .Ldiv_underflow
	/* Return +/- Infinity.  */
	addi	a8, a4, 1	/* 0xff */
	slli	a2, a8, 23
	j	.Ldiv_addsign

.Ldiv_exactlyhalf:
	/* Remainder is exactly half the divisor.  Round even.  */
	srli	a10, a10, 1
	slli	a10, a10, 1
	j	.Ldiv_rounded

.Ldiv_underflow:
	/* Create a subnormal value, where the exponent field contains zero,
	   but the effective exponent is 1.  The value of a8 is one less than
	   the actual exponent, so just negate it to get the shift amount.  */
	neg	a8, a8
	ssr	a8
	bgeui	a8, 32, .Ldiv_flush_to_zero
	
	/* Shift a10 right.  Any bits that are shifted out of a10 are
	   saved in a6 for rounding the result.  */
	sll	a6, a10
	srl	a10, a10

	/* Set the exponent to zero.  */
	movi	a8, 0

	/* Pack any nonzero remainder (in a2) into a6.  */
	beqz	a2, 1f
	movi	a9, 1
	or	a6, a6, a9
	
	/* Round a10 based on the bits shifted out into a6.  */
1:	bgez	a6, .Ldiv_rounded
	addi	a10, a10, 1
	slli	a6, a6, 1
	bnez	a6, .Ldiv_rounded
	srli	a10, a10, 1
	slli	a10, a10, 1
	j	.Ldiv_rounded

.Ldiv_flush_to_zero:
	/* Return zero with the appropriate sign bit.  */
	srli	a2, a7, 31
	slli	a2, a2, 31
1157
	leaf_return
bwilson's avatar
bwilson committed
1158

1159 1160
#endif /* XCHAL_HAVE_FP_DIV */

bwilson's avatar
bwilson committed
1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172
#endif /* L_divsf3 */

#ifdef L_cmpsf2

	/* Equal and Not Equal */

	.align	4
	.global	__eqsf2
	.global	__nesf2
	.set	__nesf2, __eqsf2
	.type	__eqsf2, @function
__eqsf2:
1173
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1174 1175 1176 1177 1178 1179 1180 1181
	bne	a2, a3, 4f

	/* The values are equal but NaN != NaN.  Check the exponent.  */
	movi	a6, 0x7f800000
	ball	a2, a6, 3f

	/* Equal.  */
	movi	a2, 0
1182
	leaf_return
bwilson's avatar
bwilson committed
1183 1184 1185

	/* Not equal.  */
2:	movi	a2, 1
1186
	leaf_return
bwilson's avatar
bwilson committed
1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200

	/* Check if the mantissas are nonzero.  */
3:	slli	a7, a2, 9
	j	5f

	/* Check if x and y are zero with different signs.  */
4:	or	a7, a2, a3
	slli	a7, a7, 1

	/* Equal if a7 == 0, where a7 is either abs(x | y) or the mantissa
	   or x when exponent(x) = 0x7f8 and x == y.  */
5:	movi	a2, 0
	movi	a3, 1
	movnez	a2, a3, a7	
1201
	leaf_return
bwilson's avatar
bwilson committed
1202 1203 1204 1205 1206 1207 1208 1209


	/* Greater Than */

	.align	4
	.global	__gtsf2
	.type	__gtsf2, @function
__gtsf2:
1210
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1211 1212 1213 1214 1215 1216 1217 1218
	movi	a6, 0x7f800000
	ball	a2, a6, 2f
1:	bnall	a3, a6, .Lle_cmp

	/* Check if y is a NaN.  */
	slli	a7, a3, 9
	beqz	a7, .Lle_cmp
	movi	a2, 0
1219
	leaf_return
bwilson's avatar
bwilson committed
1220 1221 1222 1223 1224

	/* Check if x is a NaN.  */
2:	slli	a7, a2, 9
	beqz	a7, 1b
	movi	a2, 0
1225
	leaf_return
bwilson's avatar
bwilson committed
1226 1227 1228 1229 1230 1231 1232 1233


	/* Less Than or Equal */

	.align	4
	.global	__lesf2
	.type	__lesf2, @function
__lesf2:
1234
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1235 1236 1237 1238 1239 1240 1241 1242
	movi	a6, 0x7f800000
	ball	a2, a6, 2f
1:	bnall	a3, a6, .Lle_cmp

	/* Check if y is a NaN.  */
	slli	a7, a3, 9
	beqz	a7, .Lle_cmp
	movi	a2, 1
1243
	leaf_return
bwilson's avatar
bwilson committed
1244 1245 1246 1247 1248

	/* Check if x is a NaN.  */
2:	slli	a7, a2, 9
	beqz	a7, 1b
	movi	a2, 1
1249
	leaf_return
bwilson's avatar
bwilson committed
1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261

.Lle_cmp:
	/* Check if x and y have different signs.  */
	xor	a7, a2, a3
	bltz	a7, .Lle_diff_signs

	/* Check if x is negative.  */
	bltz	a2, .Lle_xneg

	/* Check if x <= y.  */
	bltu	a3, a2, 5f
4:	movi	a2, 0
1262
	leaf_return
bwilson's avatar
bwilson committed
1263 1264 1265 1266 1267

.Lle_xneg:
	/* Check if y <= x.  */
	bgeu	a2, a3, 4b
5:	movi	a2, 1
1268
	leaf_return
bwilson's avatar
bwilson committed
1269 1270 1271 1272 1273 1274 1275 1276 1277 1278

.Lle_diff_signs:
	bltz	a2, 4b

	/* Check if both x and y are zero.  */
	or	a7, a2, a3
	slli	a7, a7, 1
	movi	a2, 1
	movi	a3, 0
	moveqz	a2, a3, a7
1279
	leaf_return
bwilson's avatar
bwilson committed
1280 1281 1282 1283 1284 1285 1286 1287


	/* Greater Than or Equal */

	.align	4
	.global	__gesf2
	.type	__gesf2, @function
__gesf2:
1288
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1289 1290 1291 1292 1293 1294 1295 1296
	movi	a6, 0x7f800000
	ball	a2, a6, 2f
1:	bnall	a3, a6, .Llt_cmp

	/* Check if y is a NaN.  */
	slli	a7, a3, 9
	beqz	a7, .Llt_cmp
	movi	a2, -1
1297
	leaf_return
bwilson's avatar
bwilson committed
1298 1299 1300 1301 1302

	/* Check if x is a NaN.  */
2:	slli	a7, a2, 9
	beqz	a7, 1b
	movi	a2, -1
1303
	leaf_return
bwilson's avatar
bwilson committed
1304 1305 1306 1307 1308 1309 1310 1311


	/* Less Than */

	.align	4
	.global	__ltsf2
	.type	__ltsf2, @function
__ltsf2:
1312
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1313 1314 1315 1316 1317 1318 1319 1320
	movi	a6, 0x7f800000
	ball	a2, a6, 2f
1:	bnall	a3, a6, .Llt_cmp

	/* Check if y is a NaN.  */
	slli	a7, a3, 9
	beqz	a7, .Llt_cmp
	movi	a2, 0
1321
	leaf_return
bwilson's avatar
bwilson committed
1322 1323 1324 1325 1326

	/* Check if x is a NaN.  */
2:	slli	a7, a2, 9
	beqz	a7, 1b
	movi	a2, 0
1327
	leaf_return
bwilson's avatar
bwilson committed
1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339

.Llt_cmp:
	/* Check if x and y have different signs.  */
	xor	a7, a2, a3
	bltz	a7, .Llt_diff_signs

	/* Check if x is negative.  */
	bltz	a2, .Llt_xneg

	/* Check if x < y.  */
	bgeu	a2, a3, 5f
4:	movi	a2, -1
1340
	leaf_return
bwilson's avatar
bwilson committed
1341 1342 1343 1344 1345

.Llt_xneg:
	/* Check if y < x.  */
	bltu	a3, a2, 4b
5:	movi	a2, 0
1346
	leaf_return
bwilson's avatar
bwilson committed
1347 1348 1349 1350 1351 1352 1353 1354 1355 1356

.Llt_diff_signs:
	bgez	a2, 5b

	/* Check if both x and y are nonzero.  */
	or	a7, a2, a3
	slli	a7, a7, 1
	movi	a2, 0
	movi	a3, -1
	movnez	a2, a3, a7
1357
	leaf_return
bwilson's avatar
bwilson committed
1358 1359 1360 1361 1362 1363 1364 1365


	/* Unordered */

	.align	4
	.global	__unordsf2
	.type	__unordsf2, @function
__unordsf2:
1366
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1367 1368 1369 1370
	movi	a6, 0x7f800000
	ball	a2, a6, 3f
1:	ball	a3, a6, 4f
2:	movi	a2, 0
1371
	leaf_return
bwilson's avatar
bwilson committed
1372 1373 1374 1375

3:	slli	a7, a2, 9
	beqz	a7, 1b
	movi	a2, 1
1376
	leaf_return
bwilson's avatar
bwilson committed
1377 1378 1379 1380

4:	slli	a7, a3, 9
	beqz	a7, 2b
	movi	a2, 1
1381
	leaf_return
bwilson's avatar
bwilson committed
1382 1383 1384 1385 1386 1387 1388 1389 1390

#endif /* L_cmpsf2 */

#ifdef L_fixsfsi

	.align	4
	.global	__fixsfsi
	.type	__fixsfsi, @function
__fixsfsi:
1391
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413

	/* Check for NaN and Infinity.  */
	movi	a6, 0x7f800000
	ball	a2, a6, .Lfixsfsi_nan_or_inf

	/* Extract the exponent and check if 0 < (exp - 0x7e) < 32.  */
	extui	a4, a2, 23, 8
	addi	a4, a4, -0x7e
	bgei	a4, 32, .Lfixsfsi_maxint
	blti	a4, 1, .Lfixsfsi_zero

	/* Add explicit "1.0" and shift << 8.  */
	or	a7, a2, a6
	slli	a5, a7, 8

	/* Shift back to the right, based on the exponent.  */
	ssl	a4		/* shift by 32 - a4 */
	srl	a5, a5

	/* Negate the result if sign != 0.  */
	neg	a2, a5
	movgez	a2, a5, a7
1414
	leaf_return
bwilson's avatar
bwilson committed
1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428

.Lfixsfsi_nan_or_inf:
	/* Handle Infinity and NaN.  */
	slli	a4, a2, 9
	beqz	a4, .Lfixsfsi_maxint

	/* Translate NaN to +maxint.  */
	movi	a2, 0

.Lfixsfsi_maxint:
	slli	a4, a6, 8	/* 0x80000000 */
	addi	a5, a4, -1	/* 0x7fffffff */
	movgez	a4, a5, a2
	mov	a2, a4
1429
	leaf_return
bwilson's avatar
bwilson committed
1430 1431 1432

.Lfixsfsi_zero:
	movi	a2, 0
1433
	leaf_return
bwilson's avatar
bwilson committed
1434 1435 1436 1437 1438 1439 1440 1441 1442

#endif /* L_fixsfsi */

#ifdef L_fixsfdi

	.align	4
	.global	__fixsfdi
	.type	__fixsfdi, @function
__fixsfdi:
1443
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471

	/* Check for NaN and Infinity.  */
	movi	a6, 0x7f800000
	ball	a2, a6, .Lfixsfdi_nan_or_inf

	/* Extract the exponent and check if 0 < (exp - 0x7e) < 64.  */
	extui	a4, a2, 23, 8
	addi	a4, a4, -0x7e
	bgei	a4, 64, .Lfixsfdi_maxint
	blti	a4, 1, .Lfixsfdi_zero

	/* Add explicit "1.0" and shift << 8.  */
	or	a7, a2, a6
	slli	xh, a7, 8

	/* Shift back to the right, based on the exponent.  */
	ssl	a4		/* shift by 64 - a4 */
	bgei	a4, 32, .Lfixsfdi_smallshift
	srl	xl, xh
	movi	xh, 0

.Lfixsfdi_shifted:	
	/* Negate the result if sign != 0.  */
	bgez	a7, 1f
	neg	xl, xl
	neg	xh, xh
	beqz	xl, 1f
	addi	xh, xh, -1
1472
1:	leaf_return
bwilson's avatar
bwilson committed
1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492

.Lfixsfdi_smallshift:
	movi	xl, 0
	sll	xl, xh
	srl	xh, xh
	j	.Lfixsfdi_shifted

.Lfixsfdi_nan_or_inf:
	/* Handle Infinity and NaN.  */
	slli	a4, a2, 9
	beqz	a4, .Lfixsfdi_maxint

	/* Translate NaN to +maxint.  */
	movi	a2, 0

.Lfixsfdi_maxint:
	slli	a7, a6, 8	/* 0x80000000 */
	bgez	a2, 1f
	mov	xh, a7
	movi	xl, 0
1493
	leaf_return
bwilson's avatar
bwilson committed
1494 1495 1496

1:	addi	xh, a7, -1	/* 0x7fffffff */
	movi	xl, -1
1497
	leaf_return
bwilson's avatar
bwilson committed
1498 1499 1500 1501

.Lfixsfdi_zero:
	movi	xh, 0
	movi	xl, 0
1502
	leaf_return
bwilson's avatar
bwilson committed
1503 1504 1505 1506 1507 1508 1509 1510 1511

#endif /* L_fixsfdi */

#ifdef L_fixunssfsi

	.align	4
	.global	__fixunssfsi
	.type	__fixunssfsi, @function
__fixunssfsi:
1512
	leaf_entry sp, 16
bwilson's avatar
bwilson committed
1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536

	/* Check for NaN and Infinity.  */
	movi	a6, 0x7f800000
	ball	a2, a6, .Lfixunssfsi_nan_or_inf

	/* Extract the exponent and check if 0 <= (exp - 0x7f) < 32.  */
	extui	a4, a2, 23, 8
	addi	a4, a4, -0x7f
	bgei	a4, 32, .Lfixunssfsi_maxint
	bltz	a4, .Lfixunssfsi_zero

	/* Add explicit "1.0" and shift << 8.  */
	or	a7, a2, a6
	slli	a5, a7, 8

	/* Shift back to the right, based on the exponent.  */
	addi	a4, a4, 1
	beqi	a4, 32, .Lfixunssfsi_bigexp
	ssl	a4		/* shift by 32 - a4 */
	srl	a5, a5

	/* Negate the result if sign != 0.  */
	neg	a2, a5
	movgez	a2, a5, a7