1 /* 1
2 * linux/arch/arm/vfp/vfpdouble.c
3 *
4 * This code is derived in part from John R. H
5 * carries the following notice:
6 *
7 * ===========================================
8 * This C source file is part of the SoftFloat
9 * Arithmetic Package, Release 2.
10 *
11 * Written by John R. Hauser. This work was m
12 * International Computer Science Institute, l
13 * Street, Berkeley, California 94704. Fundin
14 * National Science Foundation under grant MIP
15 * of this code was written as part of a proje
16 * processor in collaboration with the Univers
17 * overseen by Profs. Nelson Morgan and John W
18 * is available through the web page `http://H
19 * arithmetic/softfloat.html'.
20 *
21 * THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FRE
22 * has been made to avoid it, THIS SOFTWARE MA
23 * TIMES RESULT IN INCORRECT BEHAVIOR. USE OF
24 * PERSONS AND ORGANIZATIONS WHO CAN AND WILL
25 * AND ALL LOSSES, COSTS, OR OTHER PROBLEMS AR
26 *
27 * Derivative works are acceptable, even for c
28 * (1) they include prominent notice that the
29 * include prominent notice akin to these thre
30 * this code that are retained.
31 * ===========================================
32 */
33 #include <linux/kernel.h>
34 #include <linux/bitops.h>
35
36 #include <asm/div64.h>
37 #include <asm/vfp.h>
38
39 #include "vfpinstr.h"
40 #include "vfp.h"
41
42 static struct vfp_double vfp_double_default_qn
43 .exponent = 2047,
44 .sign = 0,
45 .significand = VFP_DOUBLE_SIGNIFICA
46 };
47
48 static void vfp_double_dump(const char *str, s
49 {
50 pr_debug("VFP: %s: sign=%d exponent=%d
51 str, d->sign != 0, d->exponen
52 }
53
54 static void vfp_double_normalise_denormal(stru
55 {
56 int bits = 31 - fls(vd->significand >>
57 if (bits == 31)
58 bits = 63 - fls(vd->significan
59
60 vfp_double_dump("normalise_denormal: i
61
62 if (bits) {
63 vd->exponent -= bits - 1;
64 vd->significand <<= bits;
65 }
66
67 vfp_double_dump("normalise_denormal: o
68 }
69
70 u32 vfp_double_normaliseround(int dd, struct v
71 {
72 u64 significand, incr;
73 int exponent, shift, underflow;
74 u32 rmode;
75
76 vfp_double_dump("pack: in", vd);
77
78 /*
79 * Infinities and NaNs are a special c
80 */
81 if (vd->exponent == 2047 && (vd->signi
82 goto pack;
83
84 /*
85 * Special-case zero.
86 */
87 if (vd->significand == 0) {
88 vd->exponent = 0;
89 goto pack;
90 }
91
92 exponent = vd->exponent;
93 significand = vd->significand;
94
95 shift = 32 - fls(significand >> 32);
96 if (shift == 32)
97 shift = 64 - fls(significand);
98 if (shift) {
99 exponent -= shift;
100 significand <<= shift;
101 }
102
103 #ifdef DEBUG
104 vd->exponent = exponent;
105 vd->significand = significand;
106 vfp_double_dump("pack: normalised", vd
107 #endif
108
109 /*
110 * Tiny number?
111 */
112 underflow = exponent < 0;
113 if (underflow) {
114 significand = vfp_shiftright64
115 exponent = 0;
116 #ifdef DEBUG
117 vd->exponent = exponent;
118 vd->significand = significand;
119 vfp_double_dump("pack: tiny nu
120 #endif
121 if (!(significand & ((1ULL <<
122 underflow = 0;
123 }
124
125 /*
126 * Select rounding increment.
127 */
128 incr = 0;
129 rmode = fpscr & FPSCR_RMODE_MASK;
130
131 if (rmode == FPSCR_ROUND_NEAREST) {
132 incr = 1ULL << VFP_DOUBLE_LOW_
133 if ((significand & (1ULL << (V
134 incr -= 1;
135 } else if (rmode == FPSCR_ROUND_TOZERO
136 incr = 0;
137 } else if ((rmode == FPSCR_ROUND_PLUSI
138 incr = (1ULL << (VFP_DOUBLE_LO
139
140 pr_debug("VFP: rounding increment = 0x
141
142 /*
143 * Is our rounding going to overflow?
144 */
145 if ((significand + incr) < significand
146 exponent += 1;
147 significand = (significand >>
148 incr >>= 1;
149 #ifdef DEBUG
150 vd->exponent = exponent;
151 vd->significand = significand;
152 vfp_double_dump("pack: overflo
153 #endif
154 }
155
156 /*
157 * If any of the low bits (which will
158 * number) are non-zero, the result is
159 */
160 if (significand & ((1 << (VFP_DOUBLE_L
161 exceptions |= FPSCR_IXC;
162
163 /*
164 * Do our rounding.
165 */
166 significand += incr;
167
168 /*
169 * Infinity?
170 */
171 if (exponent >= 2046) {
172 exceptions |= FPSCR_OFC | FPSC
173 if (incr == 0) {
174 vd->exponent = 2045;
175 vd->significand = 0x7f
176 } else {
177 vd->exponent = 2047;
178 vd->significand = 0;
179 }
180 } else {
181 if (significand >> (VFP_DOUBLE
182 exponent = 0;
183 if (exponent || significand >
184 underflow = 0;
185 if (underflow)
186 exceptions |= FPSCR_UF
187 vd->exponent = exponent;
188 vd->significand = significand
189 }
190
191 pack:
192 vfp_double_dump("pack: final", vd);
193 {
194 s64 d = vfp_double_pack(vd);
195 pr_debug("VFP: %s: d(d%d)=%016
196 dd, d, exceptions);
197 vfp_put_double(d, dd);
198 }
199 return exceptions;
200 }
201
202 /*
203 * Propagate the NaN, setting exceptions if it
204 * 'n' is always a NaN. 'm' may be a number,
205 */
206 static u32
207 vfp_propagate_nan(struct vfp_double *vdd, stru
208 struct vfp_double *vdm, u32
209 {
210 struct vfp_double *nan;
211 int tn, tm = 0;
212
213 tn = vfp_double_type(vdn);
214
215 if (vdm)
216 tm = vfp_double_type(vdm);
217
218 if (fpscr & FPSCR_DEFAULT_NAN)
219 /*
220 * Default NaN mode - always r
221 */
222 nan = &vfp_double_default_qnan
223 else {
224 /*
225 * Contemporary mode - select
226 * NAN, or if neither are sign
227 * quiet NAN.
228 */
229 if (tn == VFP_SNAN || (tm != V
230 nan = vdn;
231 else
232 nan = vdm;
233 /*
234 * Make the NaN quiet.
235 */
236 nan->significand |= VFP_DOUBLE
237 }
238
239 *vdd = *nan;
240
241 /*
242 * If one was a signalling NAN, raise
243 */
244 return tn == VFP_SNAN || tm == VFP_SNA
245 }
246
247 /*
248 * Extended operations
249 */
250 static u32 vfp_double_fabs(int dd, int unused,
251 {
252 vfp_put_double(vfp_double_packed_abs(v
253 return 0;
254 }
255
256 static u32 vfp_double_fcpy(int dd, int unused,
257 {
258 vfp_put_double(vfp_get_double(dm), dd)
259 return 0;
260 }
261
262 static u32 vfp_double_fneg(int dd, int unused,
263 {
264 vfp_put_double(vfp_double_packed_negat
265 return 0;
266 }
267
268 static u32 vfp_double_fsqrt(int dd, int unused
269 {
270 struct vfp_double vdm, vdd;
271 int ret, tm;
272
273 vfp_double_unpack(&vdm, vfp_get_double
274 tm = vfp_double_type(&vdm);
275 if (tm & (VFP_NAN|VFP_INFINITY)) {
276 struct vfp_double *vdp = &vdd;
277
278 if (tm & VFP_NAN)
279 ret = vfp_propagate_na
280 else if (vdm.sign == 0) {
281 sqrt_copy:
282 vdp = &vdm;
283 ret = 0;
284 } else {
285 sqrt_invalid:
286 vdp = &vfp_double_defa
287 ret = FPSCR_IOC;
288 }
289 vfp_put_double(vfp_double_pack
290 return ret;
291 }
292
293 /*
294 * sqrt(+/- 0) == +/- 0
295 */
296 if (tm & VFP_ZERO)
297 goto sqrt_copy;
298
299 /*
300 * Normalise a denormalised number
301 */
302 if (tm & VFP_DENORMAL)
303 vfp_double_normalise_denormal(
304
305 /*
306 * sqrt(<0) = invalid
307 */
308 if (vdm.sign)
309 goto sqrt_invalid;
310
311 vfp_double_dump("sqrt", &vdm);
312
313 /*
314 * Estimate the square root.
315 */
316 vdd.sign = 0;
317 vdd.exponent = ((vdm.exponent - 1023)
318 vdd.significand = (u64)vfp_estimate_sq
319
320 vfp_double_dump("sqrt estimate1", &vdd
321
322 vdm.significand >>= 1 + (vdm.exponent
323 vdd.significand += 2 + vfp_estimate_di
324
325 vfp_double_dump("sqrt estimate2", &vdd
326
327 /*
328 * And now adjust.
329 */
330 if ((vdd.significand & VFP_DOUBLE_LOW_
331 if (vdd.significand < 2) {
332 vdd.significand = ~0UL
333 } else {
334 u64 termh, terml, remh
335 vdm.significand <<= 2;
336 mul64to128(&termh, &te
337 sub128(&remh, &reml, v
338 while ((s64)remh < 0)
339 vdd.significan
340 shift64left(&t
341 terml |= 1;
342 add128(&remh,
343 }
344 vdd.significand |= (re
345 }
346 }
347 vdd.significand = vfp_shiftright64jamm
348
349 return vfp_double_normaliseround(dd, &
350 }
351
352 /*
353 * Equal := ZC
354 * Less than := N
355 * Greater than := C
356 * Unordered := CV
357 */
358 static u32 vfp_compare(int dd, int signal_on_q
359 {
360 s64 d, m;
361 u32 ret = 0;
362
363 m = vfp_get_double(dm);
364 if (vfp_double_packed_exponent(m) == 2
365 ret |= FPSCR_C | FPSCR_V;
366 if (signal_on_qnan || !(vfp_do
367 /*
368 * Signalling NaN, or
369 */
370 ret |= FPSCR_IOC;
371 }
372
373 d = vfp_get_double(dd);
374 if (vfp_double_packed_exponent(d) == 2
375 ret |= FPSCR_C | FPSCR_V;
376 if (signal_on_qnan || !(vfp_do
377 /*
378 * Signalling NaN, or
379 */
380 ret |= FPSCR_IOC;
381 }
382
383 if (ret == 0) {
384 if (d == m || vfp_double_packe
385 /*
386 * equal
387 */
388 ret |= FPSCR_Z | FPSCR
389 } else if (vfp_double_packed_s
390 /*
391 * different signs
392 */
393 if (vfp_double_packed_
394 /*
395 * d is negati
396 */
397 ret |= FPSCR_N
398 else
399 /*
400 * d is positi
401 */
402 ret |= FPSCR_C
403 } else if ((vfp_double_packed_
404 /*
405 * d < m
406 */
407 ret |= FPSCR_N;
408 } else if ((vfp_double_packed_
409 /*
410 * d > m
411 */
412 ret |= FPSCR_C;
413 }
414 }
415
416 return ret;
417 }
418
419 static u32 vfp_double_fcmp(int dd, int unused,
420 {
421 return vfp_compare(dd, 0, dm, fpscr);
422 }
423
424 static u32 vfp_double_fcmpe(int dd, int unused
425 {
426 return vfp_compare(dd, 1, dm, fpscr);
427 }
428
429 static u32 vfp_double_fcmpz(int dd, int unused
430 {
431 return vfp_compare(dd, 0, VFP_REG_ZERO
432 }
433
434 static u32 vfp_double_fcmpez(int dd, int unuse
435 {
436 return vfp_compare(dd, 1, VFP_REG_ZERO
437 }
438
439 static u32 vfp_double_fcvts(int sd, int unused
440 {
441 struct vfp_double vdm;
442 struct vfp_single vsd;
443 int tm;
444 u32 exceptions = 0;
445
446 vfp_double_unpack(&vdm, vfp_get_double
447
448 tm = vfp_double_type(&vdm);
449
450 /*
451 * If we have a signalling NaN, signal
452 */
453 if (tm == VFP_SNAN)
454 exceptions = FPSCR_IOC;
455
456 if (tm & VFP_DENORMAL)
457 vfp_double_normalise_denormal(
458
459 vsd.sign = vdm.sign;
460 vsd.significand = vfp_hi64to32jamming(
461
462 /*
463 * If we have an infinity or a NaN, th
464 */
465 if (tm & (VFP_INFINITY|VFP_NAN)) {
466 vsd.exponent = 255;
467 if (tm == VFP_QNAN)
468 vsd.significand |= VFP
469 goto pack_nan;
470 } else if (tm & VFP_ZERO)
471 vsd.exponent = 0;
472 else
473 vsd.exponent = vdm.exponent -
474
475 return vfp_single_normaliseround(sd, &
476
477 pack_nan:
478 vfp_put_float(vfp_single_pack(&vsd), s
479 return exceptions;
480 }
481
482 static u32 vfp_double_fuito(int dd, int unused
483 {
484 struct vfp_double vdm;
485 u32 m = vfp_get_float(dm);
486
487 vdm.sign = 0;
488 vdm.exponent = 1023 + 63 - 1;
489 vdm.significand = (u64)m;
490
491 return vfp_double_normaliseround(dd, &
492 }
493
494 static u32 vfp_double_fsito(int dd, int unused
495 {
496 struct vfp_double vdm;
497 u32 m = vfp_get_float(dm);
498
499 vdm.sign = (m & 0x80000000) >> 16;
500 vdm.exponent = 1023 + 63 - 1;
501 vdm.significand = vdm.sign ? -m : m;
502
503 return vfp_double_normaliseround(dd, &
504 }
505
506 static u32 vfp_double_ftoui(int sd, int unused
507 {
508 struct vfp_double vdm;
509 u32 d, exceptions = 0;
510 int rmode = fpscr & FPSCR_RMODE_MASK;
511 int tm;
512
513 vfp_double_unpack(&vdm, vfp_get_double
514
515 /*
516 * Do we have a denormalised number?
517 */
518 tm = vfp_double_type(&vdm);
519 if (tm & VFP_DENORMAL)
520 exceptions |= FPSCR_IDC;
521
522 if (tm & VFP_NAN)
523 vdm.sign = 0;
524
525 if (vdm.exponent >= 1023 + 32) {
526 d = vdm.sign ? 0 : 0xffffffff;
527 exceptions = FPSCR_IOC;
528 } else if (vdm.exponent >= 1023 - 1) {
529 int shift = 1023 + 63 - vdm.ex
530 u64 rem, incr = 0;
531
532 /*
533 * 2^0 <= m < 2^32-2^8
534 */
535 d = (vdm.significand << 1) >>
536 rem = vdm.significand << (65 -
537
538 if (rmode == FPSCR_ROUND_NEARE
539 incr = 0x8000000000000
540 if ((d & 1) == 0)
541 incr -= 1;
542 } else if (rmode == FPSCR_ROUN
543 incr = 0;
544 } else if ((rmode == FPSCR_ROU
545 incr = ~0ULL;
546 }
547
548 if ((rem + incr) < rem) {
549 if (d < 0xffffffff)
550 d += 1;
551 else
552 exceptions |=
553 }
554
555 if (d && vdm.sign) {
556 d = 0;
557 exceptions |= FPSCR_IO
558 } else if (rem)
559 exceptions |= FPSCR_IX
560 } else {
561 d = 0;
562 if (vdm.exponent | vdm.signifi
563 exceptions |= FPSCR_IX
564 if (rmode == FPSCR_ROU
565 d = 1;
566 else if (rmode == FPSC
567 d = 0;
568 exceptions |=
569 }
570 }
571 }
572
573 pr_debug("VFP: ftoui: d(s%d)=%08x exce
574
575 vfp_put_float(d, sd);
576
577 return exceptions;
578 }
579
580 static u32 vfp_double_ftouiz(int sd, int unuse
581 {
582 return vfp_double_ftoui(sd, unused, dm
583 }
584
585 static u32 vfp_double_ftosi(int sd, int unused
586 {
587 struct vfp_double vdm;
588 u32 d, exceptions = 0;
589 int rmode = fpscr & FPSCR_RMODE_MASK;
590 int tm;
591
592 vfp_double_unpack(&vdm, vfp_get_double
593 vfp_double_dump("VDM", &vdm);
594
595 /*
596 * Do we have denormalised number?
597 */
598 tm = vfp_double_type(&vdm);
599 if (tm & VFP_DENORMAL)
600 exceptions |= FPSCR_IDC;
601
602 if (tm & VFP_NAN) {
603 d = 0;
604 exceptions |= FPSCR_IOC;
605 } else if (vdm.exponent >= 1023 + 32)
606 d = 0x7fffffff;
607 if (vdm.sign)
608 d = ~d;
609 exceptions |= FPSCR_IOC;
610 } else if (vdm.exponent >= 1023 - 1) {
611 int shift = 1023 + 63 - vdm.ex
612 u64 rem, incr = 0;
613
614 d = (vdm.significand << 1) >>
615 rem = vdm.significand << (65 -
616
617 if (rmode == FPSCR_ROUND_NEARE
618 incr = 0x8000000000000
619 if ((d & 1) == 0)
620 incr -= 1;
621 } else if (rmode == FPSCR_ROUN
622 incr = 0;
623 } else if ((rmode == FPSCR_ROU
624 incr = ~0ULL;
625 }
626
627 if ((rem + incr) < rem && d <
628 d += 1;
629 if (d > 0x7fffffff + (vdm.sign
630 d = 0x7fffffff + (vdm.
631 exceptions |= FPSCR_IO
632 } else if (rem)
633 exceptions |= FPSCR_IX
634
635 if (vdm.sign)
636 d = -d;
637 } else {
638 d = 0;
639 if (vdm.exponent | vdm.signifi
640 exceptions |= FPSCR_IX
641 if (rmode == FPSCR_ROU
642 d = 1;
643 else if (rmode == FPSC
644 d = -1;
645 }
646 }
647
648 pr_debug("VFP: ftosi: d(s%d)=%08x exce
649
650 vfp_put_float((s32)d, sd);
651
652 return exceptions;
653 }
654
655 static u32 vfp_double_ftosiz(int dd, int unuse
656 {
657 return vfp_double_ftosi(dd, unused, dm
658 }
659
660
661 static struct op fops_ext[32] = {
662 [FEXT_TO_IDX(FEXT_FCPY)] = { vf
663 [FEXT_TO_IDX(FEXT_FABS)] = { vf
664 [FEXT_TO_IDX(FEXT_FNEG)] = { vf
665 [FEXT_TO_IDX(FEXT_FSQRT)] = { vf
666 [FEXT_TO_IDX(FEXT_FCMP)] = { vf
667 [FEXT_TO_IDX(FEXT_FCMPE)] = { vf
668 [FEXT_TO_IDX(FEXT_FCMPZ)] = { vf
669 [FEXT_TO_IDX(FEXT_FCMPEZ)] = { vf
670 [FEXT_TO_IDX(FEXT_FCVT)] = { vf
671 [FEXT_TO_IDX(FEXT_FUITO)] = { vf
672 [FEXT_TO_IDX(FEXT_FSITO)] = { vf
673 [FEXT_TO_IDX(FEXT_FTOUI)] = { vf
674 [FEXT_TO_IDX(FEXT_FTOUIZ)] = { vf
675 [FEXT_TO_IDX(FEXT_FTOSI)] = { vf
676 [FEXT_TO_IDX(FEXT_FTOSIZ)] = { vf
677 };
678
679
680
681
682 static u32
683 vfp_double_fadd_nonnumber(struct vfp_double *v
684 struct vfp_double *v
685 {
686 struct vfp_double *vdp;
687 u32 exceptions = 0;
688 int tn, tm;
689
690 tn = vfp_double_type(vdn);
691 tm = vfp_double_type(vdm);
692
693 if (tn & tm & VFP_INFINITY) {
694 /*
695 * Two infinities. Are they d
696 */
697 if (vdn->sign ^ vdm->sign) {
698 /*
699 * different signs ->
700 */
701 exceptions = FPSCR_IOC
702 vdp = &vfp_double_defa
703 } else {
704 /*
705 * same signs -> valid
706 */
707 vdp = vdn;
708 }
709 } else if (tn & VFP_INFINITY && tm & V
710 /*
711 * One infinity and one number
712 */
713 vdp = vdn;
714 } else {
715 /*
716 * 'n' is a NaN of some type
717 */
718 return vfp_propagate_nan(vdd,
719 }
720 *vdd = *vdp;
721 return exceptions;
722 }
723
724 static u32
725 vfp_double_add(struct vfp_double *vdd, struct
726 struct vfp_double *vdm, u32 fps
727 {
728 u32 exp_diff;
729 u64 m_sig;
730
731 if (vdn->significand & (1ULL << 63) ||
732 vdm->significand & (1ULL << 63)) {
733 pr_info("VFP: bad FP values in
734 vfp_double_dump("VDN", vdn);
735 vfp_double_dump("VDM", vdm);
736 }
737
738 /*
739 * Ensure that 'n' is the largest magn
740 * if 'n' and 'm' have equal exponents
741 * This ensures that NaN propagation w
742 */
743 if (vdn->exponent < vdm->exponent) {
744 struct vfp_double *t = vdn;
745 vdn = vdm;
746 vdm = t;
747 }
748
749 /*
750 * Is 'n' an infinity or a NaN? Note
751 * infinity or a NaN here.
752 */
753 if (vdn->exponent == 2047)
754 return vfp_double_fadd_nonnumb
755
756 /*
757 * We have two proper numbers, where '
758 *
759 * Copy 'n' to 'd' before doing the ar
760 */
761 *vdd = *vdn;
762
763 /*
764 * Align 'm' with the result.
765 */
766 exp_diff = vdn->exponent - vdm->expone
767 m_sig = vfp_shiftright64jamming(vdm->s
768
769 /*
770 * If the signs are different, we are
771 */
772 if (vdn->sign ^ vdm->sign) {
773 m_sig = vdn->significand - m_s
774 if ((s64)m_sig < 0) {
775 vdd->sign = vfp_sign_n
776 m_sig = -m_sig;
777 } else if (m_sig == 0) {
778 vdd->sign = (fpscr & F
779 FPSCR_RO
780 }
781 } else {
782 m_sig += vdn->significand;
783 }
784 vdd->significand = m_sig;
785
786 return 0;
787 }
788
789 static u32
790 vfp_double_multiply(struct vfp_double *vdd, st
791 struct vfp_double *vdm, u3
792 {
793 vfp_double_dump("VDN", vdn);
794 vfp_double_dump("VDM", vdm);
795
796 /*
797 * Ensure that 'n' is the largest magn
798 * if 'n' and 'm' have equal exponents
799 * This ensures that NaN propagation w
800 */
801 if (vdn->exponent < vdm->exponent) {
802 struct vfp_double *t = vdn;
803 vdn = vdm;
804 vdm = t;
805 pr_debug("VFP: swapping M <->
806 }
807
808 vdd->sign = vdn->sign ^ vdm->sign;
809
810 /*
811 * If 'n' is an infinity or NaN, handl
812 */
813 if (vdn->exponent == 2047) {
814 if (vdn->significand || (vdm->
815 return vfp_propagate_n
816 if ((vdm->exponent | vdm->sign
817 *vdd = vfp_double_defa
818 return FPSCR_IOC;
819 }
820 vdd->exponent = vdn->exponent;
821 vdd->significand = 0;
822 return 0;
823 }
824
825 /*
826 * If 'm' is zero, the result is alway
827 * 'n' may be zero or a number, but it
828 */
829 if ((vdm->exponent | vdm->significand)
830 vdd->exponent = 0;
831 vdd->significand = 0;
832 return 0;
833 }
834
835 /*
836 * We add 2 to the destination exponen
837 * as the addition case - though this
838 * each input operand.
839 */
840 vdd->exponent = vdn->exponent + vdm->e
841 vdd->significand = vfp_hi64multiply64(
842
843 vfp_double_dump("VDD", vdd);
844 return 0;
845 }
846
847 #define NEG_MULTIPLY (1 << 0)
848 #define NEG_SUBTRACT (1 << 1)
849
850 static u32
851 vfp_double_multiply_accumulate(int dd, int dn,
852 {
853 struct vfp_double vdd, vdp, vdn, vdm;
854 u32 exceptions;
855
856 vfp_double_unpack(&vdn, vfp_get_double
857 if (vdn.exponent == 0 && vdn.significa
858 vfp_double_normalise_denormal(
859
860 vfp_double_unpack(&vdm, vfp_get_double
861 if (vdm.exponent == 0 && vdm.significa
862 vfp_double_normalise_denormal(
863
864 exceptions = vfp_double_multiply(&vdp,
865 if (negate & NEG_MULTIPLY)
866 vdp.sign = vfp_sign_negate(vdp
867
868 vfp_double_unpack(&vdn, vfp_get_double
869 if (vdn.exponent == 0 && vdn.significa
870 vfp_double_normalise_denormal(
871 if (negate & NEG_SUBTRACT)
872 vdn.sign = vfp_sign_negate(vdn
873
874 exceptions |= vfp_double_add(&vdd, &vd
875
876 return vfp_double_normaliseround(dd, &
877 }
878
879 /*
880 * Standard operations
881 */
882
883 /*
884 * sd = sd + (sn * sm)
885 */
886 static u32 vfp_double_fmac(int dd, int dn, int
887 {
888 return vfp_double_multiply_accumulate(
889 }
890
891 /*
892 * sd = sd - (sn * sm)
893 */
894 static u32 vfp_double_fnmac(int dd, int dn, in
895 {
896 return vfp_double_multiply_accumulate(
897 }
898
899 /*
900 * sd = -sd + (sn * sm)
901 */
902 static u32 vfp_double_fmsc(int dd, int dn, int
903 {
904 return vfp_double_multiply_accumulate(
905 }
906
907 /*
908 * sd = -sd - (sn * sm)
909 */
910 static u32 vfp_double_fnmsc(int dd, int dn, in
911 {
912 return vfp_double_multiply_accumulate(
913 }
914
915 /*
916 * sd = sn * sm
917 */
918 static u32 vfp_double_fmul(int dd, int dn, int
919 {
920 struct vfp_double vdd, vdn, vdm;
921 u32 exceptions;
922
923 vfp_double_unpack(&vdn, vfp_get_double
924 if (vdn.exponent == 0 && vdn.significa
925 vfp_double_normalise_denormal(
926
927 vfp_double_unpack(&vdm, vfp_get_double
928 if (vdm.exponent == 0 && vdm.significa
929 vfp_double_normalise_denormal(
930
931 exceptions = vfp_double_multiply(&vdd,
932 return vfp_double_normaliseround(dd, &
933 }
934
935 /*
936 * sd = -(sn * sm)
937 */
938 static u32 vfp_double_fnmul(int dd, int dn, in
939 {
940 struct vfp_double vdd, vdn, vdm;
941 u32 exceptions;
942
943 vfp_double_unpack(&vdn, vfp_get_double
944 if (vdn.exponent == 0 && vdn.significa
945 vfp_double_normalise_denormal(
946
947 vfp_double_unpack(&vdm, vfp_get_double
948 if (vdm.exponent == 0 && vdm.significa
949 vfp_double_normalise_denormal(
950
951 exceptions = vfp_double_multiply(&vdd,
952 vdd.sign = vfp_sign_negate(vdd.sign);
953
954 return vfp_double_normaliseround(dd, &
955 }
956
957 /*
958 * sd = sn + sm
959 */
960 static u32 vfp_double_fadd(int dd, int dn, int
961 {
962 struct vfp_double vdd, vdn, vdm;
963 u32 exceptions;
964
965 vfp_double_unpack(&vdn, vfp_get_double
966 if (vdn.exponent == 0 && vdn.significa
967 vfp_double_normalise_denormal(
968
969 vfp_double_unpack(&vdm, vfp_get_double
970 if (vdm.exponent == 0 && vdm.significa
971 vfp_double_normalise_denormal(
972
973 exceptions = vfp_double_add(&vdd, &vdn
974
975 return vfp_double_normaliseround(dd, &
976 }
977
978 /*
979 * sd = sn - sm
980 */
981 static u32 vfp_double_fsub(int dd, int dn, int
982 {
983 struct vfp_double vdd, vdn, vdm;
984 u32 exceptions;
985
986 vfp_double_unpack(&vdn, vfp_get_double
987 if (vdn.exponent == 0 && vdn.significa
988 vfp_double_normalise_denormal(
989
990 vfp_double_unpack(&vdm, vfp_get_double
991 if (vdm.exponent == 0 && vdm.significa
992 vfp_double_normalise_denormal(
993
994 /*
995 * Subtraction is like addition, but w
996 */
997 vdm.sign = vfp_sign_negate(vdm.sign);
998
999 exceptions = vfp_double_add(&vdd, &vdn
1000
1001 return vfp_double_normaliseround(dd,
1002 }
1003
1004 /*
1005 * sd = sn / sm
1006 */
1007 static u32 vfp_double_fdiv(int dd, int dn, in
1008 {
1009 struct vfp_double vdd, vdn, vdm;
1010 u32 exceptions = 0;
1011 int tm, tn;
1012
1013 vfp_double_unpack(&vdn, vfp_get_doubl
1014 vfp_double_unpack(&vdm, vfp_get_doubl
1015
1016 vdd.sign = vdn.sign ^ vdm.sign;
1017
1018 tn = vfp_double_type(&vdn);
1019 tm = vfp_double_type(&vdm);
1020
1021 /*
1022 * Is n a NAN?
1023 */
1024 if (tn & VFP_NAN)
1025 goto vdn_nan;
1026
1027 /*
1028 * Is m a NAN?
1029 */
1030 if (tm & VFP_NAN)
1031 goto vdm_nan;
1032
1033 /*
1034 * If n and m are infinity, the resul
1035 * If n and m are zero, the result is
1036 */
1037 if (tm & tn & (VFP_INFINITY|VFP_ZERO)
1038 goto invalid;
1039
1040 /*
1041 * If n is infinity, the result is in
1042 */
1043 if (tn & VFP_INFINITY)
1044 goto infinity;
1045
1046 /*
1047 * If m is zero, raise div0 exception
1048 */
1049 if (tm & VFP_ZERO)
1050 goto divzero;
1051
1052 /*
1053 * If m is infinity, or n is zero, th
1054 */
1055 if (tm & VFP_INFINITY || tn & VFP_ZER
1056 goto zero;
1057
1058 if (tn & VFP_DENORMAL)
1059 vfp_double_normalise_denormal
1060 if (tm & VFP_DENORMAL)
1061 vfp_double_normalise_denormal
1062
1063 /*
1064 * Ok, we have two numbers, we can pe
1065 */
1066 vdd.exponent = vdn.exponent - vdm.exp
1067 vdm.significand <<= 1;
1068 if (vdm.significand <= (2 * vdn.signi
1069 vdn.significand >>= 1;
1070 vdd.exponent++;
1071 }
1072 vdd.significand = vfp_estimate_div128
1073 if ((vdd.significand & 0x1ff) <= 2) {
1074 u64 termh, terml, remh, reml;
1075 mul64to128(&termh, &terml, vd
1076 sub128(&remh, &reml, vdn.sign
1077 while ((s64)remh < 0) {
1078 vdd.significand -= 1;
1079 add128(&remh, &reml,
1080 }
1081 vdd.significand |= (reml != 0
1082 }
1083 return vfp_double_normaliseround(dd,
1084
1085 vdn_nan:
1086 exceptions = vfp_propagate_nan(&vdd,
1087 pack:
1088 vfp_put_double(vfp_double_pack(&vdd),
1089 return exceptions;
1090
1091 vdm_nan:
1092 exceptions = vfp_propagate_nan(&vdd,
1093 goto pack;
1094
1095 zero:
1096 vdd.exponent = 0;
1097 vdd.significand = 0;
1098 goto pack;
1099
1100 divzero:
1101 exceptions = FPSCR_DZC;
1102 infinity:
1103 vdd.exponent = 2047;
1104 vdd.significand = 0;
1105 goto pack;
1106
1107 invalid:
1108 vfp_put_double(vfp_double_pack(&vfp_d
1109 return FPSCR_IOC;
1110 }
1111
1112 static struct op fops[16] = {
1113 [FOP_TO_IDX(FOP_FMAC)] = { vfp_doubl
1114 [FOP_TO_IDX(FOP_FNMAC)] = { vfp_doubl
1115 [FOP_TO_IDX(FOP_FMSC)] = { vfp_doubl
1116 [FOP_TO_IDX(FOP_FNMSC)] = { vfp_doubl
1117 [FOP_TO_IDX(FOP_FMUL)] = { vfp_doubl
1118 [FOP_TO_IDX(FOP_FNMUL)] = { vfp_doubl
1119 [FOP_TO_IDX(FOP_FADD)] = { vfp_doubl
1120 [FOP_TO_IDX(FOP_FSUB)] = { vfp_doubl
1121 [FOP_TO_IDX(FOP_FDIV)] = { vfp_doubl
1122 };
1123
1124 #define FREG_BANK(x) ((x) & 0x0c)
1125 #define FREG_IDX(x) ((x) & 3)
1126
1127 u32 vfp_double_cpdo(u32 inst, u32 fpscr)
1128 {
1129 u32 op = inst & FOP_MASK;
1130 u32 exceptions = 0;
1131 unsigned int dest;
1132 unsigned int dn = vfp_get_dn(inst);
1133 unsigned int dm;
1134 unsigned int vecitr, veclen, vecstrid
1135 struct op *fop;
1136
1137 vecstride = (1 + ((fpscr & FPSCR_STRI
1138
1139 fop = (op == FOP_EXT) ? &fops_ext[FEX
1140
1141 /*
1142 * fcvtds takes an sN register number
1143 * It also always operates on scalars
1144 */
1145 if (fop->flags & OP_SD)
1146 dest = vfp_get_sd(inst);
1147 else
1148 dest = vfp_get_dd(inst);
1149
1150 /*
1151 * f[us]ito takes a sN operand, not a
1152 */
1153 if (fop->flags & OP_SM)
1154 dm = vfp_get_sm(inst);
1155 else
1156 dm = vfp_get_dm(inst);
1157
1158 /*
1159 * If destination bank is zero, vecto
1160 * ARM DDI0100F C5.1.3, C5.3.2.
1161 */
1162 if ((fop->flags & OP_SCALAR) || (FREG
1163 veclen = 0;
1164 else
1165 veclen = fpscr & FPSCR_LENGTH
1166
1167 pr_debug("VFP: vecstride=%u veclen=%u
1168 (veclen >> FPSCR_LENGTH_BIT)
1169
1170 if (!fop->fn)
1171 goto invalid;
1172
1173 for (vecitr = 0; vecitr <= veclen; ve
1174 u32 except;
1175 char type;
1176
1177 type = fop->flags & OP_SD ? '
1178 if (op == FOP_EXT)
1179 pr_debug("VFP: itr%d
1180 vecitr >> FP
1181 type, dest,
1182 else
1183 pr_debug("VFP: itr%d
1184 vecitr >> FP
1185 type, dest,
1186
1187 except = fop->fn(dest, dn, dm
1188 pr_debug("VFP: itr%d: excepti
1189 vecitr >> FPSCR_LENG
1190
1191 exceptions |= except;
1192
1193 /*
1194 * CHECK: It appears to be un
1195 * we encounter an exception.
1196 */
1197 dest = FREG_BANK(dest) + ((FR
1198 dn = FREG_BANK(dn) + ((FREG_I
1199 if (FREG_BANK(dm) != 0)
1200 dm = FREG_BANK(dm) +
1201 }
1202 return exceptions;
1203
1204 invalid:
1205 return ~0;
1206 }
1207
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.