Merge remote-tracking branch 'upstream/master' into kvm-devel
[linux-2.6/kvm.git] / arch / arm / vfp / vfpdouble.c
blob6cac43bd1d86c63638993bafa8ba989fea52e91f
1 /*
2 * linux/arch/arm/vfp/vfpdouble.c
4 * This code is derived in part from John R. Housers softfloat library, which
5 * carries the following notice:
7 * ===========================================================================
8 * This C source file is part of the SoftFloat IEC/IEEE Floating-point
9 * Arithmetic Package, Release 2.
11 * Written by John R. Hauser. This work was made possible in part by the
12 * International Computer Science Institute, located at Suite 600, 1947 Center
13 * Street, Berkeley, California 94704. Funding was partially provided by the
14 * National Science Foundation under grant MIP-9311980. The original version
15 * of this code was written as part of a project to build a fixed-point vector
16 * processor in collaboration with the University of California at Berkeley,
17 * overseen by Profs. Nelson Morgan and John Wawrzynek. More information
18 * is available through the web page `http://HTTP.CS.Berkeley.EDU/~jhauser/
19 * arithmetic/softfloat.html'.
21 * THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE. Although reasonable effort
22 * has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT
23 * TIMES RESULT IN INCORRECT BEHAVIOR. USE OF THIS SOFTWARE IS RESTRICTED TO
24 * PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY
25 * AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE.
27 * Derivative works are acceptable, even for commercial purposes, so long as
28 * (1) they include prominent notice that the work is derivative, and (2) they
29 * include prominent notice akin to these three paragraphs for those parts of
30 * this code that are retained.
31 * ===========================================================================
33 #include <linux/kernel.h>
34 #include <linux/bitops.h>
36 #include <asm/div64.h>
37 #include <asm/vfp.h>
39 #include "vfpinstr.h"
40 #include "vfp.h"
42 static struct vfp_double vfp_double_default_qnan = {
43 .exponent = 2047,
44 .sign = 0,
45 .significand = VFP_DOUBLE_SIGNIFICAND_QNAN,
48 static void vfp_double_dump(const char *str, struct vfp_double *d)
50 pr_debug("VFP: %s: sign=%d exponent=%d significand=%016llx\n",
51 str, d->sign != 0, d->exponent, d->significand);
54 static void vfp_double_normalise_denormal(struct vfp_double *vd)
56 int bits = 31 - fls(vd->significand >> 32);
57 if (bits == 31)
58 bits = 63 - fls(vd->significand);
60 vfp_double_dump("normalise_denormal: in", vd);
62 if (bits) {
63 vd->exponent -= bits - 1;
64 vd->significand <<= bits;
67 vfp_double_dump("normalise_denormal: out", vd);
70 u32 vfp_double_normaliseround(int dd, struct vfp_double *vd, u32 fpscr, u32 exceptions, const char *func)
72 u64 significand, incr;
73 int exponent, shift, underflow;
74 u32 rmode;
76 vfp_double_dump("pack: in", vd);
79 * Infinities and NaNs are a special case.
81 if (vd->exponent == 2047 && (vd->significand == 0 || exceptions))
82 goto pack;
85 * Special-case zero.
87 if (vd->significand == 0) {
88 vd->exponent = 0;
89 goto pack;
92 exponent = vd->exponent;
93 significand = vd->significand;
95 shift = 32 - fls(significand >> 32);
96 if (shift == 32)
97 shift = 64 - fls(significand);
98 if (shift) {
99 exponent -= shift;
100 significand <<= shift;
103 #ifdef DEBUG
104 vd->exponent = exponent;
105 vd->significand = significand;
106 vfp_double_dump("pack: normalised", vd);
107 #endif
110 * Tiny number?
112 underflow = exponent < 0;
113 if (underflow) {
114 significand = vfp_shiftright64jamming(significand, -exponent);
115 exponent = 0;
116 #ifdef DEBUG
117 vd->exponent = exponent;
118 vd->significand = significand;
119 vfp_double_dump("pack: tiny number", vd);
120 #endif
121 if (!(significand & ((1ULL << (VFP_DOUBLE_LOW_BITS + 1)) - 1)))
122 underflow = 0;
126 * Select rounding increment.
128 incr = 0;
129 rmode = fpscr & FPSCR_RMODE_MASK;
131 if (rmode == FPSCR_ROUND_NEAREST) {
132 incr = 1ULL << VFP_DOUBLE_LOW_BITS;
133 if ((significand & (1ULL << (VFP_DOUBLE_LOW_BITS + 1))) == 0)
134 incr -= 1;
135 } else if (rmode == FPSCR_ROUND_TOZERO) {
136 incr = 0;
137 } else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vd->sign != 0))
138 incr = (1ULL << (VFP_DOUBLE_LOW_BITS + 1)) - 1;
140 pr_debug("VFP: rounding increment = 0x%08llx\n", incr);
143 * Is our rounding going to overflow?
145 if ((significand + incr) < significand) {
146 exponent += 1;
147 significand = (significand >> 1) | (significand & 1);
148 incr >>= 1;
149 #ifdef DEBUG
150 vd->exponent = exponent;
151 vd->significand = significand;
152 vfp_double_dump("pack: overflow", vd);
153 #endif
157 * If any of the low bits (which will be shifted out of the
158 * number) are non-zero, the result is inexact.
160 if (significand & ((1 << (VFP_DOUBLE_LOW_BITS + 1)) - 1))
161 exceptions |= FPSCR_IXC;
164 * Do our rounding.
166 significand += incr;
169 * Infinity?
171 if (exponent >= 2046) {
172 exceptions |= FPSCR_OFC | FPSCR_IXC;
173 if (incr == 0) {
174 vd->exponent = 2045;
175 vd->significand = 0x7fffffffffffffffULL;
176 } else {
177 vd->exponent = 2047; /* infinity */
178 vd->significand = 0;
180 } else {
181 if (significand >> (VFP_DOUBLE_LOW_BITS + 1) == 0)
182 exponent = 0;
183 if (exponent || significand > 0x8000000000000000ULL)
184 underflow = 0;
185 if (underflow)
186 exceptions |= FPSCR_UFC;
187 vd->exponent = exponent;
188 vd->significand = significand >> 1;
191 pack:
192 vfp_double_dump("pack: final", vd);
194 s64 d = vfp_double_pack(vd);
195 pr_debug("VFP: %s: d(d%d)=%016llx exceptions=%08x\n", func,
196 dd, d, exceptions);
197 vfp_put_double(d, dd);
199 return exceptions;
203 * Propagate the NaN, setting exceptions if it is signalling.
204 * 'n' is always a NaN. 'm' may be a number, NaN or infinity.
206 static u32
207 vfp_propagate_nan(struct vfp_double *vdd, struct vfp_double *vdn,
208 struct vfp_double *vdm, u32 fpscr)
210 struct vfp_double *nan;
211 int tn, tm = 0;
213 tn = vfp_double_type(vdn);
215 if (vdm)
216 tm = vfp_double_type(vdm);
218 if (fpscr & FPSCR_DEFAULT_NAN)
220 * Default NaN mode - always returns a quiet NaN
222 nan = &vfp_double_default_qnan;
223 else {
225 * Contemporary mode - select the first signalling
226 * NAN, or if neither are signalling, the first
227 * quiet NAN.
229 if (tn == VFP_SNAN || (tm != VFP_SNAN && tn == VFP_QNAN))
230 nan = vdn;
231 else
232 nan = vdm;
234 * Make the NaN quiet.
236 nan->significand |= VFP_DOUBLE_SIGNIFICAND_QNAN;
239 *vdd = *nan;
242 * If one was a signalling NAN, raise invalid operation.
244 return tn == VFP_SNAN || tm == VFP_SNAN ? FPSCR_IOC : VFP_NAN_FLAG;
248 * Extended operations
250 static u32 vfp_double_fabs(int dd, int unused, int dm, u32 fpscr)
252 vfp_put_double(vfp_double_packed_abs(vfp_get_double(dm)), dd);
253 return 0;
256 static u32 vfp_double_fcpy(int dd, int unused, int dm, u32 fpscr)
258 vfp_put_double(vfp_get_double(dm), dd);
259 return 0;
262 static u32 vfp_double_fneg(int dd, int unused, int dm, u32 fpscr)
264 vfp_put_double(vfp_double_packed_negate(vfp_get_double(dm)), dd);
265 return 0;
268 static u32 vfp_double_fsqrt(int dd, int unused, int dm, u32 fpscr)
270 struct vfp_double vdm, vdd;
271 int ret, tm;
273 vfp_double_unpack(&vdm, vfp_get_double(dm));
274 tm = vfp_double_type(&vdm);
275 if (tm & (VFP_NAN|VFP_INFINITY)) {
276 struct vfp_double *vdp = &vdd;
278 if (tm & VFP_NAN)
279 ret = vfp_propagate_nan(vdp, &vdm, NULL, fpscr);
280 else if (vdm.sign == 0) {
281 sqrt_copy:
282 vdp = &vdm;
283 ret = 0;
284 } else {
285 sqrt_invalid:
286 vdp = &vfp_double_default_qnan;
287 ret = FPSCR_IOC;
289 vfp_put_double(vfp_double_pack(vdp), dd);
290 return ret;
294 * sqrt(+/- 0) == +/- 0
296 if (tm & VFP_ZERO)
297 goto sqrt_copy;
300 * Normalise a denormalised number
302 if (tm & VFP_DENORMAL)
303 vfp_double_normalise_denormal(&vdm);
306 * sqrt(<0) = invalid
308 if (vdm.sign)
309 goto sqrt_invalid;
311 vfp_double_dump("sqrt", &vdm);
314 * Estimate the square root.
316 vdd.sign = 0;
317 vdd.exponent = ((vdm.exponent - 1023) >> 1) + 1023;
318 vdd.significand = (u64)vfp_estimate_sqrt_significand(vdm.exponent, vdm.significand >> 32) << 31;
320 vfp_double_dump("sqrt estimate1", &vdd);
322 vdm.significand >>= 1 + (vdm.exponent & 1);
323 vdd.significand += 2 + vfp_estimate_div128to64(vdm.significand, 0, vdd.significand);
325 vfp_double_dump("sqrt estimate2", &vdd);
328 * And now adjust.
330 if ((vdd.significand & VFP_DOUBLE_LOW_BITS_MASK) <= 5) {
331 if (vdd.significand < 2) {
332 vdd.significand = ~0ULL;
333 } else {
334 u64 termh, terml, remh, reml;
335 vdm.significand <<= 2;
336 mul64to128(&termh, &terml, vdd.significand, vdd.significand);
337 sub128(&remh, &reml, vdm.significand, 0, termh, terml);
338 while ((s64)remh < 0) {
339 vdd.significand -= 1;
340 shift64left(&termh, &terml, vdd.significand);
341 terml |= 1;
342 add128(&remh, &reml, remh, reml, termh, terml);
344 vdd.significand |= (remh | reml) != 0;
347 vdd.significand = vfp_shiftright64jamming(vdd.significand, 1);
349 return vfp_double_normaliseround(dd, &vdd, fpscr, 0, "fsqrt");
353 * Equal := ZC
354 * Less than := N
355 * Greater than := C
356 * Unordered := CV
358 static u32 vfp_compare(int dd, int signal_on_qnan, int dm, u32 fpscr)
360 s64 d, m;
361 u32 ret = 0;
363 m = vfp_get_double(dm);
364 if (vfp_double_packed_exponent(m) == 2047 && vfp_double_packed_mantissa(m)) {
365 ret |= FPSCR_C | FPSCR_V;
366 if (signal_on_qnan || !(vfp_double_packed_mantissa(m) & (1ULL << (VFP_DOUBLE_MANTISSA_BITS - 1))))
368 * Signalling NaN, or signalling on quiet NaN
370 ret |= FPSCR_IOC;
373 d = vfp_get_double(dd);
374 if (vfp_double_packed_exponent(d) == 2047 && vfp_double_packed_mantissa(d)) {
375 ret |= FPSCR_C | FPSCR_V;
376 if (signal_on_qnan || !(vfp_double_packed_mantissa(d) & (1ULL << (VFP_DOUBLE_MANTISSA_BITS - 1))))
378 * Signalling NaN, or signalling on quiet NaN
380 ret |= FPSCR_IOC;
383 if (ret == 0) {
384 if (d == m || vfp_double_packed_abs(d | m) == 0) {
386 * equal
388 ret |= FPSCR_Z | FPSCR_C;
389 } else if (vfp_double_packed_sign(d ^ m)) {
391 * different signs
393 if (vfp_double_packed_sign(d))
395 * d is negative, so d < m
397 ret |= FPSCR_N;
398 else
400 * d is positive, so d > m
402 ret |= FPSCR_C;
403 } else if ((vfp_double_packed_sign(d) != 0) ^ (d < m)) {
405 * d < m
407 ret |= FPSCR_N;
408 } else if ((vfp_double_packed_sign(d) != 0) ^ (d > m)) {
410 * d > m
412 ret |= FPSCR_C;
416 return ret;
419 static u32 vfp_double_fcmp(int dd, int unused, int dm, u32 fpscr)
421 return vfp_compare(dd, 0, dm, fpscr);
424 static u32 vfp_double_fcmpe(int dd, int unused, int dm, u32 fpscr)
426 return vfp_compare(dd, 1, dm, fpscr);
429 static u32 vfp_double_fcmpz(int dd, int unused, int dm, u32 fpscr)
431 return vfp_compare(dd, 0, VFP_REG_ZERO, fpscr);
434 static u32 vfp_double_fcmpez(int dd, int unused, int dm, u32 fpscr)
436 return vfp_compare(dd, 1, VFP_REG_ZERO, fpscr);
439 static u32 vfp_double_fcvts(int sd, int unused, int dm, u32 fpscr)
441 struct vfp_double vdm;
442 struct vfp_single vsd;
443 int tm;
444 u32 exceptions = 0;
446 vfp_double_unpack(&vdm, vfp_get_double(dm));
448 tm = vfp_double_type(&vdm);
451 * If we have a signalling NaN, signal invalid operation.
453 if (tm == VFP_SNAN)
454 exceptions = FPSCR_IOC;
456 if (tm & VFP_DENORMAL)
457 vfp_double_normalise_denormal(&vdm);
459 vsd.sign = vdm.sign;
460 vsd.significand = vfp_hi64to32jamming(vdm.significand);
463 * If we have an infinity or a NaN, the exponent must be 255
465 if (tm & (VFP_INFINITY|VFP_NAN)) {
466 vsd.exponent = 255;
467 if (tm == VFP_QNAN)
468 vsd.significand |= VFP_SINGLE_SIGNIFICAND_QNAN;
469 goto pack_nan;
470 } else if (tm & VFP_ZERO)
471 vsd.exponent = 0;
472 else
473 vsd.exponent = vdm.exponent - (1023 - 127);
475 return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fcvts");
477 pack_nan:
478 vfp_put_float(vfp_single_pack(&vsd), sd);
479 return exceptions;
482 static u32 vfp_double_fuito(int dd, int unused, int dm, u32 fpscr)
484 struct vfp_double vdm;
485 u32 m = vfp_get_float(dm);
487 vdm.sign = 0;
488 vdm.exponent = 1023 + 63 - 1;
489 vdm.significand = (u64)m;
491 return vfp_double_normaliseround(dd, &vdm, fpscr, 0, "fuito");
494 static u32 vfp_double_fsito(int dd, int unused, int dm, u32 fpscr)
496 struct vfp_double vdm;
497 u32 m = vfp_get_float(dm);
499 vdm.sign = (m & 0x80000000) >> 16;
500 vdm.exponent = 1023 + 63 - 1;
501 vdm.significand = vdm.sign ? -m : m;
503 return vfp_double_normaliseround(dd, &vdm, fpscr, 0, "fsito");
506 static u32 vfp_double_ftoui(int sd, int unused, int dm, u32 fpscr)
508 struct vfp_double vdm;
509 u32 d, exceptions = 0;
510 int rmode = fpscr & FPSCR_RMODE_MASK;
511 int tm;
513 vfp_double_unpack(&vdm, vfp_get_double(dm));
516 * Do we have a denormalised number?
518 tm = vfp_double_type(&vdm);
519 if (tm & VFP_DENORMAL)
520 exceptions |= FPSCR_IDC;
522 if (tm & VFP_NAN)
523 vdm.sign = 0;
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.exponent;
530 u64 rem, incr = 0;
533 * 2^0 <= m < 2^32-2^8
535 d = (vdm.significand << 1) >> shift;
536 rem = vdm.significand << (65 - shift);
538 if (rmode == FPSCR_ROUND_NEAREST) {
539 incr = 0x8000000000000000ULL;
540 if ((d & 1) == 0)
541 incr -= 1;
542 } else if (rmode == FPSCR_ROUND_TOZERO) {
543 incr = 0;
544 } else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vdm.sign != 0)) {
545 incr = ~0ULL;
548 if ((rem + incr) < rem) {
549 if (d < 0xffffffff)
550 d += 1;
551 else
552 exceptions |= FPSCR_IOC;
555 if (d && vdm.sign) {
556 d = 0;
557 exceptions |= FPSCR_IOC;
558 } else if (rem)
559 exceptions |= FPSCR_IXC;
560 } else {
561 d = 0;
562 if (vdm.exponent | vdm.significand) {
563 exceptions |= FPSCR_IXC;
564 if (rmode == FPSCR_ROUND_PLUSINF && vdm.sign == 0)
565 d = 1;
566 else if (rmode == FPSCR_ROUND_MINUSINF && vdm.sign) {
567 d = 0;
568 exceptions |= FPSCR_IOC;
573 pr_debug("VFP: ftoui: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
575 vfp_put_float(d, sd);
577 return exceptions;
580 static u32 vfp_double_ftouiz(int sd, int unused, int dm, u32 fpscr)
582 return vfp_double_ftoui(sd, unused, dm, FPSCR_ROUND_TOZERO);
585 static u32 vfp_double_ftosi(int sd, int unused, int dm, u32 fpscr)
587 struct vfp_double vdm;
588 u32 d, exceptions = 0;
589 int rmode = fpscr & FPSCR_RMODE_MASK;
590 int tm;
592 vfp_double_unpack(&vdm, vfp_get_double(dm));
593 vfp_double_dump("VDM", &vdm);
596 * Do we have denormalised number?
598 tm = vfp_double_type(&vdm);
599 if (tm & VFP_DENORMAL)
600 exceptions |= FPSCR_IDC;
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.exponent; /* 58 */
612 u64 rem, incr = 0;
614 d = (vdm.significand << 1) >> shift;
615 rem = vdm.significand << (65 - shift);
617 if (rmode == FPSCR_ROUND_NEAREST) {
618 incr = 0x8000000000000000ULL;
619 if ((d & 1) == 0)
620 incr -= 1;
621 } else if (rmode == FPSCR_ROUND_TOZERO) {
622 incr = 0;
623 } else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vdm.sign != 0)) {
624 incr = ~0ULL;
627 if ((rem + incr) < rem && d < 0xffffffff)
628 d += 1;
629 if (d > 0x7fffffff + (vdm.sign != 0)) {
630 d = 0x7fffffff + (vdm.sign != 0);
631 exceptions |= FPSCR_IOC;
632 } else if (rem)
633 exceptions |= FPSCR_IXC;
635 if (vdm.sign)
636 d = -d;
637 } else {
638 d = 0;
639 if (vdm.exponent | vdm.significand) {
640 exceptions |= FPSCR_IXC;
641 if (rmode == FPSCR_ROUND_PLUSINF && vdm.sign == 0)
642 d = 1;
643 else if (rmode == FPSCR_ROUND_MINUSINF && vdm.sign)
644 d = -1;
648 pr_debug("VFP: ftosi: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
650 vfp_put_float((s32)d, sd);
652 return exceptions;
655 static u32 vfp_double_ftosiz(int dd, int unused, int dm, u32 fpscr)
657 return vfp_double_ftosi(dd, unused, dm, FPSCR_ROUND_TOZERO);
661 static struct op fops_ext[32] = {
662 [FEXT_TO_IDX(FEXT_FCPY)] = { vfp_double_fcpy, 0 },
663 [FEXT_TO_IDX(FEXT_FABS)] = { vfp_double_fabs, 0 },
664 [FEXT_TO_IDX(FEXT_FNEG)] = { vfp_double_fneg, 0 },
665 [FEXT_TO_IDX(FEXT_FSQRT)] = { vfp_double_fsqrt, 0 },
666 [FEXT_TO_IDX(FEXT_FCMP)] = { vfp_double_fcmp, OP_SCALAR },
667 [FEXT_TO_IDX(FEXT_FCMPE)] = { vfp_double_fcmpe, OP_SCALAR },
668 [FEXT_TO_IDX(FEXT_FCMPZ)] = { vfp_double_fcmpz, OP_SCALAR },
669 [FEXT_TO_IDX(FEXT_FCMPEZ)] = { vfp_double_fcmpez, OP_SCALAR },
670 [FEXT_TO_IDX(FEXT_FCVT)] = { vfp_double_fcvts, OP_SCALAR|OP_SD },
671 [FEXT_TO_IDX(FEXT_FUITO)] = { vfp_double_fuito, OP_SCALAR|OP_SM },
672 [FEXT_TO_IDX(FEXT_FSITO)] = { vfp_double_fsito, OP_SCALAR|OP_SM },
673 [FEXT_TO_IDX(FEXT_FTOUI)] = { vfp_double_ftoui, OP_SCALAR|OP_SD },
674 [FEXT_TO_IDX(FEXT_FTOUIZ)] = { vfp_double_ftouiz, OP_SCALAR|OP_SD },
675 [FEXT_TO_IDX(FEXT_FTOSI)] = { vfp_double_ftosi, OP_SCALAR|OP_SD },
676 [FEXT_TO_IDX(FEXT_FTOSIZ)] = { vfp_double_ftosiz, OP_SCALAR|OP_SD },
682 static u32
683 vfp_double_fadd_nonnumber(struct vfp_double *vdd, struct vfp_double *vdn,
684 struct vfp_double *vdm, u32 fpscr)
686 struct vfp_double *vdp;
687 u32 exceptions = 0;
688 int tn, tm;
690 tn = vfp_double_type(vdn);
691 tm = vfp_double_type(vdm);
693 if (tn & tm & VFP_INFINITY) {
695 * Two infinities. Are they different signs?
697 if (vdn->sign ^ vdm->sign) {
699 * different signs -> invalid
701 exceptions = FPSCR_IOC;
702 vdp = &vfp_double_default_qnan;
703 } else {
705 * same signs -> valid
707 vdp = vdn;
709 } else if (tn & VFP_INFINITY && tm & VFP_NUMBER) {
711 * One infinity and one number -> infinity
713 vdp = vdn;
714 } else {
716 * 'n' is a NaN of some type
718 return vfp_propagate_nan(vdd, vdn, vdm, fpscr);
720 *vdd = *vdp;
721 return exceptions;
724 static u32
725 vfp_double_add(struct vfp_double *vdd, struct vfp_double *vdn,
726 struct vfp_double *vdm, u32 fpscr)
728 u32 exp_diff;
729 u64 m_sig;
731 if (vdn->significand & (1ULL << 63) ||
732 vdm->significand & (1ULL << 63)) {
733 pr_info("VFP: bad FP values in %s\n", __func__);
734 vfp_double_dump("VDN", vdn);
735 vfp_double_dump("VDM", vdm);
739 * Ensure that 'n' is the largest magnitude number. Note that
740 * if 'n' and 'm' have equal exponents, we do not swap them.
741 * This ensures that NaN propagation works correctly.
743 if (vdn->exponent < vdm->exponent) {
744 struct vfp_double *t = vdn;
745 vdn = vdm;
746 vdm = t;
750 * Is 'n' an infinity or a NaN? Note that 'm' may be a number,
751 * infinity or a NaN here.
753 if (vdn->exponent == 2047)
754 return vfp_double_fadd_nonnumber(vdd, vdn, vdm, fpscr);
757 * We have two proper numbers, where 'vdn' is the larger magnitude.
759 * Copy 'n' to 'd' before doing the arithmetic.
761 *vdd = *vdn;
764 * Align 'm' with the result.
766 exp_diff = vdn->exponent - vdm->exponent;
767 m_sig = vfp_shiftright64jamming(vdm->significand, exp_diff);
770 * If the signs are different, we are really subtracting.
772 if (vdn->sign ^ vdm->sign) {
773 m_sig = vdn->significand - m_sig;
774 if ((s64)m_sig < 0) {
775 vdd->sign = vfp_sign_negate(vdd->sign);
776 m_sig = -m_sig;
777 } else if (m_sig == 0) {
778 vdd->sign = (fpscr & FPSCR_RMODE_MASK) ==
779 FPSCR_ROUND_MINUSINF ? 0x8000 : 0;
781 } else {
782 m_sig += vdn->significand;
784 vdd->significand = m_sig;
786 return 0;
789 static u32
790 vfp_double_multiply(struct vfp_double *vdd, struct vfp_double *vdn,
791 struct vfp_double *vdm, u32 fpscr)
793 vfp_double_dump("VDN", vdn);
794 vfp_double_dump("VDM", vdm);
797 * Ensure that 'n' is the largest magnitude number. Note that
798 * if 'n' and 'm' have equal exponents, we do not swap them.
799 * This ensures that NaN propagation works correctly.
801 if (vdn->exponent < vdm->exponent) {
802 struct vfp_double *t = vdn;
803 vdn = vdm;
804 vdm = t;
805 pr_debug("VFP: swapping M <-> N\n");
808 vdd->sign = vdn->sign ^ vdm->sign;
811 * If 'n' is an infinity or NaN, handle it. 'm' may be anything.
813 if (vdn->exponent == 2047) {
814 if (vdn->significand || (vdm->exponent == 2047 && vdm->significand))
815 return vfp_propagate_nan(vdd, vdn, vdm, fpscr);
816 if ((vdm->exponent | vdm->significand) == 0) {
817 *vdd = vfp_double_default_qnan;
818 return FPSCR_IOC;
820 vdd->exponent = vdn->exponent;
821 vdd->significand = 0;
822 return 0;
826 * If 'm' is zero, the result is always zero. In this case,
827 * 'n' may be zero or a number, but it doesn't matter which.
829 if ((vdm->exponent | vdm->significand) == 0) {
830 vdd->exponent = 0;
831 vdd->significand = 0;
832 return 0;
836 * We add 2 to the destination exponent for the same reason
837 * as the addition case - though this time we have +1 from
838 * each input operand.
840 vdd->exponent = vdn->exponent + vdm->exponent - 1023 + 2;
841 vdd->significand = vfp_hi64multiply64(vdn->significand, vdm->significand);
843 vfp_double_dump("VDD", vdd);
844 return 0;
847 #define NEG_MULTIPLY (1 << 0)
848 #define NEG_SUBTRACT (1 << 1)
850 static u32
851 vfp_double_multiply_accumulate(int dd, int dn, int dm, u32 fpscr, u32 negate, char *func)
853 struct vfp_double vdd, vdp, vdn, vdm;
854 u32 exceptions;
856 vfp_double_unpack(&vdn, vfp_get_double(dn));
857 if (vdn.exponent == 0 && vdn.significand)
858 vfp_double_normalise_denormal(&vdn);
860 vfp_double_unpack(&vdm, vfp_get_double(dm));
861 if (vdm.exponent == 0 && vdm.significand)
862 vfp_double_normalise_denormal(&vdm);
864 exceptions = vfp_double_multiply(&vdp, &vdn, &vdm, fpscr);
865 if (negate & NEG_MULTIPLY)
866 vdp.sign = vfp_sign_negate(vdp.sign);
868 vfp_double_unpack(&vdn, vfp_get_double(dd));
869 if (negate & NEG_SUBTRACT)
870 vdn.sign = vfp_sign_negate(vdn.sign);
872 exceptions |= vfp_double_add(&vdd, &vdn, &vdp, fpscr);
874 return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, func);
878 * Standard operations
882 * sd = sd + (sn * sm)
884 static u32 vfp_double_fmac(int dd, int dn, int dm, u32 fpscr)
886 return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, 0, "fmac");
890 * sd = sd - (sn * sm)
892 static u32 vfp_double_fnmac(int dd, int dn, int dm, u32 fpscr)
894 return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, NEG_MULTIPLY, "fnmac");
898 * sd = -sd + (sn * sm)
900 static u32 vfp_double_fmsc(int dd, int dn, int dm, u32 fpscr)
902 return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, NEG_SUBTRACT, "fmsc");
906 * sd = -sd - (sn * sm)
908 static u32 vfp_double_fnmsc(int dd, int dn, int dm, u32 fpscr)
910 return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, NEG_SUBTRACT | NEG_MULTIPLY, "fnmsc");
914 * sd = sn * sm
916 static u32 vfp_double_fmul(int dd, int dn, int dm, u32 fpscr)
918 struct vfp_double vdd, vdn, vdm;
919 u32 exceptions;
921 vfp_double_unpack(&vdn, vfp_get_double(dn));
922 if (vdn.exponent == 0 && vdn.significand)
923 vfp_double_normalise_denormal(&vdn);
925 vfp_double_unpack(&vdm, vfp_get_double(dm));
926 if (vdm.exponent == 0 && vdm.significand)
927 vfp_double_normalise_denormal(&vdm);
929 exceptions = vfp_double_multiply(&vdd, &vdn, &vdm, fpscr);
930 return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fmul");
934 * sd = -(sn * sm)
936 static u32 vfp_double_fnmul(int dd, int dn, int dm, u32 fpscr)
938 struct vfp_double vdd, vdn, vdm;
939 u32 exceptions;
941 vfp_double_unpack(&vdn, vfp_get_double(dn));
942 if (vdn.exponent == 0 && vdn.significand)
943 vfp_double_normalise_denormal(&vdn);
945 vfp_double_unpack(&vdm, vfp_get_double(dm));
946 if (vdm.exponent == 0 && vdm.significand)
947 vfp_double_normalise_denormal(&vdm);
949 exceptions = vfp_double_multiply(&vdd, &vdn, &vdm, fpscr);
950 vdd.sign = vfp_sign_negate(vdd.sign);
952 return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fnmul");
956 * sd = sn + sm
958 static u32 vfp_double_fadd(int dd, int dn, int dm, u32 fpscr)
960 struct vfp_double vdd, vdn, vdm;
961 u32 exceptions;
963 vfp_double_unpack(&vdn, vfp_get_double(dn));
964 if (vdn.exponent == 0 && vdn.significand)
965 vfp_double_normalise_denormal(&vdn);
967 vfp_double_unpack(&vdm, vfp_get_double(dm));
968 if (vdm.exponent == 0 && vdm.significand)
969 vfp_double_normalise_denormal(&vdm);
971 exceptions = vfp_double_add(&vdd, &vdn, &vdm, fpscr);
973 return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fadd");
977 * sd = sn - sm
979 static u32 vfp_double_fsub(int dd, int dn, int dm, u32 fpscr)
981 struct vfp_double vdd, vdn, vdm;
982 u32 exceptions;
984 vfp_double_unpack(&vdn, vfp_get_double(dn));
985 if (vdn.exponent == 0 && vdn.significand)
986 vfp_double_normalise_denormal(&vdn);
988 vfp_double_unpack(&vdm, vfp_get_double(dm));
989 if (vdm.exponent == 0 && vdm.significand)
990 vfp_double_normalise_denormal(&vdm);
993 * Subtraction is like addition, but with a negated operand.
995 vdm.sign = vfp_sign_negate(vdm.sign);
997 exceptions = vfp_double_add(&vdd, &vdn, &vdm, fpscr);
999 return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fsub");
1003 * sd = sn / sm
1005 static u32 vfp_double_fdiv(int dd, int dn, int dm, u32 fpscr)
1007 struct vfp_double vdd, vdn, vdm;
1008 u32 exceptions = 0;
1009 int tm, tn;
1011 vfp_double_unpack(&vdn, vfp_get_double(dn));
1012 vfp_double_unpack(&vdm, vfp_get_double(dm));
1014 vdd.sign = vdn.sign ^ vdm.sign;
1016 tn = vfp_double_type(&vdn);
1017 tm = vfp_double_type(&vdm);
1020 * Is n a NAN?
1022 if (tn & VFP_NAN)
1023 goto vdn_nan;
1026 * Is m a NAN?
1028 if (tm & VFP_NAN)
1029 goto vdm_nan;
1032 * If n and m are infinity, the result is invalid
1033 * If n and m are zero, the result is invalid
1035 if (tm & tn & (VFP_INFINITY|VFP_ZERO))
1036 goto invalid;
1039 * If n is infinity, the result is infinity
1041 if (tn & VFP_INFINITY)
1042 goto infinity;
1045 * If m is zero, raise div0 exceptions
1047 if (tm & VFP_ZERO)
1048 goto divzero;
1051 * If m is infinity, or n is zero, the result is zero
1053 if (tm & VFP_INFINITY || tn & VFP_ZERO)
1054 goto zero;
1056 if (tn & VFP_DENORMAL)
1057 vfp_double_normalise_denormal(&vdn);
1058 if (tm & VFP_DENORMAL)
1059 vfp_double_normalise_denormal(&vdm);
1062 * Ok, we have two numbers, we can perform division.
1064 vdd.exponent = vdn.exponent - vdm.exponent + 1023 - 1;
1065 vdm.significand <<= 1;
1066 if (vdm.significand <= (2 * vdn.significand)) {
1067 vdn.significand >>= 1;
1068 vdd.exponent++;
1070 vdd.significand = vfp_estimate_div128to64(vdn.significand, 0, vdm.significand);
1071 if ((vdd.significand & 0x1ff) <= 2) {
1072 u64 termh, terml, remh, reml;
1073 mul64to128(&termh, &terml, vdm.significand, vdd.significand);
1074 sub128(&remh, &reml, vdn.significand, 0, termh, terml);
1075 while ((s64)remh < 0) {
1076 vdd.significand -= 1;
1077 add128(&remh, &reml, remh, reml, 0, vdm.significand);
1079 vdd.significand |= (reml != 0);
1081 return vfp_double_normaliseround(dd, &vdd, fpscr, 0, "fdiv");
1083 vdn_nan:
1084 exceptions = vfp_propagate_nan(&vdd, &vdn, &vdm, fpscr);
1085 pack:
1086 vfp_put_double(vfp_double_pack(&vdd), dd);
1087 return exceptions;
1089 vdm_nan:
1090 exceptions = vfp_propagate_nan(&vdd, &vdm, &vdn, fpscr);
1091 goto pack;
1093 zero:
1094 vdd.exponent = 0;
1095 vdd.significand = 0;
1096 goto pack;
1098 divzero:
1099 exceptions = FPSCR_DZC;
1100 infinity:
1101 vdd.exponent = 2047;
1102 vdd.significand = 0;
1103 goto pack;
1105 invalid:
1106 vfp_put_double(vfp_double_pack(&vfp_double_default_qnan), dd);
1107 return FPSCR_IOC;
1110 static struct op fops[16] = {
1111 [FOP_TO_IDX(FOP_FMAC)] = { vfp_double_fmac, 0 },
1112 [FOP_TO_IDX(FOP_FNMAC)] = { vfp_double_fnmac, 0 },
1113 [FOP_TO_IDX(FOP_FMSC)] = { vfp_double_fmsc, 0 },
1114 [FOP_TO_IDX(FOP_FNMSC)] = { vfp_double_fnmsc, 0 },
1115 [FOP_TO_IDX(FOP_FMUL)] = { vfp_double_fmul, 0 },
1116 [FOP_TO_IDX(FOP_FNMUL)] = { vfp_double_fnmul, 0 },
1117 [FOP_TO_IDX(FOP_FADD)] = { vfp_double_fadd, 0 },
1118 [FOP_TO_IDX(FOP_FSUB)] = { vfp_double_fsub, 0 },
1119 [FOP_TO_IDX(FOP_FDIV)] = { vfp_double_fdiv, 0 },
1122 #define FREG_BANK(x) ((x) & 0x0c)
1123 #define FREG_IDX(x) ((x) & 3)
1125 u32 vfp_double_cpdo(u32 inst, u32 fpscr)
1127 u32 op = inst & FOP_MASK;
1128 u32 exceptions = 0;
1129 unsigned int dest;
1130 unsigned int dn = vfp_get_dn(inst);
1131 unsigned int dm;
1132 unsigned int vecitr, veclen, vecstride;
1133 struct op *fop;
1135 vecstride = (1 + ((fpscr & FPSCR_STRIDE_MASK) == FPSCR_STRIDE_MASK));
1137 fop = (op == FOP_EXT) ? &fops_ext[FEXT_TO_IDX(inst)] : &fops[FOP_TO_IDX(op)];
1140 * fcvtds takes an sN register number as destination, not dN.
1141 * It also always operates on scalars.
1143 if (fop->flags & OP_SD)
1144 dest = vfp_get_sd(inst);
1145 else
1146 dest = vfp_get_dd(inst);
1149 * f[us]ito takes a sN operand, not a dN operand.
1151 if (fop->flags & OP_SM)
1152 dm = vfp_get_sm(inst);
1153 else
1154 dm = vfp_get_dm(inst);
1157 * If destination bank is zero, vector length is always '1'.
1158 * ARM DDI0100F C5.1.3, C5.3.2.
1160 if ((fop->flags & OP_SCALAR) || (FREG_BANK(dest) == 0))
1161 veclen = 0;
1162 else
1163 veclen = fpscr & FPSCR_LENGTH_MASK;
1165 pr_debug("VFP: vecstride=%u veclen=%u\n", vecstride,
1166 (veclen >> FPSCR_LENGTH_BIT) + 1);
1168 if (!fop->fn)
1169 goto invalid;
1171 for (vecitr = 0; vecitr <= veclen; vecitr += 1 << FPSCR_LENGTH_BIT) {
1172 u32 except;
1173 char type;
1175 type = fop->flags & OP_SD ? 's' : 'd';
1176 if (op == FOP_EXT)
1177 pr_debug("VFP: itr%d (%c%u) = op[%u] (d%u)\n",
1178 vecitr >> FPSCR_LENGTH_BIT,
1179 type, dest, dn, dm);
1180 else
1181 pr_debug("VFP: itr%d (%c%u) = (d%u) op[%u] (d%u)\n",
1182 vecitr >> FPSCR_LENGTH_BIT,
1183 type, dest, dn, FOP_TO_IDX(op), dm);
1185 except = fop->fn(dest, dn, dm, fpscr);
1186 pr_debug("VFP: itr%d: exceptions=%08x\n",
1187 vecitr >> FPSCR_LENGTH_BIT, except);
1189 exceptions |= except;
1192 * CHECK: It appears to be undefined whether we stop when
1193 * we encounter an exception. We continue.
1195 dest = FREG_BANK(dest) + ((FREG_IDX(dest) + vecstride) & 3);
1196 dn = FREG_BANK(dn) + ((FREG_IDX(dn) + vecstride) & 3);
1197 if (FREG_BANK(dm) != 0)
1198 dm = FREG_BANK(dm) + ((FREG_IDX(dm) + vecstride) & 3);
1200 return exceptions;
1202 invalid:
1203 return ~0;