xref: /linux/arch/arm/vfp/vfpdouble.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
1 /*
2  *  linux/arch/arm/vfp/vfpdouble.c
3  *
4  * This code is derived in part from John R. Housers softfloat library, which
5  * carries the following notice:
6  *
7  * ===========================================================================
8  * This C source file is part of the SoftFloat IEC/IEEE Floating-point
9  * Arithmetic Package, Release 2.
10  *
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'.
20  *
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.
26  *
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  * ===========================================================================
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_qnan = {
43 	.exponent	= 2047,
44 	.sign		= 0,
45 	.significand	= VFP_DOUBLE_SIGNIFICAND_QNAN,
46 };
47 
48 static void vfp_double_dump(const char *str, struct vfp_double *d)
49 {
50 	pr_debug("VFP: %s: sign=%d exponent=%d significand=%016llx\n",
51 		 str, d->sign != 0, d->exponent, d->significand);
52 }
53 
54 static void vfp_double_normalise_denormal(struct vfp_double *vd)
55 {
56 	int bits = 31 - fls(vd->significand >> 32);
57 	if (bits == 31)
58 		bits = 63 - fls(vd->significand);
59 
60 	vfp_double_dump("normalise_denormal: in", vd);
61 
62 	if (bits) {
63 		vd->exponent -= bits - 1;
64 		vd->significand <<= bits;
65 	}
66 
67 	vfp_double_dump("normalise_denormal: out", vd);
68 }
69 
70 u32 vfp_double_normaliseround(int dd, struct vfp_double *vd, u32 fpscr, u32 exceptions, const char *func)
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 case.
80 	 */
81 	if (vd->exponent == 2047 && (vd->significand == 0 || exceptions))
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_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;
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_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;
139 
140 	pr_debug("VFP: rounding increment = 0x%08llx\n", incr);
141 
142 	/*
143 	 * Is our rounding going to overflow?
144 	 */
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
154 	}
155 
156 	/*
157 	 * If any of the low bits (which will be shifted out of the
158 	 * number) are non-zero, the result is inexact.
159 	 */
160 	if (significand & ((1 << (VFP_DOUBLE_LOW_BITS + 1)) - 1))
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 | 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;
179 		}
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;
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)=%016llx exceptions=%08x\n", func,
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 is signalling.
204  * 'n' is always a NaN.  'm' may be a number, NaN or infinity.
205  */
206 static u32
207 vfp_propagate_nan(struct vfp_double *vdd, struct vfp_double *vdn,
208 		  struct vfp_double *vdm, u32 fpscr)
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 returns a quiet NaN
221 		 */
222 		nan = &vfp_double_default_qnan;
223 	else {
224 		/*
225 		 * Contemporary mode - select the first signalling
226 		 * NAN, or if neither are signalling, the first
227 		 * quiet NAN.
228 		 */
229 		if (tn == VFP_SNAN || (tm != VFP_SNAN && tn == VFP_QNAN))
230 			nan = vdn;
231 		else
232 			nan = vdm;
233 		/*
234 		 * Make the NaN quiet.
235 		 */
236 		nan->significand |= VFP_DOUBLE_SIGNIFICAND_QNAN;
237 	}
238 
239 	*vdd = *nan;
240 
241 	/*
242 	 * If one was a signalling NAN, raise invalid operation.
243 	 */
244 	return tn == VFP_SNAN || tm == VFP_SNAN ? FPSCR_IOC : VFP_NAN_FLAG;
245 }
246 
247 /*
248  * Extended operations
249  */
250 static u32 vfp_double_fabs(int dd, int unused, int dm, u32 fpscr)
251 {
252 	vfp_put_double(vfp_double_packed_abs(vfp_get_double(dm)), dd);
253 	return 0;
254 }
255 
256 static u32 vfp_double_fcpy(int dd, int unused, int dm, u32 fpscr)
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, int dm, u32 fpscr)
263 {
264 	vfp_put_double(vfp_double_packed_negate(vfp_get_double(dm)), dd);
265 	return 0;
266 }
267 
268 static u32 vfp_double_fsqrt(int dd, int unused, int dm, u32 fpscr)
269 {
270 	struct vfp_double vdm, vdd;
271 	int ret, tm;
272 
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;
277 
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;
288 		}
289 		vfp_put_double(vfp_double_pack(vdp), dd);
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(&vdm);
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) >> 1) + 1023;
318 	vdd.significand = (u64)vfp_estimate_sqrt_significand(vdm.exponent, vdm.significand >> 32) << 31;
319 
320 	vfp_double_dump("sqrt estimate1", &vdd);
321 
322 	vdm.significand >>= 1 + (vdm.exponent & 1);
323 	vdd.significand += 2 + vfp_estimate_div128to64(vdm.significand, 0, vdd.significand);
324 
325 	vfp_double_dump("sqrt estimate2", &vdd);
326 
327 	/*
328 	 * And now adjust.
329 	 */
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);
343 			}
344 			vdd.significand |= (remh | reml) != 0;
345 		}
346 	}
347 	vdd.significand = vfp_shiftright64jamming(vdd.significand, 1);
348 
349 	return vfp_double_normaliseround(dd, &vdd, fpscr, 0, "fsqrt");
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_qnan, int dm, u32 fpscr)
359 {
360 	s64 d, m;
361 	u32 ret = 0;
362 
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))))
367 			/*
368 			 * Signalling NaN, or signalling on quiet NaN
369 			 */
370 			ret |= FPSCR_IOC;
371 	}
372 
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))))
377 			/*
378 			 * Signalling NaN, or signalling on quiet NaN
379 			 */
380 			ret |= FPSCR_IOC;
381 	}
382 
383 	if (ret == 0) {
384 		if (d == m || vfp_double_packed_abs(d | m) == 0) {
385 			/*
386 			 * equal
387 			 */
388 			ret |= FPSCR_Z | FPSCR_C;
389 		} else if (vfp_double_packed_sign(d ^ m)) {
390 			/*
391 			 * different signs
392 			 */
393 			if (vfp_double_packed_sign(d))
394 				/*
395 				 * d is negative, so d < m
396 				 */
397 				ret |= FPSCR_N;
398 			else
399 				/*
400 				 * d is positive, so d > m
401 				 */
402 				ret |= FPSCR_C;
403 		} else if ((vfp_double_packed_sign(d) != 0) ^ (d < m)) {
404 			/*
405 			 * d < m
406 			 */
407 			ret |= FPSCR_N;
408 		} else if ((vfp_double_packed_sign(d) != 0) ^ (d > m)) {
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, int dm, u32 fpscr)
420 {
421 	return vfp_compare(dd, 0, dm, fpscr);
422 }
423 
424 static u32 vfp_double_fcmpe(int dd, int unused, int dm, u32 fpscr)
425 {
426 	return vfp_compare(dd, 1, dm, fpscr);
427 }
428 
429 static u32 vfp_double_fcmpz(int dd, int unused, int dm, u32 fpscr)
430 {
431 	return vfp_compare(dd, 0, VFP_REG_ZERO, fpscr);
432 }
433 
434 static u32 vfp_double_fcmpez(int dd, int unused, int dm, u32 fpscr)
435 {
436 	return vfp_compare(dd, 1, VFP_REG_ZERO, fpscr);
437 }
438 
439 static u32 vfp_double_fcvts(int sd, int unused, int dm, u32 fpscr)
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(dm));
447 
448 	tm = vfp_double_type(&vdm);
449 
450 	/*
451 	 * If we have a signalling NaN, signal invalid operation.
452 	 */
453 	if (tm == VFP_SNAN)
454 		exceptions = FPSCR_IOC;
455 
456 	if (tm & VFP_DENORMAL)
457 		vfp_double_normalise_denormal(&vdm);
458 
459 	vsd.sign = vdm.sign;
460 	vsd.significand = vfp_hi64to32jamming(vdm.significand);
461 
462 	/*
463 	 * If we have an infinity or a NaN, the exponent must be 255
464 	 */
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);
474 
475 	return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fcvts");
476 
477  pack_nan:
478 	vfp_put_float(vfp_single_pack(&vsd), sd);
479 	return exceptions;
480 }
481 
482 static u32 vfp_double_fuito(int dd, int unused, int dm, u32 fpscr)
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, &vdm, fpscr, 0, "fuito");
492 }
493 
494 static u32 vfp_double_fsito(int dd, int unused, int dm, u32 fpscr)
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, &vdm, fpscr, 0, "fsito");
504 }
505 
506 static u32 vfp_double_ftoui(int sd, int unused, int dm, u32 fpscr)
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(dm));
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.exponent;
530 		u64 rem, incr = 0;
531 
532 		/*
533 		 * 2^0 <= m < 2^32-2^8
534 		 */
535 		d = (vdm.significand << 1) >> shift;
536 		rem = vdm.significand << (65 - shift);
537 
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;
546 		}
547 
548 		if ((rem + incr) < rem) {
549 			if (d < 0xffffffff)
550 				d += 1;
551 			else
552 				exceptions |= FPSCR_IOC;
553 		}
554 
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;
569 			}
570 		}
571 	}
572 
573 	pr_debug("VFP: ftoui: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
574 
575 	vfp_put_float(d, sd);
576 
577 	return exceptions;
578 }
579 
580 static u32 vfp_double_ftouiz(int sd, int unused, int dm, u32 fpscr)
581 {
582 	return vfp_double_ftoui(sd, unused, dm, FPSCR_ROUND_TOZERO);
583 }
584 
585 static u32 vfp_double_ftosi(int sd, int unused, int dm, u32 fpscr)
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(dm));
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.exponent;	/* 58 */
612 		u64 rem, incr = 0;
613 
614 		d = (vdm.significand << 1) >> shift;
615 		rem = vdm.significand << (65 - shift);
616 
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;
625 		}
626 
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;
634 
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;
645 		}
646 	}
647 
648 	pr_debug("VFP: ftosi: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
649 
650 	vfp_put_float((s32)d, sd);
651 
652 	return exceptions;
653 }
654 
655 static u32 vfp_double_ftosiz(int dd, int unused, int dm, u32 fpscr)
656 {
657 	return vfp_double_ftosi(dd, unused, dm, FPSCR_ROUND_TOZERO);
658 }
659 
660 
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 },
677 };
678 
679 
680 
681 
682 static u32
683 vfp_double_fadd_nonnumber(struct vfp_double *vdd, struct vfp_double *vdn,
684 			  struct vfp_double *vdm, u32 fpscr)
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 different signs?
696 		 */
697 		if (vdn->sign ^ vdm->sign) {
698 			/*
699 			 * different signs -> invalid
700 			 */
701 			exceptions = FPSCR_IOC;
702 			vdp = &vfp_double_default_qnan;
703 		} else {
704 			/*
705 			 * same signs -> valid
706 			 */
707 			vdp = vdn;
708 		}
709 	} else if (tn & VFP_INFINITY && tm & VFP_NUMBER) {
710 		/*
711 		 * One infinity and one number -> infinity
712 		 */
713 		vdp = vdn;
714 	} else {
715 		/*
716 		 * 'n' is a NaN of some type
717 		 */
718 		return vfp_propagate_nan(vdd, vdn, vdm, fpscr);
719 	}
720 	*vdd = *vdp;
721 	return exceptions;
722 }
723 
724 static u32
725 vfp_double_add(struct vfp_double *vdd, struct vfp_double *vdn,
726 	       struct vfp_double *vdm, u32 fpscr)
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 %s\n", __func__);
734 		vfp_double_dump("VDN", vdn);
735 		vfp_double_dump("VDM", vdm);
736 	}
737 
738 	/*
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.
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 that 'm' may be a number,
751 	 * infinity or a NaN here.
752 	 */
753 	if (vdn->exponent == 2047)
754 		return vfp_double_fadd_nonnumber(vdd, vdn, vdm, fpscr);
755 
756 	/*
757 	 * We have two proper numbers, where 'vdn' is the larger magnitude.
758 	 *
759 	 * Copy 'n' to 'd' before doing the arithmetic.
760 	 */
761 	*vdd = *vdn;
762 
763 	/*
764 	 * Align 'm' with the result.
765 	 */
766 	exp_diff = vdn->exponent - vdm->exponent;
767 	m_sig = vfp_shiftright64jamming(vdm->significand, exp_diff);
768 
769 	/*
770 	 * If the signs are different, we are really subtracting.
771 	 */
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;
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, struct vfp_double *vdn,
791 		    struct vfp_double *vdm, u32 fpscr)
792 {
793 	vfp_double_dump("VDN", vdn);
794 	vfp_double_dump("VDM", vdm);
795 
796 	/*
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.
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 <-> N\n");
806 	}
807 
808 	vdd->sign = vdn->sign ^ vdm->sign;
809 
810 	/*
811 	 * If 'n' is an infinity or NaN, handle it.  'm' may be anything.
812 	 */
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;
819 		}
820 		vdd->exponent = vdn->exponent;
821 		vdd->significand = 0;
822 		return 0;
823 	}
824 
825 	/*
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.
828 	 */
829 	if ((vdm->exponent | vdm->significand) == 0) {
830 		vdd->exponent = 0;
831 		vdd->significand = 0;
832 		return 0;
833 	}
834 
835 	/*
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.
839 	 */
840 	vdd->exponent = vdn->exponent + vdm->exponent - 1023 + 2;
841 	vdd->significand = vfp_hi64multiply64(vdn->significand, vdm->significand);
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, int dm, u32 fpscr, u32 negate, char *func)
852 {
853 	struct vfp_double vdd, vdp, vdn, vdm;
854 	u32 exceptions;
855 
856 	vfp_double_unpack(&vdn, vfp_get_double(dn));
857 	if (vdn.exponent == 0 && vdn.significand)
858 		vfp_double_normalise_denormal(&vdn);
859 
860 	vfp_double_unpack(&vdm, vfp_get_double(dm));
861 	if (vdm.exponent == 0 && vdm.significand)
862 		vfp_double_normalise_denormal(&vdm);
863 
864 	exceptions = vfp_double_multiply(&vdp, &vdn, &vdm, fpscr);
865 	if (negate & NEG_MULTIPLY)
866 		vdp.sign = vfp_sign_negate(vdp.sign);
867 
868 	vfp_double_unpack(&vdn, vfp_get_double(dd));
869 	if (vdn.exponent == 0 && vdn.significand)
870 		vfp_double_normalise_denormal(&vdn);
871 	if (negate & NEG_SUBTRACT)
872 		vdn.sign = vfp_sign_negate(vdn.sign);
873 
874 	exceptions |= vfp_double_add(&vdd, &vdn, &vdp, fpscr);
875 
876 	return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, func);
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 dm, u32 fpscr)
887 {
888 	return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, 0, "fmac");
889 }
890 
891 /*
892  * sd = sd - (sn * sm)
893  */
894 static u32 vfp_double_fnmac(int dd, int dn, int dm, u32 fpscr)
895 {
896 	return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, NEG_MULTIPLY, "fnmac");
897 }
898 
899 /*
900  * sd = -sd + (sn * sm)
901  */
902 static u32 vfp_double_fmsc(int dd, int dn, int dm, u32 fpscr)
903 {
904 	return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, NEG_SUBTRACT, "fmsc");
905 }
906 
907 /*
908  * sd = -sd - (sn * sm)
909  */
910 static u32 vfp_double_fnmsc(int dd, int dn, int dm, u32 fpscr)
911 {
912 	return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, NEG_SUBTRACT | NEG_MULTIPLY, "fnmsc");
913 }
914 
915 /*
916  * sd = sn * sm
917  */
918 static u32 vfp_double_fmul(int dd, int dn, int dm, u32 fpscr)
919 {
920 	struct vfp_double vdd, vdn, vdm;
921 	u32 exceptions;
922 
923 	vfp_double_unpack(&vdn, vfp_get_double(dn));
924 	if (vdn.exponent == 0 && vdn.significand)
925 		vfp_double_normalise_denormal(&vdn);
926 
927 	vfp_double_unpack(&vdm, vfp_get_double(dm));
928 	if (vdm.exponent == 0 && vdm.significand)
929 		vfp_double_normalise_denormal(&vdm);
930 
931 	exceptions = vfp_double_multiply(&vdd, &vdn, &vdm, fpscr);
932 	return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fmul");
933 }
934 
935 /*
936  * sd = -(sn * sm)
937  */
938 static u32 vfp_double_fnmul(int dd, int dn, int dm, u32 fpscr)
939 {
940 	struct vfp_double vdd, vdn, vdm;
941 	u32 exceptions;
942 
943 	vfp_double_unpack(&vdn, vfp_get_double(dn));
944 	if (vdn.exponent == 0 && vdn.significand)
945 		vfp_double_normalise_denormal(&vdn);
946 
947 	vfp_double_unpack(&vdm, vfp_get_double(dm));
948 	if (vdm.exponent == 0 && vdm.significand)
949 		vfp_double_normalise_denormal(&vdm);
950 
951 	exceptions = vfp_double_multiply(&vdd, &vdn, &vdm, fpscr);
952 	vdd.sign = vfp_sign_negate(vdd.sign);
953 
954 	return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fnmul");
955 }
956 
957 /*
958  * sd = sn + sm
959  */
960 static u32 vfp_double_fadd(int dd, int dn, int dm, u32 fpscr)
961 {
962 	struct vfp_double vdd, vdn, vdm;
963 	u32 exceptions;
964 
965 	vfp_double_unpack(&vdn, vfp_get_double(dn));
966 	if (vdn.exponent == 0 && vdn.significand)
967 		vfp_double_normalise_denormal(&vdn);
968 
969 	vfp_double_unpack(&vdm, vfp_get_double(dm));
970 	if (vdm.exponent == 0 && vdm.significand)
971 		vfp_double_normalise_denormal(&vdm);
972 
973 	exceptions = vfp_double_add(&vdd, &vdn, &vdm, fpscr);
974 
975 	return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fadd");
976 }
977 
978 /*
979  * sd = sn - sm
980  */
981 static u32 vfp_double_fsub(int dd, int dn, int dm, u32 fpscr)
982 {
983 	struct vfp_double vdd, vdn, vdm;
984 	u32 exceptions;
985 
986 	vfp_double_unpack(&vdn, vfp_get_double(dn));
987 	if (vdn.exponent == 0 && vdn.significand)
988 		vfp_double_normalise_denormal(&vdn);
989 
990 	vfp_double_unpack(&vdm, vfp_get_double(dm));
991 	if (vdm.exponent == 0 && vdm.significand)
992 		vfp_double_normalise_denormal(&vdm);
993 
994 	/*
995 	 * Subtraction is like addition, but with a negated operand.
996 	 */
997 	vdm.sign = vfp_sign_negate(vdm.sign);
998 
999 	exceptions = vfp_double_add(&vdd, &vdn, &vdm, fpscr);
1000 
1001 	return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fsub");
1002 }
1003 
1004 /*
1005  * sd = sn / sm
1006  */
1007 static u32 vfp_double_fdiv(int dd, int dn, int dm, u32 fpscr)
1008 {
1009 	struct vfp_double vdd, vdn, vdm;
1010 	u32 exceptions = 0;
1011 	int tm, tn;
1012 
1013 	vfp_double_unpack(&vdn, vfp_get_double(dn));
1014 	vfp_double_unpack(&vdm, vfp_get_double(dm));
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 result is invalid
1035 	 * If n and m are zero, the result is invalid
1036 	 */
1037 	if (tm & tn & (VFP_INFINITY|VFP_ZERO))
1038 		goto invalid;
1039 
1040 	/*
1041 	 * If n is infinity, the result is infinity
1042 	 */
1043 	if (tn & VFP_INFINITY)
1044 		goto infinity;
1045 
1046 	/*
1047 	 * If m is zero, raise div0 exceptions
1048 	 */
1049 	if (tm & VFP_ZERO)
1050 		goto divzero;
1051 
1052 	/*
1053 	 * If m is infinity, or n is zero, the result is zero
1054 	 */
1055 	if (tm & VFP_INFINITY || tn & VFP_ZERO)
1056 		goto zero;
1057 
1058 	if (tn & VFP_DENORMAL)
1059 		vfp_double_normalise_denormal(&vdn);
1060 	if (tm & VFP_DENORMAL)
1061 		vfp_double_normalise_denormal(&vdm);
1062 
1063 	/*
1064 	 * Ok, we have two numbers, we can perform division.
1065 	 */
1066 	vdd.exponent = vdn.exponent - vdm.exponent + 1023 - 1;
1067 	vdm.significand <<= 1;
1068 	if (vdm.significand <= (2 * vdn.significand)) {
1069 		vdn.significand >>= 1;
1070 		vdd.exponent++;
1071 	}
1072 	vdd.significand = vfp_estimate_div128to64(vdn.significand, 0, vdm.significand);
1073 	if ((vdd.significand & 0x1ff) <= 2) {
1074 		u64 termh, terml, remh, reml;
1075 		mul64to128(&termh, &terml, vdm.significand, vdd.significand);
1076 		sub128(&remh, &reml, vdn.significand, 0, termh, terml);
1077 		while ((s64)remh < 0) {
1078 			vdd.significand -= 1;
1079 			add128(&remh, &reml, remh, reml, 0, vdm.significand);
1080 		}
1081 		vdd.significand |= (reml != 0);
1082 	}
1083 	return vfp_double_normaliseround(dd, &vdd, fpscr, 0, "fdiv");
1084 
1085  vdn_nan:
1086 	exceptions = vfp_propagate_nan(&vdd, &vdn, &vdm, fpscr);
1087  pack:
1088 	vfp_put_double(vfp_double_pack(&vdd), dd);
1089 	return exceptions;
1090 
1091  vdm_nan:
1092 	exceptions = vfp_propagate_nan(&vdd, &vdm, &vdn, fpscr);
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_double_default_qnan), dd);
1109 	return FPSCR_IOC;
1110 }
1111 
1112 static struct op fops[16] = {
1113 	[FOP_TO_IDX(FOP_FMAC)]	= { vfp_double_fmac,  0 },
1114 	[FOP_TO_IDX(FOP_FNMAC)]	= { vfp_double_fnmac, 0 },
1115 	[FOP_TO_IDX(FOP_FMSC)]	= { vfp_double_fmsc,  0 },
1116 	[FOP_TO_IDX(FOP_FNMSC)]	= { vfp_double_fnmsc, 0 },
1117 	[FOP_TO_IDX(FOP_FMUL)]	= { vfp_double_fmul,  0 },
1118 	[FOP_TO_IDX(FOP_FNMUL)]	= { vfp_double_fnmul, 0 },
1119 	[FOP_TO_IDX(FOP_FADD)]	= { vfp_double_fadd,  0 },
1120 	[FOP_TO_IDX(FOP_FSUB)]	= { vfp_double_fsub,  0 },
1121 	[FOP_TO_IDX(FOP_FDIV)]	= { vfp_double_fdiv,  0 },
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, vecstride;
1135 	struct op *fop;
1136 
1137 	vecstride = (1 + ((fpscr & FPSCR_STRIDE_MASK) == FPSCR_STRIDE_MASK));
1138 
1139 	fop = (op == FOP_EXT) ? &fops_ext[FEXT_TO_IDX(inst)] : &fops[FOP_TO_IDX(op)];
1140 
1141 	/*
1142 	 * fcvtds takes an sN register number as destination, not dN.
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 dN operand.
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, vector length is always '1'.
1160 	 * ARM DDI0100F C5.1.3, C5.3.2.
1161 	 */
1162 	if ((fop->flags & OP_SCALAR) || (FREG_BANK(dest) == 0))
1163 		veclen = 0;
1164 	else
1165 		veclen = fpscr & FPSCR_LENGTH_MASK;
1166 
1167 	pr_debug("VFP: vecstride=%u veclen=%u\n", vecstride,
1168 		 (veclen >> FPSCR_LENGTH_BIT) + 1);
1169 
1170 	if (!fop->fn)
1171 		goto invalid;
1172 
1173 	for (vecitr = 0; vecitr <= veclen; vecitr += 1 << FPSCR_LENGTH_BIT) {
1174 		u32 except;
1175 		char type;
1176 
1177 		type = fop->flags & OP_SD ? 's' : 'd';
1178 		if (op == FOP_EXT)
1179 			pr_debug("VFP: itr%d (%c%u) = op[%u] (d%u)\n",
1180 				 vecitr >> FPSCR_LENGTH_BIT,
1181 				 type, dest, dn, dm);
1182 		else
1183 			pr_debug("VFP: itr%d (%c%u) = (d%u) op[%u] (d%u)\n",
1184 				 vecitr >> FPSCR_LENGTH_BIT,
1185 				 type, dest, dn, FOP_TO_IDX(op), dm);
1186 
1187 		except = fop->fn(dest, dn, dm, fpscr);
1188 		pr_debug("VFP: itr%d: exceptions=%08x\n",
1189 			 vecitr >> FPSCR_LENGTH_BIT, except);
1190 
1191 		exceptions |= except;
1192 
1193 		/*
1194 		 * CHECK: It appears to be undefined whether we stop when
1195 		 * we encounter an exception.  We continue.
1196 		 */
1197 		dest = FREG_BANK(dest) + ((FREG_IDX(dest) + vecstride) & 3);
1198 		dn = FREG_BANK(dn) + ((FREG_IDX(dn) + vecstride) & 3);
1199 		if (FREG_BANK(dm) != 0)
1200 			dm = FREG_BANK(dm) + ((FREG_IDX(dm) + vecstride) & 3);
1201 	}
1202 	return exceptions;
1203 
1204  invalid:
1205 	return ~0;
1206 }
1207