xref: /linux/arch/arm/vfp/vfpsingle.c (revision f24e9f586b377749dff37554696cf3a105540c94)
1 /*
2  *  linux/arch/arm/vfp/vfpsingle.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/ptrace.h>
38 #include <asm/vfp.h>
39 
40 #include "vfpinstr.h"
41 #include "vfp.h"
42 
43 static struct vfp_single vfp_single_default_qnan = {
44 	.exponent	= 255,
45 	.sign		= 0,
46 	.significand	= VFP_SINGLE_SIGNIFICAND_QNAN,
47 };
48 
49 static void vfp_single_dump(const char *str, struct vfp_single *s)
50 {
51 	pr_debug("VFP: %s: sign=%d exponent=%d significand=%08x\n",
52 		 str, s->sign != 0, s->exponent, s->significand);
53 }
54 
55 static void vfp_single_normalise_denormal(struct vfp_single *vs)
56 {
57 	int bits = 31 - fls(vs->significand);
58 
59 	vfp_single_dump("normalise_denormal: in", vs);
60 
61 	if (bits) {
62 		vs->exponent -= bits - 1;
63 		vs->significand <<= bits;
64 	}
65 
66 	vfp_single_dump("normalise_denormal: out", vs);
67 }
68 
69 #ifndef DEBUG
70 #define vfp_single_normaliseround(sd,vsd,fpscr,except,func) __vfp_single_normaliseround(sd,vsd,fpscr,except)
71 u32 __vfp_single_normaliseround(int sd, struct vfp_single *vs, u32 fpscr, u32 exceptions)
72 #else
73 u32 vfp_single_normaliseround(int sd, struct vfp_single *vs, u32 fpscr, u32 exceptions, const char *func)
74 #endif
75 {
76 	u32 significand, incr, rmode;
77 	int exponent, shift, underflow;
78 
79 	vfp_single_dump("pack: in", vs);
80 
81 	/*
82 	 * Infinities and NaNs are a special case.
83 	 */
84 	if (vs->exponent == 255 && (vs->significand == 0 || exceptions))
85 		goto pack;
86 
87 	/*
88 	 * Special-case zero.
89 	 */
90 	if (vs->significand == 0) {
91 		vs->exponent = 0;
92 		goto pack;
93 	}
94 
95 	exponent = vs->exponent;
96 	significand = vs->significand;
97 
98 	/*
99 	 * Normalise first.  Note that we shift the significand up to
100 	 * bit 31, so we have VFP_SINGLE_LOW_BITS + 1 below the least
101 	 * significant bit.
102 	 */
103 	shift = 32 - fls(significand);
104 	if (shift < 32 && shift) {
105 		exponent -= shift;
106 		significand <<= shift;
107 	}
108 
109 #ifdef DEBUG
110 	vs->exponent = exponent;
111 	vs->significand = significand;
112 	vfp_single_dump("pack: normalised", vs);
113 #endif
114 
115 	/*
116 	 * Tiny number?
117 	 */
118 	underflow = exponent < 0;
119 	if (underflow) {
120 		significand = vfp_shiftright32jamming(significand, -exponent);
121 		exponent = 0;
122 #ifdef DEBUG
123 		vs->exponent = exponent;
124 		vs->significand = significand;
125 		vfp_single_dump("pack: tiny number", vs);
126 #endif
127 		if (!(significand & ((1 << (VFP_SINGLE_LOW_BITS + 1)) - 1)))
128 			underflow = 0;
129 	}
130 
131 	/*
132 	 * Select rounding increment.
133 	 */
134 	incr = 0;
135 	rmode = fpscr & FPSCR_RMODE_MASK;
136 
137 	if (rmode == FPSCR_ROUND_NEAREST) {
138 		incr = 1 << VFP_SINGLE_LOW_BITS;
139 		if ((significand & (1 << (VFP_SINGLE_LOW_BITS + 1))) == 0)
140 			incr -= 1;
141 	} else if (rmode == FPSCR_ROUND_TOZERO) {
142 		incr = 0;
143 	} else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vs->sign != 0))
144 		incr = (1 << (VFP_SINGLE_LOW_BITS + 1)) - 1;
145 
146 	pr_debug("VFP: rounding increment = 0x%08x\n", incr);
147 
148 	/*
149 	 * Is our rounding going to overflow?
150 	 */
151 	if ((significand + incr) < significand) {
152 		exponent += 1;
153 		significand = (significand >> 1) | (significand & 1);
154 		incr >>= 1;
155 #ifdef DEBUG
156 		vs->exponent = exponent;
157 		vs->significand = significand;
158 		vfp_single_dump("pack: overflow", vs);
159 #endif
160 	}
161 
162 	/*
163 	 * If any of the low bits (which will be shifted out of the
164 	 * number) are non-zero, the result is inexact.
165 	 */
166 	if (significand & ((1 << (VFP_SINGLE_LOW_BITS + 1)) - 1))
167 		exceptions |= FPSCR_IXC;
168 
169 	/*
170 	 * Do our rounding.
171 	 */
172 	significand += incr;
173 
174 	/*
175 	 * Infinity?
176 	 */
177 	if (exponent >= 254) {
178 		exceptions |= FPSCR_OFC | FPSCR_IXC;
179 		if (incr == 0) {
180 			vs->exponent = 253;
181 			vs->significand = 0x7fffffff;
182 		} else {
183 			vs->exponent = 255;		/* infinity */
184 			vs->significand = 0;
185 		}
186 	} else {
187 		if (significand >> (VFP_SINGLE_LOW_BITS + 1) == 0)
188 			exponent = 0;
189 		if (exponent || significand > 0x80000000)
190 			underflow = 0;
191 		if (underflow)
192 			exceptions |= FPSCR_UFC;
193 		vs->exponent = exponent;
194 		vs->significand = significand >> 1;
195 	}
196 
197  pack:
198 	vfp_single_dump("pack: final", vs);
199 	{
200 		s32 d = vfp_single_pack(vs);
201 		pr_debug("VFP: %s: d(s%d)=%08x exceptions=%08x\n", func,
202 			 sd, d, exceptions);
203 		vfp_put_float(d, sd);
204 	}
205 
206 	return exceptions;
207 }
208 
209 /*
210  * Propagate the NaN, setting exceptions if it is signalling.
211  * 'n' is always a NaN.  'm' may be a number, NaN or infinity.
212  */
213 static u32
214 vfp_propagate_nan(struct vfp_single *vsd, struct vfp_single *vsn,
215 		  struct vfp_single *vsm, u32 fpscr)
216 {
217 	struct vfp_single *nan;
218 	int tn, tm = 0;
219 
220 	tn = vfp_single_type(vsn);
221 
222 	if (vsm)
223 		tm = vfp_single_type(vsm);
224 
225 	if (fpscr & FPSCR_DEFAULT_NAN)
226 		/*
227 		 * Default NaN mode - always returns a quiet NaN
228 		 */
229 		nan = &vfp_single_default_qnan;
230 	else {
231 		/*
232 		 * Contemporary mode - select the first signalling
233 		 * NAN, or if neither are signalling, the first
234 		 * quiet NAN.
235 		 */
236 		if (tn == VFP_SNAN || (tm != VFP_SNAN && tn == VFP_QNAN))
237 			nan = vsn;
238 		else
239 			nan = vsm;
240 		/*
241 		 * Make the NaN quiet.
242 		 */
243 		nan->significand |= VFP_SINGLE_SIGNIFICAND_QNAN;
244 	}
245 
246 	*vsd = *nan;
247 
248 	/*
249 	 * If one was a signalling NAN, raise invalid operation.
250 	 */
251 	return tn == VFP_SNAN || tm == VFP_SNAN ? FPSCR_IOC : VFP_NAN_FLAG;
252 }
253 
254 
255 /*
256  * Extended operations
257  */
258 static u32 vfp_single_fabs(int sd, int unused, s32 m, u32 fpscr)
259 {
260 	vfp_put_float(vfp_single_packed_abs(m), sd);
261 	return 0;
262 }
263 
264 static u32 vfp_single_fcpy(int sd, int unused, s32 m, u32 fpscr)
265 {
266 	vfp_put_float(m, sd);
267 	return 0;
268 }
269 
270 static u32 vfp_single_fneg(int sd, int unused, s32 m, u32 fpscr)
271 {
272 	vfp_put_float(vfp_single_packed_negate(m), sd);
273 	return 0;
274 }
275 
276 static const u16 sqrt_oddadjust[] = {
277 	0x0004, 0x0022, 0x005d, 0x00b1, 0x011d, 0x019f, 0x0236, 0x02e0,
278 	0x039c, 0x0468, 0x0545, 0x0631, 0x072b, 0x0832, 0x0946, 0x0a67
279 };
280 
281 static const u16 sqrt_evenadjust[] = {
282 	0x0a2d, 0x08af, 0x075a, 0x0629, 0x051a, 0x0429, 0x0356, 0x029e,
283 	0x0200, 0x0179, 0x0109, 0x00af, 0x0068, 0x0034, 0x0012, 0x0002
284 };
285 
286 u32 vfp_estimate_sqrt_significand(u32 exponent, u32 significand)
287 {
288 	int index;
289 	u32 z, a;
290 
291 	if ((significand & 0xc0000000) != 0x40000000) {
292 		printk(KERN_WARNING "VFP: estimate_sqrt: invalid significand\n");
293 	}
294 
295 	a = significand << 1;
296 	index = (a >> 27) & 15;
297 	if (exponent & 1) {
298 		z = 0x4000 + (a >> 17) - sqrt_oddadjust[index];
299 		z = ((a / z) << 14) + (z << 15);
300 		a >>= 1;
301 	} else {
302 		z = 0x8000 + (a >> 17) - sqrt_evenadjust[index];
303 		z = a / z + z;
304 		z = (z >= 0x20000) ? 0xffff8000 : (z << 15);
305 		if (z <= a)
306 			return (s32)a >> 1;
307 	}
308 	{
309 		u64 v = (u64)a << 31;
310 		do_div(v, z);
311 		return v + (z >> 1);
312 	}
313 }
314 
315 static u32 vfp_single_fsqrt(int sd, int unused, s32 m, u32 fpscr)
316 {
317 	struct vfp_single vsm, vsd;
318 	int ret, tm;
319 
320 	vfp_single_unpack(&vsm, m);
321 	tm = vfp_single_type(&vsm);
322 	if (tm & (VFP_NAN|VFP_INFINITY)) {
323 		struct vfp_single *vsp = &vsd;
324 
325 		if (tm & VFP_NAN)
326 			ret = vfp_propagate_nan(vsp, &vsm, NULL, fpscr);
327 		else if (vsm.sign == 0) {
328  sqrt_copy:
329 			vsp = &vsm;
330 			ret = 0;
331 		} else {
332  sqrt_invalid:
333 			vsp = &vfp_single_default_qnan;
334 			ret = FPSCR_IOC;
335 		}
336 		vfp_put_float(vfp_single_pack(vsp), sd);
337 		return ret;
338 	}
339 
340 	/*
341 	 * sqrt(+/- 0) == +/- 0
342 	 */
343 	if (tm & VFP_ZERO)
344 		goto sqrt_copy;
345 
346 	/*
347 	 * Normalise a denormalised number
348 	 */
349 	if (tm & VFP_DENORMAL)
350 		vfp_single_normalise_denormal(&vsm);
351 
352 	/*
353 	 * sqrt(<0) = invalid
354 	 */
355 	if (vsm.sign)
356 		goto sqrt_invalid;
357 
358 	vfp_single_dump("sqrt", &vsm);
359 
360 	/*
361 	 * Estimate the square root.
362 	 */
363 	vsd.sign = 0;
364 	vsd.exponent = ((vsm.exponent - 127) >> 1) + 127;
365 	vsd.significand = vfp_estimate_sqrt_significand(vsm.exponent, vsm.significand) + 2;
366 
367 	vfp_single_dump("sqrt estimate", &vsd);
368 
369 	/*
370 	 * And now adjust.
371 	 */
372 	if ((vsd.significand & VFP_SINGLE_LOW_BITS_MASK) <= 5) {
373 		if (vsd.significand < 2) {
374 			vsd.significand = 0xffffffff;
375 		} else {
376 			u64 term;
377 			s64 rem;
378 			vsm.significand <<= !(vsm.exponent & 1);
379 			term = (u64)vsd.significand * vsd.significand;
380 			rem = ((u64)vsm.significand << 32) - term;
381 
382 			pr_debug("VFP: term=%016llx rem=%016llx\n", term, rem);
383 
384 			while (rem < 0) {
385 				vsd.significand -= 1;
386 				rem += ((u64)vsd.significand << 1) | 1;
387 			}
388 			vsd.significand |= rem != 0;
389 		}
390 	}
391 	vsd.significand = vfp_shiftright32jamming(vsd.significand, 1);
392 
393 	return vfp_single_normaliseround(sd, &vsd, fpscr, 0, "fsqrt");
394 }
395 
396 /*
397  * Equal	:= ZC
398  * Less than	:= N
399  * Greater than	:= C
400  * Unordered	:= CV
401  */
402 static u32 vfp_compare(int sd, int signal_on_qnan, s32 m, u32 fpscr)
403 {
404 	s32 d;
405 	u32 ret = 0;
406 
407 	d = vfp_get_float(sd);
408 	if (vfp_single_packed_exponent(m) == 255 && vfp_single_packed_mantissa(m)) {
409 		ret |= FPSCR_C | FPSCR_V;
410 		if (signal_on_qnan || !(vfp_single_packed_mantissa(m) & (1 << (VFP_SINGLE_MANTISSA_BITS - 1))))
411 			/*
412 			 * Signalling NaN, or signalling on quiet NaN
413 			 */
414 			ret |= FPSCR_IOC;
415 	}
416 
417 	if (vfp_single_packed_exponent(d) == 255 && vfp_single_packed_mantissa(d)) {
418 		ret |= FPSCR_C | FPSCR_V;
419 		if (signal_on_qnan || !(vfp_single_packed_mantissa(d) & (1 << (VFP_SINGLE_MANTISSA_BITS - 1))))
420 			/*
421 			 * Signalling NaN, or signalling on quiet NaN
422 			 */
423 			ret |= FPSCR_IOC;
424 	}
425 
426 	if (ret == 0) {
427 		if (d == m || vfp_single_packed_abs(d | m) == 0) {
428 			/*
429 			 * equal
430 			 */
431 			ret |= FPSCR_Z | FPSCR_C;
432 		} else if (vfp_single_packed_sign(d ^ m)) {
433 			/*
434 			 * different signs
435 			 */
436 			if (vfp_single_packed_sign(d))
437 				/*
438 				 * d is negative, so d < m
439 				 */
440 				ret |= FPSCR_N;
441 			else
442 				/*
443 				 * d is positive, so d > m
444 				 */
445 				ret |= FPSCR_C;
446 		} else if ((vfp_single_packed_sign(d) != 0) ^ (d < m)) {
447 			/*
448 			 * d < m
449 			 */
450 			ret |= FPSCR_N;
451 		} else if ((vfp_single_packed_sign(d) != 0) ^ (d > m)) {
452 			/*
453 			 * d > m
454 			 */
455 			ret |= FPSCR_C;
456 		}
457 	}
458 	return ret;
459 }
460 
461 static u32 vfp_single_fcmp(int sd, int unused, s32 m, u32 fpscr)
462 {
463 	return vfp_compare(sd, 0, m, fpscr);
464 }
465 
466 static u32 vfp_single_fcmpe(int sd, int unused, s32 m, u32 fpscr)
467 {
468 	return vfp_compare(sd, 1, m, fpscr);
469 }
470 
471 static u32 vfp_single_fcmpz(int sd, int unused, s32 m, u32 fpscr)
472 {
473 	return vfp_compare(sd, 0, 0, fpscr);
474 }
475 
476 static u32 vfp_single_fcmpez(int sd, int unused, s32 m, u32 fpscr)
477 {
478 	return vfp_compare(sd, 1, 0, fpscr);
479 }
480 
481 static u32 vfp_single_fcvtd(int dd, int unused, s32 m, u32 fpscr)
482 {
483 	struct vfp_single vsm;
484 	struct vfp_double vdd;
485 	int tm;
486 	u32 exceptions = 0;
487 
488 	vfp_single_unpack(&vsm, m);
489 
490 	tm = vfp_single_type(&vsm);
491 
492 	/*
493 	 * If we have a signalling NaN, signal invalid operation.
494 	 */
495 	if (tm == VFP_SNAN)
496 		exceptions = FPSCR_IOC;
497 
498 	if (tm & VFP_DENORMAL)
499 		vfp_single_normalise_denormal(&vsm);
500 
501 	vdd.sign = vsm.sign;
502 	vdd.significand = (u64)vsm.significand << 32;
503 
504 	/*
505 	 * If we have an infinity or NaN, the exponent must be 2047.
506 	 */
507 	if (tm & (VFP_INFINITY|VFP_NAN)) {
508 		vdd.exponent = 2047;
509 		if (tm == VFP_QNAN)
510 			vdd.significand |= VFP_DOUBLE_SIGNIFICAND_QNAN;
511 		goto pack_nan;
512 	} else if (tm & VFP_ZERO)
513 		vdd.exponent = 0;
514 	else
515 		vdd.exponent = vsm.exponent + (1023 - 127);
516 
517 	return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fcvtd");
518 
519  pack_nan:
520 	vfp_put_double(vfp_double_pack(&vdd), dd);
521 	return exceptions;
522 }
523 
524 static u32 vfp_single_fuito(int sd, int unused, s32 m, u32 fpscr)
525 {
526 	struct vfp_single vs;
527 
528 	vs.sign = 0;
529 	vs.exponent = 127 + 31 - 1;
530 	vs.significand = (u32)m;
531 
532 	return vfp_single_normaliseround(sd, &vs, fpscr, 0, "fuito");
533 }
534 
535 static u32 vfp_single_fsito(int sd, int unused, s32 m, u32 fpscr)
536 {
537 	struct vfp_single vs;
538 
539 	vs.sign = (m & 0x80000000) >> 16;
540 	vs.exponent = 127 + 31 - 1;
541 	vs.significand = vs.sign ? -m : m;
542 
543 	return vfp_single_normaliseround(sd, &vs, fpscr, 0, "fsito");
544 }
545 
546 static u32 vfp_single_ftoui(int sd, int unused, s32 m, u32 fpscr)
547 {
548 	struct vfp_single vsm;
549 	u32 d, exceptions = 0;
550 	int rmode = fpscr & FPSCR_RMODE_MASK;
551 	int tm;
552 
553 	vfp_single_unpack(&vsm, m);
554 	vfp_single_dump("VSM", &vsm);
555 
556 	/*
557 	 * Do we have a denormalised number?
558 	 */
559 	tm = vfp_single_type(&vsm);
560 	if (tm & VFP_DENORMAL)
561 		exceptions |= FPSCR_IDC;
562 
563 	if (tm & VFP_NAN)
564 		vsm.sign = 0;
565 
566 	if (vsm.exponent >= 127 + 32) {
567 		d = vsm.sign ? 0 : 0xffffffff;
568 		exceptions = FPSCR_IOC;
569 	} else if (vsm.exponent >= 127 - 1) {
570 		int shift = 127 + 31 - vsm.exponent;
571 		u32 rem, incr = 0;
572 
573 		/*
574 		 * 2^0 <= m < 2^32-2^8
575 		 */
576 		d = (vsm.significand << 1) >> shift;
577 		rem = vsm.significand << (33 - shift);
578 
579 		if (rmode == FPSCR_ROUND_NEAREST) {
580 			incr = 0x80000000;
581 			if ((d & 1) == 0)
582 				incr -= 1;
583 		} else if (rmode == FPSCR_ROUND_TOZERO) {
584 			incr = 0;
585 		} else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vsm.sign != 0)) {
586 			incr = ~0;
587 		}
588 
589 		if ((rem + incr) < rem) {
590 			if (d < 0xffffffff)
591 				d += 1;
592 			else
593 				exceptions |= FPSCR_IOC;
594 		}
595 
596 		if (d && vsm.sign) {
597 			d = 0;
598 			exceptions |= FPSCR_IOC;
599 		} else if (rem)
600 			exceptions |= FPSCR_IXC;
601 	} else {
602 		d = 0;
603 		if (vsm.exponent | vsm.significand) {
604 			exceptions |= FPSCR_IXC;
605 			if (rmode == FPSCR_ROUND_PLUSINF && vsm.sign == 0)
606 				d = 1;
607 			else if (rmode == FPSCR_ROUND_MINUSINF && vsm.sign) {
608 				d = 0;
609 				exceptions |= FPSCR_IOC;
610 			}
611 		}
612 	}
613 
614 	pr_debug("VFP: ftoui: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
615 
616 	vfp_put_float(d, sd);
617 
618 	return exceptions;
619 }
620 
621 static u32 vfp_single_ftouiz(int sd, int unused, s32 m, u32 fpscr)
622 {
623 	return vfp_single_ftoui(sd, unused, m, FPSCR_ROUND_TOZERO);
624 }
625 
626 static u32 vfp_single_ftosi(int sd, int unused, s32 m, u32 fpscr)
627 {
628 	struct vfp_single vsm;
629 	u32 d, exceptions = 0;
630 	int rmode = fpscr & FPSCR_RMODE_MASK;
631 	int tm;
632 
633 	vfp_single_unpack(&vsm, m);
634 	vfp_single_dump("VSM", &vsm);
635 
636 	/*
637 	 * Do we have a denormalised number?
638 	 */
639 	tm = vfp_single_type(&vsm);
640 	if (vfp_single_type(&vsm) & VFP_DENORMAL)
641 		exceptions |= FPSCR_IDC;
642 
643 	if (tm & VFP_NAN) {
644 		d = 0;
645 		exceptions |= FPSCR_IOC;
646 	} else if (vsm.exponent >= 127 + 32) {
647 		/*
648 		 * m >= 2^31-2^7: invalid
649 		 */
650 		d = 0x7fffffff;
651 		if (vsm.sign)
652 			d = ~d;
653 		exceptions |= FPSCR_IOC;
654 	} else if (vsm.exponent >= 127 - 1) {
655 		int shift = 127 + 31 - vsm.exponent;
656 		u32 rem, incr = 0;
657 
658 		/* 2^0 <= m <= 2^31-2^7 */
659 		d = (vsm.significand << 1) >> shift;
660 		rem = vsm.significand << (33 - shift);
661 
662 		if (rmode == FPSCR_ROUND_NEAREST) {
663 			incr = 0x80000000;
664 			if ((d & 1) == 0)
665 				incr -= 1;
666 		} else if (rmode == FPSCR_ROUND_TOZERO) {
667 			incr = 0;
668 		} else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vsm.sign != 0)) {
669 			incr = ~0;
670 		}
671 
672 		if ((rem + incr) < rem && d < 0xffffffff)
673 			d += 1;
674 		if (d > 0x7fffffff + (vsm.sign != 0)) {
675 			d = 0x7fffffff + (vsm.sign != 0);
676 			exceptions |= FPSCR_IOC;
677 		} else if (rem)
678 			exceptions |= FPSCR_IXC;
679 
680 		if (vsm.sign)
681 			d = -d;
682 	} else {
683 		d = 0;
684 		if (vsm.exponent | vsm.significand) {
685 			exceptions |= FPSCR_IXC;
686 			if (rmode == FPSCR_ROUND_PLUSINF && vsm.sign == 0)
687 				d = 1;
688 			else if (rmode == FPSCR_ROUND_MINUSINF && vsm.sign)
689 				d = -1;
690 		}
691 	}
692 
693 	pr_debug("VFP: ftosi: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
694 
695 	vfp_put_float((s32)d, sd);
696 
697 	return exceptions;
698 }
699 
700 static u32 vfp_single_ftosiz(int sd, int unused, s32 m, u32 fpscr)
701 {
702 	return vfp_single_ftosi(sd, unused, m, FPSCR_ROUND_TOZERO);
703 }
704 
705 static u32 (* const fop_extfns[32])(int sd, int unused, s32 m, u32 fpscr) = {
706 	[FEXT_TO_IDX(FEXT_FCPY)]	= vfp_single_fcpy,
707 	[FEXT_TO_IDX(FEXT_FABS)]	= vfp_single_fabs,
708 	[FEXT_TO_IDX(FEXT_FNEG)]	= vfp_single_fneg,
709 	[FEXT_TO_IDX(FEXT_FSQRT)]	= vfp_single_fsqrt,
710 	[FEXT_TO_IDX(FEXT_FCMP)]	= vfp_single_fcmp,
711 	[FEXT_TO_IDX(FEXT_FCMPE)]	= vfp_single_fcmpe,
712 	[FEXT_TO_IDX(FEXT_FCMPZ)]	= vfp_single_fcmpz,
713 	[FEXT_TO_IDX(FEXT_FCMPEZ)]	= vfp_single_fcmpez,
714 	[FEXT_TO_IDX(FEXT_FCVT)]	= vfp_single_fcvtd,
715 	[FEXT_TO_IDX(FEXT_FUITO)]	= vfp_single_fuito,
716 	[FEXT_TO_IDX(FEXT_FSITO)]	= vfp_single_fsito,
717 	[FEXT_TO_IDX(FEXT_FTOUI)]	= vfp_single_ftoui,
718 	[FEXT_TO_IDX(FEXT_FTOUIZ)]	= vfp_single_ftouiz,
719 	[FEXT_TO_IDX(FEXT_FTOSI)]	= vfp_single_ftosi,
720 	[FEXT_TO_IDX(FEXT_FTOSIZ)]	= vfp_single_ftosiz,
721 };
722 
723 
724 
725 
726 
727 static u32
728 vfp_single_fadd_nonnumber(struct vfp_single *vsd, struct vfp_single *vsn,
729 			  struct vfp_single *vsm, u32 fpscr)
730 {
731 	struct vfp_single *vsp;
732 	u32 exceptions = 0;
733 	int tn, tm;
734 
735 	tn = vfp_single_type(vsn);
736 	tm = vfp_single_type(vsm);
737 
738 	if (tn & tm & VFP_INFINITY) {
739 		/*
740 		 * Two infinities.  Are they different signs?
741 		 */
742 		if (vsn->sign ^ vsm->sign) {
743 			/*
744 			 * different signs -> invalid
745 			 */
746 			exceptions = FPSCR_IOC;
747 			vsp = &vfp_single_default_qnan;
748 		} else {
749 			/*
750 			 * same signs -> valid
751 			 */
752 			vsp = vsn;
753 		}
754 	} else if (tn & VFP_INFINITY && tm & VFP_NUMBER) {
755 		/*
756 		 * One infinity and one number -> infinity
757 		 */
758 		vsp = vsn;
759 	} else {
760 		/*
761 		 * 'n' is a NaN of some type
762 		 */
763 		return vfp_propagate_nan(vsd, vsn, vsm, fpscr);
764 	}
765 	*vsd = *vsp;
766 	return exceptions;
767 }
768 
769 static u32
770 vfp_single_add(struct vfp_single *vsd, struct vfp_single *vsn,
771 	       struct vfp_single *vsm, u32 fpscr)
772 {
773 	u32 exp_diff, m_sig;
774 
775 	if (vsn->significand & 0x80000000 ||
776 	    vsm->significand & 0x80000000) {
777 		pr_info("VFP: bad FP values in %s\n", __func__);
778 		vfp_single_dump("VSN", vsn);
779 		vfp_single_dump("VSM", vsm);
780 	}
781 
782 	/*
783 	 * Ensure that 'n' is the largest magnitude number.  Note that
784 	 * if 'n' and 'm' have equal exponents, we do not swap them.
785 	 * This ensures that NaN propagation works correctly.
786 	 */
787 	if (vsn->exponent < vsm->exponent) {
788 		struct vfp_single *t = vsn;
789 		vsn = vsm;
790 		vsm = t;
791 	}
792 
793 	/*
794 	 * Is 'n' an infinity or a NaN?  Note that 'm' may be a number,
795 	 * infinity or a NaN here.
796 	 */
797 	if (vsn->exponent == 255)
798 		return vfp_single_fadd_nonnumber(vsd, vsn, vsm, fpscr);
799 
800 	/*
801 	 * We have two proper numbers, where 'vsn' is the larger magnitude.
802 	 *
803 	 * Copy 'n' to 'd' before doing the arithmetic.
804 	 */
805 	*vsd = *vsn;
806 
807 	/*
808 	 * Align both numbers.
809 	 */
810 	exp_diff = vsn->exponent - vsm->exponent;
811 	m_sig = vfp_shiftright32jamming(vsm->significand, exp_diff);
812 
813 	/*
814 	 * If the signs are different, we are really subtracting.
815 	 */
816 	if (vsn->sign ^ vsm->sign) {
817 		m_sig = vsn->significand - m_sig;
818 		if ((s32)m_sig < 0) {
819 			vsd->sign = vfp_sign_negate(vsd->sign);
820 			m_sig = -m_sig;
821 		} else if (m_sig == 0) {
822 			vsd->sign = (fpscr & FPSCR_RMODE_MASK) ==
823 				      FPSCR_ROUND_MINUSINF ? 0x8000 : 0;
824 		}
825 	} else {
826 		m_sig = vsn->significand + m_sig;
827 	}
828 	vsd->significand = m_sig;
829 
830 	return 0;
831 }
832 
833 static u32
834 vfp_single_multiply(struct vfp_single *vsd, struct vfp_single *vsn, struct vfp_single *vsm, u32 fpscr)
835 {
836 	vfp_single_dump("VSN", vsn);
837 	vfp_single_dump("VSM", vsm);
838 
839 	/*
840 	 * Ensure that 'n' is the largest magnitude number.  Note that
841 	 * if 'n' and 'm' have equal exponents, we do not swap them.
842 	 * This ensures that NaN propagation works correctly.
843 	 */
844 	if (vsn->exponent < vsm->exponent) {
845 		struct vfp_single *t = vsn;
846 		vsn = vsm;
847 		vsm = t;
848 		pr_debug("VFP: swapping M <-> N\n");
849 	}
850 
851 	vsd->sign = vsn->sign ^ vsm->sign;
852 
853 	/*
854 	 * If 'n' is an infinity or NaN, handle it.  'm' may be anything.
855 	 */
856 	if (vsn->exponent == 255) {
857 		if (vsn->significand || (vsm->exponent == 255 && vsm->significand))
858 			return vfp_propagate_nan(vsd, vsn, vsm, fpscr);
859 		if ((vsm->exponent | vsm->significand) == 0) {
860 			*vsd = vfp_single_default_qnan;
861 			return FPSCR_IOC;
862 		}
863 		vsd->exponent = vsn->exponent;
864 		vsd->significand = 0;
865 		return 0;
866 	}
867 
868 	/*
869 	 * If 'm' is zero, the result is always zero.  In this case,
870 	 * 'n' may be zero or a number, but it doesn't matter which.
871 	 */
872 	if ((vsm->exponent | vsm->significand) == 0) {
873 		vsd->exponent = 0;
874 		vsd->significand = 0;
875 		return 0;
876 	}
877 
878 	/*
879 	 * We add 2 to the destination exponent for the same reason as
880 	 * the addition case - though this time we have +1 from each
881 	 * input operand.
882 	 */
883 	vsd->exponent = vsn->exponent + vsm->exponent - 127 + 2;
884 	vsd->significand = vfp_hi64to32jamming((u64)vsn->significand * vsm->significand);
885 
886 	vfp_single_dump("VSD", vsd);
887 	return 0;
888 }
889 
890 #define NEG_MULTIPLY	(1 << 0)
891 #define NEG_SUBTRACT	(1 << 1)
892 
893 static u32
894 vfp_single_multiply_accumulate(int sd, int sn, s32 m, u32 fpscr, u32 negate, char *func)
895 {
896 	struct vfp_single vsd, vsp, vsn, vsm;
897 	u32 exceptions;
898 	s32 v;
899 
900 	v = vfp_get_float(sn);
901 	pr_debug("VFP: s%u = %08x\n", sn, v);
902 	vfp_single_unpack(&vsn, v);
903 	if (vsn.exponent == 0 && vsn.significand)
904 		vfp_single_normalise_denormal(&vsn);
905 
906 	vfp_single_unpack(&vsm, m);
907 	if (vsm.exponent == 0 && vsm.significand)
908 		vfp_single_normalise_denormal(&vsm);
909 
910 	exceptions = vfp_single_multiply(&vsp, &vsn, &vsm, fpscr);
911 	if (negate & NEG_MULTIPLY)
912 		vsp.sign = vfp_sign_negate(vsp.sign);
913 
914 	v = vfp_get_float(sd);
915 	pr_debug("VFP: s%u = %08x\n", sd, v);
916 	vfp_single_unpack(&vsn, v);
917 	if (negate & NEG_SUBTRACT)
918 		vsn.sign = vfp_sign_negate(vsn.sign);
919 
920 	exceptions |= vfp_single_add(&vsd, &vsn, &vsp, fpscr);
921 
922 	return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, func);
923 }
924 
925 /*
926  * Standard operations
927  */
928 
929 /*
930  * sd = sd + (sn * sm)
931  */
932 static u32 vfp_single_fmac(int sd, int sn, s32 m, u32 fpscr)
933 {
934 	return vfp_single_multiply_accumulate(sd, sn, m, fpscr, 0, "fmac");
935 }
936 
937 /*
938  * sd = sd - (sn * sm)
939  */
940 static u32 vfp_single_fnmac(int sd, int sn, s32 m, u32 fpscr)
941 {
942 	return vfp_single_multiply_accumulate(sd, sn, m, fpscr, NEG_MULTIPLY, "fnmac");
943 }
944 
945 /*
946  * sd = -sd + (sn * sm)
947  */
948 static u32 vfp_single_fmsc(int sd, int sn, s32 m, u32 fpscr)
949 {
950 	return vfp_single_multiply_accumulate(sd, sn, m, fpscr, NEG_SUBTRACT, "fmsc");
951 }
952 
953 /*
954  * sd = -sd - (sn * sm)
955  */
956 static u32 vfp_single_fnmsc(int sd, int sn, s32 m, u32 fpscr)
957 {
958 	return vfp_single_multiply_accumulate(sd, sn, m, fpscr, NEG_SUBTRACT | NEG_MULTIPLY, "fnmsc");
959 }
960 
961 /*
962  * sd = sn * sm
963  */
964 static u32 vfp_single_fmul(int sd, int sn, s32 m, u32 fpscr)
965 {
966 	struct vfp_single vsd, vsn, vsm;
967 	u32 exceptions;
968 	s32 n = vfp_get_float(sn);
969 
970 	pr_debug("VFP: s%u = %08x\n", sn, n);
971 
972 	vfp_single_unpack(&vsn, n);
973 	if (vsn.exponent == 0 && vsn.significand)
974 		vfp_single_normalise_denormal(&vsn);
975 
976 	vfp_single_unpack(&vsm, m);
977 	if (vsm.exponent == 0 && vsm.significand)
978 		vfp_single_normalise_denormal(&vsm);
979 
980 	exceptions = vfp_single_multiply(&vsd, &vsn, &vsm, fpscr);
981 	return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fmul");
982 }
983 
984 /*
985  * sd = -(sn * sm)
986  */
987 static u32 vfp_single_fnmul(int sd, int sn, s32 m, u32 fpscr)
988 {
989 	struct vfp_single vsd, vsn, vsm;
990 	u32 exceptions;
991 	s32 n = vfp_get_float(sn);
992 
993 	pr_debug("VFP: s%u = %08x\n", sn, n);
994 
995 	vfp_single_unpack(&vsn, n);
996 	if (vsn.exponent == 0 && vsn.significand)
997 		vfp_single_normalise_denormal(&vsn);
998 
999 	vfp_single_unpack(&vsm, m);
1000 	if (vsm.exponent == 0 && vsm.significand)
1001 		vfp_single_normalise_denormal(&vsm);
1002 
1003 	exceptions = vfp_single_multiply(&vsd, &vsn, &vsm, fpscr);
1004 	vsd.sign = vfp_sign_negate(vsd.sign);
1005 	return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fnmul");
1006 }
1007 
1008 /*
1009  * sd = sn + sm
1010  */
1011 static u32 vfp_single_fadd(int sd, int sn, s32 m, u32 fpscr)
1012 {
1013 	struct vfp_single vsd, vsn, vsm;
1014 	u32 exceptions;
1015 	s32 n = vfp_get_float(sn);
1016 
1017 	pr_debug("VFP: s%u = %08x\n", sn, n);
1018 
1019 	/*
1020 	 * Unpack and normalise denormals.
1021 	 */
1022 	vfp_single_unpack(&vsn, n);
1023 	if (vsn.exponent == 0 && vsn.significand)
1024 		vfp_single_normalise_denormal(&vsn);
1025 
1026 	vfp_single_unpack(&vsm, m);
1027 	if (vsm.exponent == 0 && vsm.significand)
1028 		vfp_single_normalise_denormal(&vsm);
1029 
1030 	exceptions = vfp_single_add(&vsd, &vsn, &vsm, fpscr);
1031 
1032 	return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fadd");
1033 }
1034 
1035 /*
1036  * sd = sn - sm
1037  */
1038 static u32 vfp_single_fsub(int sd, int sn, s32 m, u32 fpscr)
1039 {
1040 	/*
1041 	 * Subtraction is addition with one sign inverted.
1042 	 */
1043 	return vfp_single_fadd(sd, sn, vfp_single_packed_negate(m), fpscr);
1044 }
1045 
1046 /*
1047  * sd = sn / sm
1048  */
1049 static u32 vfp_single_fdiv(int sd, int sn, s32 m, u32 fpscr)
1050 {
1051 	struct vfp_single vsd, vsn, vsm;
1052 	u32 exceptions = 0;
1053 	s32 n = vfp_get_float(sn);
1054 	int tm, tn;
1055 
1056 	pr_debug("VFP: s%u = %08x\n", sn, n);
1057 
1058 	vfp_single_unpack(&vsn, n);
1059 	vfp_single_unpack(&vsm, m);
1060 
1061 	vsd.sign = vsn.sign ^ vsm.sign;
1062 
1063 	tn = vfp_single_type(&vsn);
1064 	tm = vfp_single_type(&vsm);
1065 
1066 	/*
1067 	 * Is n a NAN?
1068 	 */
1069 	if (tn & VFP_NAN)
1070 		goto vsn_nan;
1071 
1072 	/*
1073 	 * Is m a NAN?
1074 	 */
1075 	if (tm & VFP_NAN)
1076 		goto vsm_nan;
1077 
1078 	/*
1079 	 * If n and m are infinity, the result is invalid
1080 	 * If n and m are zero, the result is invalid
1081 	 */
1082 	if (tm & tn & (VFP_INFINITY|VFP_ZERO))
1083 		goto invalid;
1084 
1085 	/*
1086 	 * If n is infinity, the result is infinity
1087 	 */
1088 	if (tn & VFP_INFINITY)
1089 		goto infinity;
1090 
1091 	/*
1092 	 * If m is zero, raise div0 exception
1093 	 */
1094 	if (tm & VFP_ZERO)
1095 		goto divzero;
1096 
1097 	/*
1098 	 * If m is infinity, or n is zero, the result is zero
1099 	 */
1100 	if (tm & VFP_INFINITY || tn & VFP_ZERO)
1101 		goto zero;
1102 
1103 	if (tn & VFP_DENORMAL)
1104 		vfp_single_normalise_denormal(&vsn);
1105 	if (tm & VFP_DENORMAL)
1106 		vfp_single_normalise_denormal(&vsm);
1107 
1108 	/*
1109 	 * Ok, we have two numbers, we can perform division.
1110 	 */
1111 	vsd.exponent = vsn.exponent - vsm.exponent + 127 - 1;
1112 	vsm.significand <<= 1;
1113 	if (vsm.significand <= (2 * vsn.significand)) {
1114 		vsn.significand >>= 1;
1115 		vsd.exponent++;
1116 	}
1117 	{
1118 		u64 significand = (u64)vsn.significand << 32;
1119 		do_div(significand, vsm.significand);
1120 		vsd.significand = significand;
1121 	}
1122 	if ((vsd.significand & 0x3f) == 0)
1123 		vsd.significand |= ((u64)vsm.significand * vsd.significand != (u64)vsn.significand << 32);
1124 
1125 	return vfp_single_normaliseround(sd, &vsd, fpscr, 0, "fdiv");
1126 
1127  vsn_nan:
1128 	exceptions = vfp_propagate_nan(&vsd, &vsn, &vsm, fpscr);
1129  pack:
1130 	vfp_put_float(vfp_single_pack(&vsd), sd);
1131 	return exceptions;
1132 
1133  vsm_nan:
1134 	exceptions = vfp_propagate_nan(&vsd, &vsm, &vsn, fpscr);
1135 	goto pack;
1136 
1137  zero:
1138 	vsd.exponent = 0;
1139 	vsd.significand = 0;
1140 	goto pack;
1141 
1142  divzero:
1143 	exceptions = FPSCR_DZC;
1144  infinity:
1145 	vsd.exponent = 255;
1146 	vsd.significand = 0;
1147 	goto pack;
1148 
1149  invalid:
1150 	vfp_put_float(vfp_single_pack(&vfp_single_default_qnan), sd);
1151 	return FPSCR_IOC;
1152 }
1153 
1154 static u32 (* const fop_fns[16])(int sd, int sn, s32 m, u32 fpscr) = {
1155 	[FOP_TO_IDX(FOP_FMAC)]	= vfp_single_fmac,
1156 	[FOP_TO_IDX(FOP_FNMAC)]	= vfp_single_fnmac,
1157 	[FOP_TO_IDX(FOP_FMSC)]	= vfp_single_fmsc,
1158 	[FOP_TO_IDX(FOP_FNMSC)]	= vfp_single_fnmsc,
1159 	[FOP_TO_IDX(FOP_FMUL)]	= vfp_single_fmul,
1160 	[FOP_TO_IDX(FOP_FNMUL)]	= vfp_single_fnmul,
1161 	[FOP_TO_IDX(FOP_FADD)]	= vfp_single_fadd,
1162 	[FOP_TO_IDX(FOP_FSUB)]	= vfp_single_fsub,
1163 	[FOP_TO_IDX(FOP_FDIV)]	= vfp_single_fdiv,
1164 };
1165 
1166 #define FREG_BANK(x)	((x) & 0x18)
1167 #define FREG_IDX(x)	((x) & 7)
1168 
1169 u32 vfp_single_cpdo(u32 inst, u32 fpscr)
1170 {
1171 	u32 op = inst & FOP_MASK;
1172 	u32 exceptions = 0;
1173 	unsigned int dest;
1174 	unsigned int sn = vfp_get_sn(inst);
1175 	unsigned int sm = vfp_get_sm(inst);
1176 	unsigned int vecitr, veclen, vecstride;
1177 	u32 (*fop)(int, int, s32, u32);
1178 
1179 	veclen = fpscr & FPSCR_LENGTH_MASK;
1180 	vecstride = 1 + ((fpscr & FPSCR_STRIDE_MASK) == FPSCR_STRIDE_MASK);
1181 
1182 	/*
1183 	 * fcvtsd takes a dN register number as destination, not sN.
1184 	 * Technically, if bit 0 of dd is set, this is an invalid
1185 	 * instruction.  However, we ignore this for efficiency.
1186 	 * It also only operates on scalars.
1187 	 */
1188 	if ((inst & FEXT_MASK) == FEXT_FCVT) {
1189 		veclen = 0;
1190 		dest = vfp_get_dd(inst);
1191 	} else
1192 		dest = vfp_get_sd(inst);
1193 
1194 	/*
1195 	 * If destination bank is zero, vector length is always '1'.
1196 	 * ARM DDI0100F C5.1.3, C5.3.2.
1197 	 */
1198 	if (FREG_BANK(dest) == 0)
1199 		veclen = 0;
1200 
1201 	pr_debug("VFP: vecstride=%u veclen=%u\n", vecstride,
1202 		 (veclen >> FPSCR_LENGTH_BIT) + 1);
1203 
1204 	fop = (op == FOP_EXT) ? fop_extfns[FEXT_TO_IDX(inst)] : fop_fns[FOP_TO_IDX(op)];
1205 	if (!fop)
1206 		goto invalid;
1207 
1208 	for (vecitr = 0; vecitr <= veclen; vecitr += 1 << FPSCR_LENGTH_BIT) {
1209 		s32 m = vfp_get_float(sm);
1210 		u32 except;
1211 
1212 		if (op == FOP_EXT && (inst & FEXT_MASK) == FEXT_FCVT)
1213 			pr_debug("VFP: itr%d (d%u) = op[%u] (s%u=%08x)\n",
1214 				 vecitr >> FPSCR_LENGTH_BIT, dest, sn, sm, m);
1215 		else if (op == FOP_EXT)
1216 			pr_debug("VFP: itr%d (s%u) = op[%u] (s%u=%08x)\n",
1217 				 vecitr >> FPSCR_LENGTH_BIT, dest, sn, sm, m);
1218 		else
1219 			pr_debug("VFP: itr%d (s%u) = (s%u) op[%u] (s%u=%08x)\n",
1220 				 vecitr >> FPSCR_LENGTH_BIT, dest, sn,
1221 				 FOP_TO_IDX(op), sm, m);
1222 
1223 		except = fop(dest, sn, m, fpscr);
1224 		pr_debug("VFP: itr%d: exceptions=%08x\n",
1225 			 vecitr >> FPSCR_LENGTH_BIT, except);
1226 
1227 		exceptions |= except;
1228 
1229 		/*
1230 		 * This ensures that comparisons only operate on scalars;
1231 		 * comparisons always return with one FPSCR status bit set.
1232 		 */
1233 		if (except & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
1234 			break;
1235 
1236 		/*
1237 		 * CHECK: It appears to be undefined whether we stop when
1238 		 * we encounter an exception.  We continue.
1239 		 */
1240 
1241 		dest = FREG_BANK(dest) + ((FREG_IDX(dest) + vecstride) & 7);
1242 		sn = FREG_BANK(sn) + ((FREG_IDX(sn) + vecstride) & 7);
1243 		if (FREG_BANK(sm) != 0)
1244 			sm = FREG_BANK(sm) + ((FREG_IDX(sm) + vecstride) & 7);
1245 	}
1246 	return exceptions;
1247 
1248  invalid:
1249 	return (u32)-1;
1250 }
1251