xref: /freebsd/contrib/ntp/libntp/ntp_calendar.c (revision ab00ac327a66a53edaac95b536b209db3ae2cd9f)
1 /*
2  * ntp_calendar.c - calendar and helper functions
3  *
4  * Written by Juergen Perlinger (perlinger@ntp.org) for the NTP project.
5  * The contents of 'html/copyright.html' apply.
6  *
7  * --------------------------------------------------------------------
8  * Some notes on the implementation:
9  *
10  * Calendar algorithms thrive on the division operation, which is one of
11  * the slowest numerical operations in any CPU. What saves us here from
12  * abysmal performance is the fact that all divisions are divisions by
13  * constant numbers, and most compilers can do this by a multiplication
14  * operation.  But this might not work when using the div/ldiv/lldiv
15  * function family, because many compilers are not able to do inline
16  * expansion of the code with following optimisation for the
17  * constant-divider case.
18  *
19  * Also div/ldiv/lldiv are defined in terms of int/long/longlong, which
20  * are inherently target dependent. Nothing that could not be cured with
21  * autoconf, but still a mess...
22  *
23  * Furthermore, we need floor division in many places. C either leaves
24  * the division behaviour undefined (< C99) or demands truncation to
25  * zero (>= C99), so additional steps are required to make sure the
26  * algorithms work. The {l,ll}div function family is requested to
27  * truncate towards zero, which is also the wrong direction for our
28  * purpose.
29  *
30  * For all this, all divisions by constant are coded manually, even when
31  * there is a joined div/mod operation: The optimiser should sort that
32  * out, if possible. Most of the calculations are done with unsigned
33  * types, explicitely using two's complement arithmetics where
34  * necessary. This minimises the dependecies to compiler and target,
35  * while still giving reasonable to good performance.
36  *
37  * The implementation uses a few tricks that exploit properties of the
38  * two's complement: Floor division on negative dividents can be
39  * executed by using the one's complement of the divident. One's
40  * complement can be easily created using XOR and a mask.
41  *
42  * Finally, check for overflow conditions is minimal. There are only two
43  * calculation steps in the whole calendar that suffer from an internal
44  * overflow, and these conditions are checked: errno is set to EDOM and
45  * the results are clamped/saturated in this case.  All other functions
46  * do not suffer from internal overflow and simply return the result
47  * truncated to 32 bits.
48  *
49  * This is a sacrifice made for execution speed.  Since a 32-bit day
50  * counter covers +/- 5,879,610 years and the clamp limits the effective
51  * range to +/-2.9 million years, this should not pose a problem here.
52  *
53  */
54 
55 #include <config.h>
56 #include <sys/types.h>
57 
58 #include "ntp_types.h"
59 #include "ntp_calendar.h"
60 #include "ntp_stdlib.h"
61 #include "ntp_fp.h"
62 #include "ntp_unixtime.h"
63 
64 /* For now, let's take the conservative approach: if the target property
65  * macros are not defined, check a few well-known compiler/architecture
66  * settings. Default is to assume that the representation of signed
67  * integers is unknown and shift-arithmetic-right is not available.
68  */
69 #ifndef TARGET_HAS_2CPL
70 # if defined(__GNUC__)
71 #  if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
72 #   define TARGET_HAS_2CPL 1
73 #  else
74 #   define TARGET_HAS_2CPL 0
75 #  endif
76 # elif defined(_MSC_VER)
77 #  if defined(_M_IX86) || defined(_M_X64) || defined(_M_ARM)
78 #   define TARGET_HAS_2CPL 1
79 #  else
80 #   define TARGET_HAS_2CPL 0
81 #  endif
82 # else
83 #  define TARGET_HAS_2CPL 0
84 # endif
85 #endif
86 
87 #ifndef TARGET_HAS_SAR
88 # define TARGET_HAS_SAR 0
89 #endif
90 
91 /*
92  *---------------------------------------------------------------------
93  * replacing the 'time()' function
94  *---------------------------------------------------------------------
95  */
96 
97 static systime_func_ptr systime_func = &time;
98 static inline time_t now(void);
99 
100 
101 systime_func_ptr
102 ntpcal_set_timefunc(
103 	systime_func_ptr nfunc
104 	)
105 {
106 	systime_func_ptr res;
107 
108 	res = systime_func;
109 	if (NULL == nfunc)
110 		nfunc = &time;
111 	systime_func = nfunc;
112 
113 	return res;
114 }
115 
116 
117 static inline time_t
118 now(void)
119 {
120 	return (*systime_func)(NULL);
121 }
122 
123 /*
124  *---------------------------------------------------------------------
125  * Get sign extension mask and unsigned 2cpl rep for a signed integer
126  *---------------------------------------------------------------------
127  */
128 
129 static inline uint32_t
130 int32_sflag(
131 	const int32_t v)
132 {
133 #   if TARGET_HAS_2CPL && TARGET_HAS_SAR && SIZEOF_INT >= 4
134 
135 	/* Let's assume that shift is the fastest way to get the sign
136 	 * extension of of a signed integer. This might not always be
137 	 * true, though -- On 8bit CPUs or machines without barrel
138 	 * shifter this will kill the performance. So we make sure
139 	 * we do this only if 'int' has at least 4 bytes.
140 	 */
141 	return (uint32_t)(v >> 31);
142 
143 #   else
144 
145 	/* This should be a rather generic approach for getting a sign
146 	 * extension mask...
147 	 */
148 	return UINT32_C(0) - (uint32_t)(v < 0);
149 
150 #   endif
151 }
152 
153 static inline uint32_t
154 int32_to_uint32_2cpl(
155 	const int32_t v)
156 {
157 	uint32_t vu;
158 
159 #   if TARGET_HAS_2CPL
160 
161 	/* Just copy through the 32 bits from the signed value if we're
162 	 * on a two's complement target.
163 	 */
164 	vu = (uint32_t)v;
165 
166 #   else
167 
168 	/* Convert from signed int to unsigned int two's complement. Do
169 	 * not make any assumptions about the representation of signed
170 	 * integers, but make sure signed integer overflow cannot happen
171 	 * here. A compiler on a two's complement target *might* find
172 	 * out that this is just a complicated cast (as above), but your
173 	 * mileage might vary.
174 	 */
175 	if (v < 0)
176 		vu = ~(uint32_t)(-(v + 1));
177 	else
178 		vu = (uint32_t)v;
179 
180 #   endif
181 
182 	return vu;
183 }
184 
185 static inline int32_t
186 uint32_2cpl_to_int32(
187 	const uint32_t vu)
188 {
189 	int32_t v;
190 
191 #   if TARGET_HAS_2CPL
192 
193 	/* Just copy through the 32 bits from the unsigned value if
194 	 * we're on a two's complement target.
195 	 */
196 	v = (int32_t)vu;
197 
198 #   else
199 
200 	/* Convert to signed integer, making sure signed integer
201 	 * overflow cannot happen. Again, the optimiser might or might
202 	 * not find out that this is just a copy of 32 bits on a target
203 	 * with two's complement representation for signed integers.
204 	 */
205 	if (vu > INT32_MAX)
206 		v = -(int32_t)(~vu) - 1;
207 	else
208 		v = (int32_t)vu;
209 
210 #   endif
211 
212 	return v;
213 }
214 
215 /* Some of the calculations need to multiply the input by 4 before doing
216  * a division. This can cause overflow and strange results. Therefore we
217  * clamp / saturate the input operand. And since we do the calculations
218  * in unsigned int with an extra sign flag/mask, we only loose one bit
219  * of the input value range.
220  */
221 static inline uint32_t
222 uint32_saturate(
223 	uint32_t vu,
224 	uint32_t mu)
225 {
226 	static const uint32_t limit = UINT32_MAX/4u;
227 	if ((mu ^ vu) > limit) {
228 		vu    = mu ^ limit;
229 		errno = EDOM;
230 	}
231 	return vu;
232 }
233 
234 /*
235  *---------------------------------------------------------------------
236  * Convert between 'time_t' and 'vint64'
237  *---------------------------------------------------------------------
238  */
239 vint64
240 time_to_vint64(
241 	const time_t * ptt
242 	)
243 {
244 	vint64 res;
245 	time_t tt;
246 
247 	tt = *ptt;
248 
249 #   if SIZEOF_TIME_T <= 4
250 
251 	res.D_s.hi = 0;
252 	if (tt < 0) {
253 		res.D_s.lo = (uint32_t)-tt;
254 		M_NEG(res.D_s.hi, res.D_s.lo);
255 	} else {
256 		res.D_s.lo = (uint32_t)tt;
257 	}
258 
259 #   elif defined(HAVE_INT64)
260 
261 	res.q_s = tt;
262 
263 #   else
264 	/*
265 	 * shifting negative signed quantities is compiler-dependent, so
266 	 * we better avoid it and do it all manually. And shifting more
267 	 * than the width of a quantity is undefined. Also a don't do!
268 	 */
269 	if (tt < 0) {
270 		tt = -tt;
271 		res.D_s.lo = (uint32_t)tt;
272 		res.D_s.hi = (uint32_t)(tt >> 32);
273 		M_NEG(res.D_s.hi, res.D_s.lo);
274 	} else {
275 		res.D_s.lo = (uint32_t)tt;
276 		res.D_s.hi = (uint32_t)(tt >> 32);
277 	}
278 
279 #   endif
280 
281 	return res;
282 }
283 
284 
285 time_t
286 vint64_to_time(
287 	const vint64 *tv
288 	)
289 {
290 	time_t res;
291 
292 #   if SIZEOF_TIME_T <= 4
293 
294 	res = (time_t)tv->D_s.lo;
295 
296 #   elif defined(HAVE_INT64)
297 
298 	res = (time_t)tv->q_s;
299 
300 #   else
301 
302 	res = ((time_t)tv->d_s.hi << 32) | tv->D_s.lo;
303 
304 #   endif
305 
306 	return res;
307 }
308 
309 /*
310  *---------------------------------------------------------------------
311  * Get the build date & time
312  *---------------------------------------------------------------------
313  */
314 int
315 ntpcal_get_build_date(
316 	struct calendar * jd
317 	)
318 {
319 	/* The C standard tells us the format of '__DATE__':
320 	 *
321 	 * __DATE__ The date of translation of the preprocessing
322 	 * translation unit: a character string literal of the form "Mmm
323 	 * dd yyyy", where the names of the months are the same as those
324 	 * generated by the asctime function, and the first character of
325 	 * dd is a space character if the value is less than 10. If the
326 	 * date of translation is not available, an
327 	 * implementation-defined valid date shall be supplied.
328 	 *
329 	 * __TIME__ The time of translation of the preprocessing
330 	 * translation unit: a character string literal of the form
331 	 * "hh:mm:ss" as in the time generated by the asctime
332 	 * function. If the time of translation is not available, an
333 	 * implementation-defined valid time shall be supplied.
334 	 *
335 	 * Note that MSVC declares DATE and TIME to be in the local time
336 	 * zone, while neither the C standard nor the GCC docs make any
337 	 * statement about this. As a result, we may be +/-12hrs off
338 	 * UTC.  But for practical purposes, this should not be a
339 	 * problem.
340 	 *
341 	 */
342 #   ifdef MKREPRO_DATE
343 	static const char build[] = MKREPRO_TIME "/" MKREPRO_DATE;
344 #   else
345 	static const char build[] = __TIME__ "/" __DATE__;
346 #   endif
347 	static const char mlist[] = "JanFebMarAprMayJunJulAugSepOctNovDec";
348 
349 	char		  monstr[4];
350 	const char *	  cp;
351 	unsigned short	  hour, minute, second, day, year;
352  	/* Note: The above quantities are used for sscanf 'hu' format,
353 	 * so using 'uint16_t' is contra-indicated!
354 	 */
355 
356 #   ifdef DEBUG
357 	static int        ignore  = 0;
358 #   endif
359 
360 	ZERO(*jd);
361 	jd->year     = 1970;
362 	jd->month    = 1;
363 	jd->monthday = 1;
364 
365 #   ifdef DEBUG
366 	/* check environment if build date should be ignored */
367 	if (0 == ignore) {
368 	    const char * envstr;
369 	    envstr = getenv("NTPD_IGNORE_BUILD_DATE");
370 	    ignore = 1 + (envstr && (!*envstr || !strcasecmp(envstr, "yes")));
371 	}
372 	if (ignore > 1)
373 	    return FALSE;
374 #   endif
375 
376 	if (6 == sscanf(build, "%hu:%hu:%hu/%3s %hu %hu",
377 			&hour, &minute, &second, monstr, &day, &year)) {
378 		cp = strstr(mlist, monstr);
379 		if (NULL != cp) {
380 			jd->year     = year;
381 			jd->month    = (uint8_t)((cp - mlist) / 3 + 1);
382 			jd->monthday = (uint8_t)day;
383 			jd->hour     = (uint8_t)hour;
384 			jd->minute   = (uint8_t)minute;
385 			jd->second   = (uint8_t)second;
386 
387 			return TRUE;
388 		}
389 	}
390 
391 	return FALSE;
392 }
393 
394 
395 /*
396  *---------------------------------------------------------------------
397  * basic calendar stuff
398  *---------------------------------------------------------------------
399  */
400 
401 /* month table for a year starting with March,1st */
402 static const uint16_t shift_month_table[13] = {
403 	0, 31, 61, 92, 122, 153, 184, 214, 245, 275, 306, 337, 366
404 };
405 
406 /* month tables for years starting with January,1st; regular & leap */
407 static const uint16_t real_month_table[2][13] = {
408 	/* -*- table for regular years -*- */
409 	{ 0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334, 365 },
410 	/* -*- table for leap years -*- */
411 	{ 0, 31, 60, 91, 121, 152, 182, 213, 244, 274, 305, 335, 366 }
412 };
413 
414 /*
415  * Some notes on the terminology:
416  *
417  * We use the proleptic Gregorian calendar, which is the Gregorian
418  * calendar extended in both directions ad infinitum. This totally
419  * disregards the fact that this calendar was invented in 1582, and
420  * was adopted at various dates over the world; sometimes even after
421  * the start of the NTP epoch.
422  *
423  * Normally date parts are given as current cycles, while time parts
424  * are given as elapsed cycles:
425  *
426  * 1970-01-01/03:04:05 means 'IN the 1970st. year, IN the first month,
427  * ON the first day, with 3hrs, 4minutes and 5 seconds elapsed.
428  *
429  * The basic calculations for this calendar implementation deal with
430  * ELAPSED date units, which is the number of full years, full months
431  * and full days before a date: 1970-01-01 would be (1969, 0, 0) in
432  * that notation.
433  *
434  * To ease the numeric computations, month and day values outside the
435  * normal range are acceptable: 2001-03-00 will be treated as the day
436  * before 2001-03-01, 2000-13-32 will give the same result as
437  * 2001-02-01 and so on.
438  *
439  * 'rd' or 'RD' is used as an abbreviation for the latin 'rata die'
440  * (day number).  This is the number of days elapsed since 0000-12-31
441  * in the proleptic Gregorian calendar. The begin of the Christian Era
442  * (0001-01-01) is RD(1).
443  */
444 
445 /*
446  * ====================================================================
447  *
448  * General algorithmic stuff
449  *
450  * ====================================================================
451  */
452 
453 /*
454  *---------------------------------------------------------------------
455  * Do a periodic extension of 'value' around 'pivot' with a period of
456  * 'cycle'.
457  *
458  * The result 'res' is a number that holds to the following properties:
459  *
460  *   1)	 res MOD cycle == value MOD cycle
461  *   2)	 pivot <= res < pivot + cycle
462  *	 (replace </<= with >/>= for negative cycles)
463  *
464  * where 'MOD' denotes the modulo operator for FLOOR DIVISION, which
465  * is not the same as the '%' operator in C: C requires division to be
466  * a truncated division, where remainder and dividend have the same
467  * sign if the remainder is not zero, whereas floor division requires
468  * divider and modulus to have the same sign for a non-zero modulus.
469  *
470  * This function has some useful applications:
471  *
472  * + let Y be a calendar year and V a truncated 2-digit year: then
473  *	periodic_extend(Y-50, V, 100)
474  *   is the closest expansion of the truncated year with respect to
475  *   the full year, that is a 4-digit year with a difference of less
476  *   than 50 years to the year Y. ("century unfolding")
477  *
478  * + let T be a UN*X time stamp and V be seconds-of-day: then
479  *	perodic_extend(T-43200, V, 86400)
480  *   is a time stamp that has the same seconds-of-day as the input
481  *   value, with an absolute difference to T of <= 12hrs.  ("day
482  *   unfolding")
483  *
484  * + Wherever you have a truncated periodic value and a non-truncated
485  *   base value and you want to match them somehow...
486  *
487  * Basically, the function delivers 'pivot + (value - pivot) % cycle',
488  * but the implementation takes some pains to avoid internal signed
489  * integer overflows in the '(value - pivot) % cycle' part and adheres
490  * to the floor division convention.
491  *
492  * If 64bit scalars where available on all intended platforms, writing a
493  * version that uses 64 bit ops would be easy; writing a general
494  * division routine for 64bit ops on a platform that can only do
495  * 32/16bit divisions and is still performant is a bit more
496  * difficult. Since most usecases can be coded in a way that does only
497  * require the 32-bit version a 64bit version is NOT provided here.
498  *---------------------------------------------------------------------
499  */
500 int32_t
501 ntpcal_periodic_extend(
502 	int32_t pivot,
503 	int32_t value,
504 	int32_t cycle
505 	)
506 {
507 	uint32_t diff;
508 	char	 cpl = 0; /* modulo complement flag */
509 	char	 neg = 0; /* sign change flag	    */
510 
511 	/* make the cycle positive and adjust the flags */
512 	if (cycle < 0) {
513 		cycle = - cycle;
514 		neg ^= 1;
515 		cpl ^= 1;
516 	}
517 	/* guard against div by zero or one */
518 	if (cycle > 1) {
519 		/*
520 		 * Get absolute difference as unsigned quantity and
521 		 * the complement flag. This is done by always
522 		 * subtracting the smaller value from the bigger
523 		 * one.
524 		 */
525 		if (value >= pivot) {
526 			diff = int32_to_uint32_2cpl(value)
527 			     - int32_to_uint32_2cpl(pivot);
528 		} else {
529 			diff = int32_to_uint32_2cpl(pivot)
530 			     - int32_to_uint32_2cpl(value);
531 			cpl ^= 1;
532 		}
533 		diff %= (uint32_t)cycle;
534 		if (diff) {
535 			if (cpl)
536 				diff = (uint32_t)cycle - diff;
537 			if (neg)
538 				diff = ~diff + 1;
539 			pivot += uint32_2cpl_to_int32(diff);
540 		}
541 	}
542 	return pivot;
543 }
544 
545 /*---------------------------------------------------------------------
546  * Note to the casual reader
547  *
548  * In the next two functions you will find (or would have found...)
549  * the expression
550  *
551  *   res.Q_s -= 0x80000000;
552  *
553  * There was some ruckus about a possible programming error due to
554  * integer overflow and sign propagation.
555  *
556  * This assumption is based on a lack of understanding of the C
557  * standard. (Though this is admittedly not one of the most 'natural'
558  * aspects of the 'C' language and easily to get wrong.)
559  *
560  * see
561  *	http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1570.pdf
562  *	"ISO/IEC 9899:201x Committee Draft — April 12, 2011"
563  *	6.4.4.1 Integer constants, clause 5
564  *
565  * why there is no sign extension/overflow problem here.
566  *
567  * But to ease the minds of the doubtful, I added back the 'u' qualifiers
568  * that somehow got lost over the last years.
569  */
570 
571 
572 /*
573  *---------------------------------------------------------------------
574  * Convert a timestamp in NTP scale to a 64bit seconds value in the UN*X
575  * scale with proper epoch unfolding around a given pivot or the current
576  * system time. This function happily accepts negative pivot values as
577  * timestamps befor 1970-01-01, so be aware of possible trouble on
578  * platforms with 32bit 'time_t'!
579  *
580  * This is also a periodic extension, but since the cycle is 2^32 and
581  * the shift is 2^31, we can do some *very* fast math without explicit
582  * divisions.
583  *---------------------------------------------------------------------
584  */
585 vint64
586 ntpcal_ntp_to_time(
587 	uint32_t	ntp,
588 	const time_t *	pivot
589 	)
590 {
591 	vint64 res;
592 
593 #   if defined(HAVE_INT64)
594 
595 	res.q_s = (pivot != NULL)
596 		      ? *pivot
597 		      : now();
598 	res.Q_s -= 0x80000000u;		/* unshift of half range */
599 	ntp	-= (uint32_t)JAN_1970;	/* warp into UN*X domain */
600 	ntp	-= res.D_s.lo;		/* cycle difference	 */
601 	res.Q_s += (uint64_t)ntp;	/* get expanded time	 */
602 
603 #   else /* no 64bit scalars */
604 
605 	time_t tmp;
606 
607 	tmp = (pivot != NULL)
608 		  ? *pivot
609 		  : now();
610 	res = time_to_vint64(&tmp);
611 	M_SUB(res.D_s.hi, res.D_s.lo, 0, 0x80000000u);
612 	ntp -= (uint32_t)JAN_1970;	/* warp into UN*X domain */
613 	ntp -= res.D_s.lo;		/* cycle difference	 */
614 	M_ADD(res.D_s.hi, res.D_s.lo, 0, ntp);
615 
616 #   endif /* no 64bit scalars */
617 
618 	return res;
619 }
620 
621 /*
622  *---------------------------------------------------------------------
623  * Convert a timestamp in NTP scale to a 64bit seconds value in the NTP
624  * scale with proper epoch unfolding around a given pivot or the current
625  * system time.
626  *
627  * Note: The pivot must be given in the UN*X time domain!
628  *
629  * This is also a periodic extension, but since the cycle is 2^32 and
630  * the shift is 2^31, we can do some *very* fast math without explicit
631  * divisions.
632  *---------------------------------------------------------------------
633  */
634 vint64
635 ntpcal_ntp_to_ntp(
636 	uint32_t      ntp,
637 	const time_t *pivot
638 	)
639 {
640 	vint64 res;
641 
642 #   if defined(HAVE_INT64)
643 
644 	res.q_s = (pivot)
645 		      ? *pivot
646 		      : now();
647 	res.Q_s -= 0x80000000u;		/* unshift of half range */
648 	res.Q_s += (uint32_t)JAN_1970;	/* warp into NTP domain	 */
649 	ntp	-= res.D_s.lo;		/* cycle difference	 */
650 	res.Q_s += (uint64_t)ntp;	/* get expanded time	 */
651 
652 #   else /* no 64bit scalars */
653 
654 	time_t tmp;
655 
656 	tmp = (pivot)
657 		  ? *pivot
658 		  : now();
659 	res = time_to_vint64(&tmp);
660 	M_SUB(res.D_s.hi, res.D_s.lo, 0, 0x80000000u);
661 	M_ADD(res.D_s.hi, res.D_s.lo, 0, (uint32_t)JAN_1970);/*into NTP */
662 	ntp -= res.D_s.lo;		/* cycle difference	 */
663 	M_ADD(res.D_s.hi, res.D_s.lo, 0, ntp);
664 
665 #   endif /* no 64bit scalars */
666 
667 	return res;
668 }
669 
670 
671 /*
672  * ====================================================================
673  *
674  * Splitting values to composite entities
675  *
676  * ====================================================================
677  */
678 
679 /*
680  *---------------------------------------------------------------------
681  * Split a 64bit seconds value into elapsed days in 'res.hi' and
682  * elapsed seconds since midnight in 'res.lo' using explicit floor
683  * division. This function happily accepts negative time values as
684  * timestamps before the respective epoch start.
685  *---------------------------------------------------------------------
686  */
687 ntpcal_split
688 ntpcal_daysplit(
689 	const vint64 *ts
690 	)
691 {
692 	ntpcal_split res;
693 	uint32_t Q;
694 
695 #   if defined(HAVE_INT64)
696 
697 	/* Manual floor division by SECSPERDAY. This uses the one's
698 	 * complement trick, too, but without an extra flag value: The
699 	 * flag would be 64bit, and that's a bit of overkill on a 32bit
700 	 * target that has to use a register pair for a 64bit number.
701 	 */
702 	if (ts->q_s < 0)
703 		Q = ~(uint32_t)(~ts->Q_s / SECSPERDAY);
704 	else
705 		Q = (uint32_t)(ts->Q_s / SECSPERDAY);
706 
707 #   else
708 
709 	uint32_t ah, al, sflag, A;
710 
711 	/* get operand into ah/al (either ts or ts' one's complement,
712 	 * for later floor division)
713 	 */
714 	sflag = int32_sflag(ts->d_s.hi);
715 	ah = sflag ^ ts->D_s.hi;
716 	al = sflag ^ ts->D_s.lo;
717 
718 	/* Since 86400 == 128*675 we can drop the least 7 bits and
719 	 * divide by 675 instead of 86400. Then the maximum remainder
720 	 * after each devision step is 674, and we need 10 bits for
721 	 * that. So in the next step we can shift in 22 bits from the
722 	 * numerator.
723 	 *
724 	 * Therefore we load the accu with the top 13 bits (51..63) in
725 	 * the first shot. We don't have to remember the quotient -- it
726 	 * would be shifted out anyway.
727 	 */
728 	A = ah >> 19;
729 	if (A >= 675)
730 		A = (A % 675u);
731 
732 	/* Now assemble the remainder with bits 29..50 from the
733 	 * numerator and divide. This creates the upper ten bits of the
734 	 * quotient. (Well, the top 22 bits of a 44bit result. But that
735 	 * will be truncated to 32 bits anyway.)
736 	 */
737 	A = (A << 19) | (ah & 0x0007FFFFu);
738 	A = (A <<  3) | (al >> 29);
739 	Q = A / 675u;
740 	A = A % 675u;
741 
742 	/* Now assemble the remainder with bits 7..28 from the numerator
743 	 * and do a final division step.
744 	 */
745 	A = (A << 22) | ((al >> 7) & 0x003FFFFFu);
746 	Q = (Q << 22) | (A / 675u);
747 
748 	/* The last 7 bits get simply dropped, as they have no affect on
749 	 * the quotient when dividing by 86400.
750 	 */
751 
752 	/* apply sign correction and calculate the true floor
753 	 * remainder.
754 	 */
755 	Q ^= sflag;
756 
757 #   endif
758 
759 	res.hi = uint32_2cpl_to_int32(Q);
760 	res.lo = ts->D_s.lo - Q * SECSPERDAY;
761 
762 	return res;
763 }
764 
765 /*
766  *---------------------------------------------------------------------
767  * Split a 32bit seconds value into h/m/s and excessive days.  This
768  * function happily accepts negative time values as timestamps before
769  * midnight.
770  *---------------------------------------------------------------------
771  */
772 static int32_t
773 priv_timesplit(
774 	int32_t split[3],
775 	int32_t ts
776 	)
777 {
778 	/* Do 3 chained floor divisions by positive constants, using the
779 	 * one's complement trick and factoring out the intermediate XOR
780 	 * ops to reduce the number of operations.
781 	 */
782 	uint32_t us, um, uh, ud, sflag;
783 
784 	sflag = int32_sflag(ts);
785 	us    = int32_to_uint32_2cpl(ts);
786 
787 	um = (sflag ^ us) / SECSPERMIN;
788 	uh = um / MINSPERHR;
789 	ud = uh / HRSPERDAY;
790 
791 	um ^= sflag;
792 	uh ^= sflag;
793 	ud ^= sflag;
794 
795 	split[0] = (int32_t)(uh - ud * HRSPERDAY );
796 	split[1] = (int32_t)(um - uh * MINSPERHR );
797 	split[2] = (int32_t)(us - um * SECSPERMIN);
798 
799 	return uint32_2cpl_to_int32(ud);
800 }
801 
802 /*
803  *---------------------------------------------------------------------
804  * Given the number of elapsed days in the calendar era, split this
805  * number into the number of elapsed years in 'res.hi' and the number
806  * of elapsed days of that year in 'res.lo'.
807  *
808  * if 'isleapyear' is not NULL, it will receive an integer that is 0 for
809  * regular years and a non-zero value for leap years.
810  *---------------------------------------------------------------------
811  */
812 ntpcal_split
813 ntpcal_split_eradays(
814 	int32_t days,
815 	int  *isleapyear
816 	)
817 {
818 	/* Use the fast cyclesplit algorithm here, to calculate the
819 	 * centuries and years in a century with one division each. This
820 	 * reduces the number of division operations to two, but is
821 	 * susceptible to internal range overflow. We make sure the
822 	 * input operands are in the safe range; this still gives us
823 	 * approx +/-2.9 million years.
824 	 */
825 	ntpcal_split res;
826 	int32_t	 n100, n001; /* calendar year cycles */
827 	uint32_t uday, Q, sflag;
828 
829 	/* split off centuries first */
830 	sflag = int32_sflag(days);
831 	uday  = uint32_saturate(int32_to_uint32_2cpl(days), sflag);
832 	uday  = (4u * uday) | 3u;
833 	Q    = sflag ^ ((sflag ^ uday) / GREGORIAN_CYCLE_DAYS);
834 	uday = uday - Q * GREGORIAN_CYCLE_DAYS;
835 	n100 = uint32_2cpl_to_int32(Q);
836 
837 	/* Split off years in century -- days >= 0 here, and we're far
838 	 * away from integer overflow trouble now. */
839 	uday |= 3;
840 	n001 = uday / GREGORIAN_NORMAL_LEAP_CYCLE_DAYS;
841 	uday = uday % GREGORIAN_NORMAL_LEAP_CYCLE_DAYS;
842 
843 	/* Assemble the year and day in year */
844 	res.hi = n100 * 100 + n001;
845 	res.lo = uday / 4u;
846 
847 	/* Eventually set the leap year flag. Note: 0 <= n001 <= 99 and
848 	 * Q is still the two's complement representation of the
849 	 * centuries: The modulo 4 ops can be done with masking here.
850 	 * We also shift the year and the century by one, so the tests
851 	 * can be done against zero instead of 3.
852 	 */
853 	if (isleapyear)
854 		*isleapyear = !((n001+1) & 3)
855 		    && ((n001 != 99) || !((Q+1) & 3));
856 
857 	return res;
858 }
859 
860 /*
861  *---------------------------------------------------------------------
862  * Given a number of elapsed days in a year and a leap year indicator,
863  * split the number of elapsed days into the number of elapsed months in
864  * 'res.hi' and the number of elapsed days of that month in 'res.lo'.
865  *
866  * This function will fail and return {-1,-1} if the number of elapsed
867  * days is not in the valid range!
868  *---------------------------------------------------------------------
869  */
870 ntpcal_split
871 ntpcal_split_yeardays(
872 	int32_t eyd,
873 	int     isleapyear
874 	)
875 {
876 	ntpcal_split    res;
877 	const uint16_t *lt;	/* month length table	*/
878 
879 	/* check leap year flag and select proper table */
880 	lt = real_month_table[(isleapyear != 0)];
881 	if (0 <= eyd && eyd < lt[12]) {
882 		/* get zero-based month by approximation & correction step */
883 		res.hi = eyd >> 5;	   /* approx month; might be 1 too low */
884 		if (lt[res.hi + 1] <= eyd) /* fixup approximative month value  */
885 			res.hi += 1;
886 		res.lo = eyd - lt[res.hi];
887 	} else {
888 		res.lo = res.hi = -1;
889 	}
890 
891 	return res;
892 }
893 
894 /*
895  *---------------------------------------------------------------------
896  * Convert a RD into the date part of a 'struct calendar'.
897  *---------------------------------------------------------------------
898  */
899 int
900 ntpcal_rd_to_date(
901 	struct calendar *jd,
902 	int32_t		 rd
903 	)
904 {
905 	ntpcal_split split;
906 	int	     leapy;
907 	u_int	     ymask;
908 
909 	/* Get day-of-week first. Since rd is signed, the remainder can
910 	 * be in the range [-6..+6], but the assignment to an unsigned
911 	 * variable maps the negative values to positive values >=7.
912 	 * This makes the sign correction look strange, but adding 7
913 	 * causes the needed wrap-around into the desired value range of
914 	 * zero to six, both inclusive.
915 	 */
916 	jd->weekday = rd % DAYSPERWEEK;
917 	if (jd->weekday >= DAYSPERWEEK)	/* weekday is unsigned! */
918 		jd->weekday += DAYSPERWEEK;
919 
920 	split = ntpcal_split_eradays(rd - 1, &leapy);
921 	/* Get year and day-of-year, with overflow check. If any of the
922 	 * upper 16 bits is set after shifting to unity-based years, we
923 	 * will have an overflow when converting to an unsigned 16bit
924 	 * year. Shifting to the right is OK here, since it does not
925 	 * matter if the shift is logic or arithmetic.
926 	 */
927 	split.hi += 1;
928 	ymask = 0u - ((split.hi >> 16) == 0);
929 	jd->year = (uint16_t)(split.hi & ymask);
930 	jd->yearday = (uint16_t)split.lo + 1;
931 
932 	/* convert to month and mday */
933 	split = ntpcal_split_yeardays(split.lo, leapy);
934 	jd->month    = (uint8_t)split.hi + 1;
935 	jd->monthday = (uint8_t)split.lo + 1;
936 
937 	return ymask ? leapy : -1;
938 }
939 
940 /*
941  *---------------------------------------------------------------------
942  * Convert a RD into the date part of a 'struct tm'.
943  *---------------------------------------------------------------------
944  */
945 int
946 ntpcal_rd_to_tm(
947 	struct tm  *utm,
948 	int32_t	    rd
949 	)
950 {
951 	ntpcal_split split;
952 	int	     leapy;
953 
954 	/* get day-of-week first */
955 	utm->tm_wday = rd % DAYSPERWEEK;
956 	if (utm->tm_wday < 0)
957 		utm->tm_wday += DAYSPERWEEK;
958 
959 	/* get year and day-of-year */
960 	split = ntpcal_split_eradays(rd - 1, &leapy);
961 	utm->tm_year = split.hi - 1899;
962 	utm->tm_yday = split.lo;	/* 0-based */
963 
964 	/* convert to month and mday */
965 	split = ntpcal_split_yeardays(split.lo, leapy);
966 	utm->tm_mon  = split.hi;	/* 0-based */
967 	utm->tm_mday = split.lo + 1;	/* 1-based */
968 
969 	return leapy;
970 }
971 
972 /*
973  *---------------------------------------------------------------------
974  * Take a value of seconds since midnight and split it into hhmmss in a
975  * 'struct calendar'.
976  *---------------------------------------------------------------------
977  */
978 int32_t
979 ntpcal_daysec_to_date(
980 	struct calendar *jd,
981 	int32_t		sec
982 	)
983 {
984 	int32_t days;
985 	int   ts[3];
986 
987 	days = priv_timesplit(ts, sec);
988 	jd->hour   = (uint8_t)ts[0];
989 	jd->minute = (uint8_t)ts[1];
990 	jd->second = (uint8_t)ts[2];
991 
992 	return days;
993 }
994 
995 /*
996  *---------------------------------------------------------------------
997  * Take a value of seconds since midnight and split it into hhmmss in a
998  * 'struct tm'.
999  *---------------------------------------------------------------------
1000  */
1001 int32_t
1002 ntpcal_daysec_to_tm(
1003 	struct tm *utm,
1004 	int32_t	   sec
1005 	)
1006 {
1007 	int32_t days;
1008 	int32_t ts[3];
1009 
1010 	days = priv_timesplit(ts, sec);
1011 	utm->tm_hour = ts[0];
1012 	utm->tm_min  = ts[1];
1013 	utm->tm_sec  = ts[2];
1014 
1015 	return days;
1016 }
1017 
1018 /*
1019  *---------------------------------------------------------------------
1020  * take a split representation for day/second-of-day and day offset
1021  * and convert it to a 'struct calendar'. The seconds will be normalised
1022  * into the range of a day, and the day will be adjusted accordingly.
1023  *
1024  * returns >0 if the result is in a leap year, 0 if in a regular
1025  * year and <0 if the result did not fit into the calendar struct.
1026  *---------------------------------------------------------------------
1027  */
1028 int
1029 ntpcal_daysplit_to_date(
1030 	struct calendar	   *jd,
1031 	const ntpcal_split *ds,
1032 	int32_t		    dof
1033 	)
1034 {
1035 	dof += ntpcal_daysec_to_date(jd, ds->lo);
1036 	return ntpcal_rd_to_date(jd, ds->hi + dof);
1037 }
1038 
1039 /*
1040  *---------------------------------------------------------------------
1041  * take a split representation for day/second-of-day and day offset
1042  * and convert it to a 'struct tm'. The seconds will be normalised
1043  * into the range of a day, and the day will be adjusted accordingly.
1044  *
1045  * returns 1 if the result is in a leap year and zero if in a regular
1046  * year.
1047  *---------------------------------------------------------------------
1048  */
1049 int
1050 ntpcal_daysplit_to_tm(
1051 	struct tm	   *utm,
1052 	const ntpcal_split *ds ,
1053 	int32_t		    dof
1054 	)
1055 {
1056 	dof += ntpcal_daysec_to_tm(utm, ds->lo);
1057 
1058 	return ntpcal_rd_to_tm(utm, ds->hi + dof);
1059 }
1060 
1061 /*
1062  *---------------------------------------------------------------------
1063  * Take a UN*X time and convert to a calendar structure.
1064  *---------------------------------------------------------------------
1065  */
1066 int
1067 ntpcal_time_to_date(
1068 	struct calendar	*jd,
1069 	const vint64	*ts
1070 	)
1071 {
1072 	ntpcal_split ds;
1073 
1074 	ds = ntpcal_daysplit(ts);
1075 	ds.hi += ntpcal_daysec_to_date(jd, ds.lo);
1076 	ds.hi += DAY_UNIX_STARTS;
1077 
1078 	return ntpcal_rd_to_date(jd, ds.hi);
1079 }
1080 
1081 
1082 /*
1083  * ====================================================================
1084  *
1085  * merging composite entities
1086  *
1087  * ====================================================================
1088  */
1089 
1090 /*
1091  *---------------------------------------------------------------------
1092  * Merge a number of days and a number of seconds into seconds,
1093  * expressed in 64 bits to avoid overflow.
1094  *---------------------------------------------------------------------
1095  */
1096 vint64
1097 ntpcal_dayjoin(
1098 	int32_t days,
1099 	int32_t secs
1100 	)
1101 {
1102 	vint64 res;
1103 
1104 #   if defined(HAVE_INT64)
1105 
1106 	res.q_s	 = days;
1107 	res.q_s *= SECSPERDAY;
1108 	res.q_s += secs;
1109 
1110 #   else
1111 
1112 	uint32_t p1, p2;
1113 	int	 isneg;
1114 
1115 	/*
1116 	 * res = days *86400 + secs, using manual 16/32 bit
1117 	 * multiplications and shifts.
1118 	 */
1119 	isneg = (days < 0);
1120 	if (isneg)
1121 		days = -days;
1122 
1123 	/* assemble days * 675 */
1124 	res.D_s.lo = (days & 0xFFFF) * 675u;
1125 	res.D_s.hi = 0;
1126 	p1 = (days >> 16) * 675u;
1127 	p2 = p1 >> 16;
1128 	p1 = p1 << 16;
1129 	M_ADD(res.D_s.hi, res.D_s.lo, p2, p1);
1130 
1131 	/* mul by 128, using shift */
1132 	res.D_s.hi = (res.D_s.hi << 7) | (res.D_s.lo >> 25);
1133 	res.D_s.lo = (res.D_s.lo << 7);
1134 
1135 	/* fix sign */
1136 	if (isneg)
1137 		M_NEG(res.D_s.hi, res.D_s.lo);
1138 
1139 	/* properly add seconds */
1140 	p2 = 0;
1141 	if (secs < 0) {
1142 		p1 = (uint32_t)-secs;
1143 		M_NEG(p2, p1);
1144 	} else {
1145 		p1 = (uint32_t)secs;
1146 	}
1147 	M_ADD(res.D_s.hi, res.D_s.lo, p2, p1);
1148 
1149 #   endif
1150 
1151 	return res;
1152 }
1153 
1154 /*
1155  *---------------------------------------------------------------------
1156  * get leap years since epoch in elapsed years
1157  *---------------------------------------------------------------------
1158  */
1159 int32_t
1160 ntpcal_leapyears_in_years(
1161 	int32_t years
1162 	)
1163 {
1164 	/* We use the in-out-in algorithm here, using the one's
1165 	 * complement division trick for negative numbers. The chained
1166 	 * division sequence by 4/25/4 gives the compiler the chance to
1167 	 * get away with only one true division and doing shifts otherwise.
1168 	 */
1169 
1170 	uint32_t sflag, sum, uyear;
1171 
1172 	sflag = int32_sflag(years);
1173 	uyear = int32_to_uint32_2cpl(years);
1174 	uyear ^= sflag;
1175 
1176 	sum  = (uyear /=  4u);	/*   4yr rule --> IN  */
1177 	sum -= (uyear /= 25u);	/* 100yr rule --> OUT */
1178 	sum += (uyear /=  4u);	/* 400yr rule --> IN  */
1179 
1180 	/* Thanks to the alternation of IN/OUT/IN we can do the sum
1181 	 * directly and have a single one's complement operation
1182 	 * here. (Only if the years are negative, of course.) Otherwise
1183 	 * the one's complement would have to be done when
1184 	 * adding/subtracting the terms.
1185 	 */
1186 	return uint32_2cpl_to_int32(sflag ^ sum);
1187 }
1188 
1189 /*
1190  *---------------------------------------------------------------------
1191  * Convert elapsed years in Era into elapsed days in Era.
1192  *---------------------------------------------------------------------
1193  */
1194 int32_t
1195 ntpcal_days_in_years(
1196 	int32_t years
1197 	)
1198 {
1199 	return years * DAYSPERYEAR + ntpcal_leapyears_in_years(years);
1200 }
1201 
1202 /*
1203  *---------------------------------------------------------------------
1204  * Convert a number of elapsed month in a year into elapsed days in year.
1205  *
1206  * The month will be normalized, and 'res.hi' will contain the
1207  * excessive years that must be considered when converting the years,
1208  * while 'res.lo' will contain the number of elapsed days since start
1209  * of the year.
1210  *
1211  * This code uses the shifted-month-approach to convert month to days,
1212  * because then there is no need to have explicit leap year
1213  * information.	 The slight disadvantage is that for most month values
1214  * the result is a negative value, and the year excess is one; the
1215  * conversion is then simply based on the start of the following year.
1216  *---------------------------------------------------------------------
1217  */
1218 ntpcal_split
1219 ntpcal_days_in_months(
1220 	int32_t m
1221 	)
1222 {
1223 	ntpcal_split res;
1224 
1225 	/* Add ten months and correct if needed. (It likely is...) */
1226 	res.lo  = m + 10;
1227 	res.hi  = (res.lo >= 12);
1228 	if (res.hi)
1229 		res.lo -= 12;
1230 
1231 	/* if still out of range, normalise by floor division ... */
1232 	if (res.lo < 0 || res.lo >= 12) {
1233 		uint32_t mu, Q, sflag;
1234 		sflag = int32_sflag(res.lo);
1235 		mu    = int32_to_uint32_2cpl(res.lo);
1236 		Q     = sflag ^ ((sflag ^ mu) / 12u);
1237 		res.hi += uint32_2cpl_to_int32(Q);
1238 		res.lo  = mu - Q * 12u;
1239 	}
1240 
1241 	/* get cummulated days in year with unshift */
1242 	res.lo = shift_month_table[res.lo] - 306;
1243 
1244 	return res;
1245 }
1246 
1247 /*
1248  *---------------------------------------------------------------------
1249  * Convert ELAPSED years/months/days of gregorian calendar to elapsed
1250  * days in Gregorian epoch.
1251  *
1252  * If you want to convert years and days-of-year, just give a month of
1253  * zero.
1254  *---------------------------------------------------------------------
1255  */
1256 int32_t
1257 ntpcal_edate_to_eradays(
1258 	int32_t years,
1259 	int32_t mons,
1260 	int32_t mdays
1261 	)
1262 {
1263 	ntpcal_split tmp;
1264 	int32_t	     res;
1265 
1266 	if (mons) {
1267 		tmp = ntpcal_days_in_months(mons);
1268 		res = ntpcal_days_in_years(years + tmp.hi) + tmp.lo;
1269 	} else
1270 		res = ntpcal_days_in_years(years);
1271 	res += mdays;
1272 
1273 	return res;
1274 }
1275 
1276 /*
1277  *---------------------------------------------------------------------
1278  * Convert ELAPSED years/months/days of gregorian calendar to elapsed
1279  * days in year.
1280  *
1281  * Note: This will give the true difference to the start of the given
1282  * year, even if months & days are off-scale.
1283  *---------------------------------------------------------------------
1284  */
1285 int32_t
1286 ntpcal_edate_to_yeardays(
1287 	int32_t years,
1288 	int32_t mons,
1289 	int32_t mdays
1290 	)
1291 {
1292 	ntpcal_split tmp;
1293 
1294 	if (0 <= mons && mons < 12) {
1295 		years += 1;
1296 		mdays += real_month_table[is_leapyear(years)][mons];
1297 	} else {
1298 		tmp = ntpcal_days_in_months(mons);
1299 		mdays += tmp.lo
1300 		       + ntpcal_days_in_years(years + tmp.hi)
1301 		       - ntpcal_days_in_years(years);
1302 	}
1303 
1304 	return mdays;
1305 }
1306 
1307 /*
1308  *---------------------------------------------------------------------
1309  * Convert elapsed days and the hour/minute/second information into
1310  * total seconds.
1311  *
1312  * If 'isvalid' is not NULL, do a range check on the time specification
1313  * and tell if the time input is in the normal range, permitting for a
1314  * single leapsecond.
1315  *---------------------------------------------------------------------
1316  */
1317 int32_t
1318 ntpcal_etime_to_seconds(
1319 	int32_t hours,
1320 	int32_t minutes,
1321 	int32_t seconds
1322 	)
1323 {
1324 	int32_t res;
1325 
1326 	res = (hours * MINSPERHR + minutes) * SECSPERMIN + seconds;
1327 
1328 	return res;
1329 }
1330 
1331 /*
1332  *---------------------------------------------------------------------
1333  * Convert the date part of a 'struct tm' (that is, year, month,
1334  * day-of-month) into the RD of that day.
1335  *---------------------------------------------------------------------
1336  */
1337 int32_t
1338 ntpcal_tm_to_rd(
1339 	const struct tm *utm
1340 	)
1341 {
1342 	return ntpcal_edate_to_eradays(utm->tm_year + 1899,
1343 				       utm->tm_mon,
1344 				       utm->tm_mday - 1) + 1;
1345 }
1346 
1347 /*
1348  *---------------------------------------------------------------------
1349  * Convert the date part of a 'struct calendar' (that is, year, month,
1350  * day-of-month) into the RD of that day.
1351  *---------------------------------------------------------------------
1352  */
1353 int32_t
1354 ntpcal_date_to_rd(
1355 	const struct calendar *jd
1356 	)
1357 {
1358 	return ntpcal_edate_to_eradays((int32_t)jd->year - 1,
1359 				       (int32_t)jd->month - 1,
1360 				       (int32_t)jd->monthday - 1) + 1;
1361 }
1362 
1363 /*
1364  *---------------------------------------------------------------------
1365  * convert a year number to rata die of year start
1366  *---------------------------------------------------------------------
1367  */
1368 int32_t
1369 ntpcal_year_to_ystart(
1370 	int32_t year
1371 	)
1372 {
1373 	return ntpcal_days_in_years(year - 1) + 1;
1374 }
1375 
1376 /*
1377  *---------------------------------------------------------------------
1378  * For a given RD, get the RD of the associated year start,
1379  * that is, the RD of the last January,1st on or before that day.
1380  *---------------------------------------------------------------------
1381  */
1382 int32_t
1383 ntpcal_rd_to_ystart(
1384 	int32_t rd
1385 	)
1386 {
1387 	/*
1388 	 * Rather simple exercise: split the day number into elapsed
1389 	 * years and elapsed days, then remove the elapsed days from the
1390 	 * input value. Nice'n sweet...
1391 	 */
1392 	return rd - ntpcal_split_eradays(rd - 1, NULL).lo;
1393 }
1394 
1395 /*
1396  *---------------------------------------------------------------------
1397  * For a given RD, get the RD of the associated month start.
1398  *---------------------------------------------------------------------
1399  */
1400 int32_t
1401 ntpcal_rd_to_mstart(
1402 	int32_t rd
1403 	)
1404 {
1405 	ntpcal_split split;
1406 	int	     leaps;
1407 
1408 	split = ntpcal_split_eradays(rd - 1, &leaps);
1409 	split = ntpcal_split_yeardays(split.lo, leaps);
1410 
1411 	return rd - split.lo;
1412 }
1413 
1414 /*
1415  *---------------------------------------------------------------------
1416  * take a 'struct calendar' and get the seconds-of-day from it.
1417  *---------------------------------------------------------------------
1418  */
1419 int32_t
1420 ntpcal_date_to_daysec(
1421 	const struct calendar *jd
1422 	)
1423 {
1424 	return ntpcal_etime_to_seconds(jd->hour, jd->minute,
1425 				       jd->second);
1426 }
1427 
1428 /*
1429  *---------------------------------------------------------------------
1430  * take a 'struct tm' and get the seconds-of-day from it.
1431  *---------------------------------------------------------------------
1432  */
1433 int32_t
1434 ntpcal_tm_to_daysec(
1435 	const struct tm *utm
1436 	)
1437 {
1438 	return ntpcal_etime_to_seconds(utm->tm_hour, utm->tm_min,
1439 				       utm->tm_sec);
1440 }
1441 
1442 /*
1443  *---------------------------------------------------------------------
1444  * take a 'struct calendar' and convert it to a 'time_t'
1445  *---------------------------------------------------------------------
1446  */
1447 time_t
1448 ntpcal_date_to_time(
1449 	const struct calendar *jd
1450 	)
1451 {
1452 	vint64  join;
1453 	int32_t days, secs;
1454 
1455 	days = ntpcal_date_to_rd(jd) - DAY_UNIX_STARTS;
1456 	secs = ntpcal_date_to_daysec(jd);
1457 	join = ntpcal_dayjoin(days, secs);
1458 
1459 	return vint64_to_time(&join);
1460 }
1461 
1462 
1463 /*
1464  * ====================================================================
1465  *
1466  * extended and unchecked variants of caljulian/caltontp
1467  *
1468  * ====================================================================
1469  */
1470 int
1471 ntpcal_ntp64_to_date(
1472 	struct calendar *jd,
1473 	const vint64    *ntp
1474 	)
1475 {
1476 	ntpcal_split ds;
1477 
1478 	ds = ntpcal_daysplit(ntp);
1479 	ds.hi += ntpcal_daysec_to_date(jd, ds.lo);
1480 
1481 	return ntpcal_rd_to_date(jd, ds.hi + DAY_NTP_STARTS);
1482 }
1483 
1484 int
1485 ntpcal_ntp_to_date(
1486 	struct calendar *jd,
1487 	uint32_t	 ntp,
1488 	const time_t	*piv
1489 	)
1490 {
1491 	vint64	ntp64;
1492 
1493 	/*
1494 	 * Unfold ntp time around current time into NTP domain. Split
1495 	 * into days and seconds, shift days into CE domain and
1496 	 * process the parts.
1497 	 */
1498 	ntp64 = ntpcal_ntp_to_ntp(ntp, piv);
1499 	return ntpcal_ntp64_to_date(jd, &ntp64);
1500 }
1501 
1502 
1503 vint64
1504 ntpcal_date_to_ntp64(
1505 	const struct calendar *jd
1506 	)
1507 {
1508 	/*
1509 	 * Convert date to NTP. Ignore yearday, use d/m/y only.
1510 	 */
1511 	return ntpcal_dayjoin(ntpcal_date_to_rd(jd) - DAY_NTP_STARTS,
1512 			      ntpcal_date_to_daysec(jd));
1513 }
1514 
1515 
1516 uint32_t
1517 ntpcal_date_to_ntp(
1518 	const struct calendar *jd
1519 	)
1520 {
1521 	/*
1522 	 * Get lower half of 64-bit NTP timestamp from date/time.
1523 	 */
1524 	return ntpcal_date_to_ntp64(jd).d_s.lo;
1525 }
1526 
1527 
1528 
1529 /*
1530  * ====================================================================
1531  *
1532  * day-of-week calculations
1533  *
1534  * ====================================================================
1535  */
1536 /*
1537  * Given a RataDie and a day-of-week, calculate a RDN that is reater-than,
1538  * greater-or equal, closest, less-or-equal or less-than the given RDN
1539  * and denotes the given day-of-week
1540  */
1541 int32_t
1542 ntpcal_weekday_gt(
1543 	int32_t rdn,
1544 	int32_t dow
1545 	)
1546 {
1547 	return ntpcal_periodic_extend(rdn+1, dow, 7);
1548 }
1549 
1550 int32_t
1551 ntpcal_weekday_ge(
1552 	int32_t rdn,
1553 	int32_t dow
1554 	)
1555 {
1556 	return ntpcal_periodic_extend(rdn, dow, 7);
1557 }
1558 
1559 int32_t
1560 ntpcal_weekday_close(
1561 	int32_t rdn,
1562 	int32_t dow
1563 	)
1564 {
1565 	return ntpcal_periodic_extend(rdn-3, dow, 7);
1566 }
1567 
1568 int32_t
1569 ntpcal_weekday_le(
1570 	int32_t rdn,
1571 	int32_t dow
1572 	)
1573 {
1574 	return ntpcal_periodic_extend(rdn, dow, -7);
1575 }
1576 
1577 int32_t
1578 ntpcal_weekday_lt(
1579 	int32_t rdn,
1580 	int32_t dow
1581 	)
1582 {
1583 	return ntpcal_periodic_extend(rdn-1, dow, -7);
1584 }
1585 
1586 /*
1587  * ====================================================================
1588  *
1589  * ISO week-calendar conversions
1590  *
1591  * The ISO8601 calendar defines a calendar of years, weeks and weekdays.
1592  * It is related to the Gregorian calendar, and a ISO year starts at the
1593  * Monday closest to Jan,1st of the corresponding Gregorian year.  A ISO
1594  * calendar year has always 52 or 53 weeks, and like the Grogrian
1595  * calendar the ISO8601 calendar repeats itself every 400 years, or
1596  * 146097 days, or 20871 weeks.
1597  *
1598  * While it is possible to write ISO calendar functions based on the
1599  * Gregorian calendar functions, the following implementation takes a
1600  * different approach, based directly on years and weeks.
1601  *
1602  * Analysis of the tabulated data shows that it is not possible to
1603  * interpolate from years to weeks over a full 400 year range; cyclic
1604  * shifts over 400 years do not provide a solution here. But it *is*
1605  * possible to interpolate over every single century of the 400-year
1606  * cycle. (The centennial leap year rule seems to be the culprit here.)
1607  *
1608  * It can be shown that a conversion from years to weeks can be done
1609  * using a linear transformation of the form
1610  *
1611  *   w = floor( y * a + b )
1612  *
1613  * where the slope a must hold to
1614  *
1615  *  52.1780821918 <= a < 52.1791044776
1616  *
1617  * and b must be chosen according to the selected slope and the number
1618  * of the century in a 400-year period.
1619  *
1620  * The inverse calculation can also be done in this way. Careful scaling
1621  * provides an unlimited set of integer coefficients a,k,b that enable
1622  * us to write the calulation in the form
1623  *
1624  *   w = (y * a	 + b ) / k
1625  *   y = (w * a' + b') / k'
1626  *
1627  * In this implementation the values of k and k' are chosen to be
1628  * smallest possible powers of two, so the division can be implemented
1629  * as shifts if the optimiser chooses to do so.
1630  *
1631  * ====================================================================
1632  */
1633 
1634 /*
1635  * Given a number of elapsed (ISO-)years since the begin of the
1636  * christian era, return the number of elapsed weeks corresponding to
1637  * the number of years.
1638  */
1639 int32_t
1640 isocal_weeks_in_years(
1641 	int32_t years
1642 	)
1643 {
1644 	/*
1645 	 * use: w = (y * 53431 + b[c]) / 1024 as interpolation
1646 	 */
1647 	static const uint16_t bctab[4] = { 157, 449, 597, 889 };
1648 
1649 	int32_t  cs, cw;
1650 	uint32_t cc, ci, yu, sflag;
1651 
1652 	sflag = int32_sflag(years);
1653 	yu    = int32_to_uint32_2cpl(years);
1654 
1655 	/* split off centuries, using floor division */
1656 	cc  = sflag ^ ((sflag ^ yu) / 100u);
1657 	yu -= cc * 100u;
1658 
1659 	/* calculate century cycles shift and cycle index:
1660 	 * Assuming a century is 5217 weeks, we have to add a cycle
1661 	 * shift that is 3 for every 4 centuries, because 3 of the four
1662 	 * centuries have 5218 weeks. So '(cc*3 + 1) / 4' is the actual
1663 	 * correction, and the second century is the defective one.
1664 	 *
1665 	 * Needs floor division by 4, which is done with masking and
1666 	 * shifting.
1667 	 */
1668 	ci = cc * 3u + 1;
1669 	cs = uint32_2cpl_to_int32(sflag ^ ((sflag ^ ci) / 4u));
1670 	ci = ci % 4u;
1671 
1672 	/* Get weeks in century. Can use plain division here as all ops
1673 	 * are >= 0,  and let the compiler sort out the possible
1674 	 * optimisations.
1675 	 */
1676 	cw = (yu * 53431u + bctab[ci]) / 1024u;
1677 
1678 	return uint32_2cpl_to_int32(cc) * 5217 + cs + cw;
1679 }
1680 
1681 /*
1682  * Given a number of elapsed weeks since the begin of the christian
1683  * era, split this number into the number of elapsed years in res.hi
1684  * and the excessive number of weeks in res.lo. (That is, res.lo is
1685  * the number of elapsed weeks in the remaining partial year.)
1686  */
1687 ntpcal_split
1688 isocal_split_eraweeks(
1689 	int32_t weeks
1690 	)
1691 {
1692 	/*
1693 	 * use: y = (w * 157 + b[c]) / 8192 as interpolation
1694 	 */
1695 
1696 	static const uint16_t bctab[4] = { 85, 130, 17, 62 };
1697 
1698 	ntpcal_split res;
1699 	int32_t  cc, ci;
1700 	uint32_t sw, cy, Q, sflag;
1701 
1702 	/* Use two fast cycle-split divisions here. This is again
1703 	 * susceptible to internal overflow, so we check the range. This
1704 	 * still permits more than +/-20 million years, so this is
1705 	 * likely a pure academical problem.
1706 	 *
1707 	 * We want to execute '(weeks * 4 + 2) /% 20871' under floor
1708 	 * division rules in the first step.
1709 	 */
1710 	sflag = int32_sflag(weeks);
1711 	sw  = uint32_saturate(int32_to_uint32_2cpl(weeks), sflag);
1712 	sw  = 4u * sw + 2;
1713 	Q   = sflag ^ ((sflag ^ sw) / GREGORIAN_CYCLE_WEEKS);
1714 	sw -= Q * GREGORIAN_CYCLE_WEEKS;
1715 	ci  = Q % 4u;
1716 	cc  = uint32_2cpl_to_int32(Q);
1717 
1718 	/* Split off years; sw >= 0 here! The scaled weeks in the years
1719 	 * are scaled up by 157 afterwards.
1720 	 */
1721 	sw  = (sw / 4u) * 157u + bctab[ci];
1722 	cy  = sw / 8192u;	/* ws >> 13 , let the compiler sort it out */
1723 	sw  = sw % 8192u;	/* ws & 8191, let the compiler sort it out */
1724 
1725 	/* assemble elapsed years and downscale the elapsed weeks in
1726 	 * the year.
1727 	 */
1728 	res.hi = 100*cc + cy;
1729 	res.lo = sw / 157u;
1730 
1731 	return res;
1732 }
1733 
1734 /*
1735  * Given a second in the NTP time scale and a pivot, expand the NTP
1736  * time stamp around the pivot and convert into an ISO calendar time
1737  * stamp.
1738  */
1739 int
1740 isocal_ntp64_to_date(
1741 	struct isodate *id,
1742 	const vint64   *ntp
1743 	)
1744 {
1745 	ntpcal_split ds;
1746 	int32_t      ts[3];
1747 	uint32_t     uw, ud, sflag;
1748 
1749 	/*
1750 	 * Split NTP time into days and seconds, shift days into CE
1751 	 * domain and process the parts.
1752 	 */
1753 	ds = ntpcal_daysplit(ntp);
1754 
1755 	/* split time part */
1756 	ds.hi += priv_timesplit(ts, ds.lo);
1757 	id->hour   = (uint8_t)ts[0];
1758 	id->minute = (uint8_t)ts[1];
1759 	id->second = (uint8_t)ts[2];
1760 
1761 	/* split days into days and weeks, using floor division in unsigned */
1762 	ds.hi += DAY_NTP_STARTS - 1; /* shift from NTP to RDN */
1763 	sflag = int32_sflag(ds.hi);
1764 	ud  = int32_to_uint32_2cpl(ds.hi);
1765 	uw  = sflag ^ ((sflag ^ ud) / DAYSPERWEEK);
1766 	ud -= uw * DAYSPERWEEK;
1767 	ds.hi = uint32_2cpl_to_int32(uw);
1768 	ds.lo = ud;
1769 
1770 	id->weekday = (uint8_t)ds.lo + 1;	/* weekday result    */
1771 
1772 	/* get year and week in year */
1773 	ds = isocal_split_eraweeks(ds.hi);	/* elapsed years&week*/
1774 	id->year = (uint16_t)ds.hi + 1;		/* shift to current  */
1775 	id->week = (uint8_t )ds.lo + 1;
1776 
1777 	return (ds.hi >= 0 && ds.hi < 0x0000FFFF);
1778 }
1779 
1780 int
1781 isocal_ntp_to_date(
1782 	struct isodate *id,
1783 	uint32_t	ntp,
1784 	const time_t   *piv
1785 	)
1786 {
1787 	vint64	ntp64;
1788 
1789 	/*
1790 	 * Unfold ntp time around current time into NTP domain, then
1791 	 * convert the full time stamp.
1792 	 */
1793 	ntp64 = ntpcal_ntp_to_ntp(ntp, piv);
1794 	return isocal_ntp64_to_date(id, &ntp64);
1795 }
1796 
1797 /*
1798  * Convert a ISO date spec into a second in the NTP time scale,
1799  * properly truncated to 32 bit.
1800  */
1801 vint64
1802 isocal_date_to_ntp64(
1803 	const struct isodate *id
1804 	)
1805 {
1806 	int32_t weeks, days, secs;
1807 
1808 	weeks = isocal_weeks_in_years((int32_t)id->year - 1)
1809 	      + (int32_t)id->week - 1;
1810 	days = weeks * 7 + (int32_t)id->weekday;
1811 	/* days is RDN of ISO date now */
1812 	secs = ntpcal_etime_to_seconds(id->hour, id->minute, id->second);
1813 
1814 	return ntpcal_dayjoin(days - DAY_NTP_STARTS, secs);
1815 }
1816 
1817 uint32_t
1818 isocal_date_to_ntp(
1819 	const struct isodate *id
1820 	)
1821 {
1822 	/*
1823 	 * Get lower half of 64-bit NTP timestamp from date/time.
1824 	 */
1825 	return isocal_date_to_ntp64(id).d_s.lo;
1826 }
1827 
1828 /* -*-EOF-*- */
1829