Home | History | Annotate | Line # | Download | only in libntp
ntp_calendar.c revision 1.1.1.8
      1 /*	$NetBSD: ntp_calendar.c,v 1.1.1.8 2016/01/08 21:21:24 christos Exp $	*/
      2 
      3 /*
      4  * ntp_calendar.c - calendar and helper functions
      5  *
      6  * Written by Juergen Perlinger (perlinger (at) ntp.org) for the NTP project.
      7  * The contents of 'html/copyright.html' apply.
      8  *
      9  * --------------------------------------------------------------------
     10  * Some notes on the implementation:
     11  *
     12  * Calendar algorithms thrive on the division operation, which is one of
     13  * the slowest numerical operations in any CPU. What saves us here from
     14  * abysmal performance is the fact that all divisions are divisions by
     15  * constant numbers, and most compilers can do this by a multiplication
     16  * operation.  But this might not work when using the div/ldiv/lldiv
     17  * function family, because many compilers are not able to do inline
     18  * expansion of the code with following optimisation for the
     19  * constant-divider case.
     20  *
     21  * Also div/ldiv/lldiv are defined in terms of int/long/longlong, which
     22  * are inherently target dependent. Nothing that could not be cured with
     23  * autoconf, but still a mess...
     24  *
     25  * Furthermore, we need floor division in many places. C either leaves
     26  * the division behaviour undefined (< C99) or demands truncation to
     27  * zero (>= C99), so additional steps are required to make sure the
     28  * algorithms work. The {l,ll}div function family is requested to
     29  * truncate towards zero, which is also the wrong direction for our
     30  * purpose.
     31  *
     32  * For all this, all divisions by constant are coded manually, even when
     33  * there is a joined div/mod operation: The optimiser should sort that
     34  * out, if possible. Most of the calculations are done with unsigned
     35  * types, explicitely using two's complement arithmetics where
     36  * necessary. This minimises the dependecies to compiler and target,
     37  * while still giving reasonable to good performance.
     38  *
     39  * The implementation uses a few tricks that exploit properties of the
     40  * two's complement: Floor division on negative dividents can be
     41  * executed by using the one's complement of the divident. One's
     42  * complement can be easily created using XOR and a mask.
     43  *
     44  * Finally, check for overflow conditions is minimal. There are only two
     45  * calculation steps in the whole calendar that suffer from an internal
     46  * overflow, and these conditions are checked: errno is set to EDOM and
     47  * the results are clamped/saturated in this case.  All other functions
     48  * do not suffer from internal overflow and simply return the result
     49  * truncated to 32 bits.
     50  *
     51  * This is a sacrifice made for execution speed.  Since a 32-bit day
     52  * counter covers +/- 5,879,610 years and the clamp limits the effective
     53  * range to +/-2.9 million years, this should not pose a problem here.
     54  *
     55  */
     56 
     57 #include <config.h>
     58 #include <sys/types.h>
     59 
     60 #include "ntp_types.h"
     61 #include "ntp_calendar.h"
     62 #include "ntp_stdlib.h"
     63 #include "ntp_fp.h"
     64 #include "ntp_unixtime.h"
     65 
     66 /* For now, let's take the conservative approach: if the target property
     67  * macros are not defined, check a few well-known compiler/architecture
     68  * settings. Default is to assume that the representation of signed
     69  * integers is unknown and shift-arithmetic-right is not available.
     70  */
     71 #ifndef TARGET_HAS_2CPL
     72 # if defined(__GNUC__)
     73 #  if defined(__i386__) || defined(__x86_64__) || defined(__arm__)
     74 #   define TARGET_HAS_2CPL 1
     75 #  else
     76 #   define TARGET_HAS_2CPL 0
     77 #  endif
     78 # elif defined(_MSC_VER)
     79 #  if defined(_M_IX86) || defined(_M_X64) || defined(_M_ARM)
     80 #   define TARGET_HAS_2CPL 1
     81 #  else
     82 #   define TARGET_HAS_2CPL 0
     83 #  endif
     84 # else
     85 #  define TARGET_HAS_2CPL 0
     86 # endif
     87 #endif
     88 
     89 #ifndef TARGET_HAS_SAR
     90 # define TARGET_HAS_SAR 0
     91 #endif
     92 
     93 /*
     94  *---------------------------------------------------------------------
     95  * replacing the 'time()' function
     96  * --------------------------------------------------------------------
     97  */
     98 
     99 static systime_func_ptr systime_func = &time;
    100 static inline time_t now(void);
    101 
    102 
    103 systime_func_ptr
    104 ntpcal_set_timefunc(
    105 	systime_func_ptr nfunc
    106 	)
    107 {
    108 	systime_func_ptr res;
    109 
    110 	res = systime_func;
    111 	if (NULL == nfunc)
    112 		nfunc = &time;
    113 	systime_func = nfunc;
    114 
    115 	return res;
    116 }
    117 
    118 
    119 static inline time_t
    120 now(void)
    121 {
    122 	return (*systime_func)(NULL);
    123 }
    124 
    125 /*
    126  *---------------------------------------------------------------------
    127  * Get sign extension mask and unsigned 2cpl rep for a signed integer
    128  *---------------------------------------------------------------------
    129  */
    130 
    131 static inline uint32_t
    132 int32_sflag(
    133 	const int32_t v)
    134 {
    135 #   if TARGET_HAS_2CPL && TARGET_HAS_SAR && SIZEOF_INT >= 4
    136 
    137 	/* Let's assume that shift is the fastest way to get the sign
    138 	 * extension of of a signed integer. This might not always be
    139 	 * true, though -- On 8bit CPUs or machines without barrel
    140 	 * shifter this will kill the performance. So we make sure
    141 	 * we do this only if 'int' has at least 4 bytes.
    142 	 */
    143 	return (uint32_t)(v >> 31);
    144 
    145 #   else
    146 
    147 	/* This should be a rather generic approach for getting a sign
    148 	 * extension mask...
    149 	 */
    150 	return UINT32_C(0) - (uint32_t)(v < 0);
    151 
    152 #   endif
    153 }
    154 
    155 static inline uint32_t
    156 int32_to_uint32_2cpl(
    157 	const int32_t v)
    158 {
    159 	uint32_t vu;
    160 
    161 #   if TARGET_HAS_2CPL
    162 
    163 	/* Just copy through the 32 bits from the signed value if we're
    164 	 * on a two's complement target.
    165 	 */
    166 	vu = (uint32_t)v;
    167 
    168 #   else
    169 
    170 	/* Convert from signed int to unsigned int two's complement. Do
    171 	 * not make any assumptions about the representation of signed
    172 	 * integers, but make sure signed integer overflow cannot happen
    173 	 * here. A compiler on a two's complement target *might* find
    174 	 * out that this is just a complicated cast (as above), but your
    175 	 * mileage might vary.
    176 	 */
    177 	if (v < 0)
    178 		vu = ~(uint32_t)(-(v + 1));
    179 	else
    180 		vu = (uint32_t)v;
    181 
    182 #   endif
    183 
    184 	return vu;
    185 }
    186 
    187 static inline int32_t
    188 uint32_2cpl_to_int32(
    189 	const uint32_t vu)
    190 {
    191 	int32_t v;
    192 
    193 #   if TARGET_HAS_2CPL
    194 
    195 	/* Just copy through the 32 bits from the unsigned value if
    196 	 * we're on a two's complement target.
    197 	 */
    198 	v = (int32_t)vu;
    199 
    200 #   else
    201 
    202 	/* Convert to signed integer, making sure signed integer
    203 	 * overflow cannot happen. Again, the optimiser might or might
    204 	 * not find out that this is just a copy of 32 bits on a target
    205 	 * with two's complement representation for signed integers.
    206 	 */
    207 	if (vu > INT32_MAX)
    208 		v = -(int32_t)(~vu) - 1;
    209 	else
    210 		v = (int32_t)vu;
    211 
    212 #   endif
    213 
    214 	return v;
    215 }
    216 
    217 /* Some of the calculations need to multiply the input by 4 before doing
    218  * a division. This can cause overflow and strange results. Therefore we
    219  * clamp / saturate the input operand. And since we do the calculations
    220  * in unsigned int with an extra sign flag/mask, we only loose one bit
    221  * of the input value range.
    222  */
    223 static inline uint32_t
    224 uint32_saturate(
    225 	uint32_t vu,
    226 	uint32_t mu)
    227 {
    228 	static const uint32_t limit = UINT32_MAX/4u;
    229 	if ((mu ^ vu) > limit) {
    230 		vu    = mu ^ limit;
    231 		errno = EDOM;
    232 	}
    233 	return vu;
    234 }
    235 
    236 /*
    237  *---------------------------------------------------------------------
    238  * Convert between 'time_t' and 'vint64'
    239  *---------------------------------------------------------------------
    240  */
    241 vint64
    242 time_to_vint64(
    243 	const time_t * ptt
    244 	)
    245 {
    246 	vint64 res;
    247 	time_t tt;
    248 
    249 	tt = *ptt;
    250 
    251 #   if SIZEOF_TIME_T <= 4
    252 
    253 	res.D_s.hi = 0;
    254 	if (tt < 0) {
    255 		res.D_s.lo = (uint32_t)-tt;
    256 		M_NEG(res.D_s.hi, res.D_s.lo);
    257 	} else {
    258 		res.D_s.lo = (uint32_t)tt;
    259 	}
    260 
    261 #   elif defined(HAVE_INT64)
    262 
    263 	res.q_s = tt;
    264 
    265 #   else
    266 	/*
    267 	 * shifting negative signed quantities is compiler-dependent, so
    268 	 * we better avoid it and do it all manually. And shifting more
    269 	 * than the width of a quantity is undefined. Also a don't do!
    270 	 */
    271 	if (tt < 0) {
    272 		tt = -tt;
    273 		res.D_s.lo = (uint32_t)tt;
    274 		res.D_s.hi = (uint32_t)(tt >> 32);
    275 		M_NEG(res.D_s.hi, res.D_s.lo);
    276 	} else {
    277 		res.D_s.lo = (uint32_t)tt;
    278 		res.D_s.hi = (uint32_t)(tt >> 32);
    279 	}
    280 
    281 #   endif
    282 
    283 	return res;
    284 }
    285 
    286 
    287 time_t
    288 vint64_to_time(
    289 	const vint64 *tv
    290 	)
    291 {
    292 	time_t res;
    293 
    294 #   if SIZEOF_TIME_T <= 4
    295 
    296 	res = (time_t)tv->D_s.lo;
    297 
    298 #   elif defined(HAVE_INT64)
    299 
    300 	res = (time_t)tv->q_s;
    301 
    302 #   else
    303 
    304 	res = ((time_t)tv->d_s.hi << 32) | tv->D_s.lo;
    305 
    306 #   endif
    307 
    308 	return res;
    309 }
    310 
    311 /*
    312  *---------------------------------------------------------------------
    313  * Get the build date & time
    314  *---------------------------------------------------------------------
    315  */
    316 int
    317 ntpcal_get_build_date(
    318 	struct calendar * jd
    319 	)
    320 {
    321 	/* The C standard tells us the format of '__DATE__':
    322 	 *
    323 	 * __DATE__ The date of translation of the preprocessing
    324 	 * translation unit: a character string literal of the form "Mmm
    325 	 * dd yyyy", where the names of the months are the same as those
    326 	 * generated by the asctime function, and the first character of
    327 	 * dd is a space character if the value is less than 10. If the
    328 	 * date of translation is not available, an
    329 	 * implementation-defined valid date shall be supplied.
    330 	 *
    331 	 * __TIME__ The time of translation of the preprocessing
    332 	 * translation unit: a character string literal of the form
    333 	 * "hh:mm:ss" as in the time generated by the asctime
    334 	 * function. If the time of translation is not available, an
    335 	 * implementation-defined valid time shall be supplied.
    336 	 *
    337 	 * Note that MSVC declares DATE and TIME to be in the local time
    338 	 * zone, while neither the C standard nor the GCC docs make any
    339 	 * statement about this. As a result, we may be +/-12hrs off
    340 	 * UTC.  But for practical purposes, this should not be a
    341 	 * problem.
    342 	 *
    343 	 */
    344 #   ifdef MKREPRO_DATE
    345 	static const char build[] = MKREPRO_TIME "/" MKREPRO_DATE;
    346 #   else
    347 	static const char build[] = __TIME__ "/" __DATE__;
    348 #   endif
    349 	static const char mlist[] = "JanFebMarAprMayJunJulAugSepOctNovDec";
    350 
    351 	char		  monstr[4];
    352 	const char *	  cp;
    353 	unsigned short	  hour, minute, second, day, year;
    354  	/* Note: The above quantities are used for sscanf 'hu' format,
    355 	 * so using 'uint16_t' is contra-indicated!
    356 	 */
    357 
    358 #   ifdef DEBUG
    359 	static int        ignore  = 0;
    360 #   endif
    361 
    362 	ZERO(*jd);
    363 	jd->year     = 1970;
    364 	jd->month    = 1;
    365 	jd->monthday = 1;
    366 
    367 #   ifdef DEBUG
    368 	/* check environment if build date should be ignored */
    369 	if (0 == ignore) {
    370 	    const char * envstr;
    371 	    envstr = getenv("NTPD_IGNORE_BUILD_DATE");
    372 	    ignore = 1 + (envstr && (!*envstr || !strcasecmp(envstr, "yes")));
    373 	}
    374 	if (ignore > 1)
    375 	    return FALSE;
    376 #   endif
    377 
    378 	if (6 == sscanf(build, "%hu:%hu:%hu/%3s %hu %hu",
    379 			&hour, &minute, &second, monstr, &day, &year)) {
    380 		cp = strstr(mlist, monstr);
    381 		if (NULL != cp) {
    382 			jd->year     = year;
    383 			jd->month    = (uint8_t)((cp - mlist) / 3 + 1);
    384 			jd->monthday = (uint8_t)day;
    385 			jd->hour     = (uint8_t)hour;
    386 			jd->minute   = (uint8_t)minute;
    387 			jd->second   = (uint8_t)second;
    388 
    389 			return TRUE;
    390 		}
    391 	}
    392 
    393 	return FALSE;
    394 }
    395 
    396 
    397 /*
    398  *---------------------------------------------------------------------
    399  * basic calendar stuff
    400  * --------------------------------------------------------------------
    401  */
    402 
    403 /* month table for a year starting with March,1st */
    404 static const uint16_t shift_month_table[13] = {
    405 	0, 31, 61, 92, 122, 153, 184, 214, 245, 275, 306, 337, 366
    406 };
    407 
    408 /* month tables for years starting with January,1st; regular & leap */
    409 static const uint16_t real_month_table[2][13] = {
    410 	/* -*- table for regular years -*- */
    411 	{ 0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334, 365 },
    412 	/* -*- table for leap years -*- */
    413 	{ 0, 31, 60, 91, 121, 152, 182, 213, 244, 274, 305, 335, 366 }
    414 };
    415 
    416 /*
    417  * Some notes on the terminology:
    418  *
    419  * We use the proleptic Gregorian calendar, which is the Gregorian
    420  * calendar extended in both directions ad infinitum. This totally
    421  * disregards the fact that this calendar was invented in 1582, and
    422  * was adopted at various dates over the world; sometimes even after
    423  * the start of the NTP epoch.
    424  *
    425  * Normally date parts are given as current cycles, while time parts
    426  * are given as elapsed cycles:
    427  *
    428  * 1970-01-01/03:04:05 means 'IN the 1970st. year, IN the first month,
    429  * ON the first day, with 3hrs, 4minutes and 5 seconds elapsed.
    430  *
    431  * The basic calculations for this calendar implementation deal with
    432  * ELAPSED date units, which is the number of full years, full months
    433  * and full days before a date: 1970-01-01 would be (1969, 0, 0) in
    434  * that notation.
    435  *
    436  * To ease the numeric computations, month and day values outside the
    437  * normal range are acceptable: 2001-03-00 will be treated as the day
    438  * before 2001-03-01, 2000-13-32 will give the same result as
    439  * 2001-02-01 and so on.
    440  *
    441  * 'rd' or 'RD' is used as an abbreviation for the latin 'rata die'
    442  * (day number).  This is the number of days elapsed since 0000-12-31
    443  * in the proleptic Gregorian calendar. The begin of the Christian Era
    444  * (0001-01-01) is RD(1).
    445  */
    446 
    447 /*
    448  * ==================================================================
    449  *
    450  * General algorithmic stuff
    451  *
    452  * ==================================================================
    453  */
    454 
    455 /*
    456  *---------------------------------------------------------------------
    457  * Do a periodic extension of 'value' around 'pivot' with a period of
    458  * 'cycle'.
    459  *
    460  * The result 'res' is a number that holds to the following properties:
    461  *
    462  *   1)	 res MOD cycle == value MOD cycle
    463  *   2)	 pivot <= res < pivot + cycle
    464  *	 (replace </<= with >/>= for negative cycles)
    465  *
    466  * where 'MOD' denotes the modulo operator for FLOOR DIVISION, which
    467  * is not the same as the '%' operator in C: C requires division to be
    468  * a truncated division, where remainder and dividend have the same
    469  * sign if the remainder is not zero, whereas floor division requires
    470  * divider and modulus to have the same sign for a non-zero modulus.
    471  *
    472  * This function has some useful applications:
    473  *
    474  * + let Y be a calendar year and V a truncated 2-digit year: then
    475  *	periodic_extend(Y-50, V, 100)
    476  *   is the closest expansion of the truncated year with respect to
    477  *   the full year, that is a 4-digit year with a difference of less
    478  *   than 50 years to the year Y. ("century unfolding")
    479  *
    480  * + let T be a UN*X time stamp and V be seconds-of-day: then
    481  *	perodic_extend(T-43200, V, 86400)
    482  *   is a time stamp that has the same seconds-of-day as the input
    483  *   value, with an absolute difference to T of <= 12hrs.  ("day
    484  *   unfolding")
    485  *
    486  * + Wherever you have a truncated periodic value and a non-truncated
    487  *   base value and you want to match them somehow...
    488  *
    489  * Basically, the function delivers 'pivot + (value - pivot) % cycle',
    490  * but the implementation takes some pains to avoid internal signed
    491  * integer overflows in the '(value - pivot) % cycle' part and adheres
    492  * to the floor division convention.
    493  *
    494  * If 64bit scalars where available on all intended platforms, writing a
    495  * version that uses 64 bit ops would be easy; writing a general
    496  * division routine for 64bit ops on a platform that can only do
    497  * 32/16bit divisions and is still performant is a bit more
    498  * difficult. Since most usecases can be coded in a way that does only
    499  * require the 32-bit version a 64bit version is NOT provided here.
    500  * ---------------------------------------------------------------------
    501  */
    502 int32_t
    503 ntpcal_periodic_extend(
    504 	int32_t pivot,
    505 	int32_t value,
    506 	int32_t cycle
    507 	)
    508 {
    509 	uint32_t diff;
    510 	char	 cpl = 0; /* modulo complement flag */
    511 	char	 neg = 0; /* sign change flag	    */
    512 
    513 	/* make the cycle positive and adjust the flags */
    514 	if (cycle < 0) {
    515 		cycle = - cycle;
    516 		neg ^= 1;
    517 		cpl ^= 1;
    518 	}
    519 	/* guard against div by zero or one */
    520 	if (cycle > 1) {
    521 		/*
    522 		 * Get absolute difference as unsigned quantity and
    523 		 * the complement flag. This is done by always
    524 		 * subtracting the smaller value from the bigger
    525 		 * one.
    526 		 */
    527 		if (value >= pivot) {
    528 			diff = int32_to_uint32_2cpl(value)
    529 			     - int32_to_uint32_2cpl(pivot);
    530 		} else {
    531 			diff = int32_to_uint32_2cpl(pivot)
    532 			     - int32_to_uint32_2cpl(value);
    533 			cpl ^= 1;
    534 		}
    535 		diff %= (uint32_t)cycle;
    536 		if (diff) {
    537 			if (cpl)
    538 				diff = (uint32_t)cycle - diff;
    539 			if (neg)
    540 				diff = ~diff + 1;
    541 			pivot += uint32_2cpl_to_int32(diff);
    542 		}
    543 	}
    544 	return pivot;
    545 }
    546 
    547 /*
    548  *-------------------------------------------------------------------
    549  * Convert a timestamp in NTP scale to a 64bit seconds value in the UN*X
    550  * scale with proper epoch unfolding around a given pivot or the current
    551  * system time. This function happily accepts negative pivot values as
    552  * timestamps befor 1970-01-01, so be aware of possible trouble on
    553  * platforms with 32bit 'time_t'!
    554  *
    555  * This is also a periodic extension, but since the cycle is 2^32 and
    556  * the shift is 2^31, we can do some *very* fast math without explicit
    557  * divisions.
    558  *-------------------------------------------------------------------
    559  */
    560 vint64
    561 ntpcal_ntp_to_time(
    562 	uint32_t	ntp,
    563 	const time_t *	pivot
    564 	)
    565 {
    566 	vint64 res;
    567 
    568 #   if defined(HAVE_INT64)
    569 
    570 	res.q_s = (pivot != NULL)
    571 		      ? *pivot
    572 		      : now();
    573 	res.Q_s -= 0x80000000;		/* unshift of half range */
    574 	ntp	-= (uint32_t)JAN_1970;	/* warp into UN*X domain */
    575 	ntp	-= res.D_s.lo;		/* cycle difference	 */
    576 	res.Q_s += (uint64_t)ntp;	/* get expanded time	 */
    577 
    578 #   else /* no 64bit scalars */
    579 
    580 	time_t tmp;
    581 
    582 	tmp = (pivot != NULL)
    583 		  ? *pivot
    584 		  : now();
    585 	res = time_to_vint64(&tmp);
    586 	M_SUB(res.D_s.hi, res.D_s.lo, 0, 0x80000000);
    587 	ntp -= (uint32_t)JAN_1970;	/* warp into UN*X domain */
    588 	ntp -= res.D_s.lo;		/* cycle difference	 */
    589 	M_ADD(res.D_s.hi, res.D_s.lo, 0, ntp);
    590 
    591 #   endif /* no 64bit scalars */
    592 
    593 	return res;
    594 }
    595 
    596 /*
    597  *-------------------------------------------------------------------
    598  * Convert a timestamp in NTP scale to a 64bit seconds value in the NTP
    599  * scale with proper epoch unfolding around a given pivot or the current
    600  * system time.
    601  *
    602  * Note: The pivot must be given in the UN*X time domain!
    603  *
    604  * This is also a periodic extension, but since the cycle is 2^32 and
    605  * the shift is 2^31, we can do some *very* fast math without explicit
    606  * divisions.
    607  *-------------------------------------------------------------------
    608  */
    609 vint64
    610 ntpcal_ntp_to_ntp(
    611 	uint32_t      ntp,
    612 	const time_t *pivot
    613 	)
    614 {
    615 	vint64 res;
    616 
    617 #   if defined(HAVE_INT64)
    618 
    619 	res.q_s = (pivot)
    620 		      ? *pivot
    621 		      : now();
    622 	res.Q_s -= 0x80000000;		/* unshift of half range */
    623 	res.Q_s += (uint32_t)JAN_1970;	/* warp into NTP domain	 */
    624 	ntp	-= res.D_s.lo;		/* cycle difference	 */
    625 	res.Q_s += (uint64_t)ntp;	/* get expanded time	 */
    626 
    627 #   else /* no 64bit scalars */
    628 
    629 	time_t tmp;
    630 
    631 	tmp = (pivot)
    632 		  ? *pivot
    633 		  : now();
    634 	res = time_to_vint64(&tmp);
    635 	M_SUB(res.D_s.hi, res.D_s.lo, 0, 0x80000000u);
    636 	M_ADD(res.D_s.hi, res.D_s.lo, 0, (uint32_t)JAN_1970);/*into NTP */
    637 	ntp -= res.D_s.lo;		/* cycle difference	 */
    638 	M_ADD(res.D_s.hi, res.D_s.lo, 0, ntp);
    639 
    640 #   endif /* no 64bit scalars */
    641 
    642 	return res;
    643 }
    644 
    645 
    646 /*
    647  * ==================================================================
    648  *
    649  * Splitting values to composite entities
    650  *
    651  * ==================================================================
    652  */
    653 
    654 /*
    655  *-------------------------------------------------------------------
    656  * Split a 64bit seconds value into elapsed days in 'res.hi' and
    657  * elapsed seconds since midnight in 'res.lo' using explicit floor
    658  * division. This function happily accepts negative time values as
    659  * timestamps before the respective epoch start.
    660  * -------------------------------------------------------------------
    661  */
    662 ntpcal_split
    663 ntpcal_daysplit(
    664 	const vint64 *ts
    665 	)
    666 {
    667 	ntpcal_split res;
    668 	uint32_t Q;
    669 
    670 #   if defined(HAVE_INT64)
    671 
    672 	/* Manual floor division by SECSPERDAY. This uses the one's
    673 	 * complement trick, too, but without an extra flag value: The
    674 	 * flag would be 64bit, and that's a bit of overkill on a 32bit
    675 	 * target that has to use a register pair for a 64bit number.
    676 	 */
    677 	if (ts->q_s < 0)
    678 		Q = ~(uint32_t)(~ts->Q_s / SECSPERDAY);
    679 	else
    680 		Q = (uint32_t)(ts->Q_s / SECSPERDAY);
    681 
    682 #   else
    683 
    684 	uint32_t ah, al, sflag, A;
    685 
    686 	/* get operand into ah/al (either ts or ts' one's complement,
    687 	 * for later floor division)
    688 	 */
    689 	sflag = int32_sflag(ts->d_s.hi);
    690 	ah = sflag ^ ts->D_s.hi;
    691 	al = sflag ^ ts->D_s.lo;
    692 
    693 	/* Since 86400 == 128*675 we can drop the least 7 bits and
    694 	 * divide by 675 instead of 86400. Then the maximum remainder
    695 	 * after each devision step is 674, and we need 10 bits for
    696 	 * that. So in the next step we can shift in 22 bits from the
    697 	 * numerator.
    698 	 *
    699 	 * Therefore we load the accu with the top 13 bits (51..63) in
    700 	 * the first shot. We don't have to remember the quotient -- it
    701 	 * would be shifted out anyway.
    702 	 */
    703 	A = ah >> 19;
    704 	if (A >= 675)
    705 		A = (A % 675u);
    706 
    707 	/* Now assemble the remainder with bits 29..50 from the
    708 	 * numerator and divide. This creates the upper ten bits of the
    709 	 * quotient. (Well, the top 22 bits of a 44bit result. But that
    710 	 * will be truncated to 32 bits anyway.)
    711 	 */
    712 	A = (A << 19) | (ah & 0x0007FFFFu);
    713 	A = (A <<  3) | (al >> 29);
    714 	Q = A / 675u;
    715 	A = A % 675u;
    716 
    717 	/* Now assemble the remainder with bits 7..28 from the numerator
    718 	 * and do a final division step.
    719 	 */
    720 	A = (A << 22) | ((al >> 7) & 0x003FFFFFu);
    721 	Q = (Q << 22) | (A / 675u);
    722 
    723 	/* The last 7 bits get simply dropped, as they have no affect on
    724 	 * the quotient when dividing by 86400.
    725 	 */
    726 
    727 	/* apply sign correction and calculate the true floor
    728 	 * remainder.
    729 	 */
    730 	Q ^= sflag;
    731 
    732 #   endif
    733 
    734 	res.hi = uint32_2cpl_to_int32(Q);
    735 	res.lo = ts->D_s.lo - Q * SECSPERDAY;
    736 
    737 	return res;
    738 }
    739 
    740 /*
    741  *-------------------------------------------------------------------
    742  * Split a 32bit seconds value into h/m/s and excessive days.  This
    743  * function happily accepts negative time values as timestamps before
    744  * midnight.
    745  * -------------------------------------------------------------------
    746  */
    747 static int32_t
    748 priv_timesplit(
    749 	int32_t split[3],
    750 	int32_t ts
    751 	)
    752 {
    753 	/* Do 3 chained floor divisions by positive constants, using the
    754 	 * one's complement trick and factoring out the intermediate XOR
    755 	 * ops to reduce the number of operations.
    756 	 */
    757 	uint32_t us, um, uh, ud, sflag;
    758 
    759 	sflag = int32_sflag(ts);
    760 	us    = int32_to_uint32_2cpl(ts);
    761 
    762 	um = (sflag ^ us) / SECSPERMIN;
    763 	uh = um / MINSPERHR;
    764 	ud = uh / HRSPERDAY;
    765 
    766 	um ^= sflag;
    767 	uh ^= sflag;
    768 	ud ^= sflag;
    769 
    770 	split[0] = (int32_t)(uh - ud * HRSPERDAY );
    771 	split[1] = (int32_t)(um - uh * MINSPERHR );
    772 	split[2] = (int32_t)(us - um * SECSPERMIN);
    773 
    774 	return uint32_2cpl_to_int32(ud);
    775 }
    776 
    777 /*
    778  * ---------------------------------------------------------------------
    779  * Given the number of elapsed days in the calendar era, split this
    780  * number into the number of elapsed years in 'res.hi' and the number
    781  * of elapsed days of that year in 'res.lo'.
    782  *
    783  * if 'isleapyear' is not NULL, it will receive an integer that is 0 for
    784  * regular years and a non-zero value for leap years.
    785  *---------------------------------------------------------------------
    786  */
    787 ntpcal_split
    788 ntpcal_split_eradays(
    789 	int32_t days,
    790 	int  *isleapyear
    791 	)
    792 {
    793 	/* Use the fast cyclesplit algorithm here, to calculate the
    794 	 * centuries and years in a century with one division each. This
    795 	 * reduces the number of division operations to two, but is
    796 	 * susceptible to internal range overflow. We make sure the
    797 	 * input operands are in the safe range; this still gives us
    798 	 * approx +/-2.9 million years.
    799 	 */
    800 	ntpcal_split res;
    801 	int32_t	 n100, n001; /* calendar year cycles */
    802 	uint32_t uday, Q, sflag;
    803 
    804 	/* split off centuries first */
    805 	sflag = int32_sflag(days);
    806 	uday  = uint32_saturate(int32_to_uint32_2cpl(days), sflag);
    807 	uday  = (4u * uday) | 3u;
    808 	Q    = sflag ^ ((sflag ^ uday) / GREGORIAN_CYCLE_DAYS);
    809 	uday = uday - Q * GREGORIAN_CYCLE_DAYS;
    810 	n100 = uint32_2cpl_to_int32(Q);
    811 
    812 	/* Split off years in century -- days >= 0 here, and we're far
    813 	 * away from integer overflow trouble now. */
    814 	uday |= 3;
    815 	n001 = uday / GREGORIAN_NORMAL_LEAP_CYCLE_DAYS;
    816 	uday = uday % GREGORIAN_NORMAL_LEAP_CYCLE_DAYS;
    817 
    818 	/* Assemble the year and day in year */
    819 	res.hi = n100 * 100 + n001;
    820 	res.lo = uday / 4u;
    821 
    822 	/* Eventually set the leap year flag. Note: 0 <= n001 <= 99 and
    823 	 * Q is still the two's complement representation of the
    824 	 * centuries: The modulo 4 ops can be done with masking here.
    825 	 * We also shift the year and the century by one, so the tests
    826 	 * can be done against zero instead of 3.
    827 	 */
    828 	if (isleapyear)
    829 		*isleapyear = !((n001+1) & 3)
    830 		    && ((n001 != 99) || !((Q+1) & 3));
    831 
    832 	return res;
    833 }
    834 
    835 /*
    836  *---------------------------------------------------------------------
    837  * Given a number of elapsed days in a year and a leap year indicator,
    838  * split the number of elapsed days into the number of elapsed months in
    839  * 'res.hi' and the number of elapsed days of that month in 'res.lo'.
    840  *
    841  * This function will fail and return {-1,-1} if the number of elapsed
    842  * days is not in the valid range!
    843  *---------------------------------------------------------------------
    844  */
    845 ntpcal_split
    846 ntpcal_split_yeardays(
    847 	int32_t eyd,
    848 	int     isleapyear
    849 	)
    850 {
    851 	ntpcal_split    res;
    852 	const uint16_t *lt;	/* month length table	*/
    853 
    854 	/* check leap year flag and select proper table */
    855 	lt = real_month_table[(isleapyear != 0)];
    856 	if (0 <= eyd && eyd < lt[12]) {
    857 		/* get zero-based month by approximation & correction step */
    858 		res.hi = eyd >> 5;	   /* approx month; might be 1 too low */
    859 		if (lt[res.hi + 1] <= eyd) /* fixup approximative month value  */
    860 			res.hi += 1;
    861 		res.lo = eyd - lt[res.hi];
    862 	} else {
    863 		res.lo = res.hi = -1;
    864 	}
    865 
    866 	return res;
    867 }
    868 
    869 /*
    870  *---------------------------------------------------------------------
    871  * Convert a RD into the date part of a 'struct calendar'.
    872  *---------------------------------------------------------------------
    873  */
    874 int
    875 ntpcal_rd_to_date(
    876 	struct calendar *jd,
    877 	int32_t		 rd
    878 	)
    879 {
    880 	ntpcal_split split;
    881 	int	     leapy;
    882 	u_int	     ymask;
    883 
    884 	/* Get day-of-week first. Since rd is signed, the remainder can
    885 	 * be in the range [-6..+6], but the assignment to an unsigned
    886 	 * variable maps the negative values to positive values >=7.
    887 	 * This makes the sign correction look strange, but adding 7
    888 	 * causes the needed wrap-around into the desired value range of
    889 	 * zero to six, both inclusive.
    890 	 */
    891 	jd->weekday = rd % DAYSPERWEEK;
    892 	if (jd->weekday >= DAYSPERWEEK)	/* weekday is unsigned! */
    893 		jd->weekday += DAYSPERWEEK;
    894 
    895 	split = ntpcal_split_eradays(rd - 1, &leapy);
    896 	/* Get year and day-of-year, with overflow check. If any of the
    897 	 * upper 16 bits is set after shifting to unity-based years, we
    898 	 * will have an overflow when converting to an unsigned 16bit
    899 	 * year. Shifting to the right is OK here, since it does not
    900 	 * matter if the shift is logic or arithmetic.
    901 	 */
    902 	split.hi += 1;
    903 	ymask = 0u - ((split.hi >> 16) == 0);
    904 	jd->year = (uint16_t)(split.hi & ymask);
    905 	jd->yearday = (uint16_t)split.lo + 1;
    906 
    907 	/* convert to month and mday */
    908 	split = ntpcal_split_yeardays(split.lo, leapy);
    909 	jd->month    = (uint8_t)split.hi + 1;
    910 	jd->monthday = (uint8_t)split.lo + 1;
    911 
    912 	return ymask ? leapy : -1;
    913 }
    914 
    915 /*
    916  *---------------------------------------------------------------------
    917  * Convert a RD into the date part of a 'struct tm'.
    918  *---------------------------------------------------------------------
    919  */
    920 int
    921 ntpcal_rd_to_tm(
    922 	struct tm  *utm,
    923 	int32_t	    rd
    924 	)
    925 {
    926 	ntpcal_split split;
    927 	int	     leapy;
    928 
    929 	/* get day-of-week first */
    930 	utm->tm_wday = rd % DAYSPERWEEK;
    931 	if (utm->tm_wday < 0)
    932 		utm->tm_wday += DAYSPERWEEK;
    933 
    934 	/* get year and day-of-year */
    935 	split = ntpcal_split_eradays(rd - 1, &leapy);
    936 	utm->tm_year = split.hi - 1899;
    937 	utm->tm_yday = split.lo;	/* 0-based */
    938 
    939 	/* convert to month and mday */
    940 	split = ntpcal_split_yeardays(split.lo, leapy);
    941 	utm->tm_mon  = split.hi;	/* 0-based */
    942 	utm->tm_mday = split.lo + 1;	/* 1-based */
    943 
    944 	return leapy;
    945 }
    946 
    947 /*
    948  *---------------------------------------------------------------------
    949  * Take a value of seconds since midnight and split it into hhmmss in a
    950  * 'struct calendar'.
    951  *---------------------------------------------------------------------
    952  */
    953 int32_t
    954 ntpcal_daysec_to_date(
    955 	struct calendar *jd,
    956 	int32_t		sec
    957 	)
    958 {
    959 	int32_t days;
    960 	int   ts[3];
    961 
    962 	days = priv_timesplit(ts, sec);
    963 	jd->hour   = (uint8_t)ts[0];
    964 	jd->minute = (uint8_t)ts[1];
    965 	jd->second = (uint8_t)ts[2];
    966 
    967 	return days;
    968 }
    969 
    970 /*
    971  *---------------------------------------------------------------------
    972  * Take a value of seconds since midnight and split it into hhmmss in a
    973  * 'struct tm'.
    974  *---------------------------------------------------------------------
    975  */
    976 int32_t
    977 ntpcal_daysec_to_tm(
    978 	struct tm *utm,
    979 	int32_t	   sec
    980 	)
    981 {
    982 	int32_t days;
    983 	int32_t ts[3];
    984 
    985 	days = priv_timesplit(ts, sec);
    986 	utm->tm_hour = ts[0];
    987 	utm->tm_min  = ts[1];
    988 	utm->tm_sec  = ts[2];
    989 
    990 	return days;
    991 }
    992 
    993 /*
    994  *---------------------------------------------------------------------
    995  * take a split representation for day/second-of-day and day offset
    996  * and convert it to a 'struct calendar'. The seconds will be normalised
    997  * into the range of a day, and the day will be adjusted accordingly.
    998  *
    999  * returns >0 if the result is in a leap year, 0 if in a regular
   1000  * year and <0 if the result did not fit into the calendar struct.
   1001  *---------------------------------------------------------------------
   1002  */
   1003 int
   1004 ntpcal_daysplit_to_date(
   1005 	struct calendar	   *jd,
   1006 	const ntpcal_split *ds,
   1007 	int32_t		    dof
   1008 	)
   1009 {
   1010 	dof += ntpcal_daysec_to_date(jd, ds->lo);
   1011 	return ntpcal_rd_to_date(jd, ds->hi + dof);
   1012 }
   1013 
   1014 /*
   1015  *---------------------------------------------------------------------
   1016  * take a split representation for day/second-of-day and day offset
   1017  * and convert it to a 'struct tm'. The seconds will be normalised
   1018  * into the range of a day, and the day will be adjusted accordingly.
   1019  *
   1020  * returns 1 if the result is in a leap year and zero if in a regular
   1021  * year.
   1022  *---------------------------------------------------------------------
   1023  */
   1024 int
   1025 ntpcal_daysplit_to_tm(
   1026 	struct tm	   *utm,
   1027 	const ntpcal_split *ds ,
   1028 	int32_t		    dof
   1029 	)
   1030 {
   1031 	dof += ntpcal_daysec_to_tm(utm, ds->lo);
   1032 
   1033 	return ntpcal_rd_to_tm(utm, ds->hi + dof);
   1034 }
   1035 
   1036 /*
   1037  *---------------------------------------------------------------------
   1038  * Take a UN*X time and convert to a calendar structure.
   1039  *---------------------------------------------------------------------
   1040  */
   1041 int
   1042 ntpcal_time_to_date(
   1043 	struct calendar	*jd,
   1044 	const vint64	*ts
   1045 	)
   1046 {
   1047 	ntpcal_split ds;
   1048 
   1049 	ds = ntpcal_daysplit(ts);
   1050 	ds.hi += ntpcal_daysec_to_date(jd, ds.lo);
   1051 	ds.hi += DAY_UNIX_STARTS;
   1052 
   1053 	return ntpcal_rd_to_date(jd, ds.hi);
   1054 }
   1055 
   1056 
   1057 /*
   1058  * ==================================================================
   1059  *
   1060  * merging composite entities
   1061  *
   1062  * ==================================================================
   1063  */
   1064 
   1065 /*
   1066  *---------------------------------------------------------------------
   1067  * Merge a number of days and a number of seconds into seconds,
   1068  * expressed in 64 bits to avoid overflow.
   1069  *---------------------------------------------------------------------
   1070  */
   1071 vint64
   1072 ntpcal_dayjoin(
   1073 	int32_t days,
   1074 	int32_t secs
   1075 	)
   1076 {
   1077 	vint64 res;
   1078 
   1079 #   if defined(HAVE_INT64)
   1080 
   1081 	res.q_s	 = days;
   1082 	res.q_s *= SECSPERDAY;
   1083 	res.q_s += secs;
   1084 
   1085 #   else
   1086 
   1087 	uint32_t p1, p2;
   1088 	int	 isneg;
   1089 
   1090 	/*
   1091 	 * res = days *86400 + secs, using manual 16/32 bit
   1092 	 * multiplications and shifts.
   1093 	 */
   1094 	isneg = (days < 0);
   1095 	if (isneg)
   1096 		days = -days;
   1097 
   1098 	/* assemble days * 675 */
   1099 	res.D_s.lo = (days & 0xFFFF) * 675u;
   1100 	res.D_s.hi = 0;
   1101 	p1 = (days >> 16) * 675u;
   1102 	p2 = p1 >> 16;
   1103 	p1 = p1 << 16;
   1104 	M_ADD(res.D_s.hi, res.D_s.lo, p2, p1);
   1105 
   1106 	/* mul by 128, using shift */
   1107 	res.D_s.hi = (res.D_s.hi << 7) | (res.D_s.lo >> 25);
   1108 	res.D_s.lo = (res.D_s.lo << 7);
   1109 
   1110 	/* fix sign */
   1111 	if (isneg)
   1112 		M_NEG(res.D_s.hi, res.D_s.lo);
   1113 
   1114 	/* properly add seconds */
   1115 	p2 = 0;
   1116 	if (secs < 0) {
   1117 		p1 = (uint32_t)-secs;
   1118 		M_NEG(p2, p1);
   1119 	} else {
   1120 		p1 = (uint32_t)secs;
   1121 	}
   1122 	M_ADD(res.D_s.hi, res.D_s.lo, p2, p1);
   1123 
   1124 #   endif
   1125 
   1126 	return res;
   1127 }
   1128 
   1129 /*
   1130  *---------------------------------------------------------------------
   1131  * get leap years since epoch in elapsed years
   1132  *---------------------------------------------------------------------
   1133  */
   1134 int32_t
   1135 ntpcal_leapyears_in_years(
   1136 	int32_t years
   1137 	)
   1138 {
   1139 	/* We use the in-out-in algorithm here, using the one's
   1140 	 * complement division trick for negative numbers. The chained
   1141 	 * division sequence by 4/25/4 gives the compiler the chance to
   1142 	 * get away with only one true division and doing shifts otherwise.
   1143 	 */
   1144 
   1145 	uint32_t sflag, sum, uyear;
   1146 
   1147 	sflag = int32_sflag(years);
   1148 	uyear = int32_to_uint32_2cpl(years);
   1149 	uyear ^= sflag;
   1150 
   1151 	sum  = (uyear /=  4u);	/*   4yr rule --> IN  */
   1152 	sum -= (uyear /= 25u);	/* 100yr rule --> OUT */
   1153 	sum += (uyear /=  4u);	/* 400yr rule --> IN  */
   1154 
   1155 	/* Thanks to the alternation of IN/OUT/IN we can do the sum
   1156 	 * directly and have a single one's complement operation
   1157 	 * here. (Only if the years are negative, of course.) Otherwise
   1158 	 * the one's complement would have to be done when
   1159 	 * adding/subtracting the terms.
   1160 	 */
   1161 	return uint32_2cpl_to_int32(sflag ^ sum);
   1162 }
   1163 
   1164 /*
   1165  *---------------------------------------------------------------------
   1166  * Convert elapsed years in Era into elapsed days in Era.
   1167  *---------------------------------------------------------------------
   1168  */
   1169 int32_t
   1170 ntpcal_days_in_years(
   1171 	int32_t years
   1172 	)
   1173 {
   1174 	return years * DAYSPERYEAR + ntpcal_leapyears_in_years(years);
   1175 }
   1176 
   1177 /*
   1178  *---------------------------------------------------------------------
   1179  * Convert a number of elapsed month in a year into elapsed days in year.
   1180  *
   1181  * The month will be normalized, and 'res.hi' will contain the
   1182  * excessive years that must be considered when converting the years,
   1183  * while 'res.lo' will contain the number of elapsed days since start
   1184  * of the year.
   1185  *
   1186  * This code uses the shifted-month-approach to convert month to days,
   1187  * because then there is no need to have explicit leap year
   1188  * information.	 The slight disadvantage is that for most month values
   1189  * the result is a negative value, and the year excess is one; the
   1190  * conversion is then simply based on the start of the following year.
   1191  *---------------------------------------------------------------------
   1192  */
   1193 ntpcal_split
   1194 ntpcal_days_in_months(
   1195 	int32_t m
   1196 	)
   1197 {
   1198 	ntpcal_split res;
   1199 
   1200 	/* Add ten months and correct if needed. (It likely is...) */
   1201 	res.lo  = m + 10;
   1202 	res.hi  = (res.lo >= 12);
   1203 	if (res.hi)
   1204 		res.lo -= 12;
   1205 
   1206 	/* if still out of range, normalise by floor division ... */
   1207 	if (res.lo < 0 || res.lo >= 12) {
   1208 		uint32_t mu, Q, sflag;
   1209 		sflag = int32_sflag(res.lo);
   1210 		mu    = int32_to_uint32_2cpl(res.lo);
   1211 		Q     = sflag ^ ((sflag ^ mu) / 12u);
   1212 		res.hi += uint32_2cpl_to_int32(Q);
   1213 		res.lo  = mu - Q * 12u;
   1214 	}
   1215 
   1216 	/* get cummulated days in year with unshift */
   1217 	res.lo = shift_month_table[res.lo] - 306;
   1218 
   1219 	return res;
   1220 }
   1221 
   1222 /*
   1223  *---------------------------------------------------------------------
   1224  * Convert ELAPSED years/months/days of gregorian calendar to elapsed
   1225  * days in Gregorian epoch.
   1226  *
   1227  * If you want to convert years and days-of-year, just give a month of
   1228  * zero.
   1229  *---------------------------------------------------------------------
   1230  */
   1231 int32_t
   1232 ntpcal_edate_to_eradays(
   1233 	int32_t years,
   1234 	int32_t mons,
   1235 	int32_t mdays
   1236 	)
   1237 {
   1238 	ntpcal_split tmp;
   1239 	int32_t	     res;
   1240 
   1241 	if (mons) {
   1242 		tmp = ntpcal_days_in_months(mons);
   1243 		res = ntpcal_days_in_years(years + tmp.hi) + tmp.lo;
   1244 	} else
   1245 		res = ntpcal_days_in_years(years);
   1246 	res += mdays;
   1247 
   1248 	return res;
   1249 }
   1250 
   1251 /*
   1252  *---------------------------------------------------------------------
   1253  * Convert ELAPSED years/months/days of gregorian calendar to elapsed
   1254  * days in year.
   1255  *
   1256  * Note: This will give the true difference to the start of the given year,
   1257  * even if months & days are off-scale.
   1258  *---------------------------------------------------------------------
   1259  */
   1260 int32_t
   1261 ntpcal_edate_to_yeardays(
   1262 	int32_t years,
   1263 	int32_t mons,
   1264 	int32_t mdays
   1265 	)
   1266 {
   1267 	ntpcal_split tmp;
   1268 
   1269 	if (0 <= mons && mons < 12) {
   1270 		years += 1;
   1271 		mdays += real_month_table[is_leapyear(years)][mons];
   1272 	} else {
   1273 		tmp = ntpcal_days_in_months(mons);
   1274 		mdays += tmp.lo
   1275 		       + ntpcal_days_in_years(years + tmp.hi)
   1276 		       - ntpcal_days_in_years(years);
   1277 	}
   1278 
   1279 	return mdays;
   1280 }
   1281 
   1282 /*
   1283  *---------------------------------------------------------------------
   1284  * Convert elapsed days and the hour/minute/second information into
   1285  * total seconds.
   1286  *
   1287  * If 'isvalid' is not NULL, do a range check on the time specification
   1288  * and tell if the time input is in the normal range, permitting for a
   1289  * single leapsecond.
   1290  *---------------------------------------------------------------------
   1291  */
   1292 int32_t
   1293 ntpcal_etime_to_seconds(
   1294 	int32_t hours,
   1295 	int32_t minutes,
   1296 	int32_t seconds
   1297 	)
   1298 {
   1299 	int32_t res;
   1300 
   1301 	res = (hours * MINSPERHR + minutes) * SECSPERMIN + seconds;
   1302 
   1303 	return res;
   1304 }
   1305 
   1306 /*
   1307  *---------------------------------------------------------------------
   1308  * Convert the date part of a 'struct tm' (that is, year, month,
   1309  * day-of-month) into the RD of that day.
   1310  *---------------------------------------------------------------------
   1311  */
   1312 int32_t
   1313 ntpcal_tm_to_rd(
   1314 	const struct tm *utm
   1315 	)
   1316 {
   1317 	return ntpcal_edate_to_eradays(utm->tm_year + 1899,
   1318 				       utm->tm_mon,
   1319 				       utm->tm_mday - 1) + 1;
   1320 }
   1321 
   1322 /*
   1323  *---------------------------------------------------------------------
   1324  * Convert the date part of a 'struct calendar' (that is, year, month,
   1325  * day-of-month) into the RD of that day.
   1326  *---------------------------------------------------------------------
   1327  */
   1328 int32_t
   1329 ntpcal_date_to_rd(
   1330 	const struct calendar *jd
   1331 	)
   1332 {
   1333 	return ntpcal_edate_to_eradays((int32_t)jd->year - 1,
   1334 				       (int32_t)jd->month - 1,
   1335 				       (int32_t)jd->monthday - 1) + 1;
   1336 }
   1337 
   1338 /*
   1339  *---------------------------------------------------------------------
   1340  * convert a year number to rata die of year start
   1341  *---------------------------------------------------------------------
   1342  */
   1343 int32_t
   1344 ntpcal_year_to_ystart(
   1345 	int32_t year
   1346 	)
   1347 {
   1348 	return ntpcal_days_in_years(year - 1) + 1;
   1349 }
   1350 
   1351 /*
   1352  *---------------------------------------------------------------------
   1353  * For a given RD, get the RD of the associated year start,
   1354  * that is, the RD of the last January,1st on or before that day.
   1355  *---------------------------------------------------------------------
   1356  */
   1357 int32_t
   1358 ntpcal_rd_to_ystart(
   1359 	int32_t rd
   1360 	)
   1361 {
   1362 	/*
   1363 	 * Rather simple exercise: split the day number into elapsed
   1364 	 * years and elapsed days, then remove the elapsed days from the
   1365 	 * input value. Nice'n sweet...
   1366 	 */
   1367 	return rd - ntpcal_split_eradays(rd - 1, NULL).lo;
   1368 }
   1369 
   1370 /*
   1371  *---------------------------------------------------------------------
   1372  * For a given RD, get the RD of the associated month start.
   1373  *---------------------------------------------------------------------
   1374  */
   1375 int32_t
   1376 ntpcal_rd_to_mstart(
   1377 	int32_t rd
   1378 	)
   1379 {
   1380 	ntpcal_split split;
   1381 	int	     leaps;
   1382 
   1383 	split = ntpcal_split_eradays(rd - 1, &leaps);
   1384 	split = ntpcal_split_yeardays(split.lo, leaps);
   1385 
   1386 	return rd - split.lo;
   1387 }
   1388 
   1389 /*
   1390  *---------------------------------------------------------------------
   1391  * take a 'struct calendar' and get the seconds-of-day from it.
   1392  *---------------------------------------------------------------------
   1393  */
   1394 int32_t
   1395 ntpcal_date_to_daysec(
   1396 	const struct calendar *jd
   1397 	)
   1398 {
   1399 	return ntpcal_etime_to_seconds(jd->hour, jd->minute,
   1400 				       jd->second);
   1401 }
   1402 
   1403 /*
   1404  *---------------------------------------------------------------------
   1405  * take a 'struct tm' and get the seconds-of-day from it.
   1406  *---------------------------------------------------------------------
   1407  */
   1408 int32_t
   1409 ntpcal_tm_to_daysec(
   1410 	const struct tm *utm
   1411 	)
   1412 {
   1413 	return ntpcal_etime_to_seconds(utm->tm_hour, utm->tm_min,
   1414 				       utm->tm_sec);
   1415 }
   1416 
   1417 /*
   1418  *---------------------------------------------------------------------
   1419  * take a 'struct calendar' and convert it to a 'time_t'
   1420  *---------------------------------------------------------------------
   1421  */
   1422 time_t
   1423 ntpcal_date_to_time(
   1424 	const struct calendar *jd
   1425 	)
   1426 {
   1427 	vint64  join;
   1428 	int32_t days, secs;
   1429 
   1430 	days = ntpcal_date_to_rd(jd) - DAY_UNIX_STARTS;
   1431 	secs = ntpcal_date_to_daysec(jd);
   1432 	join = ntpcal_dayjoin(days, secs);
   1433 
   1434 	return vint64_to_time(&join);
   1435 }
   1436 
   1437 
   1438 /*
   1439  * ==================================================================
   1440  *
   1441  * extended and unchecked variants of caljulian/caltontp
   1442  *
   1443  * ==================================================================
   1444  */
   1445 int
   1446 ntpcal_ntp64_to_date(
   1447 	struct calendar *jd,
   1448 	const vint64    *ntp
   1449 	)
   1450 {
   1451 	ntpcal_split ds;
   1452 
   1453 	ds = ntpcal_daysplit(ntp);
   1454 	ds.hi += ntpcal_daysec_to_date(jd, ds.lo);
   1455 
   1456 	return ntpcal_rd_to_date(jd, ds.hi + DAY_NTP_STARTS);
   1457 }
   1458 
   1459 int
   1460 ntpcal_ntp_to_date(
   1461 	struct calendar *jd,
   1462 	uint32_t	 ntp,
   1463 	const time_t	*piv
   1464 	)
   1465 {
   1466 	vint64	ntp64;
   1467 
   1468 	/*
   1469 	 * Unfold ntp time around current time into NTP domain. Split
   1470 	 * into days and seconds, shift days into CE domain and
   1471 	 * process the parts.
   1472 	 */
   1473 	ntp64 = ntpcal_ntp_to_ntp(ntp, piv);
   1474 	return ntpcal_ntp64_to_date(jd, &ntp64);
   1475 }
   1476 
   1477 
   1478 vint64
   1479 ntpcal_date_to_ntp64(
   1480 	const struct calendar *jd
   1481 	)
   1482 {
   1483 	/*
   1484 	 * Convert date to NTP. Ignore yearday, use d/m/y only.
   1485 	 */
   1486 	return ntpcal_dayjoin(ntpcal_date_to_rd(jd) - DAY_NTP_STARTS,
   1487 			      ntpcal_date_to_daysec(jd));
   1488 }
   1489 
   1490 
   1491 uint32_t
   1492 ntpcal_date_to_ntp(
   1493 	const struct calendar *jd
   1494 	)
   1495 {
   1496 	/*
   1497 	 * Get lower half of 64-bit NTP timestamp from date/time.
   1498 	 */
   1499 	return ntpcal_date_to_ntp64(jd).d_s.lo;
   1500 }
   1501 
   1502 
   1503 
   1504 /*
   1505  * ==================================================================
   1506  *
   1507  * day-of-week calculations
   1508  *
   1509  * ==================================================================
   1510  */
   1511 /*
   1512  * Given a RataDie and a day-of-week, calculate a RDN that is reater-than,
   1513  * greater-or equal, closest, less-or-equal or less-than the given RDN
   1514  * and denotes the given day-of-week
   1515  */
   1516 int32_t
   1517 ntpcal_weekday_gt(
   1518 	int32_t rdn,
   1519 	int32_t dow
   1520 	)
   1521 {
   1522 	return ntpcal_periodic_extend(rdn+1, dow, 7);
   1523 }
   1524 
   1525 int32_t
   1526 ntpcal_weekday_ge(
   1527 	int32_t rdn,
   1528 	int32_t dow
   1529 	)
   1530 {
   1531 	return ntpcal_periodic_extend(rdn, dow, 7);
   1532 }
   1533 
   1534 int32_t
   1535 ntpcal_weekday_close(
   1536 	int32_t rdn,
   1537 	int32_t dow
   1538 	)
   1539 {
   1540 	return ntpcal_periodic_extend(rdn-3, dow, 7);
   1541 }
   1542 
   1543 int32_t
   1544 ntpcal_weekday_le(
   1545 	int32_t rdn,
   1546 	int32_t dow
   1547 	)
   1548 {
   1549 	return ntpcal_periodic_extend(rdn, dow, -7);
   1550 }
   1551 
   1552 int32_t
   1553 ntpcal_weekday_lt(
   1554 	int32_t rdn,
   1555 	int32_t dow
   1556 	)
   1557 {
   1558 	return ntpcal_periodic_extend(rdn-1, dow, -7);
   1559 }
   1560 
   1561 /*
   1562  * ==================================================================
   1563  *
   1564  * ISO week-calendar conversions
   1565  *
   1566  * The ISO8601 calendar defines a calendar of years, weeks and weekdays.
   1567  * It is related to the Gregorian calendar, and a ISO year starts at the
   1568  * Monday closest to Jan,1st of the corresponding Gregorian year.  A ISO
   1569  * calendar year has always 52 or 53 weeks, and like the Grogrian
   1570  * calendar the ISO8601 calendar repeats itself every 400 years, or
   1571  * 146097 days, or 20871 weeks.
   1572  *
   1573  * While it is possible to write ISO calendar functions based on the
   1574  * Gregorian calendar functions, the following implementation takes a
   1575  * different approach, based directly on years and weeks.
   1576  *
   1577  * Analysis of the tabulated data shows that it is not possible to
   1578  * interpolate from years to weeks over a full 400 year range; cyclic
   1579  * shifts over 400 years do not provide a solution here. But it *is*
   1580  * possible to interpolate over every single century of the 400-year
   1581  * cycle. (The centennial leap year rule seems to be the culprit here.)
   1582  *
   1583  * It can be shown that a conversion from years to weeks can be done
   1584  * using a linear transformation of the form
   1585  *
   1586  *   w = floor( y * a + b )
   1587  *
   1588  * where the slope a must hold to
   1589  *
   1590  *  52.1780821918 <= a < 52.1791044776
   1591  *
   1592  * and b must be chosen according to the selected slope and the number
   1593  * of the century in a 400-year period.
   1594  *
   1595  * The inverse calculation can also be done in this way. Careful scaling
   1596  * provides an unlimited set of integer coefficients a,k,b that enable
   1597  * us to write the calulation in the form
   1598  *
   1599  *   w = (y * a	 + b ) / k
   1600  *   y = (w * a' + b') / k'
   1601  *
   1602  * In this implementation the values of k and k' are chosen to be
   1603  * smallest possible powers of two, so the division can be implemented
   1604  * as shifts if the optimiser chooses to do so.
   1605  *
   1606  * ==================================================================
   1607  */
   1608 
   1609 /*
   1610  * Given a number of elapsed (ISO-)years since the begin of the
   1611  * christian era, return the number of elapsed weeks corresponding to
   1612  * the number of years.
   1613  */
   1614 int32_t
   1615 isocal_weeks_in_years(
   1616 	int32_t years
   1617 	)
   1618 {
   1619 	/*
   1620 	 * use: w = (y * 53431 + b[c]) / 1024 as interpolation
   1621 	 */
   1622 	static const uint16_t bctab[4] = { 157, 449, 597, 889 };
   1623 
   1624 	int32_t  cs, cw;
   1625 	uint32_t cc, ci, yu, sflag;
   1626 
   1627 	sflag = int32_sflag(years);
   1628 	yu    = int32_to_uint32_2cpl(years);
   1629 
   1630 	/* split off centuries, using floor division */
   1631 	cc  = sflag ^ ((sflag ^ yu) / 100u);
   1632 	yu -= cc * 100u;
   1633 
   1634 	/* calculate century cycles shift and cycle index:
   1635 	 * Assuming a century is 5217 weeks, we have to add a cycle
   1636 	 * shift that is 3 for every 4 centuries, because 3 of the four
   1637 	 * centuries have 5218 weeks. So '(cc*3 + 1) / 4' is the actual
   1638 	 * correction, and the second century is the defective one.
   1639 	 *
   1640 	 * Needs floor division by 4, which is done with masking and
   1641 	 * shifting.
   1642 	 */
   1643 	ci = cc * 3u + 1;
   1644 	cs = uint32_2cpl_to_int32(sflag ^ ((sflag ^ ci) / 4u));
   1645 	ci = ci % 4u;
   1646 
   1647 	/* Get weeks in century. Can use plain division here as all ops
   1648 	 * are >= 0,  and let the compiler sort out the possible
   1649 	 * optimisations.
   1650 	 */
   1651 	cw = (yu * 53431u + bctab[ci]) / 1024u;
   1652 
   1653 	return uint32_2cpl_to_int32(cc) * 5217 + cs + cw;
   1654 }
   1655 
   1656 /*
   1657  * Given a number of elapsed weeks since the begin of the christian
   1658  * era, split this number into the number of elapsed years in res.hi
   1659  * and the excessive number of weeks in res.lo. (That is, res.lo is
   1660  * the number of elapsed weeks in the remaining partial year.)
   1661  */
   1662 ntpcal_split
   1663 isocal_split_eraweeks(
   1664 	int32_t weeks
   1665 	)
   1666 {
   1667 	/*
   1668 	 * use: y = (w * 157 + b[c]) / 8192 as interpolation
   1669 	 */
   1670 
   1671 	static const uint16_t bctab[4] = { 85, 130, 17, 62 };
   1672 
   1673 	ntpcal_split res;
   1674 	int32_t  cc, ci;
   1675 	uint32_t sw, cy, Q, sflag;
   1676 
   1677 	/* Use two fast cycle-split divisions here. This is again
   1678 	 * susceptible to internal overflow, so we check the range. This
   1679 	 * still permits more than +/-20 million years, so this is
   1680 	 * likely a pure academical problem.
   1681 	 *
   1682 	 * We want to execute '(weeks * 4 + 2) /% 20871' under floor
   1683 	 * division rules in the first step.
   1684 	 */
   1685 	sflag = int32_sflag(weeks);
   1686 	sw  = uint32_saturate(int32_to_uint32_2cpl(weeks), sflag);
   1687 	sw  = 4u * sw + 2;
   1688 	Q   = sflag ^ ((sflag ^ sw) / GREGORIAN_CYCLE_WEEKS);
   1689 	sw -= Q * GREGORIAN_CYCLE_WEEKS;
   1690 	ci  = Q % 4u;
   1691 	cc  = uint32_2cpl_to_int32(Q);
   1692 
   1693 	/* Split off years; sw >= 0 here! The scaled weeks in the years
   1694 	 * are scaled up by 157 afterwards.
   1695 	 */
   1696 	sw  = (sw / 4u) * 157u + bctab[ci];
   1697 	cy  = sw / 8192u;	/* ws >> 13 , let the compiler sort it out */
   1698 	sw  = sw % 8192u;	/* ws & 8191, let the compiler sort it out */
   1699 
   1700 	/* assemble elapsed years and downscale the elapsed weeks in
   1701 	 * the year.
   1702 	 */
   1703 	res.hi = 100*cc + cy;
   1704 	res.lo = sw / 157u;
   1705 
   1706 	return res;
   1707 }
   1708 
   1709 /*
   1710  * Given a second in the NTP time scale and a pivot, expand the NTP
   1711  * time stamp around the pivot and convert into an ISO calendar time
   1712  * stamp.
   1713  */
   1714 int
   1715 isocal_ntp64_to_date(
   1716 	struct isodate *id,
   1717 	const vint64   *ntp
   1718 	)
   1719 {
   1720 	ntpcal_split ds;
   1721 	int32_t      ts[3];
   1722 	uint32_t     uw, ud, sflag;
   1723 
   1724 	/*
   1725 	 * Split NTP time into days and seconds, shift days into CE
   1726 	 * domain and process the parts.
   1727 	 */
   1728 	ds = ntpcal_daysplit(ntp);
   1729 
   1730 	/* split time part */
   1731 	ds.hi += priv_timesplit(ts, ds.lo);
   1732 	id->hour   = (uint8_t)ts[0];
   1733 	id->minute = (uint8_t)ts[1];
   1734 	id->second = (uint8_t)ts[2];
   1735 
   1736 	/* split days into days and weeks, using floor division in unsigned */
   1737 	ds.hi += DAY_NTP_STARTS - 1; /* shift from NTP to RDN */
   1738 	sflag = int32_sflag(ds.hi);
   1739 	ud  = int32_to_uint32_2cpl(ds.hi);
   1740 	uw  = sflag ^ ((sflag ^ ud) / DAYSPERWEEK);
   1741 	ud -= uw * DAYSPERWEEK;
   1742 	ds.hi = uint32_2cpl_to_int32(uw);
   1743 	ds.lo = ud;
   1744 
   1745 	id->weekday = (uint8_t)ds.lo + 1;	/* weekday result    */
   1746 
   1747 	/* get year and week in year */
   1748 	ds = isocal_split_eraweeks(ds.hi);	/* elapsed years&week*/
   1749 	id->year = (uint16_t)ds.hi + 1;		/* shift to current  */
   1750 	id->week = (uint8_t )ds.lo + 1;
   1751 
   1752 	return (ds.hi >= 0 && ds.hi < 0x0000FFFF);
   1753 }
   1754 
   1755 int
   1756 isocal_ntp_to_date(
   1757 	struct isodate *id,
   1758 	uint32_t	ntp,
   1759 	const time_t   *piv
   1760 	)
   1761 {
   1762 	vint64	ntp64;
   1763 
   1764 	/*
   1765 	 * Unfold ntp time around current time into NTP domain, then
   1766 	 * convert the full time stamp.
   1767 	 */
   1768 	ntp64 = ntpcal_ntp_to_ntp(ntp, piv);
   1769 	return isocal_ntp64_to_date(id, &ntp64);
   1770 }
   1771 
   1772 /*
   1773  * Convert a ISO date spec into a second in the NTP time scale,
   1774  * properly truncated to 32 bit.
   1775  */
   1776 vint64
   1777 isocal_date_to_ntp64(
   1778 	const struct isodate *id
   1779 	)
   1780 {
   1781 	int32_t weeks, days, secs;
   1782 
   1783 	weeks = isocal_weeks_in_years((int32_t)id->year - 1)
   1784 	      + (int32_t)id->week - 1;
   1785 	days = weeks * 7 + (int32_t)id->weekday;
   1786 	/* days is RDN of ISO date now */
   1787 	secs = ntpcal_etime_to_seconds(id->hour, id->minute, id->second);
   1788 
   1789 	return ntpcal_dayjoin(days - DAY_NTP_STARTS, secs);
   1790 }
   1791 
   1792 uint32_t
   1793 isocal_date_to_ntp(
   1794 	const struct isodate *id
   1795 	)
   1796 {
   1797 	/*
   1798 	 * Get lower half of 64-bit NTP timestamp from date/time.
   1799 	 */
   1800 	return isocal_date_to_ntp64(id).d_s.lo;
   1801 }
   1802 
   1803 /* -*-EOF-*- */
   1804