Possible bug using POSIX::strftime Digital UNIX Perl 5.005_03 #514
Comments
From phil@finchcomputer.comI believe I have found a bug in Perl > 5.004_04 (possibly only Perl The problem seems to be that strftime C module calls mktime() which Should I not use strftime? Is strftime supposed to call mktime? It The *real* world where I've run into problems, gmtime sets tm_isdst to % /opt/dtv/bin/perl -e '@tmp=gmtime;use POSIX; print scalar(gmtime), % /opt/dtv/bin/perl -e '@tmp=gmtime;use POSIX; print scalar(gmtime), The equivalent C code that works without this problem. Details, including output from perlbug below, and I believe I've found Phil Details... * The system (Digital UNIX 4, Perl 5.005_03): * The problematic type of Perl program (NOTE: note 1hr difference!): * Sample C code that works... * Output from the system date call and from the C code. The first time % date -u * POSSIBLE FIX: Output from diff directory is perl5.005_03/ext/POSIX... * bambam% /opt/dtv/perl/5.005_03/bin/perlbug -d -v Configured by lobbesp at Mon Jul 19 10:56:06 PDT 1999. Summary of my perl5 (5.0 patchlevel 5 subversion 3) configuration: Locally applied patches: @INC for perl 5.00503: Environment for perl 5.00503: Complete configuration data for perl 5.00503: Author='' |
From [Unknown Contact. See original ticket]On Sun, 19 Sep 1999 15:30:31 -0700, Gurusamy Sarathy wrote (in part): Sarathy> The mktime() call was added (change#1914) to fix the Sarathy> I'd say this needs to be controlled via the hints only I know of no vendor strftime() which doesn't need help to avoid Sometime today I'll submit a patch which substitutes a /spider |
From @timbunceOn Thu, Sep 23, 1999 at 09:55:26AM -0400, spider-perl@Orb.Nashua.NH.US wrote:
Random observation: I've found that mktime() can be _very slow, so Tim. |
From [Unknown Contact. See original ticket]On Thu, 23 Sep 1999 16:14:30 +0100, Tim Bunce wrote (in part): Tim> Random observation: I've found that mktime() can be _very The mini-mktime won't be doing anything with tzset() or its /spider |
From [Unknown Contact. See original ticket]Hi, Just to clarify things a bit, I believe taking out mktime in this case I'd personally have my program SEGV if I pass improper data to Maybe we should add a test to POSIX (not sure how portable this is but use POSIX; Phil
spider> On Sun, 19 Sep 1999 15:30:31 -0700, Gurusamy Sarathy wrote (in Sarathy> I'd say this needs to be controlled via the hints only for spider> I know of no vendor strftime() which doesn't need help to spider> Sometime today I'll submit a patch which substitutes a spider> /spider |
From [Unknown Contact. See original ticket]On Thu, 23 Sep 1999 09:55:26 -0400, I wrote (in part): sb> Sometime today I'll submit a patch which substitutes a Here it is. /spider Inline Patch--- ext/POSIX/POSIX.pod.DIST Tue Jul 20 13:17:56 1999
+++ ext/POSIX/POSIX.pod Thu Sep 23 17:47:51 1999
@@ -1024,7 +1024,8 @@ If you want your code to be portable, yo
should use only the conversion specifiers defined by the ANSI C
standard. These are C<aAbBcdHIjmMpSUwWxXyYZ%>.
The given arguments are made consistent
-by calling C<mktime()> before calling your system's C<strftime()> function.
+as though by calling C<mktime()> before calling your system's
+C<strftime()> function, except that the C<isdst> value is not affected.
The string for Tuesday, December 12, 1995.
--- ext/POSIX/POSIX.xs.DIST Fri Aug 20 11:51:30 1999
+++ ext/POSIX/POSIX.xs Thu Sep 23 15:39:30 1999
@@ -332,6 +332,196 @@ init_tm(struct tm *ptm) /* see mktime,
# define init_tm(ptm)
#endif
+/*
+ * mini_mktime - normalise struct tm values without the localtime()
+ * semantics (and overhead) of mktime().
+ */
+static void
+mini_mktime(struct tm *ptm)
+{
+ int yearday;
+ int secs;
+ int month, mday, year, jday;
+ int odd_cent, odd_year;
+
+#define DAYS_PER_YEAR 365
+#define DAYS_PER_QYEAR (4*DAYS_PER_YEAR+1)
+#define DAYS_PER_CENT (25*DAYS_PER_QYEAR-1)
+#define DAYS_PER_QCENT (4*DAYS_PER_CENT+1)
+#define SECS_PER_HOUR (60*60)
+#define SECS_PER_DAY (24*SECS_PER_HOUR)
+/* parentheses deliberately absent on these two, otherwise they don't work */
+#define MONTH_TO_DAYS 153/5
+#define DAYS_TO_MONTH 5/153
+/* offset to bias by March (month 4) 1st between month/mday & year finding */
+#define YEAR_ADJUST (4*MONTH_TO_DAYS+1)
+/* as used here, the algorithm leaves Sunday as day 1 unless we adjust it */
+#define WEEKDAY_BIAS 6 /* (1+6)%7 makes Sunday 0 again */
+
+/*
+ * Year/day algorithm notes:
+ *
+ * With a suitable offset for numeric value of the month, one can find
+ * an offset into the year by considering months to have 30.6 (153/5) days,
+ * using integer arithmetic (i.e., with truncation). To avoid too much
+ * messing about with leap days, we consider January and February to be
+ * the 13th and 14th month of the previous year. After that transformation,
+ * we need the month index we use to be high by 1 from 'normal human' usage,
+ * so the month index values we use run from 4 through 15.
+ *
+ * Given that, and the rules for the Gregorian calendar (leap years are those
+ * divisible by 4 unless also divisible by 100, when they must be divisible
+ * by 400 instead), we can simply calculate the number of days since some
+ * arbitrary 'beginning of time' by futzing with the (adjusted) year number,
+ * the days we derive from our month index, and adding in the day of the
+ * month. The value used here is not adjusted for the actual origin which
+ * it normally would use (1 January A.D. 1), since we're not exposing it.
+ * We're only building the value so we can turn around and get the
+ * normalised values for the year, month, day-of-month, and day-of-year.
+ *
+ * For going backward, we need to bias the value we're using so that we find
+ * the right year value. (Basically, we don't want the contribution of
+ * March 1st to the number to apply while deriving the year). Having done
+ * that, we 'count up' the contribution to the year number by accounting for
+ * full quadracenturies (400-year periods) with their extra leap days, plus
+ * the contribution from full centuries (to avoid counting in the lost leap
+ * days), plus the contribution from full quad-years (to count in the normal
+ * leap days), plus the leftover contribution from any non-leap years.
+ * At this point, if we were working with an actual leap day, we'll have 0
+ * days left over. This is also true for March 1st, however. So, we have
+ * to special-case that result, and (earlier) keep track of the 'odd'
+ * century and year contributions. If we got 4 extra centuries in a qcent,
+ * or 4 extra years in a qyear, then it's a leap day and we call it 29 Feb.
+ * Otherwise, we add back in the earlier bias we removed (the 123 from
+ * figuring in March 1st), find the month index (integer division by 30.6),
+ * and the remainder is the day-of-month. We then have to convert back to
+ * 'real' months (including fixing January and February from being 14/15 in
+ * the previous year to being in the proper year). After that, to get
+ * tm_yday, we work with the normalised year and get a new yearday value for
+ * January 1st, which we subtract from the yearday value we had earlier,
+ * representing the date we've re-built. This is done from January 1
+ * because tm_yday is 0-origin.
+ *
+ * Since POSIX time routines are only guaranteed to work for times since the
+ * UNIX epoch (00:00:00 1 Jan 1970 UTC), the fact that this algorithm
+ * applies Gregorian calendar rules even to dates before the 16th century
+ * doesn't bother me. Besides, you'd need cultural context for a given
+ * date to know whether it was Julian or Gregorian calendar, and that's
+ * outside the scope for this routine. Since we convert back based on the
+ * same rules we used to build the yearday, you'll only get strange results
+ * for input which needed normalising, or for the 'odd' century years which
+ * were leap years in the Julian calander but not in the Gregorian one.
+ * I can live with that.
+ *
+ * This algorithm also fails to handle years before A.D. 1 gracefully, but
+ * that's still outside the scope for POSIX time manipulation, so I don't
+ * care.
+ */
+
+ year = 1900 + ptm->tm_year;
+ month = ptm->tm_mon;
+ mday = ptm->tm_mday;
+ /* allow given yday with no month & mday to dominate the result */
+ if (ptm->tm_yday >= 0 && mday <= 0 && month <= 0) {
+ month = 0;
+ mday = 0;
+ jday = 1 + ptm->tm_yday;
+ }
+ else {
+ jday = 0;
+ }
+ if (month >= 2)
+ month+=2;
+ else
+ month+=14, year--;
+ yearday = DAYS_PER_YEAR * year + year/4 - year/100 + year/400;
+ yearday += month*MONTH_TO_DAYS + mday + jday;
+ /*
+ * Note that we don't know when leap-seconds were or will be,
+ * so we have to trust the user if we get something which looks
+ * like a sensible leap-second. Wild values for seconds will
+ * be rationalised, however.
+ */
+ if ((unsigned) ptm->tm_sec <= 60) {
+ secs = 0;
+ }
+ else {
+ secs = ptm->tm_sec;
+ ptm->tm_sec = 0;
+ }
+ secs += 60 * ptm->tm_min;
+ secs += SECS_PER_HOUR * ptm->tm_hour;
+ if (secs < 0) {
+ if (secs-(secs/SECS_PER_DAY*SECS_PER_DAY) < 0) {
+ /* got negative remainder, but need positive time */
+ /* back off an extra day to compensate */
+ yearday += (secs/SECS_PER_DAY)-1;
+ secs -= SECS_PER_DAY * (secs/SECS_PER_DAY - 1);
+ }
+ else {
+ yearday += (secs/SECS_PER_DAY);
+ secs -= SECS_PER_DAY * (secs/SECS_PER_DAY);
+ }
+ }
+ else if (secs >= SECS_PER_DAY) {
+ yearday += (secs/SECS_PER_DAY);
+ secs %= SECS_PER_DAY;
+ }
+ ptm->tm_hour = secs/SECS_PER_HOUR;
+ secs %= SECS_PER_HOUR;
+ ptm->tm_min = secs/60;
+ secs %= 60;
+ ptm->tm_sec += secs;
+ /* done with time of day effects */
+ /*
+ * The algorithm for yearday has (so far) left it high by 428.
+ * To avoid mistaking a legitimate Feb 29 as Mar 1, we need to
+ * bias it by 123 while trying to figure out what year it
+ * really represents. Even with this tweak, the reverse
+ * translation fails for years before A.D. 0001.
+ * It would still fail for Feb 29, but we catch that one below.
+ */
+ jday = yearday; /* save for later fixup vis-a-vis Jan 1 */
+ yearday -= YEAR_ADJUST;
+ year = (yearday / DAYS_PER_QCENT) * 400;
+ yearday %= DAYS_PER_QCENT;
+ odd_cent = yearday / DAYS_PER_CENT;
+ year += odd_cent * 100;
+ yearday %= DAYS_PER_CENT;
+ year += (yearday / DAYS_PER_QYEAR) * 4;
+ yearday %= DAYS_PER_QYEAR;
+ odd_year = yearday / DAYS_PER_YEAR;
+ year += odd_year;
+ yearday %= DAYS_PER_YEAR;
+ if (!yearday && (odd_cent==4 || odd_year==4)) { /* catch Feb 29 */
+ month = 1;
+ yearday = 29;
+ }
+ else {
+ yearday += YEAR_ADJUST; /* recover March 1st crock */
+ month = yearday*DAYS_TO_MONTH;
+ yearday -= month*MONTH_TO_DAYS;
+ /* recover other leap-year adjustment */
+ if (month > 13) {
+ month-=14;
+ year++;
+ }
+ else {
+ month-=2;
+ }
+ }
+ ptm->tm_year = year - 1900;
+ ptm->tm_mon = month;
+ ptm->tm_mday = yearday;
+ /* re-build yearday based on Jan 1 to get tm_yday */
+ year--;
+ yearday = year*DAYS_PER_YEAR + year/4 - year/100 + year/400;
+ yearday += 14*MONTH_TO_DAYS + 1;
+ ptm->tm_yday = jday - yearday;
+ /* fix tm_wday if not overridden by caller */
+ if ((unsigned)ptm->tm_wday > 6)
+ ptm->tm_wday = (jday + WEEKDAY_BIAS) % 7;
+}
#ifdef HAS_LONG_DOUBLE
# if LONG_DOUBLESIZE > DOUBLESIZE
@@ -3650,7 +3840,7 @@ strftime(fmt, sec, min, hour, mday, mon,
mytm.tm_wday = wday;
mytm.tm_yday = yday;
mytm.tm_isdst = isdst;
- (void) mktime(&mytm);
+ mini_mktime(&mytm);
len = strftime(tmpbuf, sizeof tmpbuf, fmt, &mytm);
/*
** The following is needed to handle to the situation where
--- t/lib/posix.t.DIST Tue Jul 20 13:18:13 1999
+++ t/lib/posix.t Thu Sep 23 17:00:43 1999
@@ -14,7 +14,7 @@
use strict subs;
$| = 1;
-print "1..18\n";
+print "1..26\n";
$Is_W32 = $^O eq 'MSWin32';
@@ -94,6 +94,31 @@
# -DSTRUCT_TM_HASZONE to your cflags when compiling ext/POSIX/POSIX.c.
# See ext/POSIX/hints/sunos_4.pl and ext/POSIX/hints/linux.pl
print POSIX::strftime("ok 18 # %H:%M, on %D\n", localtime());
+
+# If that worked, validate the mini_mktime() routine's normalisation of
+# input fields to strftime().
+sub try_strftime {
+ my $num = shift;
+ my $expect = shift;
+ my $got = POSIX::strftime("%a %b %d %H:%M:%S %Y %j", @_);
+ if ($got eq $expect) {
+ print "ok $num\n";
+ }
+ else {
+ print "# expected: $expect\n# got: $got\nnot ok $num\n";
+ }
+}
+
+$lc = &POSIX::setlocale(&POSIX::LC_TIME, 'C') if $Config{d_setlocale};
+try_strftime(19, "Wed Feb 28 00:00:00 1996 059", 0,0,0, 28,1,96);
+try_strftime(20, "Thu Feb 29 00:00:60 1996 060", 60,0,-24, 30,1,96);
+try_strftime(21, "Fri Mar 01 00:00:00 1996 061", 0,0,-24, 31,1,96);
+try_strftime(22, "Sun Feb 28 00:00:00 1999 059", 0,0,0, 28,1,99);
+try_strftime(23, "Mon Mar 01 00:00:00 1999 060", 0,0,24, 28,1,99);
+try_strftime(24, "Mon Feb 28 00:00:00 2000 059", 0,0,0, 28,1,100);
+try_strftime(25, "Tue Feb 29 00:00:00 2000 060", 0,0,0, 0,2,100);
+try_strftime(26, "Wed Mar 01 00:00:00 2000 061", 0,0,0, 1,2,100);
+&POSIX::setlocale(&POSIX::LC_TIME, $lc) if $Config{d_setlocale};
$| = 0;
# The following line assumes buffered output, which may be not true with EMX: |
From @gsarOn Mon, 13 Sep 1999 14:51:17 PDT, Phil Lobbes wrote:
The mktime() call was added (change#1914) to fix the behavior of http://www.xray.mpe.mpg.de/mailing-lists/perl5-porters/1998-09/msg01660.html I'd say this needs to be controlled via the hints only for the I've checked in the attached change. Someone will need to add a hints
Sarathy Inline Patch-----------------------------------8<-----------------------------------
Change 4196 by gsar@auger on 1999/09/19 22:14:29
control change#1914 via hints (causes problems on some platforms)
Affected files ...
... //depot/perl/ext/POSIX/POSIX.pod#12 edit
... //depot/perl/ext/POSIX/POSIX.xs#49 edit
... //depot/perl/ext/POSIX/hints/linux.pl#3 edit
Differences ...
==== //depot/perl/ext/POSIX/POSIX.pod#12 (text) ====
Index: perl/ext/POSIX/POSIX.pod
--- perl/ext/POSIX/POSIX.pod.~1~ Sun Sep 19 15:14:34 1999
+++ perl/ext/POSIX/POSIX.pod Sun Sep 19 15:14:34 1999
@@ -1023,7 +1023,7 @@
If you want your code to be portable, your format (C<fmt>) argument
should use only the conversion specifiers defined by the ANSI C
standard. These are C<aAbBcdHIjmMpSUwWxXyYZ%>.
-The given arguments are made consistent
+On platforms that need it, the given arguments are made consistent
by calling C<mktime()> before calling your system's C<strftime()> function.
The string for Tuesday, December 12, 1995.
==== //depot/perl/ext/POSIX/POSIX.xs#49 (text) ====
Index: perl/ext/POSIX/POSIX.xs
--- perl/ext/POSIX/POSIX.xs.~1~ Sun Sep 19 15:14:34 1999
+++ perl/ext/POSIX/POSIX.xs Sun Sep 19 15:14:34 1999
@@ -3652,7 +3652,9 @@
mytm.tm_wday = wday;
mytm.tm_yday = yday;
mytm.tm_isdst = isdst;
+#if defined(HINT_STRFTIME_NEEDS_MKTIME)
(void) mktime(&mytm);
+#endif
len = strftime(tmpbuf, sizeof tmpbuf, fmt, &mytm);
/*
** The following is needed to handle to the situation where
==== //depot/perl/ext/POSIX/hints/linux.pl#3 (text) ====
Index: perl/ext/POSIX/hints/linux.pl
--- perl/ext/POSIX/hints/linux.pl.~1~ Sun Sep 19 15:14:34 1999
+++ perl/ext/POSIX/hints/linux.pl Sun Sep 19 15:14:34 1999
@@ -2,4 +2,6 @@
# Thanks to Bart Schuller <schuller@Lunatech.com>
# See Message-ID: <19971009002636.50729@tanglefoot>
# XXX A Configure test is needed.
-$self->{CCFLAGS} = $Config{ccflags} . ' -DSTRUCT_TM_HASZONE -DHINT_SC_EXIST' ;
+$self->{CCFLAGS} = $Config{ccflags}
+ . ' -DHINT_STRFTIME_NEEDS_MKTIME'
+ . ' -DSTRUCT_TM_HASZONE -DHINT_SC_EXIST' ;
End of Patch. |
Migrated from rt.perl.org#1361 (status was 'resolved')
Searchable as RT1361$
The text was updated successfully, but these errors were encountered: