xref: /freebsd/contrib/one-true-awk/FIXES.1e (revision d37eb51047221dc3322b34db1038ff3aa533883f)

/****************************************************************
Copyright (C) Lucent Technologies 1997
All Rights Reserved

Permission to use, copy, modify, and distribute this software and
its documentation for any purpose and without fee is hereby
granted, provided that the above copyright notice appear in all
copies and that both that the copyright notice and this
permission notice and warranty disclaimer appear in supporting
documentation, and that the name Lucent Technologies or any of
its entities not be used in advertising or publicity pertaining
to distribution of the software without specific, written prior
permission.

LUCENT DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS.
IN NO EVENT SHALL LUCENT OR ANY OF ITS ENTITIES BE LIABLE FOR ANY
SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER
IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION,
ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF
THIS SOFTWARE.
****************************************************************/

This file lists all bug fixes, changes, etc., made since the AWK book
was sent to the printers in August 1987.

Sep 06, 2023:
	Fix edge case where FS is changed on commandline. Thanks to
	Gordon Shephard and Miguel Pineiro Jr.

	Fix regular expression clobbering in the lexer, where lexer does
	not make a copy of regexp literals. also makedfa memory leaks have
	been plugged. Thanks to Miguel Pineiro Jr.

Dec 15, 2022:
	Force hex escapes in strings to be no more than two characters,
	as they already are in regular expressions. This brings internal
	consistency, as well as consistency with gawk. Thanks to
	Arnold Robbins.

Sep 12, 2022:
	adjbuf minlen error (cannot be 0) in cat, resulting in NULL pbuf.
	discovered by todd miller. also use-after-free issue with
	tempfree in cat, thanks to Miguel Pineiro Jr and valgrind.

Aug 30, 2022:
	Various leaks and use-after-free issues plugged/fixed.
	Thanks to Miguel Pineiro Jr. <mpj@pineiro.cc>.

May 23, 2022:
	Memory leak when assigning a string to some of the built-in
	variables. allocated string erroneously marked DONTFREE.
	Thanks to Miguel Pineiro Jr. <mpj@pineiro.cc>.

Mar 14, 2022:
	Historic bug: command-line "name=value" assignment had been
	truncating its entry in ARGV. (circa 1989) Thanks to
	Miguel Pineiro Jr. <mpj@pineiro.cc>.

Mar 3, 2022:
	Fixed file management memory leak that appears to have been
	there since the files array was first initialized with stdin,
	stdout, and stderr (circa 1992). Thanks to Miguel Pineiro Jr.
	<mpj@pineiro.cc>.

December 8, 2021:
	The error handling in closefile and closeall was mangled. Long
	standing warnings had been made fatal and some fatal errors went
	undetected. Thanks to Miguel Pineiro Jr. <mpj@pineiro.cc>.

Nov 03, 2021:
        getline accesses uninitialized data after getrec()
	returns 0 on EOF and leaves the contents of buf unchanged.
	Thanks to Volodymyr Gubarkov, and Todd C Miller.

Oct 12, 2021:
	The fix for #83 changed the code to insert 2 chars, but the
	call to adjbuf just above it only allows for 1 char. This can
	cause a heap buffer overflow.

July 27, 2021:
	As per IEEE Std 1003.1-2008, -F "str" is now consistent with
	-v FS="str" when str is null. Thanks to Warner Losh.

July 24, 2021:
	Fix readrec's definition of a record. This fixes an issue
	with NetBSD's RS regular expression support that can cause
	an infinite read loop. Thanks to Miguel Pineiro Jr.

	Fix regular expression RS ^-anchoring. RS ^-anchoring needs to
	know if it is reading the first record of a file. This change
	restores a missing line that was overlooked when porting NetBSD's
	RS regex functionality. Thanks to Miguel Pineiro Jr.

	Fix size computation in replace_repeat() for special case
	REPEAT_WITH_Q. Thanks to Todd C. Miller.

February 15, 2021:
	Small fix so that awk will compile again with g++. Thanks to
	Arnold Robbins.

January 06, 2021:
	Fix a decision bug with trailing stuff in lib.c:is_valid_number
	after recent changes. Thanks to Ozan Yigit.

December 18, 2020:
	Fix problems converting inf and NaN values in lib.c:is_valid_number.
	Enhance number to string conversion to do the right thing for
	NaN and inf values.  Things are now pretty much the same as in
	gawk.  (Found a gawk bug while we're at it.) Added a torture
	test for these values.  Thanks to Arnold Robbins.  Allows closing
	of PR #101.

December 15, 2020:
	Merge PR #99, which gets the right header for strcasecmp.
	Thanks to GitHub user michaelforney.

December 8, 2020:
	Merge PR #98: Disallow hex data. Allow only +nan, -nan,
	+inf, -inf (case independent) to give NaN and infinity values.
	Improve things so that string to double conversion is only
	done once, yielding something of a speedup.  This obviate
	PR #95. Thanks to Arnold Robbins.

December 3, 2020:
	Fix to argument parsing to avoid printing spurious newlines.
	Thanks to Todd Miller. Merges PR #97.

October 13, 2020:
	Add casts before all the calls to malloc/calloc/realloc in order
	to get it to compile with g++. Thanks to Arnold Robbins.

August 16, 2020:
	Additional fixes for DJGPP. Thanks to Eli Zaretskii for
	the testing.

August 7, 2020:
	Merge PR #93, which adds casts to (void*) for debug prints
	using the %p format specifier. Thanks to GitHub user YongHaoWu
	("Chris") for the fixes.

August 4, 2020:
	In run.c, use non-restartable multibyte routines to attain
	portability to DJGPP. Should fix Issue 92. Thanks to Albert Wik
	for the report and to Todd Miller for the suggested fix.

July 30, 2020:
	Merge PRs 88-91 which fix small bugs. Thanks to Todd Miller and
	Tim van der Molen for the fixes.

	In order to make life easier, we move exclusively to bison
	as the parser generator.

July 2, 2020:
	Merge PRs 85 and 86 which fix regressions. Thanks to
	Tim van der Molen for the fixes.

June 25, 2020:
	Merge PRs 82 and 84. The latter fixes issue #83. Thanks to
	Todd Miller and awkfan77.

June 12, 2020:
	Clear errno before calling errcheck to avoid any spurious errors
	left over from previous calls that may have set it. Thanks to
	Todd Miller for the fix, from PR #80.

	Fix Issue #78 by allowing \r to follow floating point numbers in
	lib.c:is_number. Thanks to GitHub user ajcarr for the report
	and to Arnold Robbins for the fix.

June 5, 2020:
	In fldbld(), make sure that inputFS is set before trying to
	use it. Thanks to  Steffen Nurpmeso <steffen@sdaoden.eu>
	for the report.

May 5, 2020:
	Fix checks for compilers that can handle noreturn. Thanks to
	GitHub user enh-google for pointing it out. Closes Issue #79.

April 16, 2020:
	Handle old compilers that don't support C11 (for noreturn).
	Thanks to Arnold Robbins.

April 5, 2020:
	Use <stdnoreturn.h> and noreturn instead of GCC attributes.
	Thanks to GitHub user awkfan77. Closes PR #77.

February 28, 2020:
	More cleanups from Christos Zoulas: notably backslash continuation
	inside strings removes the newline and a fix for RS = "^a".
	Fix for address sanitizer-found problem. Thanks to GitHub user
	enh-google.

February 19, 2020:
	More small cleanups from Christos Zoulas.

February 18, 2020:
	Additional cleanups from Christos Zoulas. It's no longer necessary
	to use the -y flag to bison.

February 6, 2020:
	Additional small cleanups from Christos Zoulas. awk is now
	a little more robust about reporting I/O errors upon exit.

January 31, 2020:
	Merge PR #70, which avoids use of variable length arrays. Thanks
	to GitHub user michaelforney.  Fix issue #60 ({0} in interval
	expressions doesn't work).  Also get all tests working again.
	Thanks to Arnold Robbins.

January 24, 2020:
	A number of small cleanups from Christos Zoulas.  Add the close
	on exec flag to files/pipes opened for redirection; courtesy of
	Arnold Robbins.

January 19, 2020:
	If POSIXLY_CORRECT is set in the environment, then sub and gsub
	use POSIX rules for multiple backslashes.  This fixes Issue #66,
	while maintaining backwards compatibility.

January 9, 2020:
	Input/output errors on closing files are now fatal instead of
	mere warnings. Thanks to Martijn Dekker <martijn@inlv.org>.

January 5, 2020:
	Fix a bug in the concatentation of two string constants into
	one done in the grammar.  Fixes GitHub issue #61.  Thanks
	to GitHub user awkfan77 for pointing out the direction for
	the fix.  New test T.concat added to the test suite.
	Fix a few memory leaks reported by valgrind, as well.

December 27, 2019:
	Fix a bug whereby a{0,3} could match four a's.  Thanks to
	"Anonymous AWK fan" for the report.

December 11, 2019:
	Further printf-related fixes for 32 bit systems.
	Thanks again to Christos Zoulas.

December 8, 2019:
	Fix the return value of sprintf("%d") on 32 bit systems.
	Thanks to Jim Lowe for the report and to Christos Zoulas
	for the fix.

November 10, 2019:
	Convert a number of Boolean integer variables into
	actual bools. Convert compile_time variable into an
	enum and simplify some of the related code.  Thanks
	to Arnold Robbins.

November 8, 2019:
	Fix from Ori Bernstein to get UTF-8 characters instead of
	bytes when FS = "".  This is currently the only bit of
	the One True Awk that understands multibyte characters.
	From Arnold Robbins, apply some cleanups in the test suite.

October 25, 2019:
	More fixes and cleanups from NetBSD, courtesy of Christos
	Zoulas. Merges PRs 54 and 55.

October 24, 2019:
	Import second round of code cleanups from NetBSD. Much thanks
	to Christos Zoulas (GitHub user zoulasc). Merges PR 53.
	Add an optimization for string concatenation, also from
	Christos.

October 17, 2019:
	Import code cleanups from NetBSD. Much thanks to Christos
	Zoulas (GitHub user zoulasc). Merges PR 51.

October 6, 2019:
	Import code from NetBSD awk that implements RS as a regular
	expression.

September 10, 2019:
	Fixes for various array / memory overruns found via gcc's
	-fsanitize=unknown. Thanks to Alexander Richardson (GitHub
	user arichardson). Merges PRs 47 and 48.

July 28, 2019:
	Import grammar optimization from NetBSD: Two string constants
	concatenated together get turned into a single string.

July 26, 2019:
	Support POSIX-specified C-style escape sequences "\a" (alarm)
	and "\v" (vertical tab) in command line arguments and regular
	expressions, further to the support for them in strings added on
	Apr 9, 1989. These now no longer match as literal "a" and "v"
	characters (as they don't on other awk implementations).
	Thanks to Martijn Dekker.

July 17, 2019:
	Pull in a number of code cleanups and minor fixes from
	Warner Losh's bsd-ota branch.  The only user visible change
	is the use of random(3) as the random number generator.
	Thanks to Warner Losh for collecting all these fixes in
	one easy place to get them from.

July 16, 2019:
	Fix field splitting to use FS value as of the time a record
	was read or assigned to.  Thanks to GitHub user Cody Mello (melloc)
	for the fix. (Merged from his branch, via PR #42.) Updated
	testdir/T.split per said PR as well.

June 24, 2019:
	Extract awktest.tar into testdir directory. Add some very
	simple mechanics to the makefile for running the tests and
	for cleaning up. No changes to awk itself.

June 17, 2019:
	Disallow deleting SYMTAB and its elements, which creates
	use-after-free bugs. Thanks to GitHub user Cody Mello (melloc)
	for the fix. (Merged from PR #43.)

June 5, 2019:
	Allow unmatched right parenthesis in a regular expression to
	be treated literally. Fixes Issue #40. Thanks to GitHub user
	Warner Losh (bsdimp) for the report. Thanks to Arnold Robbins
	for the fix.

May 29,2019:
	Fix check for command line arguments to no longer require that
	first character after '=' not be another '='. Reverts change of
	August 11, 1989. Thanks to GitHub user Jamie Landeg Jones for
	pointing out the issue; from Issue #38.

Apr 7, 2019:
	Update awktest.tar(p.50) to use modern options to sort. Needed
	for Android development. Thanks to GitHub user mohd-akram (Mohamed
	Akram).  From Issue #33.

Mar 12, 2019:
	Added very simplistic support for cross-compiling in the
	makefile.  We are NOT going to go in the direction of the
	autotools, though.  Thanks to GitHub user nee-san for
	the basic change. (Merged from PR #34.)

Mar 5, 2019:
	Added support for POSIX-standard interval expressions (a.k.a.
	bounds, a.k.a. repetition expressions) in regular expressions,
	backported (via NetBSD) from Apple awk-24 (20070501).
	Thanks to Martijn Dekker <martijn@inlv.org> for the port.
	(Merged from PR #30.)

Mar 3, 2019:
	Merge PRs as follows:
	#12: Avoid undefined behaviour when using ctype(3) functions in
	     relex(). Thanks to GitHub user iamleot.
	#31: Make getline handle numeric strings, and update FIXES. Thanks
	     to GitHub user Arnold Robbins (arnoldrobbins)
	#32: maketab: support build systems with read-only source. Thanks
	     to GitHub user enh.

Jan 25, 2019:
	Make getline handle numeric strings properly in all cases.
	(Thanks, Arnold.)

Jan 21, 2019:
	Merged a number of small fixes from GitHub pull requests.
	Thanks to GitHub users Arnold Robbins (arnoldrobbins),
	Cody Mello (melloc) and Christoph Junghans (junghans).
	PR numbers: 13-21, 23, 24, 27.

Oct 25, 2018:
	Added test in maketab.c to prevent generating a proctab entry
	for YYSTYPE_IS_DEFINED.  It was harmless but some gcc settings
	generated a warning message.  Thanks to Nan Xiao for report.

Aug 27, 2018:
	Disallow '$' in printf formats; arguments evaluated in order
	and printed in order.

	Added some casts to silence warnings on debugging printfs.
	(Thanks, Arnold.)

Aug 23, 2018:
        A long list of fixes courtesy of Arnold Robbins,
        to whom profound thanks.

        1. ofs-rebuild: OFS value used to rebuild the record was incorrect.
        Fixed August 19, 2014. Revised fix August 2018.

        2. system-status: Instead of a floating-point division by 256, use
        the wait(2) macros to create a reasonable exit status.
        Fixed March 12, 2016.

        3. space: Use provided xisblank() function instead of ispace() for
        matching [[:blank:]].

        4. a-format: Add POSIX standard %a and %A to supported formats. Check
        at runtime that this format is available.

        5. decr-NF: Decrementing NF did not change $0. This is a decades-old
        bug. There are interactions with the old and new value of OFS as well.
        Most of the fix came from the NetBSD awk.

        6. string-conv: String conversions of scalars were sticky.  Once a
        conversion to string happened, even with OFMT, that value was used until
        a new numeric value was assigned, even if OFMT differed from CONVFMT,
        and also if CONVFMT changed.

        7. unary-plus: Unary plus on a string constant returned the string.
        Instead, it should convert the value to numeric and give that value.

	Also added Arnold's tests for these to awktest.tar as T.arnold.

Aug 15, 2018:
	fixed mangled awktest.tar (thanks, Arnold), posted all
	current (very minor) fixes to github / onetrueawk

Jun 7, 2018:
	(yes, a long layoff)
	Updated some broken tests (beebe.tar, T.lilly)
	[thanks to Arnold Robbins]

Mar 26, 2015:
	buffer overflow in error reporting; thanks to tobias ulmer
	and john-mark gurney for spotting it and the fix.

Feb 4, 2013:
	cleaned up a handful of tests that didn't seem to actually
	test for correct behavior: T.latin1, T.gawk.

Jan 5, 2013:
	added ,NULL initializer to static Cells in run.c; not really
	needed but cleaner.  Thanks to Michael Bombardieri.

Dec 20, 2012:
	fiddled makefile to get correct yacc and bison flags.  pick yacc
	(linux) or bison (mac) as necessary.

	added  __attribute__((__noreturn__)) to a couple of lines in
	proto.h, to silence someone's enthusiastic checker.

	fixed obscure call by value bug in split(a[1],a) reported on
	9fans.  the management of temporary values is just a mess; i
	took a shortcut by making an extra string copy.  thanks
	to paul patience and arnold robbins for passing it on and for
	proposed patches.

	tiny fiddle in setfval to eliminate -0 results in T.expr, which
	has irritated me for 20+ years.

Aug 10, 2011:
	another fix to avoid core dump with delete(ARGV); again, many thanks
	to ruslan ermilov.

Aug 7, 2011:
	split(s, a, //) now behaves the same as split(s, a, "")

Jun 12, 2011:
	/pat/, \n /pat/ {...} is now legal, though bad style to use.

	added checks to new -v code that permits -vnospace; thanks to
	ruslan ermilov for spotting this and providing the patch.

	removed fixed limit on number of open files; thanks to aleksey
	cheusov and christos zoulos.

	fixed day 1 bug that resurrected deleted elements of ARGV when
	used as filenames (in lib.c).

	minor type fiddles to make gcc -Wall -pedantic happier (but not
	totally so); turned on -fno-strict-aliasing in makefile.

May 6, 2011:
	added #ifdef for isblank.
	now allows -ffoo as well as -f foo arguments.
	(thanks, ruslan)

May 1, 2011:
	after advice from todd miller, kevin lo, ruslan ermilov,
	and arnold robbins, changed srand() to return the previous
	seed (which is 1 on the first call of srand).  the seed is
	an Awkfloat internally though converted to unsigned int to
	pass to the library srand().  thanks, everyone.

	fixed a subtle (and i hope low-probability) overflow error
	in fldbld, by adding space for one extra \0.  thanks to
	robert bassett for spotting this one and providing a fix.

	removed the files related to compilation on windows.  i no
	longer have anything like a current windows environment, so
	i can't test any of it.

May 23, 2010:
	fixed long-standing overflow bug in run.c; many thanks to
	nelson beebe for spotting it and providing the fix.

	fixed bug that didn't parse -vd=1 properly; thanks to santiago
	vila for spotting it.

Feb 8, 2010:
	i give up.  replaced isblank with isspace in b.c; there are
	no consistent header files.

Nov 26, 2009:
	fixed a long-standing issue with when FS takes effect.  a
	change to FS is now noticed immediately for subsequent splits.

	changed the name getline() to awkgetline() to avoid yet another
	name conflict somewhere.

Feb 11, 2009:
	temporarily for now defined HAS_ISBLANK, since that seems to
	be the best way through the thicket.  isblank arrived in C99,
	but seems to be arriving at different systems at different
	times.

Oct 8, 2008:
	fixed typo in b.c that set tmpvec wrongly.  no one had ever
	run into the problem, apparently.  thanks to alistair crooks.

Oct 23, 2007:
	minor fix in lib.c: increase inputFS to 100, change malloc
	for fields to n+1.

	fixed memory fault caused by out of order test in setsval.

	thanks to david o'brien, freebsd, for both fixes.

May 1, 2007:
	fiddle in makefile to fix for BSD make; thanks to igor sobrado.

Mar 31, 2007:
	fixed some null pointer refs calling adjbuf.

Feb 21, 2007:
	fixed a bug in matching the null RE in sub and gsub.  thanks to al aho
	who actually did the fix (in b.c), and to wolfgang seeberg for finding
	it and providing a very compact test case.

	fixed quotation in b.c; thanks to Hal Pratt and the Princeton Dante
	Project.

	removed some no-effect asserts in run.c.

	fiddled maketab.c to not complain about bison-generated values.

	removed the obsolete -V argument; fixed --version to print the
	version and exit.

	fixed wording and an outright error in the usage message; thanks to igor
	sobrado and jason mcintyre.

	fixed a bug in -d that caused core dump if no program followed.

Jan 1, 2007:
	dropped mac.code from makefile; there are few non-MacOSX
	mac's these days.

Jan 17, 2006:
	system() not flagged as unsafe in the unadvertised -safe option.
	found it while enhancing tests before shipping the ;login: article.
	practice what you preach.

	removed the 9-years-obsolete -mr and -mf flags.

	added -version and --version options.

	core dump on linux with BEGIN {nextfile}, now fixed.

	removed some #ifdef's in run.c and lex.c that appear to no
	longer be necessary.

Apr 24, 2005:
	modified lib.c so that values of $0 et al are preserved in the END
	block, apparently as required by posix.  thanks to havard eidnes
	for the report and code.

Jan 14, 2005:
	fixed infinite loop in parsing, originally found by brian tsang.
	thanks to arnold robbins for a suggestion that started me
	rethinking it.

Dec 31, 2004:
	prevent overflow of -f array in main, head off potential error in
	call of SYNTAX(), test malloc return in lib.c, all with thanks to
	todd miller.

Dec 22, 2004:
	cranked up size of NCHARS; coverity thinks it can be overrun with
	smaller size, and i think that's right.  added some assertions to b.c
	to catch places where it might overrun.  the RE code is still fragile.

Dec 5, 2004:
	fixed a couple of overflow problems with ridiculous field numbers:
	e.g., print $(2^32-1).  thanks to ruslan ermilov, giorgos keramidas
	and david o'brien at freebsd.org for patches.  this really should
	be re-done from scratch.

Nov 21, 2004:
	fixed another 25-year-old RE bug, in split.  it's another failure
	to (re-)initialize.  thanks to steve fisher for spotting this and
	providing a good test case.

Nov 22, 2003:
	fixed a bug in regular expressions that dates (so help me) from 1977;
	it's been there from the beginning.  an anchored longest match that
	was longer than the number of states triggered a failure to initialize
	the machine properly.  many thanks to moinak ghosh for not only finding
	this one but for providing a fix, in some of the most mysterious
	code known to man.

	fixed a storage leak in call() that appears to have been there since
	1983 or so -- a function without an explicit return that assigns a
	string to a parameter leaked a Cell.  thanks to moinak ghosh for
	spotting this very subtle one.

Jul 31, 2003:
	fixed, thanks to andrey chernov and ruslan ermilov, a bug in lex.c
	that mis-handled the character 255 in input.  (it was being compared
	to EOF with a signed comparison.)

Jul 29, 2003:
	fixed (i think) the long-standing botch that included the beginning of
	line state ^ for RE's in the set of valid characters; this led to a
	variety of odd problems, including failure to properly match certain
	regular expressions in non-US locales.  thanks to ruslan for keeping
	at this one.

Jul 28, 2003:
	n-th try at getting internationalization right, with thanks to volker
	kiefel, arnold robbins and ruslan ermilov for advice, though they
	should not be blamed for the outcome.  according to posix, "."  is the
	radix character in programs and command line arguments regardless of
	the locale; otherwise, the locale should prevail for input and output
	of numbers.  so it's intended to work that way.

	i have rescinded the attempt to use strcoll in expanding shorthands in
	regular expressions (cclenter).  its properties are much too
	surprising; for example [a-c] matches aAbBc in locale en_US but abBcC
	in locale fr_CA.  i can see how this might arise by implementation
	but i cannot explain it to a human user.  (this behavior can be seen
	in gawk as well; we're leaning on the same library.)

	the issue appears to be that strcoll is meant for sorting, where
	merging upper and lower case may make sense (though note that unix
	sort does not do this by default either).  it is not appropriate
	for regular expressions, where the goal is to match specific
	patterns of characters.  in any case, the notations [:lower:], etc.,
	are available in awk, and they are more likely to work correctly in
	most locales.

	a moratorium is hereby declared on internationalization changes.
	i apologize to friends and colleagues in other parts of the world.
	i would truly like to get this "right", but i don't know what
	that is, and i do not want to keep making changes until it's clear.

Jul 4, 2003:
	fixed bug that permitted non-terminated RE, as in "awk /x".

Jun 1, 2003:
	subtle change to split: if source is empty, number of elems
	is always 0 and the array is not set.

Mar 21, 2003:
	added some parens to isblank, in another attempt to make things
	internationally portable.

Mar 14, 2003:
	the internationalization changes, somewhat modified, are now
	reinstated.  in theory awk will now do character comparisons
	and case conversions in national language, but "." will always
	be the decimal point separator on input and output regardless
	of national language.  isblank(){} has an #ifndef.

	this no longer compiles on windows: LC_MESSAGES isn't defined
	in vc6++.

	fixed subtle behavior in field and record splitting: if FS is
	a single character and RS is not empty, \n is NOT a separator.
	this tortuous reading is found in the awk book; behavior now
	matches gawk and mawk.

Dec 13, 2002:
	for the moment, the internationalization changes of nov 29 are
	rolled back -- programs like x = 1.2 don't work in some locales,
	because the parser is expecting x = 1,2.  until i understand this
	better, this will have to wait.

Nov 29, 2002:
	modified b.c (with tiny changes in main and run) to support
	locales, using strcoll and iswhatever tests for posix character
	classes.  thanks to ruslan ermilov (ru@freebsd.org) for code.
	the function isblank doesn't seem to have propagated to any
	header file near me, so it's there explicitly.  not properly
	tested on non-ascii character sets by me.

Jun 28, 2002:
	modified run/format() and tran/getsval() to do a slightly better
	job on using OFMT for output from print and CONVFMT for other
	number->string conversions, as promised by posix and done by
	gawk and mawk.  there are still places where it doesn't work
	right if CONVFMT is changed; by then the STR attribute of the
	variable has been irrevocably set.  thanks to arnold robbins for
	code and examples.

	fixed subtle bug in format that could get core dump.  thanks to
	Jaromir Dolecek <jdolecek@NetBSD.org> for finding and fixing.
	minor cleanup in run.c / format() at the same time.

	added some tests for null pointers to debugging printf's, which
	were never intended for external consumption.  thanks to dave
	kerns (dkerns@lucent.com) for pointing this out.

	GNU compatibility: an empty regexp matches anything (thanks to
	dag-erling smorgrav, des@ofug.org).  subject to reversion if
	this does more harm than good.

	pervasive small changes to make things more const-correct, as
	reported by gcc's -Wwrite-strings.  as it says in the gcc manual,
	this may be more nuisance than useful.  provoked by a suggestion
	and code from arnaud desitter, arnaud@nimbus.geog.ox.ac.uk

	minor documentation changes to note that this now compiles out
	of the box on Mac OS X.

Feb 10, 2002:
	changed types in posix chars structure to quiet solaris cc.

Jan 1, 2002:
	fflush() or fflush("") flushes all files and pipes.

	length(arrayname) returns number of elements; thanks to
	arnold robbins for suggestion.

	added a makefile.win to make it easier to build on windows.
	based on dan allen's buildwin.bat.

Nov 16, 2001:
	added support for posix character class names like [:digit:],
	which are not exactly shorter than [0-9] and perhaps no more
	portable.  thanks to dag-erling smorgrav for code.

Feb 16, 2001:
	removed -m option; no longer needed, and it was actually
	broken (noted thanks to volker kiefel).

Feb 10, 2001:
	fixed an appalling bug in gettok: any sequence of digits, +,-, E, e,
	and period was accepted as a valid number if it started with a period.
	this would never have happened with the lex version.

	other 1-character botches, now fixed, include a bare $ and a
	bare " at the end of the input.

Feb 7, 2001:
	more (const char *) casts in b.c and tran.c to silence warnings.

Nov 15, 2000:
	fixed a bug introduced in august 1997 that caused expressions
	like $f[1] to be syntax errors.  thanks to arnold robbins for
	noticing this and providing a fix.

Oct 30, 2000:
	fixed some nextfile bugs: not handling all cases.  thanks to
	arnold robbins for pointing this out.  new regressions added.

	close() is now a function.  it returns whatever the library
	fclose returns, and -1 for closing a file or pipe that wasn't
	opened.

Sep 24, 2000:
	permit \n explicitly in character classes; won't work right
	if comes in as "[\n]" but ok as /[\n]/, because of multiple
	processing of \'s.  thanks to arnold robbins.

July 5, 2000:
	minor fiddles in tran.c to keep compilers happy about uschar.
	thanks to norman wilson.

May 25, 2000:
	yet another attempt at making 8-bit input work, with another
	band-aid in b.c (member()), and some (uschar) casts to head
	off potential errors in subscripts (like isdigit).  also
	changed HAT to NCHARS-2.  thanks again to santiago vila.

	changed maketab.c to ignore apparently out of range definitions
	instead of halting; new freeBSD generates one.  thanks to
	jon snader <jsnader@ix.netcom.com> for pointing out the problem.

May 2, 2000:
	fixed an 8-bit problem in b.c by making several char*'s into
	unsigned char*'s.  not clear i have them all yet.  thanks to
	Santiago Vila <sanvila@unex.es> for the bug report.

Apr 21, 2000:
	finally found and fixed a memory leak in function call; it's
	been there since functions were added ~1983.  thanks to
	jon bentley for the test case that found it.

	added test in envinit to catch environment "variables" with
	names beginning with '='; thanks to Berend Hasselman.

Jul 28, 1999:
	added test in defn() to catch function foo(foo), which
	otherwise recurses until core dump.  thanks to arnold
	robbins for noticing this.

Jun 20, 1999:
	added *bp in gettok in lex.c; appears possible to exit function
	without terminating the string.  thanks to russ cox.

Jun 2, 1999:
	added function stdinit() to run to initialize files[] array,
	in case stdin, etc., are not constants; some compilers care.

May 10, 1999:
	replaced the ERROR ... FATAL, etc., macros with functions
	based on vprintf, to avoid problems caused by overrunning
	fixed-size errbuf array.  thanks to ralph corderoy for the
	impetus, and for pointing out a string termination bug in
	qstring as well.

Apr 21, 1999:
	fixed bug that caused occasional core dumps with commandline
	variable with value ending in \.  (thanks to nelson beebe for
	the test case.)

Apr 16, 1999:
	with code kindly provided by Bruce Lilly, awk now parses
	/=/ and similar constructs more sensibly in more places.
	Bruce also provided some helpful test cases.

Apr 5, 1999:
	changed true/false to True/False in run.c to make it
	easier to compile with C++.  Added some casts on malloc
	and realloc to be honest about casts; ditto.  changed
	ltype int to long in struct rrow to reduce some 64-bit
	complaints; other changes scattered throughout for the
	same purpose.  thanks to Nelson Beebe for these portability
	improvements.

	removed some horrible pointer-int casting in b.c and elsewhere
	by adding ptoi and itonp to localize the casts, which are
	all benign.  fixed one incipient bug that showed up on sgi
	in 64-bit mode.

	reset lineno for new source file; include filename in error
	message.  also fixed line number error in continuation lines.
	(thanks to Nelson Beebe for both of these.)

Mar 24, 1999:
	Nelson Beebe notes that irix 5.3 yacc dies with a bogus
	error; use a newer version or switch to bison, since sgi
	is unlikely to fix it.

Mar 5, 1999:
	changed isnumber to is_number to avoid the problem caused by
	versions of ctype.h that include the name isnumber.

	distribution now includes a script for building on a Mac,
	thanks to Dan Allen.

Feb 20, 1999:
	fixed memory leaks in run.c (call) and tran.c (setfval).
	thanks to Stephen Nutt for finding these and providing the fixes.

Jan 13, 1999:
	replaced srand argument by (unsigned int) in run.c;
	avoids problem on Mac and potentially on Unix & Windows.
	thanks to Dan Allen.

	added a few (int) casts to silence useless compiler warnings.
	e.g., errorflag= in run.c jump().

	added proctab.c to the bundle outout; one less thing
	to have to compile out of the box.

	added calls to _popen and _pclose to the win95 stub for
	pipes (thanks to Steve Adams for this helpful suggestion).
	seems to work, though properties are not well understood
	by me, and it appears that under some circumstances the
	pipe output is truncated.  Be careful.

Oct 19, 1998:
	fixed a couple of bugs in getrec: could fail to update $0
	after a getline var; because inputFS wasn't initialized,
	could split $0 on every character, a misleading diversion.

	fixed caching bug in makedfa: LRU was actually removing
	least often used.

	thanks to ross ridge for finding these, and for providing
	great bug reports.

May 12, 1998:
	fixed potential bug in readrec: might fail to update record
	pointer after growing.  thanks to dan levy for spotting this
	and suggesting the fix.

Mar 12, 1998:
	added -V to print version number and die.

[notify dave kerns, dkerns@dacsoup.ih.lucent.com]

Feb 11, 1998:
	subtle silent bug in lex.c: if the program ended with a number
	longer than 1 digit, part of the input would be pushed back and
	parsed again because token buffer wasn't terminated right.
	example:  awk 'length($0) > 10'.  blush.  at least i found it
	myself.

Aug 31, 1997:
	s/adelete/awkdelete/: SGI uses this in malloc.h.
	thanks to nelson beebe for pointing this one out.

Aug 21, 1997:
	fixed some bugs in sub and gsub when replacement includes \\.
	this is a dark, horrible corner, but at least now i believe that
	the behavior is the same as gawk and the intended posix standard.
	thanks to arnold robbins for advice here.

Aug 9, 1997:
	somewhat regretfully, replaced the ancient lex-based lexical
	analyzer with one written in C.  it's longer, generates less code,
	and more portable; the old one depended too much on mysterious
	properties of lex that were not preserved in other environments.
	in theory these recognize the same language.

	now using strtod to test whether a string is a number, instead of
	the convoluted original function.  should be more portable and
	reliable if strtod is implemented right.

	removed now-pointless optimization in makefile that tries to avoid
	recompilation when awkgram.y is changed but symbols are not.

	removed most fixed-size arrays, though a handful remain, some
	of which are unchecked.  you have been warned.

Aug 4, 1997:
	with some trepidation, replaced the ancient code that managed
	fields and $0 in fixed-size arrays with arrays that grow on
	demand.  there is still some tension between trying to make this
	run fast and making it clean; not sure it's right yet.

	the ill-conceived -mr and -mf arguments are now useful only
	for debugging.  previous dynamic string code removed.

	numerous other minor cleanups along the way.

Jul 30, 1997:
	using code provided by dan levy (to whom profuse thanks), replaced
	fixed-size arrays and awkward kludges by a fairly uniform mechanism
	to grow arrays as needed for printf, sub, gsub, etc.

Jul 23, 1997:
	falling off the end of a function returns "" and 0, not 0.
	thanks to arnold robbins.

Jun 17, 1997:
	replaced several fixed-size arrays by dynamically-created ones
	in run.c; added overflow tests to some previously unchecked cases.
	getline, toupper, tolower.

	getline code is still broken in that recursive calls may wind
	up using the same space.  [fixed later]

	increased RECSIZE to 8192 to push problems further over the horizon.

	added \r to \n as input line separator for programs, not data.
	damn CRLFs.

	modified format() to permit explicit printf("%c", 0) to include
	a null byte in output.  thanks to ken stailey for the fix.

	added a "-safe" argument that disables file output (print >,
	print >>), process creation (cmd|getline, print |, system), and
	access to the environment (ENVIRON).  this is a first approximation
	to a "safe" version of awk, but don't rely on it too much.  thanks
	to joan feigenbaum and matt blaze for the inspiration long ago.

Jul 8, 1996:
	fixed long-standing bug in sub, gsub(/a/, "\\\\&"); thanks to
	ralph corderoy.

Jun 29, 1996:
	fixed awful bug in new field splitting; didn't get all the places
	where input was done.

Jun 28, 1996:
	changed field-splitting to conform to posix definition: fields are
	split using the value of FS at the time of input; it used to be
	the value when the field or NF was first referred to, a much less
	predictable definition.  thanks to arnold robbins for encouragement
	to do the right thing.

May 28, 1996:
	fixed appalling but apparently unimportant bug in parsing octal
	numbers in reg exprs.

	explicit hex in reg exprs now limited to 2 chars: \xa, \xaa.

May 27, 1996:
	cleaned up some declarations so gcc -Wall is now almost silent.

	makefile now includes backup copies of ytab.c and lexyy.c in case
	one makes before looking; it also avoids recreating lexyy.c unless
	really needed.

	s/aprintf/awkprint, s/asprintf/awksprintf/ to avoid some name clashes
	with unwisely-written header files.

	thanks to jeffrey friedl for several of these.

May 26, 1996:
	an attempt to rationalize the (unsigned) char issue.  almost all
	instances of unsigned char have been removed; the handful of places
	in b.c where chars are used as table indices have been hand-crafted.
	added some latin-1 tests to the regression, but i'm not confident;
	none of my compilers seem to care much.  thanks to nelson beebe for
	pointing out some others that do care.

May 2, 1996:
	removed all register declarations.

	enhanced split(), as in gawk, etc:  split(s, a, "") splits s into
	a[1]...a[length(s)] with each character a single element.

	made the same changes for field-splitting if FS is "".

	added nextfile, as in gawk: causes immediate advance to next
	input file. (thanks to arnold robbins for inspiration and code).

	small fixes to regexpr code:  can now handle []], [[], and
	variants;  [] is now a syntax error, rather than matching
	everything;  [z-a] is now empty, not z.  far from complete
	or correct, however.  (thanks to jeffrey friedl for pointing out
	some awful behaviors.)

Apr 29, 1996:
	replaced uchar by uschar everywhere; apparently some compilers
	usurp this name and this causes conflicts.

	fixed call to time in run.c (bltin); arg is time_t *.

	replaced horrible pointer/long punning in b.c by a legitimate
	union.  should be safer on 64-bit machines and cleaner everywhere.
	(thanks to nelson beebe for pointing out some of these problems.)

	replaced nested comments by #if 0...#endif in run.c, lib.c.

	removed getsval, setsval, execute macros from run.c and lib.c.
	machines are 100x faster than they were when these macros were
	first used.

	revised filenames: awk.g.y => awkgram.y, awk.lx.l => awklex.l,
	y.tab.[ch] => ytab.[ch], lex.yy.c => lexyy.c, all in the aid of
	portability to nameless systems.

	"make bundle" now includes yacc and lex output files for recipients
	who don't have yacc or lex.

Aug 15, 1995:
	initialized Cells in setsymtab more carefully; some fields
	were not set.  (thanks to purify, all of whose complaints i
	think i now understand.)

	fixed at least one error in gsub that looked at -1-th element
	of an array when substituting for a null match (e.g., $).

	delete arrayname is now legal; it clears the elements but leaves
	the array, which may not be the right behavior.

	modified makefile: my current make can't cope with the test used
	to avoid unnecessary yacc invocations.

Jul 17, 1995:
	added dynamically growing strings to awk.lx.l and b.c
	to permit regular expressions to be much bigger.
	the state arrays can still overflow.

Aug 24, 1994:
	detect duplicate arguments in function definitions (mdm).

May 11, 1994:
	trivial fix to printf to limit string size in sub().

Apr 22, 1994:
	fixed yet another subtle self-assignment problem:
	$1 = $2; $1 = $1 clobbered $1.

	Regression tests now use private echo, to avoid quoting problems.

Feb 2, 1994:
	changed error() to print line number as %d, not %g.

Jul 23, 1993:
	cosmetic changes: increased sizes of some arrays,
	reworded some error messages.

	added CONVFMT as in posix (just replaced OFMT in getsval)

	FILENAME is now "" until the first thing that causes a file
	to be opened.

Nov 28, 1992:
	deleted yyunput and yyoutput from proto.h;
	different versions of lex give these different declarations.

May 31, 1992:
	added -mr N and -mf N options: more record and fields.
	these really ought to adjust automatically.

	cleaned up some error messages; "out of space" now means
	malloc returned NULL in all cases.

	changed rehash so that if it runs out, it just returns;
	things will continue to run slow, but maybe a bit longer.

Apr 24, 1992:
	remove redundant close of stdin when using -f -.

	got rid of core dump with -d; awk -d just prints date.

Apr 12, 1992:
	added explicit check for /dev/std(in,out,err) in redirection.
	unlike gawk, no /dev/fd/n yet.

	added (file/pipe) builtin.  hard to test satisfactorily.
	not posix.

Feb 20, 1992:
	recompile after abortive changes;  should be unchanged.

Dec 2, 1991:
	die-casting time:  converted to ansi C, installed that.

Nov 30, 1991:
	fixed storage leak in freefa, failing to recover [N]CCL.
	thanks to Bill Jones (jones@cs.usask.ca)

Nov 19, 1991:
	use RAND_MAX instead of literal in builtin().

Nov 12, 1991:
	cranked up some fixed-size arrays in b.c, and added a test for
	overflow in penter.  thanks to mark larsen.

Sep 24, 1991:
	increased buffer in gsub.  a very crude fix to a general problem.
	and again on Sep 26.

Aug 18, 1991:
	enforce variable name syntax for commandline variables: has to
	start with letter or _.

Jul 27, 1991:
	allow newline after ; in for statements.

Jul 21, 1991:
	fixed so that in self-assignment like $1=$1, side effects
	like recomputing $0 take place.  (this is getting subtle.)

Jun 30, 1991:
	better test for detecting too-long output record.

Jun 2, 1991:
	better defense against very long printf strings.
	made break and continue illegal outside of loops.

May 13, 1991:
	removed extra arg on gettemp, tempfree.  minor error message rewording.

May 6, 1991:
	fixed silly bug in hex parsing in hexstr().
	removed an apparently unnecessary test in isnumber().
	warn about weird printf conversions.
	fixed unchecked array overwrite in relex().

	changed for (i in array) to access elements in sorted order.
	then unchanged it -- it really does run slower in too many cases.
	left the code in place, commented out.

Feb 10, 1991:
	check error status on all writes, to avoid banging on full disks.

Jan 28, 1991:
	awk -f - reads the program from stdin.

Jan 11, 1991:
	failed to set numeric state on $0 in cmd|getline context in run.c.

Nov 2, 1990:
	fixed sleazy test for integrality in getsval;  use modf.

Oct 29, 1990:
	fixed sleazy buggy code in lib.c that looked (incorrectly) for
	too long input lines.

Oct 14, 1990:
	fixed the bug on p. 198 in which it couldn't deduce that an
	argument was an array in some contexts.  replaced the error
	message in intest() by code that damn well makes it an array.

Oct 8, 1990:
	fixed horrible bug:  types and values were not preserved in
	some kinds of self-assignment. (in assign().)

Aug 24, 1990:
	changed NCHARS to 256 to handle 8-bit characters in strings
	presented to match(), etc.

Jun 26, 1990:
	changed struct rrow (awk.h) to use long instead of int for lval,
	since cfoll() stores a pointer in it.  now works better when int's
	are smaller than pointers!

May 6, 1990:
	AVA fixed the grammar so that ! is uniformly of the same precedence as
	unary + and -.  This renders illegal some constructs like !x=y, which
	now has to be parenthesized as !(x=y), and makes others work properly:
	!x+y is (!x)+y, and x!y is x !y, not two pattern-action statements.
	(These problems were pointed out by Bob Lenk of Posix.)

	Added \x to regular expressions (already in strings).
	Limited octal to octal digits; \8 and \9 are not octal.
	Centralized the code for parsing escapes in regular expressions.
	Added a bunch of tests to T.re and T.sub to verify some of this.

Feb 9, 1990:
	fixed null pointer dereference bug in main.c:  -F[nothing].  sigh.

	restored srand behavior:  it returns the current seed.

Jan 18, 1990:
	srand now returns previous seed value (0 to start).

Jan 5, 1990:
	fix potential problem in tran.c -- something was freed,
	then used in freesymtab.

Oct 18, 1989:
	another try to get the max number of open files set with
	relatively machine-independent code.

	small fix to input() in case of multiple reads after EOF.

Oct 11, 1989:
	FILENAME is now defined in the BEGIN block -- too many old
	programs broke.

	"-" means stdin in getline as well as on the commandline.

	added a bunch of casts to the code to tell the truth about
	char * vs. unsigned char *, a right royal pain.  added a
	setlocale call to the front of main, though probably no one
	has it usefully implemented yet.

Aug 24, 1989:
	removed redundant relational tests against nullnode if parse
	tree already had a relational at that point.

Aug 11, 1989:
	fixed bug:  commandline variable assignment has to look like
	var=something.  (consider the man page for =, in file =.1)

	changed number of arguments to functions to static arrays
	to avoid repeated malloc calls.

Aug 2, 1989:
	restored -F (space) separator

Jul 30, 1989:
	added -v x=1 y=2 ... for immediate commandline variable assignment;
	done before the BEGIN block for sure.  they have to precede the
	program if the program is on the commandline.
	Modified Aug 2 to require a separate -v for each assignment.

Jul 10, 1989:
	fixed ref-thru-zero bug in environment code in tran.c

Jun 23, 1989:
	add newline to usage message.

Jun 14, 1989:
	added some missing ansi printf conversion letters: %i %X %E %G.
	no sensible meaning for h or L, so they may not do what one expects.

	made %* conversions work.

	changed x^y so that if n is a positive integer, it's done
	by explicit multiplication, thus achieving maximum accuracy.
	(this should be done by pow() but it seems not to be locally.)
	done to x ^= y as well.

Jun 4, 1989:
	ENVIRON array contains environment: if shell variable V=thing,
		ENVIRON["V"] is "thing"

	multiple -f arguments permitted.  error reporting is naive.
	(they were permitted before, but only the last was used.)

	fixed a really stupid botch in the debugging macro dprintf

	fixed order of evaluation of commandline assignments to match
	what the book claims:  an argument of the form x=e is evaluated
	at the time it would have been opened if it were a filename (p 63).
	this invalidates the suggested answer to ex 4-1 (p 195).

	removed some code that permitted -F (space) fieldseparator,
	since it didn't quite work right anyway.  (restored aug 2)

Apr 27, 1989:
	Line number now accumulated correctly for comment lines.

Apr 26, 1989:
	Debugging output now includes a version date,
	if one compiles it into the source each time.

Apr 9, 1989:
	Changed grammar to prohibit constants as 3rd arg of sub and gsub;
	prevents class of overwriting-a-constant errors.  (Last one?)
	This invalidates the "banana" example on page 43 of the book.

	Added \a ("alert"), \v (vertical tab), \xhhh (hexadecimal),
	as in ANSI, for strings.  Rescinded the sloppiness that permitted
	non-octal digits in \ooo.  Warning:  not all compilers and libraries
	will be able to deal with \x correctly.

Jan 9, 1989:
	Fixed bug that caused tempcell list to contain a duplicate.
	The fix is kludgy.

Dec 17, 1988:
	Catches some more commandline errors in main.
	Removed redundant decl of modf in run.c (confuses some compilers).
	Warning:  there's no single declaration of malloc, etc., in awk.h
	that seems to satisfy all compilers.

Dec 7, 1988:
	Added a bit of code to error printing to avoid printing nulls.
	(Not clear that it actually would.)

Nov 27, 1988:
	With fear and trembling, modified the grammar to permit
	multiple pattern-action statements on one line without
	an explicit separator.  By definition, this capitulation
	to the ghost of ancient implementations remains undefined
	and thus subject to change without notice or apology.
	DO NOT COUNT ON IT.

Oct 30, 1988:
	Fixed bug in call() that failed to recover storage.

	A warning is now generated if there are more arguments
	in the call than in the definition (in lieu of fixing
	another storage leak).

Oct 20, 1988:
	Fixed %c:  if expr is numeric, use numeric value;
	otherwise print 1st char of string value.  still
	doesn't work if the value is 0 -- won't print \0.

	Added a few more checks for running out of malloc.

Oct 12, 1988:
	Fixed bug in call() that freed local arrays twice.

	Fixed to handle deletion of non-existent array right;
	complains about attempt to delete non-array element.

Sep 30, 1988:
	Now guarantees to evaluate all arguments of built-in
	functions, as in C;  the appearance is that arguments
	are evaluated before the function is called.  Places
	affected are sub (gsub was ok), substr, printf, and
	all the built-in arithmetic functions in bltin().
	A warning is generated if a bltin() is called with
	the wrong number of arguments.

	This requires changing makeprof on p167 of the book.

Aug 23, 1988:
	setting FILENAME in BEGIN caused core dump, apparently
	because it was freeing space not allocated by malloc.

July 24, 1988:
	fixed egregious error in toupper/tolower functions.
	still subject to rescinding, however.

July 2, 1988:
	flush stdout before opening file or pipe

July 2, 1988:
	performance bug in b.c/cgoto(): not freeing some sets of states.
	partial fix only right now, and the number of states increased
	to make it less obvious.

June 1, 1988:
	check error status on close

May 28, 1988:
	srand returns seed value it's using.
	see 1/18/90

May 22, 1988:
	Removed limit on depth of function calls.

May 10, 1988:
	Fixed lib.c to permit _ in commandline variable names.

Mar 25, 1988:
	main.c fixed to recognize -- as terminator of command-
	line options.  Illegal options flagged.
	Error reporting slightly cleaned up.

Dec 2, 1987:
	Newer C compilers apply a strict scope rule to extern
	declarations within functions.  Two extern declarations in
	lib.c and tran.c have been moved to obviate this problem.

Oct xx, 1987:
	Reluctantly added toupper and tolower functions.
	Subject to rescinding without notice.

Sep 17, 1987:
	Error-message printer had printf(s) instead of
	printf("%s",s);  got core dumps when the message
	included a %.

Sep 12, 1987:
	Very long printf strings caused core dump;
	fixed aprintf, asprintf, format to catch them.
	Can still get a core dump in printf itself.