New in alpha3.8:  Bug fix for signed/unsigned mixup, found and fixed
by the FreeBSD folks.

New in alpha3.7:  A bit of cleanup aimed at maximizing portability,
possibly at slight cost in efficiency.  "ul" suffixes and "unsigned long"
no longer appear, in particular.

New in alpha3.6:  A couple more portability glitches fixed.

New in alpha3.5:  Active development of this code has been stopped --
I'm working on a complete reimplementation -- but folks have found some
minor portability glitches and the like, hence this release to fix them.
One penalty:  slightly reduced compatibility with old compilers, because
the ANSI C `unsigned long' type and `ul' constant suffix are used in a
few places (I could avoid this but it would be considerably more work).

New in alpha3.4:  The complex bug alluded to below has been fixed (in a
slightly kludgey temporary way that may hurt efficiency a bit; this is
another "get it out the door for 4.4" release).  The tests at the end of
the tests file have accordingly been uncommented.  The primary sign of
the bug was that something like a?b matching ab matched b rather than ab.
(The bug was essentially specific to this exact situation, else it would
have shown up earlier.)

New in alpha3.3:  The definition of word boundaries has been altered
slightly, to more closely match the usual programming notion that "_"
is an alphabetic.  Stuff used for pre-ANSI systems is now in a subdir,
and the makefile no longer alludes to it in mysterious ways.  The
makefile has generally been cleaned up some.  Fixes have been made
(again!) so that the regression test will run without -DREDEBUG, at
the cost of weaker checking.  A workaround for a bug in some folks'
<assert.h> has been added.  And some more things have been added to
tests, including a couple right at the end which are commented out
because the code currently flunks them (complex bug; fix coming).
Plus the usual minor cleanup.

New in alpha3.2:  Assorted bits of cleanup and portability improvement
(the development base is now a BSDI system using GCC instead of an ancient
Sun system, and the newer compiler exposed some glitches).  Fix for a
serious bug that affected REs using many [] (including REG_ICASE REs
because of the way they are implemented), *sometimes*, depending on
memory-allocation patterns.  The header-file prototypes no longer name
the parameters, avoiding possible name conflicts.  The possibility that
some clot has defined CHAR_MIN as (say) `-128' instead of `(-128)' is
now handled gracefully.  "uchar" is no longer used as an internal type
name (too many people have the same idea).  Still the same old lousy
performance, alas.

New in alpha3.1:  Basically nothing, this release is just a bookkeeping
convenience.  Stay tuned.

New in alpha3.0:  Performance is no better, alas, but some fixes have been
made and some functionality has been added.  (This is basically the "get
it out the door in time for 4.4" release.)  One bug fix:  regfree() didn't
free the main internal structure (how embarrassing).  It is now possible
to put NULs in either the RE or the target string, using (resp.) a new
REG_PEND flag and the old REG_STARTEND flag.  The REG_NOSPEC flag to
regcomp() makes all characters ordinary, so you can match a literal
string easily (this will become more useful when performance improves!).
There are now primitives to match beginnings and ends of words, although
the syntax is disgusting and so is the implementation.  The REG_ATOI
debugging interface has changed a bit.  And there has been considerable
internal cleanup of various kinds.

New in alpha2.3:  Split change list out of README, and moved flags notes
into Makefile.  Macro-ized the name of regex(7) in regex(3), since it has
to change for 4.4BSD.  Cleanup work in engine.c, and some new regression
tests to catch tricky cases thereof.

New in alpha2.2:  Out-of-date manpages updated.  Regerror() acquires two
small extensions -- REG_ITOA and REG_ATOI -- which avoid debugging kludges
in my own test program and might be useful to others for similar purposes.
The regression test will now compile (and run) without REDEBUG.  The
BRE \$ bug is fixed.  Most uses of "uchar" are gone; it's all chars now.
Char/uchar parameters are now written int/unsigned, to avoid possible
portability problems with unpromoted parameters.  Some unsigned casts have
been introduced to minimize portability problems with shifting into sign
bits.

New in alpha2.1:  Lots of little stuff, cleanup and fixes.  The one big
thing is that regex.h is now generated, using mkh, rather than being
supplied in the distribution; due to circularities in dependencies,
you have to build regex.h explicitly by "make h".  The two known bugs
have been fixed (and the regression test now checks for them), as has a
problem with assertions not being suppressed in the absence of REDEBUG.
No performance work yet.

New in alpha2:  Backslash-anything is an ordinary character, not an
error (except, of course, for the handful of backslashed metacharacters
in BREs), which should reduce script breakage.  The regression test
checks *where* null strings are supposed to match, and has generally
been tightened up somewhat.  Small bug fixes in parameter passing (not
harmful, but technically errors) and some other areas.  Debugging
invoked by defining REDEBUG rather than not defining NDEBUG.

New in alpha+3:  full prototyping for internal routines, using a little
helper program, mkh, which extracts prototypes given in stylized comments.
More minor cleanup.  Buglet fix:  it's CHAR_BIT, not CHAR_BITS.  Simple
pre-screening of input when a literal string is known to be part of the
RE; this does wonders for performance.

New in alpha+2:  minor bits of cleanup.  Notably, the number "32" for the
word width isn't hardwired into regexec.c any more, the public header
file prototypes the functions if __STDC__ is defined, and some small typos
in the manpages have been fixed.

New in alpha+1:  improvements to the manual pages, and an important
extension, the REG_STARTEND option to regexec().