Return to CHANGES CVS log

Up to [DVB] / dietlibc

import <scsi/*> from glibc.

0.10:
  update getpagesize.
  add personality.
  add "%*s" and "%*.*s" support to vsnprintf.
  add putchar (apparently arm-linux-gcc will substitute printf("\n")
    with putchar('\n') behind your back...?!
  import include/scsi/ from glibc.

0.9:
  fixed initgroups (oops)
  Cleanup appending file open for stdio (thanks, James Antill)
  imported vfprintf fix from James Antill.
  add fnmatch
  change index, rindex and bcmp as weak ELF symbols (were #defined before)
  Olaf added character classes to *scanf.
  Put object and library files into a separate directory (bin-$ARCH).
  Add "cross" make target to make all architectures.
  Add "diet" program and man page.
  Add getgrnam, getgrgid
  Add cfmakeraw (Thanks, Antonio Musumeci)
  localtime won't call gettimeofday when the tzfile parser is used.

0.8:
  vsnprintf(0,...) estimated the length of strings with a padding width
    incorrectly.
  new reboot() from Olaf.
  added prototypes for getsockopt and setsockopt.
  added experimental sem* and shm*
  define EXIT_SUCCESS and EXIT_FAILURE
  added setvbuf (only changes mode, not buffer)
  added dietwarning.h (woohoo, linker warnings!) and used it to warn of
    insecurity (system, sprintf) or bloat (printf and stdio).
  added qsort() and bsearch().  Benchmark on 1000 random ints:
    glibc: 679414 cycles qsort and 3352 cycles bsearch
    diet libc: 300010 cycles qsort, 553 cycles bsearch
    I have no idea why glibc is so slow.
  added __isnan (references by libm.a from glibc).
  added popen() and pclose().  Now the diet libc can link slsh (from
    S-Lang!)
  corrected usleep.  It was more of an msleep until now.
  corrected __dtostr (it looped when trying to format "0.0").
  several fixes to stdio (thanks to S-Lang's excellent test suite!)
  added ungetc (does it work?)
  added inet_aton (yuck!)
  added gethostbyname, gethostbyaddr, gethostbyname_r and gethostbyaddr_r
  added h_errno and __h_errno_location
  added inet_ntoa
  added setpgrp and getpgrp, strsignal and execv (thanks, Olaf)
  add bcopy, index, rindex macros
  add isgraph and isxdigit
  add strftime (ugh!!!)
  Olaf fixed strftime, gmtime_r leap year bug.
  Jeff Garzik contributed a spec file.  I put it in contrib and hope
    everbody knows what to do with it...
  Olaf checked in a ton of libpthread stuff.
  added mkstemp using /dev/urandom (harder to guess than getpid()).
  add large file backwards compatibility support (so gzip will run on
    kernel 2.2 and 2.0)
  added shutdown
  added nice
  added pseudo-locale support for is* (they are now weak aliases,
    isupper -> __isupper_ascii etc).  So, for example, liblatin1 can
    overwrite the default is* functions.
  added liblatin1.a
  added tzfile implementation (does it work for anyone else?)
  added syslog(3) (thanks to Rene Müller!)
  Jeff Garzik contributed getgr*, getpw* and getsp* in one tiny package.
    Thanks!
  merged duplicate code in execv and execvp into exec_lib.o

0.7.2:
  added several prototypes to already implemented functions and
    <sys/reboot.h>.  Thanks for Abraham for pointing this out.
  added prototypes for process group and dup syscalls.  Thanks to
    Abraham again.  He also contributed an implementation for daemon(),
    an ugly BSD hack.  I decided to diversify the lib/ directory a
    little and now created a libugly/ directory for stuff like system()
    and daemon() ;-)
  added libstdio and moved all that stdio crap from lib/ there.  The
    idea is to start separating now what we will separate for shared
    libraries anyway.
  optimized x86 unified-syscall by Olaf.
  added limits.h, removed __ARCHBITS from endian.h (use __WORDSIZE from
    limits.h instead)
  added remove.c and enough cruft to limits.h to make it compile lzo,
    lzop and bzip2.
  added mktime, asctime_r, asctime and ctime (needed for gzip).
  Someone will need to implement a reader for the glibc localtime file
    format some day (man tzfile, ugh!) :-(
  added unlink prototype to unistd.h.
  added gmtime and gmtime_r.
  fixed sparc unified syscall error handling (please upgrade!).
  removed readdir glibc compatibility cruft.
  removed stat glibc compatibility cruft.
  oops, my fread was broken for pipes.
  Olaf found a bug in pipe() for sparc, too.

0.7.1:
  optimized away the static arrays from __ltostr and __lltostr.
  optimized away the static arrays (and more!) from strtoul and strtoull.
  added system (by Olaf).
  added tolower and toupper
  added getopt from Olaf (~600 bytes on x86!)
  added clone for mips (by Olaf).
  fixed longjmp and setjmp on sparc not to require glibc include files.
  added pause for all (by Olaf).
  added __attribute__ ((__const__)) to ctype.h
  added isupper and islower.
  added putc and putchar as macros.
  added strcasecmp and strncasecmp (thanks to Abraham vd Merwe for
    insightful discussions and patches I partially used for the last three)
  added getpriority and setpriority (also by Abraham)
  oops, I broke strtoul in a recent optimization attempt.
  oi, setjmp and longjmp were broken on other platforms, too!
    Thanks to Sébastien Côté for reporting this.

0.7:
  clone support for i386,alpha,arm and sparc by Olaf.
  Initial dynamic loading glue for i386, also by Olaf.
  We don't have a dynamic loader yet, so this is not yet usable.
  added readv, sigdelset, sigaltstack, sigfillset, sigismember, usleep,
    vsprintf, writev, all by Olaf.
  imported readdir fix from Guillaume.  Mhh, can the kernel interface
    really be this broken?
  Found a kludgy way to detect whether the kernel includes define struct
    stat64 (they also define STAT64_HAS_BROKEN_ST_INO).  If this does
    not work for you, please tell me!
  Removed bogus __seek_types enum, it's a bunch of #defines now.
  Fixed stdio line buffering.  The code was already there but I forgot
    to mark stdio and stdout as line buffered.  This should really be
    done dynamically using isatty, though.
  Removed superfluous "if (1)" in execvp.c

0.6.14:
  *scanf did not append the 0 byte for %s.
  added creat and changed creat to call open directly instead of open64.
  did for fread what 0.6.10 did for fwrite.
  repaired buffered stdio a little (line buffering is still not
    supported).
  [All of these bugs have been reported by Guillaume.  Thanks!]
  added __pure__ to strchr and strrchr in <string.h>

0.6.13:
  included COPYING to make clear that the diet libc is covered by the
    GNU General Public License (and _not_ the LGPL).  That means that
    you need to obtain a license from me if you want to use the diet
    libc in a proprietary program that you want to distribute.
  The x86 unified syscall swapped arguments #4 and #5 and noone
    noticed until now!  Thanks to Guillaume Cottenceau for reporting
    this!
  The malloc had a subtle bug with small allocations that could cause
    segfault.  Guillaume reported it, Olaf fixed it.

0.6.12:
  added assert, statfs and fstatfs
  added include/net/if.h so fget compiles again.

0.6.11:
  added putenv.

0.6.10:
  fixed fwrite.  It returned the number of bytes written, not the number
    of records.  Thanks to Albert D. Cahalan for pointing this out.
  fixed fgets.  It returned EOF on empty lines.  Thanks, Erik Frey.
  added brk and sbrk.  harold@nb.com.sg made me do it.

0.6.9:
  fixed execvp not to return on ENOENT
  did for {open|read|close|seek|tell}dir what I did for stat before.
    The glibc compatibility part is currently non-functional.
  fixed white space handling in sscanf " (" would not match " (".
  fixed printf to accept 'l' flag.
  fixed getpwuid (ignored last line)

0.6.8:
  Olaf contributed initial thread-safe syscalls for i386, alpha, sparc,
    mips and arm.  I did the one for ppc.  They can be enabled in dietfeatures.h.
  added strlcpy and strlcat from OpenBSD.
  added stat64, fstat64 and lstat64.
  added endian.h and made ending of strlen.c endianness-aware
  added string routines size tweak to dietfeatures.h.  Most of the
    string routines have been unrolled.  This is up to three times as
    fast but creates up to three times larger code.  Now you can disable
    the unrolling.
  added sendfile.
  I also added pread.  It works on x86 and sparc but not on ppc and mips.
    I have no idea what's going on.  strace disagrees with the kernel.
    Can anyone help?  I asked the Linux kernel mailing list for help, too.
  added a few aliases of the type __libc_open for open.  nm on
    libpthreads.so indicates that we will need them.

0.6.7:
  the sources now compile without warnings with -Wall.
  printf now prints "(null)" when %s is passed NULL.  This can be
    removed with WANT_NULL_PRINTF in dietfeatures.h
  added vfprintf, execl, ttyname.
  stat, lstat and fstat are now #defined to __dietstat, __dietlstat and
    __dietfstat so they can use the normal kernel struct stat.
    stat, lstat and fstat are now C wrappers that convert to the glibc
    struct stat.  So we can avoid that overhead for programs that use
    dietlibc headers.

0.6.6:
  changed the Makefiles so you can now set CFLAGS on the command line
  added memccmp (analogous to memccpy) and strncmp.
  Olaf fixed another bug in the sparc unified syscall.
  Paul Clifford contributed a C version of his strlen.S that is much
    more efficient than the previous strlen on all platforms!  It also
    looks like technology from Roswell. ;-}
  added a "real" stdio.  Well, almost.  Please contribute!
  added dietfeatures.h so you can remove features you don't need.
  removed debug code from vsnprintf that changed \0 to ' ' (argh!).
  stdio now works with simple programs and minigzip from libz.
  stdio uses some major trickery to avoid linking the stdio, stderr and
    stdout (and reserving space for them) when they are not used.

0.6.5:
  Red Plait <redplait@ixcelerator.com> found several bugs in diet libc,
    most of them bugs in the new header files, but also missing sigset
    functions.
  Paul Clifford contributed an assembly strlen.S for ARM and fixed
    several bugs.
  Fixed ppc/setjmp and mips/pipe.
  Olaf contributed a new sparc unified syscall.

0.6.4:
  declared is* static inline in getservent.
  added assert_fail, strtoul, isalpha, isdigit, isalnum, isascii.
  changed strlen to return 0 when passed NULL.
  new, much smaller unified syscall for MIPS.  Thanks to Olaf the Mad
    Scientist who actually implemented this without access to a MIPS
    box, just from reading the architecture manual.  And it worked out
    of the box.
  started a set of system includes, mainly so that I can use lcc and
    my alpha-linux cross compiler (which is unable to cross-compile
    glibc).  diet libc can now be compiled without any include files
    from a normal libc.  The includes do declare more than diet libc
    currently delivers and they are still far from complete for real
    applications.
  I will only add assembly versions that are smaller _and_ faster than
    the C version.  Larger routines are only accepted if they are called
    very often and are substantially faster.
  added i386 assembly strchr, which is smaller and faster than the old
    version.  It is, however, larger than the version contributed by
    proton (thanks, anyway).
  added i386 strlen (31% faster, 14% smaller)
  "make t" will now create a map file called "mapfile".

0.6.3:
  added sys_errlist, strerror and perror
  added isblank
  added atol

0.6.2:
  mmap for ARM didn't compile.  Thanks, Paul!

0.6.1:
  split mmap into the architecture specific subdirectories.
  getenv now copes with environ==NULL, thanks Paul Clifford.
  Paul also contributed a smaller ARM startup code.

0.6:
  strcat returned the wrong result.  Thanks, Dietz Pröpper.
  strtod now understands a negative exponent (oops, thanks Bertram Barth)
  Port to arm-linux-gnu, but on the Netwinder I use for testing the
    __dtostr does not work (I have no idea why!)
  The MIPS port now uses (much smaller) non-PIC code.  That means that
    the applications you link against diet libc also have to be compiled
    non-PIC.  I suggest copying the CFLAGS from the diet libc Makefile.
    Thanks to Ralf Bächle for helping me with this!
  I made subdirectories for the architectures and use VPATH to override
    VPATH so that make finds the source file automatically.  That should
    simplify the sources greatly.
  Olaf Dreesen contributed Alpha support including setjmp and longjmp!
  ARM and MIPS now also have setjmp and longjmp
  "compile" and "load" are now make targets.  Use them for djb programs.
  added contrib/elftrunc.c which will remove unnecessary ELF headers.
    Again, contributed from Olaf.  Great work!

0.5.12:
  printf also does signed numbers.
  If you don't use atexit, dietlibc now does not link exit, only _exit.
  *printf now correctly returns the number of bytes written.
  Olaf contributed experimental sscanf and vsscanf implementations.
  If passed NULL as buffer, snprintf will not write anything but still
    return the number of bytes it would have written.
  Initial MIPS port!   (Oh, the agony!)
    No setjmp and longjmp support yet!  Please contribute!
  I even implemented unified syscalls for MIPS.  Still, MIPS code is
    almost twice the size of SPARC code.  If anyone knows why: please
    tell me!

0.5.11:
  I implemented new unified syscalls for x86, sparc and ppc, this time based
    on .s files and not .c files, and I moved the syscalls into
    subdirectores to clean the dietlibc sources up a little.
  A binary consisting of printf("%s is %d\n","olaf",23) is now
    2864 bytes on sparc
    2488 bytes on intel
  I kludgily implemented fprintf(stdout,... and fprintf(stderr,...
    to make a few more applications work.
  Olaf Dreesen also implemented some unified syscall stuff for x86.
  He discovered that with our unified syscall interface it actually
    costs <10 bytes total to make all system calls thread safe!
  I implemented atexit() (can register up to 4 callbacks).
  I implemented strtod and __dtostr (the opposite).  Now we can add
    floating point support to vsnprintf and sscanf!

0.5.10:
  I actually saw that I can not only merge the errno handling code of
    the system calls on x86, I can also merge the rest (including the
    arguments) except for setting the system call number.  All those
    system calls are now just a jump to a unified system call handler.
    I got rid of x86openclose again.  The savings are substantial:
    chown from embutils went from 7664 to 7184 bytes!
    If I'd move the system call wrappers to assembly language, I could
    even reduce the alignment (does not matter since it's just a jump
    anyway) and get rid of the "ret" behind each jump (one byte per
    system call!)

0.5.9:
  added memchr, strpbrk, strstr, strtol, isspace (hehe)
  fixed strdup
  fixed return values for strcpy and strcat (thanks to Norbert Berzen)
  Olaf Dreesen contributed a strtol and initial {sn|vsn|}printf implementation
    (no signed integers, only strings and unsigned integers (but
    supporting octal, hex and decimal).  Thanks, Olaf!  (by the way:
    that code must be wonderful, I don't understand it at all *bg*)
    Olaf's printf does understand padding like in "%08d" and "%8d" and
    automatically pads pointers with '0'.
  Rewrote the x86 start code, old: 22 instructions, new: 12 instructions.
  Thanks to proton for inspiration on this.
  The new x86 start code also does not reference exit any more, thus saving
  64 bytes for executables that don't call exit explicitly.
    [insert maniacal laughter] EVERY BYTE COUNTS!1!! ;-)
  Moved errno and environ to start.S, saving no byte binary size but
    speeding up compilation and slightly shortening dietlibc.a ;-)
  For x86: joined open and close into one assembler file, sharing the
    errno handling.  This is a feasibility test and it actually saves a
    few bytes.  I think I will reimplement the _syscall[1-6] macros on
    all platforms now to share the errno handling code for them.

0.5.8:
  fixed strchr to be able to look for 0.

  added _llseek, ftruncate, getpgid, getresgid, getresuid, getsid,
  memccpy, memmove, mprotect, setregid, setresgid, setresuid, setreuid,
  strncpy, swapon, truncate, strtok, strtok_r, strspn, strcspn (all
  contributed by Olaf Dreesen)

  added execvp, getcwd

  fixed __xmknod

  Note: can it be that the ftw interface really is so broken that I
  cannot implement it without having to implement some searching data
  structure?  I included an experimental ftw implementation that will
  not follow symlinks.

0.5.7:
  oops, if_nametoindex was broken!

0.5.6:
  ported to ppc-linux.
  fixed i386 sigsetjmp (I mistyped the function name)
  included sigjmp.c

0.5.5:
  added wait, sys_siglist, longjmp/setjmp/sigsetjmp for i386 and sparc
  actually, it wasn't fork that wasn't working for sparc, it was pipe.
  I fixed it now.
  dietlibc/SPARC assumes -msupersparc (does not provide div, mul, etc)
  fixed readdir to use getdents and not the intel inline asm

0.5.4:
  ported to sparclinux (sparc32 only).  Beware: does not work yet.
  added raise, abort, readlink, strcat, geteuid, geteuid, wait3, access

0.5.3:
  added tcsetattr and getenv to compile e3

0.5.2:
  added vhangup, tcgetattr, isatty and memcmp for fgetty
    (http://www.fefe.de/fgetty/)
  added localtime from uC-libc.

0.5.1:
  stat and friends actually work now.  Yuck!  Another case of
    translation between kernel and userland.  Why can't the kernel
    people and the libc people simply agree on a standard?
  removed many unnecessary includes to speed up compilation.
  added -fomit-frame-pointer and i386 compilation target to reduce code size.
  inlined socketcall to reduce code size.
  "load" and "compile" are examples for djb code.

0.5:
  Olaf Dreesen contributed a much smaller implementation of malloc and friends.
  each object file is now treated with "strip -x -R .note -R .comment"
  split each system call into a separate object file
  added ntohs, htons, alarm, if_indextoname and if_nametoindex

0.4.1:
  oops, the strchr implementation was wrong.  Thanks Jens Laas!

0.4:
  copied opendir and friends from uC-libc.
  added getservent and getservby* in gerservent.c

0.3:
  remove readdir system call
  added getdents system call
  added getpwnam/getpwuid implementation in getpwnam.c
  added a few string functions (in str*.c)