aboutsummaryrefslogtreecommitdiff
Commit message (Collapse)AuthorAgeFilesLines
* df,lsattr: support Unix-like mount pointsmount_pointRon Yorston2026-06-153-1/+24
| | | | | | | | | | | | | | Update 'lsattr' to distinguish between mount points and junctions. Use the new 'mntent' functions in 'df' to display mount points as well as drive letters. If mount point is associated with a drive letter this is displayed as the filesystem. The Windows-specific '-w' option forces the volume name to be displayed instead, if avaialable. Adds 87-111 bytes. Signed-off-by: Ron Yorston <rmy@pobox.com>
* win32: Unix-like support for volume mountsRon Yorston2026-06-155-98/+229
| | | | | | | | | | | | | | | | | | Add code to distinguish between volume mount points and junctions. Volume mount points are to be treated like directries while junctions are like symlinks. Update the 'mntent' functions to scan drive letters first, then volume names and their associated mount points. By default the filesystem is returned as a drive letter, if possible, though the volume name is also returned. The function 'find_mount_point()' has been updated to use the new 'mntent' routines. Adds 772-803 bytes. Signed-off-by: Ron Yorston <rmy@pobox.com>
* ash: scan drives directly for 'pwd -a'Ron Yorston2026-06-151-8/+6
| | | | | | | | | | | The shell builtin 'pwd' takes a Windows-specific '-a' option to display the current directory on each drive. Use Windows APIs directly to scan for valid drives instead of 'getmntent()'. Future changes will result in 'getmntent()' returning mounted volumes as well as drives. Signed-off-by: Ron Yorston <rmy@pobox.com>
* win32: extend handling of virtual disksRon Yorston2026-06-083-7/+20
| | | | | | | | | | | | | | | | | | | | | | | | | Commit 62e4c5d29 (win32: special treatment for virtual hard disk) added special treatment for mount points associated with virtual hard disks. One statement made there was incorrect: a VHD without a drive letter containing no files *does* have a '.' directory. The error message for this case was misleading. The changes made to handle VHDs were insufficiently general: - lstat(2) only resolved symlinks for the explicit path '.'. It should also do so for paths of the form 'path/to/.'. - opendir(3) only generated entries for '.' and '..' directories when FindFirstFileA() failed for the path '.'. It should also do so for any path which resolves to a volume mount point. Make the necessary changes. Adds 64-80 bytes. (GitHub issue #597) Signed-off-by: Ron Yorston <rmy@pobox.com>
* win32: special treatment for virtual hard diskRon Yorston2026-06-072-4/+14
| | | | | | | | | | | | | | | | | | | | Virtual hard disks (VHD) without a drive letter gave incorrect results in 'ls' and 'stat'. There were two problems: - If the disk is empty it doesn't have a '.' directory. - If it contains files the '.' directory is actually the mount point, which looks like a symbolic link and can't be opened as a directory. Modify our 'stat(2)' and 'opendir(3)' implementations to handle these quirks. Adds 78-80 bytes. (GitHub issue #597) Signed-off-by: Ron Yorston <rmy@pobox.com>
* win32: update current directory to target of symlinkRon Yorston2026-06-061-0/+4
| | | | | | | | | | | | | | | | If busybox.exe was started from a Windows command line in a directory which was the target of a symlink (or equivalent) 'ls' and 'stat' were unable to report the correct details of the current directory. Call 'chdir()' from the BusyBox main program to update our notion of our current directory. Adds 16 bytes. (GitHub issue #597) Signed-off-by: Ron Yorston <rmy@pobox.com>
* win32: avoid problems with forward slashes in path to cmd.exeRon Yorston2026-06-061-3/+5
| | | | | | | | | | | | | | | Running 'C:/Windows/System32/cmd.exe' from the shell was found to have the unxpected result of creating the directory '.exe'. This is due to the weirdness of cmd.exe. A similar problem with batch files was already being handled in `spawnveq()` by converting forward slashes to backslashes in the path to the executable file. Doing this unconditionally will fix the above case too. (GitHub issue #598) Signed-off-by: Ron Yorston <rmy@pobox.com>
* lsattr: clarify output for reparse/mount pointsRon Yorston2026-06-052-4/+4
| | | | | | | | | | | | | | | | | | The 'lsattr' applet identified three specific types of reparse points: junctions, symbolic links and app exec links. Any other type was just shown as a reparse point. A junction is actually a specific type of mount point; a volume mount is another. The description 'junction' has been changed to the more generic 'mount point'. 'lsattr' doesn't distinguish between volume mounts and junctions. Change the help text to say 'Unidentified reparse point' for reparse point which aren't a mount point, symlink or app exec link. Adds 29 bytes. Signed-off-by: Ron Yorston <rmy@pobox.com>
* win32: special treatment for mounted volumesRon Yorston2026-06-041-0/+9
| | | | | | | | | | | | | It was reported that a mounted volume looked like a broken symlink and that changing to the mount directory failed. Preserve the format of the volume name when resolving the link. Adds 84-96 bytes. (GitHub issue #597) Signed-off-by: Ron Yorston <rmy@pobox.com>
* ash: don't treat process substitutions as background jobsRon Yorston2026-05-311-3/+7
| | | | | | | | | | | | | | | | | Commit bda604a70 (ash: prevent leakage of process handles) added job tracking for process substitutions. This had the unwanted side-effect that if the user tried to exit from the shell after a command involving process substitution the shell reported that background jobs were present and refused to exit. Use the flag introduced in commit e6c716317 (ash: don't report completion of process substitution) to avoid this. Adds 16 bytes. (GitHub issue #587) Signed-off-by: Ron Yorston <rmy@pobox.com>
* stty: code shrinkRon Yorston2026-05-262-4/+12
| | | | | | | | | | The current implementation of 'stty' for Windows only uses a limited subset of the flags in 'struct termios'. Remove some unused features. Saves 16-40 bytes. Signed-off-by: Ron Yorston <rmy@pobox.com>
* ash: read built-in should respect stty -echoRon Yorston2026-05-241-1/+11
| | | | | | | | | | | | If echo has been disabled by the command 'stty -echo' the shell 'read' built-in should match its behaviour on Linux and not echo keyboard input. Adds 32-48 bytes. (GitHub issue #594) Signed-off-by: Ron Yorston <rmy@pobox.com>
* win32: fix incorrect change to tcgetattr()Ron Yorston2026-05-221-1/+1
| | | | | | | | | | | | | | Commit a560fdf23 (win32: better handling of console state) changed to using CONIN$/CONOUT$ to handle terminal modes. In tcgetattr() CONOUT$ was used instead of CONIN$, leading to incorrect mode bits being fetched. Saves 8-16 bytes. (GitHub issue #594) Signed-off-by: Ron Yorston <rmy@pobox.com>
* win32: better handling of console stateRon Yorston2026-05-222-25/+35
| | | | | | | | | | | | | | | | | | | When 'stty size' was run in a command substitution with stderr redirected or 'stty -echo' with stdin redirected, it failed to work properly. The redirections broke its connection to the terminal. To fix this, ioctl/tcsetattr/tcgetattr have been modified to use CONIN$ or CONOUT$ to access the screen buffer. Using CON isn't an adequate substitute in this case. Also, if ioctl(fd, TIOCGWINSZ) fails, set errno. Adds 112-116 bytes. (GitHub issue #594) Signed-off-by: Ron Yorston <rmy@pobox.com>
* win32: skip check for execute with tab-completionmbartlett212026-05-201-0/+8
| | | | | Tab completion as arguments no longer reads a bit of each file to detect whether it is executable.
* crond: fix reboot special timeRon Yorston2026-05-191-0/+31
| | | | | | | | | | | | | | | The 'reboot' special time didn't work because the file to check for a reboot couldn't be created. And in any case, the mechanism used on Linux wouldn't work: Windows has no /var/run directory which is cleared on reboot. The crond.reboot file is now placed in the cron directory. If it doesn't exist when crond starts, it's created but the reboot job (if any) isn't run. If the file exists its modification time is compared with the time of the current boot. If it's older update its time and run the reboot job. Adds 144-224 bytes.
* crond: drop '-S' optionRon Yorston2026-05-191-5/+22
| | | | | | | Since we don't use syslog on Windows there's no point in having an option to set it. Log to stderr by default instead. Saves 88-96 bytes.
* crond,crontab: executable-relative cron directoriesRon Yorston2026-05-192-4/+40
| | | | | | | | | | | | | | | | | Some users want a 'portable' busybox-w32 installation on removable media. To help with this, certain files can be positioned relative to the location of the executable. crond and crontab now look for var/spool/cron/crontabs relative to the binary, and use that as their working directory if it exists. If it doesn't, fall back to the hardcoded absolute path. The '-c' command line option can still be used to override the automatic choice. The '-c' option in crond now updates the cron directory variable as well as that for crontabs. Adds 128-144 bytes.
* win32: don't use built-in memmoveRon Yorston2026-05-191-1/+1
| | | | | | | GCC 16 has an optimised memmove which inflates 64-bit builds by 20KB. This shouldn't happen with -Os, but it did. It's been fixed in GCC, but until the new version trickles down to Fedora 44 explictly disabled the memmove built-in.
* Add markup to README.mdRon Yorston2026-05-181-1/+1
|
* Update README.mdRon Yorston2026-05-181-0/+1
| | | | | Add a note about excluding busybox.exe from monitoring by Windows security.
* win32: allow stat(2) to skip costly check for executeRon Yorston2026-05-184-20/+43
| | | | | | | | | | | | | | | | | If 'ls' is invoked without the '-l' option (or something that's equivalent) and colour output isn't required, there's no need to perform the expensive check for execute permission. Allow 'ls' to inform stat(2) not to call has_exec_format(). This has been combined with the existing mechanism to enable counting subdirectories. With this change the command 'ls | wc -l' used in issue #589 doesn't perform the test which upset Windows' anti-malware thing. Adds 32-48 bytes. (GitHub issue #589)
* ash: don't report completion of process substitutionRon Yorston2026-05-171-2/+20
| | | | | | | | | | | | | Commit bda604a70 (ash: prevent leakage of process handles) added job tracking for process substitutions. This had the unwanted side-effect that the completion of such processes was then reported in interactive shells. Set a flag in such jobs so their completion isn't reported. Adds 32 bytes. (GitHub issue #587)
* testsuite: skip tests for handling of non-printable charactersRon Yorston2026-05-171-1/+8
| | | | | | | | | | | Among the tests for 'ls' there are some for how it displays a filename containing non-printable characters. These fail on Windows because most of the characters used aren't allowed in filenames there. This can cause the test process to hang. Skip these tests on Windows. (GitHub commit #592)
* ash: prevent leakage of process handlesRon Yorston2026-05-131-4/+7
| | | | | | | | | | | | | | When the shell invoked process substitution it retained a process handle to the child. These handles could accumulate without limit. The shell in upstream BusyBox doesn't bother to create a job structure for process substitutions or here documents. This isn't appropriate on Windows, where we need to track the child process handles. Create job structures as required. Saves 32-48 bytes. (GitHub issue #587)
* dd: remove note about /dev/urandom and /dev/zeroRon Yorston2026-05-121-3/+0
| | | | | | | Now that /dev/urandom and /dev/zero can be used by all applets there's no need to mention that dd has support for them. Saves 32-56 bytes.
* reset: code shrink and clarificationRon Yorston2026-05-121-4/+3
| | | | | | | | | | | | The 'ESC]J' sequence in 'reset' is unnecessary. One or other of 'ESC]3J' and the explicit call to reset_screen() should be sufficient, but the Win 10/11 terminal requires the former while the Win 10 console requires the latter. Saves 4 bytes in the 32-bit build. (GitHub issue #161)
* reset: add support for stty and Windows TerminalRon Yorston2026-05-112-23/+22
| | | | | | | | | | | | | | | | | | The 'reset' applet hasn't kept pace with developments elsewhere. We now have support for 'stty sane', so the code in 'reset' which calls that can be enabled. Windows 10 and 11 have updated Terminal and Console programs, with varying features and behaviours. Add ANSI emulation code to handle clearing the scrollback buffer and use ANSI sequences in 'reset' to do that. It was also necessary to retain the explicit call to reset_screen(), or 'reset' didn't work properly in the Terminal I'd installed in Windows 10. Adds 92-128 bytes. (GitHub issue #161)
* libbb: reverse performance regression in bb_copyfd_*Ron Yorston2026-05-101-1/+4
| | | | | | | | | | | | | Commit c4f24ec9d (libbb: try to mitigate 'cat /dev/urandom') resulted in a 'cat' to the terminal taking ~25% longer than before. Raising the isatty() call out of the loop reduces the penalty to ~1%. Adds 16 bytes in the 32-bit build. (GitHub issue #585)
* libbb: try to mitigate 'cat /dev/urandom'Ron Yorston2026-05-081-0/+11
| | | | | | | | | | | | | | | | | | | | | | | Now that '/dev/urandom' can be used directly, people with a sense of curiosity and adventure have tried running 'cat /dev/urandom' or 'cat </dev/urandom'. The 'cat' applet in BusyBox is so efficient and ANSI escape handling is so inefficient that this overwhelms the capabilities of the Windows console or terminal to the extent that it's unable to process Ctrl-C requests in a timely manner. To give it a chance to catch up, force 'cat' to take a short nap from time to time (very short, zero length) but only when it's writing to a tty. The call to Sleep() is in a separate function to avoid unnecessary bloat in 32-bit builds, where its presence upsets the stack and requires much larger code for stack access. Adds 48 bytes. (GitHub issue #585)
* Update README.mdRon Yorston2026-05-071-1/+1
|
* ash: further fixes to merge of upstream changesFRP-6075-g169694ebdRon Yorston2026-05-061-24/+60
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Commits e23652908 and 686a0803f (ash: fix execution of applets via Unix-style path) were necessary following a major revision of the shell upstream. Another problem was found: $ sh $ PATH="/usr/bin;$PATH" exec non-existent resulted in a segfault. This happened because during a PATH search tryexec() modified argv[0] when it detected a Unix-style path (in this case '/usr/bin/non-existent') but failed to restore it if the execution failed. There were other issues: - The logic of the new code failed to match the original: the test for a Unix-style path should only have happened if an attempt to execute the full path had already failed. - The test for a script running an interpreter which is an applet also modified argv. If the execution failed argv should have been restored to its original state. - During a PATH search the 'path' pointer walked through the elements of the path. This pointer was also used to determine if an applet was overridden by an executable. This is wrong: the full PATH variable should have been used. The code around tryexec()/shellexec() has been rewritten to take these issues into account. Adds 32-48 bytes. (GitHub issue #584)
* ash: fix build failure if SH_STANDALONE is disabledRon Yorston2026-05-051-0/+4
| | | | | | | | It wasn't possible to build ash if FEATURE_SH_STANDALONE was disabled. This issue was introduced during the large merge of upstream changes to ash in commit e23652908.
* win32: silence warning in ioctl(2)Ron Yorston2026-05-051-0/+2
| | | | | Silence a compiler warning that 'arg' is an unused variable if neither stty nor ttysize is defined.
* win32: add FAST_FUNC annotationsRon Yorston2026-05-0423-238/+250
| | | | | | | | | I've been far too lax in applying FAST_FUNC annotations. These can make function calls smaller and faster, but only on 32-bit systems and for non-static, non-variadic functions with non-void arguments. Saves 4680 bytes in the 32-bit build; 64-bit builds are unaffected.
* crond: build tweaksRon Yorston2026-05-038-42/+42
| | | | | | | | | | | | | | | Using 'depends on FEATURE_SYSLOG' is unreliable because FEATURE_SYSLOG can't be set directly. A POSIX build with all applets that need syslog disabled *except* crond may fail. Revert to using 'select FEATURE_SYSLOG' in crond and exclude the unused syslog code on Windows. If crond is unable to run sendmail report this in the error message. Update all default Windows configurations. (GitHub PR #561)
* Add crond and crontab (#561)mbartlett212026-05-035-21/+227
| | | | | | | | | | | | | | | | | | | | | * Enable crontab * Enable crond * Allow sendmail to be used in crond This also adds another option to the mingw_popen_fd function to capture stderr as well. And also added the options to support WNOHANG as well as an unused status in mingw_wait3 * Check current uid when running crond * Only show tail for mailto jobs * Use system drive for cron * crond: add missing feature check * crond: use spawn_detach for background operation
* win32: disable urandom/zero devices if dd is disabledRon Yorston2026-04-272-2/+12
| | | | | | | | | Support for /dev/urandom and /dev/zero relies on having a working dd. If the dd applet is disabled attempts to access them should fail. Theoretically a third-party dd utility could also work, but let's not encourage that sort of heresy.
* win32: improve urandom/zero device handlingRon Yorston2026-04-263-14/+14
| | | | | | Rearrange the code to avoid a superfluous process creation. Saves 16 bytes in the 64-bit build, adds 24 in the 32-bit.
* win32: expose /dev/urandom and /dev/zeroRon Yorston2026-04-261-8/+23
| | | | | | | | | | | | | | | Previously the special devices /dev/urandom and /dev/zero could only be used via the 'dd' applet. They were also used internally by some applets. Expose /dev/urandom and /dev/zero so they can be used as file arguments to applets and a source of redirections in the shell. They're also now visible to 'stat' and 'ls', but only if given as an explicit argument. Adds 122-144 bytes. (GitHub issues #98, #282)
* tls: fix double free of HMAC contextRon Yorston2026-04-251-4/+1
| | | | | | | | | | | | | | | | | Commit aed19625ff (Post-merge fixes) attempted to handle some upstream code shuffling which broke support for https in wget with FEATURE_USE_CNG_API enabled and CONFIG_FEATURE_TLS_SCHANNEL disabled (i.e. using the upstream internal TLS code). Unfortunately it resulted in the HMAC context being freed twice which caused failures in certain cases. Fix this by not freeing the context in hmac_blocks(), thus making that function match upstream again. Thanks to @avih for identifying the problem. (GitHub issue #582)
* wget: allow HTTPS certificate check to be skippedRon Yorston2026-04-244-2/+36
| | | | | | | | | | | The Microsoft Windows Schannel implementation of TLS validates the server certificate. Enable the --no-check-certificate option to wget to allow these checks to be skipped. This may be useful to connect to badly configured websites. Adds 202 bytes to the x86_64 build with Schannel enabled. (GitHub issue #581)
* Win32: trivial refinements (#577)avih2026-04-184-11/+13
| | | | | | | | | | | | | | | | | * win32: fnmatch2 minor refinements (no-op) With beacket, add '\' to the switch/case, and at fnmatch, pre-calc the two most used flags (CASEFOLD and PATHNAME) once on init. Negligible perf impact, and happens to save few bytes in x64. * win32: actype: trivial refinments (no-op) NULL was used in actype.h but requires stddef.h - which we don't include, so use (int*)0 which also expresses it more clearly. Remove "#include <string.h>" at actype.c, which we don't need anymore. * scripts/patbench.sh: make executable (chmod +x)
* patch: improved handling of files with no final newlineRon Yorston2026-04-162-9/+94
| | | | | | | | | | | | | Commit 2275a53f0 (patch: handle files with no final newline) was incomplete. It only detected a line with no newline in the first section of a hunk. Add the code necessary to detect the '\ No newline at end of file' warning in the second section. Also add some tests. Adds 48-80 bytes. (GitHub issue #575)
* win32: fnmatch2: replace if-else-... with switch/caseAvi Halachmi (:avih)2026-04-161-17/+28
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This was done to be more concise, but it also happens to speed up the two slowest test patterns by about ~ 30%, and without measurable impact on any of the other test patterns (nor the correctness). Maybe the stars aligned for compiler optimizations. We'll take it. Here are results of scripts/patbench.sh with recent iterations of fnmatch, from most recent to the original old fnmatch: CURRENT results: [ 1] 107 ms OK (M:-----) '' [ 2] 105 ms OK (M:12345) '*' [ 3] 106 ms OK (M:-----) '@' [ 4] 111 ms OK (M:1----) 'Lorem*' [ 5] 122 ms OK (M:-2---) '*quis' [ 6] 133 ms OK (M:--3--) '*nisi*' [ 7] 229 ms OK (M:--3--) '*[fobar]xyz*' [ 8] 451 ms OK (M:--3--) '*[foobarfoobarfoobarfoobar]xyz*' [ 9] 463 ms OK (M:-2---) '*[!foobarfoobarfoobarfoobar]xyz*' [10] 136 ms OK (M:12345) '*[hello world, this is a test]*' [11] 126 ms OK (M:12345) '*[!hello world, this is a test]*' [12] 169 ms OK (M:1-3--) '*[*' [13] 175 ms OK (M:1----) '*[]*' [14] 171 ms OK (M:--3--) '*[!]*' [15] 162 ms OK (M:-23-5) '*[!]].*' [16] 243 ms OK (M:1----) '*[![:print:]]*' [17] 135 ms OK (M:-23--) '*z*' [18] 131 ms OK (M:---4-) '*-*' [19] 157 ms OK (M:---4-) '*[-]*' [20] 165 ms OK (M:-234-) '*[-z]*' [21] 176 ms OK (M:-234-) '*[z-]*' [22] 182 ms OK (M:-234-) '*[-z-]*' [23] 130 ms OK (M:1-3--) '*]*' [24] 153 ms OK (M:1-3--) '*[]]*' [25] 164 ms OK (M:1-34-) '*[]-]*' [26] 167 ms OK (M:123--) '*[]-_]*' [27] 168 ms OK (M:1234-) '*[]-_-]*' [28] 122 ms OK (M:12-45) '*[0-9A-Za-z][0-9A-Za-z]' [29] 126 ms OK (M:12-45) '*[[:alnum:]][[:alnum:]]' [30] 118 ms OK (M:12-45) '*[a-z][a-z]' [31] 127 ms OK (M:12-45) '*[a-bc-de-z][a-bc-de-z]' [32] 131 ms OK (M:12-45) '*[[:lower:]][[:lower:]]' [33] 136 ms OK (M:-----) '*[0-9A-Fa-ef][0-9A-Fa-ef]' [34] 124 ms OK (M:-----) '*[[:xdigit:]][[:xdigit:]]' [35] 128 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]' [36] 129 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]' [37] 127 ms OK (M:12-45) '*[a-z]*[a-z]' [38] 123 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]' [39] 124 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]' [40] 136 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [41] 134 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]' [42] 136 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]*[0-9A-Za-z]' [43] 131 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]*[[:alnum:]]' [44] 124 ms OK (M:12-45) '*[a-z]*[a-z]*[a-z]' [45] 137 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]*[a-bc-de-z]' [46] 129 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]*[[:lower:]]' [47] 160 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [48] 164 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]*[[:xdigit:]]' [49] 152 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]*[0-9A-Za-z]*[0-9A-Za-z]' [50] 152 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]*[[:alnum:]]*[[:alnum:]]' [51] 130 ms OK (M:12-45) '*[a-z]*[a-z]*[a-z]*[a-z]' [52] 165 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]*[a-bc-de-z]*[a-bc-de-z]' [53] 146 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]*[[:lower:]]*[[:lower:]]' [54] 180 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [55] 189 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]*[[:xdigit:]]*[[:xdigit:]]' Previous commit (if-else-...) is mainly slower in patterns 8 and 9: [ 6] 143 ms OK (M:--3--) '*nisi*' [ 7] 251 ms OK (M:--3--) '*[fobar]xyz*' [ 8] 638 ms OK (M:--3--) '*[foobarfoobarfoobarfoobar]xyz*' [ 9] 651 ms OK (M:-2---) '*[!foobarfoobarfoobarfoobar]xyz*' [10] 138 ms OK (M:12345) '*[hello world, this is a test]*' [11] 137 ms OK (M:12345) '*[!hello world, this is a test]*' Initial fnmatch2 commit (without any optimizations): [ 1] 101 ms OK (M:-----) '' [ 2] 123 ms OK (M:12345) '*' [ 3] 105 ms OK (M:-----) '@' [ 4] 115 ms OK (M:1----) 'Lorem*' [ 5] 134 ms OK (M:-2---) '*quis' [ 6] 131 ms OK (M:--3--) '*nisi*' [ 7] 235 ms OK (M:--3--) '*[fobar]xyz*' [ 8] 586 ms OK (M:--3--) '*[foobarfoobarfoobarfoobar]xyz*' [ 9] 613 ms OK (M:-2---) '*[!foobarfoobarfoobarfoobar]xyz*' [10] 153 ms OK (M:12345) '*[hello world, this is a test]*' [11] 160 ms OK (M:12345) '*[!hello world, this is a test]*' [12] 164 ms OK (M:1-3--) '*[*' [13] 165 ms OK (M:1----) '*[]*' [14] 157 ms OK (M:--3--) '*[!]*' [15] 157 ms OK (M:-23-5) '*[!]].*' [16] 244 ms OK (M:1----) '*[![:print:]]*' [17] 132 ms OK (M:-23--) '*z*' [18] 134 ms OK (M:---4-) '*-*' [19] 160 ms OK (M:---4-) '*[-]*' [20] 157 ms OK (M:-234-) '*[-z]*' [21] 168 ms OK (M:-234-) '*[z-]*' [22] 166 ms OK (M:-234-) '*[-z-]*' [23] 130 ms OK (M:1-3--) '*]*' [24] 150 ms OK (M:1-3--) '*[]]*' [25] 151 ms OK (M:1-34-) '*[]-]*' [26] 157 ms OK (M:123--) '*[]-_]*' [27] 156 ms OK (M:1234-) '*[]-_-]*' [28] 402 ms OK (M:12-45) '*[0-9A-Za-z][0-9A-Za-z]' [29] 416 ms OK (M:12-45) '*[[:alnum:]][[:alnum:]]' [30] 265 ms OK (M:12-45) '*[a-z][a-z]' [31] 410 ms OK (M:12-45) '*[a-bc-de-z][a-bc-de-z]' [32] 426 ms OK (M:12-45) '*[[:lower:]][[:lower:]]' [33] 347 ms OK (M:-----) '*[0-9A-Fa-ef][0-9A-Fa-ef]' [34] 370 ms OK (M:-----) '*[[:xdigit:]][[:xdigit:]]' [35] 294 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]' [36] 301 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]' [37] 214 ms OK (M:12-45) '*[a-z]*[a-z]' [38] 304 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]' [39] 302 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]' [40] 306 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [41] 291 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]' [42] 291 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]*[0-9A-Za-z]' [43] 302 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]*[[:alnum:]]' [44] 208 ms OK (M:12-45) '*[a-z]*[a-z]*[a-z]' [45] 314 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]*[a-bc-de-z]' [46] 288 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]*[[:lower:]]' [47] 315 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [48] 340 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]*[[:xdigit:]]' [49] 309 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]*[0-9A-Za-z]*[0-9A-Za-z]' [50] 304 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]*[[:alnum:]]*[[:alnum:]]' [51] 217 ms OK (M:12-45) '*[a-z]*[a-z]*[a-z]*[a-z]' [52] 317 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]*[a-bc-de-z]*[a-bc-de-z]' [53] 322 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]*[[:lower:]]*[[:lower:]]' [54] 315 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [55] 348 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]*[[:xdigit:]]*[[:xdigit:]]' Old fnmatch with actype/isactype in O(1): [ 1] 103 ms OK (M:-----) '' [ 2] 116 ms OK (M:12345) '*' [ 3] 108 ms OK (M:-----) '@' [ 4] 112 ms OK (M:1----) 'Lorem*' [ 5] 122 ms OK (M:-2---) '*quis' [ 6] 123 ms OK (M:--3--) '*nisi*' [ 7] 270 ms OK (M:--3--) '*[fobar]xyz*' [ 8] 650 ms OK (M:--3--) '*[foobarfoobarfoobarfoobar]xyz*' [ 9] 681 ms OK (M:-2---) '*[!foobarfoobarfoobarfoobar]xyz*' [10] 133 ms OK (M:12345) '*[hello world, this is a test]*' [11] 135 ms OK (M:12345) '*[!hello world, this is a test]*' [12] 191 ms E:1-3-- (M:-----) '*[*' [13] 213 ms E:1---- (M:-----) '*[]*' [14] 215 ms E:--3-- (M:-----) '*[!]*' [15] 182 ms OK (M:-23-5) '*[!]].*' [16] 296 ms OK (M:1----) '*[![:print:]]*' [17] 121 ms OK (M:-23--) '*z*' [18] 115 ms OK (M:---4-) '*-*' [19] 167 ms OK (M:---4-) '*[-]*' [20] 189 ms OK (M:-234-) '*[-z]*' [21] 194 ms OK (M:-234-) '*[z-]*' [22] 203 ms OK (M:-234-) '*[-z-]*' [23] 122 ms OK (M:1-3--) '*]*' [24] 166 ms OK (M:1-3--) '*[]]*' [25] 185 ms OK (M:1-34-) '*[]-]*' [26] 163 ms OK (M:123--) '*[]-_]*' [27] 176 ms OK (M:1234-) '*[]-_-]*' [28] 381 ms OK (M:12-45) '*[0-9A-Za-z][0-9A-Za-z]' [29] 510 ms OK (M:12-45) '*[[:alnum:]][[:alnum:]]' [30] 274 ms OK (M:12-45) '*[a-z][a-z]' [31] 379 ms OK (M:12-45) '*[a-bc-de-z][a-bc-de-z]' [32] 488 ms OK (M:12-45) '*[[:lower:]][[:lower:]]' [33] 340 ms OK (M:-----) '*[0-9A-Fa-ef][0-9A-Fa-ef]' [34] 398 ms OK (M:-----) '*[[:xdigit:]][[:xdigit:]]' [35] 1402 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]' [36] 1962 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]' [37] 939 ms OK (M:12-45) '*[a-z]*[a-z]' [38] 1453 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]' [39] 1965 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]' [40] 1683 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [41] 2117 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]' (aborted manually due to exponential slowness) Old fnmatch (before actype/isactype): [ 1] 107 ms OK (M:-----) '' [ 2] 108 ms OK (M:12345) '*' [ 3] 104 ms OK (M:-----) '@' [ 4] 112 ms OK (M:1----) 'Lorem*' [ 5] 120 ms OK (M:-2---) '*quis' [ 6] 127 ms OK (M:--3--) '*nisi*' [ 7] 244 ms OK (M:--3--) '*[fobar]xyz*' [ 8] 544 ms OK (M:--3--) '*[foobarfoobarfoobarfoobar]xyz*' [ 9] 588 ms OK (M:-2---) '*[!foobarfoobarfoobarfoobar]xyz*' [10] 129 ms OK (M:12345) '*[hello world, this is a test]*' [11] 136 ms OK (M:12345) '*[!hello world, this is a test]*' [12] 192 ms E:1-3-- (M:-----) '*[*' [13] 201 ms E:1---- (M:-----) '*[]*' [14] 203 ms E:--3-- (M:-----) '*[!]*' [15] 182 ms OK (M:-23-5) '*[!]].*' [16] 454 ms OK (M:1----) '*[![:print:]]*' [17] 124 ms OK (M:-23--) '*z*' [18] 123 ms OK (M:---4-) '*-*' [19] 160 ms OK (M:---4-) '*[-]*' [20] 175 ms OK (M:-234-) '*[-z]*' [21] 177 ms OK (M:-234-) '*[z-]*' [22] 197 ms OK (M:-234-) '*[-z-]*' [23] 126 ms OK (M:1-3--) '*]*' [24] 168 ms OK (M:1-3--) '*[]]*' [25] 173 ms OK (M:1-34-) '*[]-]*' [26] 168 ms OK (M:123--) '*[]-_]*' [27] 170 ms OK (M:1234-) '*[]-_-]*' [28] 365 ms OK (M:12-45) '*[0-9A-Za-z][0-9A-Za-z]' [29] 513 ms OK (M:12-45) '*[[:alnum:]][[:alnum:]]' [30] 263 ms OK (M:12-45) '*[a-z][a-z]' [31] 376 ms OK (M:12-45) '*[a-bc-de-z][a-bc-de-z]' [32] 815 ms OK (M:12-45) '*[[:lower:]][[:lower:]]' [33] 311 ms OK (M:-----) '*[0-9A-Fa-ef][0-9A-Fa-ef]' [34] 756 ms OK (M:-----) '*[[:xdigit:]][[:xdigit:]]' [35] 1305 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]' [36] 1928 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]' [37] 919 ms OK (M:12-45) '*[a-z]*[a-z]' [38] 1446 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]' [39] 3192 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]' [40] 1586 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [41] 4601 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]' [42] 22491 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]*[0-9A-Za-z]' (aborted manually due to exponential slowness)
* win32: fnmatch2.c: optimize last '*'Avi Halachmi (:avih)2026-04-161-1/+87
| | | | | | | | This adds two fairly cheap and effective optimizations at the tail of the pattern (after the last '*'). Speeds up many cases by ~ x2+. Details at the comment inside. Adds 192 bytes in x64.
* win32: switch to new fnmatch2.cAvi Halachmi (:avih)2026-04-163-0/+233
| | | | | | | | | | | | | | | | | | | | | | | | | win32/fnmatch.c has several issues: - *[* doesn't match foo[bar - but it should (non-special '['). - FMN_CASEFOLD is not handled corrrectly. It does tolower on pattern and string, which doesn't always work, e.g. *[@-[]* correctly matches '@', A-Z, and '[', but as case-insensitive it only matches '@' and '[' (tolower on the string breaks all the alpha matches): set -o nocaseglob; echo *[@-[]* set +o nocaseglob; echo *[@-[]* - According to posix, negative bracket range, like [z-a], should be either invalid or empty, but instead it matches only 'z'. - There could be more issues hiding - the code is not easy to follow. - It's exponential time in the number of '*' in pattern. This commit adds win32/fnmatch2.c and disables win32/fnmatch.c (currently using "#if 0"), but still keeps it in the tree. The new implementation is written from scratch, non-recursive and linear time, improves POSIX compliance, and hopefully more readable and fixable if needed. Also, it's about half the size at the binary (saves 908 bytes in x64). See implementation details and choices as comment at fnmatch2.c .
* win32: actype: remove size-optimization option (no-op)Avi Halachmi (:avih)2026-04-161-23/+9
| | | | | | It was already disabled (ASH_OPTIMIZE_FOR_SIZE is 0 by default), and it saved ~ 40 bytes but with big hit on performance, so not really worth it. This also makes actype.c less noisy.
* win32: actype: code shrinkAvi Halachmi (:avih)2026-04-162-18/+13
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Unify actype/actail into a single function, and make actype a wrapper macro which invokes actail(str, NULL) (to avoid calling a function actype which then trampolines to actail with NULL). This concludes the integration of actype/isactype. Overall: - Unified/standard-ish char-class handling in fnmatch/regcomp/tr. - Saved about 1400 bytes at the binary as x64. - regcomp.c is negligibly faster (actype/isactype are O(1)). - tr.c now also supports [:graph:] and [:print:], and not slower. fnmatch.c (tested using scripts/patbench.sh): Before actype was integrated, alnum was fastest and xdigit slowest due to their order in the names strings list and in the switch/case: [35] 1305 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]' [36] 1928 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]' [38] 1446 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]' [39] 3192 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]' [40] 1586 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [41] 4601 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]' Now all classes are similar relative to an equivalent range, and just barely slower than "alnum" before actype was added (still slower than range due to additional temp name buffer): [35] 1402 ms OK (M:12-45) '*[0-9A-Za-z]*[0-9A-Za-z]' [36] 1962 ms OK (M:12-45) '*[[:alnum:]]*[[:alnum:]]' [38] 1453 ms OK (M:12-45) '*[a-bc-de-z]*[a-bc-de-z]' [39] 1965 ms OK (M:12-45) '*[[:lower:]]*[[:lower:]]' [40] 1683 ms OK (M:---4-) '*[0-9A-Fa-ef]*[0-9A-Fa-ef]' [41] 2117 ms OK (M:---4-) '*[[:xdigit:]]*[[:xdigit:]]'
* win32: actype/isactype: optimize both to O(1)Avi Halachmi (:avih)2026-04-162-40/+154
| | | | | | | | | | | | | | | | | | | | | The premise, and implementation, is simple: instead of doing string search in a list of strings for the class name - which is fast for the first strings in the list but slow for the last strings, use perfect hash to map the string directly to a potential class to match in O(1). The switch/case in isactype is also changed to a function table in O(1) (the compiler might have optimized this switch to O(1) jump-table, but now we don't leave it to chance). Two implementations are supported: one which allows testing actype_t value against specific classes like AC_ALNUM etc, and one which doesn't allow that (just like wctype/iswctype) and is more efficient. Since nothing uses specific class matches anymore, default is the more efficient choice - where actype_t is a function pointer to isalnum etc. And while at it, also change the enum names CCLASS_ALNUM etc to more appropriate AC_ALNUM etc, to use the AC "namespace".
generated by cgit v1.2.3-55-g6feb (git 2.39.1) at 2026-08-27 18:47:11 +0000