Hi, On Fri, 31 Aug 2007, Marius Storm-Olsen wrote: > We really need to streamline the patches needed for MinGW, and get those > merged into upstream when 1.5.3 is released. Okay, with 1.5.3 just about to break free, I think we can think about what to do to flood the mailing list with MinGW compatibility patches. So I went through the diff between mingw.git and git.git (I did not use 4msysgit.git as a reference yet). There are many, many changes that are not really dependent on each other, so many, many people can participate ;-) (And of course, development on mingw.git and/or 4msysgit.git will not stall through that time.) AFAICS these are the topics that need to be whipped into shape as individual patch series: - NO_ST_BLOCKS (struct stat has no "blocks" member on MinGW) - handle_line() in fmt-merge-msg.c should handle CRLF, too. Maybe this does not need to be special cased for MinGW at all. - we have a special handling for relative template paths in MinGW. I am not sure if we need it any more. - struct lock_file needs an "fd" member (assigned in lock_file()) so it can close the fd before renaming in remove_lock_file(), commit_lock_file() and commit_locked_index(), since Windows does not allow renaming of a still open file. This does not need to be special cased for MinGW IMHO. - is_absolute_path() needs to be aware of <DRIVE>:/. We should probably do this for Windows in general, not just MinGW (as pointed out by Robin in a recent mail, Cygwin suffers from this, too) - For some reason, we use "/usr/bin/sort" instead of "sort" in check-builtins.h, generate-cmdlist.sh, git-ls-remote.sh, git-mergetool.sh and some tests. It should be checked if this is still necessary (AFAIR it was some Vista fsckup) - A similar issue is /usr/bin/find in git-clone.sh, git-ls-remote.sh, git-repack.sh, and some tests. - fnmatch.[ch] were added to compat/. IIRC these were "borrowed" from GNU libc. - A lot of functions were added to compat/mingw.c; someone should check which are still needed, and which still need implementation (some functions just return an error) - regex.[ch] were added to compat/. These were "borrowed" from GNU regex. - We have git_etc_config() instead of ETC_GITCONFIG. Probably this should be changed to "#define ETC_GITCONFIG git_etc_config()" for MinGW only. (AFAIR this is a workaround for setenv() on MinGW refreshing _all_ environment variables, invalidating the previous getenv()ed addresses. - Hannes implemented a spawn-pipe.h interface, to replace the fork() in git_connect(), spawn_prog(), get_pack(). This is probably the most difficult part to merge into git.git. I am not clear about how much of this should be moved into run-command.h instead, for example. - We have a cpio "emulator" script, since MinGW lacks this tool. - gettimeofday() in compat/mingw.c needs my_mktime() from date.c to be public. I did not form an opinion yet if compat/mingw.c should be introduced in one go, or in incremental steps, implementing one function at a time. - In exec_cmd.c, we introduced a _function_ builtin_exec_path() to allow determining the path from the executable name. This is needed when calling the git programs from cmd.exe, _without_ going through sh first. IMHO this is wrong, since those programs can no longer call shell scripts, as needed e.g. in hooks, merge drivers or git-gc. So we probably should get rid of this part (but I imagine before this, there will be a fierce discussion about that), and force going through /bin/sh. - Many places do not call is_absolute_path() (because it was not there earlier). These have to be replaced. (This patch is orthogonal to MinGW.) - We have a spawnv_git_cmd(), which is probably a duplication of the run-command.h API. We should get rid of that function, and adjust the existing API if necessary. - For some reason, support for side-band-64k and side-band is commented out for MinGW. Find out the reasoning, and if it still holds true, introduce NO_SIDE_BAND into the Makefile and fetch-pack.c. - In git-clone, we call "pwd -W" (which is utterly unportable...). Find out if it is still necessary. - In git-compat-util.h, quite a bunch of header files have to be guarded between #ifndef __MINGW32__ ... #endif. - A bunch of #defines are in git-compat-util.h, which would have been defined in the headers that do not exist on MinGW. These should go into an #else clause, and it should be documented which header would have had those definitions. - There are a couple of workarounds in git-compat-util.h, like a mkdir() replacement which takes an "int mode" parameter, so that we do not need a dozen special cases scattered throughout the source code (MinGW's mkdir() does not take a "mode" parameter, since MSCRT's mkdir() doesn't) - We have no "sync" on MinGW, and in git-repack.sh, this is worked around by "sync 2> /dev/null". This should be changed to something like test ! -z "@@NO_SYNC@@" || sync - In prepend_to_path(), we need to replace the PATH delimiter ":" by ";" - In run_command(), we return "status & 0xff", because Windows' version of exit() apparently replaces all values outside of 0-0xff with 0! We probably need this in handle_internal_command(), too. - In handle_internal_command(), we need to strip the extension ".exe". This is done by defining STRIP_EXTENSION to ".exe" in the Makefile. - In git.c's main(), we check for "\\" in addition to "/". This is probably not the only place needing this... (For example, in setup.c we roll an own version of it, limited to setup.c, which is bad.) - In gitk, we still delete ~/.gitk everytime. Check if this is still necessary. - help.c #includes <sys/ioctl.h>. Why? - We have NO_ETC_PASSWD, but this is a misnomer. It would be more appropriately named NEEDS_GECOS, and there should be a replacement function "getpwuid()" in compat/mingw.c, which calls NetUserGetInfo() and fills a (custom) struct passwd. The changes in ident.c and path.c should be undone. - We have special pager handling in pager.c - In path.c, we fall back to TMP or TEMP when TMPDIR is not set. Since I propose to always run in a shell, I am torn if we should keep this code. - For some reason or other, we need some changes in perl/Makefile. - We have special progress handling in progress.c - In refs.c, we catch an erroneous errno (open can return EACCES when EISDIR would be appropriate). IMHO this should go into git-compat-util.h, where the #define'd open() should be replaced by a git_open() thing a la git_mkdir(). - In prefix_path(), we replace all "\\" to "/" in the path name. - In prefix_path(), there is an #ifndef __MINGW32__, where the #else clause uses is_absolute_path(). IMHO this is wrong, do not special case that for MinGW. - In setup_git_directory_gently(), there is special code for absolute paths of the form <DRIVE>:/bla. IMHO this should be changed so that such paths are converted to /<DRIVE>/bla first. - In sha1_file.c, we have yet another is_path_absolute() implementation. Use is_absolute_path() here instead. - In sha1_file.c, some fcntl()s are guarded (IIRC since MinGW does not have fcntl() at all). Make this a NO_FCNTL Makefile variable. - Plenty of tests have to be adjusted for the fact that there are no symlinks in MinGW. This probably benefits from the recent addition of a test for symlinks in git-init... (Needs adjusting, though, e.g. in test-lib.sh). - t0001 uses "pwd -W". - t1002 implements a "sum" function if it was not found (this is a patch which is independent of MinGW). - t1003, t4109 and t4110 implement cmp via git diff. Maybe change them over to git diff? - t1301 skips some tests. Look into that. - t3100 uses "diff" instead of "git diff". Why? - t3402m t3404, t4116 and t6200 sleep 1 second. Probably because of some file date granularity thing. - t3404 replaces HEAD^ by HEAD~1, and something weird seems to happen with "@" in the command line, when passed to scripts. - t3800 replaces one egrep with grep, because egrep is not available. IMHO egrep should not be used in git anyway. - t3901 skips iconv tests (in msysgit, we recently added iconv, so this might be worth fixing in MinGW). - t3902 tests if funny filenames are allowed on the current filesystem anyway. If not, the tests are skipped. - t5300 works around the lack of /dev/zero. It also splits a big test case into nice mouthfuls. - t5301 works around the lack of /dev/random. - t5302 just skips a few tests without a comment. Find out why. - t5502 works around a missing "git remote" (probably because we did not have perl scripts running for quite some time). Probably unnecessary. - t5515 escapes curly brackets, but since git-tag is no longer a shell script, this should not be necessary. - t7500 is skipped. Find out why "vi is not cooperating". - t7501 uses an unportable "sed -i". (Should be fixed in git.git anyway.) - t7501 also skips tests without explanation. Fix. - t9001 says it cannot invoke fake.sendmail. Fix. - We have a NOEXECTEMPL to make up for the lack of chmod a+x in MinGW (scripts are executable if and only if the first line starts with a #!). Rename this to NOEXEC_SUFFIX. - upload-pack.c works around the only fork() that does not exec() in git.git, by defining a function and spawning a thread. We should get rid of all fork() calls outside run-command.c, and have something like BEGINTHREAD=_beginthread in the Makefile. This is a little involved. - maybe_flush_or_die() special cases MinGW for returning EINVAL from fflush() sometimes. Either change this to _always_ check for EINVAL, or do the git_fflush() trick in git-compat-util.h to replace errno with EPIPE when it was set to EINVAL by fflush(). - imap-send is not built. - convert-objects still is not built. I think it is that date function which we could replace by some call to date.c. Phew. Ciao, Dscho - To unsubscribe from this list: send the line "unsubscribe git" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
