On Thu, May 26, 2016 at 10:36:44AM -0700, Junio C Hamano wrote: > Jeff King <peff@xxxxxxxx> writes: > > > The problem is running test-config inside of a git alias. The > > bin-wrappers will set the exec-path to the root-level of git's build > > directory, which the git binary will then stick at the front of the > > $PATH. > > I was wondering why exec-path does not point at bin-wrappers in the > first place. > > A wrapper script needs to know where the real thing lives in order > to "exec" (or "exec gdb") anyway, and it hardcodes the path to it. > It happens to use GIT_EXEC_PATH to shorten the hardcoded path it > uses when it does "exec", but it does not have to. > > Wouldn't we want to see "git" use any of these wrapped ones when it > invokes a non-builtin subcommand when it does so normally, honoring > GIT_EXEC_PATH? Pointing GIT_EXEC_PATH at the top-level means that > wrappers are bypassed for such an invocation (if what is run happens > to have executable at the top-level), and possibly a totally wrong > thing is run (when we start generating the binaries in different > directories, which is what we are seeing here). I think the issue is that bin-wrappers serves two purposes. One is for testing, but the other is for people who run git directly without installing. For us to set GIT_EXEC_PATH to bin-wrappers, it would have to have all of the git-* external programs, which would then put them all in the $PATH of people doing the no-install thing. That's certainly not insurmountable. Either we can tell them to live with it, or we can break out a separate wrapper directory that serves as a pseudo-exec-path. > > So the simplest solution really is: don't do that. The only debate > > in my mind is whether this is rare enough that it won't bite > > somebody again in the future, or if we should look into a solution > > that makes this Just Work. > > I think it was you who alluded to revamping the test framework along > the lines of preparing a "test installation" (aka "make install" > with DESTDIR set to somewhere) and making bin-wrappers to point into > that installation (or if we are testing an installed Git that may be > different from what we have source for, that final installed > location). An installed version of Git will not have test-* helpers > so they need to come from a freshly built source tree, not from > "test installation" or "installed Git". There may be other details > that need to be worked out, but as a longer term direction that may > not be a bad idea. I think you can make it even simpler by not really doing a "make install", but just linking or bin-wrappering a fake exec-path. It would be great if we could truly just "make install" into a fake area and test that (dropping bin-wrappers entirely), but git cares too much about some hard-coded paths, I think. We'd have to first have a truly relocatable binary. -Peff -- 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