Java software tends to come with its own mini file layout including
bin, lib and other folders. The lib folder typically holds jars,
which can conflict between packages if linked into HOMEBREW_PREFIX/lib
and aren't needed to compile other .dylib based software anyway.
The recommendation for Java software is to install to "libexec" and then
symlink or wrap binaries from libexec/bin to bin in the Cellar.
If the top-level formula being installed has any keg-only
dependencies, the lib/include/bin/pkg-config paths for those
deps will be added twice to the ENV vars.
This doesn't break anything, but does make debug output somewhat longer
than it needs to be.
Tidy this up by calling uniq on deps before iterating.
(The cause of this duplication is related to how top-level formula are
run, to support keeping patches in __END__ blocks.)
FixesHomebrew/homebrew#1110
If a package contains a folder that has the same name as one of our
expected meta files, skip trying to install that folder.
(Otherwise we install an empty folder and can get odd permission errors
when trying to summarize after install.)
When an "install -d formula" fails, and the user gets dropped into an
interactive shell, set the HOMEBREW_DEBUG_INSTALL env var to the name
for the formula that failed. Also set variable if the user requests an
interactive install in the first place.
Note that this may be different than the formula being installed, since
it may have been a dep that failed.
Also remove todo in utils; users can now look for HOMEBREW_DEBUG_INSTALL
in their prompt command, and adjust their prompts accordingly.
If an exception is thrown while parsing the formula file before the install.rb at_exit then it will be in $!. Marshal it back to the parent brew process.
Thus install.rb gets the path to the instantiating brew process rather than whichever is first in the PATH. Not to mention Homebrew doesn't *have* to be in the PATH.
Couldn't make this atomic, apologies.
Fixes a few things, like deps failing to build not aborting the install.
--force now works properly again.
Overall more robust code. I went back over it all and gave it a lot of
thought.
Cleaner separation of logic. Less code in brew, now the only code there is
ARGV handling, and basic sanity checks.
Not extending ARGV or ENV in global now as that would propagate to other tools
or utilities you may write.
This rewrite attempts to sort out where the Prefix, Cellar, and
Repository are relative to the real and symlinked 'brew' command.
Also included is a --config option which dumps all of these variables.
Any top-level script must define a "BREW_FILE" that gives the path
to brew as it exists in the path. 'brew' itself just uses __FILE__ and
install.rb does a `which brew` (there may be a better way?)
The Prefix is always relative to the location of brew as it exists in
the path. Thus, whether or not /usr/local/bin/brew is a symlink or real
file, the Prefix is always /usr/local. If you have brew in some other
prefix, such as /nonstandard/bin/brew, then '/nonstandard/ will be
managed by brew instead.
The Repository, Cellar, and "Library/Homebrew" required code is always
found relative to the "real" path or brew. If brew is a real file in
/usr/local/bin/brew, then everything else will be found in /usr/local
and we'll expect a /usr/local/.git
Otherwise, we dereference brew's symlink and look for everything else
relative to that path instead.
Using an error pipe.
The use of ENV[HOMEBREW_ERROR_PIPE] feels wrong, but I wasn't sure how else to
proxy the file descriptor to the child process since the fork immediately
calls exec.
The classes better reflect their contents. I'm sure this change may be
contentious, but I am a sucker for trying to create source bases that are easy
to get to grips with and easy to navigate.
brewkit.rb is now a deprecated file.
This works better as it comes after any keg-only deps change ENV, but also it needs to happen after brewkit is required and that only happens in install now.
brewkit.rb changes ENV destructively, so lets not do that everytime a formula
is required. Now it's possible for other tools to require a formula
description without worrying about side-effects.
Axel tries to accelerate downloads by using multiple connections (possibly to
multiple servers) for one download. Because of its size, it might be very
useful on bootdisks or other small systems as a wget replacement.
Eg gettext gets added into LDFLAGS, INCLUDE and that. I hope I got everything
that is typical. Prolly not. But we'll find out.
Made readline keg_only because the BSD version is provided by OS X, and I
don't want bug reports that are tricky to solve due to unexpected differences
between the two.
GNU GetText breaks eg. Ruby 1.9 builds, and some other formula I have been building too. But it is required by eg. glib. So to solve this we are going to by default not symlink gettext into the Homebrew prefix.
Formula that depend on GetText will have the gettext paths added to the brewing environment automatically. Neat.
Eg. sbin may be part of the formula, but that isn't in the default Mac PATH. Also will avoid bug reports for users who forget to amend their PATH and stick Homebrew somewhere different.
For this to work the "running script" must be the formulae file. Making this
so wasn't so hard, there is now an install.rb script which is included with
the -r flag to the ruby executable. An at_exit handler calls the install
function.
Having the install logic in its own file made it feel like there was so much
space that I added extra error handling. So there is something to be said for
separating functionality out into its own files.
Still the error handling sucks, we'll need to marshall the exception back to
the bin/brew command. Which is another PITA.
Still overall I think this will prove worthwhile. But if it doesn't we'll
revert.
As a first usage, you can put a diff after __END__ and return DATA from
Formula::patches to make Homebrew aware of it.
