inside of a sysroot targeting a system+sysroot which is "similar" or
"compatible" with the host system. This shows up when trying to build
system images on largely compatible hardware as-if fully cross compiled.
The problem is that previously we *perfectly* mimiced GCC here, and it
turns out GCC has a bug that no one has really stumbled across. GCC will
try to look in thy system prefix ('/usr/local' f.ex.) into which it is
instaled to find libraries installed along side GCC that should be
preferred to the base system libraries ('/usr' f.ex.). This seems not
unreasonable, but it has a very unfortunate consequence when combined
with a '--sysroot' which does *not* contain the GCC installation we're
using to complete the toolchain. That results in some of the host
system's library directories being searched during the link.
Now, it so happens that most folks doing stuff like this use
'--with-sysroot' and '--disable-multilib' when configuring GCC. Even
better, they're usually not cross-compiling to a target that is similar
to the host. As a result, searching the host for libraries doesn't
really matter -- most of the time weird directories get appended that
don't exist (no arm triple lib directory, etc). Even if you're
cross-compiling from 32-bit to 64-bit x86 or vice-versa, disabling
multilib makes it less likely that you'll actually find viable libraries
on the host. But that's just luck. We shouldn't rely on this, and this
patch disables looking in the system prefix containing the GCC
installation if that system prefix is *outside* of the sysroot. For
empty sysroots, this has no effect. Similarly, when using the GCC
*inside* of the sysroot, we still track wherever it is installed within
the sysroot and look there for libraries. But now we can use a cross
compiler GCC installation outside the system root, and only look for the
crtbegin.o in the GCC installation, and look for all the other libraries
inside the system root.
This should fix PR12478, allowing Clang to be used when building
a ChromiumOS image without polluting the image with libraries from the
host system.
llvm-svn: 154176
Patch from Michel Dänzer, sent our way via Jeremy Huddleston who added
64-bit support. I just added one other place where powerpc64-linux-gnu
was missing (we only had powerpc64-unknown-linux-gnu).
I've also added a tree to test out the debian multiarch stuff. I don't
use debian regularly, so I'm not certain this is entirely accurate. If
anyone wants to check it against a debian system and fix any
inaccuracies, fire away. This way at least folks can see how this is
*supposed* to be tested.
It'd be particularly good to get the Debian MIPS toolchains tested in
this way.
llvm-svn: 151482
version of Ubuntu. It has a very broken multiarch configuration, and so
we need special logic to handle it correctly. Fixing and testing this
uncovered a few other trivial issues with the logic that are fixed as
well.
I added tests to cover this as it is hard to notice if you install
recent versions of the OS.
llvm-svn: 144165
edge cases and have better behavior. Specifically, we should actually
prefer the general '4.6' version string over the '4.6.1' string, as
'4.6.2' should be able to replace it without breaking rpaths or any
other place that these paths have been embedded. Debian-based
distributions are already using a path structure with symlinks to
achieve in-place upgrades for patch versions. Now our parsing reflects
this and we select the shorter paths instead of the longer paths.
A separate issue was that we would not parse a leading patch version
number even in the presence of a suffix. The above change makes this
more problematic as it would cause a suffix being added to make us treat
the entire thing as patch-version-agnostic, which it isn't. This changes
the logic to distinguish between '4.4.x' and 4.4.1-x', and retain that
the latter has *some* patch number information. Currently, we always
bias toward the shorter and more canonical version strings. If it
becomes important we can add more Debian like rules to produce sequences
such as '4.4.1b' > '4.4.1' > '4.4.1-rc3' > '4.4.1-rc2' > '4.4.1-pre5',
but I'm very doubtful this will ever matter or be desirable.
I've made the tests for this logic a bit more interesting, and added
some specific tests for logic that is now different.
llvm-svn: 143841
this saga. Teach the driver to detect a GCC installed along side Clang
using the existing InstalledDir support in the Clang driver. This makes
a lot of Clang's behavior more automatic when it is installed along side
GCC.
Also include the first test cases (more to come, honest) which test both
the install directory behavior, and the version sorting behavior to show
that we're actually searching for the best candidate GCC installation
now.
llvm-svn: 141145
configuration, although the test still stubs out more directories than
are necessary or common in order to exercise all of the lookup paths
observed with upstream GCC.
This finishes testing the distribution-independent and
GCC-installation-independent parts of the library path search logic.
More testing is still needed for the triple detection, GCC-installation
detection, and handling distributions with unusual configurations.
llvm-svn: 141000
enabled for debian hosts, which is quite odd. I think all restriction on
when Clang attempts to use a multilib installation should go away. Clang
is fundamentally a cross compiler. It behaves more like GCC when built
as a cross compiler, and so it should just use multilib installs when
they are present on the system. However, there is a very specific
exemption for Exherbo, which I can't test on, so I'm leaving that in
place.
With this, check in a generic test tree for multilib on a 32-bit host.
This stubs out many directories that most distributions don't use but
that uptsream GCC supports. This is intended to be an agnostic test that
the driver behaves properly compared with the GCC driver it aims for
compatibility with.
Also, fix a bug in the driver that this testing exposed (see!) where it
was incorrectly testing the target architecture rather than the host
architecture.
If anyone is having trouble with the tree-structure stubs I'm creating
to test this, let me know and I can revisit the design. I chose this
over (for example) a tar-ball in order to make tests run faster at the
small, hopefully amortized VCS cost.
llvm-svn: 140999
include *any* path on crtbegin.o unless we actually find such a file via
one of the search paths. We still strictly check the search paths right
after this, so we'll catch any issues there.
The reason for this is that the driver does some normalization of the
path on the actual object file, and this changes the textual format of
the string on Windows. It no longer matches the textual format of the
sysroot flag.
llvm-svn: 140998
This requires fixing a latent bug -- if we used the default host triple
instead of an autodetected triple to locate GCC's installation, we
didn't go back and fix the GCC triple. Correct that with a pile of
hacks. This entire routine needs a major refactoring which I'm saving
for a subsequent commit. Essentially, the detection of the GCC triple
should be hoisted into the same routine as we locate the GCC
installation: the first is intrinsically tied to the latter. Then the
routine will just return the triple and base directory.
Also start to bring the rest of the library search path logic under
test, including locating crtbegin.o. Still need to test the multilib and
other behaviors, but there are also bugs in the way of that.
llvm-svn: 140995
This is still very much a WIP, but sysroot was completely broken before
this so we are moving closer to correctness.
The crux of this is that 'ld' (on Linux, the only place I'm touching
here) doesn't apply the sysroot to any flags given to it. Instead, the
driver must translate all the paths it adds to the link step with the
system root. This is easily observed by building a GCC that supports
sysroot, and checking its driver output.
This patch just fixes the non-multilib library search paths. We should
also use this in many other places, but first things first.
This also allows us to make the Linux 'ld' test independent of the host
system. This in turn will allow me to check in test tree configurations
based on various different distro's configuration. Again, WIP.
llvm-svn: 140990
precisely match the pattern and logic used by the GCC driver on Linux as
of a recent SVN checkout.
This happens to follow a *much* more principled approach. There is
a strict hierarchy of paths examined, first with multilib-suffixing,
second without such suffixing. Any and all of these directories which
exist will be added to the library search path when using GCC.
There were many places where Clang followed different paths, omitted
critical entries, and worst of all (in terms of challenges to debugging)
got the entries in a subtly wrong order.
If this breaks Clang on a distro you use, please let me know, and I'll
work with you to figure out what is needed to work on that distro. I've
checked the behavior of the latest release of Ubuntu, OpenSUSE, Fedora,
and Gentoo. I'll be testing it on those as well as Debian stable and
unstable and ArchLinux. I may even dig out a Slackware install.
No real regression tests yet, those will follow once I add enough
support for sysroot to simulate various distro layouts in the testsuite.
llvm-svn: 140981
Make the suffixes optional everywhere, and just make sure they have the
right value. The suffixes aren't the interesting part of this test
anyways.
Sorry for the churn as I let the bots try out various patterns.
llvm-svn: 140927
Evgeniy Stepanov