Throw the switch and move all Linux header search over to the GCC
detection system that is providing the library paths and crt object
files.
This, modulo any bugs that need to be shaken out, resolves numerous bugs
with how we handle header paths. Here are a few that I know of:
- We no longer need to enumerate all GCC versions searched.
- OpenSUSE searched GCC versions in the wrong order.
- There were typos when selecting various patterns, etc.
- We aren't stating quite some many directories now.
- SysRoot didn't always work in a reasonable way.
I'm working on tests for this, but the tests are making me and Lit sad.
The real testing for this type of driver change is to try it out on
various distributions. I'll hit the common ones right away, and start
more thorough testing tomorrow after some sleep.
Lift the GCCVersion type into the header file and start persisting it in
the detected GCC installation. This allows us to expose another aspect
of what we detected: the GCC version. This will be used shortly.
John McCall [Sun, 6 Nov 2011 09:01:30 +0000 (09:01 +0000)]
Change the AST representation of operations on Objective-C
property references to use a new PseudoObjectExpr
expression which pairs a syntactic form of the expression
with a set of semantic expressions implementing it.
This should significantly reduce the complexity required
elsewhere in the compiler to deal with these kinds of
expressions (e.g. IR generation's special l-value kind,
the static analyzer's Message abstraction), at the lower
cost of specifically dealing with the odd AST structure
of these expressions. It should also greatly simplify
efforts to implement similar language features in the
future, most notably Managed C++'s properties and indexed
properties.
Most of the effort here is in dealing with the various
clients of the AST. I've gone ahead and simplified the
ObjC rewriter's use of properties; other clients, like
IR-gen and the static analyzer, have all the old
complexity *and* all the new complexity, at least
temporarily. Many thanks to Ted for writing and advising
on the necessary changes to the static analyzer.
I've xfailed a small diagnostics regression in the static
analyzer at Ted's request.
Take a better approach to detecting and selecting multiarch include
directories. This way we stop at the first multiarch directory found on
the system. This achieves the real intended result of pruning
non-existent directories.
Revert r143864, as the problem I was actually looking at can be solved
a better way. The more I think about it the more worried I am that this
hammer is simply too large. We should only be reaching out to the
filesystem when doing interesting "detection" things, not gratuitously.
Original commit message:
Start pruning down the set of flags passed to CC1 for header search.
This cleans up the CC1 invocations, and reduces the overhead there.
We're still hammering the filesystem looking for the C++ standard
libraries though.
The only reservation I have about this policy is the case of virtualized
files inside of CC1, but it's not clear what the best way to solve that
is. The Driver consistently queries the actual filesystem to make its
decisions. Changing that would be a very large undertaking. It might be
worthwhile, but it's not an immediate goal.
Start pruning down the set of flags passed to CC1 for header search.
This cleans up the CC1 invocations, and reduces the overhead there.
We're still hammering the filesystem looking for the C++ standard
libraries though.
The only reservation I have about this policy is the case of virtualized
files inside of CC1, but it's not clear what the best way to solve that
is. The Driver consistently queries the actual filesystem to make its
decisions. Changing that would be a very large undertaking. It might be
worthwhile, but it's not an immediate goal.
This test was assuming that /usr/include was in the system header search
path. That assumption should never have been true, but it was until
I fixed it. Now that its fixed, add a triple here to get correct
behavior even on Windows.
Switch Lit to directly query the driver for the builtin inclue path.
Thanks to Peter for pointing out how easy this is to do. I'm now much
happier with this solution.
Enhance the GCC version parsing and comparison logic to handle some more
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.
Remove a pointless member. I have no idea why I made this not a local
variable to begin with... As I'm planning to add include root
information to this object, this would have caused confusion. It didn't
even *actually* hold the include root by the time we were done with it.
Move the GCC installation detection helper a member of the Linux
toolchain instead of merely using it in the constructor. This will allow
us to query it when building include paths as well as the file search
paths built in the constructor. I've lifted as little of it as I could
into the header file.
Eventually this will likely sink down into some of the Generic
toolchains and be used on more platforms, but I'm starting on Linux so
I can work out all the APIs needed there, where it is easiest to test
and we have the most pressing need.
Remove support for Gentoo subversion-ebuild installed llvm-gcc libstdc++
headers. As llvm-gcc is dead, and I have no idea if this ever really
worked, I think it's time for it to go. More importantly, it makes it
harder to generalize the include search logic. If someone really wants
these to work, they can set the CPLUS_INCLUDE_PATH environment variable.
Teach lit to ask the Clang it is running what version string to use
rather than presuming that it is 3.0. This is extra important as the
version should be 3.1, but CMake hasn't caught up with the times.
That'll be fixed in a separate commit.
Move the Linux header searching from the Frontend to the Driver. This is
the first (and diff-noisiest) step to making Linux header searching
tremendously more principled and less brittle. Note that this step
should have essentially no functional impact. We still search the exact
same set of paths in the exact same order. The only change here is where
the code implementing such a search lives.
This has one obvious negative impact -- we now pass a ludicrous number
of flags to the CC1 layer. That should go away as I re-base this logic
on the logic to detect a GCC installation. I want to do this in two
phases so the bots can tell me if this step alone breaks something, and
so that the diffs of the refactoring make more sense.
Change this test to reflect the state we are moving in. The Clang
builtin headers are no longer going to receive the old 'implicit extern
"C" block' semantics. This hint is actually ignored by both Clang and
GCC at this point, and Clang's own builtin headers can simply be changed
if there is any issue with this. Clang should be free to include these
however it wants, and so shorter and simpler is better.
Note: *nothing* is changing about the *system* stddef.h include. That
should always have the exact same include semantics, whether with Clang
or GCC or any other compiler. Only the compiler-builtin header search
path is changing.
If anyone knows of some risk that this introduces that I've not thought
of, please chime in. So far, only Windows has switched to the Brave New
World, but others should be switching soon.
Switch these two tests to use the Clang driver instead of CC1. They want
to do "realistic" includes, and so need the header search logic now in
the driver. This in turn requires switching the CC1 options to the
actual driver options, and passing -Xclang where there is no analogy.
Teach Lit to pass the CC1 invocation the builtin include directory. This
is a pretty gross hack, but I don't have any significantly cleaner ideas
for this. There are several things obviously gross about it:
1) Lit shouldn't know that Clang needs this. This really that bad, as
Lit already knows about CC1 and other internal details.
2) This hard codes the '3.0' version number, which is pretty lame.
3) This hard codes every other aspect of the resource dir structure
which is less lame than the version number, but still not great.
However, it should bring the MSVC tests back to life, and it should
unblock the rest of the move from Frontend to Driver, so I think it's
worth a bit of grossness that is isolated in our testing infrastructure
while we figure out the best long term approach. I have the following
ideas, some of which only solve part of the problem (and thus might need
to be combined with other ideas):
a) Create a symlink or other convenience path instead of a version
number.
b) Run 'clang' directly in the lit.cfg, look at its resource dir, and use
that.
c) Switch all the tests to use the driver instead of CC1.
d) Hack the frontend to synthesize builtin include directories when none
are provided by the driver.
I don't like (d) because it feels very hackish and likely to break. We
can only solve a small part of the problem with (a). I wanted to vote
for (c), but lots of the tests in this bucket are really heavily using
internal-only flags like -verify and -triple. I'm loath to complicate
them with the full driver layer. Also, switching them to the driver adds
more than just builtin headers, but all of the rest of the system
headers!
This leaves me with (b). If others like (b), I'll switch to it, but it
felt a bit icky. Nothing concrete, and the other options look
significantly worse, but I felt icky enough that I wanted to start with
a more brain-dead patch to stop the bleeding, and gauge others' feelings
here.
Fix a significant oversight in my move of MSVC includes to the driver:
actually manage the builtin header file includes as well as the system
ones.
This one is actually debatable whether it belongs in the driver or not,
as the builtin includes are really an internal bit of implementation
goop for Clang. However, they must be included at *exactly* the right
point in the sequence of header files, which makes it essentially
impossible to have this be managed by the Frontend and the rest by the
Driver. I have terrible ideas that would "work", but I think they're
worse than putting this in the driver and making the Frontend library
even more ignorant of the environment and system on which it is being
run.
Also fix the fact that we weren't properly respecting the flags which
suppress standard system include directories.
Note that this still leaves all of the Clang tests which run CC1
directly and include builtin header files broken on Windows. I'm working
on a followup patch to address that.
Add two flags to the CC1 layer that I was hoping to avoid. We need to
encode the *exact* semantics which the header search paths internally
built by the Frontend layer have had, which is both non-user-provided,
and at times adding the implicit extern "C" bit to the directory entry.
There are lots of CC1 options that are very close, but none do quite
this, and they are all already overloaded for other purposes. In some
senses this makes the command lines more clean as it clearly indicates
which flags are exclusively used to implement internal detection of
"standard" header search paths.
Lots of the implementation of this is really crufty, due to the
surrounding cruft. It doesn't seem worth investing lots of time cleaning
this up as it isn't new, and hopefully *lots* of this code will melt
away as header search inside of the frontend becomes increasingly
trivial.
Ted Kremenek [Sat, 5 Nov 2011 07:34:28 +0000 (07:34 +0000)]
Tweak LookThroughStmt() in LiveVariables to properly look through alternativing ParenExprs and OpaqueValueExprs. Thanks to Anna and Argiris for iterating on this function. My original patch embarssingly didn't even pass the Clang tests.
Ted Kremenek [Sat, 5 Nov 2011 00:10:04 +0000 (00:10 +0000)]
Serialized diagnostics: pull category name serialization into diagnostic blocks. The goal is to remove BLOCK_STRINGS so that the bitcode file can potentially be streamed.
Begin the migration of header search logic to the driver, starting with
Windows. There are still FIXMEs and lots of problems with this code.
Some of them will be addressed shortly by my follow-up patches, but most
are going to wait until we isolate this code and can fix it properly.
This version should be no worse than what we had before.
Switch the C++ include interface in the ToolChain to use the same naming
as the system include interface before I start adding implementations of
it to individual ToolChain implementations.
In ARC, when applying an ownership to a non-objc pointer, instead of ignoring it
create an attributed type with same type as the original type.
We effectively retain the source info that an ownership attribute was present but the attribute
is ignored by not modifying the type that it was applied to.
Sink the strange '-stdlib=...' flag handling into the C++ include
handling logic of the generic ToolChain. This flag, despite its name,
has *nothing* to do with the GCC flag '-nostdlib' that relates
(exclusively) to the linking behavior. It is a most unfortunate name in
that regard...
It is used to tell InitHeaderSearch.cpp *which* set of C++ standard
library header search paths to use -- those for libstdc++ from GCC's
installation, or those from a libc++ installation. As this logic is
hoisted out of the Frontend, and into the Driver as part of this
ToolChain, the generic method will be overridden for the platform, where
it can implement this logic directly. As such, hiding the CC1 option
passing in the generic space is a natural fit despite the odd naming.
Also, expand on the comments to clarify whats going on, and tidy up the
Tools.cpp code now that its simpler.
Sink the handling of -fobjc-arc-cxxlib to live with the other -fobjc-arc
implementation in the driver. This cleans up the signature and semantics
of the include flag adding component of the toolchain. Another step to
ready it for holding all the InitHeaderSearch logic.
Add a FIXME to cleanup the CPATH handling. It's currently burried with
the rest of the mess in InitHeaderSearch.cpp. We could hoist it into the
driver profitably, removing more noise from the driver -> frontend
communication.
Add a system include management interface to the toolchain, and call it
and the C++ include management routine from the proper place when
forming preprocessor options in the driver. This is the first step to
teaching the driver to manage all of the header search paths. Currently,
these methods remain just stubs in the abstract toolchain. Subsequent
patches will flesh them out with implementations for various toolchains
based on the current code in InitHeaderSearch.cpp.
Richard Smith [Fri, 4 Nov 2011 05:33:44 +0000 (05:33 +0000)]
Constant expression evaluation: refactor to start the groundwork for coping with
initializations which refer indirectly to elements of the object being
initialized.
Richard Smith [Fri, 4 Nov 2011 02:25:55 +0000 (02:25 +0000)]
Constant expression evaluation: track the manner in which an lvalue was written,
to allow us to implement the C++11 rule that a non-active union member can't be
read, and use it to implement subobject access for string literals.
Richard Smith [Fri, 4 Nov 2011 01:10:57 +0000 (01:10 +0000)]
Constant expression evaluation: although we don't know whether a literal will
be at the same address as another object, we do know it won't alias a null
pointer.
Chad Rosier [Thu, 3 Nov 2011 21:23:39 +0000 (21:23 +0000)]
Parse the warning options twice. The first pass sets diagnostic state, while
the second pass reports warnings/errors. This has the effect that we follow
the more canonical "last option wins" paradigm when there are conflicting
options.
rdar://10383776
Previously, we parsed the warning options in order. This caused non-intuitive
behavior:
projects / clang / log