Douglas Gregor [Fri, 27 Jan 2012 19:52:33 +0000 (19:52 +0000)]
Introduce module attributes into the module map grammar, along with a
single attribute ("system") that allows us to mark a module as being a
"system" module. Each of the headers that makes up a system module is
considered to be a system header, so that we (for example) suppress
warnings there.
If a module is being inferred for a framework, and that framework
directory is within a system frameworks directory, infer it as a
system framework.
Douglas Gregor [Fri, 27 Jan 2012 01:47:08 +0000 (01:47 +0000)]
Reimplement (de-)serialization of Objective-C categories to eliminate
the direct serialization of the linked-list structure. Instead, use a
scheme similar to how we handle redeclarations, with redeclaration
lists on the side. This addresses several issues:
- In cases involving mixing and matching of many categories across
many modules, the linked-list structure would not be consistent
across different modules, and categories would get lost.
- If a module is loaded after the class definition and its other
categories have already been loaded, we wouldn't see any categories
in the newly-loaded module.
Richard Smith [Fri, 27 Jan 2012 01:14:48 +0000 (01:14 +0000)]
constexpr: Implement the [dcl.constexpr]p5 check for whether a constexpr
function definition can produce a constant expression. This also provides the
last few checks for [dcl.constexpr]p3 and [dcl.constexpr]p4.
Eli Friedman [Thu, 26 Jan 2012 23:34:06 +0000 (23:34 +0000)]
Turn off implicit truncation warning for compound assignment to bitfields; it might be reasonable in some cases, but it clearly doesn't make sense in some cases, like the included testcase.
Eli Friedman [Thu, 26 Jan 2012 23:11:39 +0000 (23:11 +0000)]
Make the bitfield implicit truncation warning slightly more aggressive, and make the printed warning a bit more accurate. The new behavior matches gcc's -Wconversion. <rdar://problem/10238797>.
Remove the ToolTriple logic in NetBSD, which was completely broken by
the recent refactoring. All interesting NetBSD release have a GNU as
version on i386 that supports --32, so don't bother with the conditional
setting of it.
Bob Wilson [Thu, 26 Jan 2012 22:14:27 +0000 (22:14 +0000)]
Make clz/ctz builtins defined for zero on ARM targets. rdar://10732455
ARM supports clz and ctz directly and both operations have well-defined
results for zero. There is no disadvantage in performance to using the
defined-at-zero versions of llvm.ctlz/cttz intrinsics. We're running into
ARM-specific code written with the assumption that __builtin_clz(0) == 32,
even though that value is technically undefined. The code is failing now
because of llvm optimizations that are taking advantage of the undef
behavior (specifically svn r147255). There's nothing wrong with that
optimization on x86 where any incorrect assumptions about __builtin_clz(0)
will quickly be exposed. For ARM, though, optimizations based on that undef
behavior are likely to cause subtle bugs. Other targets with defined-at-zero
clz/ctz support may want to override the default behavior as well.
Keep track of the original target the user specified before
normalization. This used to be captured in DefaultTargetTriple and is
used for the (optional) $triple-$tool lookup for cross-compilation.
Do this properly by making it an attribute of the toolchain and use it
in combination with the computed triple as index for the toolchain
lookup.
Ted Kremenek [Thu, 26 Jan 2012 21:29:00 +0000 (21:29 +0000)]
Change references to 'const ProgramState *' to typedef 'ProgramStateRef'.
At this point this is largely cosmetic, but it opens the door to replace
ProgramStateRef with a smart pointer that more eagerly acts in the role
of reclaiming unused ProgramState objects.
objc-arc: introduce -no-finalize-removal which in gc mode,
leaves "finalize' behind and in arc mode, does not
include it. This allows the migrated source to be compiled
in both gc and arc mode. // rdar://10532441
John McCall [Thu, 26 Jan 2012 20:04:03 +0000 (20:04 +0000)]
Don't suppress access-control or invalid-type diagnostics from a
declarator just because we were able to build an invalid decl
for it. The invalid-type diagnostics, in particular, are still useful
to know, and may indicate something about why the decl is invalid.
Also, recover from an illegal pointer/reference-to-unqualified-retainable
type using __strong instead of __autoreleasing; in general, a random
object is much more likely to be __strong, so this avoids unnecessary
cascading errors in the most common case.
Chandler Carruth [Thu, 26 Jan 2012 09:28:50 +0000 (09:28 +0000)]
Suppress any warnings from this test. They aren't interesting, and they
end up in the same output file as the layout stuff. There may even be
a race condition which is causing this output to confuse the FileCheck
in some cases. I actually don't know how on earth the parsing of the
layout file even works given that there are diagnostics in the middle of
it. ;]
Douglas Gregor [Thu, 26 Jan 2012 07:55:45 +0000 (07:55 +0000)]
Extend the ExternalASTSource interface to allow the AST source to
provide the layout of records, rather than letting Clang compute
the layout itself. LLDB provides the motivation for this feature:
because various layout-altering attributes (packed, aligned, etc.)
don't get reliably get placed into DWARF, the record layouts computed
by LLDB from the reconstructed records differ from the actual layouts,
and badness occurs. This interface lets the DWARF data drive layout,
so we don't need the attributes preserved to get the answer write.
The testing methodology for this change is fun. I've introduced a
variant of -fdump-record-layouts called -fdump-record-layouts-simple
that always has the simple C format and provides size/alignment/field
offsets. There is also a -cc1 option -foverride-record-layout=<file>
to take the output of -fdump-record-layouts-simple and parse it to
produce a set of overridden layouts, which is introduced into the AST
via a testing-only ExternalASTSource (called
LayoutOverrideSource). Each test contains a number of records to lay
out, which use various layout-changing attributes, and then dumps the
layouts. We then run the test again, using the preprocessor to
eliminate the layout-changing attributes entirely (which would give us
different layouts for the records), but supplying the
previously-computed record layouts. Finally, we diff the layouts
produced from the two runs to be sure that they are identical.
Note that this code makes the assumption that we don't *have* to
provide the offsets of bases or virtual bases to get the layout right,
because the alignment attributes don't affect it. I believe this
assumption holds, but if it does not, we can extend
LayoutOverrideSource to also provide base offset information.
Fixes the Clang side of <rdar://problem/10169539>.
Eric Christopher [Thu, 26 Jan 2012 07:01:04 +0000 (07:01 +0000)]
Revert previous patch unifying all of the C++ record prep in one area,
the gdb testsuite complains too much about the ordering of items printed,
even if the offsets in the debug info are correct.
Bob Wilson [Thu, 26 Jan 2012 06:22:30 +0000 (06:22 +0000)]
Fix a minor bug in r148582, which made -ccc-host-triple into an alias option.
I'm not adding a testcase because -ccc-host-triple is slated to be removed,
but clang crashes if you try to use -ccc-host-triple without this fix.
Improve efficiency of Sema::MaybeBindToTemporary by working with the
canonical type directly and adding a fast path for the common case
that the type is directly a RecordType.
Eli Friedman [Thu, 26 Jan 2012 03:00:14 +0000 (03:00 +0000)]
Refactor to share code for handling return statements between lambda expressions and block literals. As it turns out, almost all the logic can be shared.
-fixit-recompile
applies fixits and recompiles the result
-fixit-to-temporary
applies fixits to temporary files
-fix-only-warnings">,
applies fixits for warnings only, not errors
Combining "-fixit-recompile -fixit-to-temporary" allows testing the result of fixits
without touching the original sources.
Rafael Espindola [Thu, 26 Jan 2012 02:02:57 +0000 (02:02 +0000)]
Fix our handling of #pragma GCC visibility.
Now the lexer just produces a token and the parser is the one responsible for
activating it.
This fixes problem like the one pr11797 where the lexer and the parser were not
in sync. This also let us be more strict on where in the file we accept
these pragmas.
Chandler Carruth [Thu, 26 Jan 2012 01:35:15 +0000 (01:35 +0000)]
Reintroduce r148981 with significantly improved regression test. Now it
both actually tests what it wants to, doesn't have bogus and broken
assertions in it, and is also formatted much more cleanly and
consistently. Probably still some more that can be improved here, but
its much better.
Original commit message:
----
Try to unbreak the FreeBSD toolchain's detection of 32-bit targets
inside a 64-bit freebsd machine with the 32-bit compatibility layer
installed. The FreeBSD image always has the /usr/lib32 directory, so
test for the more concrete existence of crt1.o. Also enhance the tests
for freebsd to clarify what these trees look like and exercise the new
code.
Thanks to all the FreeBSD folks for helping me understand what caused
the failure and how we might fix it. =] That helps a lot. Also, yay
build bots.
Ted Kremenek [Wed, 25 Jan 2012 23:47:14 +0000 (23:47 +0000)]
Rework flushing of diagnostics to PathDiagnosticConsumer. Now all the reports are batched up before being flushed
to the underlying consumer implementation. This allows us to unique reports across analyses to multiple functions (which
shows up with inlining).
Revert r148981 because it fails test/Driver/freebsd.c
Original log:
Author: chandlerc <chandlerc@91177308-0d34-0410-b5e6-96231b3b80d8>
Date: Wed Jan 25 21:32:31 2012 +0000
Try to unbreak the FreeBSD toolchain's detection of 32-bit targets
inside a 64-bit freebsd machine with the 32-bit compatibility layer
installed. The FreeBSD image always has the /usr/lib32 directory, so
test for the more concrete existence of crt1.o. Also enhance the tests
for freebsd to clarify what these trees look like and exercise the new
code.
Thanks to all the FreeBSD folks for helping me understand what caused
the failure and how we might fix it. =] That helps a lot. Also, yay
build bots.
Chandler Carruth [Wed, 25 Jan 2012 21:32:31 +0000 (21:32 +0000)]
Try to unbreak the FreeBSD toolchain's detection of 32-bit targets
inside a 64-bit freebsd machine with the 32-bit compatibility layer
installed. The FreeBSD image always has the /usr/lib32 directory, so
test for the more concrete existence of crt1.o. Also enhance the tests
for freebsd to clarify what these trees look like and exercise the new
code.
Thanks to all the FreeBSD folks for helping me understand what caused
the failure and how we might fix it. =] That helps a lot. Also, yay
build bots.
Chandler Carruth [Wed, 25 Jan 2012 21:03:58 +0000 (21:03 +0000)]
Restore a tiny bit of functionality that I completely overlooked in the
Linux toolchain selection -- sorry folks. =] This should fix the Hexagon
toolchain.
However, I would point out that I see why my testing didn't catch this
-- we have no tests for Hexagon. ;]
Kaelyn Uhrain [Wed, 25 Jan 2012 20:49:08 +0000 (20:49 +0000)]
Avoid correcting unknown identifiers to types where types aren't allowed.
Pass a typo correction callback object from ParseCastExpr to
Sema::ActOnIdExpression to be a bit more selective about what kinds of
corrections will be allowed for unknown identifiers.
Kaelyn Uhrain [Wed, 25 Jan 2012 18:37:44 +0000 (18:37 +0000)]
Add custom callback object for typo correction in BuildRecoveryCallExpr.
The new callback, in addition to limiting which keywords to include in
the pool of typo correction candidates, also filters out non-keyword
candidates that don't refer to (template) functions that accept the
number of arguments that are present for the call being recovered.
Douglas Gregor [Wed, 25 Jan 2012 16:15:54 +0000 (16:15 +0000)]
When we're substituting into a function parameter pack and expect to
get a function parameter pack (but don't due to weird substitutions),
complain. Fixes the last bit of PR11848.
Chandler Carruth [Wed, 25 Jan 2012 11:24:24 +0000 (11:24 +0000)]
The Linux pattern of adding all the search paths that exist doesn't seem
to suit the FreeBSD folks. Take them back to something closer to the old
behavior. We test whether the /usr/lib32 directory exists (within the
SysRoot), and use it if so, otherwise use /usr/lib.
FreeBSD folks, let me know if this causes any problems, or if you have
further tweaks.
Chandler Carruth [Wed, 25 Jan 2012 11:18:20 +0000 (11:18 +0000)]
Remove the 'ToolTriple' concept from the NetBSD toolchain along with my
gross hack to provide it from my previous patch removing HostInfo. This
was enshrining (and hiding from my searches) the concept of storing and
diff-ing the host and target triples. We don't have the host triple
reliably available, so we need to merely inspect the target system. I've
changed the logic in selecting library search paths for NetBSD to match
what I provided for FreeBSD -- we include both search paths, but put the
32-bit-on-64-bit-host path first so it trumps.
NetBSD maintainers, you may want to tweak this, or feel free to ask me
to tweak it. I've left a FIXME here about the challeng I see in fixing
this properly.
Chandler Carruth [Wed, 25 Jan 2012 11:01:57 +0000 (11:01 +0000)]
Delete the driver's HostInfo class. This abstraction just never really
did anything. The two big pieces of functionality it tried to provide
was to cache the ToolChain objects for each target, and to figure out
the exact target based on the flag set coming in to an invocation.
However, it had a lot of flaws even with those goals:
- Neither of these have anything to do with the host, or its info.
- The HostInfo class was setup as a full blown class *hierarchy* with
a separate implementation for each "host" OS. This required
dispatching just to create the objects in the first place.
- The hierarchy claimed to represent the host, when in fact it was
based on the target OS.
- Each leaf in the hierarchy was responsible for implementing the flag
processing and caching, resulting in a *lot* of copy-paste code and
quite a few bugs.
- The caching was consistently done based on architecture alone, even
though *any* aspect of the targeted triple might change the behavior
of the configured toolchain.
- Flag processing was already being done in the Driver proper,
separating the flag handling even more than it already is.
Instead of this, we can simply have the dispatch logic in the Driver
which previously created a HostInfo object create the ToolChain objects.
Adding caching in the Driver layer is a tiny amount of code. Finally,
pulling the flag processing into the Driver puts it where it belongs and
consolidates it in one location.
The result is that two functions, and maybe 100 lines of new code
replace over 10 classes and 800 lines of code. Woot.
This also paves the way to introduce more detailed ToolChain objects for
various OSes without threading through a new HostInfo type as well, and
the accompanying boiler plate. That, of course, was the yak I started to
shave that began this entire refactoring escapade. Wheee!
Chandler Carruth [Wed, 25 Jan 2012 10:50:34 +0000 (10:50 +0000)]
Add some really minimalist freebsd testing trees and use them in the
freebsd test so that it's behavior isn't dependent on the filesystem of
the host running the tests. This should revive the build bots at least.
The tests and the trees still need a lot of love to make them as useful
and easy to maintain as linux-ld.c.
Chandler Carruth [Wed, 25 Jan 2012 09:12:06 +0000 (09:12 +0000)]
Switch the ToolChain types to all store a Driver reference rather than
a HostInfo reference. Nothing about the HostInfo was used by any
toolchain except digging out the driver from it. This just makes that
a lot more direct. The change was accomplished entirely mechanically.
It's one step closer to removing the shim full of buggy copy/paste code
that is HostInfo.
Chandler Carruth [Wed, 25 Jan 2012 08:49:21 +0000 (08:49 +0000)]
Remove the TargetTriple object that I added to the Driver recently. This
helped stage the refactoring of things a bit, but really isn't the right
place for it. The driver may be responsible for compilations with many
different targets. In those cases, having a target triple in the driver
is actively misleading because for many of those compilations that is
not actually the triple being targeted.
This moves the last remaining users of the Driver's target triple to
instead use the ToolChain's target triple. The toolchain has a single,
concrete target it operates over, making this a more stable and natural
home for it.
Chandler Carruth [Wed, 25 Jan 2012 08:10:33 +0000 (08:10 +0000)]
Have FreeBSD use even more of the same smarts as Linux is now using for
adding search paths. Add them only when they exist, and prefix the paths
with the sysroot. This will allow targeting a FreeBSD sysroot on
a non-FreeBSD host machine, and perhaps more importantly should allow
testing the FreeBSD driver's behavior similarly to the Linux tests with
a fake tree of files in the regression test suite.
I don't have FreeBSD systems handy to build up the list of files that
should be used here, but this is the basic functionality and I'm hoping
Roman or someone from the community can contribute the actual test
cases.
Chandler Carruth [Wed, 25 Jan 2012 08:04:15 +0000 (08:04 +0000)]
Switch FreeBSD to just include both '/usr/lib32' and '/usr/lib' in the
search paths for 32-bit targets. This avoids having to detect which is
expected for the target system, and the linker should DTRT, and take the
32-bit libraries from the first one when applicable. Thanks to Roman
Divacky for sanity checking this.
Chandler Carruth [Wed, 25 Jan 2012 07:21:38 +0000 (07:21 +0000)]
Make a major refactoring to how the GCC installation detection works.
The fundamental shift here is to stop making *any* assumptions about the
*host* triple. Where these assumptions you ask? Why, they were in one of
the two target triples referenced of course. This was the single biggest
place where the previously named "host triple" was actually used as
such. ;] The reason we were reasoning about the host is in order to
detect the use of '-m32' or '-m64' flags to change the target. These
flags shift the default target only slightly, which typically means
a slight deviation from the host. When using these flags, the GCC
installation is under a different triple from the one actually targeted
in the compilation, and we used the host triple to find it.
Too bad that wasn't even correct. Consider an x86 Linux host which has
a PPC64 cross-compiling GCC toolchain installed. This toolchain is also
configured for multiarch compiling and can target PPC32 with eth '-m32'
flag. When targeting 'powerpc-linux-gnu' or some other PPC32 triple, we
have to look for the PPC64 variant of the triple to find the GCC
install, and that triple is neither the host nor target.
The new logic computes the multiarch's alternate triple from the target
triple, and looks under both sides. It also looks more aggressively for
the correct subdirectory of the GCC installation, and exposes the
subdirectory in a nice programmatic way. This '/32' or '/64' suffix is
something we can reuse in many other parts of the toolchain.
An important note -- while this likely fixes a large category of
cross-compile use cases, that's not my primary goal, and I've not done
testing (or added test cases) for scenarios that may now work. If
someone else wants to try more interesting PPC cross compiles, I'd love
to have reports. But my focus is on factoring away the references to the
"host" triple. The refactoring is my goal, and so I'm mostly relying on
the existing (pretty good) test coverage we have here.
Future patches will leverage this new functionality to factor out more
and more of the toolchain's triple manipulation.
Chris Lattner [Wed, 25 Jan 2012 05:34:41 +0000 (05:34 +0000)]
reapply r148902:
"use the new ConstantVector::getSplat method where it makes sense."
Also simplify a bunch of code to use the Builder->getInt32 instead
of doing it the hard and ugly way. Much more progress could be made
here, but I don't plan to do it.