Reid Kleckner [Fri, 21 Jun 2013 12:45:15 +0000 (12:45 +0000)]
[ms-cxxabi] Destroy temporary record arguments in the callee
Itanium destroys them in the caller at the end of the full expression,
but MSVC destroys them in the callee. This is further complicated by
the need to emit EH-only destructor cleanups in the caller.
This should help clang compile MSVC's debug iterators more correctly.
There is still an outstanding issue in PR5064 of a memcpy emitted by the
LLVM backend, which is not correct for C++ records.
David Blaikie [Fri, 21 Jun 2013 03:41:46 +0000 (03:41 +0000)]
DebugInfo: handle the DI asm printing change to reword '[fwd]' as '[decl]' and add '[def]' for non-decl tag types
This is to make test cases looking for definitions more legible by
making the definition explicit rather than just the absence of '[fwd]'.
This allowed the debug-info-record tests to be rephrased - and in the
interests of reducing the number of individual test cases/invocations we
have, I merged them into one file, separated them with namespaces (&
then moved them to C++ because namespaces are great). If they need to
remain 'C' only tests, they can be moved back. (I didn't group them with
'debug-info-class.cpp' because these tests only apply to
-fno-limit-debug-info)
I removed the pieces of code that would cause these tests to pass under
-flimit-debug-info to ensure the tests remain relevant to their fixes
should we ever improve -flimit-debug-info to catch that kind of code.
This commit is version locked with the corresponding change to
DebugInfo.h in LLVM. Except some transient buildbot fallout.
Jordan Rose [Fri, 21 Jun 2013 00:59:00 +0000 (00:59 +0000)]
[analyzer] Handle zeroing CXXConstructExprs.
Certain expressions can cause a constructor invocation to zero-initialize
its object even if the constructor itself does no initialization. The
analyzer now handles that before evaluating the call to the constructor,
using the same "default binding" mechanism that calloc() uses, rather
than simply ignoring the zero-initialization flag.
As a bonus, trivial default constructors are now no longer inlined; they
are instead processed explicitly by ExprEngine. This has a (positive)
effect on the generated path edges: they no longer stop at a default
constructor call unless there's a user-provided implementation.
David Blaikie [Fri, 21 Jun 2013 00:40:50 +0000 (00:40 +0000)]
Alternative fix for r184473.
This just seems a bit tidier/more principled. Based on a patch provided
by Adrian - with the only minor tweak that it needed to use
"getTypeOrNull" rather than "getCompletedTypeOrNull" since we don't
store declarations in the CompletedTypes cache.
Douglas Gregor [Fri, 21 Jun 2013 00:20:25 +0000 (00:20 +0000)]
When building a module, keep *all* declared methods in the global method pool.
As an optimization, we only kept declared methods with distinct
signatures in the global method pool, to keep the method lists
small. Under modules, however, one could have two different methods
with the same signature that occur in different (sub)modules. If only
the later submodule is important, message sends to 'id' with that
selector would fail because the first method (the only one that got
into the method pool) was hidden. When building a module, keep *all*
of the declared methods.
I did a quick check of both module build time and uses of modules, and
found no performance regression despite this causing us to keep more
methods in the global method pool. Fixes <rdar://problem/14148896>.
Chandler Carruth [Thu, 20 Jun 2013 23:37:54 +0000 (23:37 +0000)]
Update a comment to clarify that searching the target triple bin
directory for programs used by the driver is actually the standard
behavior we want to be compatible with GCC cross compilers -- it isn't
specific to SUSE or any other distro.
Also start fleshing out testing of the different cross compilation
patterns, both with a new very bare-bones tree of cross compilers and by
extending the multilib trees. Currently, we don't correctly model doing
a cross compile using the non-triple target of a bi-arch GCC install,
but I'll add support for that (and tests) next.
Richard Smith [Thu, 20 Jun 2013 22:21:56 +0000 (22:21 +0000)]
Avoid repeatedly evaluating subexpressions when checking for unsequenced
operations in the case where evaluating a subexpression fails. No functionality
change, but test/Sema/many-logical-ops.c gets ~100x faster with this change.
John McCall [Thu, 20 Jun 2013 21:37:43 +0000 (21:37 +0000)]
Call __cxa_begin_catch before std::terminate() in a noexcept trap.
r174939-40 caused us to do this in the canonical terminate lpad,
but when the EH stack has other cleanups on it we use the
terminate handler block, which wasn't doing this.
Fixes the rest of rdar://11904428 given appropriate stdlib support.
Meador Inge [Thu, 20 Jun 2013 19:47:18 +0000 (19:47 +0000)]
CodeGen: Don't set 'PMBuilder.DisableSimplifyLibCalls'
The simplify-libcalls pass has been removed from LLVM. Thus
'PMBuilder.DisableSimplifyLibCalls' does not exist anymore.
The disabling/enabling of library call simplifications is
done through the TargetLibraryInfo which is already wired
up in Clang.
Reid Kleckner [Thu, 20 Jun 2013 16:28:24 +0000 (16:28 +0000)]
Fix CodeGenCXX/debug-info.cpp test on Windows
On Windows, it looks like FlagIndirectVariable is being set in Flags for
DIBuilder::createLocalVariable(), giving us an i32 of 8192 instead of 0,
as on Linux.
Samuel Benzaquen [Thu, 20 Jun 2013 14:28:32 +0000 (14:28 +0000)]
Enhancements for the DynTypedMatcher system.
- Added conversion routines and checks in Matcher<T> that take a DynTypedMatcher.
- Added type information on the error messages for the marshallers.
- Allows future work on Polymorphic/overloaded matchers. We should be
able to disambiguate at runtime and choose the appropriate overload.
Use the same set of whitespace characters for all operations in BreakableToken.
Summary:
Fixes a problem where \t,\v or \f could lead to a crash when placed as
a first character in a line comment. The cause is that rtrim and ltrim handle
these characters, but our code didn't, so some invariants could be broken.
I've added legacy aliases for the original spellings. I've updated the
canonical tests to check both spellings, and switched all of the
-gcc-toolchain usages elsewhere in the test suite to use the new one.
I've updated some of the usages of -target to the new syntax, but will
finish that in a separate entirely mechanical change once I'm sure this
won't get rolled back for some reason (It touches a *huge* number of RUN
lines in the test suite unsurprisingly).
A nice result is that the three most common flags I end up using when
doing cross compiles are all now consistent: --target=, --sysroot=, and
--gcc-toolchain=.
Pavel Labath [Thu, 20 Jun 2013 07:45:01 +0000 (07:45 +0000)]
Fix static analyzer crash when casting from an incomplete type
Summary:
When doing a reinterpret+dynamic cast from an incomplete type, the analyzer
would crash (bug #16308). This fix makes the dynamic cast evaluator ignore
incomplete types, as they can never be used in a dynamic_cast. Also adding a
regression test.
Chandler Carruth [Thu, 20 Jun 2013 07:06:39 +0000 (07:06 +0000)]
Add a regression test for PR16370 next to the dr7 test case since they
seem closely related. (I'm happy to move this if others have a better
idea of where to put it.)
Richard Smith [Thu, 20 Jun 2013 02:18:31 +0000 (02:18 +0000)]
Add a workaround for a libstdc++-4.2 <tr1/hashtable> bug. This header uses
return false;
in a function returning a pointer. 'false' was a null pointer constant in C++98
but is not in C++11. Punch a very small hole in the initialization rules in
C++11 mode to allow this specific case in system headers.
[clang-lit] Added the run_long_tests param option/long_tests feature to toggle execution of long running FileCheck tests.
This will allow for longer running FileCheck based tests to be committed to
clang for use on buildbots, preventing the normal make-check cycle from
increasing in time significantly.
This is a necessary change in order to commit the end-to-end arm neon intrinsic
tests since FileCheck takes ~20 seconds to run said test due to the large amount
of neon intrinsics.
To force a test to run only when --param run_long_tests=true is passed in use
the following requires statement:
David Blaikie [Wed, 19 Jun 2013 21:53:53 +0000 (21:53 +0000)]
Debug Info: PR14763/r183329 - specify that non-trivial pass-by-value parameters are stored indirectly
This is to fix the location information for such parameters to refer to
the object accessible through the pointer rather than to the pointer
parameter itself.
Fixed long-standing issue with incorrect length calculation of multi-line comments.
Summary:
A trailing block comment having multiple lines would cause extremely
high penalties if the summary length of its lines is more than the column limit.
Fixed by always considering only the last line of a multi-line block comment.
Removed a long-standing FIXME from relevant tests and added a motivating test
modelled after problem cases from real code.
Stephen Lin [Wed, 19 Jun 2013 18:10:35 +0000 (18:10 +0000)]
Corrections to r184205 ('this'-return optimization) due to the wrong version of the patch being committed originally.
1) Removed useless return value of CGCXXABI::EmitConstructorCall and CGCXXABI::EmitVirtualDestructorCall and implementations
2) Corrected last portion of CodeGenCXX/constructor-destructor-return-this to correctly test for non-'this'-return of virtual destructor calls
Manuel Klimek [Wed, 19 Jun 2013 15:42:45 +0000 (15:42 +0000)]
Completely revamp node binding for AST matchers.
This is in preparation for the backwards references to bound
nodes, which will expose a lot more about how matches occur. Main
changes:
- instead of building the tree of bound nodes, we build a "set" of bound
nodes and explode all possible match combinations while running
through the matchers; this will allow us to also implement matchers
that filter down the current set of matches, like "equalsBoundNode"
- take the set of bound nodes at the start of the match into
consideration when doing memoization; as part of that, reevaluated
that memoization gives us benefits that are large enough (it still
does - the effect on common match patterns is up to an order of
magnitude)
- reset the bound nodes when a node does not match, thus never leaking
information from partial sub-matcher matches for failing matchers
Effects:
- we can now correctly "explode" combinatorial matches, for example:
allOf(forEachDescendant(...bind("a")),
forEachDescendant(...bind("b"))) will now trigger matches for all
combinations of matching "a" and "b"s.
- we now never expose bound nodes from partial matches in matchers that
did not match in the end - this fixes a long-standing issue
FIXMEs:
- rename BoundNodesTreeBuilder to BoundNodesBuilder or
BoundNodesSetBuilder, as we don't build a tree any more; this is out
of scope for this change, though
- we're seeing some performance regressions (around 10%), but I expect
some performance tuning will get that back, and it's easily worth
the increase in expressiveness for now
Reid Kleckner [Wed, 19 Jun 2013 15:20:38 +0000 (15:20 +0000)]
[ms-cxxabi] Emit and install appropriately mangled vbtables
In Itanium, dynamic classes have one vtable with several different
address points for dynamic base classes that can't share vtables.
In the MS C++ ABI, each vbtable that can't be shared gets its own
symbol, similar to how ctor vtables work in Itanium. However, instead
of mangling the subobject offset into the symbol, the unique portions of
the inheritance path are mangled into the symbol to make it unique.
This patch implements the MSVC 2012 scheme for forming unique vbtable
symbol names. MSVC 2010 use the same mangling with a different subset
of the path. Implementing that mangling and possibly others is TODO.
Each vbtable is an array of i32 offsets from the vbptr that points to it
to another virtual base subobject. The first entry of a vbtable always
points to the base of the current subobject, implying that it is the
same no matter which parent class contains it.
Reid Kleckner [Wed, 19 Jun 2013 15:09:06 +0000 (15:09 +0000)]
[Driver] Don't forward source file input args to gcc
gcc's inputs are already added by the InputInfoList passed to
Action::ConstructJob.
Fixes a regression from r183989. This was manifesting when targetting
mingw as an extra input argument to gcc when assembling. It presumably
affects other situations where clang calls gcc.
Prior to r183989, forwardToGCC() was returning false because the INPUT
option defined in OptParser.td had the DriverOption flag set on it.
LLVM's Option library does not set this flag for INPUT.
Manman Ren [Wed, 19 Jun 2013 01:46:49 +0000 (01:46 +0000)]
Debug Info: support for gdwarf-2 gdwarf-3 gdwarf-4
These options will add a module flag with name "Dwarf Version".
The behavior flag is currently set to Warning, so when two values disagree,
a warning will be emitted.
Anna Zaks [Tue, 18 Jun 2013 23:16:15 +0000 (23:16 +0000)]
[analyzer] Do not report uninitialized value warnings inside swap functions.
This silences warnings that could occur when one is swapping partially initialized structs. We suppress
not only the assignments of uninitialized members, but any values inside swap because swap could
potentially be used as a subroutine to swap class members.
This silences a warning from std::try::function::swap() on partially initialized objects.
Eli Friedman [Tue, 18 Jun 2013 22:41:37 +0000 (22:41 +0000)]
Introduce a new mangling for protocol-qualified ObjC types in C++. This allows
to provide proper overloading, and also prevents mangling conflicts with
template arguments of protocol-qualified type.
This is a non-backward-compatible mangling change, but per discussion with
John, the benefits outweigh this cost.
the lookup for 'OldImage' will return the 'NewImage' decl ("@class NewImage").
In such a case, when creating the decl for "@class OldImage" use the real declaration name ("NewImage"),
instead of the alias one ("OldImage"), otherwise we will break IdentifierResolver and redecls-chain invariants.