Richard Smith [Tue, 29 Jan 2013 10:02:16 +0000 (10:02 +0000)]
Downgrade 'attribute ignored when parsing type' from error to warning, to match
the diagnostic's warn_ name. Switch some places (notably C++11 attributes)
which really wanted an error over to a different diagnostic. Finally, suppress
the diagnostic entirely for __ptr32, __ptr64 and __w64, to avoid producing
diagnostics in important system headers.
Richard Smith [Tue, 29 Jan 2013 04:55:18 +0000 (04:55 +0000)]
C++11 status:
* Mark 'sequence points' as done now we have a warning for unsequenced operations
* Mark 'memory model' as done now we correctly model bitfield operations
Richard Smith [Tue, 29 Jan 2013 04:13:32 +0000 (04:13 +0000)]
PR15017: A '>' can appear after a type-specifier in a template-argument-list.
It turns out that there's no correctness bug here (because we can't have a type
definition in this location), but there was a diagnostic bug.
Richard Smith [Tue, 29 Jan 2013 01:48:07 +0000 (01:48 +0000)]
Treat alignas and _Alignas as keyword attributes. This allows us to
pretty-print them properly (modulo the more general badness in alignment
attribute printing).
Richard Smith [Tue, 29 Jan 2013 01:38:41 +0000 (01:38 +0000)]
Actually remove the hack which was blocking the Borland-style attributes from
working, and add the missing attribute spellings. This brings _pascal,
_fastcall, _stdcall and _cdecl to life in -fborland-extensions mode.
Richard Smith [Tue, 29 Jan 2013 01:24:26 +0000 (01:24 +0000)]
Replace AS_MSTypespec with AS_Keyword, for representing any attribute spelled
as a keyword. Rationalize existing attributes to use it as appropriate, and to
not lie about some __declspec attributes being GNU attributes. In passing,
remove a gross hack which was discarding attributes which we could handle. This
results in us actually respecting the __pascal keyword again.
Richard Smith [Mon, 28 Jan 2013 22:42:45 +0000 (22:42 +0000)]
Finish semantic analysis for [[carries_dependency]] attribute.
This required plumbing through a new flag to determine whether a ParmVarDecl is
actually a parameter of a function declaration (as opposed to a function
typedef etc, where the attribute is prohibited). Weirdly, this attribute (just
like [[noreturn]]) cannot be applied to a function type, just to a function
declaration (and its parameters).
Douglas Gregor [Mon, 28 Jan 2013 18:38:02 +0000 (18:38 +0000)]
Enable the global module index by default. Introduce the
-fno-modules-global-index -cc1 option to allow one to disable the
index for performance testing purposes, but with a 10% win in
-fsyntax-only time, there is no reason a user would do this.
Douglas Gregor [Mon, 28 Jan 2013 18:29:39 +0000 (18:29 +0000)]
ASTReader and profiling statistics indicate that implementing a method
pool in the global module index is not worthwhile. Update comments to
limit the scope of the global module index to identifiers.
Douglas Gregor [Mon, 28 Jan 2013 16:46:33 +0000 (16:46 +0000)]
Eliminate memory allocation from most invocations of
ModuleManager::visit() by keeping a free list of the two data
structures used to store state (a preallocated stack and a visitation
number vector). Improves -fsyntax-only performance for my modules test
case by 2.8%. Modules has pulled ahead by almost 10% with the global
module index.
Daniel Jasper [Mon, 28 Jan 2013 15:40:20 +0000 (15:40 +0000)]
Don't put a function's return type on its own line in Google style.
This would be against the style guide:
http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml?showone=Function_Declarations_and_Definitions#Function_Declarations_and_Definitions
Not sure what to do as a last resort if the function signature does not
fit onto a single line in Google style ..
Douglas Gregor [Mon, 28 Jan 2013 15:39:10 +0000 (15:39 +0000)]
Note that SrcMgr::SLocEntry's are POD-like, so SmallVector can do a
better job with them. Shaves off 0.7% of -fparse-only time for a
modules test case. Sure makes you wonder...
Hal Finkel [Mon, 28 Jan 2013 04:37:37 +0000 (04:37 +0000)]
Fix the indentation of the first line of preprocessor output
The -E output from clang did not produce the correct indentation on the first line.
This is because MoveToLine returned false, and when this happens,
the regular process for producing initial indentation is skipped.
Thanks to Eli for suggesting a way to simplify this to a one-line change.
Since ObjCARC has been refactored into its own library with its own declaration header, we need to include the declaration header alongside Scalar.h in BackendUtil.
Richard Smith [Mon, 28 Jan 2013 00:54:05 +0000 (00:54 +0000)]
Add a -pedantic warning: an anonymous union within an anonymous union is not
permitted in standard C++, despite being silently accepted by many (all?) major
C++ implementations.
NAKAMURA Takumi [Sun, 27 Jan 2013 12:20:50 +0000 (12:20 +0000)]
[CMake][Lit][unittests] Deprecate CMAKE_BUILD_TYPE in each build directory for unittests.
For example,
cur) unittests/ADT/Release/ADTTests
new) unittests/ADT/ADTTests
RUNTIME_BUILD_MODE can be substituted to CMAKE_CFG_INTDIR.
With Make and Ninja, the tree is not built with multiple configurations.
Then, including the build type in target directory doesn't make sense.
See also "How can I build multiple modes without switching?"
http://www.cmake.org/Wiki/CMake_FAQ
CMAKE_CFG_INTDIR is set to "."
With multiple-configuration-aware build system, like Visual Studio, each unittest is built on appropriate directory, for example,
unittests/ADT/Release/ADTTests.exe
CMAKE_CFG_INTDIR is set to build system's variable, like "$(Configuration)" or "$(OutDir)".
David Blaikie [Sat, 26 Jan 2013 22:16:26 +0000 (22:16 +0000)]
PR14566: Debug Info: avoid top level lexical blocks in functions
One of the gotchas (see changes to CodeGenFunction) was due to the fix in
r139416 (for PR10829). This only worked previously because the top level
lexical block would set the location to the end of the function, the debug
location would be updated (as per r139416), the location would be set to
the end of the function again (but that would no-op, since it was the same
as the previous location), then the return instruction would be emitted using
the debug location.
Once the top level lexical block was no longer emitted, the end-of-function
location change was causing the debug loc to be updated, regressing that bug.
Jordan Rose [Sat, 26 Jan 2013 03:16:31 +0000 (03:16 +0000)]
[analyzer] C++ initializers may require cleanups; look through these.
When the analyzer sees an initializer, it checks if the initializer
contains a CXXConstructExpr. If so, it trusts that the CXXConstructExpr
does the necessary work to initialize the object, and performs no further
initialization.
This patch looks through any implicit wrapping expressions like
ExprWithCleanups to find the CXXConstructExpr inside.
Richard Trieu [Sat, 26 Jan 2013 02:31:38 +0000 (02:31 +0000)]
Give a more informative error message when the dot or arrow operator is used
on a type. Currently, it gives a generic "expected unqualified-id" error.
The new error message is "cannot use (dot|arrow) operator on a type".
Richard Trieu [Sat, 26 Jan 2013 01:31:20 +0000 (01:31 +0000)]
Highlight various parts of the AST dump with color. Colors are controlled by
-f(no-)color-diagnostics. In addition, dumpColor() function calls are added
to force color printing. No structural changes to -ast-dump.
Jordan Rose [Sat, 26 Jan 2013 01:28:23 +0000 (01:28 +0000)]
[analyzer] Track null object lvalues back through C++ method calls.
The expression 'a->b.c()' contains a call to the 'c' method of 'a->b'.
We emit an error if 'a' is NULL, but previously didn't actually track
the null value back through the 'a->b' expression, which caused us to
miss important false-positive-suppression cases, including
<rdar://problem/12676053>.
Douglas Gregor [Sat, 26 Jan 2013 00:55:12 +0000 (00:55 +0000)]
Since we're stuck with realpath for the header <-> module mapping,
factor the realpath calls into FileManager::getCanonicalName() so we
can cache the results of this epically slow operation. 5% speedup on
my modules test, and realpath drops out of the profile.
patch for PR9027 and // rdar://11861085
Title: [PR9027] volatile struct bug: member is not loaded at -O;
This is caused by last flag passed to @llvm.memcpy being false,
not honoring that aggregate has at least one 'volatile' data member
(even though aggregate itself has not been qualified as 'volatile'.
As a result, optimization optimizes away the memcpy altogether.
Patch review by John MaCall (I still need to fix up a test though).
John McCall [Fri, 25 Jan 2013 23:36:19 +0000 (23:36 +0000)]
ARM says that the array cookie should always be eight bytes.
ARM is not thinking about over-aligned structures.
Overrule ARM in both our generic-ARM and iOS ABI implementations.