Anders Carlsson [Sun, 17 Apr 2011 21:56:13 +0000 (21:56 +0000)]
When laying out bases in, always try the "base subobject" LLVM type. If it
turns out that a field or base needs to be laid out in the tail padding of
the base, CGRecordLayoutBuilder::ResizeLastBaseFieldIfNecessary will convert
it to an array of i8.
I've audited the new test results to make sure that they are still valid. I've
also verified that we pass a self-host with this change.
Chris Lattner [Sun, 17 Apr 2011 00:40:24 +0000 (00:40 +0000)]
fold memcpy/set/move_chk to llvm.memcpy/set/move when the sizes
are trivial. This exposes opportunities earlier, and allows fastisel
to do good things with these at -O0.
This addresses rdar://9289468 - clang doesn't fold memset_chk at -O0
Chris Lattner [Sat, 16 Apr 2011 23:15:35 +0000 (23:15 +0000)]
fix rdar://9289603 - clang should fold trivial ?: for enums as well as integer constants into select at -O0
by making the isCheapEnoughToEvaluateUnconditionally predicate handle anything that folds to a constant. In particular, we now fold enums.
Introduce Rewriter::IncreaseIndentation() which increase indentations for the lines between the given source range.
To determine what the indentation should be, a SourceLocation 'parentIndent' parameter is used that should be at
a source location with an indentation one degree lower than the given range.
Douglas Gregor [Fri, 15 Apr 2011 20:45:44 +0000 (20:45 +0000)]
For the purposes of overload resolution, consider a conversion from an
Objective-C pointer to void* as a "conversion to void*". This allows
us to prefer an Objective-C object pointer conversion to a superclass
object pointer over an Objective-C object pointer conversion to
cv-void*. Fixes PR9735.
Douglas Gregor [Fri, 15 Apr 2011 19:46:20 +0000 (19:46 +0000)]
Forbid the use of C++ new/delete to allocate/free objects within an
address space. I could see that this functionality would be useful,
but not in its current form (where the address space is ignored):
rather, we'd want to encode the address space into the parameter list
passed to operator new/operator delete somehow, which would require a
bunch more semantic analysis.
Nick Lewycky [Fri, 15 Apr 2011 18:33:24 +0000 (18:33 +0000)]
Reapply r129561, moving the runtime/Makefile that builds compiler-rt into
runtime/compiler-rt/Makefile paving the way to put multiple different libraries
into runtime/ and build all of them.
Douglas Gregor [Fri, 15 Apr 2011 17:59:54 +0000 (17:59 +0000)]
Implement appropriate semantics for C++ casting and conversion when
dealing with address-space- and GC-qualified pointers. Previously,
these qualifiers were being treated just like cvr-qualifiers (in some
cases) or were completely ignored, leading to uneven behavior. For
example, const_cast would allow conversion between pointers to
different address spaces.
The new semantics are fairly simple: reinterpret_cast can be used to
explicitly cast between pointers to different address spaces
(including adding/removing addresss spaces), while
static_cast/dynamic_cast/const_cast do not tolerate any changes in the
address space. C-style casts can add/remove/change address spaces
through the reinterpret_cast mechanism. Other non-CVR qualifiers
(e.g., Objective-C GC qualifiers) work similarly.
As part of this change, I tweaked the "casts away constness"
diagnostic to use the term "casts away qualifiers". The term
"constness" actually comes from the C++ standard, despite the fact
that removing "volatile" also falls under that category. In Clang, we
also have restrict, address spaces, ObjC GC attributes, etc., so the
more general "qualifiers" is clearer.
Allow shadowin of 'self' in objc methods in
cases where stand-alone ivar can be looked up
in shadowed object. To fix gcc compatibility
breakage. // rdar://9284603
Richard Smith [Fri, 15 Apr 2011 10:12:39 +0000 (10:12 +0000)]
Apply NAKAMURA Takumi's workaround to fix thses tests on -Asserts build. The labels there are numbered, not named, and numbered labels are formatted differently.
Nick Lewycky [Fri, 15 Apr 2011 06:57:32 +0000 (06:57 +0000)]
Create a compiler-rt directory and move the Makefile to it. Add a makefile that
builds the subdirs from this directory. This makes the behaviour with make match
what already happens with cmake.
Do not expand the opencl pragma names. This is so you can also have a define named the samed as the pragma and they do not interfere (ie. cl_khr_fp64).
Douglas Gregor [Thu, 14 Apr 2011 23:19:27 +0000 (23:19 +0000)]
If the declaration of a C++ member function with an inline definition
is so broken that Sema can't form a declaration for it, don't bother
trying to parse the definition later. Fixes <rdar://problem/9221993>.
Douglas Gregor [Thu, 14 Apr 2011 22:32:28 +0000 (22:32 +0000)]
When we transform a C++ exception declaration (e.g., for template
instantiation), be sure to add the transformed declaration into the
current DeclContext. Also, remove the -Wuninitialized hack that works
around this bug. Fixes <rdar://problem/9200676>.
Richard Smith [Thu, 14 Apr 2011 21:45:45 +0000 (21:45 +0000)]
Detect when the string "<::" is found in code after a cast or template name and is interpreted as "[:" because of the digraph "<:". When found, give an error with a fix-it to add whitespace between the "<" and "::".
Patch by Richard Trieu! Plus a small tweak from me to deal with one of the tokens coming from a macro.
Douglas Gregor [Thu, 14 Apr 2011 21:41:34 +0000 (21:41 +0000)]
Harden Clang's cursor visitation logic against NULL declaration,
statement, and expression pointers. While these shouldn't happen, it's
better to be safe in libclang.
Douglas Gregor [Thu, 14 Apr 2011 20:33:34 +0000 (20:33 +0000)]
When determining the "usage" type of a declaration for the purposes of code
completion, look through block pointer and function pointer types to the
result type of the block/function. Fixes <rdar://problem/9282583>.
Implement ARM pcs attribute. Basically it's another way of calling convention selection (AAPCS or
AAPCS+VFP), similar to fastcall / stdcall / whatevercall seen on x86.
In particular, all library functions should always be AAPCS regardless of floating point ABI used.
Douglas Gregor [Thu, 14 Apr 2011 17:21:19 +0000 (17:21 +0000)]
Parse an '@' in an Objective-C++ class member specification,
diagnosing it as an error rather than looping infinitely. Also,
explicitly disallow @defs in Objective-C++. Fixes <rdar://problem/9260136>.
Sebastian Redl [Thu, 14 Apr 2011 14:07:59 +0000 (14:07 +0000)]
Chained PCH: Remember when additional specializations are added to a function template from a previous PCH. Fixes the only crasher when using massive chains on Clang's Sema component. We still have some incomplete codegen there.
Sebastian Redl [Thu, 14 Apr 2011 14:07:41 +0000 (14:07 +0000)]
The ASTReader created by -chain-include used the generated PCH buffers in the wrong order. The effect was that all but the first chain-include files was ignored for subsequent compilations.
Eli Friedman [Thu, 14 Apr 2011 00:51:41 +0000 (00:51 +0000)]
Re-fix r129481 and r129465 properly. Nulls fixits shouldn't be dropped in
DiagnosticBuilder::AddFixItHint: they will be dropped along with any
other (possibly valid) fixits later.
When creating an implicit member expression through a qualified-id, check that the class
named by the nested-name-specifier is same or base of the class in which the member expression appears.
It seems we also had an ill-formed test case, mon dieu! Fixes rdar://8576107.
Anders Carlsson [Thu, 14 Apr 2011 00:41:11 +0000 (00:41 +0000)]
In C++, when initializing an array from a pascal string, it's OK if the array
is 1 element smaller than the string, because we can just strip off the last
null character. This matches GCC.
Add -no-canonical-prefixes to tests which assert on the name of the
built clang binary that is used by the test. Build systems that use
symlinks for build outputs will fail these assertions otherwise.
Teach -Wuninitialized about C++'s typeid expression, including both the
evaluated and unevaluated contexts. Add some testing of sizeof and
typeid.
Both of the typeid tests added here were triggering warnings previously.
Now the one false positive is suppressed without suppressing the warning
on actually buggy code.
In Microsoft mode, within class scope, if a CXXScopeSpec's type is equal to the type of one of the base classes then downgrade the missing typename error to a warning. Up to now this is the only case I found where MSVC doesn't require "typename" at class scope. Really strange!
This fixes 1 error when parsing the MSVC 2008 header files.
Example:
template<class T> class A {
public:
typedef int TYPE;
};
template<class T> class B : public A<T> {
public:
A<T>::TYPE a; // no typename required because A<T> is a base class.
};
Daniel Dunbar [Tue, 12 Apr 2011 23:30:52 +0000 (23:30 +0000)]
IRgen/Obj-C: Emit CFStrings and NSStrings with the alignment of the char type,
there is no reason to align them higher.
- This roughly matches llvm-gcc's r126913.
- It is an open question whether or not we should do this for cstring's in
general (code size vs optimization potential), for now we just match llvm-gcc
until someone wants to run some experiments.
John McCall [Tue, 12 Apr 2011 22:02:02 +0000 (22:02 +0000)]
We can't emit an aggregate cast as its sub-expression in general just
because the result is ignored. The particular example here is with
property l-values, but there could be all sorts of lovely casts that this
isn't safe for. Sink the check into the one case that seems to actually
be capable of honoring this.
This patch adds modeling of strcmp() to the CString checker. Validates inputs are not NULL and are real C strings, then does the comparison and binds the proper return value. Unit tests included.
Douglas Gregor [Tue, 12 Apr 2011 02:47:21 +0000 (02:47 +0000)]
Don't suggest dynamic_cast or typeid as code completion results when
RTTI is disabled. Similarly, don't suggest throw or try as code
completion results when C++ exceptions are disabled. Fixes
<rdar://problem/9193560>.
projects / clang / log