access-control diagnostics which arise from the portion of the declarator
following the scope specifier, just in case access is granted by
friending the individual method. This can also happen with in-line
member function declarations of class templates due to templated-scope
friend declarations.
We were really playing fast-and-loose before with this sort of thing,
and it turned out to work because *most* friend functions are in file
scope. Making us delay regardless of context exposed several bugs with
how we were manipulating delay. I ended up needing a concept of a
context that's independent of the declarations in which it appears,
and then I actually had to make some things save contexts correctly,
but delay should be much cleaner now.
I also encapsulated all the delayed-diagnostics machinery in a single
subobject of Sema; this is a pattern we might want to consider rolling
out to other components of Sema.
llvm-svn: 125485
there were only three virtual methods of any significance.
The primary way to grab child iterators now is with
Stmt::child_range children();
Stmt::const_child_range children() const;
where a child_range is just a std::pair of iterators suitable for
being llvm::tie'd to some locals. I've left the old child_begin()
and child_end() accessors in place, but it's probably a substantial
penalty to grab the iterators individually now, since the
switch-based dispatch is kindof inherently slower than vtable
dispatch. Grabbing them together is probably a slight win over the
status quo, although of course we could've achieved that with vtables, too.
I also reclassified SwitchCase (correctly) as an abstract Stmt
class, which (as the first such class that wasn't an Expr subclass)
required some fiddling in a few places.
There are somewhat gross metaprogramming hooks in place to ensure
that new statements/expressions continue to implement
getSourceRange() and children(). I had to work around a recent clang
bug; dgregor actually fixed it already, but I didn't want to
introduce a selfhosting dependency on ToT.
llvm-svn: 125183
instead from the Scope; Inner scopes in bodies don't have DeclContexts associated with them.
Fixes http://llvm.org/PR9160 & rdar://problem/8966163.
llvm-svn: 125097
a glvalue as a temporary. Previously, we were enumerating all of the
cases that coul return glvalues and might be called with
Sema::MaybeBindToTemporary(), but that was gross and we missed the
Objective-C property reference case.
llvm-svn: 125070
- BlockDeclRefExprs always store VarDecls
- BDREs no longer store copy expressions
- BlockDecls now store a list of captured variables, information about
how they're captured, and a copy expression if necessary
With that in hand, change IR generation to use the captures data in
blocks instead of walking the block independently.
Additionally, optimize block layout by emitting fields in descending
alignment order, with a heuristic for filling in words when alignment
of the end of the block header is insufficient for the most aligned
field.
llvm-svn: 125005
overridden via remapping. Thus, when we create a "virtual" file in the
file manager, we still stat() the real file that lives behind it so
that we can provide proper uniquing based on inodes. This helps keep
the file manager much more consistent.
To take advantage of this when reparsing files in libclang, we disable
the use of the stat() cache when reparsing or performing code
completion, since the stat() cache is very likely to be out of date in
this use case.
llvm-svn: 124971
say "out-of-line definition differ from the declaration in the return type" instead of
the silly "functions that differ only in their return type cannot be overloaded".
Addresses rdar://7980179.
llvm-svn: 124939
when selector metadata is generated, which is triggered
by at least on class implementation. This is to match gcc's
behavior. // rdar://8851684.
llvm-svn: 124909
id-expression, e.g.,
CurrentClass<T>::member
Previously, if CurrentClass<T> was dependent and not complete, we
would treat it as a dependent-scoped declaration reference expression,
even if CurrentClass<T> referred to the current instantiation.
Fixes PR8966 and improves type checking of templates.
llvm-svn: 124867
it's okay for the following template parameters to not have default
arguments (since those template parameters can still be
deduced). Also, downgrade the error about default template arguments
in function templates to an extension warning, since this is a
harmless C++0x extension.
llvm-svn: 124855
The difference with gcc is that it warns if you overload virtual methods only if
the method doesn't also override any method. This is to cut down on the number of warnings
and make it more useful like reported here: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=20423.
If we want to warn that not all overloads are overriden we can have an additional
warning like -Wpartial-override.
-Woverloaded-virtual, unlike gcc, is added to -Wmost. Addresses rdar://8757630.
llvm-svn: 124805
right for anonymous struct/union members led to me discovering some
seemingly broken code in that area of Sema, which I fixed, partly by
changing the representation of member pointer constants so that
IndirectFieldDecls aren't expanded. This led to assorted cleanups with
member pointers in CodeGen, and while I was doing that I saw some random
other things to clean up.
llvm-svn: 124785
is not defined in the current translation unit. Doing so lead to compile errors
such as PR9114.
Instead, when CodeGen is building the vtable, don't try to emit a definition
for functions that aren't marked used in the current translation unit.
Fixes PR9114.
llvm-svn: 124768
savings of 25% sounds impressive, except that this amounted to only
about 360k in our standard "large" completion result set (40,000
results). Since code completion is performance-sensitive, the 4%
slowdown due to uniquing outweighs the 360k benefit.
llvm-svn: 124737
Macros frequently contain extra '()' to make instantiation less error prone.
This warning was flagging a ton of times on postgresql because of its use of macros.
llvm-svn: 124695
BumpPtrAllocator, rather than manually new/delete'ing them. This
optimization also allows us to avoid allocating memory for and copying
constant strings (e.g., "return", "class").
This also required embedding the priority and availability of results
within the code completion string, to avoid extra memory allocation
within libclang.
llvm-svn: 124673
Implicit dtors confuse the ad hoc path-sensitivity of UninitializedValuesV2.cpp. This isn't
the ideal solution, as it will directly impact compile time, but should significantly reduce
the noise of -Wuninitialized on some code bases.
This immediately "fixes" the false positive reported in PR 9063, although this
isn't the right fix in the long run.
llvm-svn: 124667
extremely rambunctious, both on parsing and on template instantiation.
Calm it down, fixing an internal consistency assert on anonymous enum
instantiation manglings.
llvm-svn: 124653
compatible, not having the same type.
Fix rdar://8183908 in which compatible vector types weren't initialized properly leading to a crash.
llvm-svn: 124637
The rationale is that it is highly likely that the user's getter/setter isn't atomically implemented. Off by default.
Addresses rdar://8782645.
-Wcustom-atomic-properties and -Wimplicit-atomic-properties are under the -Watomic-properties group.
llvm-svn: 124609
conversions (<rdar://problem/8592139>) for overload resolution. The
conversion ranking mirrors C++'s conversion ranking fairly closely,
except that we use a same pseudo-subtyping relationship employed by
Objective-C pointer assignment rather than simple checking
derived-to-base conversions. This change covers:
- Conversions to pointers to a specific object type are better than
conversions to 'id', 'Class', qualified 'id', or qualified 'Class'
(note: GCC doesn't perform this ranking, but it matches C++'s rules
for ranking conversions to void*).
- Conversions to qualified 'id' or qualified 'Class' are better than
conversions to 'id' or 'Class', respectively.
- When two conversion sequences convert to the same type, rank the
conversions based on the relationship between the types we're
converting from.
- When two conversion sequences convert from the same non-id,
non-Class type, rank the conversions based on the relationship of
the types we're converting to. (note: GCC allows this ranking even
when converting from 'id', which is extremeley dangerous).
llvm-svn: 124591
allow ref-qualifiers on function types used as template type
arguments. GNU actually allows cv-qualifiers on function types in many
places where it shouldn't, so we currently categorize this as a GNU
extension.
llvm-svn: 124584
current translation unit as available_externally.
This helps devirtualize the second example in PR3100, comment 18:
struct S { S() {}; virtual void xyzzy(); };
inline void foo(S *s) { s->xyzzy(); }
void bar() { S s; foo(&s); }
This involved four major changes:
1. In DefineUsedVTables, always mark virtual member functions as referenced for
non-template classes and class template specializations.
2. In CodeGenVTables::ShouldEmitVTableInThisTU return true if optimizations are
enabled, even if the key function is not implemented in this translation
unit. We don't ever do this for code compiled with -fapple-kext, because we
don't ever want to devirtualize virtual member function calls in that case.
3. Give the correct linkage for vtables where the key function is not defined.
4. Update the linkage for RTTI structures when necessary.
llvm-svn: 124565
might be queried in places where we absolutely require a valid
location (e.g., for template instantiation). Fixes some major
brokenness in the use of __is_convertible_to.
llvm-svn: 124465
access control errors into SFINAE errors, so that the trait provides
enough support to implement the C++0x std::is_convertible type trait.
To get there, the SFINAETrap now knows how to set up a SFINAE context
independent of any template instantiations or template argument
deduction steps, and (separately) can set a Sema flag to translate
access control errors into SFINAE errors. The latter can also be
useful if we decide that access control errors during template argument
deduction should cause substitution failure (rather than a hard error)
as has been proposed for C++0x.
llvm-svn: 124446
semantics after the C++0x is_convertible type trait. This
implementation is not 100% complete, because it allows access errors
to be hard errors (rather than just evaluating false).
Original patch by Steven Watanabe!
llvm-svn: 124425
deallocation function has a two-argument form. Store the result of this
check in new[] and delete[] nodes.
Fixes rdar://problem/8913519
llvm-svn: 124373
argument but doesn't (because previous template parameters had default
arguments), clear out all of the default arguments so that we maintain
the invariant that a template parameter has a default argument only if
subsequence template parameters also have default arguments.
Fixes a crash-on-invalid <rdar://problem/8913649>.
llvm-svn: 124345
derived-to-base cast that also casts away constness (one of the cases
for static_cast followed by const_cast) would be treated as a bit-cast
rather than a derived-to-base class, causing miscompiles and
heartburn.
Fixes <rdar://problem/8913298>.
llvm-svn: 124340
overload a function without a ref-qualifier (C++0x
[over.load]p2). This, apparently, completes the implementation of
rvalue references for *this.
llvm-svn: 124321
reference binding is for the implicit object parameter of a member
function with a ref-qualifier. My previous comment, that we didn't
need to track this explicitly, was wrong: we do in fact get
rvalue-references-prefer-rvalues overloading with ref-qualifiers.
llvm-svn: 124313
the presence and form of a ref-qualifier. Note that we do *not* yet
implement the restriction in C++0x [over.load]p2 that requires either
all non-static functions with a given parameter-type-list to have a
ref-qualifier or none of them to have a ref-qualifier.
llvm-svn: 124297
- Add ref-qualifiers to the type system; they are part of the
canonical type. Print & profile ref-qualifiers
- Translate the ref-qualifier from the Declarator chunk for
functions to the function type.
- Diagnose mis-uses of ref-qualifiers w.r.t. static member
functions, free functions, constructors, destructors, etc.
- Add serialization and deserialization of ref-qualifiers.
llvm-svn: 124281
after a 'return' in a CFGBlock. This accidentally
was working before, but the false assumption that
'return' always appeared at the end of the block
was uncovered by a recent change.
llvm-svn: 124280
handling all CFGElement kinds. While writing
the test case, it turned out that return-noreturn.cpp
wasn't actually testing anything since it has the wrong -W
flag. That uncovered another regression with
the handling of destructors marked noreturn. WIP.
llvm-svn: 124238
for reference binding (C++ [over.rank.ics]p3b1sb4), so that we prefer
the binding of an lvalue reference to a function lvalue over the
binding of an rvalue reference. This change resolves the ambiguity
with std::forward and lvalue references to function types in a way
that seems consistent with the original rvalue references proposal.
My proposed wording for this change is shown in
isBetterReferenceBindingKind(); we'll try to get this change adopted
in the C++0x working paper as well.
llvm-svn: 124236
authors to write
class __attribute__((forbid_temporaries)) Name { ... };
when they want to force users to name all variables of the type. This protects
people from doing things like creating a scoped_lock that only lives for a
single statement instead of an entire scope.
The warning produced by this attribute can be disabled by
-Wno-forbid-temporaries.
llvm-svn: 124217
(C++0x [over.ics.rank]p3) when one binding is an lvalue reference and
the other is an rvalue reference that binds to an rvalue. In
particular, we were using the predict "is an rvalue reference" rather
than "is an rvalue reference that binds to an rvalue", which was
incorrect in the one case where an rvalue reference can bind to an
lvalue: function references.
This particular issue cropped up with std::forward, where Clang was
picking an std::forward overload while forwarding an (lvalue)
reference to a function. However (and unfortunately!), the right
answer for this code is that the call to std::forward is
ambiguous. Clang now gets that right, but we need to revisit the
std::forward implementation in libc++.
llvm-svn: 124216
generate meaningful [*] template argument location information.
[*] Well, as meaningful as possible, given that this entire code path
is a hack for when we've lost type-source information.
llvm-svn: 124211
during template instantiation. This code needs to eventually die, but
this little tweak fixes PR8629, where bad location information slipped
through to the location of a class template instantiation.
llvm-svn: 124199
T) when taking the address of an overloaded function or matching a
specialization to a template (C++0x [temp.deduct.type]p10). Fixes
PR9044.
llvm-svn: 124197
implementation used by overload resolution to support rvalue
references. The original commits caused PR9026 and some
hard-to-reproduce self-host breakage.
The only (crucial!) difference between this commit and the previous
commits is that we now properly check the SuppressUserConversions flag
before attempting to perform a second user-defined conversion in
reference binding, breaking the infinite recursion chain of
user-defined conversions.
Rvalue references should be working a bit better now.
llvm-svn: 124121
error: no super class declared in @interface for 'XXX'
to be:
error: 'X' cannot use 'super' because it is a root class
The latter explains what the user actually did wrong.
Fixes: <rdar://problem/8904409>
llvm-svn: 124074
clang's -Wuninitialized-experimental warning.
While these don't look like real bugs, clang's
-Wuninitialized-experimental analysis is stricter
than GCC's, and these fixes have the benefit
of being general nice cleanups.
llvm-svn: 124072