type trait. The previous implementation suffered from several problems:
1) It implemented all of the logic in RecordType by walking over every
base and field in a CXXRecordDecl and validating the constraints of
the standard. This made for very straightforward code, but is
extremely inefficient. It also is conceptually wrong, the logic tied
to the C++ definition of standard-layout classes should be in
CXXRecordDecl, not RecordType.
2) To address the performance problems with #1, a cache bit was added to
CXXRecordDecl, and at the completion of every C++ class, the
RecordType was queried to determine if it was a standard layout
class, and that state was cached. Two things went very very wrong
with this. First, the caching version of the query *was never
called*. Even within the recursive steps of the walk over all fields
and bases the caching variant was not called, making each query
a full *recursive* walk. Second, despite the cache not being used, it
was computed for every class declared, even when the trait was never
used in the program. This probably significantly regressed compile
time performance for edge-case files.
3) An ASTContext was required merely to query the type trait because
querying it performed the actual computations.
4) The caching bit wasn't managed correctly (uninitialized).
The new implementation follows the system for all the other traits on
C++ classes by encoding all the state needed in the definition data and
building up the trait incrementally as each base and member are added to
the definition of the class.
The idiosyncracies of the specification of standard-layout classes
requires more state than I would like; currently 5 bits. I could
eliminate one of the bits easily at the expense of both clarity and
resilience of the code. I might be able to eliminate one of the other
bits by computing its state in terms of other state bits in the
definition. I've already done that in one place where there was a fairly
simple way to achieve it.
It's possible some of the bits could be moved out of the definition data
and into some other structure which isn't serialized if the serialized
bloat is a problem. That would preclude serialization of a partial class
declaration, but that's likely already precluded.
Comments on any of these issues welcome.
llvm-svn: 130601
operators in C++ record declarations.
This patch starts off by updating a bunch of the standard citations to
refer to the draft 0x standard so that the semantics intended for move
varianst is clear. Where necessary these are duplicated so they'll be
available in doxygen.
It adds bit fields to keep track of the state for the move constructs,
and updates all the code necessary to track this state (I think) as
members are declared for a class. It also wires the state into the
various trait-like accessors in the AST's API, and tests that the type
trait expressions now behave correctly in the presence of move
constructors and move assignment operators.
This isn't complete yet due to these glaring FIXMEs:
1) No synthesis of implicit move constructors or assignment operators.
2) I don't think we correctly enforce the new logic for both copy and
move trivial checks: that the *selected* copy/move
constructor/operator is trivial. Currently this requires *all* of them
to be trivial.
3) Some of the trait logic needs to be folded into the fine-grained
trivial bits to more closely match the wording of the standard. For
example, many of the places we currently set a bit to track POD-ness
could be removed by querying other more fine grained traits on
demand.
llvm-svn: 130076
gcc's unused warnings which don't get emitted if the function is referenced even in an unevaluated context
(e.g. in templates, sizeof, etc.). Also, saying that a function is 'unused' because it won't get codegen'ed
is somewhat misleading.
- Don't emit 'unused' warnings for functions that are referenced in any part of the user's code.
- A warning that an internal function/variable won't get emitted is useful though, so introduce
-Wunneeded-internal-declaration which will warn if a function/variable with internal linkage is not
"needed" ('used' from the codegen perspective), e.g:
static void foo() { }
template <int>
void bar() {
foo();
}
test.cpp:1:13: warning: function 'foo' is not needed and will not be emitted
static void foo() { }
^
Addresses rdar://8733476.
llvm-svn: 129794
use the translation unit as its declaration context, then deserialize
the actual lexical and semantic DeclContexts after the template
parameter is complete. This avoids problems when the DeclContext
itself (e.g., a class template) is dependent on the template parameter
(e.g., for the injected-class-name).
llvm-svn: 127056
DeclContext once we've created it. This mirrors what we do for
function parameters, where the parameters start out with
translation-unit context and then are adopted by the appropriate
DeclContext when it is created. Also give template parameters public
access and make sure that they don't show up for the purposes of name
lookup.
Fixes PR9400, a regression introduced by r126920, which implemented
substitution of default template arguments provided in template
template parameters (C++ core issue 150).
How on earth could the DeclContext of a template parameter affect the
handling of default template arguments?
I'm so glad you asked! The link is
Sema::getTemplateInstantiationArgs(), which determines the outer
template argument lists that correspond to a given declaration. When
we're instantiating a default template argument for a template
template parameter within the body of a template definition (not it's
instantiation, per core issue 150), we weren't getting any outer
template arguments because the context of the template template
parameter was the translation unit. Now that the context of the
template template parameter is its owning template, we get the
template arguments from the injected-class-name of the owning
template, so substitution works as it should.
llvm-svn: 127004
UnresolvedUsingValueDecl to use NestedNameSpecifierLoc rather than the
extremely-lossy NestedNameSpecifier/SourceRange pair it used to use,
improving source-location information.
Various infrastructure updates to support NestedNameSpecifierLoc:
- AST/PCH (de-)serialization
- Recursive AST visitor
- libclang traversal (including the first tests of this
functionality)
llvm-svn: 126459
making them be template instantiated in a more normal way and
make them handle attributes like other decls.
This fixes the used/unused label handling stuff, making it use
the same infrastructure as other decls.
llvm-svn: 125771
LabelDecl and LabelStmt. There is a 1-1 correspondence between the
two, but this simplifies a bunch of code by itself. This is because
labels are the only place where we previously had references to random
other statements, causing grief for AST serialization and other stuff.
This does cause one regression (attr(unused) doesn't silence unused
label warnings) which I'll address next.
This does fix some minor bugs:
1. "The only valid attribute " diagnostic was capitalized.
2. Various diagnostics printed as ''labelname'' instead of 'labelname'
3. This reduces duplication of label checking between functions and blocks.
Review appreciated, particularly for the cindex and template bits.
llvm-svn: 125733
Temporarily set the first (canonical) declaration as the previous one, which is the one that
matters, and mark the real previous DeclID to be loaded & attached later on.
Fixes rdar://8956193.
llvm-svn: 125434
- 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
Inheritable attributes on declarations may be inherited by any later
redeclaration at merge time. By contrast, a non-inheritable attribute
will not be inherited by later redeclarations. Non-inheritable
attributes may be semantically analysed early, allowing them to
influence the redeclaration/overloading process.
Before this change, the "overloadable" attribute received special
handling to be treated as non-inheritable, while all other attributes
were treated as inheritable. This patch generalises the concept,
while removing a FIXME. Some CUDA location attributes are also marked
as non-inheritable in order to support special overloading semantics
(to be introduced in a later patch).
The patch introduces a new Attr subclass, InheritableAttr, from
which all inheritable attributes derive. Non-inheritable attributes
simply derive from Attr.
N.B. I did not review every attribute to determine whether it should
be marked non-inheritable. This can be done later on an incremental
basis, as this change does not affect default functionality.
llvm-svn: 123959
a pack expansion, e.g., the parameter pack Values in:
template<typename ...Types>
struct Outer {
template<Types ...Values>
struct Inner;
};
This new implementation approach introduces the notion of an
"expanded" non-type template parameter pack, for which we have already
expanded the types of the parameter pack (to, say, "int*, float*",
for Outer<int*, float*>) but have not yet expanded the values. Aside
from creating these expanded non-type template parameter packs, this
patch updates template argument checking and non-type template
parameter pack instantiation to make use of the appropriate types in
the parameter pack.
llvm-svn: 123845
FunctionDecl::setPure crashed a poor user's code.
Remove the use of this accessor when deserializing, along with several other in the neighborhood. Fixes rdar://8759653.
llvm-svn: 122756
packs, e.g.,
template<typename T, unsigned ...Dims> struct multi_array;
along with semantic analysis support for finding unexpanded non-type
template parameter packs in types, expressions, and so on.
Template instantiation involving non-type template parameter packs
probably doesn't work yet. That'll come soon.
llvm-svn: 122527
common base for ExtQuals and Type that stores the underlying type
pointer. This results in a 2% performance win for -emit-llvm on a
typical C file, with 1% memory growth in the AST.
Note that there is an API change in this optimization:
QualType::getTypePtr() can no longer be invoked on a NULL
QualType. If the QualType might be NULL, use
QualType::getTypePtrOrNull(). I've audited all uses of getTypePtr() in
the code base and changed the appropriate uses over to
getTypePtrOrNull().
A future optimization opportunity would be to distinguish between
cast/dyn_cast and cast_or_null/dyn_cast_or_null; for the former, we
could use getTypePtr() rather than getTypePtrOrNull(), to take another
branch out of the cast/dyn_cast implementation.
llvm-svn: 121489
"inline", we weren't giving the definition weak linkage because the
"inline" bit wasn't propagated. This was a longstanding FIXME that,
somehow, hadn't triggered a bug in the wild. Fix this problem by
tracking whether any declaration was marked "inline", and clean up the
semantics of GNU's "extern inline" semantics calculation based on this
change.
Fixes <rdar://problem/8740363>.
llvm-svn: 121373
My previous attempt at solving the compile-time problem with many
redeclarations of the same entity cached both linkage and visibility,
while this patch only tackles linkage. There are several reasons for
this difference:
- Linkage is a language concept, and is evaluated many times during
semantic analysis and codegen, while visibility is only a
code-generation concept that is evaluated only once per (unique)
declaration. Hence, we *must* optimize linkage calculations but
don't need to optimize visibility computation.
- Once we know the linkage of a declaration, subsequent
redeclarations can't change that linkage. Hence, cache
invalidation is far simpler than for visibility, where a later
redeclaration can completely change the visibility.
- We have 3 spare bits in Decl to store the linkage cache, so the
cache doesn't increase the size of declarations. With the
visibility+linkage cache, NamedDecl got larger.
llvm-svn: 121023
and visibility of declarations, because it was extremely messy and it
increased the size of NamedDecl.
An improved implementation is forthcoming.
llvm-svn: 121012
declarations.
The motivation for this patch is that linkage/visibility computations
are linear in the number of redeclarations of an entity, and we've run
into a case where a single translation unit has > 6500 redeclarations
of the same (unused!) external variable. Since each redeclaration
involves a linkage check, the resulting quadratic behavior makes Clang
slow to a crawl. With this change, a simple test with 512
redeclarations of a variable syntax-checks ~20x faster than
before.
That said, I hate this change, and will probably end up reverting it
in a few hours. Reasons to hate it:
- It makes NamedDecl larger, since we don't have enough free bits in
Decl to squeeze in the extra information about caching.
- There are way too many places where we need to invalidate this
cache, because the visibility of a declaration can change due to
redeclarations (!). Despite self-hosting and passing the testsuite,
I have no confidence that I've found all of places where this cache
needs to be invalidated.
llvm-svn: 120808
A new AST node is introduced:
def IndirectField : DDecl<Value>;
IndirectFields are injected into the anonymous's parent scope and chain back to
the original field. Name lookup for anonymous entities now result in an
IndirectFieldDecl instead of a FieldDecl.
There is no functionality change, the code generated should be the same.
llvm-svn: 119919
