The macros required for DeclNodes use have changed to match the use of
StmtNodes. The FooFirst enumerator constants have been named firstFoo
to match usage elsewhere.
llvm-svn: 105165
isn't possible to compute.
This patch is mostly refactoring; the key change is the addition of the code
starting with the comment, "Check whether the function has a computable LLVM
signature." The solution here is essentially the same as the way the
vtable code handles such functions.
llvm-svn: 105151
This class only supports name mangling (which is apparently used during C/ObjC
codegen). For now only the Itanium C++ ABI is supported. Patches to add a
second C++ ABI are forthcoming.
llvm-svn: 104630
variables within blocks. We loosely follow GCC's mangling, but since
these are always internal symbols the names don't really matter. I
intend to revisit block mangling later, because GCC's mangling is
rather verbose. <rdar://problem/8015719>.
llvm-svn: 104610
variables should have that linkage. Otherwise, its static local
variables should have internal linkage. To avoid computing this excessively,
set a function's linkage before we emit code for it.
Previously we were assigning weak linkage to the static variables of
static inline functions in C++, with predictably terrible results. This
fixes that and also gives better linkage than 'weak' when merging is required.
llvm-svn: 104581
This works around a crash where malloc reused the memory of an erased BB for a
new BB leaving old cleanup information pointing at the new block.
llvm-svn: 104472
temporaries. There are actually several interrelated fixes here:
- When converting an object to a base class, it's only an lvalue
cast when the original object was an lvalue and we aren't casting
pointer-to-derived to pointer-to-base. Previously, we were
misclassifying derived-to-base casts of class rvalues as lvalues,
causing various oddities (including problems with reference binding
not extending the lifetimes of some temporaries).
- Teach the code for emitting a reference binding how to look
through no-op casts and parentheses directly, since
Expr::IgnoreParenNoOpCasts is just plain wrong for this. Also, make
sure that we properly look through multiple levels of indirection
from the temporary object, but destroy the actual temporary object;
this fixes the reference-binding issue mentioned above.
- Teach Objective-C message sends to bind the result as a temporary
when needed. This is actually John's change, but it triggered the
reference-binding problem above, so it's included here. Now John
can actually test his return-slot improvements.
llvm-svn: 104434
not make copies non-POD arguments or arguments passed by reference:
just copy the pointers directly. This eliminates another source of the
dreaded memcpy-of-non-PODs. Fixes PR7188.
llvm-svn: 104327
'self' variable arising from uses of the 'super' keyword. Also reorganize
some code so that BlockInfo (now CGBlockInfo) can be opaque outside of
CGBlocks.cpp.
Fixes rdar://problem/8010633.
llvm-svn: 104312
particular issue was the cause of the Boost.Interprocess failures, and
in general will lead to horrendous, hard-to-diagnose miscompiles. The
assertion itself has survives self-host and a full Boost build, so we
are close to eradicating this problem in C++.
Note that the assertion is *not* turned on for Objective-C++, where we
still have problems with introducing memcpy's of non-POD class
types. That part of the assertion will go away as soon as we fix the
known issues in Objective-C++.
llvm-svn: 104227
subobject. Previously, we could only properly bind to a base class
subobject while extending the lifetime of the complete object (of a
derived type); for non-static data member subobjects, we could memcpy
(!) the result and bind to that, which is rather broken.
Now, we pull apart the expression that we're binding to, to figure out
which subobject we're accessing, then construct the temporary object
(adding a destruction if needed) and, finally, dig out the subobject
we actually meant to access.
This fixes yet another instance where we were memcpy'ing rather than
doing the right thing. However, note the FIXME in references.cpp:
there's more work to be done for binding to subobjects, since the AST
is incorrectly modeling some member accesses in base classes as
lvalues when they are really rvalues.
llvm-svn: 104219
class type (that uses a return slot), pass the return slot to the
callee directly rather than allocating new storage and trying to copy
the object. This appears to have been the cause of the remaining two
Boost.Interprocess failures.
llvm-svn: 104215
This is still probably wrong for Objective-C++ and adds a couple of lines in CGException that should probably be in the CGObjCRuntime subclass. The personality function is now only looked up in one place in CGException though, so this should be easier to fix in the future.
llvm-svn: 103938
__cxa_guard_abort along the exceptional edge into (in effect) a nested
"try" that rethrows after aborting. Fixes PR7144 and the remaining
Boost.ProgramOptions failures, along with the regressions that r103880
caused.
The crucial difference between this and r103880 is that we now follow
LLVM's little dance with the llvm.eh.exception and llvm.eh.selector
calls, then use _Unwind_Resume_or_Rethrow to rethrow.
llvm-svn: 103892
__cxa_guard_abort along the exceptional edge into (in effect) a nested
"try" that rethrows after aborting. Fixes PR7144 and the remaining
Boost.ProgramOptions failures.
llvm-svn: 103880
ObjCObjectType, which is basically just a pair of
one of {primitive-id, primitive-Class, user-defined @class}
with
a list of protocols.
An ObjCObjectPointerType is therefore just a pointer which always points to
one of these types (possibly sugared). ObjCInterfaceType is now just a kind
of ObjCObjectType which happens to not carry any protocols.
Alter a rather large number of use sites to use ObjCObjectType instead of
ObjCInterfaceType. Store an ObjCInterfaceType as a pointer on the decl rather
than hashing them in a FoldingSet. Remove some number of methods that are no
longer used, at least after this patch.
By simplifying ObjCObjectPointerType, we are now able to easily remove and apply
pointers to Objective-C types, which is crucial for a certain kind of ObjC++
metaprogramming common in WebKit.
llvm-svn: 103870
return statements. We perform NRVO only when all of the return
statements in the function return the same variable. Fixes some link
failures in Boost.Interprocess (which is relying on NRVO), and
probably improves performance for some C++ applications.
llvm-svn: 103867
return value optimization. Sema marks return statements with their
NRVO candidates (which may or may not end up using the NRVO), then, at
the end of a function body, computes and marks those variables that
can be allocated into the return slot.
I've checked this locally with some debugging statements (not
committed), but there won't be any tests until CodeGen comes along.
llvm-svn: 103865
"return" statement and mark the corresponding CXXConstructExpr as
elidable. Teach CodeGen that eliding a temporary is different from
eliding an object construction.
This is just a baby step toward NRVO.
llvm-svn: 103849
throw, it should use invoke when needed. The fixes the
Boost.Statechrt failures that motivated PR7132, but there are a few
side issues to tackle as well.
llvm-svn: 103803
methods for which the key function is guaranteed to be in another
translation unit. Unfortunately, this guarantee isn't the case when
dealing with shared libraries that fail to export these virtual method
definitions.
I'm reopening PR6747 so we can consider this again at a later point in
time.
llvm-svn: 103741
"used" (e.g., we will refer to the vtable in the generated code) and
when they are defined (i.e., because we've seen the key function
definition). Previously, we were effectively tracking "potential
definitions" rather than uses, so we were a bit too eager about emitting
vtables for classes without key functions.
The new scheme:
- For every use of a vtable, Sema calls MarkVTableUsed() to indicate
the use. For example, this occurs when calling a virtual member
function of the class, defining a constructor of that class type,
dynamic_cast'ing from that type to a derived class, casting
to/through a virtual base class, etc.
- For every definition of a vtable, Sema calls MarkVTableUsed() to
indicate the definition. This happens at the end of the translation
unit for classes whose key function has been defined (so we can
delay computation of the key function; see PR6564), and will also
occur with explicit template instantiation definitions.
- For every vtable defined/used, we mark all of the virtual member
functions of that vtable as defined/used, unless we know that the key
function is in another translation unit. This instantiates virtual
member functions when needed.
- At the end of the translation unit, Sema tells CodeGen (via the
ASTConsumer) which vtables must be defined (CodeGen will define
them) and which may be used (for which CodeGen will define the
vtables lazily).
From a language perspective, both the old and the new schemes are
permissible: we're allowed to instantiate virtual member functions
whenever we want per the standard. However, all other C++ compilers
were more lazy than we were, and our eagerness was both a performance
issue (we instantiated too much) and a portability problem (we broke
Boost test cases, which now pass).
Notes:
(1) There's a ton of churn in the tests, because the order in which
vtables get emitted to IR has changed. I've tried to isolate some of
the larger tests from these issues.
(2) Some diagnostics related to
implicitly-instantiated/implicitly-defined virtual member functions
have moved to the point of first use/definition. It's better this
way.
(3) I could use a review of the places where we MarkVTableUsed, to
see if I missed any place where the language effectively requires a
vtable.
Fixes PR7114 and PR6564.
llvm-svn: 103718
This fixes recent regressions reported by gdb testsuite.
Tighter verification of debug info generated by FE found these regressions.
Refactor code to extract line number and column number from SourceLocation.
llvm-svn: 103678
available_externally linkage, since they may not have been given a
strong definition in another translation unit. Without this patch, the
following test case fails to link with a GCC-compiled libstdc++:
#include <sstream>
int main() { std::basic_stringbuf<char> bs; }
Fixes the last problem with the Boost.IO library.
llvm-svn: 103208
function attributes like byval get applied to the function
definition. This fixes PR7058 and makes i386 llvm/clang bootstrap
pass all the same tests as x86-64 bootstrap for me (the llvmc
tests still fail in both).
llvm-svn: 103131
reference type, make sure that the initializer we build is the
of the appropriate type for the *reference*, not for the thing that it
refers to. Fixes PR7050.
llvm-svn: 103115
destructors, place the __cxa_atexit call after the __cxa_guard_release
call, mimicking GCC/LLVM-GCC behavior. Noticed while debugging
something related.
llvm-svn: 103088
implicitly-generated copy constructor. Previously, Sema would perform
some checking and instantiation to determine which copy constructors,
etc., would be called, then CodeGen would attempt to figure out which
copy constructor to call... but would get it wrong, or poke at an
uninstantiated default argument, or fail in other ways.
The new scheme is similar to what we now do for the implicit
copy-assignment operator, where Sema performs all of the semantic
analysis and builds specific ASTs that look similar to the ASTs we'd
get from explicitly writing the copy constructor, so that CodeGen need
only do a direct translation.
However, it's not quite that simple because one cannot explicit write
elementwise copy-construction of an array. So, I've extended
CXXBaseOrMemberInitializer to contain a list of indexing variables
used to copy-construct the elements. For example, if we have:
struct A { A(const A&); };
struct B {
A array[2][3];
};
then we generate an implicit copy assignment operator for B that looks
something like this:
B::B(const B &other) : array[i0][i1](other.array[i0][i1]) { }
CodeGen will loop over the invented variables i0 and i1 to visit all
elements in the array, so that each element in the destination array
will be copy-constructed from the corresponding element in the source
array. Of course, if we're dealing with arrays of scalars or class
types with trivial copy-assignment operators, we just generate a
memcpy rather than a loop.
Fixes PR6928, PR5989, and PR6887. Boost.Regex now compiles and passes
all of its regression tests.
Conspicuously missing from this patch is handling for the exceptional
case, where we need to destruct those objects that we have
constructed. I'll address that case separately.
llvm-svn: 103079
they're unreachable. This matters because (if they're POD, or if this is C)
the scope containing the variable might be reachable even if the variable
isn't. Fixes PR7044.
llvm-svn: 103052
typedef int functype(int, int);
functype func;
also instantiate the synthesized function parameters for the resulting
function declaration.
With this change, Boost.Wave builds and passes all of its regression
tests.
llvm-svn: 103025
not just the inner expression. This is important if the expression has any
temporaries. Fixes PR 7028.
Basically a symptom of really tragic method names.
llvm-svn: 102998
assignment operators.
Previously, Sema provided type-checking and template instantiation for
copy assignment operators, then CodeGen would synthesize the actual
body of the copy constructor. Unfortunately, the two were not in sync,
and CodeGen might pick a copy-assignment operator that is different
from what Sema chose, leading to strange failures, e.g., link-time
failures when CodeGen called a copy-assignment operator that was not
instantiation, run-time failures when copy-assignment operators were
overloaded for const/non-const references and the wrong one was
picked, and run-time failures when by-value copy-assignment operators
did not have their arguments properly copy-initialized.
This implementation synthesizes the implicitly-defined copy assignment
operator bodies in Sema, so that the resulting ASTs encode exactly
what CodeGen needs to do; there is no longer any special code in
CodeGen to synthesize copy-assignment operators. The synthesis of the
body is relatively simple, and we generate one of three different
kinds of copy statements for each base or member:
- For a class subobject, call the appropriate copy-assignment
operator, after overload resolution has determined what that is.
- For an array of scalar types or an array of class types that have
trivial copy assignment operators, construct a call to
__builtin_memcpy.
- For an array of class types with non-trivial copy assignment
operators, synthesize a (possibly nested!) for loop whose inner
statement calls the copy constructor.
- For a scalar type, use built-in assignment.
This patch fixes at least a few tests cases in Boost.Spirit that were
failing because CodeGen picked the wrong copy-assignment operator
(leading to link-time failures), and I suspect a number of undiagnosed
problems will also go away with this change.
Some of the diagnostics we had previously have gotten worse with this
change, since we're going through generic code for our
type-checking. I will improve this in a subsequent patch.
llvm-svn: 102853
Amadini.
This change introduces a new expression node type, OffsetOfExpr, that
describes __builtin_offsetof. Previously, __builtin_offsetof was
implemented using a unary operator whose subexpression involved
various synthesized array-subscript and member-reference expressions,
which was ugly and made it very hard to instantiate as a
template. OffsetOfExpr represents the AST more faithfully, with proper
type source information and a more compact representation.
OffsetOfExpr also has support for dependent __builtin_offsetof
expressions; it can be value-dependent, but will never be
type-dependent (like sizeof or alignof). This commit introduces
template instantiation for __builtin_offsetof as well.
There are two major caveats to this patch:
1) CodeGen cannot handle the case where __builtin_offsetof is not a
constant expression, so it produces an error. So, to avoid
regressing in C, we retain the old UnaryOperator-based
__builtin_offsetof implementation in C while using the shiny new
OffsetOfExpr implementation in C++. The old implementation can go
away once we have proper CodeGen support for this case, which we
expect won't cause much trouble in C++.
2) __builtin_offsetof doesn't work well with non-POD class types,
particularly when the designated field is found within a base
class. I will address this in a subsequent patch.
Fixes PR5880 and a bunch of assertions when building Boost.Python
tests.
llvm-svn: 102542
Emitted some metadata on message sends to allow a later pass to do some speculative inlining of class methods (GNU runtime). Speculative inlining of instance methods requires type feedback to be useful (work in progress), but for class methods it works quite nicely.
llvm-svn: 102514
T::template apply<U>), handling a few cases where we previously failed
and performing substitutions on such dependent names. Fixes a crash in
Boost.PropertyTree.
llvm-svn: 102490
This works around stack corruption / crashes resulting from PR6944, and also
works around people who expect 'what works on my machine' to work everywhere
(GCC crashes in a number of cases on SPARC that should now work correctly with
clang).
llvm-svn: 102430
of a class template or class template partial specialization. That is to
say, in
template <class T> class A { ... };
or
template <class T> class B<const T*> { ... };
make 'A<T>' and 'B<const T*>' sugar for the corresponding InjectedClassNameType
when written inside the appropriate context. This allows us to track the
current instantiation appropriately even inside AST routines. It also allows
us to compute a DeclContext for a type much more efficiently, at some extra
cost every time we write a template specialization (which can be optimized,
but I've left it simple in this patch).
llvm-svn: 102407
- Fix some places that had the alignment hard coded.
- Use ABI type alignment, not preferred type alignment -- neither of this is exactly right, as we really want the C type alignment as required by the runtime, but the ABI alignment is a more correct choice.
This should be equivalent for x86_64, but fixes the alignment for ARM.
llvm-svn: 102314
- Replace -cc1 level -fobjc-legacy-dispatch with -fobjc-dispatch-method={legacy,non-legacy,mixed}.
- Lift "mixed" vs "non-mixed" policy choice up to driver level, instead of being buried in CGObjCMac.cpp.
- No intended functionality change.
llvm-svn: 102255
statements. Instead of the @try having a single @catch, where all of
the @catch's were chained (using an O(n^2) algorithm nonetheless),
@try just holds an array of its @catch blocks. The resulting AST is
slightly more compact (not important) and better represents the actual
language semantics (good).
llvm-svn: 102221
input and output types when the smaller value isn't mentioned in the
asm string. Extend this support from integers to also allowing
fp values to be mismatched (if not mentioned in the asm string).
llvm-svn: 102188
T::apply <U>::type
Fixes PR6899, although I want to dig a little deeper into the FIXME
for dependent template names that refer to operators.
llvm-svn: 102167
- This fixes some pedantic bugs with packed structures, as well as major problems with -fno-bitfield-type-align.
- Fixes PR5591, PR5567, and all known -fno-bitfield-type-align issues.
- Review appreciated.
llvm-svn: 102045
in a throw expression. Use EmitAnyExprToMem to emit the throw expression,
which magically elides the final copy-constructor call (which raises a new
strict-compliance bug, but baby steps). Give __cxa_throw a destructor pointer
if the exception type has a non-trivial destructor.
llvm-svn: 102039
(if there's a current block). The chief advantage of doing this is that it
lets us pick blocks (e.g. EH blocks) to push to the end of the function so
that fallthrough happens consistently --- i.e. it gives us the flexibility
of ordering blocks as we please without having to change the order in which
we generate code. There are standard (?) optimization passes which can do some
of that for us, but better to generate reasonable code to begin with.
llvm-svn: 101997
expressions, to improve source-location information, clarify the
actual receiver of the message, and pave the way for proper C++
support. The ObjCMessageExpr node represents four different kinds of
message sends in a single AST node:
1) Send to a object instance described by an expression (e.g., [x method:5])
2) Send to a class described by the class name (e.g., [NSString method:5])
3) Send to a superclass class (e.g, [super method:5] in class method)
4) Send to a superclass instance (e.g., [super method:5] in instance method)
Previously these four cases where tangled together. Now, they have
more distinct representations. Specific changes:
1) Unchanged; the object instance is represented by an Expr*.
2) Previously stored the ObjCInterfaceDecl* referring to the class
receiving the message. Now stores a TypeSourceInfo* so that we know
how the class was spelled. This both maintains typedef information
and opens the door for more complicated C++ types (e.g., dependent
types). There was an alternative, unused representation of these
sends by naming the class via an IdentifierInfo *. In practice, we
either had an ObjCInterfaceDecl *, from which we would get the
IdentifierInfo *, or we fell into the case below...
3) Previously represented by a class message whose IdentifierInfo *
referred to "super". Sema and CodeGen would use isStr("super") to
determine if they had a send to super. Now represented as a
"class super" send, where we have both the location of the "super"
keyword and the ObjCInterfaceDecl* of the superclass we're
targetting (statically).
4) Previously represented by an instance message whose receiver is a
an ObjCSuperExpr, which Sema and CodeGen would check for via
isa<ObjCSuperExpr>(). Now represented as an "instance super" send,
where we have both the location of the "super" keyword and the
ObjCInterfaceDecl* of the superclass we're targetting
(statically). Note that ObjCSuperExpr only has one remaining use in
the AST, which is for "super.prop" references.
The new representation of ObjCMessageExpr is 2 pointers smaller than
the old one, since it combines more storage. It also eliminates a leak
when we loaded message-send expressions from a precompiled header. The
representation also feels much cleaner to me; comments welcome!
This patch attempts to maintain the same semantics we previously had
with Objective-C message sends. In several places, there are massive
changes that boil down to simply replacing a nested-if structure such
as:
if (message has a receiver expression) {
// instance message
if (isa<ObjCSuperExpr>(...)) {
// send to super
} else {
// send to an object
}
} else {
// class message
if (name->isStr("super")) {
// class send to super
} else {
// send to class
}
}
with a switch
switch (E->getReceiverKind()) {
case ObjCMessageExpr::SuperInstance: ...
case ObjCMessageExpr::Instance: ...
case ObjCMessageExpr::SuperClass: ...
case ObjCMessageExpr::Class:...
}
There are quite a few places (particularly in the checkers) where
send-to-super is effectively ignored. I've placed FIXMEs in most of
them, and attempted to address send-to-super in a reasonable way. This
could use some review.
llvm-svn: 101972
This mirror's Dan's patch for llvm-gcc in r97989, and
fixes the miscompilation in PR6525. There is some contention
over whether this is the right thing to do, but it is the
conservative answer and demonstrably fixes a miscompilation.
llvm-svn: 101877
function declaration, since it may end up being changed (e.g.,
"extern" can become "static" if a prior declaration was static). Patch
by Enea Zaffanella and Paolo Bolzoni.
llvm-svn: 101826
struct may cause it to shrink more than one byte. Before
my recent changes we compiled the new test into:
%0 = type { [6 x i8] }
@x = global %0 { [6 x i8] undef }, align 2 ; <%0*> [#uses=0]
which is obviously bogus. Now we compile it into:
%0 = type <{ i32, i8, i8 }>
@x = global %0 zeroinitializer, align 2 ; <%0*> [#uses=0]
Where the last byte only is tail padding.
llvm-svn: 101536
- Sadly, this doesn't seem to give any .ll size win so far. It is possible to make this routine significantly smarter & avoid various shifting, masking, and zext/sext, but I'm not really convinced it is worth it. It is tricky, and this is really instcombine's job.
- No intended functionality change; the test case is just to increase coverage & serves as a demo file, it worked before this commit.
The new fixes from r101222 are:
1. The shift to the target position needs to occur after the value is extended to the correct size. This broke Clang bootstrap, among other things no doubt.
2. Swap the order of arguments to OR, to get a tad more constant folding.
llvm-svn: 101339
Stop multiplying constant by 8 accordingly in the header and change
intrinsic definition for what types we expect.
Add to existing palignr test to check that we're emitting the correct things.
llvm-svn: 101332
- Sadly, this doesn't seem to give any .ll size win so far. It is possible to make this routine significantly smarter & avoid various shifting, masking, and zext/sext, but I'm not really convinced it is worth it. It is tricky, and this is really instcombine's job.
- No intended functionality change; the test case is just to increase coverage & serves as a demo file, it worked before this commit.
llvm-svn: 101222
- This lets the method focus slightly more on emitting clean IR to honor the policy which has been selected. On 403.gcc's combine.c, x86_64, -O0, this reduces the number of lines in the .ll file (~= # of instructions) by 2.5%.
- No intended functionality change -- at -O3 this should produce equivalent if not identical output. On 403.gcc's combine.c, x86_64, -O3, this isn't quite true and some of the changes are regressions, but I'm not going to worry about that until we move to a new access policy.
- There is still some room for improvement in the generated IR, in particular we can usually fold the sign-extension of the bit-field into one of the component access. See the FIXME.
llvm-svn: 101192
- For now, these policies are computed to match the current IRgen strategy, although the new information isn't being used yet (except in -fdump-record-layouts).
- Design comments appreciated.
llvm-svn: 101178
elements with explicit zero values instead of with tail padding.
On an example like this:
struct foo { int a; int b; };
struct foo fooarray[] = {
{1, 2},
{4},
};
We now lay this out as:
@fooarray = global [2 x %struct.foo] [%struct.foo { i32 1, i32 2 }, %struct.foo { i32 4, i32 0 }]
instead of as:
@fooarray = global %0 <{ %struct.foo { i32 1, i32 2 }, %1 { i32 4, [4 x i8] zeroinitializer } }>
Preserving both the struct type of the second element, but also the array type of the entire thing.
llvm-svn: 101155
trailing fields may not be represented in initializer lists, they
are being handled as padding and those fields *must* be zero
initialized.
llvm-svn: 101067
__cxxabiv1::__fundamental_type_info in every translation
unit. Previously, we would perform name lookup for
__cxxabiv1::__fundamental_type_info at the end of IRGen for a each
translation unit, to determine whether it was present. If so, we we
produce type information for all of the fundamental types. However,
this name lookup causes PCH deserialization of a significant part of the
translation unit, which has a woeful impact on performance.
With this change, we now look at each record type after we've
generated its vtable to see if it is
__cxxabiv1::__fundamental_type_info. If so, we generate type info for
all of the fundamental types. This works because
__cxxabiv1::__fundamental_type_info should always have a key function
(typically the virtual destructor), that will be defined once in the
support library. The fundamental type information will end up there.
Fixes <rdar://problem/7840011>.
llvm-svn: 100772
have the code generate slap a srcloc metadata on inline asm nodes.
This allows us to diagnose invalid inline asms with such nice
diagnostics as:
<inline asm>:1:2: error: unrecognized instruction
abc incl %eax
^
asm.c:2:12: note: generated from here
__asm__ ("abc incl %0" : "+r" (X));
^
2 diagnostics generated.
llvm-svn: 100608
- Unfortunately, this requires some horrible code in CGObjCMac which always
allocats a CGBitFieldInfo because we don't currently build a proper layout
for Objective-C classes. It needs to be cleaned up, but I don't want the
bit-field cleanups to be blocked on that.
llvm-svn: 100474
poor (and wrong) approximation of the actual rules governing when to
build a copy and when it can be elided.
The correct implementation is actually simpler than the
approximation. When we only enumerate constructors as part of
initialization (e.g., for direct initialization or when we're copying
from a class type or one of its derived classes), we don't create a
copy. When we enumerate all conversion functions, we do create a
copy. Before, we created some extra copies and missed some
others. The new test copy-initialization.cpp shows a case where we
missed creating a (required, non-elidable) copy as part of a
user-defined conversion, which resulted in a miscompile. This commit
also fixes PR6757, where the missing copy made us reject well-formed
code in the ternary operator.
This commit also cleans up our handling of copy elision in the case
where we create an extra copy of a temporary object, which became
necessary now that we produce the right copies. The code that seeks to
find the temporary object being copied has moved into
Expr::getTemporaryObject(); it used to have two different
not-quite-the-same implementations, one in Sema and one in CodeGen.
Note that we still do not attempt to perform the named return value
optimization, so we miss copy elisions for return values and throw
expressions.
llvm-svn: 100196
the existing (and already well-tested) linkage computation for types,
with minor tweaks for dynamic classes and (pointers to) incomplete
types. Fixes PR6597.
llvm-svn: 99968
This introduces FunctionType::ExtInfo to hold the calling convention and the
noreturn attribute. The next patch will extend it to include the regparm
attribute and fix the bug.
llvm-svn: 99920
Nate Begeman