llvm-project/clang/lib/CodeGen/CGObjCGNU.cpp

4124 lines
168 KiB
C++
Raw Normal View History

//===------- CGObjCGNU.cpp - Emit LLVM Code from ASTs for a Module --------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This provides Objective-C code generation targeting the GNU runtime. The
// class in this file generates structures used by the GNU Objective-C runtime
// library. These structures are defined in objc/objc.h and objc/objc-api.h in
// the GNU runtime distribution.
//
//===----------------------------------------------------------------------===//
#include "CGObjCRuntime.h"
#include "CGCleanup.h"
#include "CodeGenFunction.h"
#include "CodeGenModule.h"
#include "CGCXXABI.h"
#include "clang/CodeGen/ConstantInitBuilder.h"
#include "clang/AST/ASTContext.h"
#include "clang/AST/Decl.h"
#include "clang/AST/DeclObjC.h"
2009-07-19 03:43:29 +08:00
#include "clang/AST/RecordLayout.h"
#include "clang/AST/StmtObjC.h"
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
#include "clang/Basic/FileManager.h"
#include "clang/Basic/SourceManager.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringMap.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Module.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/ConvertUTF.h"
#include <cctype>
using namespace clang;
using namespace CodeGen;
namespace {
std::string SymbolNameForMethod( StringRef ClassName,
StringRef CategoryName, const Selector MethodName,
bool isClassMethod) {
std::string MethodNameColonStripped = MethodName.getAsString();
std::replace(MethodNameColonStripped.begin(), MethodNameColonStripped.end(),
':', '_');
return (Twine(isClassMethod ? "_c_" : "_i_") + ClassName + "_" +
CategoryName + "_" + MethodNameColonStripped).str();
}
/// Class that lazily initialises the runtime function. Avoids inserting the
/// types and the function declaration into a module if they're not used, and
/// avoids constructing the type more than once if it's used more than once.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
class LazyRuntimeFunction {
CodeGenModule *CGM;
llvm::FunctionType *FTy;
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
const char *FunctionName;
llvm::FunctionCallee Function;
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
2015-05-19 06:51:39 +08:00
public:
/// Constructor leaves this class uninitialized, because it is intended to
/// be used as a field in another class and not all of the types that are
/// used as arguments will necessarily be available at construction time.
LazyRuntimeFunction()
: CGM(nullptr), FunctionName(nullptr), Function(nullptr) {}
2015-05-19 06:51:39 +08:00
/// Initialises the lazy function with the name, return type, and the types
/// of the arguments.
template <typename... Tys>
void init(CodeGenModule *Mod, const char *name, llvm::Type *RetTy,
Tys *... Types) {
2015-05-19 06:51:39 +08:00
CGM = Mod;
FunctionName = name;
Function = nullptr;
if(sizeof...(Tys)) {
SmallVector<llvm::Type *, 8> ArgTys({Types...});
FTy = llvm::FunctionType::get(RetTy, ArgTys, false);
}
else {
FTy = llvm::FunctionType::get(RetTy, None, false);
}
2015-05-19 06:51:39 +08:00
}
llvm::FunctionType *getType() { return FTy; }
2015-05-19 06:51:39 +08:00
/// Overloaded cast operator, allows the class to be implicitly cast to an
/// LLVM constant.
operator llvm::FunctionCallee() {
2015-05-19 06:51:39 +08:00
if (!Function) {
if (!FunctionName)
return nullptr;
Function = CGM->CreateRuntimeFunction(FTy, FunctionName);
2015-05-19 06:51:39 +08:00
}
return Function;
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
};
/// GNU Objective-C runtime code generation. This class implements the parts of
/// Objective-C support that are specific to the GNU family of runtimes (GCC,
/// GNUstep and ObjFW).
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
class CGObjCGNU : public CGObjCRuntime {
protected:
/// The LLVM module into which output is inserted
llvm::Module &TheModule;
/// strut objc_super. Used for sending messages to super. This structure
/// contains the receiver (object) and the expected class.
llvm::StructType *ObjCSuperTy;
/// struct objc_super*. The type of the argument to the superclass message
/// lookup functions.
llvm::PointerType *PtrToObjCSuperTy;
/// LLVM type for selectors. Opaque pointer (i8*) unless a header declaring
/// SEL is included in a header somewhere, in which case it will be whatever
/// type is declared in that header, most likely {i8*, i8*}.
llvm::PointerType *SelectorTy;
/// LLVM i8 type. Cached here to avoid repeatedly getting it in all of the
/// places where it's used
llvm::IntegerType *Int8Ty;
/// Pointer to i8 - LLVM type of char*, for all of the places where the
/// runtime needs to deal with C strings.
llvm::PointerType *PtrToInt8Ty;
/// struct objc_protocol type
llvm::StructType *ProtocolTy;
/// Protocol * type.
llvm::PointerType *ProtocolPtrTy;
/// Instance Method Pointer type. This is a pointer to a function that takes,
/// at a minimum, an object and a selector, and is the generic type for
/// Objective-C methods. Due to differences between variadic / non-variadic
/// calling conventions, it must always be cast to the correct type before
/// actually being used.
llvm::PointerType *IMPTy;
/// Type of an untyped Objective-C object. Clang treats id as a built-in type
/// when compiling Objective-C code, so this may be an opaque pointer (i8*),
/// but if the runtime header declaring it is included then it may be a
/// pointer to a structure.
llvm::PointerType *IdTy;
/// Pointer to a pointer to an Objective-C object. Used in the new ABI
/// message lookup function and some GC-related functions.
llvm::PointerType *PtrToIdTy;
/// The clang type of id. Used when using the clang CGCall infrastructure to
/// call Objective-C methods.
CanQualType ASTIdTy;
/// LLVM type for C int type.
llvm::IntegerType *IntTy;
/// LLVM type for an opaque pointer. This is identical to PtrToInt8Ty, but is
/// used in the code to document the difference between i8* meaning a pointer
/// to a C string and i8* meaning a pointer to some opaque type.
llvm::PointerType *PtrTy;
/// LLVM type for C long type. The runtime uses this in a lot of places where
/// it should be using intptr_t, but we can't fix this without breaking
/// compatibility with GCC...
llvm::IntegerType *LongTy;
/// LLVM type for C size_t. Used in various runtime data structures.
llvm::IntegerType *SizeTy;
/// LLVM type for C intptr_t.
llvm::IntegerType *IntPtrTy;
/// LLVM type for C ptrdiff_t. Mainly used in property accessor functions.
llvm::IntegerType *PtrDiffTy;
/// LLVM type for C int*. Used for GCC-ABI-compatible non-fragile instance
/// variables.
llvm::PointerType *PtrToIntTy;
/// LLVM type for Objective-C BOOL type.
llvm::Type *BoolTy;
/// 32-bit integer type, to save us needing to look it up every time it's used.
llvm::IntegerType *Int32Ty;
/// 64-bit integer type, to save us needing to look it up every time it's used.
llvm::IntegerType *Int64Ty;
/// The type of struct objc_property.
llvm::StructType *PropertyMetadataTy;
/// Metadata kind used to tie method lookups to message sends. The GNUstep
/// runtime provides some LLVM passes that can use this to do things like
/// automatic IMP caching and speculative inlining.
unsigned msgSendMDKind;
/// Does the current target use SEH-based exceptions? False implies
/// Itanium-style DWARF unwinding.
bool usesSEHExceptions;
/// Helper to check if we are targeting a specific runtime version or later.
bool isRuntime(ObjCRuntime::Kind kind, unsigned major, unsigned minor=0) {
const ObjCRuntime &R = CGM.getLangOpts().ObjCRuntime;
return (R.getKind() == kind) &&
(R.getVersion() >= VersionTuple(major, minor));
}
std::string ManglePublicSymbol(StringRef Name) {
return (StringRef(CGM.getTriple().isOSBinFormatCOFF() ? "$_" : "._") + Name).str();
}
std::string SymbolForProtocol(Twine Name) {
return (ManglePublicSymbol("OBJC_PROTOCOL_") + Name).str();
}
std::string SymbolForProtocolRef(StringRef Name) {
return (ManglePublicSymbol("OBJC_REF_PROTOCOL_") + Name).str();
}
/// Helper function that generates a constant string and returns a pointer to
/// the start of the string. The result of this function can be used anywhere
/// where the C code specifies const char*.
llvm::Constant *MakeConstantString(StringRef Str, const char *Name = "") {
ConstantAddress Array = CGM.GetAddrOfConstantCString(Str, Name);
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
return llvm::ConstantExpr::getGetElementPtr(Array.getElementType(),
Array.getPointer(), Zeros);
}
/// Emits a linkonce_odr string, whose name is the prefix followed by the
/// string value. This allows the linker to combine the strings between
/// different modules. Used for EH typeinfo names, selector strings, and a
/// few other things.
llvm::Constant *ExportUniqueString(const std::string &Str,
const std::string &prefix,
bool Private=false) {
std::string name = prefix + Str;
auto *ConstStr = TheModule.getGlobalVariable(name);
if (!ConstStr) {
llvm::Constant *value = llvm::ConstantDataArray::getString(VMContext,Str);
auto *GV = new llvm::GlobalVariable(TheModule, value->getType(), true,
llvm::GlobalValue::LinkOnceODRLinkage, value, name);
GV->setComdat(TheModule.getOrInsertComdat(name));
if (Private)
GV->setVisibility(llvm::GlobalValue::HiddenVisibility);
ConstStr = GV;
}
return llvm::ConstantExpr::getGetElementPtr(ConstStr->getValueType(),
ConstStr, Zeros);
}
/// Returns a property name and encoding string.
llvm::Constant *MakePropertyEncodingString(const ObjCPropertyDecl *PD,
const Decl *Container) {
assert(!isRuntime(ObjCRuntime::GNUstep, 2));
if (isRuntime(ObjCRuntime::GNUstep, 1, 6)) {
std::string NameAndAttributes;
std::string TypeStr =
CGM.getContext().getObjCEncodingForPropertyDecl(PD, Container);
NameAndAttributes += '\0';
NameAndAttributes += TypeStr.length() + 3;
NameAndAttributes += TypeStr;
NameAndAttributes += '\0';
NameAndAttributes += PD->getNameAsString();
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
return MakeConstantString(NameAndAttributes);
}
return MakeConstantString(PD->getNameAsString());
}
/// Push the property attributes into two structure fields.
void PushPropertyAttributes(ConstantStructBuilder &Fields,
const ObjCPropertyDecl *property, bool isSynthesized=true, bool
isDynamic=true) {
int attrs = property->getPropertyAttributes();
// For read-only properties, clear the copy and retain flags
if (attrs & ObjCPropertyDecl::OBJC_PR_readonly) {
attrs &= ~ObjCPropertyDecl::OBJC_PR_copy;
attrs &= ~ObjCPropertyDecl::OBJC_PR_retain;
attrs &= ~ObjCPropertyDecl::OBJC_PR_weak;
attrs &= ~ObjCPropertyDecl::OBJC_PR_strong;
}
// The first flags field has the same attribute values as clang uses internally
Fields.addInt(Int8Ty, attrs & 0xff);
attrs >>= 8;
attrs <<= 2;
// For protocol properties, synthesized and dynamic have no meaning, so we
// reuse these flags to indicate that this is a protocol property (both set
// has no meaning, as a property can't be both synthesized and dynamic)
attrs |= isSynthesized ? (1<<0) : 0;
attrs |= isDynamic ? (1<<1) : 0;
// The second field is the next four fields left shifted by two, with the
// low bit set to indicate whether the field is synthesized or dynamic.
Fields.addInt(Int8Ty, attrs & 0xff);
// Two padding fields
Fields.addInt(Int8Ty, 0);
Fields.addInt(Int8Ty, 0);
}
virtual llvm::Constant *GenerateCategoryProtocolList(const
ObjCCategoryDecl *OCD);
virtual ConstantArrayBuilder PushPropertyListHeader(ConstantStructBuilder &Fields,
int count) {
// int count;
Fields.addInt(IntTy, count);
// int size; (only in GNUstep v2 ABI.
if (isRuntime(ObjCRuntime::GNUstep, 2)) {
llvm::DataLayout td(&TheModule);
Fields.addInt(IntTy, td.getTypeSizeInBits(PropertyMetadataTy) /
CGM.getContext().getCharWidth());
}
// struct objc_property_list *next;
Fields.add(NULLPtr);
// struct objc_property properties[]
return Fields.beginArray(PropertyMetadataTy);
}
virtual void PushProperty(ConstantArrayBuilder &PropertiesArray,
const ObjCPropertyDecl *property,
const Decl *OCD,
bool isSynthesized=true, bool
isDynamic=true) {
auto Fields = PropertiesArray.beginStruct(PropertyMetadataTy);
ASTContext &Context = CGM.getContext();
Fields.add(MakePropertyEncodingString(property, OCD));
PushPropertyAttributes(Fields, property, isSynthesized, isDynamic);
auto addPropertyMethod = [&](const ObjCMethodDecl *accessor) {
if (accessor) {
std::string TypeStr = Context.getObjCEncodingForMethodDecl(accessor);
llvm::Constant *TypeEncoding = MakeConstantString(TypeStr);
Fields.add(MakeConstantString(accessor->getSelector().getAsString()));
Fields.add(TypeEncoding);
} else {
Fields.add(NULLPtr);
Fields.add(NULLPtr);
}
};
addPropertyMethod(property->getGetterMethodDecl());
addPropertyMethod(property->getSetterMethodDecl());
Fields.finishAndAddTo(PropertiesArray);
}
/// Ensures that the value has the required type, by inserting a bitcast if
/// required. This function lets us avoid inserting bitcasts that are
/// redundant.
llvm::Value* EnforceType(CGBuilderTy &B, llvm::Value *V, llvm::Type *Ty) {
if (V->getType() == Ty) return V;
return B.CreateBitCast(V, Ty);
}
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address EnforceType(CGBuilderTy &B, Address V, llvm::Type *Ty) {
if (V.getType() == Ty) return V;
return B.CreateBitCast(V, Ty);
}
// Some zeros used for GEPs in lots of places.
llvm::Constant *Zeros[2];
/// Null pointer value. Mainly used as a terminator in various arrays.
llvm::Constant *NULLPtr;
/// LLVM context.
llvm::LLVMContext &VMContext;
protected:
/// Placeholder for the class. Lots of things refer to the class before we've
/// actually emitted it. We use this alias as a placeholder, and then replace
/// it with a pointer to the class structure before finally emitting the
/// module.
llvm::GlobalAlias *ClassPtrAlias;
/// Placeholder for the metaclass. Lots of things refer to the class before
/// we've / actually emitted it. We use this alias as a placeholder, and then
/// replace / it with a pointer to the metaclass structure before finally
/// emitting the / module.
llvm::GlobalAlias *MetaClassPtrAlias;
/// All of the classes that have been generated for this compilation units.
std::vector<llvm::Constant*> Classes;
/// All of the categories that have been generated for this compilation units.
std::vector<llvm::Constant*> Categories;
/// All of the Objective-C constant strings that have been generated for this
/// compilation units.
std::vector<llvm::Constant*> ConstantStrings;
/// Map from string values to Objective-C constant strings in the output.
/// Used to prevent emitting Objective-C strings more than once. This should
/// not be required at all - CodeGenModule should manage this list.
llvm::StringMap<llvm::Constant*> ObjCStrings;
/// All of the protocols that have been declared.
llvm::StringMap<llvm::Constant*> ExistingProtocols;
/// For each variant of a selector, we store the type encoding and a
/// placeholder value. For an untyped selector, the type will be the empty
/// string. Selector references are all done via the module's selector table,
/// so we create an alias as a placeholder and then replace it with the real
/// value later.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
typedef std::pair<std::string, llvm::GlobalAlias*> TypedSelector;
/// Type of the selector map. This is roughly equivalent to the structure
/// used in the GNUstep runtime, which maintains a list of all of the valid
/// types for a selector in a table.
typedef llvm::DenseMap<Selector, SmallVector<TypedSelector, 2> >
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
SelectorMap;
/// A map from selectors to selector types. This allows us to emit all
/// selectors of the same name and type together.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
SelectorMap SelectorTable;
/// Selectors related to memory management. When compiling in GC mode, we
/// omit these.
Selector RetainSel, ReleaseSel, AutoreleaseSel;
/// Runtime functions used for memory management in GC mode. Note that clang
/// supports code generation for calling these functions, but neither GNU
/// runtime actually supports this API properly yet.
LazyRuntimeFunction IvarAssignFn, StrongCastAssignFn, MemMoveFn, WeakReadFn,
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
WeakAssignFn, GlobalAssignFn;
typedef std::pair<std::string, std::string> ClassAliasPair;
/// All classes that have aliases set for them.
std::vector<ClassAliasPair> ClassAliases;
protected:
/// Function used for throwing Objective-C exceptions.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
LazyRuntimeFunction ExceptionThrowFn;
/// Function used for rethrowing exceptions, used at the end of \@finally or
/// \@synchronize blocks.
LazyRuntimeFunction ExceptionReThrowFn;
/// Function called when entering a catch function. This is required for
/// differentiating Objective-C exceptions and foreign exceptions.
LazyRuntimeFunction EnterCatchFn;
/// Function called when exiting from a catch block. Used to do exception
/// cleanup.
LazyRuntimeFunction ExitCatchFn;
/// Function called when entering an \@synchronize block. Acquires the lock.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
LazyRuntimeFunction SyncEnterFn;
/// Function called when exiting an \@synchronize block. Releases the lock.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
LazyRuntimeFunction SyncExitFn;
private:
/// Function called if fast enumeration detects that the collection is
/// modified during the update.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
LazyRuntimeFunction EnumerationMutationFn;
/// Function for implementing synthesized property getters that return an
/// object.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
LazyRuntimeFunction GetPropertyFn;
/// Function for implementing synthesized property setters that return an
/// object.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
LazyRuntimeFunction SetPropertyFn;
/// Function used for non-object declared property getters.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
LazyRuntimeFunction GetStructPropertyFn;
/// Function used for non-object declared property setters.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
LazyRuntimeFunction SetStructPropertyFn;
protected:
/// The version of the runtime that this class targets. Must match the
/// version in the runtime.
int RuntimeVersion;
/// The version of the protocol class. Used to differentiate between ObjC1
/// and ObjC2 protocols. Objective-C 1 protocols can not contain optional
/// components and can not contain declared properties. We always emit
/// Objective-C 2 property structures, but we have to pretend that they're
/// Objective-C 1 property structures when targeting the GCC runtime or it
/// will abort.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
const int ProtocolVersion;
/// The version of the class ABI. This value is used in the class structure
/// and indicates how various fields should be interpreted.
const int ClassABIVersion;
/// Generates an instance variable list structure. This is a structure
/// containing a size and an array of structures containing instance variable
/// metadata. This is used purely for introspection in the fragile ABI. In
/// the non-fragile ABI, it's used for instance variable fixup.
virtual llvm::Constant *GenerateIvarList(ArrayRef<llvm::Constant *> IvarNames,
ArrayRef<llvm::Constant *> IvarTypes,
ArrayRef<llvm::Constant *> IvarOffsets,
ArrayRef<llvm::Constant *> IvarAlign,
ArrayRef<Qualifiers::ObjCLifetime> IvarOwnership);
/// Generates a method list structure. This is a structure containing a size
/// and an array of structures containing method metadata.
///
/// This structure is used by both classes and categories, and contains a next
/// pointer allowing them to be chained together in a linked list.
llvm::Constant *GenerateMethodList(StringRef ClassName,
StringRef CategoryName,
ArrayRef<const ObjCMethodDecl*> Methods,
bool isClassMethodList);
/// Emits an empty protocol. This is used for \@protocol() where no protocol
/// is found. The runtime will (hopefully) fix up the pointer to refer to the
/// real protocol.
virtual llvm::Constant *GenerateEmptyProtocol(StringRef ProtocolName);
/// Generates a list of property metadata structures. This follows the same
/// pattern as method and instance variable metadata lists.
llvm::Constant *GeneratePropertyList(const Decl *Container,
const ObjCContainerDecl *OCD,
bool isClassProperty=false,
bool protocolOptionalProperties=false);
/// Generates a list of referenced protocols. Classes, categories, and
/// protocols all use this structure.
llvm::Constant *GenerateProtocolList(ArrayRef<std::string> Protocols);
/// To ensure that all protocols are seen by the runtime, we add a category on
/// a class defined in the runtime, declaring no methods, but adopting the
/// protocols. This is a horribly ugly hack, but it allows us to collect all
/// of the protocols without changing the ABI.
void GenerateProtocolHolderCategory();
/// Generates a class structure.
llvm::Constant *GenerateClassStructure(
llvm::Constant *MetaClass,
llvm::Constant *SuperClass,
unsigned info,
const char *Name,
llvm::Constant *Version,
llvm::Constant *InstanceSize,
llvm::Constant *IVars,
llvm::Constant *Methods,
llvm::Constant *Protocols,
llvm::Constant *IvarOffsets,
llvm::Constant *Properties,
llvm::Constant *StrongIvarBitmap,
llvm::Constant *WeakIvarBitmap,
bool isMeta=false);
/// Generates a method list. This is used by protocols to define the required
/// and optional methods.
virtual llvm::Constant *GenerateProtocolMethodList(
ArrayRef<const ObjCMethodDecl*> Methods);
/// Emits optional and required method lists.
template<class T>
void EmitProtocolMethodList(T &&Methods, llvm::Constant *&Required,
llvm::Constant *&Optional) {
SmallVector<const ObjCMethodDecl*, 16> RequiredMethods;
SmallVector<const ObjCMethodDecl*, 16> OptionalMethods;
for (const auto *I : Methods)
if (I->isOptional())
OptionalMethods.push_back(I);
else
RequiredMethods.push_back(I);
Required = GenerateProtocolMethodList(RequiredMethods);
Optional = GenerateProtocolMethodList(OptionalMethods);
}
/// Returns a selector with the specified type encoding. An empty string is
/// used to return an untyped selector (with the types field set to NULL).
virtual llvm::Value *GetTypedSelector(CodeGenFunction &CGF, Selector Sel,
const std::string &TypeEncoding);
/// Returns the name of ivar offset variables. In the GNUstep v1 ABI, this
/// contains the class and ivar names, in the v2 ABI this contains the type
/// encoding as well.
virtual std::string GetIVarOffsetVariableName(const ObjCInterfaceDecl *ID,
const ObjCIvarDecl *Ivar) {
const std::string Name = "__objc_ivar_offset_" + ID->getNameAsString()
+ '.' + Ivar->getNameAsString();
return Name;
}
/// Returns the variable used to store the offset of an instance variable.
llvm::GlobalVariable *ObjCIvarOffsetVariable(const ObjCInterfaceDecl *ID,
const ObjCIvarDecl *Ivar);
/// Emits a reference to a class. This allows the linker to object if there
/// is no class of the matching name.
void EmitClassRef(const std::string &className);
/// Emits a pointer to the named class
virtual llvm::Value *GetClassNamed(CodeGenFunction &CGF,
const std::string &Name, bool isWeak);
/// Looks up the method for sending a message to the specified object. This
/// mechanism differs between the GCC and GNU runtimes, so this method must be
/// overridden in subclasses.
virtual llvm::Value *LookupIMP(CodeGenFunction &CGF,
llvm::Value *&Receiver,
llvm::Value *cmd,
llvm::MDNode *node,
MessageSendInfo &MSI) = 0;
/// Looks up the method for sending a message to a superclass. This
/// mechanism differs between the GCC and GNU runtimes, so this method must
/// be overridden in subclasses.
virtual llvm::Value *LookupIMPSuper(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address ObjCSuper,
llvm::Value *cmd,
MessageSendInfo &MSI) = 0;
/// Libobjc2 uses a bitfield representation where small(ish) bitfields are
/// stored in a 64-bit value with the low bit set to 1 and the remaining 63
/// bits set to their values, LSB first, while larger ones are stored in a
/// structure of this / form:
///
/// struct { int32_t length; int32_t values[length]; };
///
/// The values in the array are stored in host-endian format, with the least
/// significant bit being assumed to come first in the bitfield. Therefore,
/// a bitfield with the 64th bit set will be (int64_t)&{ 2, [0, 1<<31] },
/// while a bitfield / with the 63rd bit set will be 1<<64.
llvm::Constant *MakeBitField(ArrayRef<bool> bits);
public:
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
CGObjCGNU(CodeGenModule &cgm, unsigned runtimeABIVersion,
unsigned protocolClassVersion, unsigned classABI=1);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
ConstantAddress GenerateConstantString(const StringLiteral *) override;
RValue
GenerateMessageSend(CodeGenFunction &CGF, ReturnValueSlot Return,
QualType ResultType, Selector Sel,
llvm::Value *Receiver, const CallArgList &CallArgs,
const ObjCInterfaceDecl *Class,
const ObjCMethodDecl *Method) override;
RValue
GenerateMessageSendSuper(CodeGenFunction &CGF, ReturnValueSlot Return,
QualType ResultType, Selector Sel,
const ObjCInterfaceDecl *Class,
bool isCategoryImpl, llvm::Value *Receiver,
bool IsClassMessage, const CallArgList &CallArgs,
const ObjCMethodDecl *Method) override;
llvm::Value *GetClass(CodeGenFunction &CGF,
const ObjCInterfaceDecl *OID) override;
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *GetSelector(CodeGenFunction &CGF, Selector Sel) override;
Address GetAddrOfSelector(CodeGenFunction &CGF, Selector Sel) override;
llvm::Value *GetSelector(CodeGenFunction &CGF,
const ObjCMethodDecl *Method) override;
virtual llvm::Constant *GetConstantSelector(Selector Sel,
const std::string &TypeEncoding) {
llvm_unreachable("Runtime unable to generate constant selector");
}
llvm::Constant *GetConstantSelector(const ObjCMethodDecl *M) {
return GetConstantSelector(M->getSelector(),
CGM.getContext().getObjCEncodingForMethodDecl(M));
}
llvm::Constant *GetEHType(QualType T) override;
llvm::Function *GenerateMethod(const ObjCMethodDecl *OMD,
const ObjCContainerDecl *CD) override;
void GenerateCategory(const ObjCCategoryImplDecl *CMD) override;
void GenerateClass(const ObjCImplementationDecl *ClassDecl) override;
void RegisterAlias(const ObjCCompatibleAliasDecl *OAD) override;
llvm::Value *GenerateProtocolRef(CodeGenFunction &CGF,
const ObjCProtocolDecl *PD) override;
void GenerateProtocol(const ObjCProtocolDecl *PD) override;
llvm::Function *ModuleInitFunction() override;
llvm::FunctionCallee GetPropertyGetFunction() override;
llvm::FunctionCallee GetPropertySetFunction() override;
llvm::FunctionCallee GetOptimizedPropertySetFunction(bool atomic,
bool copy) override;
llvm::FunctionCallee GetSetStructFunction() override;
llvm::FunctionCallee GetGetStructFunction() override;
llvm::FunctionCallee GetCppAtomicObjectGetFunction() override;
llvm::FunctionCallee GetCppAtomicObjectSetFunction() override;
llvm::FunctionCallee EnumerationMutationFunction() override;
void EmitTryStmt(CodeGenFunction &CGF,
const ObjCAtTryStmt &S) override;
void EmitSynchronizedStmt(CodeGenFunction &CGF,
const ObjCAtSynchronizedStmt &S) override;
void EmitThrowStmt(CodeGenFunction &CGF,
const ObjCAtThrowStmt &S,
bool ClearInsertionPoint=true) override;
llvm::Value * EmitObjCWeakRead(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address AddrWeakObj) override;
void EmitObjCWeakAssign(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *src, Address dst) override;
void EmitObjCGlobalAssign(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *src, Address dest,
bool threadlocal=false) override;
void EmitObjCIvarAssign(CodeGenFunction &CGF, llvm::Value *src,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address dest, llvm::Value *ivarOffset) override;
void EmitObjCStrongCastAssign(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *src, Address dest) override;
void EmitGCMemmoveCollectable(CodeGenFunction &CGF, Address DestPtr,
Address SrcPtr,
llvm::Value *Size) override;
LValue EmitObjCValueForIvar(CodeGenFunction &CGF, QualType ObjectTy,
llvm::Value *BaseValue, const ObjCIvarDecl *Ivar,
unsigned CVRQualifiers) override;
llvm::Value *EmitIvarOffset(CodeGenFunction &CGF,
const ObjCInterfaceDecl *Interface,
const ObjCIvarDecl *Ivar) override;
llvm::Value *EmitNSAutoreleasePoolClassRef(CodeGenFunction &CGF) override;
llvm::Constant *BuildGCBlockLayout(CodeGenModule &CGM,
const CGBlockInfo &blockInfo) override {
return NULLPtr;
}
llvm::Constant *BuildRCBlockLayout(CodeGenModule &CGM,
const CGBlockInfo &blockInfo) override {
return NULLPtr;
}
llvm::Constant *BuildByrefLayout(CodeGenModule &CGM, QualType T) override {
return NULLPtr;
}
};
/// Class representing the legacy GCC Objective-C ABI. This is the default when
/// -fobjc-nonfragile-abi is not specified.
///
/// The GCC ABI target actually generates code that is approximately compatible
/// with the new GNUstep runtime ABI, but refrains from using any features that
/// would not work with the GCC runtime. For example, clang always generates
/// the extended form of the class structure, and the extra fields are simply
/// ignored by GCC libobjc.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
class CGObjCGCC : public CGObjCGNU {
/// The GCC ABI message lookup function. Returns an IMP pointing to the
/// method implementation for this message.
LazyRuntimeFunction MsgLookupFn;
/// The GCC ABI superclass message lookup function. Takes a pointer to a
/// structure describing the receiver and the class, and a selector as
/// arguments. Returns the IMP for the corresponding method.
LazyRuntimeFunction MsgLookupSuperFn;
protected:
llvm::Value *LookupIMP(CodeGenFunction &CGF, llvm::Value *&Receiver,
llvm::Value *cmd, llvm::MDNode *node,
MessageSendInfo &MSI) override {
CGBuilderTy &Builder = CGF.Builder;
llvm::Value *args[] = {
EnforceType(Builder, Receiver, IdTy),
EnforceType(Builder, cmd, SelectorTy) };
llvm::CallBase *imp = CGF.EmitRuntimeCallOrInvoke(MsgLookupFn, args);
imp->setMetadata(msgSendMDKind, node);
return imp;
}
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *LookupIMPSuper(CodeGenFunction &CGF, Address ObjCSuper,
llvm::Value *cmd, MessageSendInfo &MSI) override {
CGBuilderTy &Builder = CGF.Builder;
llvm::Value *lookupArgs[] = {EnforceType(Builder, ObjCSuper,
PtrToObjCSuperTy).getPointer(), cmd};
return CGF.EmitNounwindRuntimeCall(MsgLookupSuperFn, lookupArgs);
}
public:
CGObjCGCC(CodeGenModule &Mod) : CGObjCGNU(Mod, 8, 2) {
// IMP objc_msg_lookup(id, SEL);
MsgLookupFn.init(&CGM, "objc_msg_lookup", IMPTy, IdTy, SelectorTy);
// IMP objc_msg_lookup_super(struct objc_super*, SEL);
MsgLookupSuperFn.init(&CGM, "objc_msg_lookup_super", IMPTy,
PtrToObjCSuperTy, SelectorTy);
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
};
/// Class used when targeting the new GNUstep runtime ABI.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
class CGObjCGNUstep : public CGObjCGNU {
/// The slot lookup function. Returns a pointer to a cacheable structure
/// that contains (among other things) the IMP.
LazyRuntimeFunction SlotLookupFn;
/// The GNUstep ABI superclass message lookup function. Takes a pointer to
/// a structure describing the receiver and the class, and a selector as
/// arguments. Returns the slot for the corresponding method. Superclass
/// message lookup rarely changes, so this is a good caching opportunity.
LazyRuntimeFunction SlotLookupSuperFn;
/// Specialised function for setting atomic retain properties
LazyRuntimeFunction SetPropertyAtomic;
/// Specialised function for setting atomic copy properties
LazyRuntimeFunction SetPropertyAtomicCopy;
/// Specialised function for setting nonatomic retain properties
LazyRuntimeFunction SetPropertyNonAtomic;
/// Specialised function for setting nonatomic copy properties
LazyRuntimeFunction SetPropertyNonAtomicCopy;
/// Function to perform atomic copies of C++ objects with nontrivial copy
/// constructors from Objective-C ivars.
LazyRuntimeFunction CxxAtomicObjectGetFn;
/// Function to perform atomic copies of C++ objects with nontrivial copy
/// constructors to Objective-C ivars.
LazyRuntimeFunction CxxAtomicObjectSetFn;
/// Type of an slot structure pointer. This is returned by the various
/// lookup functions.
llvm::Type *SlotTy;
public:
llvm::Constant *GetEHType(QualType T) override;
protected:
llvm::Value *LookupIMP(CodeGenFunction &CGF, llvm::Value *&Receiver,
llvm::Value *cmd, llvm::MDNode *node,
MessageSendInfo &MSI) override {
CGBuilderTy &Builder = CGF.Builder;
llvm::FunctionCallee LookupFn = SlotLookupFn;
// Store the receiver on the stack so that we can reload it later
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address ReceiverPtr =
CGF.CreateTempAlloca(Receiver->getType(), CGF.getPointerAlign());
Builder.CreateStore(Receiver, ReceiverPtr);
llvm::Value *self;
if (isa<ObjCMethodDecl>(CGF.CurCodeDecl)) {
self = CGF.LoadObjCSelf();
} else {
self = llvm::ConstantPointerNull::get(IdTy);
}
// The lookup function is guaranteed not to capture the receiver pointer.
if (auto *LookupFn2 = dyn_cast<llvm::Function>(LookupFn.getCallee()))
LookupFn2->addParamAttr(0, llvm::Attribute::NoCapture);
llvm::Value *args[] = {
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
EnforceType(Builder, ReceiverPtr.getPointer(), PtrToIdTy),
EnforceType(Builder, cmd, SelectorTy),
EnforceType(Builder, self, IdTy) };
llvm::CallBase *slot = CGF.EmitRuntimeCallOrInvoke(LookupFn, args);
slot->setOnlyReadsMemory();
slot->setMetadata(msgSendMDKind, node);
// Load the imp from the slot
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *imp = Builder.CreateAlignedLoad(
Builder.CreateStructGEP(nullptr, slot, 4), CGF.getPointerAlign());
// The lookup function may have changed the receiver, so make sure we use
// the new one.
Receiver = Builder.CreateLoad(ReceiverPtr, true);
return imp;
}
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *LookupIMPSuper(CodeGenFunction &CGF, Address ObjCSuper,
llvm::Value *cmd,
MessageSendInfo &MSI) override {
CGBuilderTy &Builder = CGF.Builder;
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *lookupArgs[] = {ObjCSuper.getPointer(), cmd};
llvm::CallInst *slot =
CGF.EmitNounwindRuntimeCall(SlotLookupSuperFn, lookupArgs);
slot->setOnlyReadsMemory();
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
return Builder.CreateAlignedLoad(Builder.CreateStructGEP(nullptr, slot, 4),
CGF.getPointerAlign());
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
public:
CGObjCGNUstep(CodeGenModule &Mod) : CGObjCGNUstep(Mod, 9, 3, 1) {}
CGObjCGNUstep(CodeGenModule &Mod, unsigned ABI, unsigned ProtocolABI,
unsigned ClassABI) :
CGObjCGNU(Mod, ABI, ProtocolABI, ClassABI) {
const ObjCRuntime &R = CGM.getLangOpts().ObjCRuntime;
llvm::StructType *SlotStructTy =
llvm::StructType::get(PtrTy, PtrTy, PtrTy, IntTy, IMPTy);
SlotTy = llvm::PointerType::getUnqual(SlotStructTy);
// Slot_t objc_msg_lookup_sender(id *receiver, SEL selector, id sender);
SlotLookupFn.init(&CGM, "objc_msg_lookup_sender", SlotTy, PtrToIdTy,
SelectorTy, IdTy);
// Slot_t objc_slot_lookup_super(struct objc_super*, SEL);
SlotLookupSuperFn.init(&CGM, "objc_slot_lookup_super", SlotTy,
PtrToObjCSuperTy, SelectorTy);
// If we're in ObjC++ mode, then we want to make
if (usesSEHExceptions) {
llvm::Type *VoidTy = llvm::Type::getVoidTy(VMContext);
// void objc_exception_rethrow(void)
ExceptionReThrowFn.init(&CGM, "objc_exception_rethrow", VoidTy);
} else if (CGM.getLangOpts().CPlusPlus) {
llvm::Type *VoidTy = llvm::Type::getVoidTy(VMContext);
// void *__cxa_begin_catch(void *e)
EnterCatchFn.init(&CGM, "__cxa_begin_catch", PtrTy, PtrTy);
// void __cxa_end_catch(void)
ExitCatchFn.init(&CGM, "__cxa_end_catch", VoidTy);
// void _Unwind_Resume_or_Rethrow(void*)
ExceptionReThrowFn.init(&CGM, "_Unwind_Resume_or_Rethrow", VoidTy,
PtrTy);
} else if (R.getVersion() >= VersionTuple(1, 7)) {
llvm::Type *VoidTy = llvm::Type::getVoidTy(VMContext);
// id objc_begin_catch(void *e)
EnterCatchFn.init(&CGM, "objc_begin_catch", IdTy, PtrTy);
// void objc_end_catch(void)
ExitCatchFn.init(&CGM, "objc_end_catch", VoidTy);
// void _Unwind_Resume_or_Rethrow(void*)
ExceptionReThrowFn.init(&CGM, "objc_exception_rethrow", VoidTy, PtrTy);
}
llvm::Type *VoidTy = llvm::Type::getVoidTy(VMContext);
SetPropertyAtomic.init(&CGM, "objc_setProperty_atomic", VoidTy, IdTy,
SelectorTy, IdTy, PtrDiffTy);
SetPropertyAtomicCopy.init(&CGM, "objc_setProperty_atomic_copy", VoidTy,
IdTy, SelectorTy, IdTy, PtrDiffTy);
SetPropertyNonAtomic.init(&CGM, "objc_setProperty_nonatomic", VoidTy,
IdTy, SelectorTy, IdTy, PtrDiffTy);
SetPropertyNonAtomicCopy.init(&CGM, "objc_setProperty_nonatomic_copy",
VoidTy, IdTy, SelectorTy, IdTy, PtrDiffTy);
// void objc_setCppObjectAtomic(void *dest, const void *src, void
// *helper);
CxxAtomicObjectSetFn.init(&CGM, "objc_setCppObjectAtomic", VoidTy, PtrTy,
PtrTy, PtrTy);
// void objc_getCppObjectAtomic(void *dest, const void *src, void
// *helper);
CxxAtomicObjectGetFn.init(&CGM, "objc_getCppObjectAtomic", VoidTy, PtrTy,
PtrTy, PtrTy);
}
llvm::FunctionCallee GetCppAtomicObjectGetFunction() override {
// The optimised functions were added in version 1.7 of the GNUstep
// runtime.
assert (CGM.getLangOpts().ObjCRuntime.getVersion() >=
VersionTuple(1, 7));
return CxxAtomicObjectGetFn;
}
llvm::FunctionCallee GetCppAtomicObjectSetFunction() override {
// The optimised functions were added in version 1.7 of the GNUstep
// runtime.
assert (CGM.getLangOpts().ObjCRuntime.getVersion() >=
VersionTuple(1, 7));
return CxxAtomicObjectSetFn;
}
llvm::FunctionCallee GetOptimizedPropertySetFunction(bool atomic,
bool copy) override {
// The optimised property functions omit the GC check, and so are not
// safe to use in GC mode. The standard functions are fast in GC mode,
// so there is less advantage in using them.
assert ((CGM.getLangOpts().getGC() == LangOptions::NonGC));
// The optimised functions were added in version 1.7 of the GNUstep
// runtime.
assert (CGM.getLangOpts().ObjCRuntime.getVersion() >=
VersionTuple(1, 7));
if (atomic) {
if (copy) return SetPropertyAtomicCopy;
return SetPropertyAtomic;
}
return copy ? SetPropertyNonAtomicCopy : SetPropertyNonAtomic;
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
};
/// GNUstep Objective-C ABI version 2 implementation.
/// This is the ABI that provides a clean break with the legacy GCC ABI and
/// cleans up a number of things that were added to work around 1980s linkers.
class CGObjCGNUstep2 : public CGObjCGNUstep {
enum SectionKind
{
SelectorSection = 0,
ClassSection,
ClassReferenceSection,
CategorySection,
ProtocolSection,
ProtocolReferenceSection,
ClassAliasSection,
ConstantStringSection
};
static const char *const SectionsBaseNames[8];
static const char *const PECOFFSectionsBaseNames[8];
template<SectionKind K>
std::string sectionName() {
if (CGM.getTriple().isOSBinFormatCOFF()) {
std::string name(PECOFFSectionsBaseNames[K]);
name += "$m";
return name;
}
return SectionsBaseNames[K];
}
/// The GCC ABI superclass message lookup function. Takes a pointer to a
/// structure describing the receiver and the class, and a selector as
/// arguments. Returns the IMP for the corresponding method.
LazyRuntimeFunction MsgLookupSuperFn;
/// A flag indicating if we've emitted at least one protocol.
/// If we haven't, then we need to emit an empty protocol, to ensure that the
/// __start__objc_protocols and __stop__objc_protocols sections exist.
bool EmittedProtocol = false;
/// A flag indicating if we've emitted at least one protocol reference.
/// If we haven't, then we need to emit an empty protocol, to ensure that the
/// __start__objc_protocol_refs and __stop__objc_protocol_refs sections
/// exist.
bool EmittedProtocolRef = false;
/// A flag indicating if we've emitted at least one class.
/// If we haven't, then we need to emit an empty protocol, to ensure that the
/// __start__objc_classes and __stop__objc_classes sections / exist.
bool EmittedClass = false;
/// Generate the name of a symbol for a reference to a class. Accesses to
/// classes should be indirected via this.
typedef std::pair<std::string, std::pair<llvm::Constant*, int>> EarlyInitPair;
std::vector<EarlyInitPair> EarlyInitList;
std::string SymbolForClassRef(StringRef Name, bool isWeak) {
if (isWeak)
return (ManglePublicSymbol("OBJC_WEAK_REF_CLASS_") + Name).str();
else
return (ManglePublicSymbol("OBJC_REF_CLASS_") + Name).str();
}
/// Generate the name of a class symbol.
std::string SymbolForClass(StringRef Name) {
return (ManglePublicSymbol("OBJC_CLASS_") + Name).str();
}
void CallRuntimeFunction(CGBuilderTy &B, StringRef FunctionName,
ArrayRef<llvm::Value*> Args) {
SmallVector<llvm::Type *,8> Types;
for (auto *Arg : Args)
Types.push_back(Arg->getType());
llvm::FunctionType *FT = llvm::FunctionType::get(B.getVoidTy(), Types,
false);
llvm::FunctionCallee Fn = CGM.CreateRuntimeFunction(FT, FunctionName);
B.CreateCall(Fn, Args);
}
ConstantAddress GenerateConstantString(const StringLiteral *SL) override {
auto Str = SL->getString();
CharUnits Align = CGM.getPointerAlign();
// Look for an existing one
llvm::StringMap<llvm::Constant*>::iterator old = ObjCStrings.find(Str);
if (old != ObjCStrings.end())
return ConstantAddress(old->getValue(), Align);
bool isNonASCII = SL->containsNonAscii();
auto LiteralLength = SL->getLength();
if ((CGM.getTarget().getPointerWidth(0) == 64) &&
(LiteralLength < 9) && !isNonASCII) {
// Tiny strings are only used on 64-bit platforms. They store 8 7-bit
// ASCII characters in the high 56 bits, followed by a 4-bit length and a
// 3-bit tag (which is always 4).
uint64_t str = 0;
// Fill in the characters
for (unsigned i=0 ; i<LiteralLength ; i++)
str |= ((uint64_t)SL->getCodeUnit(i)) << ((64 - 4 - 3) - (i*7));
// Fill in the length
str |= LiteralLength << 3;
// Set the tag
str |= 4;
auto *ObjCStr = llvm::ConstantExpr::getIntToPtr(
llvm::ConstantInt::get(Int64Ty, str), IdTy);
ObjCStrings[Str] = ObjCStr;
return ConstantAddress(ObjCStr, Align);
}
StringRef StringClass = CGM.getLangOpts().ObjCConstantStringClass;
if (StringClass.empty()) StringClass = "NSConstantString";
std::string Sym = SymbolForClass(StringClass);
llvm::Constant *isa = TheModule.getNamedGlobal(Sym);
if (!isa) {
isa = new llvm::GlobalVariable(TheModule, IdTy, /* isConstant */false,
llvm::GlobalValue::ExternalLinkage, nullptr, Sym);
if (CGM.getTriple().isOSBinFormatCOFF()) {
cast<llvm::GlobalValue>(isa)->setDLLStorageClass(llvm::GlobalValue::DLLImportStorageClass);
}
} else if (isa->getType() != PtrToIdTy)
isa = llvm::ConstantExpr::getBitCast(isa, PtrToIdTy);
// struct
// {
// Class isa;
// uint32_t flags;
// uint32_t length; // Number of codepoints
// uint32_t size; // Number of bytes
// uint32_t hash;
// const char *data;
// };
ConstantInitBuilder Builder(CGM);
auto Fields = Builder.beginStruct();
if (!CGM.getTriple().isOSBinFormatCOFF()) {
Fields.add(isa);
} else {
Fields.addNullPointer(PtrTy);
}
// For now, all non-ASCII strings are represented as UTF-16. As such, the
// number of bytes is simply double the number of UTF-16 codepoints. In
// ASCII strings, the number of bytes is equal to the number of non-ASCII
// codepoints.
if (isNonASCII) {
unsigned NumU8CodeUnits = Str.size();
// A UTF-16 representation of a unicode string contains at most the same
// number of code units as a UTF-8 representation. Allocate that much
// space, plus one for the final null character.
SmallVector<llvm::UTF16, 128> ToBuf(NumU8CodeUnits + 1);
const llvm::UTF8 *FromPtr = (const llvm::UTF8 *)Str.data();
llvm::UTF16 *ToPtr = &ToBuf[0];
(void)llvm::ConvertUTF8toUTF16(&FromPtr, FromPtr + NumU8CodeUnits,
&ToPtr, ToPtr + NumU8CodeUnits, llvm::strictConversion);
uint32_t StringLength = ToPtr - &ToBuf[0];
// Add null terminator
*ToPtr = 0;
// Flags: 2 indicates UTF-16 encoding
Fields.addInt(Int32Ty, 2);
// Number of UTF-16 codepoints
Fields.addInt(Int32Ty, StringLength);
// Number of bytes
Fields.addInt(Int32Ty, StringLength * 2);
// Hash. Not currently initialised by the compiler.
Fields.addInt(Int32Ty, 0);
// pointer to the data string.
auto Arr = llvm::makeArrayRef(&ToBuf[0], ToPtr+1);
auto *C = llvm::ConstantDataArray::get(VMContext, Arr);
auto *Buffer = new llvm::GlobalVariable(TheModule, C->getType(),
/*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, C, ".str");
Buffer->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
Fields.add(Buffer);
} else {
// Flags: 0 indicates ASCII encoding
Fields.addInt(Int32Ty, 0);
// Number of UTF-16 codepoints, each ASCII byte is a UTF-16 codepoint
Fields.addInt(Int32Ty, Str.size());
// Number of bytes
Fields.addInt(Int32Ty, Str.size());
// Hash. Not currently initialised by the compiler.
Fields.addInt(Int32Ty, 0);
// Data pointer
Fields.add(MakeConstantString(Str));
}
std::string StringName;
bool isNamed = !isNonASCII;
if (isNamed) {
StringName = ".objc_str_";
for (int i=0,e=Str.size() ; i<e ; ++i) {
unsigned char c = Str[i];
if (isalnum(c))
StringName += c;
else if (c == ' ')
StringName += '_';
else {
isNamed = false;
break;
}
}
}
auto *ObjCStrGV =
Fields.finishAndCreateGlobal(
isNamed ? StringRef(StringName) : ".objc_string",
Align, false, isNamed ? llvm::GlobalValue::LinkOnceODRLinkage
: llvm::GlobalValue::PrivateLinkage);
ObjCStrGV->setSection(sectionName<ConstantStringSection>());
if (isNamed) {
ObjCStrGV->setComdat(TheModule.getOrInsertComdat(StringName));
ObjCStrGV->setVisibility(llvm::GlobalValue::HiddenVisibility);
}
if (CGM.getTriple().isOSBinFormatCOFF()) {
std::pair<llvm::Constant*, int> v{ObjCStrGV, 0};
EarlyInitList.emplace_back(Sym, v);
}
llvm::Constant *ObjCStr = llvm::ConstantExpr::getBitCast(ObjCStrGV, IdTy);
ObjCStrings[Str] = ObjCStr;
ConstantStrings.push_back(ObjCStr);
return ConstantAddress(ObjCStr, Align);
}
void PushProperty(ConstantArrayBuilder &PropertiesArray,
const ObjCPropertyDecl *property,
const Decl *OCD,
bool isSynthesized=true, bool
isDynamic=true) override {
// struct objc_property
// {
// const char *name;
// const char *attributes;
// const char *type;
// SEL getter;
// SEL setter;
// };
auto Fields = PropertiesArray.beginStruct(PropertyMetadataTy);
ASTContext &Context = CGM.getContext();
Fields.add(MakeConstantString(property->getNameAsString()));
std::string TypeStr =
CGM.getContext().getObjCEncodingForPropertyDecl(property, OCD);
Fields.add(MakeConstantString(TypeStr));
std::string typeStr;
Context.getObjCEncodingForType(property->getType(), typeStr);
Fields.add(MakeConstantString(typeStr));
auto addPropertyMethod = [&](const ObjCMethodDecl *accessor) {
if (accessor) {
std::string TypeStr = Context.getObjCEncodingForMethodDecl(accessor);
Fields.add(GetConstantSelector(accessor->getSelector(), TypeStr));
} else {
Fields.add(NULLPtr);
}
};
addPropertyMethod(property->getGetterMethodDecl());
addPropertyMethod(property->getSetterMethodDecl());
Fields.finishAndAddTo(PropertiesArray);
}
llvm::Constant *
GenerateProtocolMethodList(ArrayRef<const ObjCMethodDecl*> Methods) override {
// struct objc_protocol_method_description
// {
// SEL selector;
// const char *types;
// };
llvm::StructType *ObjCMethodDescTy =
llvm::StructType::get(CGM.getLLVMContext(),
{ PtrToInt8Ty, PtrToInt8Ty });
ASTContext &Context = CGM.getContext();
ConstantInitBuilder Builder(CGM);
// struct objc_protocol_method_description_list
// {
// int count;
// int size;
// struct objc_protocol_method_description methods[];
// };
auto MethodList = Builder.beginStruct();
// int count;
MethodList.addInt(IntTy, Methods.size());
// int size; // sizeof(struct objc_method_description)
llvm::DataLayout td(&TheModule);
MethodList.addInt(IntTy, td.getTypeSizeInBits(ObjCMethodDescTy) /
CGM.getContext().getCharWidth());
// struct objc_method_description[]
auto MethodArray = MethodList.beginArray(ObjCMethodDescTy);
for (auto *M : Methods) {
auto Method = MethodArray.beginStruct(ObjCMethodDescTy);
Method.add(CGObjCGNU::GetConstantSelector(M));
Method.add(GetTypeString(Context.getObjCEncodingForMethodDecl(M, true)));
Method.finishAndAddTo(MethodArray);
}
MethodArray.finishAndAddTo(MethodList);
return MethodList.finishAndCreateGlobal(".objc_protocol_method_list",
CGM.getPointerAlign());
}
llvm::Constant *GenerateCategoryProtocolList(const ObjCCategoryDecl *OCD)
override {
SmallVector<llvm::Constant*, 16> Protocols;
for (const auto *PI : OCD->getReferencedProtocols())
Protocols.push_back(
llvm::ConstantExpr::getBitCast(GenerateProtocolRef(PI),
ProtocolPtrTy));
return GenerateProtocolList(Protocols);
}
llvm::Value *LookupIMPSuper(CodeGenFunction &CGF, Address ObjCSuper,
llvm::Value *cmd, MessageSendInfo &MSI) override {
// Don't access the slot unless we're trying to cache the result.
CGBuilderTy &Builder = CGF.Builder;
llvm::Value *lookupArgs[] = {CGObjCGNU::EnforceType(Builder, ObjCSuper,
PtrToObjCSuperTy).getPointer(), cmd};
return CGF.EmitNounwindRuntimeCall(MsgLookupSuperFn, lookupArgs);
}
llvm::GlobalVariable *GetClassVar(StringRef Name, bool isWeak=false) {
std::string SymbolName = SymbolForClassRef(Name, isWeak);
auto *ClassSymbol = TheModule.getNamedGlobal(SymbolName);
if (ClassSymbol)
return ClassSymbol;
ClassSymbol = new llvm::GlobalVariable(TheModule,
IdTy, false, llvm::GlobalValue::ExternalLinkage,
nullptr, SymbolName);
// If this is a weak symbol, then we are creating a valid definition for
// the symbol, pointing to a weak definition of the real class pointer. If
// this is not a weak reference, then we are expecting another compilation
// unit to provide the real indirection symbol.
if (isWeak)
ClassSymbol->setInitializer(new llvm::GlobalVariable(TheModule,
Int8Ty, false, llvm::GlobalValue::ExternalWeakLinkage,
nullptr, SymbolForClass(Name)));
else {
if (CGM.getTriple().isOSBinFormatCOFF()) {
IdentifierInfo &II = CGM.getContext().Idents.get(Name);
TranslationUnitDecl *TUDecl = CGM.getContext().getTranslationUnitDecl();
DeclContext *DC = TranslationUnitDecl::castToDeclContext(TUDecl);
const ObjCInterfaceDecl *OID = nullptr;
for (const auto &Result : DC->lookup(&II))
if ((OID = dyn_cast<ObjCInterfaceDecl>(Result)))
break;
// The first Interface we find may be a @class,
// which should only be treated as the source of
// truth in the absence of a true declaration.
const ObjCInterfaceDecl *OIDDef = OID->getDefinition();
if (OIDDef != nullptr)
OID = OIDDef;
auto Storage = llvm::GlobalValue::DefaultStorageClass;
if (OID->hasAttr<DLLImportAttr>())
Storage = llvm::GlobalValue::DLLImportStorageClass;
else if (OID->hasAttr<DLLExportAttr>())
Storage = llvm::GlobalValue::DLLExportStorageClass;
cast<llvm::GlobalValue>(ClassSymbol)->setDLLStorageClass(Storage);
}
}
assert(ClassSymbol->getName() == SymbolName);
return ClassSymbol;
}
llvm::Value *GetClassNamed(CodeGenFunction &CGF,
const std::string &Name,
bool isWeak) override {
return CGF.Builder.CreateLoad(Address(GetClassVar(Name, isWeak),
CGM.getPointerAlign()));
}
int32_t FlagsForOwnership(Qualifiers::ObjCLifetime Ownership) {
// typedef enum {
// ownership_invalid = 0,
// ownership_strong = 1,
// ownership_weak = 2,
// ownership_unsafe = 3
// } ivar_ownership;
int Flag;
switch (Ownership) {
case Qualifiers::OCL_Strong:
Flag = 1;
break;
case Qualifiers::OCL_Weak:
Flag = 2;
break;
case Qualifiers::OCL_ExplicitNone:
Flag = 3;
break;
case Qualifiers::OCL_None:
case Qualifiers::OCL_Autoreleasing:
assert(Ownership != Qualifiers::OCL_Autoreleasing);
Flag = 0;
}
return Flag;
}
llvm::Constant *GenerateIvarList(ArrayRef<llvm::Constant *> IvarNames,
ArrayRef<llvm::Constant *> IvarTypes,
ArrayRef<llvm::Constant *> IvarOffsets,
ArrayRef<llvm::Constant *> IvarAlign,
ArrayRef<Qualifiers::ObjCLifetime> IvarOwnership) override {
llvm_unreachable("Method should not be called!");
}
llvm::Constant *GenerateEmptyProtocol(StringRef ProtocolName) override {
std::string Name = SymbolForProtocol(ProtocolName);
auto *GV = TheModule.getGlobalVariable(Name);
if (!GV) {
// Emit a placeholder symbol.
GV = new llvm::GlobalVariable(TheModule, ProtocolTy, false,
llvm::GlobalValue::ExternalLinkage, nullptr, Name);
GV->setAlignment(CGM.getPointerAlign().getQuantity());
}
return llvm::ConstantExpr::getBitCast(GV, ProtocolPtrTy);
}
/// Existing protocol references.
llvm::StringMap<llvm::Constant*> ExistingProtocolRefs;
llvm::Value *GenerateProtocolRef(CodeGenFunction &CGF,
const ObjCProtocolDecl *PD) override {
auto Name = PD->getNameAsString();
auto *&Ref = ExistingProtocolRefs[Name];
if (!Ref) {
auto *&Protocol = ExistingProtocols[Name];
if (!Protocol)
Protocol = GenerateProtocolRef(PD);
std::string RefName = SymbolForProtocolRef(Name);
assert(!TheModule.getGlobalVariable(RefName));
// Emit a reference symbol.
auto GV = new llvm::GlobalVariable(TheModule, ProtocolPtrTy,
false, llvm::GlobalValue::LinkOnceODRLinkage,
llvm::ConstantExpr::getBitCast(Protocol, ProtocolPtrTy), RefName);
GV->setComdat(TheModule.getOrInsertComdat(RefName));
GV->setSection(sectionName<ProtocolReferenceSection>());
GV->setAlignment(CGM.getPointerAlign().getQuantity());
Ref = GV;
}
EmittedProtocolRef = true;
return CGF.Builder.CreateAlignedLoad(Ref, CGM.getPointerAlign());
}
llvm::Constant *GenerateProtocolList(ArrayRef<llvm::Constant*> Protocols) {
llvm::ArrayType *ProtocolArrayTy = llvm::ArrayType::get(ProtocolPtrTy,
Protocols.size());
llvm::Constant * ProtocolArray = llvm::ConstantArray::get(ProtocolArrayTy,
Protocols);
ConstantInitBuilder builder(CGM);
auto ProtocolBuilder = builder.beginStruct();
ProtocolBuilder.addNullPointer(PtrTy);
ProtocolBuilder.addInt(SizeTy, Protocols.size());
ProtocolBuilder.add(ProtocolArray);
return ProtocolBuilder.finishAndCreateGlobal(".objc_protocol_list",
CGM.getPointerAlign(), false, llvm::GlobalValue::InternalLinkage);
}
void GenerateProtocol(const ObjCProtocolDecl *PD) override {
// Do nothing - we only emit referenced protocols.
}
llvm::Constant *GenerateProtocolRef(const ObjCProtocolDecl *PD) {
std::string ProtocolName = PD->getNameAsString();
auto *&Protocol = ExistingProtocols[ProtocolName];
if (Protocol)
return Protocol;
EmittedProtocol = true;
auto SymName = SymbolForProtocol(ProtocolName);
auto *OldGV = TheModule.getGlobalVariable(SymName);
// Use the protocol definition, if there is one.
if (const ObjCProtocolDecl *Def = PD->getDefinition())
PD = Def;
else {
// If there is no definition, then create an external linkage symbol and
// hope that someone else fills it in for us (and fail to link if they
// don't).
assert(!OldGV);
Protocol = new llvm::GlobalVariable(TheModule, ProtocolTy,
/*isConstant*/false,
llvm::GlobalValue::ExternalLinkage, nullptr, SymName);
return Protocol;
}
SmallVector<llvm::Constant*, 16> Protocols;
for (const auto *PI : PD->protocols())
Protocols.push_back(
llvm::ConstantExpr::getBitCast(GenerateProtocolRef(PI),
ProtocolPtrTy));
llvm::Constant *ProtocolList = GenerateProtocolList(Protocols);
// Collect information about methods
llvm::Constant *InstanceMethodList, *OptionalInstanceMethodList;
llvm::Constant *ClassMethodList, *OptionalClassMethodList;
EmitProtocolMethodList(PD->instance_methods(), InstanceMethodList,
OptionalInstanceMethodList);
EmitProtocolMethodList(PD->class_methods(), ClassMethodList,
OptionalClassMethodList);
// The isa pointer must be set to a magic number so the runtime knows it's
// the correct layout.
ConstantInitBuilder builder(CGM);
auto ProtocolBuilder = builder.beginStruct();
ProtocolBuilder.add(llvm::ConstantExpr::getIntToPtr(
llvm::ConstantInt::get(Int32Ty, ProtocolVersion), IdTy));
ProtocolBuilder.add(MakeConstantString(ProtocolName));
ProtocolBuilder.add(ProtocolList);
ProtocolBuilder.add(InstanceMethodList);
ProtocolBuilder.add(ClassMethodList);
ProtocolBuilder.add(OptionalInstanceMethodList);
ProtocolBuilder.add(OptionalClassMethodList);
// Required instance properties
ProtocolBuilder.add(GeneratePropertyList(nullptr, PD, false, false));
// Optional instance properties
ProtocolBuilder.add(GeneratePropertyList(nullptr, PD, false, true));
// Required class properties
ProtocolBuilder.add(GeneratePropertyList(nullptr, PD, true, false));
// Optional class properties
ProtocolBuilder.add(GeneratePropertyList(nullptr, PD, true, true));
auto *GV = ProtocolBuilder.finishAndCreateGlobal(SymName,
CGM.getPointerAlign(), false, llvm::GlobalValue::ExternalLinkage);
GV->setSection(sectionName<ProtocolSection>());
GV->setComdat(TheModule.getOrInsertComdat(SymName));
if (OldGV) {
OldGV->replaceAllUsesWith(llvm::ConstantExpr::getBitCast(GV,
OldGV->getType()));
OldGV->removeFromParent();
GV->setName(SymName);
}
Protocol = GV;
return GV;
}
llvm::Constant *EnforceType(llvm::Constant *Val, llvm::Type *Ty) {
if (Val->getType() == Ty)
return Val;
return llvm::ConstantExpr::getBitCast(Val, Ty);
}
llvm::Value *GetTypedSelector(CodeGenFunction &CGF, Selector Sel,
const std::string &TypeEncoding) override {
return GetConstantSelector(Sel, TypeEncoding);
}
llvm::Constant *GetTypeString(llvm::StringRef TypeEncoding) {
if (TypeEncoding.empty())
return NULLPtr;
std::string MangledTypes = TypeEncoding;
std::replace(MangledTypes.begin(), MangledTypes.end(),
'@', '\1');
std::string TypesVarName = ".objc_sel_types_" + MangledTypes;
auto *TypesGlobal = TheModule.getGlobalVariable(TypesVarName);
if (!TypesGlobal) {
llvm::Constant *Init = llvm::ConstantDataArray::getString(VMContext,
TypeEncoding);
auto *GV = new llvm::GlobalVariable(TheModule, Init->getType(),
true, llvm::GlobalValue::LinkOnceODRLinkage, Init, TypesVarName);
GV->setComdat(TheModule.getOrInsertComdat(TypesVarName));
GV->setVisibility(llvm::GlobalValue::HiddenVisibility);
TypesGlobal = GV;
}
return llvm::ConstantExpr::getGetElementPtr(TypesGlobal->getValueType(),
TypesGlobal, Zeros);
}
llvm::Constant *GetConstantSelector(Selector Sel,
const std::string &TypeEncoding) override {
// @ is used as a special character in symbol names (used for symbol
// versioning), so mangle the name to not include it. Replace it with a
// character that is not a valid type encoding character (and, being
// non-printable, never will be!)
std::string MangledTypes = TypeEncoding;
std::replace(MangledTypes.begin(), MangledTypes.end(),
'@', '\1');
auto SelVarName = (StringRef(".objc_selector_") + Sel.getAsString() + "_" +
MangledTypes).str();
if (auto *GV = TheModule.getNamedGlobal(SelVarName))
return EnforceType(GV, SelectorTy);
ConstantInitBuilder builder(CGM);
auto SelBuilder = builder.beginStruct();
SelBuilder.add(ExportUniqueString(Sel.getAsString(), ".objc_sel_name_",
true));
SelBuilder.add(GetTypeString(TypeEncoding));
auto *GV = SelBuilder.finishAndCreateGlobal(SelVarName,
CGM.getPointerAlign(), false, llvm::GlobalValue::LinkOnceODRLinkage);
GV->setComdat(TheModule.getOrInsertComdat(SelVarName));
GV->setVisibility(llvm::GlobalValue::HiddenVisibility);
GV->setSection(sectionName<SelectorSection>());
auto *SelVal = EnforceType(GV, SelectorTy);
return SelVal;
}
llvm::StructType *emptyStruct = nullptr;
/// Return pointers to the start and end of a section. On ELF platforms, we
/// use the __start_ and __stop_ symbols that GNU-compatible linkers will set
/// to the start and end of section names, as long as those section names are
/// valid identifiers and the symbols are referenced but not defined. On
/// Windows, we use the fact that MSVC-compatible linkers will lexically sort
/// by subsections and place everything that we want to reference in a middle
/// subsection and then insert zero-sized symbols in subsections a and z.
std::pair<llvm::Constant*,llvm::Constant*>
GetSectionBounds(StringRef Section) {
if (CGM.getTriple().isOSBinFormatCOFF()) {
if (emptyStruct == nullptr) {
emptyStruct = llvm::StructType::create(VMContext, ".objc_section_sentinel");
emptyStruct->setBody({}, /*isPacked*/true);
}
auto ZeroInit = llvm::Constant::getNullValue(emptyStruct);
auto Sym = [&](StringRef Prefix, StringRef SecSuffix) {
auto *Sym = new llvm::GlobalVariable(TheModule, emptyStruct,
/*isConstant*/false,
llvm::GlobalValue::LinkOnceODRLinkage, ZeroInit, Prefix +
Section);
Sym->setVisibility(llvm::GlobalValue::HiddenVisibility);
Sym->setSection((Section + SecSuffix).str());
Sym->setComdat(TheModule.getOrInsertComdat((Prefix +
Section).str()));
Sym->setAlignment(CGM.getPointerAlign().getQuantity());
return Sym;
};
return { Sym("__start_", "$a"), Sym("__stop", "$z") };
}
auto *Start = new llvm::GlobalVariable(TheModule, PtrTy,
/*isConstant*/false,
llvm::GlobalValue::ExternalLinkage, nullptr, StringRef("__start_") +
Section);
Start->setVisibility(llvm::GlobalValue::HiddenVisibility);
auto *Stop = new llvm::GlobalVariable(TheModule, PtrTy,
/*isConstant*/false,
llvm::GlobalValue::ExternalLinkage, nullptr, StringRef("__stop_") +
Section);
Stop->setVisibility(llvm::GlobalValue::HiddenVisibility);
return { Start, Stop };
}
CatchTypeInfo getCatchAllTypeInfo() override {
return CGM.getCXXABI().getCatchAllTypeInfo();
}
llvm::Function *ModuleInitFunction() override {
llvm::Function *LoadFunction = llvm::Function::Create(
llvm::FunctionType::get(llvm::Type::getVoidTy(VMContext), false),
llvm::GlobalValue::LinkOnceODRLinkage, ".objcv2_load_function",
&TheModule);
LoadFunction->setVisibility(llvm::GlobalValue::HiddenVisibility);
LoadFunction->setComdat(TheModule.getOrInsertComdat(".objcv2_load_function"));
llvm::BasicBlock *EntryBB =
llvm::BasicBlock::Create(VMContext, "entry", LoadFunction);
CGBuilderTy B(CGM, VMContext);
B.SetInsertPoint(EntryBB);
ConstantInitBuilder builder(CGM);
auto InitStructBuilder = builder.beginStruct();
InitStructBuilder.addInt(Int64Ty, 0);
auto &sectionVec = CGM.getTriple().isOSBinFormatCOFF() ? PECOFFSectionsBaseNames : SectionsBaseNames;
for (auto *s : sectionVec) {
auto bounds = GetSectionBounds(s);
InitStructBuilder.add(bounds.first);
InitStructBuilder.add(bounds.second);
}
auto *InitStruct = InitStructBuilder.finishAndCreateGlobal(".objc_init",
CGM.getPointerAlign(), false, llvm::GlobalValue::LinkOnceODRLinkage);
InitStruct->setVisibility(llvm::GlobalValue::HiddenVisibility);
InitStruct->setComdat(TheModule.getOrInsertComdat(".objc_init"));
CallRuntimeFunction(B, "__objc_load", {InitStruct});;
B.CreateRetVoid();
// Make sure that the optimisers don't delete this function.
CGM.addCompilerUsedGlobal(LoadFunction);
// FIXME: Currently ELF only!
// We have to do this by hand, rather than with @llvm.ctors, so that the
// linker can remove the duplicate invocations.
auto *InitVar = new llvm::GlobalVariable(TheModule, LoadFunction->getType(),
/*isConstant*/true, llvm::GlobalValue::LinkOnceAnyLinkage,
LoadFunction, ".objc_ctor");
// Check that this hasn't been renamed. This shouldn't happen, because
// this function should be called precisely once.
assert(InitVar->getName() == ".objc_ctor");
// In Windows, initialisers are sorted by the suffix. XCL is for library
// initialisers, which run before user initialisers. We are running
// Objective-C loads at the end of library load. This means +load methods
// will run before any other static constructors, but that static
// constructors can see a fully initialised Objective-C state.
if (CGM.getTriple().isOSBinFormatCOFF())
InitVar->setSection(".CRT$XCLz");
else
{
if (CGM.getCodeGenOpts().UseInitArray)
InitVar->setSection(".init_array");
else
InitVar->setSection(".ctors");
}
InitVar->setVisibility(llvm::GlobalValue::HiddenVisibility);
InitVar->setComdat(TheModule.getOrInsertComdat(".objc_ctor"));
CGM.addUsedGlobal(InitVar);
for (auto *C : Categories) {
auto *Cat = cast<llvm::GlobalVariable>(C->stripPointerCasts());
Cat->setSection(sectionName<CategorySection>());
CGM.addUsedGlobal(Cat);
}
auto createNullGlobal = [&](StringRef Name, ArrayRef<llvm::Constant*> Init,
StringRef Section) {
auto nullBuilder = builder.beginStruct();
for (auto *F : Init)
nullBuilder.add(F);
auto GV = nullBuilder.finishAndCreateGlobal(Name, CGM.getPointerAlign(),
false, llvm::GlobalValue::LinkOnceODRLinkage);
GV->setSection(Section);
GV->setComdat(TheModule.getOrInsertComdat(Name));
GV->setVisibility(llvm::GlobalValue::HiddenVisibility);
CGM.addUsedGlobal(GV);
return GV;
};
for (auto clsAlias : ClassAliases)
createNullGlobal(std::string(".objc_class_alias") +
clsAlias.second, { MakeConstantString(clsAlias.second),
GetClassVar(clsAlias.first) }, sectionName<ClassAliasSection>());
// On ELF platforms, add a null value for each special section so that we
// can always guarantee that the _start and _stop symbols will exist and be
// meaningful. This is not required on COFF platforms, where our start and
// stop symbols will create the section.
if (!CGM.getTriple().isOSBinFormatCOFF()) {
createNullGlobal(".objc_null_selector", {NULLPtr, NULLPtr},
sectionName<SelectorSection>());
if (Categories.empty())
createNullGlobal(".objc_null_category", {NULLPtr, NULLPtr,
NULLPtr, NULLPtr, NULLPtr, NULLPtr, NULLPtr},
sectionName<CategorySection>());
if (!EmittedClass) {
createNullGlobal(".objc_null_cls_init_ref", NULLPtr,
sectionName<ClassSection>());
createNullGlobal(".objc_null_class_ref", { NULLPtr, NULLPtr },
sectionName<ClassReferenceSection>());
}
if (!EmittedProtocol)
createNullGlobal(".objc_null_protocol", {NULLPtr, NULLPtr, NULLPtr,
NULLPtr, NULLPtr, NULLPtr, NULLPtr, NULLPtr, NULLPtr, NULLPtr,
NULLPtr}, sectionName<ProtocolSection>());
if (!EmittedProtocolRef)
createNullGlobal(".objc_null_protocol_ref", {NULLPtr},
sectionName<ProtocolReferenceSection>());
if (ClassAliases.empty())
createNullGlobal(".objc_null_class_alias", { NULLPtr, NULLPtr },
sectionName<ClassAliasSection>());
if (ConstantStrings.empty()) {
auto i32Zero = llvm::ConstantInt::get(Int32Ty, 0);
createNullGlobal(".objc_null_constant_string", { NULLPtr, i32Zero,
i32Zero, i32Zero, i32Zero, NULLPtr },
sectionName<ConstantStringSection>());
}
}
ConstantStrings.clear();
Categories.clear();
Classes.clear();
if (EarlyInitList.size() > 0) {
auto *Init = llvm::Function::Create(llvm::FunctionType::get(CGM.VoidTy,
{}), llvm::GlobalValue::InternalLinkage, ".objc_early_init",
&CGM.getModule());
llvm::IRBuilder<> b(llvm::BasicBlock::Create(CGM.getLLVMContext(), "entry",
Init));
for (const auto &lateInit : EarlyInitList) {
auto *global = TheModule.getGlobalVariable(lateInit.first);
if (global) {
b.CreateAlignedStore(global,
b.CreateStructGEP(lateInit.second.first, lateInit.second.second), CGM.getPointerAlign().getQuantity());
}
}
b.CreateRetVoid();
// We can't use the normal LLVM global initialisation array, because we
// need to specify that this runs early in library initialisation.
auto *InitVar = new llvm::GlobalVariable(CGM.getModule(), Init->getType(),
/*isConstant*/true, llvm::GlobalValue::InternalLinkage,
Init, ".objc_early_init_ptr");
InitVar->setSection(".CRT$XCLb");
CGM.addUsedGlobal(InitVar);
}
return nullptr;
}
/// In the v2 ABI, ivar offset variables use the type encoding in their name
/// to trigger linker failures if the types don't match.
std::string GetIVarOffsetVariableName(const ObjCInterfaceDecl *ID,
const ObjCIvarDecl *Ivar) override {
std::string TypeEncoding;
CGM.getContext().getObjCEncodingForType(Ivar->getType(), TypeEncoding);
// Prevent the @ from being interpreted as a symbol version.
std::replace(TypeEncoding.begin(), TypeEncoding.end(),
'@', '\1');
const std::string Name = "__objc_ivar_offset_" + ID->getNameAsString()
+ '.' + Ivar->getNameAsString() + '.' + TypeEncoding;
return Name;
}
llvm::Value *EmitIvarOffset(CodeGenFunction &CGF,
const ObjCInterfaceDecl *Interface,
const ObjCIvarDecl *Ivar) override {
const std::string Name = GetIVarOffsetVariableName(Ivar->getContainingInterface(), Ivar);
llvm::GlobalVariable *IvarOffsetPointer = TheModule.getNamedGlobal(Name);
if (!IvarOffsetPointer)
IvarOffsetPointer = new llvm::GlobalVariable(TheModule, IntTy, false,
llvm::GlobalValue::ExternalLinkage, nullptr, Name);
CharUnits Align = CGM.getIntAlign();
llvm::Value *Offset = CGF.Builder.CreateAlignedLoad(IvarOffsetPointer, Align);
if (Offset->getType() != PtrDiffTy)
Offset = CGF.Builder.CreateZExtOrBitCast(Offset, PtrDiffTy);
return Offset;
}
void GenerateClass(const ObjCImplementationDecl *OID) override {
ASTContext &Context = CGM.getContext();
bool IsCOFF = CGM.getTriple().isOSBinFormatCOFF();
// Get the class name
ObjCInterfaceDecl *classDecl =
const_cast<ObjCInterfaceDecl *>(OID->getClassInterface());
std::string className = classDecl->getNameAsString();
auto *classNameConstant = MakeConstantString(className);
ConstantInitBuilder builder(CGM);
auto metaclassFields = builder.beginStruct();
// struct objc_class *isa;
metaclassFields.addNullPointer(PtrTy);
// struct objc_class *super_class;
metaclassFields.addNullPointer(PtrTy);
// const char *name;
metaclassFields.add(classNameConstant);
// long version;
metaclassFields.addInt(LongTy, 0);
// unsigned long info;
// objc_class_flag_meta
metaclassFields.addInt(LongTy, 1);
// long instance_size;
// Setting this to zero is consistent with the older ABI, but it might be
// more sensible to set this to sizeof(struct objc_class)
metaclassFields.addInt(LongTy, 0);
// struct objc_ivar_list *ivars;
metaclassFields.addNullPointer(PtrTy);
// struct objc_method_list *methods
// FIXME: Almost identical code is copied and pasted below for the
// class, but refactoring it cleanly requires C++14 generic lambdas.
if (OID->classmeth_begin() == OID->classmeth_end())
metaclassFields.addNullPointer(PtrTy);
else {
SmallVector<ObjCMethodDecl*, 16> ClassMethods;
ClassMethods.insert(ClassMethods.begin(), OID->classmeth_begin(),
OID->classmeth_end());
metaclassFields.addBitCast(
GenerateMethodList(className, "", ClassMethods, true),
PtrTy);
}
// void *dtable;
metaclassFields.addNullPointer(PtrTy);
// IMP cxx_construct;
metaclassFields.addNullPointer(PtrTy);
// IMP cxx_destruct;
metaclassFields.addNullPointer(PtrTy);
// struct objc_class *subclass_list
metaclassFields.addNullPointer(PtrTy);
// struct objc_class *sibling_class
metaclassFields.addNullPointer(PtrTy);
// struct objc_protocol_list *protocols;
metaclassFields.addNullPointer(PtrTy);
// struct reference_list *extra_data;
metaclassFields.addNullPointer(PtrTy);
// long abi_version;
metaclassFields.addInt(LongTy, 0);
// struct objc_property_list *properties
metaclassFields.add(GeneratePropertyList(OID, classDecl, /*isClassProperty*/true));
auto *metaclass = metaclassFields.finishAndCreateGlobal(
ManglePublicSymbol("OBJC_METACLASS_") + className,
CGM.getPointerAlign());
auto classFields = builder.beginStruct();
// struct objc_class *isa;
classFields.add(metaclass);
// struct objc_class *super_class;
// Get the superclass name.
const ObjCInterfaceDecl * SuperClassDecl =
OID->getClassInterface()->getSuperClass();
llvm::Constant *SuperClass = nullptr;
if (SuperClassDecl) {
auto SuperClassName = SymbolForClass(SuperClassDecl->getNameAsString());
SuperClass = TheModule.getNamedGlobal(SuperClassName);
if (!SuperClass)
{
SuperClass = new llvm::GlobalVariable(TheModule, PtrTy, false,
llvm::GlobalValue::ExternalLinkage, nullptr, SuperClassName);
if (IsCOFF) {
auto Storage = llvm::GlobalValue::DefaultStorageClass;
if (SuperClassDecl->hasAttr<DLLImportAttr>())
Storage = llvm::GlobalValue::DLLImportStorageClass;
else if (SuperClassDecl->hasAttr<DLLExportAttr>())
Storage = llvm::GlobalValue::DLLExportStorageClass;
cast<llvm::GlobalValue>(SuperClass)->setDLLStorageClass(Storage);
}
}
if (!IsCOFF)
classFields.add(llvm::ConstantExpr::getBitCast(SuperClass, PtrTy));
else
classFields.addNullPointer(PtrTy);
} else
classFields.addNullPointer(PtrTy);
// const char *name;
classFields.add(classNameConstant);
// long version;
classFields.addInt(LongTy, 0);
// unsigned long info;
// !objc_class_flag_meta
classFields.addInt(LongTy, 0);
// long instance_size;
int superInstanceSize = !SuperClassDecl ? 0 :
Context.getASTObjCInterfaceLayout(SuperClassDecl).getSize().getQuantity();
// Instance size is negative for classes that have not yet had their ivar
// layout calculated.
classFields.addInt(LongTy,
0 - (Context.getASTObjCImplementationLayout(OID).getSize().getQuantity() -
superInstanceSize));
if (classDecl->all_declared_ivar_begin() == nullptr)
classFields.addNullPointer(PtrTy);
else {
int ivar_count = 0;
for (const ObjCIvarDecl *IVD = classDecl->all_declared_ivar_begin(); IVD;
IVD = IVD->getNextIvar()) ivar_count++;
llvm::DataLayout td(&TheModule);
// struct objc_ivar_list *ivars;
ConstantInitBuilder b(CGM);
auto ivarListBuilder = b.beginStruct();
// int count;
ivarListBuilder.addInt(IntTy, ivar_count);
// size_t size;
llvm::StructType *ObjCIvarTy = llvm::StructType::get(
PtrToInt8Ty,
PtrToInt8Ty,
PtrToInt8Ty,
Int32Ty,
Int32Ty);
ivarListBuilder.addInt(SizeTy, td.getTypeSizeInBits(ObjCIvarTy) /
CGM.getContext().getCharWidth());
// struct objc_ivar ivars[]
auto ivarArrayBuilder = ivarListBuilder.beginArray();
for (const ObjCIvarDecl *IVD = classDecl->all_declared_ivar_begin(); IVD;
IVD = IVD->getNextIvar()) {
auto ivarTy = IVD->getType();
auto ivarBuilder = ivarArrayBuilder.beginStruct();
// const char *name;
ivarBuilder.add(MakeConstantString(IVD->getNameAsString()));
// const char *type;
std::string TypeStr;
//Context.getObjCEncodingForType(ivarTy, TypeStr, IVD, true);
Context.getObjCEncodingForMethodParameter(Decl::OBJC_TQ_None, ivarTy, TypeStr, true);
ivarBuilder.add(MakeConstantString(TypeStr));
// int *offset;
uint64_t BaseOffset = ComputeIvarBaseOffset(CGM, OID, IVD);
uint64_t Offset = BaseOffset - superInstanceSize;
llvm::Constant *OffsetValue = llvm::ConstantInt::get(IntTy, Offset);
std::string OffsetName = GetIVarOffsetVariableName(classDecl, IVD);
llvm::GlobalVariable *OffsetVar = TheModule.getGlobalVariable(OffsetName);
if (OffsetVar)
OffsetVar->setInitializer(OffsetValue);
else
OffsetVar = new llvm::GlobalVariable(TheModule, IntTy,
false, llvm::GlobalValue::ExternalLinkage,
OffsetValue, OffsetName);
auto ivarVisibility =
(IVD->getAccessControl() == ObjCIvarDecl::Private ||
IVD->getAccessControl() == ObjCIvarDecl::Package ||
classDecl->getVisibility() == HiddenVisibility) ?
llvm::GlobalValue::HiddenVisibility :
llvm::GlobalValue::DefaultVisibility;
OffsetVar->setVisibility(ivarVisibility);
ivarBuilder.add(OffsetVar);
// Ivar size
ivarBuilder.addInt(Int32Ty,
CGM.getContext().getTypeSizeInChars(ivarTy).getQuantity());
// Alignment will be stored as a base-2 log of the alignment.
int align = llvm::Log2_32(Context.getTypeAlignInChars(ivarTy).getQuantity());
// Objects that require more than 2^64-byte alignment should be impossible!
assert(align < 64);
// uint32_t flags;
// Bits 0-1 are ownership.
// Bit 2 indicates an extended type encoding
// Bits 3-8 contain log2(aligment)
ivarBuilder.addInt(Int32Ty,
(align << 3) | (1<<2) |
FlagsForOwnership(ivarTy.getQualifiers().getObjCLifetime()));
ivarBuilder.finishAndAddTo(ivarArrayBuilder);
}
ivarArrayBuilder.finishAndAddTo(ivarListBuilder);
auto ivarList = ivarListBuilder.finishAndCreateGlobal(".objc_ivar_list",
CGM.getPointerAlign(), /*constant*/ false,
llvm::GlobalValue::PrivateLinkage);
classFields.add(ivarList);
}
// struct objc_method_list *methods
SmallVector<const ObjCMethodDecl*, 16> InstanceMethods;
InstanceMethods.insert(InstanceMethods.begin(), OID->instmeth_begin(),
OID->instmeth_end());
for (auto *propImpl : OID->property_impls())
if (propImpl->getPropertyImplementation() ==
ObjCPropertyImplDecl::Synthesize) {
ObjCPropertyDecl *prop = propImpl->getPropertyDecl();
auto addIfExists = [&](const ObjCMethodDecl* OMD) {
if (OMD)
InstanceMethods.push_back(OMD);
};
addIfExists(prop->getGetterMethodDecl());
addIfExists(prop->getSetterMethodDecl());
}
if (InstanceMethods.size() == 0)
classFields.addNullPointer(PtrTy);
else
classFields.addBitCast(
GenerateMethodList(className, "", InstanceMethods, false),
PtrTy);
// void *dtable;
classFields.addNullPointer(PtrTy);
// IMP cxx_construct;
classFields.addNullPointer(PtrTy);
// IMP cxx_destruct;
classFields.addNullPointer(PtrTy);
// struct objc_class *subclass_list
classFields.addNullPointer(PtrTy);
// struct objc_class *sibling_class
classFields.addNullPointer(PtrTy);
// struct objc_protocol_list *protocols;
SmallVector<llvm::Constant*, 16> Protocols;
for (const auto *I : classDecl->protocols())
Protocols.push_back(
llvm::ConstantExpr::getBitCast(GenerateProtocolRef(I),
ProtocolPtrTy));
if (Protocols.empty())
classFields.addNullPointer(PtrTy);
else
classFields.add(GenerateProtocolList(Protocols));
// struct reference_list *extra_data;
classFields.addNullPointer(PtrTy);
// long abi_version;
classFields.addInt(LongTy, 0);
// struct objc_property_list *properties
classFields.add(GeneratePropertyList(OID, classDecl));
auto *classStruct =
classFields.finishAndCreateGlobal(SymbolForClass(className),
CGM.getPointerAlign(), false, llvm::GlobalValue::ExternalLinkage);
auto *classRefSymbol = GetClassVar(className);
classRefSymbol->setSection(sectionName<ClassReferenceSection>());
classRefSymbol->setInitializer(llvm::ConstantExpr::getBitCast(classStruct, IdTy));
if (IsCOFF) {
// we can't import a class struct.
if (OID->getClassInterface()->hasAttr<DLLExportAttr>()) {
cast<llvm::GlobalValue>(classStruct)->setDLLStorageClass(llvm::GlobalValue::DLLExportStorageClass);
cast<llvm::GlobalValue>(classRefSymbol)->setDLLStorageClass(llvm::GlobalValue::DLLExportStorageClass);
}
if (SuperClass) {
std::pair<llvm::Constant*, int> v{classStruct, 1};
EarlyInitList.emplace_back(SuperClass->getName(), std::move(v));
}
}
// Resolve the class aliases, if they exist.
// FIXME: Class pointer aliases shouldn't exist!
if (ClassPtrAlias) {
ClassPtrAlias->replaceAllUsesWith(
llvm::ConstantExpr::getBitCast(classStruct, IdTy));
ClassPtrAlias->eraseFromParent();
ClassPtrAlias = nullptr;
}
if (auto Placeholder =
TheModule.getNamedGlobal(SymbolForClass(className)))
if (Placeholder != classStruct) {
Placeholder->replaceAllUsesWith(
llvm::ConstantExpr::getBitCast(classStruct, Placeholder->getType()));
Placeholder->eraseFromParent();
classStruct->setName(SymbolForClass(className));
}
if (MetaClassPtrAlias) {
MetaClassPtrAlias->replaceAllUsesWith(
llvm::ConstantExpr::getBitCast(metaclass, IdTy));
MetaClassPtrAlias->eraseFromParent();
MetaClassPtrAlias = nullptr;
}
assert(classStruct->getName() == SymbolForClass(className));
auto classInitRef = new llvm::GlobalVariable(TheModule,
classStruct->getType(), false, llvm::GlobalValue::ExternalLinkage,
classStruct, ManglePublicSymbol("OBJC_INIT_CLASS_") + className);
classInitRef->setSection(sectionName<ClassSection>());
CGM.addUsedGlobal(classInitRef);
EmittedClass = true;
}
public:
CGObjCGNUstep2(CodeGenModule &Mod) : CGObjCGNUstep(Mod, 10, 4, 2) {
MsgLookupSuperFn.init(&CGM, "objc_msg_lookup_super", IMPTy,
PtrToObjCSuperTy, SelectorTy);
// struct objc_property
// {
// const char *name;
// const char *attributes;
// const char *type;
// SEL getter;
// SEL setter;
// }
PropertyMetadataTy =
llvm::StructType::get(CGM.getLLVMContext(),
{ PtrToInt8Ty, PtrToInt8Ty, PtrToInt8Ty, PtrToInt8Ty, PtrToInt8Ty });
}
};
const char *const CGObjCGNUstep2::SectionsBaseNames[8] =
{
"__objc_selectors",
"__objc_classes",
"__objc_class_refs",
"__objc_cats",
"__objc_protocols",
"__objc_protocol_refs",
"__objc_class_aliases",
"__objc_constant_string"
};
const char *const CGObjCGNUstep2::PECOFFSectionsBaseNames[8] =
{
".objcrt$SEL",
".objcrt$CLS",
".objcrt$CLR",
".objcrt$CAT",
".objcrt$PCL",
".objcrt$PCR",
".objcrt$CAL",
".objcrt$STR"
};
/// Support for the ObjFW runtime.
class CGObjCObjFW: public CGObjCGNU {
protected:
/// The GCC ABI message lookup function. Returns an IMP pointing to the
/// method implementation for this message.
LazyRuntimeFunction MsgLookupFn;
/// stret lookup function. While this does not seem to make sense at the
/// first look, this is required to call the correct forwarding function.
LazyRuntimeFunction MsgLookupFnSRet;
/// The GCC ABI superclass message lookup function. Takes a pointer to a
/// structure describing the receiver and the class, and a selector as
/// arguments. Returns the IMP for the corresponding method.
LazyRuntimeFunction MsgLookupSuperFn, MsgLookupSuperFnSRet;
llvm::Value *LookupIMP(CodeGenFunction &CGF, llvm::Value *&Receiver,
llvm::Value *cmd, llvm::MDNode *node,
MessageSendInfo &MSI) override {
CGBuilderTy &Builder = CGF.Builder;
llvm::Value *args[] = {
EnforceType(Builder, Receiver, IdTy),
EnforceType(Builder, cmd, SelectorTy) };
llvm::CallBase *imp;
if (CGM.ReturnTypeUsesSRet(MSI.CallInfo))
imp = CGF.EmitRuntimeCallOrInvoke(MsgLookupFnSRet, args);
else
imp = CGF.EmitRuntimeCallOrInvoke(MsgLookupFn, args);
imp->setMetadata(msgSendMDKind, node);
return imp;
}
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *LookupIMPSuper(CodeGenFunction &CGF, Address ObjCSuper,
llvm::Value *cmd, MessageSendInfo &MSI) override {
CGBuilderTy &Builder = CGF.Builder;
llvm::Value *lookupArgs[] = {
EnforceType(Builder, ObjCSuper.getPointer(), PtrToObjCSuperTy), cmd,
};
if (CGM.ReturnTypeUsesSRet(MSI.CallInfo))
return CGF.EmitNounwindRuntimeCall(MsgLookupSuperFnSRet, lookupArgs);
else
return CGF.EmitNounwindRuntimeCall(MsgLookupSuperFn, lookupArgs);
}
llvm::Value *GetClassNamed(CodeGenFunction &CGF, const std::string &Name,
bool isWeak) override {
if (isWeak)
return CGObjCGNU::GetClassNamed(CGF, Name, isWeak);
EmitClassRef(Name);
std::string SymbolName = "_OBJC_CLASS_" + Name;
llvm::GlobalVariable *ClassSymbol = TheModule.getGlobalVariable(SymbolName);
if (!ClassSymbol)
ClassSymbol = new llvm::GlobalVariable(TheModule, LongTy, false,
llvm::GlobalValue::ExternalLinkage,
nullptr, SymbolName);
return ClassSymbol;
}
public:
CGObjCObjFW(CodeGenModule &Mod): CGObjCGNU(Mod, 9, 3) {
// IMP objc_msg_lookup(id, SEL);
MsgLookupFn.init(&CGM, "objc_msg_lookup", IMPTy, IdTy, SelectorTy);
MsgLookupFnSRet.init(&CGM, "objc_msg_lookup_stret", IMPTy, IdTy,
SelectorTy);
// IMP objc_msg_lookup_super(struct objc_super*, SEL);
MsgLookupSuperFn.init(&CGM, "objc_msg_lookup_super", IMPTy,
PtrToObjCSuperTy, SelectorTy);
MsgLookupSuperFnSRet.init(&CGM, "objc_msg_lookup_super_stret", IMPTy,
PtrToObjCSuperTy, SelectorTy);
}
};
} // end anonymous namespace
/// Emits a reference to a dummy variable which is emitted with each class.
/// This ensures that a linker error will be generated when trying to link
/// together modules where a referenced class is not defined.
void CGObjCGNU::EmitClassRef(const std::string &className) {
std::string symbolRef = "__objc_class_ref_" + className;
// Don't emit two copies of the same symbol
if (TheModule.getGlobalVariable(symbolRef))
return;
std::string symbolName = "__objc_class_name_" + className;
llvm::GlobalVariable *ClassSymbol = TheModule.getGlobalVariable(symbolName);
if (!ClassSymbol) {
ClassSymbol = new llvm::GlobalVariable(TheModule, LongTy, false,
llvm::GlobalValue::ExternalLinkage,
nullptr, symbolName);
}
new llvm::GlobalVariable(TheModule, ClassSymbol->getType(), true,
llvm::GlobalValue::WeakAnyLinkage, ClassSymbol, symbolRef);
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
CGObjCGNU::CGObjCGNU(CodeGenModule &cgm, unsigned runtimeABIVersion,
unsigned protocolClassVersion, unsigned classABI)
: CGObjCRuntime(cgm), TheModule(CGM.getModule()),
VMContext(cgm.getLLVMContext()), ClassPtrAlias(nullptr),
MetaClassPtrAlias(nullptr), RuntimeVersion(runtimeABIVersion),
ProtocolVersion(protocolClassVersion), ClassABIVersion(classABI) {
msgSendMDKind = VMContext.getMDKindID("GNUObjCMessageSend");
usesSEHExceptions =
cgm.getContext().getTargetInfo().getTriple().isWindowsMSVCEnvironment();
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
CodeGenTypes &Types = CGM.getTypes();
IntTy = cast<llvm::IntegerType>(
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
Types.ConvertType(CGM.getContext().IntTy));
LongTy = cast<llvm::IntegerType>(
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
Types.ConvertType(CGM.getContext().LongTy));
SizeTy = cast<llvm::IntegerType>(
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
Types.ConvertType(CGM.getContext().getSizeType()));
PtrDiffTy = cast<llvm::IntegerType>(
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
Types.ConvertType(CGM.getContext().getPointerDiffType()));
BoolTy = CGM.getTypes().ConvertType(CGM.getContext().BoolTy);
Int8Ty = llvm::Type::getInt8Ty(VMContext);
// C string type. Used in lots of places.
PtrToInt8Ty = llvm::PointerType::getUnqual(Int8Ty);
ProtocolPtrTy = llvm::PointerType::getUnqual(
Types.ConvertType(CGM.getContext().getObjCProtoType()));
Zeros[0] = llvm::ConstantInt::get(LongTy, 0);
Zeros[1] = Zeros[0];
NULLPtr = llvm::ConstantPointerNull::get(PtrToInt8Ty);
// Get the selector Type.
QualType selTy = CGM.getContext().getObjCSelType();
if (QualType() == selTy) {
SelectorTy = PtrToInt8Ty;
} else {
SelectorTy = cast<llvm::PointerType>(CGM.getTypes().ConvertType(selTy));
}
PtrToIntTy = llvm::PointerType::getUnqual(IntTy);
PtrTy = PtrToInt8Ty;
Int32Ty = llvm::Type::getInt32Ty(VMContext);
Int64Ty = llvm::Type::getInt64Ty(VMContext);
IntPtrTy =
CGM.getDataLayout().getPointerSizeInBits() == 32 ? Int32Ty : Int64Ty;
// Object type
QualType UnqualIdTy = CGM.getContext().getObjCIdType();
ASTIdTy = CanQualType();
if (UnqualIdTy != QualType()) {
ASTIdTy = CGM.getContext().getCanonicalType(UnqualIdTy);
IdTy = cast<llvm::PointerType>(CGM.getTypes().ConvertType(ASTIdTy));
} else {
IdTy = PtrToInt8Ty;
}
PtrToIdTy = llvm::PointerType::getUnqual(IdTy);
ProtocolTy = llvm::StructType::get(IdTy,
PtrToInt8Ty, // name
PtrToInt8Ty, // protocols
PtrToInt8Ty, // instance methods
PtrToInt8Ty, // class methods
PtrToInt8Ty, // optional instance methods
PtrToInt8Ty, // optional class methods
PtrToInt8Ty, // properties
PtrToInt8Ty);// optional properties
// struct objc_property_gsv1
// {
// const char *name;
// char attributes;
// char attributes2;
// char unused1;
// char unused2;
// const char *getter_name;
// const char *getter_types;
// const char *setter_name;
// const char *setter_types;
// }
PropertyMetadataTy = llvm::StructType::get(CGM.getLLVMContext(), {
PtrToInt8Ty, Int8Ty, Int8Ty, Int8Ty, Int8Ty, PtrToInt8Ty, PtrToInt8Ty,
PtrToInt8Ty, PtrToInt8Ty });
ObjCSuperTy = llvm::StructType::get(IdTy, IdTy);
PtrToObjCSuperTy = llvm::PointerType::getUnqual(ObjCSuperTy);
llvm::Type *VoidTy = llvm::Type::getVoidTy(VMContext);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// void objc_exception_throw(id);
ExceptionThrowFn.init(&CGM, "objc_exception_throw", VoidTy, IdTy);
ExceptionReThrowFn.init(&CGM, "objc_exception_throw", VoidTy, IdTy);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// int objc_sync_enter(id);
SyncEnterFn.init(&CGM, "objc_sync_enter", IntTy, IdTy);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// int objc_sync_exit(id);
SyncExitFn.init(&CGM, "objc_sync_exit", IntTy, IdTy);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// void objc_enumerationMutation (id)
EnumerationMutationFn.init(&CGM, "objc_enumerationMutation", VoidTy, IdTy);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// id objc_getProperty(id, SEL, ptrdiff_t, BOOL)
GetPropertyFn.init(&CGM, "objc_getProperty", IdTy, IdTy, SelectorTy,
PtrDiffTy, BoolTy);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// void objc_setProperty(id, SEL, ptrdiff_t, id, BOOL, BOOL)
SetPropertyFn.init(&CGM, "objc_setProperty", VoidTy, IdTy, SelectorTy,
PtrDiffTy, IdTy, BoolTy, BoolTy);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// void objc_setPropertyStruct(void*, void*, ptrdiff_t, BOOL, BOOL)
GetStructPropertyFn.init(&CGM, "objc_getPropertyStruct", VoidTy, PtrTy, PtrTy,
PtrDiffTy, BoolTy, BoolTy);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// void objc_setPropertyStruct(void*, void*, ptrdiff_t, BOOL, BOOL)
SetStructPropertyFn.init(&CGM, "objc_setPropertyStruct", VoidTy, PtrTy, PtrTy,
PtrDiffTy, BoolTy, BoolTy);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// IMP type
llvm::Type *IMPArgs[] = { IdTy, SelectorTy };
IMPTy = llvm::PointerType::getUnqual(llvm::FunctionType::get(IdTy, IMPArgs,
true));
const LangOptions &Opts = CGM.getLangOpts();
if ((Opts.getGC() != LangOptions::NonGC) || Opts.ObjCAutoRefCount)
RuntimeVersion = 10;
// Don't bother initialising the GC stuff unless we're compiling in GC mode
if (Opts.getGC() != LangOptions::NonGC) {
// This is a bit of an hack. We should sort this out by having a proper
// CGObjCGNUstep subclass for GC, but we may want to really support the old
// ABI and GC added in ObjectiveC2.framework, so we fudge it a bit for now
// Get selectors needed in GC mode
RetainSel = GetNullarySelector("retain", CGM.getContext());
ReleaseSel = GetNullarySelector("release", CGM.getContext());
AutoreleaseSel = GetNullarySelector("autorelease", CGM.getContext());
// Get functions needed in GC mode
// id objc_assign_ivar(id, id, ptrdiff_t);
IvarAssignFn.init(&CGM, "objc_assign_ivar", IdTy, IdTy, IdTy, PtrDiffTy);
// id objc_assign_strongCast (id, id*)
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
StrongCastAssignFn.init(&CGM, "objc_assign_strongCast", IdTy, IdTy,
PtrToIdTy);
// id objc_assign_global(id, id*);
GlobalAssignFn.init(&CGM, "objc_assign_global", IdTy, IdTy, PtrToIdTy);
// id objc_assign_weak(id, id*);
WeakAssignFn.init(&CGM, "objc_assign_weak", IdTy, IdTy, PtrToIdTy);
// id objc_read_weak(id*);
WeakReadFn.init(&CGM, "objc_read_weak", IdTy, PtrToIdTy);
// void *objc_memmove_collectable(void*, void *, size_t);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
MemMoveFn.init(&CGM, "objc_memmove_collectable", PtrTy, PtrTy, PtrTy,
SizeTy);
}
}
llvm::Value *CGObjCGNU::GetClassNamed(CodeGenFunction &CGF,
const std::string &Name, bool isWeak) {
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Constant *ClassName = MakeConstantString(Name);
// With the incompatible ABI, this will need to be replaced with a direct
// reference to the class symbol. For the compatible nonfragile ABI we are
// still performing this lookup at run time but emitting the symbol for the
// class externally so that we can make the switch later.
//
// Libobjc2 contains an LLVM pass that replaces calls to objc_lookup_class
// with memoized versions or with static references if it's safe to do so.
if (!isWeak)
EmitClassRef(Name);
llvm::FunctionCallee ClassLookupFn = CGM.CreateRuntimeFunction(
llvm::FunctionType::get(IdTy, PtrToInt8Ty, true), "objc_lookup_class");
return CGF.EmitNounwindRuntimeCall(ClassLookupFn, ClassName);
}
// This has to perform the lookup every time, since posing and related
// techniques can modify the name -> class mapping.
llvm::Value *CGObjCGNU::GetClass(CodeGenFunction &CGF,
const ObjCInterfaceDecl *OID) {
auto *Value =
GetClassNamed(CGF, OID->getNameAsString(), OID->isWeakImported());
if (auto *ClassSymbol = dyn_cast<llvm::GlobalVariable>(Value))
CGM.setGVProperties(ClassSymbol, OID);
return Value;
}
llvm::Value *CGObjCGNU::EmitNSAutoreleasePoolClassRef(CodeGenFunction &CGF) {
auto *Value = GetClassNamed(CGF, "NSAutoreleasePool", false);
if (CGM.getTriple().isOSBinFormatCOFF()) {
if (auto *ClassSymbol = dyn_cast<llvm::GlobalVariable>(Value)) {
IdentifierInfo &II = CGF.CGM.getContext().Idents.get("NSAutoreleasePool");
TranslationUnitDecl *TUDecl = CGM.getContext().getTranslationUnitDecl();
DeclContext *DC = TranslationUnitDecl::castToDeclContext(TUDecl);
const VarDecl *VD = nullptr;
for (const auto &Result : DC->lookup(&II))
if ((VD = dyn_cast<VarDecl>(Result)))
break;
CGM.setGVProperties(ClassSymbol, VD);
}
}
return Value;
}
llvm::Value *CGObjCGNU::GetTypedSelector(CodeGenFunction &CGF, Selector Sel,
const std::string &TypeEncoding) {
SmallVectorImpl<TypedSelector> &Types = SelectorTable[Sel];
llvm::GlobalAlias *SelValue = nullptr;
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
for (SmallVectorImpl<TypedSelector>::iterator i = Types.begin(),
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
e = Types.end() ; i!=e ; i++) {
if (i->first == TypeEncoding) {
SelValue = i->second;
break;
}
}
if (!SelValue) {
SelValue = llvm::GlobalAlias::create(
SelectorTy->getElementType(), 0, llvm::GlobalValue::PrivateLinkage,
".objc_selector_" + Sel.getAsString(), &TheModule);
Types.emplace_back(TypeEncoding, SelValue);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
}
return SelValue;
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
}
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address CGObjCGNU::GetAddrOfSelector(CodeGenFunction &CGF, Selector Sel) {
llvm::Value *SelValue = GetSelector(CGF, Sel);
// Store it to a temporary. Does this satisfy the semantics of
// GetAddrOfSelector? Hopefully.
Address tmp = CGF.CreateTempAlloca(SelValue->getType(),
CGF.getPointerAlign());
CGF.Builder.CreateStore(SelValue, tmp);
return tmp;
}
llvm::Value *CGObjCGNU::GetSelector(CodeGenFunction &CGF, Selector Sel) {
return GetTypedSelector(CGF, Sel, std::string());
}
llvm::Value *CGObjCGNU::GetSelector(CodeGenFunction &CGF,
const ObjCMethodDecl *Method) {
std::string SelTypes = CGM.getContext().getObjCEncodingForMethodDecl(Method);
return GetTypedSelector(CGF, Method->getSelector(), SelTypes);
}
llvm::Constant *CGObjCGNU::GetEHType(QualType T) {
if (T->isObjCIdType() || T->isObjCQualifiedIdType()) {
// With the old ABI, there was only one kind of catchall, which broke
// foreign exceptions. With the new ABI, we use __objc_id_typeinfo as
// a pointer indicating object catchalls, and NULL to indicate real
// catchalls
if (CGM.getLangOpts().ObjCRuntime.isNonFragile()) {
return MakeConstantString("@id");
} else {
return nullptr;
}
}
// All other types should be Objective-C interface pointer types.
const ObjCObjectPointerType *OPT = T->getAs<ObjCObjectPointerType>();
assert(OPT && "Invalid @catch type.");
const ObjCInterfaceDecl *IDecl = OPT->getObjectType()->getInterface();
assert(IDecl && "Invalid @catch type.");
return MakeConstantString(IDecl->getIdentifier()->getName());
}
llvm::Constant *CGObjCGNUstep::GetEHType(QualType T) {
if (usesSEHExceptions)
return CGM.getCXXABI().getAddrOfRTTIDescriptor(T);
if (!CGM.getLangOpts().CPlusPlus)
return CGObjCGNU::GetEHType(T);
// For Objective-C++, we want to provide the ability to catch both C++ and
// Objective-C objects in the same function.
// There's a particular fixed type info for 'id'.
if (T->isObjCIdType() ||
T->isObjCQualifiedIdType()) {
llvm::Constant *IDEHType =
CGM.getModule().getGlobalVariable("__objc_id_type_info");
if (!IDEHType)
IDEHType =
new llvm::GlobalVariable(CGM.getModule(), PtrToInt8Ty,
false,
llvm::GlobalValue::ExternalLinkage,
nullptr, "__objc_id_type_info");
return llvm::ConstantExpr::getBitCast(IDEHType, PtrToInt8Ty);
}
const ObjCObjectPointerType *PT =
T->getAs<ObjCObjectPointerType>();
assert(PT && "Invalid @catch type.");
const ObjCInterfaceType *IT = PT->getInterfaceType();
assert(IT && "Invalid @catch type.");
std::string className = IT->getDecl()->getIdentifier()->getName();
std::string typeinfoName = "__objc_eh_typeinfo_" + className;
// Return the existing typeinfo if it exists
llvm::Constant *typeinfo = TheModule.getGlobalVariable(typeinfoName);
if (typeinfo)
return llvm::ConstantExpr::getBitCast(typeinfo, PtrToInt8Ty);
// Otherwise create it.
// vtable for gnustep::libobjc::__objc_class_type_info
// It's quite ugly hard-coding this. Ideally we'd generate it using the host
// platform's name mangling.
const char *vtableName = "_ZTVN7gnustep7libobjc22__objc_class_type_infoE";
auto *Vtable = TheModule.getGlobalVariable(vtableName);
if (!Vtable) {
Vtable = new llvm::GlobalVariable(TheModule, PtrToInt8Ty, true,
llvm::GlobalValue::ExternalLinkage,
nullptr, vtableName);
}
llvm::Constant *Two = llvm::ConstantInt::get(IntTy, 2);
auto *BVtable = llvm::ConstantExpr::getBitCast(
llvm::ConstantExpr::getGetElementPtr(Vtable->getValueType(), Vtable, Two),
PtrToInt8Ty);
llvm::Constant *typeName =
ExportUniqueString(className, "__objc_eh_typename_");
ConstantInitBuilder builder(CGM);
auto fields = builder.beginStruct();
fields.add(BVtable);
fields.add(typeName);
llvm::Constant *TI =
fields.finishAndCreateGlobal("__objc_eh_typeinfo_" + className,
CGM.getPointerAlign(),
/*constant*/ false,
llvm::GlobalValue::LinkOnceODRLinkage);
return llvm::ConstantExpr::getBitCast(TI, PtrToInt8Ty);
}
/// Generate an NSConstantString object.
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
ConstantAddress CGObjCGNU::GenerateConstantString(const StringLiteral *SL) {
std::string Str = SL->getString().str();
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
CharUnits Align = CGM.getPointerAlign();
// Look for an existing one
llvm::StringMap<llvm::Constant*>::iterator old = ObjCStrings.find(Str);
if (old != ObjCStrings.end())
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
return ConstantAddress(old->getValue(), Align);
StringRef StringClass = CGM.getLangOpts().ObjCConstantStringClass;
if (StringClass.empty()) StringClass = "NSConstantString";
std::string Sym = "_OBJC_CLASS_";
Sym += StringClass;
llvm::Constant *isa = TheModule.getNamedGlobal(Sym);
if (!isa)
isa = new llvm::GlobalVariable(TheModule, IdTy, /* isConstant */false,
llvm::GlobalValue::ExternalWeakLinkage, nullptr, Sym);
else if (isa->getType() != PtrToIdTy)
isa = llvm::ConstantExpr::getBitCast(isa, PtrToIdTy);
ConstantInitBuilder Builder(CGM);
auto Fields = Builder.beginStruct();
Fields.add(isa);
Fields.add(MakeConstantString(Str));
Fields.addInt(IntTy, Str.size());
llvm::Constant *ObjCStr =
Fields.finishAndCreateGlobal(".objc_str", Align);
ObjCStr = llvm::ConstantExpr::getBitCast(ObjCStr, PtrToInt8Ty);
ObjCStrings[Str] = ObjCStr;
ConstantStrings.push_back(ObjCStr);
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
return ConstantAddress(ObjCStr, Align);
}
///Generates a message send where the super is the receiver. This is a message
///send to self with special delivery semantics indicating which class's method
///should be called.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
RValue
CGObjCGNU::GenerateMessageSendSuper(CodeGenFunction &CGF,
ReturnValueSlot Return,
QualType ResultType,
Selector Sel,
const ObjCInterfaceDecl *Class,
bool isCategoryImpl,
llvm::Value *Receiver,
bool IsClassMessage,
const CallArgList &CallArgs,
const ObjCMethodDecl *Method) {
2011-05-28 22:09:01 +08:00
CGBuilderTy &Builder = CGF.Builder;
if (CGM.getLangOpts().getGC() == LangOptions::GCOnly) {
if (Sel == RetainSel || Sel == AutoreleaseSel) {
2011-05-28 22:09:01 +08:00
return RValue::get(EnforceType(Builder, Receiver,
CGM.getTypes().ConvertType(ResultType)));
}
if (Sel == ReleaseSel) {
return RValue::get(nullptr);
}
}
llvm::Value *cmd = GetSelector(CGF, Sel);
CallArgList ActualArgs;
ActualArgs.add(RValue::get(EnforceType(Builder, Receiver, IdTy)), ASTIdTy);
ActualArgs.add(RValue::get(cmd), CGF.getContext().getObjCSelType());
ActualArgs.addFrom(CallArgs);
MessageSendInfo MSI = getMessageSendInfo(Method, ResultType, ActualArgs);
llvm::Value *ReceiverClass = nullptr;
bool isV2ABI = isRuntime(ObjCRuntime::GNUstep, 2);
if (isV2ABI) {
ReceiverClass = GetClassNamed(CGF,
Class->getSuperClass()->getNameAsString(), /*isWeak*/false);
if (IsClassMessage) {
// Load the isa pointer of the superclass is this is a class method.
ReceiverClass = Builder.CreateBitCast(ReceiverClass,
llvm::PointerType::getUnqual(IdTy));
ReceiverClass =
Builder.CreateAlignedLoad(ReceiverClass, CGF.getPointerAlign());
}
ReceiverClass = EnforceType(Builder, ReceiverClass, IdTy);
} else {
if (isCategoryImpl) {
llvm::FunctionCallee classLookupFunction = nullptr;
if (IsClassMessage) {
classLookupFunction = CGM.CreateRuntimeFunction(llvm::FunctionType::get(
IdTy, PtrTy, true), "objc_get_meta_class");
} else {
classLookupFunction = CGM.CreateRuntimeFunction(llvm::FunctionType::get(
IdTy, PtrTy, true), "objc_get_class");
}
ReceiverClass = Builder.CreateCall(classLookupFunction,
MakeConstantString(Class->getNameAsString()));
} else {
// Set up global aliases for the metaclass or class pointer if they do not
// already exist. These will are forward-references which will be set to
// pointers to the class and metaclass structure created for the runtime
// load function. To send a message to super, we look up the value of the
// super_class pointer from either the class or metaclass structure.
if (IsClassMessage) {
if (!MetaClassPtrAlias) {
MetaClassPtrAlias = llvm::GlobalAlias::create(
IdTy->getElementType(), 0, llvm::GlobalValue::InternalLinkage,
".objc_metaclass_ref" + Class->getNameAsString(), &TheModule);
}
ReceiverClass = MetaClassPtrAlias;
} else {
if (!ClassPtrAlias) {
ClassPtrAlias = llvm::GlobalAlias::create(
IdTy->getElementType(), 0, llvm::GlobalValue::InternalLinkage,
".objc_class_ref" + Class->getNameAsString(), &TheModule);
}
ReceiverClass = ClassPtrAlias;
}
}
// Cast the pointer to a simplified version of the class structure
llvm::Type *CastTy = llvm::StructType::get(IdTy, IdTy);
ReceiverClass = Builder.CreateBitCast(ReceiverClass,
llvm::PointerType::getUnqual(CastTy));
// Get the superclass pointer
ReceiverClass = Builder.CreateStructGEP(CastTy, ReceiverClass, 1);
// Load the superclass pointer
ReceiverClass =
Builder.CreateAlignedLoad(ReceiverClass, CGF.getPointerAlign());
}
// Construct the structure used to look up the IMP
llvm::StructType *ObjCSuperTy =
llvm::StructType::get(Receiver->getType(), IdTy);
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address ObjCSuper = CGF.CreateTempAlloca(ObjCSuperTy,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
CGF.getPointerAlign());
Builder.CreateStore(Receiver, Builder.CreateStructGEP(ObjCSuper, 0));
Builder.CreateStore(ReceiverClass, Builder.CreateStructGEP(ObjCSuper, 1));
ObjCSuper = EnforceType(Builder, ObjCSuper, PtrToObjCSuperTy);
// Get the IMP
llvm::Value *imp = LookupIMPSuper(CGF, ObjCSuper, cmd, MSI);
imp = EnforceType(Builder, imp, MSI.MessengerType);
llvm::Metadata *impMD[] = {
llvm::MDString::get(VMContext, Sel.getAsString()),
llvm::MDString::get(VMContext, Class->getSuperClass()->getNameAsString()),
llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
llvm::Type::getInt1Ty(VMContext), IsClassMessage))};
llvm::MDNode *node = llvm::MDNode::get(VMContext, impMD);
CGCallee callee(CGCalleeInfo(), imp);
llvm::CallBase *call;
RValue msgRet = CGF.EmitCall(MSI.CallInfo, callee, Return, ActualArgs, &call);
call->setMetadata(msgSendMDKind, node);
return msgRet;
}
/// Generate code for a message send expression.
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
RValue
CGObjCGNU::GenerateMessageSend(CodeGenFunction &CGF,
ReturnValueSlot Return,
QualType ResultType,
Selector Sel,
llvm::Value *Receiver,
const CallArgList &CallArgs,
const ObjCInterfaceDecl *Class,
const ObjCMethodDecl *Method) {
2011-05-28 22:09:01 +08:00
CGBuilderTy &Builder = CGF.Builder;
// Strip out message sends to retain / release in GC mode
if (CGM.getLangOpts().getGC() == LangOptions::GCOnly) {
if (Sel == RetainSel || Sel == AutoreleaseSel) {
2011-05-28 22:09:01 +08:00
return RValue::get(EnforceType(Builder, Receiver,
CGM.getTypes().ConvertType(ResultType)));
}
if (Sel == ReleaseSel) {
return RValue::get(nullptr);
}
}
// If the return type is something that goes in an integer register, the
// runtime will handle 0 returns. For other cases, we fill in the 0 value
// ourselves.
//
// The language spec says the result of this kind of message send is
// undefined, but lots of people seem to have forgotten to read that
// paragraph and insist on sending messages to nil that have structure
// returns. With GCC, this generates a random return value (whatever happens
// to be on the stack / in those registers at the time) on most platforms,
// and generates an illegal instruction trap on SPARC. With LLVM it corrupts
// the stack.
bool isPointerSizedReturn = (ResultType->isAnyPointerType() ||
ResultType->isIntegralOrEnumerationType() || ResultType->isVoidType());
llvm::BasicBlock *startBB = nullptr;
llvm::BasicBlock *messageBB = nullptr;
llvm::BasicBlock *continueBB = nullptr;
if (!isPointerSizedReturn) {
startBB = Builder.GetInsertBlock();
messageBB = CGF.createBasicBlock("msgSend");
continueBB = CGF.createBasicBlock("continue");
llvm::Value *isNil = Builder.CreateICmpEQ(Receiver,
llvm::Constant::getNullValue(Receiver->getType()));
Builder.CreateCondBr(isNil, continueBB, messageBB);
CGF.EmitBlock(messageBB);
}
IdTy = cast<llvm::PointerType>(CGM.getTypes().ConvertType(ASTIdTy));
llvm::Value *cmd;
if (Method)
cmd = GetSelector(CGF, Method);
else
cmd = GetSelector(CGF, Sel);
cmd = EnforceType(Builder, cmd, SelectorTy);
Receiver = EnforceType(Builder, Receiver, IdTy);
llvm::Metadata *impMD[] = {
llvm::MDString::get(VMContext, Sel.getAsString()),
llvm::MDString::get(VMContext, Class ? Class->getNameAsString() : ""),
llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
llvm::Type::getInt1Ty(VMContext), Class != nullptr))};
llvm::MDNode *node = llvm::MDNode::get(VMContext, impMD);
CallArgList ActualArgs;
ActualArgs.add(RValue::get(Receiver), ASTIdTy);
ActualArgs.add(RValue::get(cmd), CGF.getContext().getObjCSelType());
ActualArgs.addFrom(CallArgs);
MessageSendInfo MSI = getMessageSendInfo(Method, ResultType, ActualArgs);
// Get the IMP to call
llvm::Value *imp;
// If we have non-legacy dispatch specified, we try using the objc_msgSend()
// functions. These are not supported on all platforms (or all runtimes on a
// given platform), so we
switch (CGM.getCodeGenOpts().getObjCDispatchMethod()) {
case CodeGenOptions::Legacy:
imp = LookupIMP(CGF, Receiver, cmd, node, MSI);
break;
case CodeGenOptions::Mixed:
case CodeGenOptions::NonLegacy:
if (CGM.ReturnTypeUsesFPRet(ResultType)) {
imp =
CGM.CreateRuntimeFunction(llvm::FunctionType::get(IdTy, IdTy, true),
"objc_msgSend_fpret")
.getCallee();
} else if (CGM.ReturnTypeUsesSRet(MSI.CallInfo)) {
// The actual types here don't matter - we're going to bitcast the
// function anyway
imp =
CGM.CreateRuntimeFunction(llvm::FunctionType::get(IdTy, IdTy, true),
"objc_msgSend_stret")
.getCallee();
} else {
imp = CGM.CreateRuntimeFunction(
llvm::FunctionType::get(IdTy, IdTy, true), "objc_msgSend")
.getCallee();
}
}
// Reset the receiver in case the lookup modified it
ActualArgs[0] = CallArg(RValue::get(Receiver), ASTIdTy);
imp = EnforceType(Builder, imp, MSI.MessengerType);
llvm::CallBase *call;
CGCallee callee(CGCalleeInfo(), imp);
RValue msgRet = CGF.EmitCall(MSI.CallInfo, callee, Return, ActualArgs, &call);
call->setMetadata(msgSendMDKind, node);
if (!isPointerSizedReturn) {
messageBB = CGF.Builder.GetInsertBlock();
CGF.Builder.CreateBr(continueBB);
CGF.EmitBlock(continueBB);
if (msgRet.isScalar()) {
llvm::Value *v = msgRet.getScalarVal();
llvm::PHINode *phi = Builder.CreatePHI(v->getType(), 2);
phi->addIncoming(v, messageBB);
phi->addIncoming(llvm::Constant::getNullValue(v->getType()), startBB);
msgRet = RValue::get(phi);
} else if (msgRet.isAggregate()) {
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address v = msgRet.getAggregateAddress();
llvm::PHINode *phi = Builder.CreatePHI(v.getType(), 2);
llvm::Type *RetTy = v.getElementType();
Address NullVal = CGF.CreateTempAlloca(RetTy, v.getAlignment(), "null");
CGF.InitTempAlloca(NullVal, llvm::Constant::getNullValue(RetTy));
phi->addIncoming(v.getPointer(), messageBB);
phi->addIncoming(NullVal.getPointer(), startBB);
msgRet = RValue::getAggregate(Address(phi, v.getAlignment()));
} else /* isComplex() */ {
std::pair<llvm::Value*,llvm::Value*> v = msgRet.getComplexVal();
llvm::PHINode *phi = Builder.CreatePHI(v.first->getType(), 2);
phi->addIncoming(v.first, messageBB);
phi->addIncoming(llvm::Constant::getNullValue(v.first->getType()),
startBB);
llvm::PHINode *phi2 = Builder.CreatePHI(v.second->getType(), 2);
phi2->addIncoming(v.second, messageBB);
phi2->addIncoming(llvm::Constant::getNullValue(v.second->getType()),
startBB);
msgRet = RValue::getComplex(phi, phi2);
}
}
return msgRet;
}
/// Generates a MethodList. Used in construction of a objc_class and
/// objc_category structures.
llvm::Constant *CGObjCGNU::
GenerateMethodList(StringRef ClassName,
StringRef CategoryName,
ArrayRef<const ObjCMethodDecl*> Methods,
bool isClassMethodList) {
if (Methods.empty())
return NULLPtr;
ConstantInitBuilder Builder(CGM);
auto MethodList = Builder.beginStruct();
MethodList.addNullPointer(CGM.Int8PtrTy);
MethodList.addInt(Int32Ty, Methods.size());
// Get the method structure type.
llvm::StructType *ObjCMethodTy =
llvm::StructType::get(CGM.getLLVMContext(), {
PtrToInt8Ty, // Really a selector, but the runtime creates it us.
PtrToInt8Ty, // Method types
IMPTy // Method pointer
});
bool isV2ABI = isRuntime(ObjCRuntime::GNUstep, 2);
if (isV2ABI) {
// size_t size;
llvm::DataLayout td(&TheModule);
MethodList.addInt(SizeTy, td.getTypeSizeInBits(ObjCMethodTy) /
CGM.getContext().getCharWidth());
ObjCMethodTy =
llvm::StructType::get(CGM.getLLVMContext(), {
IMPTy, // Method pointer
PtrToInt8Ty, // Selector
PtrToInt8Ty // Extended type encoding
});
} else {
ObjCMethodTy =
llvm::StructType::get(CGM.getLLVMContext(), {
PtrToInt8Ty, // Really a selector, but the runtime creates it us.
PtrToInt8Ty, // Method types
IMPTy // Method pointer
});
}
auto MethodArray = MethodList.beginArray();
ASTContext &Context = CGM.getContext();
for (const auto *OMD : Methods) {
llvm::Constant *FnPtr =
TheModule.getFunction(SymbolNameForMethod(ClassName, CategoryName,
OMD->getSelector(),
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
isClassMethodList));
assert(FnPtr && "Can't generate metadata for method that doesn't exist");
auto Method = MethodArray.beginStruct(ObjCMethodTy);
if (isV2ABI) {
Method.addBitCast(FnPtr, IMPTy);
Method.add(GetConstantSelector(OMD->getSelector(),
Context.getObjCEncodingForMethodDecl(OMD)));
Method.add(MakeConstantString(Context.getObjCEncodingForMethodDecl(OMD, true)));
} else {
Method.add(MakeConstantString(OMD->getSelector().getAsString()));
Method.add(MakeConstantString(Context.getObjCEncodingForMethodDecl(OMD)));
Method.addBitCast(FnPtr, IMPTy);
}
Method.finishAndAddTo(MethodArray);
}
MethodArray.finishAndAddTo(MethodList);
// Create an instance of the structure
return MethodList.finishAndCreateGlobal(".objc_method_list",
CGM.getPointerAlign());
}
/// Generates an IvarList. Used in construction of a objc_class.
llvm::Constant *CGObjCGNU::
GenerateIvarList(ArrayRef<llvm::Constant *> IvarNames,
ArrayRef<llvm::Constant *> IvarTypes,
ArrayRef<llvm::Constant *> IvarOffsets,
ArrayRef<llvm::Constant *> IvarAlign,
ArrayRef<Qualifiers::ObjCLifetime> IvarOwnership) {
if (IvarNames.empty())
return NULLPtr;
ConstantInitBuilder Builder(CGM);
// Structure containing array count followed by array.
auto IvarList = Builder.beginStruct();
IvarList.addInt(IntTy, (int)IvarNames.size());
// Get the ivar structure type.
llvm::StructType *ObjCIvarTy =
llvm::StructType::get(PtrToInt8Ty, PtrToInt8Ty, IntTy);
// Array of ivar structures.
auto Ivars = IvarList.beginArray(ObjCIvarTy);
for (unsigned int i = 0, e = IvarNames.size() ; i < e ; i++) {
auto Ivar = Ivars.beginStruct(ObjCIvarTy);
Ivar.add(IvarNames[i]);
Ivar.add(IvarTypes[i]);
Ivar.add(IvarOffsets[i]);
Ivar.finishAndAddTo(Ivars);
}
Ivars.finishAndAddTo(IvarList);
// Create an instance of the structure
return IvarList.finishAndCreateGlobal(".objc_ivar_list",
CGM.getPointerAlign());
}
/// Generate a class structure
llvm::Constant *CGObjCGNU::GenerateClassStructure(
llvm::Constant *MetaClass,
llvm::Constant *SuperClass,
unsigned info,
const char *Name,
llvm::Constant *Version,
llvm::Constant *InstanceSize,
llvm::Constant *IVars,
llvm::Constant *Methods,
llvm::Constant *Protocols,
llvm::Constant *IvarOffsets,
llvm::Constant *Properties,
llvm::Constant *StrongIvarBitmap,
llvm::Constant *WeakIvarBitmap,
bool isMeta) {
// Set up the class structure
// Note: Several of these are char*s when they should be ids. This is
// because the runtime performs this translation on load.
//
// Fields marked New ABI are part of the GNUstep runtime. We emit them
// anyway; the classes will still work with the GNU runtime, they will just
// be ignored.
llvm::StructType *ClassTy = llvm::StructType::get(
PtrToInt8Ty, // isa
PtrToInt8Ty, // super_class
PtrToInt8Ty, // name
LongTy, // version
LongTy, // info
LongTy, // instance_size
IVars->getType(), // ivars
Methods->getType(), // methods
// These are all filled in by the runtime, so we pretend
PtrTy, // dtable
PtrTy, // subclass_list
PtrTy, // sibling_class
PtrTy, // protocols
PtrTy, // gc_object_type
// New ABI:
LongTy, // abi_version
IvarOffsets->getType(), // ivar_offsets
Properties->getType(), // properties
IntPtrTy, // strong_pointers
IntPtrTy // weak_pointers
);
ConstantInitBuilder Builder(CGM);
auto Elements = Builder.beginStruct(ClassTy);
// Fill in the structure
// isa
Elements.addBitCast(MetaClass, PtrToInt8Ty);
// super_class
Elements.add(SuperClass);
// name
Elements.add(MakeConstantString(Name, ".class_name"));
// version
Elements.addInt(LongTy, 0);
// info
Elements.addInt(LongTy, info);
// instance_size
if (isMeta) {
llvm::DataLayout td(&TheModule);
Elements.addInt(LongTy,
td.getTypeSizeInBits(ClassTy) /
CGM.getContext().getCharWidth());
} else
Elements.add(InstanceSize);
// ivars
Elements.add(IVars);
// methods
Elements.add(Methods);
// These are all filled in by the runtime, so we pretend
// dtable
Elements.add(NULLPtr);
// subclass_list
Elements.add(NULLPtr);
// sibling_class
Elements.add(NULLPtr);
// protocols
Elements.addBitCast(Protocols, PtrTy);
// gc_object_type
Elements.add(NULLPtr);
// abi_version
Elements.addInt(LongTy, ClassABIVersion);
// ivar_offsets
Elements.add(IvarOffsets);
// properties
Elements.add(Properties);
// strong_pointers
Elements.add(StrongIvarBitmap);
// weak_pointers
Elements.add(WeakIvarBitmap);
// Create an instance of the structure
// This is now an externally visible symbol, so that we can speed up class
// messages in the next ABI. We may already have some weak references to
// this, so check and fix them properly.
std::string ClassSym((isMeta ? "_OBJC_METACLASS_": "_OBJC_CLASS_") +
std::string(Name));
llvm::GlobalVariable *ClassRef = TheModule.getNamedGlobal(ClassSym);
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Constant *Class =
Elements.finishAndCreateGlobal(ClassSym, CGM.getPointerAlign(), false,
llvm::GlobalValue::ExternalLinkage);
if (ClassRef) {
ClassRef->replaceAllUsesWith(llvm::ConstantExpr::getBitCast(Class,
ClassRef->getType()));
ClassRef->removeFromParent();
Class->setName(ClassSym);
}
return Class;
}
llvm::Constant *CGObjCGNU::
GenerateProtocolMethodList(ArrayRef<const ObjCMethodDecl*> Methods) {
// Get the method structure type.
llvm::StructType *ObjCMethodDescTy =
llvm::StructType::get(CGM.getLLVMContext(), { PtrToInt8Ty, PtrToInt8Ty });
ASTContext &Context = CGM.getContext();
ConstantInitBuilder Builder(CGM);
auto MethodList = Builder.beginStruct();
MethodList.addInt(IntTy, Methods.size());
auto MethodArray = MethodList.beginArray(ObjCMethodDescTy);
for (auto *M : Methods) {
auto Method = MethodArray.beginStruct(ObjCMethodDescTy);
Method.add(MakeConstantString(M->getSelector().getAsString()));
Method.add(MakeConstantString(Context.getObjCEncodingForMethodDecl(M)));
Method.finishAndAddTo(MethodArray);
}
MethodArray.finishAndAddTo(MethodList);
return MethodList.finishAndCreateGlobal(".objc_method_list",
CGM.getPointerAlign());
}
// Create the protocol list structure used in classes, categories and so on
llvm::Constant *
CGObjCGNU::GenerateProtocolList(ArrayRef<std::string> Protocols) {
ConstantInitBuilder Builder(CGM);
auto ProtocolList = Builder.beginStruct();
ProtocolList.add(NULLPtr);
ProtocolList.addInt(LongTy, Protocols.size());
auto Elements = ProtocolList.beginArray(PtrToInt8Ty);
for (const std::string *iter = Protocols.begin(), *endIter = Protocols.end();
iter != endIter ; iter++) {
llvm::Constant *protocol = nullptr;
llvm::StringMap<llvm::Constant*>::iterator value =
ExistingProtocols.find(*iter);
if (value == ExistingProtocols.end()) {
protocol = GenerateEmptyProtocol(*iter);
} else {
protocol = value->getValue();
}
Elements.addBitCast(protocol, PtrToInt8Ty);
}
Elements.finishAndAddTo(ProtocolList);
return ProtocolList.finishAndCreateGlobal(".objc_protocol_list",
CGM.getPointerAlign());
}
llvm::Value *CGObjCGNU::GenerateProtocolRef(CodeGenFunction &CGF,
const ObjCProtocolDecl *PD) {
llvm::Constant *&protocol = ExistingProtocols[PD->getNameAsString()];
if (!protocol)
GenerateProtocol(PD);
llvm::Type *T =
CGM.getTypes().ConvertType(CGM.getContext().getObjCProtoType());
return CGF.Builder.CreateBitCast(protocol, llvm::PointerType::getUnqual(T));
}
llvm::Constant *
CGObjCGNU::GenerateEmptyProtocol(StringRef ProtocolName) {
llvm::Constant *ProtocolList = GenerateProtocolList({});
llvm::Constant *MethodList = GenerateProtocolMethodList({});
MethodList = llvm::ConstantExpr::getBitCast(MethodList, PtrToInt8Ty);
// Protocols are objects containing lists of the methods implemented and
// protocols adopted.
ConstantInitBuilder Builder(CGM);
auto Elements = Builder.beginStruct();
// The isa pointer must be set to a magic number so the runtime knows it's
// the correct layout.
Elements.add(llvm::ConstantExpr::getIntToPtr(
llvm::ConstantInt::get(Int32Ty, ProtocolVersion), IdTy));
Elements.add(MakeConstantString(ProtocolName, ".objc_protocol_name"));
Elements.add(ProtocolList); /* .protocol_list */
Elements.add(MethodList); /* .instance_methods */
Elements.add(MethodList); /* .class_methods */
Elements.add(MethodList); /* .optional_instance_methods */
Elements.add(MethodList); /* .optional_class_methods */
Elements.add(NULLPtr); /* .properties */
Elements.add(NULLPtr); /* .optional_properties */
return Elements.finishAndCreateGlobal(SymbolForProtocol(ProtocolName),
CGM.getPointerAlign());
}
void CGObjCGNU::GenerateProtocol(const ObjCProtocolDecl *PD) {
std::string ProtocolName = PD->getNameAsString();
// Use the protocol definition, if there is one.
if (const ObjCProtocolDecl *Def = PD->getDefinition())
PD = Def;
SmallVector<std::string, 16> Protocols;
for (const auto *PI : PD->protocols())
Protocols.push_back(PI->getNameAsString());
SmallVector<const ObjCMethodDecl*, 16> InstanceMethods;
SmallVector<const ObjCMethodDecl*, 16> OptionalInstanceMethods;
for (const auto *I : PD->instance_methods())
if (I->isOptional())
OptionalInstanceMethods.push_back(I);
else
InstanceMethods.push_back(I);
// Collect information about class methods:
SmallVector<const ObjCMethodDecl*, 16> ClassMethods;
SmallVector<const ObjCMethodDecl*, 16> OptionalClassMethods;
for (const auto *I : PD->class_methods())
if (I->isOptional())
OptionalClassMethods.push_back(I);
else
ClassMethods.push_back(I);
llvm::Constant *ProtocolList = GenerateProtocolList(Protocols);
llvm::Constant *InstanceMethodList =
GenerateProtocolMethodList(InstanceMethods);
llvm::Constant *ClassMethodList =
GenerateProtocolMethodList(ClassMethods);
llvm::Constant *OptionalInstanceMethodList =
GenerateProtocolMethodList(OptionalInstanceMethods);
llvm::Constant *OptionalClassMethodList =
GenerateProtocolMethodList(OptionalClassMethods);
// Property metadata: name, attributes, isSynthesized, setter name, setter
// types, getter name, getter types.
// The isSynthesized value is always set to 0 in a protocol. It exists to
// simplify the runtime library by allowing it to use the same data
// structures for protocol metadata everywhere.
llvm::Constant *PropertyList =
GeneratePropertyList(nullptr, PD, false, false);
llvm::Constant *OptionalPropertyList =
GeneratePropertyList(nullptr, PD, false, true);
// Protocols are objects containing lists of the methods implemented and
// protocols adopted.
// The isa pointer must be set to a magic number so the runtime knows it's
// the correct layout.
ConstantInitBuilder Builder(CGM);
auto Elements = Builder.beginStruct();
Elements.add(
llvm::ConstantExpr::getIntToPtr(
llvm::ConstantInt::get(Int32Ty, ProtocolVersion), IdTy));
Elements.add(MakeConstantString(ProtocolName));
Elements.add(ProtocolList);
Elements.add(InstanceMethodList);
Elements.add(ClassMethodList);
Elements.add(OptionalInstanceMethodList);
Elements.add(OptionalClassMethodList);
Elements.add(PropertyList);
Elements.add(OptionalPropertyList);
ExistingProtocols[ProtocolName] =
llvm::ConstantExpr::getBitCast(
Elements.finishAndCreateGlobal(".objc_protocol", CGM.getPointerAlign()),
IdTy);
}
void CGObjCGNU::GenerateProtocolHolderCategory() {
// Collect information about instance methods
ConstantInitBuilder Builder(CGM);
auto Elements = Builder.beginStruct();
const std::string ClassName = "__ObjC_Protocol_Holder_Ugly_Hack";
const std::string CategoryName = "AnotherHack";
Elements.add(MakeConstantString(CategoryName));
Elements.add(MakeConstantString(ClassName));
// Instance method list
Elements.addBitCast(GenerateMethodList(
ClassName, CategoryName, {}, false), PtrTy);
// Class method list
Elements.addBitCast(GenerateMethodList(
ClassName, CategoryName, {}, true), PtrTy);
// Protocol list
ConstantInitBuilder ProtocolListBuilder(CGM);
auto ProtocolList = ProtocolListBuilder.beginStruct();
ProtocolList.add(NULLPtr);
ProtocolList.addInt(LongTy, ExistingProtocols.size());
auto ProtocolElements = ProtocolList.beginArray(PtrTy);
for (auto iter = ExistingProtocols.begin(), endIter = ExistingProtocols.end();
iter != endIter ; iter++) {
ProtocolElements.addBitCast(iter->getValue(), PtrTy);
}
ProtocolElements.finishAndAddTo(ProtocolList);
Elements.addBitCast(
ProtocolList.finishAndCreateGlobal(".objc_protocol_list",
CGM.getPointerAlign()),
PtrTy);
Categories.push_back(llvm::ConstantExpr::getBitCast(
Elements.finishAndCreateGlobal("", CGM.getPointerAlign()),
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
PtrTy));
}
/// Libobjc2 uses a bitfield representation where small(ish) bitfields are
/// stored in a 64-bit value with the low bit set to 1 and the remaining 63
/// bits set to their values, LSB first, while larger ones are stored in a
/// structure of this / form:
///
/// struct { int32_t length; int32_t values[length]; };
///
/// The values in the array are stored in host-endian format, with the least
/// significant bit being assumed to come first in the bitfield. Therefore, a
/// bitfield with the 64th bit set will be (int64_t)&{ 2, [0, 1<<31] }, while a
/// bitfield / with the 63rd bit set will be 1<<64.
llvm::Constant *CGObjCGNU::MakeBitField(ArrayRef<bool> bits) {
int bitCount = bits.size();
int ptrBits = CGM.getDataLayout().getPointerSizeInBits();
if (bitCount < ptrBits) {
uint64_t val = 1;
for (int i=0 ; i<bitCount ; ++i) {
if (bits[i]) val |= 1ULL<<(i+1);
}
return llvm::ConstantInt::get(IntPtrTy, val);
}
SmallVector<llvm::Constant *, 8> values;
int v=0;
while (v < bitCount) {
int32_t word = 0;
for (int i=0 ; (i<32) && (v<bitCount) ; ++i) {
if (bits[v]) word |= 1<<i;
v++;
}
values.push_back(llvm::ConstantInt::get(Int32Ty, word));
}
ConstantInitBuilder builder(CGM);
auto fields = builder.beginStruct();
fields.addInt(Int32Ty, values.size());
auto array = fields.beginArray();
for (auto v : values) array.add(v);
array.finishAndAddTo(fields);
llvm::Constant *GS =
fields.finishAndCreateGlobal("", CharUnits::fromQuantity(4));
llvm::Constant *ptr = llvm::ConstantExpr::getPtrToInt(GS, IntPtrTy);
return ptr;
}
llvm::Constant *CGObjCGNU::GenerateCategoryProtocolList(const
ObjCCategoryDecl *OCD) {
SmallVector<std::string, 16> Protocols;
for (const auto *PD : OCD->getReferencedProtocols())
Protocols.push_back(PD->getNameAsString());
return GenerateProtocolList(Protocols);
}
void CGObjCGNU::GenerateCategory(const ObjCCategoryImplDecl *OCD) {
const ObjCInterfaceDecl *Class = OCD->getClassInterface();
std::string ClassName = Class->getNameAsString();
std::string CategoryName = OCD->getNameAsString();
// Collect the names of referenced protocols
const ObjCCategoryDecl *CatDecl = OCD->getCategoryDecl();
ConstantInitBuilder Builder(CGM);
auto Elements = Builder.beginStruct();
Elements.add(MakeConstantString(CategoryName));
Elements.add(MakeConstantString(ClassName));
// Instance method list
SmallVector<ObjCMethodDecl*, 16> InstanceMethods;
InstanceMethods.insert(InstanceMethods.begin(), OCD->instmeth_begin(),
OCD->instmeth_end());
Elements.addBitCast(
GenerateMethodList(ClassName, CategoryName, InstanceMethods, false),
PtrTy);
// Class method list
SmallVector<ObjCMethodDecl*, 16> ClassMethods;
ClassMethods.insert(ClassMethods.begin(), OCD->classmeth_begin(),
OCD->classmeth_end());
Elements.addBitCast(
GenerateMethodList(ClassName, CategoryName, ClassMethods, true),
PtrTy);
// Protocol list
Elements.addBitCast(GenerateCategoryProtocolList(CatDecl), PtrTy);
if (isRuntime(ObjCRuntime::GNUstep, 2)) {
const ObjCCategoryDecl *Category =
Class->FindCategoryDeclaration(OCD->getIdentifier());
if (Category) {
// Instance properties
Elements.addBitCast(GeneratePropertyList(OCD, Category, false), PtrTy);
// Class properties
Elements.addBitCast(GeneratePropertyList(OCD, Category, true), PtrTy);
} else {
Elements.addNullPointer(PtrTy);
Elements.addNullPointer(PtrTy);
}
}
Categories.push_back(llvm::ConstantExpr::getBitCast(
Elements.finishAndCreateGlobal(
std::string(".objc_category_")+ClassName+CategoryName,
CGM.getPointerAlign()),
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
PtrTy));
}
llvm::Constant *CGObjCGNU::GeneratePropertyList(const Decl *Container,
const ObjCContainerDecl *OCD,
bool isClassProperty,
bool protocolOptionalProperties) {
SmallVector<const ObjCPropertyDecl *, 16> Properties;
llvm::SmallPtrSet<const IdentifierInfo*, 16> PropertySet;
bool isProtocol = isa<ObjCProtocolDecl>(OCD);
ASTContext &Context = CGM.getContext();
std::function<void(const ObjCProtocolDecl *Proto)> collectProtocolProperties
= [&](const ObjCProtocolDecl *Proto) {
for (const auto *P : Proto->protocols())
collectProtocolProperties(P);
for (const auto *PD : Proto->properties()) {
if (isClassProperty != PD->isClassProperty())
continue;
// Skip any properties that are declared in protocols that this class
// conforms to but are not actually implemented by this class.
if (!isProtocol && !Context.getObjCPropertyImplDeclForPropertyDecl(PD, Container))
continue;
if (!PropertySet.insert(PD->getIdentifier()).second)
continue;
Properties.push_back(PD);
}
};
if (const ObjCInterfaceDecl *OID = dyn_cast<ObjCInterfaceDecl>(OCD))
for (const ObjCCategoryDecl *ClassExt : OID->known_extensions())
for (auto *PD : ClassExt->properties()) {
if (isClassProperty != PD->isClassProperty())
continue;
PropertySet.insert(PD->getIdentifier());
Properties.push_back(PD);
}
for (const auto *PD : OCD->properties()) {
if (isClassProperty != PD->isClassProperty())
continue;
// If we're generating a list for a protocol, skip optional / required ones
// when generating the other list.
if (isProtocol && (protocolOptionalProperties != PD->isOptional()))
continue;
// Don't emit duplicate metadata for properties that were already in a
// class extension.
if (!PropertySet.insert(PD->getIdentifier()).second)
continue;
Properties.push_back(PD);
}
if (const ObjCInterfaceDecl *OID = dyn_cast<ObjCInterfaceDecl>(OCD))
for (const auto *P : OID->all_referenced_protocols())
collectProtocolProperties(P);
else if (const ObjCCategoryDecl *CD = dyn_cast<ObjCCategoryDecl>(OCD))
for (const auto *P : CD->protocols())
collectProtocolProperties(P);
auto numProperties = Properties.size();
if (numProperties == 0)
return NULLPtr;
ConstantInitBuilder builder(CGM);
auto propertyList = builder.beginStruct();
auto properties = PushPropertyListHeader(propertyList, numProperties);
// Add all of the property methods need adding to the method list and to the
// property metadata list.
for (auto *property : Properties) {
bool isSynthesized = false;
bool isDynamic = false;
if (!isProtocol) {
auto *propertyImpl = Context.getObjCPropertyImplDeclForPropertyDecl(property, Container);
if (propertyImpl) {
isSynthesized = (propertyImpl->getPropertyImplementation() ==
ObjCPropertyImplDecl::Synthesize);
isDynamic = (propertyImpl->getPropertyImplementation() ==
ObjCPropertyImplDecl::Dynamic);
}
}
PushProperty(properties, property, Container, isSynthesized, isDynamic);
}
properties.finishAndAddTo(propertyList);
return propertyList.finishAndCreateGlobal(".objc_property_list",
CGM.getPointerAlign());
}
void CGObjCGNU::RegisterAlias(const ObjCCompatibleAliasDecl *OAD) {
// Get the class declaration for which the alias is specified.
ObjCInterfaceDecl *ClassDecl =
const_cast<ObjCInterfaceDecl *>(OAD->getClassInterface());
ClassAliases.emplace_back(ClassDecl->getNameAsString(),
OAD->getNameAsString());
}
void CGObjCGNU::GenerateClass(const ObjCImplementationDecl *OID) {
ASTContext &Context = CGM.getContext();
// Get the superclass name.
const ObjCInterfaceDecl * SuperClassDecl =
OID->getClassInterface()->getSuperClass();
std::string SuperClassName;
if (SuperClassDecl) {
SuperClassName = SuperClassDecl->getNameAsString();
EmitClassRef(SuperClassName);
}
// Get the class name
2009-04-01 10:00:48 +08:00
ObjCInterfaceDecl *ClassDecl =
const_cast<ObjCInterfaceDecl *>(OID->getClassInterface());
std::string ClassName = ClassDecl->getNameAsString();
// Emit the symbol that is used to generate linker errors if this class is
// referenced in other modules but not declared.
std::string classSymbolName = "__objc_class_name_" + ClassName;
if (auto *symbol = TheModule.getGlobalVariable(classSymbolName)) {
symbol->setInitializer(llvm::ConstantInt::get(LongTy, 0));
} else {
new llvm::GlobalVariable(TheModule, LongTy, false,
llvm::GlobalValue::ExternalLinkage,
llvm::ConstantInt::get(LongTy, 0),
classSymbolName);
}
// Get the size of instances.
int instanceSize =
Context.getASTObjCImplementationLayout(OID).getSize().getQuantity();
// Collect information about instance variables.
SmallVector<llvm::Constant*, 16> IvarNames;
SmallVector<llvm::Constant*, 16> IvarTypes;
SmallVector<llvm::Constant*, 16> IvarOffsets;
SmallVector<llvm::Constant*, 16> IvarAligns;
SmallVector<Qualifiers::ObjCLifetime, 16> IvarOwnership;
ConstantInitBuilder IvarOffsetBuilder(CGM);
auto IvarOffsetValues = IvarOffsetBuilder.beginArray(PtrToIntTy);
SmallVector<bool, 16> WeakIvars;
SmallVector<bool, 16> StrongIvars;
int superInstanceSize = !SuperClassDecl ? 0 :
Context.getASTObjCInterfaceLayout(SuperClassDecl).getSize().getQuantity();
// For non-fragile ivars, set the instance size to 0 - {the size of just this
// class}. The runtime will then set this to the correct value on load.
if (CGM.getLangOpts().ObjCRuntime.isNonFragile()) {
instanceSize = 0 - (instanceSize - superInstanceSize);
}
for (const ObjCIvarDecl *IVD = ClassDecl->all_declared_ivar_begin(); IVD;
IVD = IVD->getNextIvar()) {
// Store the name
IvarNames.push_back(MakeConstantString(IVD->getNameAsString()));
// Get the type encoding for this ivar
std::string TypeStr;
Context.getObjCEncodingForType(IVD->getType(), TypeStr, IVD);
IvarTypes.push_back(MakeConstantString(TypeStr));
IvarAligns.push_back(llvm::ConstantInt::get(IntTy,
Context.getTypeSize(IVD->getType())));
// Get the offset
uint64_t BaseOffset = ComputeIvarBaseOffset(CGM, OID, IVD);
uint64_t Offset = BaseOffset;
if (CGM.getLangOpts().ObjCRuntime.isNonFragile()) {
Offset = BaseOffset - superInstanceSize;
}
llvm::Constant *OffsetValue = llvm::ConstantInt::get(IntTy, Offset);
// Create the direct offset value
std::string OffsetName = "__objc_ivar_offset_value_" + ClassName +"." +
IVD->getNameAsString();
llvm::GlobalVariable *OffsetVar = TheModule.getGlobalVariable(OffsetName);
if (OffsetVar) {
OffsetVar->setInitializer(OffsetValue);
// If this is the real definition, change its linkage type so that
// different modules will use this one, rather than their private
// copy.
OffsetVar->setLinkage(llvm::GlobalValue::ExternalLinkage);
} else
OffsetVar = new llvm::GlobalVariable(TheModule, Int32Ty,
false, llvm::GlobalValue::ExternalLinkage,
OffsetValue, OffsetName);
IvarOffsets.push_back(OffsetValue);
IvarOffsetValues.add(OffsetVar);
Qualifiers::ObjCLifetime lt = IVD->getType().getQualifiers().getObjCLifetime();
IvarOwnership.push_back(lt);
switch (lt) {
case Qualifiers::OCL_Strong:
StrongIvars.push_back(true);
WeakIvars.push_back(false);
break;
case Qualifiers::OCL_Weak:
StrongIvars.push_back(false);
WeakIvars.push_back(true);
break;
default:
StrongIvars.push_back(false);
WeakIvars.push_back(false);
}
}
llvm::Constant *StrongIvarBitmap = MakeBitField(StrongIvars);
llvm::Constant *WeakIvarBitmap = MakeBitField(WeakIvars);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
llvm::GlobalVariable *IvarOffsetArray =
IvarOffsetValues.finishAndCreateGlobal(".ivar.offsets",
CGM.getPointerAlign());
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// Collect information about instance methods
SmallVector<const ObjCMethodDecl*, 16> InstanceMethods;
InstanceMethods.insert(InstanceMethods.begin(), OID->instmeth_begin(),
OID->instmeth_end());
SmallVector<const ObjCMethodDecl*, 16> ClassMethods;
ClassMethods.insert(ClassMethods.begin(), OID->classmeth_begin(),
OID->classmeth_end());
// Collect the same information about synthesized properties, which don't
// show up in the instance method lists.
for (auto *propertyImpl : OID->property_impls())
if (propertyImpl->getPropertyImplementation() ==
ObjCPropertyImplDecl::Synthesize) {
ObjCPropertyDecl *property = propertyImpl->getPropertyDecl();
auto addPropertyMethod = [&](const ObjCMethodDecl *accessor) {
if (accessor)
InstanceMethods.push_back(accessor);
};
addPropertyMethod(property->getGetterMethodDecl());
addPropertyMethod(property->getSetterMethodDecl());
}
llvm::Constant *Properties = GeneratePropertyList(OID, ClassDecl);
// Collect the names of referenced protocols
SmallVector<std::string, 16> Protocols;
for (const auto *I : ClassDecl->protocols())
Protocols.push_back(I->getNameAsString());
// Get the superclass pointer.
llvm::Constant *SuperClass;
if (!SuperClassName.empty()) {
SuperClass = MakeConstantString(SuperClassName, ".super_class_name");
} else {
SuperClass = llvm::ConstantPointerNull::get(PtrToInt8Ty);
}
// Empty vector used to construct empty method lists
SmallVector<llvm::Constant*, 1> empty;
// Generate the method and instance variable lists
llvm::Constant *MethodList = GenerateMethodList(ClassName, "",
InstanceMethods, false);
llvm::Constant *ClassMethodList = GenerateMethodList(ClassName, "",
ClassMethods, true);
llvm::Constant *IvarList = GenerateIvarList(IvarNames, IvarTypes,
IvarOffsets, IvarAligns, IvarOwnership);
// Irrespective of whether we are compiling for a fragile or non-fragile ABI,
// we emit a symbol containing the offset for each ivar in the class. This
// allows code compiled for the non-Fragile ABI to inherit from code compiled
// for the legacy ABI, without causing problems. The converse is also
// possible, but causes all ivar accesses to be fragile.
// Offset pointer for getting at the correct field in the ivar list when
// setting up the alias. These are: The base address for the global, the
// ivar array (second field), the ivar in this list (set for each ivar), and
// the offset (third field in ivar structure)
llvm::Type *IndexTy = Int32Ty;
llvm::Constant *offsetPointerIndexes[] = {Zeros[0],
llvm::ConstantInt::get(IndexTy, ClassABIVersion > 1 ? 2 : 1), nullptr,
llvm::ConstantInt::get(IndexTy, ClassABIVersion > 1 ? 3 : 2) };
unsigned ivarIndex = 0;
for (const ObjCIvarDecl *IVD = ClassDecl->all_declared_ivar_begin(); IVD;
IVD = IVD->getNextIvar()) {
const std::string Name = GetIVarOffsetVariableName(ClassDecl, IVD);
offsetPointerIndexes[2] = llvm::ConstantInt::get(IndexTy, ivarIndex);
// Get the correct ivar field
llvm::Constant *offsetValue = llvm::ConstantExpr::getGetElementPtr(
cast<llvm::GlobalVariable>(IvarList)->getValueType(), IvarList,
offsetPointerIndexes);
// Get the existing variable, if one exists.
llvm::GlobalVariable *offset = TheModule.getNamedGlobal(Name);
if (offset) {
offset->setInitializer(offsetValue);
// If this is the real definition, change its linkage type so that
// different modules will use this one, rather than their private
// copy.
offset->setLinkage(llvm::GlobalValue::ExternalLinkage);
} else
// Add a new alias if there isn't one already.
new llvm::GlobalVariable(TheModule, offsetValue->getType(),
false, llvm::GlobalValue::ExternalLinkage, offsetValue, Name);
++ivarIndex;
}
llvm::Constant *ZeroPtr = llvm::ConstantInt::get(IntPtrTy, 0);
//Generate metaclass for class methods
llvm::Constant *MetaClassStruct = GenerateClassStructure(
NULLPtr, NULLPtr, 0x12L, ClassName.c_str(), nullptr, Zeros[0],
NULLPtr, ClassMethodList, NULLPtr, NULLPtr,
GeneratePropertyList(OID, ClassDecl, true), ZeroPtr, ZeroPtr, true);
CGM.setGVProperties(cast<llvm::GlobalValue>(MetaClassStruct),
OID->getClassInterface());
// Generate the class structure
llvm::Constant *ClassStruct = GenerateClassStructure(
MetaClassStruct, SuperClass, 0x11L, ClassName.c_str(), nullptr,
llvm::ConstantInt::get(LongTy, instanceSize), IvarList, MethodList,
GenerateProtocolList(Protocols), IvarOffsetArray, Properties,
StrongIvarBitmap, WeakIvarBitmap);
CGM.setGVProperties(cast<llvm::GlobalValue>(ClassStruct),
OID->getClassInterface());
// Resolve the class aliases, if they exist.
if (ClassPtrAlias) {
ClassPtrAlias->replaceAllUsesWith(
llvm::ConstantExpr::getBitCast(ClassStruct, IdTy));
ClassPtrAlias->eraseFromParent();
ClassPtrAlias = nullptr;
}
if (MetaClassPtrAlias) {
MetaClassPtrAlias->replaceAllUsesWith(
llvm::ConstantExpr::getBitCast(MetaClassStruct, IdTy));
MetaClassPtrAlias->eraseFromParent();
MetaClassPtrAlias = nullptr;
}
// Add class structure to list to be added to the symtab later
ClassStruct = llvm::ConstantExpr::getBitCast(ClassStruct, PtrToInt8Ty);
Classes.push_back(ClassStruct);
}
llvm::Function *CGObjCGNU::ModuleInitFunction() {
// Only emit an ObjC load function if no Objective-C stuff has been called
if (Classes.empty() && Categories.empty() && ConstantStrings.empty() &&
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
ExistingProtocols.empty() && SelectorTable.empty())
return nullptr;
// Add all referenced protocols to a category.
GenerateProtocolHolderCategory();
llvm::StructType *selStructTy =
dyn_cast<llvm::StructType>(SelectorTy->getElementType());
llvm::Type *selStructPtrTy = SelectorTy;
if (!selStructTy) {
selStructTy = llvm::StructType::get(CGM.getLLVMContext(),
{ PtrToInt8Ty, PtrToInt8Ty });
selStructPtrTy = llvm::PointerType::getUnqual(selStructTy);
}
// Generate statics list:
llvm::Constant *statics = NULLPtr;
if (!ConstantStrings.empty()) {
llvm::GlobalVariable *fileStatics = [&] {
ConstantInitBuilder builder(CGM);
auto staticsStruct = builder.beginStruct();
StringRef stringClass = CGM.getLangOpts().ObjCConstantStringClass;
if (stringClass.empty()) stringClass = "NXConstantString";
staticsStruct.add(MakeConstantString(stringClass,
".objc_static_class_name"));
auto array = staticsStruct.beginArray();
array.addAll(ConstantStrings);
array.add(NULLPtr);
array.finishAndAddTo(staticsStruct);
return staticsStruct.finishAndCreateGlobal(".objc_statics",
CGM.getPointerAlign());
}();
ConstantInitBuilder builder(CGM);
auto allStaticsArray = builder.beginArray(fileStatics->getType());
allStaticsArray.add(fileStatics);
allStaticsArray.addNullPointer(fileStatics->getType());
statics = allStaticsArray.finishAndCreateGlobal(".objc_statics_ptr",
CGM.getPointerAlign());
statics = llvm::ConstantExpr::getBitCast(statics, PtrTy);
}
// Array of classes, categories, and constant objects.
SmallVector<llvm::GlobalAlias*, 16> selectorAliases;
unsigned selectorCount;
// Pointer to an array of selectors used in this module.
llvm::GlobalVariable *selectorList = [&] {
ConstantInitBuilder builder(CGM);
auto selectors = builder.beginArray(selStructTy);
auto &table = SelectorTable; // MSVC workaround
std::vector<Selector> allSelectors;
for (auto &entry : table)
allSelectors.push_back(entry.first);
llvm::sort(allSelectors);
for (auto &untypedSel : allSelectors) {
std::string selNameStr = untypedSel.getAsString();
llvm::Constant *selName = ExportUniqueString(selNameStr, ".objc_sel_name");
for (TypedSelector &sel : table[untypedSel]) {
llvm::Constant *selectorTypeEncoding = NULLPtr;
if (!sel.first.empty())
selectorTypeEncoding =
MakeConstantString(sel.first, ".objc_sel_types");
auto selStruct = selectors.beginStruct(selStructTy);
selStruct.add(selName);
selStruct.add(selectorTypeEncoding);
selStruct.finishAndAddTo(selectors);
// Store the selector alias for later replacement
selectorAliases.push_back(sel.second);
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
}
// Remember the number of entries in the selector table.
selectorCount = selectors.size();
// NULL-terminate the selector list. This should not actually be required,
// because the selector list has a length field. Unfortunately, the GCC
// runtime decides to ignore the length field and expects a NULL terminator,
// and GCC cooperates with this by always setting the length to 0.
auto selStruct = selectors.beginStruct(selStructTy);
selStruct.add(NULLPtr);
selStruct.add(NULLPtr);
selStruct.finishAndAddTo(selectors);
return selectors.finishAndCreateGlobal(".objc_selector_list",
CGM.getPointerAlign());
}();
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// Now that all of the static selectors exist, create pointers to them.
for (unsigned i = 0; i < selectorCount; ++i) {
llvm::Constant *idxs[] = {
Zeros[0],
llvm::ConstantInt::get(Int32Ty, i)
};
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
// FIXME: We're generating redundant loads and stores here!
llvm::Constant *selPtr = llvm::ConstantExpr::getGetElementPtr(
selectorList->getValueType(), selectorList, idxs);
// If selectors are defined as an opaque type, cast the pointer to this
// type.
selPtr = llvm::ConstantExpr::getBitCast(selPtr, SelectorTy);
selectorAliases[i]->replaceAllUsesWith(selPtr);
selectorAliases[i]->eraseFromParent();
}
llvm::GlobalVariable *symtab = [&] {
ConstantInitBuilder builder(CGM);
auto symtab = builder.beginStruct();
// Number of static selectors
symtab.addInt(LongTy, selectorCount);
symtab.addBitCast(selectorList, selStructPtrTy);
// Number of classes defined.
symtab.addInt(CGM.Int16Ty, Classes.size());
// Number of categories defined
symtab.addInt(CGM.Int16Ty, Categories.size());
// Create an array of classes, then categories, then static object instances
auto classList = symtab.beginArray(PtrToInt8Ty);
classList.addAll(Classes);
classList.addAll(Categories);
// NULL-terminated list of static object instances (mainly constant strings)
classList.add(statics);
classList.add(NULLPtr);
classList.finishAndAddTo(symtab);
// Construct the symbol table.
return symtab.finishAndCreateGlobal("", CGM.getPointerAlign());
}();
// The symbol table is contained in a module which has some version-checking
// constants
llvm::Constant *module = [&] {
llvm::Type *moduleEltTys[] = {
LongTy, LongTy, PtrToInt8Ty, symtab->getType(), IntTy
};
llvm::StructType *moduleTy =
llvm::StructType::get(CGM.getLLVMContext(),
makeArrayRef(moduleEltTys).drop_back(unsigned(RuntimeVersion < 10)));
ConstantInitBuilder builder(CGM);
auto module = builder.beginStruct(moduleTy);
// Runtime version, used for ABI compatibility checking.
module.addInt(LongTy, RuntimeVersion);
// sizeof(ModuleTy)
module.addInt(LongTy, CGM.getDataLayout().getTypeStoreSize(moduleTy));
// The path to the source file where this module was declared
SourceManager &SM = CGM.getContext().getSourceManager();
const FileEntry *mainFile = SM.getFileEntryForID(SM.getMainFileID());
std::string path =
(Twine(mainFile->getDir()->getName()) + "/" + mainFile->getName()).str();
module.add(MakeConstantString(path, ".objc_source_file_name"));
module.add(symtab);
if (RuntimeVersion >= 10) {
switch (CGM.getLangOpts().getGC()) {
case LangOptions::GCOnly:
module.addInt(IntTy, 2);
break;
case LangOptions::NonGC:
if (CGM.getLangOpts().ObjCAutoRefCount)
module.addInt(IntTy, 1);
else
module.addInt(IntTy, 0);
break;
case LangOptions::HybridGC:
module.addInt(IntTy, 1);
break;
}
}
return module.finishAndCreateGlobal("", CGM.getPointerAlign());
}();
// Create the load function calling the runtime entry point with the module
// structure
llvm::Function * LoadFunction = llvm::Function::Create(
llvm::FunctionType::get(llvm::Type::getVoidTy(VMContext), false),
llvm::GlobalValue::InternalLinkage, ".objc_load_function",
&TheModule);
llvm::BasicBlock *EntryBB =
llvm::BasicBlock::Create(VMContext, "entry", LoadFunction);
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
CGBuilderTy Builder(CGM, VMContext);
Builder.SetInsertPoint(EntryBB);
llvm::FunctionType *FT =
llvm::FunctionType::get(Builder.getVoidTy(), module->getType(), true);
llvm::FunctionCallee Register =
CGM.CreateRuntimeFunction(FT, "__objc_exec_class");
Builder.CreateCall(Register, module);
if (!ClassAliases.empty()) {
llvm::Type *ArgTypes[2] = {PtrTy, PtrToInt8Ty};
llvm::FunctionType *RegisterAliasTy =
llvm::FunctionType::get(Builder.getVoidTy(),
ArgTypes, false);
llvm::Function *RegisterAlias = llvm::Function::Create(
RegisterAliasTy,
llvm::GlobalValue::ExternalWeakLinkage, "class_registerAlias_np",
&TheModule);
llvm::BasicBlock *AliasBB =
llvm::BasicBlock::Create(VMContext, "alias", LoadFunction);
llvm::BasicBlock *NoAliasBB =
llvm::BasicBlock::Create(VMContext, "no_alias", LoadFunction);
// Branch based on whether the runtime provided class_registerAlias_np()
llvm::Value *HasRegisterAlias = Builder.CreateICmpNE(RegisterAlias,
llvm::Constant::getNullValue(RegisterAlias->getType()));
Builder.CreateCondBr(HasRegisterAlias, AliasBB, NoAliasBB);
// The true branch (has alias registration function):
Builder.SetInsertPoint(AliasBB);
// Emit alias registration calls:
for (std::vector<ClassAliasPair>::iterator iter = ClassAliases.begin();
iter != ClassAliases.end(); ++iter) {
llvm::Constant *TheClass =
TheModule.getGlobalVariable("_OBJC_CLASS_" + iter->first, true);
if (TheClass) {
TheClass = llvm::ConstantExpr::getBitCast(TheClass, PtrTy);
Builder.CreateCall(RegisterAlias,
{TheClass, MakeConstantString(iter->second)});
}
}
// Jump to end:
Builder.CreateBr(NoAliasBB);
// Missing alias registration function, just return from the function:
Builder.SetInsertPoint(NoAliasBB);
}
Builder.CreateRetVoid();
return LoadFunction;
}
llvm::Function *CGObjCGNU::GenerateMethod(const ObjCMethodDecl *OMD,
const ObjCContainerDecl *CD) {
const ObjCCategoryImplDecl *OCD =
dyn_cast<ObjCCategoryImplDecl>(OMD->getDeclContext());
StringRef CategoryName = OCD ? OCD->getName() : "";
StringRef ClassName = CD->getName();
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
Selector MethodName = OMD->getSelector();
bool isClassMethod = !OMD->isInstanceMethod();
CodeGenTypes &Types = CGM.getTypes();
llvm::FunctionType *MethodTy =
Types.GetFunctionType(Types.arrangeObjCMethodDeclaration(OMD));
std::string FunctionName = SymbolNameForMethod(ClassName, CategoryName,
MethodName, isClassMethod);
llvm::Function *Method
= llvm::Function::Create(MethodTy,
llvm::GlobalValue::InternalLinkage,
FunctionName,
&TheModule);
return Method;
}
llvm::FunctionCallee CGObjCGNU::GetPropertyGetFunction() {
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
return GetPropertyFn;
}
llvm::FunctionCallee CGObjCGNU::GetPropertySetFunction() {
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
return SetPropertyFn;
}
llvm::FunctionCallee CGObjCGNU::GetOptimizedPropertySetFunction(bool atomic,
bool copy) {
return nullptr;
}
llvm::FunctionCallee CGObjCGNU::GetGetStructFunction() {
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
return GetStructPropertyFn;
}
llvm::FunctionCallee CGObjCGNU::GetSetStructFunction() {
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
return SetStructPropertyFn;
}
llvm::FunctionCallee CGObjCGNU::GetCppAtomicObjectGetFunction() {
return nullptr;
}
llvm::FunctionCallee CGObjCGNU::GetCppAtomicObjectSetFunction() {
return nullptr;
}
llvm::FunctionCallee CGObjCGNU::EnumerationMutationFunction() {
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
return EnumerationMutationFn;
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
void CGObjCGNU::EmitSynchronizedStmt(CodeGenFunction &CGF,
const ObjCAtSynchronizedStmt &S) {
EmitAtSynchronizedStmt(CGF, S, SyncEnterFn, SyncExitFn);
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
void CGObjCGNU::EmitTryStmt(CodeGenFunction &CGF,
const ObjCAtTryStmt &S) {
// Unlike the Apple non-fragile runtimes, which also uses
// unwind-based zero cost exceptions, the GNU Objective C runtime's
// EH support isn't a veneer over C++ EH. Instead, exception
// objects are created by objc_exception_throw and destroyed by
// the personality function; this avoids the need for bracketing
// catch handlers with calls to __blah_begin_catch/__blah_end_catch
// (or even _Unwind_DeleteException), but probably doesn't
// interoperate very well with foreign exceptions.
//
// In Objective-C++ mode, we actually emit something equivalent to the C++
// exception handler.
EmitTryCatchStmt(CGF, S, EnterCatchFn, ExitCatchFn, ExceptionReThrowFn);
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
void CGObjCGNU::EmitThrowStmt(CodeGenFunction &CGF,
const ObjCAtThrowStmt &S,
bool ClearInsertionPoint) {
llvm::Value *ExceptionAsObject;
bool isRethrow = false;
if (const Expr *ThrowExpr = S.getThrowExpr()) {
llvm::Value *Exception = CGF.EmitObjCThrowOperand(ThrowExpr);
ExceptionAsObject = Exception;
} else {
assert((!CGF.ObjCEHValueStack.empty() && CGF.ObjCEHValueStack.back()) &&
"Unexpected rethrow outside @catch block.");
ExceptionAsObject = CGF.ObjCEHValueStack.back();
isRethrow = true;
}
if (isRethrow && usesSEHExceptions) {
// For SEH, ExceptionAsObject may be undef, because the catch handler is
// not passed it for catchalls and so it is not visible to the catch
// funclet. The real thrown object will still be live on the stack at this
// point and will be rethrown. If we are explicitly rethrowing the object
// that was passed into the `@catch` block, then this code path is not
// reached and we will instead call `objc_exception_throw` with an explicit
// argument.
llvm::CallBase *Throw = CGF.EmitRuntimeCallOrInvoke(ExceptionReThrowFn);
Throw->setDoesNotReturn();
}
else {
ExceptionAsObject = CGF.Builder.CreateBitCast(ExceptionAsObject, IdTy);
llvm::CallBase *Throw =
CGF.EmitRuntimeCallOrInvoke(ExceptionThrowFn, ExceptionAsObject);
Throw->setDoesNotReturn();
}
CGF.Builder.CreateUnreachable();
if (ClearInsertionPoint)
CGF.Builder.ClearInsertionPoint();
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
llvm::Value * CGObjCGNU::EmitObjCWeakRead(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address AddrWeakObj) {
CGBuilderTy &B = CGF.Builder;
AddrWeakObj = EnforceType(B, AddrWeakObj, PtrToIdTy);
return B.CreateCall(WeakReadFn, AddrWeakObj.getPointer());
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
void CGObjCGNU::EmitObjCWeakAssign(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *src, Address dst) {
CGBuilderTy &B = CGF.Builder;
src = EnforceType(B, src, IdTy);
dst = EnforceType(B, dst, PtrToIdTy);
B.CreateCall(WeakAssignFn, {src, dst.getPointer()});
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
void CGObjCGNU::EmitObjCGlobalAssign(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *src, Address dst,
bool threadlocal) {
CGBuilderTy &B = CGF.Builder;
src = EnforceType(B, src, IdTy);
dst = EnforceType(B, dst, PtrToIdTy);
// FIXME. Add threadloca assign API
assert(!threadlocal && "EmitObjCGlobalAssign - Threal Local API NYI");
B.CreateCall(GlobalAssignFn, {src, dst.getPointer()});
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
void CGObjCGNU::EmitObjCIvarAssign(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *src, Address dst,
llvm::Value *ivarOffset) {
CGBuilderTy &B = CGF.Builder;
src = EnforceType(B, src, IdTy);
dst = EnforceType(B, dst, IdTy);
B.CreateCall(IvarAssignFn, {src, dst.getPointer(), ivarOffset});
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
void CGObjCGNU::EmitObjCStrongCastAssign(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
llvm::Value *src, Address dst) {
CGBuilderTy &B = CGF.Builder;
src = EnforceType(B, src, IdTy);
dst = EnforceType(B, dst, PtrToIdTy);
B.CreateCall(StrongCastAssignFn, {src, dst.getPointer()});
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
void CGObjCGNU::EmitGCMemmoveCollectable(CodeGenFunction &CGF,
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
Address DestPtr,
Address SrcPtr,
llvm::Value *Size) {
CGBuilderTy &B = CGF.Builder;
DestPtr = EnforceType(B, DestPtr, PtrTy);
SrcPtr = EnforceType(B, SrcPtr, PtrTy);
B.CreateCall(MemMoveFn, {DestPtr.getPointer(), SrcPtr.getPointer(), Size});
}
llvm::GlobalVariable *CGObjCGNU::ObjCIvarOffsetVariable(
const ObjCInterfaceDecl *ID,
const ObjCIvarDecl *Ivar) {
const std::string Name = GetIVarOffsetVariableName(ID, Ivar);
// Emit the variable and initialize it with what we think the correct value
// is. This allows code compiled with non-fragile ivars to work correctly
// when linked against code which isn't (most of the time).
llvm::GlobalVariable *IvarOffsetPointer = TheModule.getNamedGlobal(Name);
if (!IvarOffsetPointer)
IvarOffsetPointer = new llvm::GlobalVariable(TheModule,
llvm::Type::getInt32PtrTy(VMContext), false,
llvm::GlobalValue::ExternalLinkage, nullptr, Name);
return IvarOffsetPointer;
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
LValue CGObjCGNU::EmitObjCValueForIvar(CodeGenFunction &CGF,
QualType ObjectTy,
llvm::Value *BaseValue,
const ObjCIvarDecl *Ivar,
unsigned CVRQualifiers) {
const ObjCInterfaceDecl *ID =
ObjectTy->getAs<ObjCObjectType>()->getInterface();
return EmitValueForIvarAtOffset(CGF, ID, BaseValue, Ivar, CVRQualifiers,
EmitIvarOffset(CGF, ID, Ivar));
}
static const ObjCInterfaceDecl *FindIvarInterface(ASTContext &Context,
const ObjCInterfaceDecl *OID,
const ObjCIvarDecl *OIVD) {
for (const ObjCIvarDecl *next = OID->all_declared_ivar_begin(); next;
next = next->getNextIvar()) {
if (OIVD == next)
return OID;
}
// Otherwise check in the super class.
if (const ObjCInterfaceDecl *Super = OID->getSuperClass())
return FindIvarInterface(Context, Super, OIVD);
return nullptr;
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
llvm::Value *CGObjCGNU::EmitIvarOffset(CodeGenFunction &CGF,
const ObjCInterfaceDecl *Interface,
const ObjCIvarDecl *Ivar) {
if (CGM.getLangOpts().ObjCRuntime.isNonFragile()) {
Interface = FindIvarInterface(CGM.getContext(), Interface, Ivar);
// The MSVC linker cannot have a single global defined as LinkOnceAnyLinkage
// and ExternalLinkage, so create a reference to the ivar global and rely on
// the definition being created as part of GenerateClass.
if (RuntimeVersion < 10 ||
CGF.CGM.getTarget().getTriple().isKnownWindowsMSVCEnvironment())
return CGF.Builder.CreateZExtOrBitCast(
CGF.Builder.CreateAlignedLoad(
Int32Ty, CGF.Builder.CreateAlignedLoad(
ObjCIvarOffsetVariable(Interface, Ivar),
CGF.getPointerAlign(), "ivar"),
CharUnits::fromQuantity(4)),
PtrDiffTy);
std::string name = "__objc_ivar_offset_value_" +
Interface->getNameAsString() +"." + Ivar->getNameAsString();
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
CharUnits Align = CGM.getIntAlign();
llvm::Value *Offset = TheModule.getGlobalVariable(name);
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
if (!Offset) {
auto GV = new llvm::GlobalVariable(TheModule, IntTy,
false, llvm::GlobalValue::LinkOnceAnyLinkage,
llvm::Constant::getNullValue(IntTy), name);
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
GV->setAlignment(Align.getQuantity());
Offset = GV;
}
Offset = CGF.Builder.CreateAlignedLoad(Offset, Align);
if (Offset->getType() != PtrDiffTy)
Offset = CGF.Builder.CreateZExtOrBitCast(Offset, PtrDiffTy);
return Offset;
}
uint64_t Offset = ComputeIvarBaseOffset(CGF.CGM, Interface, Ivar);
return llvm::ConstantInt::get(PtrDiffTy, Offset, /*isSigned*/true);
}
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
CGObjCRuntime *
clang::CodeGen::CreateGNUObjCRuntime(CodeGenModule &CGM) {
auto Runtime = CGM.getLangOpts().ObjCRuntime;
switch (Runtime.getKind()) {
case ObjCRuntime::GNUstep:
if (Runtime.getVersion() >= VersionTuple(2, 0))
return new CGObjCGNUstep2(CGM);
Initial work on refactoring GNU runtime code (long overdue - it's quite obvious that I hadn't used C++ for several years before writing most of this code). Still lots more to do. This set of changes includes: - Remove the distinction between typed and untyped selectors. More accurately reflect what the runtime does, by using typed selectors everywhere, with an empty type field if the types are unknown. Now we just store a small list of types for each selector (in theory, this should always be exactly one, but this constraint was not enforced back in 1986 when it should have been). - Add some consistency to how runtime functions are created. These are all generated via the LazyRuntimeFunction class (which might be useful outside CGObjCGNU - feel free to move it into a header if it is). This function stores the types of a function, looks it up the first time it's used, and caches the result. This means that we're now not wasting time constructing the llvm::FunctionType every time some of the functions are looked up, but also not inserting references to runtime functions into the module if they're not actually used. - Started separating out the fragile and non-fragile ABI behaviours into two subclasses of CGObjCGNU: CGObjCGCC for the legacy GCC runtime ABI and CGObjCGNUstep for the new GNUstep ABI. Not all of the differences in behaviour are factored out yet, but they will be in future commits. - Removed all of the CodeGen:: things: we've been using namespace CodeGen in this file for ages, so having explicit namespace specifiers is just a bit confusing. - Added a few more comments. - Used llvm::StringRef instead of std::string in a few places. - Finally got around to storing the module path in the module structure. The ABI says that the compiler should do this, although it's not used in the runtime or exposed outside the runtime, so it's pretty useless. Still to do: - We currently have two code paths for generating try blocks, one for ObjC and one for ObjC++. Not only are these substantially similar, they are also very similar to the CGObjCMac version. These need factoring out into a single parameterised implementation, either in CGObjCRuntime or CodeGenFunction. The EmitObjCXXTryStmt() function was added so that the changes to fix a bug in time for the 2.9 release would be self-contained and reduce the chances of breaking anything else, but these should be done properly as soon as possible. - Split up some large functions (e.g. GenerateClass()) into smaller functions for generating the various data structures. - The method lookup code into the two subclasses, removing the conditionals in the message send functions. - Add doxygen comments on the remaining undocumented functions. - We seem to be generating global pointer variables for selectors, then storing a pointer to the selector, then generating a load of this pointer (and then a load of the real selector later) every time a static selector is used. I can only assume I was asleep or drunk when I did this - we should just be referencing the selectors directly in the selector array. llvm-svn: 128152
2011-03-24 00:36:54 +08:00
return new CGObjCGNUstep(CGM);
case ObjCRuntime::GCC:
return new CGObjCGCC(CGM);
case ObjCRuntime::ObjFW:
return new CGObjCObjFW(CGM);
case ObjCRuntime::FragileMacOSX:
case ObjCRuntime::MacOSX:
case ObjCRuntime::iOS:
case ObjCRuntime::WatchOS:
llvm_unreachable("these runtimes are not GNU runtimes");
}
llvm_unreachable("bad runtime");
}