Hi rjmccall,
Static locals requiring initialization are not thread safe on Windows.
Unfortunately, it's possible to create static locals that are actually
externally visible with inline functions and templates. As a result, we
have to implement an initialization guard scheme that is compatible with
TUs built by MSVC, which makes thread safety prohibitively difficult.
MSVC's scheme is that every function that requires a guard gets an i32
bitfield. Each static local is assigned a bit that indicates if it has
been initialized, up to 32 bits, at which point a new bitfield is
created. MSVC rejects inline functions with more than 32 static locals,
and the externally visible mangling (?_B) only allows for one guard
variable per function.
Implements PR16888.
http://llvm-reviews.chandlerc.com/D1416
Files:
include/clang/AST/Mangle.h
lib/AST/ItaniumMangle.cpp
lib/AST/MicrosoftMangle.cpp
lib/CodeGen/CGCXXABI.cpp
lib/CodeGen/CGCXXABI.h
lib/CodeGen/CGDeclCXX.cpp
lib/CodeGen/CodeGenFunction.cpp
lib/CodeGen/CodeGenFunction.h
lib/CodeGen/ItaniumCXXABI.cpp
lib/CodeGen/MicrosoftCXXABI.cpp
test/CodeGenCXX/microsoft-abi-static-initializers.cpp
Index: include/clang/AST/Mangle.h
===================================================================
--- include/clang/AST/Mangle.h
+++ include/clang/AST/Mangle.h
@@ -139,11 +139,11 @@
void mangleObjCMethodName(const ObjCMethodDecl *MD,
raw_ostream &);
- // This is pretty lame.
- virtual void mangleItaniumGuardVariable(const VarDecl *D,
- raw_ostream &) {
- llvm_unreachable("Target does not support mangling guard variables");
- }
+ virtual void mangleStaticGuardVariable(const VarDecl *D,
+ raw_ostream &Out) = 0;
+ virtual void mangleDynamicAtExitDestructor(const VarDecl *D,
+ raw_ostream &Out) = 0;
+
// FIXME: Revisit this once we know what we need to do for MSVC compatibility.
virtual void mangleItaniumThreadLocalInit(const VarDecl *D,
raw_ostream &) {
Index: lib/AST/ItaniumMangle.cpp
===================================================================
--- lib/AST/ItaniumMangle.cpp
+++ lib/AST/ItaniumMangle.cpp
@@ -152,7 +152,8 @@
void mangleCXXDtor(const CXXDestructorDecl *D, CXXDtorType Type,
raw_ostream &);
- void mangleItaniumGuardVariable(const VarDecl *D, raw_ostream &);
+ void mangleStaticGuardVariable(const VarDecl *D, raw_ostream &);
+ void mangleDynamicAtExitDestructor(const VarDecl *D, raw_ostream &Out);
void mangleItaniumThreadLocalInit(const VarDecl *D, raw_ostream &);
void mangleItaniumThreadLocalWrapper(const VarDecl *D, raw_ostream &);
@@ -3696,15 +3697,22 @@
/// mangleGuardVariable - Returns the mangled name for a guard variable
/// for the passed in VarDecl.
-void ItaniumMangleContext::mangleItaniumGuardVariable(const VarDecl *D,
- raw_ostream &Out) {
+void ItaniumMangleContext::mangleStaticGuardVariable(const VarDecl *D,
+ raw_ostream &Out) {
// <special-name> ::= GV <object name> # Guard variable for one-time
// # initialization
CXXNameMangler Mangler(*this, Out);
Mangler.getStream() << "_ZGV";
Mangler.mangleName(D);
}
+void ItaniumMangleContext::mangleDynamicAtExitDestructor(const VarDecl *D,
+ raw_ostream &Out) {
+ // Prefix the mangling of D with __dtor_.
+ Out << "__dtor_";
+ mangleName(D, Out);
+}
+
void ItaniumMangleContext::mangleItaniumThreadLocalInit(const VarDecl *D,
raw_ostream &Out) {
// <special-name> ::= TH <object name>
Index: lib/AST/MicrosoftMangle.cpp
===================================================================
--- lib/AST/MicrosoftMangle.cpp
+++ lib/AST/MicrosoftMangle.cpp
@@ -92,15 +92,15 @@
QualifierMangleMode QMM = QMM_Mangle);
void mangleFunctionType(const FunctionType *T, const FunctionDecl *D,
bool IsStructor, bool IsInstMethod);
+ void manglePostfix(const DeclContext *DC, bool NoFunction = false);
private:
void disableBackReferences() { UseNameBackReferences = false; }
void mangleUnqualifiedName(const NamedDecl *ND) {
mangleUnqualifiedName(ND, ND->getDeclName());
}
void mangleUnqualifiedName(const NamedDecl *ND, DeclarationName Name);
void mangleSourceName(const IdentifierInfo *II);
- void manglePostfix(const DeclContext *DC, bool NoFunction=false);
void mangleOperatorName(OverloadedOperatorKind OO, SourceLocation Loc);
void mangleCXXDtorType(CXXDtorType T);
void mangleQualifiers(Qualifiers Quals, bool IsMember);
@@ -169,8 +169,10 @@
raw_ostream &);
virtual void mangleCXXDtor(const CXXDestructorDecl *D, CXXDtorType Type,
raw_ostream &);
- virtual void mangleReferenceTemporary(const clang::VarDecl *,
- raw_ostream &);
+ virtual void mangleReferenceTemporary(const VarDecl *, raw_ostream &);
+ virtual void mangleStaticGuardVariable(const VarDecl *D, raw_ostream &Out);
+ virtual void mangleDynamicAtExitDestructor(const VarDecl *D,
+ raw_ostream &Out);
};
}
@@ -1929,13 +1931,44 @@
MicrosoftCXXNameMangler mangler(*this, Out, D, Type);
mangler.mangle(D);
}
-void MicrosoftMangleContext::mangleReferenceTemporary(const clang::VarDecl *VD,
+void MicrosoftMangleContext::mangleReferenceTemporary(const VarDecl *VD,
raw_ostream &) {
unsigned DiagID = getDiags().getCustomDiagID(DiagnosticsEngine::Error,
"cannot mangle this reference temporary yet");
getDiags().Report(VD->getLocation(), DiagID);
}
+void MicrosoftMangleContext::mangleStaticGuardVariable(const VarDecl *VD,
+ raw_ostream &Out) {
+ // <guard-name> ::= ?_B <postfix> @51
+ // ::= ?$S <guard-num> @ <postfix> @4IA
+
+ // The first mangling is what MSVC uses to guard static locals in inline
+ // functions. It uses a different mangling in external functions to support
+ // guarding more than 32 variables. MSVC rejects inline functions with more
+ // than 32 static locals. We don't fully implement the second mangling
+ // because those guards are not externally visible, and instead use LLVM's
+ // default renaming when creating a new guard variable.
+ MicrosoftCXXNameMangler Mangler(*this, Out);
+
+ bool Visible = VD->isExternallyVisible();
+ // <operator-name> ::= ?_B # local static guard
+ Mangler.getStream() << (Visible ? "\01??_B" : "\01??S1@");
+ Mangler.manglePostfix(VD->getDeclContext());
+ Mangler.getStream() << (Visible ? "@51" : "@4IA");
+}
+
+void MicrosoftMangleContext::mangleDynamicAtExitDestructor(const VarDecl *D,
+ raw_ostream &Out) {
+ // <destructor-name> ::= ?__F <postfix> YAXXZ
+ MicrosoftCXXNameMangler Mangler(*this, Out);
+ Mangler.getStream() << "\01??__F";
+ Mangler.mangleName(D);
+ // This is the mangling of the function type of the stub, which is a global,
+ // non-variadic, cdecl function that returns void and takes no args.
+ Mangler.getStream() << "YAXXZ";
+}
+
MangleContext *clang::createMicrosoftMangleContext(ASTContext &Context,
DiagnosticsEngine &Diags) {
return new MicrosoftMangleContext(Context, Diags);
Index: lib/CodeGen/CGCXXABI.cpp
===================================================================
--- lib/CodeGen/CGCXXABI.cpp
+++ lib/CodeGen/CGCXXABI.cpp
@@ -17,6 +17,8 @@
using namespace clang;
using namespace CodeGen;
+CXXABIFunctionState::~CXXABIFunctionState() { }
+
CGCXXABI::~CGCXXABI() { }
void CGCXXABI::ErrorUnsupportedABI(CodeGenFunction &CGF, StringRef S) {
@@ -212,22 +214,15 @@
return llvm::ConstantInt::get(CGF.SizeTy, 0);
}
-void CGCXXABI::EmitGuardedInit(CodeGenFunction &CGF,
- const VarDecl &D,
- llvm::GlobalVariable *GV,
- bool PerformInit) {
- ErrorUnsupportedABI(CGF, "static local variable initialization");
-}
-
void CGCXXABI::registerGlobalDtor(CodeGenFunction &CGF,
const VarDecl &D,
llvm::Constant *dtor,
llvm::Constant *addr) {
if (D.getTLSKind())
CGM.ErrorUnsupported(&D, "non-trivial TLS destruction");
// The default behavior is to use atexit.
- CGF.registerGlobalDtorWithAtExit(dtor, addr);
+ CGF.registerGlobalDtorWithAtExit(D, dtor, addr);
}
/// Returns the adjustment, in bytes, required for the given
Index: lib/CodeGen/CGCXXABI.h
===================================================================
--- lib/CodeGen/CGCXXABI.h
+++ lib/CodeGen/CGCXXABI.h
@@ -97,6 +97,9 @@
return *MangleCtx;
}
+ /// Creates C++ ABI-specific state associated with a given function.
+ virtual CXXABIFunctionState *createStateForFunction() { return 0; }
+
/// Returns true if the given constructor or destructor is one of the
/// kinds that the ABI says returns 'this' (only applies when called
/// non-virtually for destructors).
@@ -383,7 +386,8 @@
/// - a static local variable
/// - a static data member of a class template instantiation
virtual void EmitGuardedInit(CodeGenFunction &CGF, const VarDecl &D,
- llvm::GlobalVariable *DeclPtr, bool PerformInit);
+ llvm::GlobalVariable *DeclPtr,
+ bool PerformInit) = 0;
/// Emit code to force the execution of a destructor during global
/// teardown. The default implementation of this uses atexit.
Index: lib/CodeGen/CGDeclCXX.cpp
===================================================================
--- lib/CodeGen/CGDeclCXX.cpp
+++ lib/CodeGen/CGDeclCXX.cpp
@@ -162,13 +162,18 @@
/// Create a stub function, suitable for being passed to atexit,
/// which passes the given address to the given destructor function.
static llvm::Constant *createAtExitStub(CodeGenModule &CGM,
+ const VarDecl &D,
llvm::Constant *dtor,
llvm::Constant *addr) {
// Get the destructor function type, void(*)(void).
llvm::FunctionType *ty = llvm::FunctionType::get(CGM.VoidTy, false);
+ SmallString<256> FnName;
+ {
+ llvm::raw_svector_ostream Out(FnName);
+ CGM.getCXXABI().getMangleContext().mangleDynamicAtExitDestructor(&D, Out);
+ }
llvm::Function *fn =
- CreateGlobalInitOrDestructFunction(CGM, ty,
- Twine("__dtor_", addr->getName()));
+ CreateGlobalInitOrDestructFunction(CGM, ty, FnName.str());
CodeGenFunction CGF(CGM);
@@ -192,10 +197,11 @@
}
/// Register a global destructor using the C atexit runtime function.
-void CodeGenFunction::registerGlobalDtorWithAtExit(llvm::Constant *dtor,
+void CodeGenFunction::registerGlobalDtorWithAtExit(const VarDecl &D,
+ llvm::Constant *dtor,
llvm::Constant *addr) {
// Create a function which calls the destructor.
- llvm::Constant *dtorStub = createAtExitStub(CGM, dtor, addr);
+ llvm::Constant *dtorStub = createAtExitStub(CGM, D, dtor, addr);
// extern "C" int atexit(void (*f)(void));
llvm::FunctionType *atexitTy =
Index: lib/CodeGen/CodeGenFunction.cpp
===================================================================
--- lib/CodeGen/CodeGenFunction.cpp
+++ lib/CodeGen/CodeGenFunction.cpp
@@ -52,6 +52,7 @@
TerminateHandler(0), TrapBB(0) {
if (!suppressNewContext)
CGM.getCXXABI().getMangleContext().startNewFunction();
+ CXXABIState.reset(CGM.getCXXABI().createStateForFunction());
llvm::FastMathFlags FMF;
if (CGM.getLangOpts().FastMath)
Index: lib/CodeGen/CodeGenFunction.h
===================================================================
--- lib/CodeGen/CodeGenFunction.h
+++ lib/CodeGen/CodeGenFunction.h
@@ -91,6 +91,12 @@
TEK_Aggregate
};
+/// \brief Holds C++ ABI specific per-function state.
+class CXXABIFunctionState {
+public:
+ virtual ~CXXABIFunctionState();
+};
+
/// CodeGenFunction - This class organizes the per-function state that is used
/// while generating LLVM code.
class CodeGenFunction : public CodeGenTypeCache {
@@ -278,6 +284,9 @@
llvm::BasicBlock *getInvokeDestImpl();
+ /// Per function state created by the C++ ABI.
+ OwningPtr<CXXABIFunctionState> CXXABIState;
+
template <class T>
typename DominatingValue<T>::saved_type saveValueInCond(T value) {
return DominatingValue<T>::save(*this, value);
@@ -2312,7 +2321,8 @@
/// Call atexit() with a function that passes the given argument to
/// the given function.
- void registerGlobalDtorWithAtExit(llvm::Constant *fn, llvm::Constant *addr);
+ void registerGlobalDtorWithAtExit(const VarDecl &D, llvm::Constant *fn,
+ llvm::Constant *addr);
/// Emit code in this function to perform a guarded variable
/// initialization. Guarded initializations are used when it's not
Index: lib/CodeGen/ItaniumCXXABI.cpp
===================================================================
--- lib/CodeGen/ItaniumCXXABI.cpp
+++ lib/CodeGen/ItaniumCXXABI.cpp
@@ -1125,7 +1125,7 @@
SmallString<256> guardName;
{
llvm::raw_svector_ostream out(guardName);
- getMangleContext().mangleItaniumGuardVariable(&D, out);
+ getMangleContext().mangleStaticGuardVariable(&D, out);
out.flush();
}
@@ -1293,7 +1293,7 @@
return CGM.AddCXXDtorEntry(dtor, addr);
}
- CGF.registerGlobalDtorWithAtExit(dtor, addr);
+ CGF.registerGlobalDtorWithAtExit(D, dtor, addr);
}
/// Get the appropriate linkage for the wrapper function. This is essentially
Index: lib/CodeGen/MicrosoftCXXABI.cpp
===================================================================
--- lib/CodeGen/MicrosoftCXXABI.cpp
+++ lib/CodeGen/MicrosoftCXXABI.cpp
@@ -27,10 +27,21 @@
namespace {
+class MicrosoftFunctionState : public CXXABIFunctionState {
+public:
+ MicrosoftFunctionState() : NumGuardedInitsEmitted(0), GuardGV(0) {}
+ int NumGuardedInitsEmitted;
+ llvm::GlobalVariable *GuardGV;
+};
+
class MicrosoftCXXABI : public CGCXXABI {
public:
MicrosoftCXXABI(CodeGenModule &CGM) : CGCXXABI(CGM) {}
+ CXXABIFunctionState *createStateForFunction() {
+ return new MicrosoftFunctionState();
+ }
+
bool HasThisReturn(GlobalDecl GD) const;
bool isReturnTypeIndirect(const CXXRecordDecl *RD) const {
@@ -619,17 +630,77 @@
}
void MicrosoftCXXABI::EmitGuardedInit(CodeGenFunction &CGF, const VarDecl &D,
- llvm::GlobalVariable *DeclPtr,
+ llvm::GlobalVariable *GV,
bool PerformInit) {
- // FIXME: this code was only tested for global initialization.
- // Not sure whether we want thread-safe static local variables as VS
- // doesn't make them thread-safe.
+ // MSVC always uses an i32 bitfield to guard initialization, which is *not*
+ // threadsafe. Since the user may be linking in inline functions compiled by
+ // cl.exe, there's no reason to provide a false sense of security by using
+ // critical sections here.
if (D.getTLSKind())
CGM.ErrorUnsupported(&D, "dynamic TLS initialization");
- // Emit the initializer and add a global destructor if appropriate.
- CGF.EmitCXXGlobalVarDeclInit(D, DeclPtr, PerformInit);
+ CGBuilderTy &Builder = CGF.Builder;
+ llvm::IntegerType *GuardTy = CGF.Int32Ty;
+ llvm::ConstantInt *Zero = llvm::ConstantInt::get(GuardTy, 0);
+
+ // Create the guard variable if we don't already have it (as we
+ // might if we're double-emitting this function body).
+ MicrosoftFunctionState *State =
+ static_cast<MicrosoftFunctionState *>(CGF.CXXABIState.get());
+ llvm::GlobalVariable *Guard = State->GuardGV;
+ int BitIdx = State->NumGuardedInitsEmitted++;
+ if (BitIdx >= 32) {
+ ErrorUnsupportedABI(CGF, "more than 32 guarded initializations");
+ BitIdx %= 32;
+ Guard = 0;
+ }
+
+ // Lazily create the i32 bitfield for this function.
+ if (!Guard) {
+ // Mangle the name for the guard.
+ SmallString<256> GuardName;
+ {
+ llvm::raw_svector_ostream Out(GuardName);
+ getMangleContext().mangleStaticGuardVariable(&D, Out);
+ Out.flush();
+ }
+
+ // Create the guard variable with a zero-initializer. Just absorb linkage
+ // and visibility from the guarded variable.
+ Guard = new llvm::GlobalVariable(CGM.getModule(), GuardTy, false,
+ GV->getLinkage(), Zero, GuardName.str());
+ Guard->setVisibility(GV->getVisibility());
+ State->GuardGV = Guard;
+ } else {
+ assert(Guard->getLinkage() == GV->getLinkage() &&
+ "static local from the same function had different linkage");
+ }
+
+ // Pseudo code for the test:
+ // if (!(GuardVar & MyGuardBit)) {
+ // GuardVar |= MyGuardBit;
+ // ... initialize the object ...;
+ // }
+
+ // Test our bit from the guard variable.
+ llvm::ConstantInt *Bit = llvm::ConstantInt::get(GuardTy, 1U << BitIdx);
+ llvm::LoadInst *LI = Builder.CreateLoad(Guard);
+ llvm::Value *IsInitialized =
+ Builder.CreateICmpNE(Builder.CreateAnd(LI, Bit), Zero);
+ llvm::BasicBlock *InitBlock = CGF.createBasicBlock("init");
+ llvm::BasicBlock *EndBlock = CGF.createBasicBlock("init.end");
+ Builder.CreateCondBr(IsInitialized, EndBlock, InitBlock);
+
+ // Set our bit in the guard variable and emit the initializer and add a global
+ // destructor if appropriate.
+ CGF.EmitBlock(InitBlock);
+ Builder.CreateStore(Builder.CreateOr(LI, Bit), Guard);
+ CGF.EmitCXXGlobalVarDeclInit(D, GV, PerformInit);
+ Builder.CreateBr(EndBlock);
+
+ // Continue.
+ CGF.EmitBlock(EndBlock);
}
// Member pointer helpers.
Index: test/CodeGenCXX/microsoft-abi-static-initializers.cpp
===================================================================
--- test/CodeGenCXX/microsoft-abi-static-initializers.cpp
+++ test/CodeGenCXX/microsoft-abi-static-initializers.cpp
@@ -7,13 +7,21 @@
// CHECK: define internal void [[INIT_s:@.*global_var.*]] [[NUW:#[0-9]+]]
// CHECK: %{{[.0-9A-Z_a-z]+}} = call x86_thiscallcc %struct.S* @"\01??0S@@QAE@XZ"
-// CHECK: call i32 @atexit(void ()* @"__dtor_\01?s@@3US@@A")
+// CHECK: call i32 @atexit(void ()* @"\01??__Fs@@YAXXZ")
// CHECK: ret void
-// CHECK: define internal void @"__dtor_\01?s@@3US@@A"() [[NUW]] {
+// CHECK: define internal void @"\01??__Fs@@YAXXZ"() [[NUW]] {
// CHECK: call x86_thiscallcc void @"\01??1S@@QAE@XZ"
// CHECK: ret void
+void StaticLocal() {
+ static S TheS;
+}
+// CHECK: define void @"\01?StaticLocal@@YAXXZ"()
+// CHECK: load i32* @"\01??S1@?1??StaticLocal@@YAXXZ@4IA"
+// CHECK: store i32 {{.*}}, i32* @"\01??S1@?1??StaticLocal@@YAXXZ@4IA"
+// CHECK: ret
+
// Force WeakODRLinkage by using templates
class A {
public:
@@ -29,20 +37,40 @@
template<typename T> A B<T>::foo;
+inline S &getS() {
+ static S TheS;
+ return TheS;
+}
+
void force_usage() {
(void)B<int>::foo; // (void) - force usage
+ getS();
}
+// CHECK-LABEL: define linkonce_odr %struct.S* @"\01?getS@@YAAAUS@@XZ"
+// CHECK: load i32* @"\01??_B?1??getS@@YAAAUS@@XZ@51"
+// CHECK: and i32 {{.*}}, 1
+// CHECK: icmp ne i32 {{.*}}, 0
+// CHECK: br i1
+// init:
+// CHECK: or i32 {{.*}}, 1
+// CHECK: store i32 {{.*}}, i32* @"\01??_B?1??getS@@YAAAUS@@XZ@51"
+// CHECK: call x86_thiscallcc %struct.S* @"\01??0S@@QAE@XZ"(%struct.S* @"\01?TheS@?1??getS@@YAAAUS@@XZ@4U2@A")
+// CHECK: call i32 @atexit(void ()* @"\01??__FTheS@?1??getS@@YAAAUS@@XZ@YAXXZ")
+// CHECK: br label
+// init.end:
+// CHECK: ret %struct.S* @"\01?TheS@?1??getS@@YAAAUS@@XZ@4U2@A"
+
// CHECK: define internal void [[INIT_foo:@.*global_var.*]] [[NUW]]
// CHECK: %{{[.0-9A-Z_a-z]+}} = call x86_thiscallcc %class.A* @"\01??0A@@QAE@XZ"
-// CHECK: call i32 @atexit(void ()* [[FOO_DTOR:@"__dtor_.*foo@.*]])
+// CHECK: call i32 @atexit(void ()* @"\01??__Ffoo@?$B@H@@YAXXZ")
// CHECK: ret void
// CHECK: define linkonce_odr x86_thiscallcc %class.A* @"\01??0A@@QAE@XZ"
// CHECK: define linkonce_odr x86_thiscallcc void @"\01??1A@@QAE@XZ"
-// CHECK: define internal void [[FOO_DTOR]]
+// CHECK: define internal void @"\01??__Ffoo@?$B@H@@YAXXZ"
// CHECK: call x86_thiscallcc void @"\01??1A@@QAE@XZ"{{.*}}foo
// CHECK: ret void
_______________________________________________
cfe-commits mailing list
[email protected]
http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
