/// Emit the actual writing-back of a writeback.
static void emitWriteback(CodeGenFunction &CGF,
const CallArgList::Writeback &writeback) {
- llvm::Value *srcAddr = writeback.Address;
+ const LValue &srcLV = writeback.Source;
+ llvm::Value *srcAddr = srcLV.getAddress();
assert(!isProvablyNull(srcAddr) &&
"shouldn't have writeback for provably null argument");
"icr.writeback-cast");
// Perform the writeback.
- QualType srcAddrType = writeback.AddressType;
- CGF.EmitStoreThroughLValue(RValue::get(value),
- CGF.MakeAddrLValue(srcAddr, srcAddrType));
+
+ // If we have a "to use" value, it's something we need to emit a use
+ // of. This has to be carefully threaded in: if it's done after the
+ // release it's potentially undefined behavior (and the optimizer
+ // will ignore it), and if it happens before the retain then the
+ // optimizer could move the release there.
+ if (writeback.ToUse) {
+ assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
+
+ // Retain the new value. No need to block-copy here: the block's
+ // being passed up the stack.
+ value = CGF.EmitARCRetainNonBlock(value);
+
+ // Emit the intrinsic use here.
+ CGF.EmitARCIntrinsicUse(writeback.ToUse);
+
+ // Load the old value (primitively).
+ llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV);
+
+ // Put the new value in place (primitively).
+ CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
+
+ // Release the old value.
+ CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
+
+ // Otherwise, we can just do a normal lvalue store.
+ } else {
+ CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
+ }
// Jump to the continuation block.
if (!provablyNonNull)
emitWriteback(CGF, *i);
}
+static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
+ if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
+ if (uop->getOpcode() == UO_AddrOf)
+ return uop->getSubExpr();
+ return 0;
+}
+
/// Emit an argument that's being passed call-by-writeback. That is,
/// we are passing the address of
static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
const ObjCIndirectCopyRestoreExpr *CRE) {
- llvm::Value *srcAddr = CGF.EmitScalarExpr(CRE->getSubExpr());
+ LValue srcLV;
+
+ // Make an optimistic effort to emit the address as an l-value.
+ // This can fail if the the argument expression is more complicated.
+ if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
+ srcLV = CGF.EmitLValue(lvExpr);
+
+ // Otherwise, just emit it as a scalar.
+ } else {
+ llvm::Value *srcAddr = CGF.EmitScalarExpr(CRE->getSubExpr());
+
+ QualType srcAddrType =
+ CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
+ srcLV = CGF.MakeNaturalAlignAddrLValue(srcAddr, srcAddrType);
+ }
+ llvm::Value *srcAddr = srcLV.getAddress();
// The dest and src types don't necessarily match in LLVM terms
// because of the crazy ObjC compatibility rules.
return;
}
- QualType srcAddrType =
- CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
-
// Create the temporary.
llvm::Value *temp = CGF.CreateTempAlloca(destType->getElementType(),
"icr.temp");
}
llvm::BasicBlock *contBB = 0;
+ llvm::BasicBlock *originBB = 0;
// If the address is *not* known to be non-null, we need to switch.
llvm::Value *finalArgument;
// If we need to copy, then the load has to be conditional, which
// means we need control flow.
if (shouldCopy) {
+ originBB = CGF.Builder.GetInsertBlock();
contBB = CGF.createBasicBlock("icr.cont");
llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
}
}
+ llvm::Value *valueToUse = 0;
+
// Perform a copy if necessary.
if (shouldCopy) {
- LValue srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
RValue srcRV = CGF.EmitLoadOfLValue(srcLV);
assert(srcRV.isScalar());
// Use an ordinary store, not a store-to-lvalue.
CGF.Builder.CreateStore(src, temp);
+
+ // If optimization is enabled, and the value was held in a
+ // __strong variable, we need to tell the optimizer that this
+ // value has to stay alive until we're doing the store back.
+ // This is because the temporary is effectively unretained,
+ // and so otherwise we can violate the high-level semantics.
+ if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
+ srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
+ valueToUse = src;
+ }
}
// Finish the control flow if we needed it.
if (shouldCopy && !provablyNonNull) {
+ llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
CGF.EmitBlock(contBB);
+
+ // Make a phi for the value to intrinsically use.
+ if (valueToUse) {
+ llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
+ "icr.to-use");
+ phiToUse->addIncoming(valueToUse, copyBB);
+ phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
+ originBB);
+ valueToUse = phiToUse;
+ }
+
condEval.end(CGF);
}
- args.addWriteback(srcAddr, srcAddrType, temp);
+ args.addWriteback(srcLV, temp, valueToUse);
args.add(RValue::get(finalArgument), CRE->getType());
}
public SmallVector<CallArg, 16> {
public:
struct Writeback {
- /// The original argument.
- llvm::Value *Address;
-
- /// The pointee type of the original argument.
- QualType AddressType;
+ /// The original argument. Note that the argument l-value
+ /// is potentially null.
+ LValue Source;
/// The temporary alloca.
llvm::Value *Temporary;
+
+ /// A value to "use" after the writeback, or null.
+ llvm::Value *ToUse;
};
void add(RValue rvalue, QualType type, bool needscopy = false) {
other.Writebacks.begin(), other.Writebacks.end());
}
- void addWriteback(llvm::Value *address, QualType addressType,
- llvm::Value *temporary) {
+ void addWriteback(LValue srcLV, llvm::Value *temporary,
+ llvm::Value *toUse) {
Writeback writeback;
- writeback.Address = address;
- writeback.AddressType = addressType;
+ writeback.Source = srcLV;
writeback.Temporary = temporary;
+ writeback.ToUse = toUse;
Writebacks.push_back(writeback);
}
return EmitARCRetainAutorelease(type, value);
}
+/// Given a number of pointers, inform the optimizer that they're
+/// being intrinsically used up until this point in the program.
+void CodeGenFunction::EmitARCIntrinsicUse(ArrayRef<llvm::Value*> values) {
+ llvm::Constant *&fn = CGM.getARCEntrypoints().clang_arc_use;
+ if (!fn) {
+ llvm::FunctionType *fnType =
+ llvm::FunctionType::get(CGM.VoidTy, ArrayRef<llvm::Type*>(), true);
+ fn = CGM.CreateRuntimeFunction(fnType, "clang.arc.use");
+ }
+
+ // This isn't really a "runtime" function, but as an intrinsic it
+ // doesn't really matter as long as we align things up.
+ EmitNounwindRuntimeCall(fn, values);
+}
+
static llvm::Constant *createARCRuntimeFunction(CodeGenModule &CGM,
llvm::FunctionType *type,
llvm::Value *EmitARCRetainScalarExpr(const Expr *expr);
llvm::Value *EmitARCRetainAutoreleaseScalarExpr(const Expr *expr);
+ void EmitARCIntrinsicUse(llvm::ArrayRef<llvm::Value*> values);
+
static Destroyer destroyARCStrongImprecise;
static Destroyer destroyARCStrongPrecise;
static Destroyer destroyARCWeak;
/// A void(void) inline asm to use to mark that the return value of
/// a call will be immediately retain.
llvm::InlineAsm *retainAutoreleasedReturnValueMarker;
+
+ /// void clang.arc.use(...);
+ llvm::Constant *clang_arc_use;
};
/// CodeGenModule - This class organizes the cross-function state that is used
// CHECK-NEXT: bitcast
// CHECK-NEXT: getelementptr
// CHECK-NEXT: [[BLOCK:%.*]] = bitcast
- // CHECK-NEXT: [[T0:%.*]] = load i8** [[STRONG]]
- // CHECK-NEXT: store i8* [[T0]], i8** [[TEMP]]
+ // CHECK-NEXT: [[V:%.*]] = load i8** [[STRONG]]
+ // CHECK-NEXT: store i8* [[V]], i8** [[TEMP]]
// CHECK-NEXT: [[F0:%.*]] = load i8**
// CHECK-NEXT: [[F1:%.*]] = bitcast i8* [[F0]] to void (i8*, i8**)*
// CHECK-NEXT: call void [[F1]](i8* [[BLOCK]], i8** [[TEMP]])
// CHECK-NEXT: [[T0:%.*]] = load i8** [[TEMP]]
// CHECK-NEXT: [[T1:%.*]] = call i8* @objc_retain(i8* [[T0]])
+ // CHECK-NEXT: call void (...)* @clang.arc.use(i8* [[V]]) [[NUW]]
// CHECK-NEXT: [[T2:%.*]] = load i8** [[STRONG]]
// CHECK-NEXT: store i8* [[T1]], i8** [[STRONG]]
// CHECK-NEXT: call void @objc_release(i8* [[T2]])
// CHECK: [[T0:%.*]] = load i8** [[ARG]]
// CHECK-NEXT: store i8* [[T0]], i8** [[TEMP1]]
// CHECK-NEXT: br label
- // CHECK: call void @test1_sink(i8** [[T1]])
+ // CHECK: [[W:%.*]] = phi i8* [ [[T0]], {{%.*}} ], [ undef, {{%.*}} ]
+ // CHECK-NEXT: call void @test1_sink(i8** [[T1]])
// CHECK-NEXT: [[T0:%.*]] = icmp eq i8** [[ARG]], null
// CHECK-NEXT: br i1 [[T0]],
// CHECK: [[T0:%.*]] = load i8** [[TEMP1]]
// CHECK-NEXT: [[T1:%.*]] = call i8* @objc_retain(i8* [[T0]])
+ // CHECK-NEXT: call void (...)* @clang.arc.use(i8* [[W]]) [[NUW]]
// CHECK-NEXT: [[T2:%.*]] = load i8** [[ARG]]
// CHECK-NEXT: store i8* [[T1]], i8** [[ARG]]
// CHECK-NEXT: call void @objc_release(i8* [[T2]])
// CHECK-NEXT: store [[A_T]]* null, [[A_T]]** [[A]]
// CHECK-NEXT: load [[TEST33]]** [[PTR]]
- // CHECK-NEXT: [[T0:%.*]] = load [[A_T]]** [[A]]
- // CHECK-NEXT: store [[A_T]]* [[T0]], [[A_T]]** [[TEMP0]]
+ // CHECK-NEXT: [[W0:%.*]] = load [[A_T]]** [[A]]
+ // CHECK-NEXT: store [[A_T]]* [[W0]], [[A_T]]** [[TEMP0]]
// CHECK-NEXT: load i8** @"\01L_OBJC_SELECTOR_REFERENCES_
// CHECK-NEXT: bitcast
// CHECK-NEXT: objc_msgSend{{.*}}, [[A_T]]** [[TEMP0]])
// CHECK-NEXT: [[T1:%.*]] = bitcast [[A_T]]* [[T0]] to i8*
// CHECK-NEXT: [[T2:%.*]] = call i8* @objc_retain(i8* [[T1]])
// CHECK-NEXT: [[T3:%.*]] = bitcast i8* [[T2]] to [[A_T]]*
+ // CHECK-NEXT: call void (...)* @clang.arc.use([[A_T]]* [[W0]]) [[NUW]]
// CHECK-NEXT: [[T4:%.*]] = load [[A_T]]** [[A]]
// CHECK-NEXT: store [[A_T]]* [[T3]], [[A_T]]** [[A]]
// CHECK-NEXT: [[T5:%.*]] = bitcast [[A_T]]* [[T4]] to i8*
// CHECK-NEXT: call void @objc_release(i8* [[T5]])
// CHECK-NEXT: load [[TEST33]]** [[PTR]]
- // CHECK-NEXT: [[T0:%.*]] = load [[A_T]]** [[A]]
- // CHECK-NEXT: store [[A_T]]* [[T0]], [[A_T]]** [[TEMP1]]
+ // CHECK-NEXT: [[W0:%.*]] = load [[A_T]]** [[A]]
+ // CHECK-NEXT: store [[A_T]]* [[W0]], [[A_T]]** [[TEMP1]]
// CHECK-NEXT: load i8** @"\01L_OBJC_SELECTOR_REFERENCES_
// CHECK-NEXT: bitcast
// CHECK-NEXT: objc_msgSend{{.*}}, [[A_T]]** [[TEMP1]])
// CHECK-NEXT: [[T1:%.*]] = bitcast [[A_T]]* [[T0]] to i8*
// CHECK-NEXT: [[T2:%.*]] = call i8* @objc_retain(i8* [[T1]])
// CHECK-NEXT: [[T3:%.*]] = bitcast i8* [[T2]] to [[A_T]]*
+ // CHECK-NEXT: call void (...)* @clang.arc.use([[A_T]]* [[W0]]) [[NUW]]
// CHECK-NEXT: [[T4:%.*]] = load [[A_T]]** [[A]]
// CHECK-NEXT: store [[A_T]]* [[T3]], [[A_T]]** [[A]]
// CHECK-NEXT: [[T5:%.*]] = bitcast [[A_T]]* [[T4]] to i8*
// CHECK-NEXT: [[TEMP:%.*]] = alloca i8*
// CHECK-NEXT: store [[TEST37]]* null, [[TEST37]]** [[VAR]]
- // CHECK-NEXT: [[T0:%.*]] = load [[TEST37]]** [[VAR]]
- // CHECK-NEXT: [[T1:%.*]] = bitcast [[TEST37]]* [[T0]] to i8*
- // CHECK-NEXT: store i8* [[T1]], i8** [[TEMP]]
+ // CHECK-NEXT: [[W0:%.*]] = load [[TEST37]]** [[VAR]]
+ // CHECK-NEXT: [[W1:%.*]] = bitcast [[TEST37]]* [[W0]] to i8*
+ // CHECK-NEXT: store i8* [[W1]], i8** [[TEMP]]
// CHECK-NEXT: call void @test37_helper(i8** [[TEMP]])
// CHECK-NEXT: [[T0:%.*]] = load i8** [[TEMP]]
// CHECK-NEXT: [[T1:%.*]] = bitcast i8* [[T0]] to [[TEST37]]*
// CHECK-NEXT: [[T2:%.*]] = bitcast [[TEST37]]* [[T1]] to i8*
// CHECK-NEXT: [[T3:%.*]] = call i8* @objc_retain(i8* [[T2]])
// CHECK-NEXT: [[T4:%.*]] = bitcast i8* [[T3]] to [[TEST37]]*
+ // CHECK-NEXT: call void (...)* @clang.arc.use(i8* [[W1]]) [[NUW]]
// CHECK-NEXT: [[T5:%.*]] = load [[TEST37]]** [[VAR]]
// CHECK-NEXT: store [[TEST37]]* [[T4]], [[TEST37]]** [[VAR]]
// CHECK-NEXT: [[T6:%.*]] = bitcast [[TEST37]]* [[T5]] to i8*
// CHECK: [[T0:%.*]] = load i8** [[ARG]]
// CHECK-NEXT: store i8* [[T0]], i8** [[TEMP1]]
// CHECK-NEXT: br label
+ // CHECK: [[W0:%.*]] = phi i8* [ [[T0]], {{%.*}} ], [ undef, {{%.*}} ]
// CHECK: call void @_Z11test34_sinkPU15__autoreleasingP11objc_object(i8** [[T1]])
// CHECK-NEXT: [[T0:%.*]] = icmp eq i8** [[ARG]], null
// CHECK-NEXT: br i1 [[T0]],
// CHECK: [[T0:%.*]] = load i8** [[TEMP1]]
// CHECK-NEXT: [[T1:%.*]] = call i8* @objc_retain(i8* [[T0]])
+ // CHECK-NEXT: call void (...)* @clang.arc.use(i8* [[W0]])
// CHECK-NEXT: [[T2:%.*]] = load i8** [[ARG]]
// CHECK-NEXT: store i8* [[T1]], i8** [[ARG]]
// CHECK-NEXT: call void @objc_release(i8* [[T2]])