Emit memmove, not memcpy, for structure copies; this is unfortunately

required for correctness in cases of copying a struct to itself or to 
an overlapping struct (itself for cases like *a = *a, and overlapping 
is possible with unions).

Hopefully, this won't end up being a perf issue; LLVM *should* be able 
to optimize memmove to memcpy in a lot of cases, and for small copies 
the generated code *should* be mostly comparable. (In reality, LLVM 
is currently horrible at optimizing memmove, but that's a bug, not a 
fundamental issue.)

gcc currently generates wrong code; that's 
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32667.

llvm-svn: 51566
This commit is contained in:
Eli Friedman 2008-05-26 12:59:39 +00:00
parent d3f21d165f
commit df649f3da5
3 changed files with 18 additions and 4 deletions

View File

@ -131,7 +131,7 @@ void AggExprEmitter::EmitAggregateCopy(llvm::Value *DestPtr,
llvm::Value *SrcPtr, QualType Ty) {
assert(!Ty->isAnyComplexType() && "Shouldn't happen for complex");
// Aggregate assignment turns into llvm.memcpy.
// Aggregate assignment turns into llvm.memmove.
const llvm::Type *BP = llvm::PointerType::getUnqual(llvm::Type::Int8Ty);
if (DestPtr->getType() != BP)
DestPtr = Builder.CreateBitCast(DestPtr, BP, "tmp");
@ -144,14 +144,14 @@ void AggExprEmitter::EmitAggregateCopy(llvm::Value *DestPtr,
// FIXME: Handle variable sized types.
const llvm::Type *IntPtr = llvm::IntegerType::get(CGF.LLVMPointerWidth);
llvm::Value *MemCpyOps[4] = {
llvm::Value *MemMoveOps[4] = {
DestPtr, SrcPtr,
// TypeInfo.first describes size in bits.
llvm::ConstantInt::get(IntPtr, TypeInfo.first/8),
llvm::ConstantInt::get(llvm::Type::Int32Ty, TypeInfo.second/8)
};
Builder.CreateCall(CGF.CGM.getMemCpyFn(), MemCpyOps, MemCpyOps+4);
Builder.CreateCall(CGF.CGM.getMemMoveFn(), MemMoveOps, MemMoveOps+4);
}

View File

@ -35,7 +35,8 @@ CodeGenModule::CodeGenModule(ASTContext &C, const LangOptions &LO,
llvm::Module &M, const llvm::TargetData &TD,
Diagnostic &diags, bool GenerateDebugInfo)
: Context(C), Features(LO), TheModule(M), TheTargetData(TD), Diags(diags),
Types(C, M, TD), MemCpyFn(0), MemSetFn(0), CFConstantStringClassRef(0) {
Types(C, M, TD), MemCpyFn(0), MemMoveFn(0), MemSetFn(0),
CFConstantStringClassRef(0) {
//TODO: Make this selectable at runtime
Runtime = CreateObjCRuntime(M,
getTypes().ConvertType(getContext().IntTy),
@ -589,6 +590,17 @@ llvm::Function *CodeGenModule::getMemCpyFn() {
return MemCpyFn = getIntrinsic(IID);
}
llvm::Function *CodeGenModule::getMemMoveFn() {
if (MemMoveFn) return MemMoveFn;
llvm::Intrinsic::ID IID;
switch (Context.Target.getPointerWidth(0)) {
default: assert(0 && "Unknown ptr width");
case 32: IID = llvm::Intrinsic::memmove_i32; break;
case 64: IID = llvm::Intrinsic::memmove_i64; break;
}
return MemMoveFn = getIntrinsic(IID);
}
llvm::Function *CodeGenModule::getMemSetFn() {
if (MemSetFn) return MemSetFn;
llvm::Intrinsic::ID IID;

View File

@ -59,6 +59,7 @@ class CodeGenModule {
CGDebugInfo *DebugInfo;
llvm::Function *MemCpyFn;
llvm::Function *MemMoveFn;
llvm::Function *MemSetFn;
llvm::DenseMap<const Decl*, llvm::Constant*> GlobalDeclMap;
std::vector<const NamedDecl*> StaticDecls;
@ -101,6 +102,7 @@ public:
/// array containing the literal. The result is pointer to array type.
llvm::Constant *GetAddrOfConstantString(const std::string& str);
llvm::Function *getMemCpyFn();
llvm::Function *getMemMoveFn();
llvm::Function *getMemSetFn();
llvm::Function *getIntrinsic(unsigned IID, const llvm::Type **Tys = 0,
unsigned NumTys = 0);