+/// isReturnSlotOptznProfitable - Determine if performing a return slot
+/// fusion with the slot dest is profitable
+static bool isReturnSlotOptznProfitable(Value* dest, MemCpyInst* cpy) {
+ // We currently consider it profitable if dest is otherwise dead.
+ SmallVector<User*, 8> useList(dest->use_begin(), dest->use_end());
+ while (!useList.empty()) {
+ User* UI = useList.back();
+
+ if (isa<GetElementPtrInst>(UI) || isa<BitCastInst>(UI)) {
+ useList.pop_back();
+ for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
+ I != E; ++I)
+ useList.push_back(*I);
+ } else if (UI == cpy)
+ useList.pop_back();
+ else
+ return false;
+ }
+
+ return true;
+}
+
+/// performReturnSlotOptzn - takes a memcpy and a call that it depends on,
+/// and checks for the possibility of a return slot optimization by having
+/// the call write its result directly into the callees return parameter
+/// rather than using memcpy
+bool GVN::performReturnSlotOptzn(MemCpyInst* cpy, CallInst* C,
+ SmallVector<Instruction*, 4>& toErase) {
+ // Deliberately get the source and destination with bitcasts stripped away,
+ // because we'll need to do type comparisons based on the underlying type.
+ Value* cpyDest = cpy->getDest();
+ Value* cpySrc = cpy->getSource();
+ CallSite CS = CallSite::get(C);
+
+ // Since this is a return slot optimization, we need to make sure that
+ // the value being copied is, in fact, in a return slot. We also need to
+ // check that the return slot parameter is marked noalias, so that we can
+ // be sure that changing it will not cause unexpected behavior changes due
+ // to it being accessed through a global or another parameter.
+ if (CS.arg_size() == 0 ||
+ cpySrc != CS.getArgument(0) ||
+ !CS.paramHasAttr(1, ParamAttr::NoAlias | ParamAttr::StructRet))
+ return false;
+
+ // Check that something sneaky is not happening involving casting
+ // return slot types around.
+ if (CS.getArgument(0)->getType() != cpyDest->getType())
+ return false;
+ // sret --> pointer
+ const PointerType* PT = cast<PointerType>(cpyDest->getType());
+
+ // We can only perform the transformation if the size of the memcpy
+ // is constant and equal to the size of the structure.
+ ConstantInt* cpyLength = dyn_cast<ConstantInt>(cpy->getLength());
+ if (!cpyLength)
+ return false;
+
+ TargetData& TD = getAnalysis<TargetData>();
+ if (TD.getTypeStoreSize(PT->getElementType()) != cpyLength->getZExtValue())
+ return false;
+
+ // We only perform the transformation if it will be profitable.
+ if (!isReturnSlotOptznProfitable(cpyDest, cpy))
+ return false;
+
+ // In addition to knowing that the call does not access the return slot
+ // in some unexpected manner, which we derive from the noalias attribute,
+ // we also need to know that it does not sneakily modify the destination
+ // slot in the caller. We don't have parameter attributes to go by
+ // for this one, so we just rely on AA to figure it out for us.
+ AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
+ if (AA.getModRefInfo(C, cpy->getRawDest(), cpyLength->getZExtValue()) !=
+ AliasAnalysis::NoModRef)
+ return false;
+
+ // If all the checks have passed, then we're alright to do the transformation.
+ CS.setArgument(0, cpyDest);
+
+ // Drop any cached information about the call, because we may have changed
+ // its dependence information by changing its parameter.
+ MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
+ MD.dropInstruction(C);
+
+ // Remove the memcpy
+ MD.removeInstruction(cpy);
+ toErase.push_back(cpy);
+
+ return true;
+}
+
+/// processMemCpy - perform simplication of memcpy's. If we have memcpy A which
+/// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be
+/// a memcpy from X to Z (or potentially a memmove, depending on circumstances).
+/// This allows later passes to remove the first memcpy altogether.
+bool GVN::processMemCpy(MemCpyInst* M, MemCpyInst* MDep,
+ SmallVector<Instruction*, 4>& toErase) {
+ // We can only transforms memcpy's where the dest of one is the source of the
+ // other
+ if (M->getSource() != MDep->getDest())
+ return false;
+
+ // Second, the length of the memcpy's must be the same, or the preceeding one
+ // must be larger than the following one.
+ ConstantInt* C1 = dyn_cast<ConstantInt>(MDep->getLength());
+ ConstantInt* C2 = dyn_cast<ConstantInt>(M->getLength());
+ if (!C1 || !C2)
+ return false;
+
+ uint64_t CpySize = C1->getValue().getZExtValue();
+ uint64_t DepSize = C2->getValue().getZExtValue();
+
+ if (DepSize < CpySize)
+ return false;
+
+ // Finally, we have to make sure that the dest of the second does not
+ // alias the source of the first
+ AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
+ if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) !=
+ AliasAnalysis::NoAlias)
+ return false;
+ else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) !=
+ AliasAnalysis::NoAlias)
+ return false;
+ else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize)
+ != AliasAnalysis::NoAlias)
+ return false;
+
+ // If all checks passed, then we can transform these memcpy's
+ Function* MemCpyFun = Intrinsic::getDeclaration(
+ M->getParent()->getParent()->getParent(),
+ M->getIntrinsicID());
+
+ std::vector<Value*> args;
+ args.push_back(M->getRawDest());
+ args.push_back(MDep->getRawSource());
+ args.push_back(M->getLength());
+ args.push_back(M->getAlignment());
+
+ CallInst* C = new CallInst(MemCpyFun, args.begin(), args.end(), "", M);
+
+ MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
+ if (MD.getDependency(C) == MDep) {
+ MD.dropInstruction(M);
+ toErase.push_back(M);
+ return true;
+ } else {
+ MD.removeInstruction(C);
+ toErase.push_back(C);
+ return false;
+ }
+}
+
+/// processInstruction - When calculating availability, handle an instruction