+void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
+ StoreSDNode *ST = cast<StoreSDNode>(Node);
+ SDValue Chain = ST->getChain();
+ SDValue Ptr = ST->getBasePtr();
+ SDLoc dl(Node);
+
+ unsigned Alignment = ST->getAlignment();
+ bool isVolatile = ST->isVolatile();
+ bool isNonTemporal = ST->isNonTemporal();
+
+ if (!ST->isTruncatingStore()) {
+ if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
+ ReplaceNode(ST, OptStore);
+ return;
+ }
+
+ {
+ SDValue Value = ST->getValue();
+ MVT VT = Value.getSimpleValueType();
+ switch (TLI.getOperationAction(ISD::STORE, VT)) {
+ default: llvm_unreachable("This action is not supported yet!");
+ case TargetLowering::Legal:
+ // If this is an unaligned store and the target doesn't support it,
+ // expand it.
+ if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
+ Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
+ if (ST->getAlignment() < ABIAlignment)
+ ExpandUnalignedStore(cast<StoreSDNode>(Node),
+ DAG, TLI, this);
+ }
+ break;
+ case TargetLowering::Custom: {
+ SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
+ if (Res.getNode())
+ ReplaceNode(SDValue(Node, 0), Res);
+ return;
+ }
+ case TargetLowering::Promote: {
+ MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
+ assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
+ "Can only promote stores to same size type");
+ Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
+ SDValue Result =
+ DAG.getStore(Chain, dl, Value, Ptr,
+ ST->getPointerInfo(), isVolatile,
+ isNonTemporal, Alignment);
+ ReplaceNode(SDValue(Node, 0), Result);
+ break;
+ }
+ }
+ return;
+ }
+ } else {
+ SDValue Value = ST->getValue();
+
+ EVT StVT = ST->getMemoryVT();
+ unsigned StWidth = StVT.getSizeInBits();
+
+ if (StWidth != StVT.getStoreSizeInBits()) {
+ // Promote to a byte-sized store with upper bits zero if not
+ // storing an integral number of bytes. For example, promote
+ // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
+ EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
+ StVT.getStoreSizeInBits());
+ Value = DAG.getZeroExtendInReg(Value, dl, StVT);
+ SDValue Result =
+ DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
+ NVT, isVolatile, isNonTemporal, Alignment);
+ ReplaceNode(SDValue(Node, 0), Result);
+ } else if (StWidth & (StWidth - 1)) {
+ // If not storing a power-of-2 number of bits, expand as two stores.
+ assert(!StVT.isVector() && "Unsupported truncstore!");
+ unsigned RoundWidth = 1 << Log2_32(StWidth);
+ assert(RoundWidth < StWidth);
+ unsigned ExtraWidth = StWidth - RoundWidth;
+ assert(ExtraWidth < RoundWidth);
+ assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
+ "Store size not an integral number of bytes!");
+ EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
+ EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
+ SDValue Lo, Hi;
+ unsigned IncrementSize;
+
+ if (TLI.isLittleEndian()) {
+ // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
+ // Store the bottom RoundWidth bits.
+ Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
+ RoundVT,
+ isVolatile, isNonTemporal, Alignment);
+
+ // Store the remaining ExtraWidth bits.
+ IncrementSize = RoundWidth / 8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getIntPtrConstant(IncrementSize));
+ Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
+ DAG.getConstant(RoundWidth,
+ TLI.getShiftAmountTy(Value.getValueType())));
+ Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
+ ST->getPointerInfo().getWithOffset(IncrementSize),
+ ExtraVT, isVolatile, isNonTemporal,
+ MinAlign(Alignment, IncrementSize));
+ } else {
+ // Big endian - avoid unaligned stores.
+ // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
+ // Store the top RoundWidth bits.
+ Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
+ DAG.getConstant(ExtraWidth,
+ TLI.getShiftAmountTy(Value.getValueType())));
+ Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
+ RoundVT, isVolatile, isNonTemporal, Alignment);
+
+ // Store the remaining ExtraWidth bits.
+ IncrementSize = RoundWidth / 8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getIntPtrConstant(IncrementSize));
+ Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
+ ST->getPointerInfo().getWithOffset(IncrementSize),
+ ExtraVT, isVolatile, isNonTemporal,
+ MinAlign(Alignment, IncrementSize));
+ }
+
+ // The order of the stores doesn't matter.
+ SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
+ ReplaceNode(SDValue(Node, 0), Result);
+ } else {
+ switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
+ StVT.getSimpleVT())) {
+ default: llvm_unreachable("This action is not supported yet!");
+ case TargetLowering::Legal:
+ // If this is an unaligned store and the target doesn't support it,
+ // expand it.
+ if (!TLI.allowsUnalignedMemoryAccesses(ST->getMemoryVT())) {
+ Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
+ if (ST->getAlignment() < ABIAlignment)
+ ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
+ }
+ break;
+ case TargetLowering::Custom: {
+ SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
+ if (Res.getNode())
+ ReplaceNode(SDValue(Node, 0), Res);
+ return;
+ }
+ case TargetLowering::Expand:
+ assert(!StVT.isVector() &&
+ "Vector Stores are handled in LegalizeVectorOps");
+
+ // TRUNCSTORE:i16 i32 -> STORE i16
+ assert(TLI.isTypeLegal(StVT) &&
+ "Do not know how to expand this store!");
+ Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
+ SDValue Result =
+ DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
+ isVolatile, isNonTemporal, Alignment);
+ ReplaceNode(SDValue(Node, 0), Result);
+ break;
+ }
+ }
+ }
+}
+
+void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
+ LoadSDNode *LD = cast<LoadSDNode>(Node);
+ SDValue Chain = LD->getChain(); // The chain.
+ SDValue Ptr = LD->getBasePtr(); // The base pointer.
+ SDValue Value; // The value returned by the load op.
+ SDLoc dl(Node);
+
+ ISD::LoadExtType ExtType = LD->getExtensionType();
+ if (ExtType == ISD::NON_EXTLOAD) {
+ MVT VT = Node->getSimpleValueType(0);
+ SDValue RVal = SDValue(Node, 0);
+ SDValue RChain = SDValue(Node, 1);
+
+ switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
+ default: llvm_unreachable("This action is not supported yet!");
+ case TargetLowering::Legal:
+ // If this is an unaligned load and the target doesn't support it,
+ // expand it.
+ if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
+ Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ unsigned ABIAlignment =
+ TLI.getDataLayout()->getABITypeAlignment(Ty);
+ if (LD->getAlignment() < ABIAlignment){
+ ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
+ }
+ }
+ break;
+ case TargetLowering::Custom: {
+ SDValue Res = TLI.LowerOperation(RVal, DAG);
+ if (Res.getNode()) {
+ RVal = Res;
+ RChain = Res.getValue(1);
+ }
+ break;
+ }
+ case TargetLowering::Promote: {
+ MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
+ assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
+ "Can only promote loads to same size type");
+
+ SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getPointerInfo(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->isInvariant(), LD->getAlignment());
+ RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
+ RChain = Res.getValue(1);
+ break;
+ }
+ }
+ if (RChain.getNode() != Node) {
+ assert(RVal.getNode() != Node && "Load must be completely replaced");
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
+ ReplacedNode(Node);
+ }
+ return;
+ }
+
+ EVT SrcVT = LD->getMemoryVT();
+ unsigned SrcWidth = SrcVT.getSizeInBits();
+ unsigned Alignment = LD->getAlignment();
+ bool isVolatile = LD->isVolatile();
+ bool isNonTemporal = LD->isNonTemporal();
+
+ if (SrcWidth != SrcVT.getStoreSizeInBits() &&
+ // Some targets pretend to have an i1 loading operation, and actually
+ // load an i8. This trick is correct for ZEXTLOAD because the top 7
+ // bits are guaranteed to be zero; it helps the optimizers understand
+ // that these bits are zero. It is also useful for EXTLOAD, since it
+ // tells the optimizers that those bits are undefined. It would be
+ // nice to have an effective generic way of getting these benefits...
+ // Until such a way is found, don't insist on promoting i1 here.
+ (SrcVT != MVT::i1 ||
+ TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
+ // Promote to a byte-sized load if not loading an integral number of
+ // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
+ unsigned NewWidth = SrcVT.getStoreSizeInBits();
+ EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
+ SDValue Ch;
+
+ // The extra bits are guaranteed to be zero, since we stored them that
+ // way. A zext load from NVT thus automatically gives zext from SrcVT.
+
+ ISD::LoadExtType NewExtType =
+ ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
+
+ SDValue Result =
+ DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
+ Chain, Ptr, LD->getPointerInfo(),
+ NVT, isVolatile, isNonTemporal, Alignment);
+
+ Ch = Result.getValue(1); // The chain.
+
+ if (ExtType == ISD::SEXTLOAD)
+ // Having the top bits zero doesn't help when sign extending.
+ Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
+ Result.getValueType(),
+ Result, DAG.getValueType(SrcVT));
+ else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
+ // All the top bits are guaranteed to be zero - inform the optimizers.
+ Result = DAG.getNode(ISD::AssertZext, dl,
+ Result.getValueType(), Result,
+ DAG.getValueType(SrcVT));
+
+ Value = Result;
+ Chain = Ch;
+ } else if (SrcWidth & (SrcWidth - 1)) {
+ // If not loading a power-of-2 number of bits, expand as two loads.
+ assert(!SrcVT.isVector() && "Unsupported extload!");
+ unsigned RoundWidth = 1 << Log2_32(SrcWidth);
+ assert(RoundWidth < SrcWidth);
+ unsigned ExtraWidth = SrcWidth - RoundWidth;
+ assert(ExtraWidth < RoundWidth);
+ assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
+ "Load size not an integral number of bytes!");
+ EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
+ EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
+ SDValue Lo, Hi, Ch;
+ unsigned IncrementSize;
+
+ if (TLI.isLittleEndian()) {
+ // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
+ // Load the bottom RoundWidth bits.
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
+ Chain, Ptr,
+ LD->getPointerInfo(), RoundVT, isVolatile,
+ isNonTemporal, Alignment);
+
+ // Load the remaining ExtraWidth bits.
+ IncrementSize = RoundWidth / 8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getIntPtrConstant(IncrementSize));
+ Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
+ LD->getPointerInfo().getWithOffset(IncrementSize),
+ ExtraVT, isVolatile, isNonTemporal,
+ MinAlign(Alignment, IncrementSize));
+
+ // Build a factor node to remember that this load is independent of
+ // the other one.
+ Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
+ Hi.getValue(1));
+
+ // Move the top bits to the right place.
+ Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
+ DAG.getConstant(RoundWidth,
+ TLI.getShiftAmountTy(Hi.getValueType())));
+
+ // Join the hi and lo parts.
+ Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
+ } else {
+ // Big endian - avoid unaligned loads.
+ // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
+ // Load the top RoundWidth bits.
+ Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
+ LD->getPointerInfo(), RoundVT, isVolatile,
+ isNonTemporal, Alignment);
+
+ // Load the remaining ExtraWidth bits.
+ IncrementSize = RoundWidth / 8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getIntPtrConstant(IncrementSize));
+ Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
+ dl, Node->getValueType(0), Chain, Ptr,
+ LD->getPointerInfo().getWithOffset(IncrementSize),
+ ExtraVT, isVolatile, isNonTemporal,
+ MinAlign(Alignment, IncrementSize));
+
+ // Build a factor node to remember that this load is independent of
+ // the other one.
+ Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
+ Hi.getValue(1));
+
+ // Move the top bits to the right place.
+ Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
+ DAG.getConstant(ExtraWidth,
+ TLI.getShiftAmountTy(Hi.getValueType())));
+
+ // Join the hi and lo parts.
+ Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
+ }
+
+ Chain = Ch;
+ } else {
+ bool isCustom = false;
+ switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
+ default: llvm_unreachable("This action is not supported yet!");
+ case TargetLowering::Custom:
+ isCustom = true;
+ // FALLTHROUGH
+ case TargetLowering::Legal: {
+ Value = SDValue(Node, 0);
+ Chain = SDValue(Node, 1);
+
+ if (isCustom) {
+ SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
+ if (Res.getNode()) {
+ Value = Res;
+ Chain = Res.getValue(1);
+ }
+ } else {
+ // If this is an unaligned load and the target doesn't support it,
+ // expand it.
+ if (!TLI.allowsUnalignedMemoryAccesses(LD->getMemoryVT())) {
+ Type *Ty =
+ LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
+ unsigned ABIAlignment =
+ TLI.getDataLayout()->getABITypeAlignment(Ty);
+ if (LD->getAlignment() < ABIAlignment){
+ ExpandUnalignedLoad(cast<LoadSDNode>(Node),
+ DAG, TLI, Value, Chain);
+ }
+ }
+ }
+ break;
+ }
+ case TargetLowering::Expand:
+ if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && TLI.isTypeLegal(SrcVT)) {
+ SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr,
+ LD->getPointerInfo(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->isInvariant(), LD->getAlignment());
+ unsigned ExtendOp;
+ switch (ExtType) {
+ case ISD::EXTLOAD:
+ ExtendOp = (SrcVT.isFloatingPoint() ?
+ ISD::FP_EXTEND : ISD::ANY_EXTEND);
+ break;
+ case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
+ case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
+ default: llvm_unreachable("Unexpected extend load type!");
+ }
+ Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
+ Chain = Load.getValue(1);
+ break;
+ }
+
+ assert(!SrcVT.isVector() &&
+ "Vector Loads are handled in LegalizeVectorOps");
+
+ // FIXME: This does not work for vectors on most targets. Sign- and
+ // zero-extend operations are currently folded into extending loads,
+ // whether they are legal or not, and then we end up here without any
+ // support for legalizing them.
+ assert(ExtType != ISD::EXTLOAD &&
+ "EXTLOAD should always be supported!");
+ // Turn the unsupported load into an EXTLOAD followed by an explicit
+ // zero/sign extend inreg.
+ SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl, Node->getValueType(0),
+ Chain, Ptr, LD->getPointerInfo(), SrcVT,
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment());
+ SDValue ValRes;
+ if (ExtType == ISD::SEXTLOAD)
+ ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
+ Result.getValueType(),
+ Result, DAG.getValueType(SrcVT));
+ else
+ ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
+ Value = ValRes;
+ Chain = Result.getValue(1);
+ break;
+ }
+ }
+
+ // Since loads produce two values, make sure to remember that we legalized
+ // both of them.
+ if (Chain.getNode() != Node) {
+ assert(Value.getNode() != Node && "Load must be completely replaced");
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
+ ReplacedNode(Node);
+ }
+}
+