aboutsummaryrefslogtreecommitdiffstats
path: root/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp')
-rw-r--r--lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp166
1 files changed, 162 insertions, 4 deletions
diff --git a/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp b/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
index ffff10ce29..7fe35306bd 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
@@ -61,6 +61,11 @@ class VectorLegalizer {
// Implements expansion for UINT_TO_FLOAT; falls back to UnrollVectorOp if
// SINT_TO_FLOAT and SHR on vectors isn't legal.
SDValue ExpandUINT_TO_FLOAT(SDValue Op);
+ // Implement vselect in terms of XOR, AND, OR when blend is not supported
+ // by the target.
+ SDValue ExpandVSELECT(SDValue Op);
+ SDValue ExpandLoad(SDValue Op);
+ SDValue ExpandStore(SDValue Op);
SDValue ExpandFNEG(SDValue Op);
// Implements vector promotion; this is essentially just bitcasting the
// operands to a different type and bitcasting the result back to the
@@ -121,6 +126,33 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
SDValue Result =
SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops.data(), Ops.size()), 0);
+ if (Op.getOpcode() == ISD::LOAD) {
+ LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
+ ISD::LoadExtType ExtType = LD->getExtensionType();
+ if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD) {
+ if (TLI.isLoadExtLegal(LD->getExtensionType(), LD->getMemoryVT()))
+ return TranslateLegalizeResults(Op, Result);
+ Changed = true;
+ return LegalizeOp(ExpandLoad(Op));
+ }
+ } else if (Op.getOpcode() == ISD::STORE) {
+ StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
+ EVT StVT = ST->getMemoryVT();
+ EVT ValVT = ST->getValue().getValueType();
+ if (StVT.isVector() && ST->isTruncatingStore())
+ switch (TLI.getTruncStoreAction(ValVT, StVT)) {
+ default: assert(0 && "This action is not supported yet!");
+ case TargetLowering::Legal:
+ return TranslateLegalizeResults(Op, Result);
+ case TargetLowering::Custom:
+ Changed = true;
+ return LegalizeOp(TLI.LowerOperation(Result, DAG));
+ case TargetLowering::Expand:
+ Changed = true;
+ return LegalizeOp(ExpandStore(Op));
+ }
+ }
+
bool HasVectorValue = false;
for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
J != E;
@@ -157,8 +189,9 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
case ISD::CTLZ:
case ISD::CTPOP:
case ISD::SELECT:
+ case ISD::VSELECT:
case ISD::SELECT_CC:
- case ISD::VSETCC:
+ case ISD::SETCC:
case ISD::ZERO_EXTEND:
case ISD::ANY_EXTEND:
case ISD::TRUNCATE:
@@ -210,11 +243,13 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
// FALL THROUGH
}
case TargetLowering::Expand:
- if (Node->getOpcode() == ISD::UINT_TO_FP)
+ if (Node->getOpcode() == ISD::VSELECT)
+ Result = ExpandVSELECT(Op);
+ else if (Node->getOpcode() == ISD::UINT_TO_FP)
Result = ExpandUINT_TO_FLOAT(Op);
else if (Node->getOpcode() == ISD::FNEG)
Result = ExpandFNEG(Op);
- else if (Node->getOpcode() == ISD::VSETCC)
+ else if (Node->getOpcode() == ISD::SETCC)
Result = UnrollVSETCC(Op);
else
Result = DAG.UnrollVectorOp(Op.getNode());
@@ -256,9 +291,132 @@ SDValue VectorLegalizer::PromoteVectorOp(SDValue Op) {
return DAG.getNode(ISD::BITCAST, dl, VT, Op);
}
-SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
+SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
+ DebugLoc dl = Op.getDebugLoc();
+ LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
+ SDValue Chain = LD->getChain();
+ SDValue BasePTR = LD->getBasePtr();
+ EVT SrcVT = LD->getMemoryVT();
+ ISD::LoadExtType ExtType = LD->getExtensionType();
+
+ SmallVector<SDValue, 8> LoadVals;
+ SmallVector<SDValue, 8> LoadChains;
+ unsigned NumElem = SrcVT.getVectorNumElements();
+ unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
+
+ for (unsigned Idx=0; Idx<NumElem; Idx++) {
+ SDValue ScalarLoad = DAG.getExtLoad(ExtType, dl,
+ Op.getNode()->getValueType(0).getScalarType(),
+ Chain, BasePTR, LD->getPointerInfo().getWithOffset(Idx * Stride),
+ SrcVT.getScalarType(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment());
+
+ BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
+ DAG.getIntPtrConstant(Stride));
+
+ LoadVals.push_back(ScalarLoad.getValue(0));
+ LoadChains.push_back(ScalarLoad.getValue(1));
+ }
+
+ SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &LoadChains[0], LoadChains.size());
+ SDValue Value = DAG.getNode(ISD::BUILD_VECTOR, dl,
+ Op.getNode()->getValueType(0), &LoadVals[0], LoadVals.size());
+
+ AddLegalizedOperand(Op.getValue(0), Value);
+ AddLegalizedOperand(Op.getValue(1), NewChain);
+
+ return (Op.getResNo() ? NewChain : Value);
+}
+SDValue VectorLegalizer::ExpandStore(SDValue Op) {
+ DebugLoc dl = Op.getDebugLoc();
+ StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
+ SDValue Chain = ST->getChain();
+ SDValue BasePTR = ST->getBasePtr();
+ SDValue Value = ST->getValue();
+ EVT StVT = ST->getMemoryVT();
+
+ unsigned Alignment = ST->getAlignment();
+ bool isVolatile = ST->isVolatile();
+ bool isNonTemporal = ST->isNonTemporal();
+
+ unsigned NumElem = StVT.getVectorNumElements();
+ // The type of the data we want to save
+ EVT RegVT = Value.getValueType();
+ EVT RegSclVT = RegVT.getScalarType();
+ // The type of data as saved in memory.
+ EVT MemSclVT = StVT.getScalarType();
+
+ // Cast floats into integers
+ unsigned ScalarSize = MemSclVT.getSizeInBits();
+
+ // Round odd types to the next pow of two.
+ if (!isPowerOf2_32(ScalarSize))
+ ScalarSize = NextPowerOf2(ScalarSize);
+
+ // Store Stride in bytes
+ unsigned Stride = ScalarSize/8;
+ // Extract each of the elements from the original vector
+ // and save them into memory individually.
+ SmallVector<SDValue, 8> Stores;
+ for (unsigned Idx = 0; Idx < NumElem; Idx++) {
+ SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
+ RegSclVT, Value, DAG.getIntPtrConstant(Idx));
+
+ // This scalar TruncStore may be illegal, but we legalize it later.
+ SDValue Store = DAG.getTruncStore(Chain, dl, Ex, BasePTR,
+ ST->getPointerInfo().getWithOffset(Idx*Stride), MemSclVT,
+ isVolatile, isNonTemporal, Alignment);
+
+ BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
+ DAG.getIntPtrConstant(Stride));
+
+ Stores.push_back(Store);
+ }
+ SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &Stores[0], Stores.size());
+ AddLegalizedOperand(Op, TF);
+ return TF;
+}
+
+SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
+ // Implement VSELECT in terms of XOR, AND, OR
+ // on platforms which do not support blend natively.
+ EVT VT = Op.getOperand(0).getValueType();
+ DebugLoc DL = Op.getDebugLoc();
+
+ SDValue Mask = Op.getOperand(0);
+ SDValue Op1 = Op.getOperand(1);
+ SDValue Op2 = Op.getOperand(2);
+
+ // If we can't even use the basic vector operations of
+ // AND,OR,XOR, we will have to scalarize the op.
+ if (!TLI.isOperationLegalOrCustom(ISD::AND, VT) ||
+ !TLI.isOperationLegalOrCustom(ISD::XOR, VT) ||
+ !TLI.isOperationLegalOrCustom(ISD::OR, VT))
+ return DAG.UnrollVectorOp(Op.getNode());
+
+ assert(VT.getSizeInBits() == Op.getOperand(1).getValueType().getSizeInBits()
+ && "Invalid mask size");
+ // Bitcast the operands to be the same type as the mask.
+ // This is needed when we select between FP types because
+ // the mask is a vector of integers.
+ Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
+ Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
+
+ SDValue AllOnes = DAG.getConstant(
+ APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), VT);
+ SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
+
+ Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
+ Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
+ return DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
+}
+
+SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
EVT VT = Op.getOperand(0).getValueType();
DebugLoc DL = Op.getDebugLoc();