//===----------------------------------------------------------------------===//
void DAGTypeLegalizer::ScalarizeVectorResult(SDNode *N, unsigned ResNo) {
- DEBUG(errs() << "Scalarize node result " << ResNo << ": ";
+ DEBUG(dbgs() << "Scalarize node result " << ResNo << ": ";
N->dump(&DAG);
- errs() << "\n");
+ dbgs() << "\n");
SDValue R = SDValue();
switch (N->getOpcode()) {
default:
#ifndef NDEBUG
- errs() << "ScalarizeVectorResult #" << ResNo << ": ";
+ dbgs() << "ScalarizeVectorResult #" << ResNo << ": ";
N->dump(&DAG);
- errs() << "\n";
+ dbgs() << "\n";
#endif
llvm_unreachable("Do not know how to scalarize the result of this operator!");
case ISD::BUILD_VECTOR: R = N->getOperand(0); break;
case ISD::CONVERT_RNDSAT: R = ScalarizeVecRes_CONVERT_RNDSAT(N); break;
case ISD::EXTRACT_SUBVECTOR: R = ScalarizeVecRes_EXTRACT_SUBVECTOR(N); break;
+ case ISD::FP_ROUND_INREG: R = ScalarizeVecRes_InregOp(N); break;
case ISD::FPOWI: R = ScalarizeVecRes_FPOWI(N); break;
case ISD::INSERT_VECTOR_ELT: R = ScalarizeVecRes_INSERT_VECTOR_ELT(N); break;
case ISD::LOAD: R = ScalarizeVecRes_LOAD(cast<LoadSDNode>(N));break;
case ISD::SCALAR_TO_VECTOR: R = ScalarizeVecRes_SCALAR_TO_VECTOR(N); break;
+ case ISD::SIGN_EXTEND_INREG: R = ScalarizeVecRes_InregOp(N); break;
case ISD::SELECT: R = ScalarizeVecRes_SELECT(N); break;
case ISD::SELECT_CC: R = ScalarizeVecRes_SELECT_CC(N); break;
case ISD::SETCC: R = ScalarizeVecRes_SETCC(N); break;
DAG.getUNDEF(N->getBasePtr().getValueType()),
N->getSrcValue(), N->getSrcValueOffset(),
N->getMemoryVT().getVectorElementType(),
- N->isVolatile(), N->getOriginalAlignment());
+ N->isVolatile(), N->isNonTemporal(),
+ N->getOriginalAlignment());
// Legalized the chain result - switch anything that used the old chain to
// use the new one.
return DAG.getNode(N->getOpcode(), N->getDebugLoc(), DestVT, Op);
}
+SDValue DAGTypeLegalizer::ScalarizeVecRes_InregOp(SDNode *N) {
+ EVT EltVT = N->getValueType(0).getVectorElementType();
+ EVT ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT().getVectorElementType();
+ SDValue LHS = GetScalarizedVector(N->getOperand(0));
+ return DAG.getNode(N->getOpcode(), N->getDebugLoc(), EltVT,
+ LHS, DAG.getValueType(ExtVT));
+}
+
SDValue DAGTypeLegalizer::ScalarizeVecRes_SCALAR_TO_VECTOR(SDNode *N) {
// If the operand is wider than the vector element type then it is implicitly
// truncated. Make that explicit here.
//===----------------------------------------------------------------------===//
bool DAGTypeLegalizer::ScalarizeVectorOperand(SDNode *N, unsigned OpNo) {
- DEBUG(errs() << "Scalarize node operand " << OpNo << ": ";
+ DEBUG(dbgs() << "Scalarize node operand " << OpNo << ": ";
N->dump(&DAG);
- errs() << "\n");
+ dbgs() << "\n");
SDValue Res = SDValue();
if (Res.getNode() == 0) {
switch (N->getOpcode()) {
default:
#ifndef NDEBUG
- errs() << "ScalarizeVectorOperand Op #" << OpNo << ": ";
+ dbgs() << "ScalarizeVectorOperand Op #" << OpNo << ": ";
N->dump(&DAG);
- errs() << "\n";
+ dbgs() << "\n";
#endif
llvm_unreachable("Do not know how to scalarize this operator's operand!");
case ISD::BIT_CONVERT:
N->getBasePtr(),
N->getSrcValue(), N->getSrcValueOffset(),
N->getMemoryVT().getVectorElementType(),
- N->isVolatile(), N->getAlignment());
+ N->isVolatile(), N->isNonTemporal(),
+ N->getAlignment());
return DAG.getStore(N->getChain(), dl, GetScalarizedVector(N->getOperand(1)),
N->getBasePtr(), N->getSrcValue(), N->getSrcValueOffset(),
- N->isVolatile(), N->getOriginalAlignment());
+ N->isVolatile(), N->isNonTemporal(),
+ N->getOriginalAlignment());
}
/// legalization, we just know that (at least) one result needs vector
/// splitting.
void DAGTypeLegalizer::SplitVectorResult(SDNode *N, unsigned ResNo) {
- DEBUG(errs() << "Split node result: ";
+ DEBUG(dbgs() << "Split node result: ";
N->dump(&DAG);
- errs() << "\n");
+ dbgs() << "\n");
SDValue Lo, Hi;
switch (N->getOpcode()) {
default:
#ifndef NDEBUG
- errs() << "SplitVectorResult #" << ResNo << ": ";
+ dbgs() << "SplitVectorResult #" << ResNo << ": ";
N->dump(&DAG);
- errs() << "\n";
+ dbgs() << "\n";
#endif
llvm_unreachable("Do not know how to split the result of this operator!");
case ISD::CONCAT_VECTORS: SplitVecRes_CONCAT_VECTORS(N, Lo, Hi); break;
case ISD::CONVERT_RNDSAT: SplitVecRes_CONVERT_RNDSAT(N, Lo, Hi); break;
case ISD::EXTRACT_SUBVECTOR: SplitVecRes_EXTRACT_SUBVECTOR(N, Lo, Hi); break;
+ case ISD::FP_ROUND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break;
case ISD::FPOWI: SplitVecRes_FPOWI(N, Lo, Hi); break;
case ISD::INSERT_VECTOR_ELT: SplitVecRes_INSERT_VECTOR_ELT(N, Lo, Hi); break;
case ISD::SCALAR_TO_VECTOR: SplitVecRes_SCALAR_TO_VECTOR(N, Lo, Hi); break;
+ case ISD::SIGN_EXTEND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break;
case ISD::LOAD:
SplitVecRes_LOAD(cast<LoadSDNode>(N), Lo, Hi);
break;
case ISD::SIGN_EXTEND:
case ISD::ZERO_EXTEND:
case ISD::ANY_EXTEND:
+ case ISD::FEXP:
+ case ISD::FEXP2:
+ case ISD::FLOG:
+ case ISD::FLOG2:
+ case ISD::FLOG10:
SplitVecRes_UnaryOp(N, Lo, Hi);
break;
Hi = DAG.getNode(ISD::FPOWI, dl, Hi.getValueType(), Hi, N->getOperand(1));
}
+void DAGTypeLegalizer::SplitVecRes_InregOp(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ SDValue LHSLo, LHSHi;
+ GetSplitVector(N->getOperand(0), LHSLo, LHSHi);
+ DebugLoc dl = N->getDebugLoc();
+
+ EVT LoVT, HiVT;
+ GetSplitDestVTs(cast<VTSDNode>(N->getOperand(1))->getVT(), LoVT, HiVT);
+
+ Lo = DAG.getNode(N->getOpcode(), dl, LHSLo.getValueType(), LHSLo,
+ DAG.getValueType(LoVT));
+ Hi = DAG.getNode(N->getOpcode(), dl, LHSHi.getValueType(), LHSHi,
+ DAG.getValueType(HiVT));
+}
+
void DAGTypeLegalizer::SplitVecRes_INSERT_VECTOR_ELT(SDNode *N, SDValue &Lo,
SDValue &Hi) {
SDValue Vec = N->getOperand(0);
EVT VecVT = Vec.getValueType();
EVT EltVT = VecVT.getVectorElementType();
SDValue StackPtr = DAG.CreateStackTemporary(VecVT);
- SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, NULL, 0);
+ SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, NULL, 0,
+ false, false, 0);
// Store the new element. This may be larger than the vector element type,
// so use a truncating store.
SDValue EltPtr = GetVectorElementPointer(StackPtr, EltVT, Idx);
+ const Type *VecType = VecVT.getTypeForEVT(*DAG.getContext());
unsigned Alignment =
- TLI.getTargetData()->getPrefTypeAlignment(VecVT.getTypeForEVT(*DAG.getContext()));
- Store = DAG.getTruncStore(Store, dl, Elt, EltPtr, NULL, 0, EltVT);
+ TLI.getTargetData()->getPrefTypeAlignment(VecType);
+ Store = DAG.getTruncStore(Store, dl, Elt, EltPtr, NULL, 0, EltVT,
+ false, false, 0);
// Load the Lo part from the stack slot.
- Lo = DAG.getLoad(Lo.getValueType(), dl, Store, StackPtr, NULL, 0);
+ Lo = DAG.getLoad(Lo.getValueType(), dl, Store, StackPtr, NULL, 0,
+ false, false, 0);
// Increment the pointer to the other part.
unsigned IncrementSize = Lo.getValueType().getSizeInBits() / 8;
// Load the Hi part from the stack slot.
Hi = DAG.getLoad(Hi.getValueType(), dl, Store, StackPtr, NULL, 0, false,
- MinAlign(Alignment, IncrementSize));
+ false, MinAlign(Alignment, IncrementSize));
}
void DAGTypeLegalizer::SplitVecRes_SCALAR_TO_VECTOR(SDNode *N, SDValue &Lo,
EVT MemoryVT = LD->getMemoryVT();
unsigned Alignment = LD->getOriginalAlignment();
bool isVolatile = LD->isVolatile();
+ bool isNonTemporal = LD->isNonTemporal();
EVT LoMemVT, HiMemVT;
GetSplitDestVTs(MemoryVT, LoMemVT, HiMemVT);
Lo = DAG.getLoad(ISD::UNINDEXED, dl, ExtType, LoVT, Ch, Ptr, Offset,
- SV, SVOffset, LoMemVT, isVolatile, Alignment);
+ SV, SVOffset, LoMemVT, isVolatile, isNonTemporal, Alignment);
unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
DAG.getIntPtrConstant(IncrementSize));
SVOffset += IncrementSize;
Hi = DAG.getLoad(ISD::UNINDEXED, dl, ExtType, HiVT, Ch, Ptr, Offset,
- SV, SVOffset, HiMemVT, isVolatile, Alignment);
+ SV, SVOffset, HiMemVT, isVolatile, isNonTemporal, Alignment);
// Build a factor node to remember that this load is independent of the
// other one.
/// result types of the node are known to be legal, but other operands of the
/// node may need legalization as well as the specified one.
bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) {
- DEBUG(errs() << "Split node operand: ";
+ DEBUG(dbgs() << "Split node operand: ";
N->dump(&DAG);
- errs() << "\n");
+ dbgs() << "\n");
SDValue Res = SDValue();
if (Res.getNode() == 0) {
switch (N->getOpcode()) {
default:
#ifndef NDEBUG
- errs() << "SplitVectorOperand Op #" << OpNo << ": ";
+ dbgs() << "SplitVectorOperand Op #" << OpNo << ": ";
N->dump(&DAG);
- errs() << "\n";
+ dbgs() << "\n";
#endif
llvm_unreachable("Do not know how to split this operator's operand!");
uint64_t LoElts = Lo.getValueType().getVectorNumElements();
if (IdxVal < LoElts)
- return DAG.UpdateNodeOperands(SDValue(N, 0), Lo, Idx);
- return DAG.UpdateNodeOperands(SDValue(N, 0), Hi,
+ return SDValue(DAG.UpdateNodeOperands(N, Lo, Idx), 0);
+ return SDValue(DAG.UpdateNodeOperands(N, Hi,
DAG.getConstant(IdxVal - LoElts,
- Idx.getValueType()));
+ Idx.getValueType())),
+ 0);
}
// Store the vector to the stack.
EVT EltVT = VecVT.getVectorElementType();
DebugLoc dl = N->getDebugLoc();
SDValue StackPtr = DAG.CreateStackTemporary(VecVT);
- const Value *SV = PseudoSourceValue::getStack();
- SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, SV, 0);
+ int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
+ const Value *SV = PseudoSourceValue::getFixedStack(SPFI);
+ SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, SV, 0,
+ false, false, 0);
// Load back the required element.
StackPtr = GetVectorElementPointer(StackPtr, EltVT, Idx);
return DAG.getExtLoad(ISD::EXTLOAD, dl, N->getValueType(0), Store, StackPtr,
- SV, 0, EltVT);
+ SV, 0, EltVT, false, false, 0);
}
SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) {
EVT MemoryVT = N->getMemoryVT();
unsigned Alignment = N->getOriginalAlignment();
bool isVol = N->isVolatile();
+ bool isNT = N->isNonTemporal();
SDValue Lo, Hi;
GetSplitVector(N->getOperand(1), Lo, Hi);
if (isTruncating)
Lo = DAG.getTruncStore(Ch, dl, Lo, Ptr, N->getSrcValue(), SVOffset,
- LoMemVT, isVol, Alignment);
+ LoMemVT, isVol, isNT, Alignment);
else
Lo = DAG.getStore(Ch, dl, Lo, Ptr, N->getSrcValue(), SVOffset,
- isVol, Alignment);
+ isVol, isNT, Alignment);
// Increment the pointer to the other half.
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
if (isTruncating)
Hi = DAG.getTruncStore(Ch, dl, Hi, Ptr, N->getSrcValue(), SVOffset,
- HiMemVT, isVol, Alignment);
+ HiMemVT, isVol, isNT, Alignment);
else
Hi = DAG.getStore(Ch, dl, Hi, Ptr, N->getSrcValue(), SVOffset,
- isVol, Alignment);
+ isVol, isNT, Alignment);
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
}
//===----------------------------------------------------------------------===//
void DAGTypeLegalizer::WidenVectorResult(SDNode *N, unsigned ResNo) {
- DEBUG(errs() << "Widen node result " << ResNo << ": ";
+ DEBUG(dbgs() << "Widen node result " << ResNo << ": ";
N->dump(&DAG);
- errs() << "\n");
- SDValue Res = SDValue();
+ dbgs() << "\n");
+
+ // See if the target wants to custom widen this node.
+ if (CustomWidenLowerNode(N, N->getValueType(ResNo)))
+ return;
+ SDValue Res = SDValue();
switch (N->getOpcode()) {
default:
#ifndef NDEBUG
- errs() << "WidenVectorResult #" << ResNo << ": ";
+ dbgs() << "WidenVectorResult #" << ResNo << ": ";
N->dump(&DAG);
- errs() << "\n";
+ dbgs() << "\n";
#endif
llvm_unreachable("Do not know how to widen the result of this operator!");
case ISD::CONCAT_VECTORS: Res = WidenVecRes_CONCAT_VECTORS(N); break;
case ISD::CONVERT_RNDSAT: Res = WidenVecRes_CONVERT_RNDSAT(N); break;
case ISD::EXTRACT_SUBVECTOR: Res = WidenVecRes_EXTRACT_SUBVECTOR(N); break;
+ case ISD::FP_ROUND_INREG: Res = WidenVecRes_InregOp(N); break;
case ISD::INSERT_VECTOR_ELT: Res = WidenVecRes_INSERT_VECTOR_ELT(N); break;
case ISD::LOAD: Res = WidenVecRes_LOAD(N); break;
case ISD::SCALAR_TO_VECTOR: Res = WidenVecRes_SCALAR_TO_VECTOR(N); break;
+ case ISD::SIGN_EXTEND_INREG: Res = WidenVecRes_InregOp(N); break;
case ISD::SELECT: Res = WidenVecRes_SELECT(N); break;
case ISD::SELECT_CC: Res = WidenVecRes_SELECT_CC(N); break;
+ case ISD::SETCC: Res = WidenVecRes_SETCC(N); break;
case ISD::UNDEF: Res = WidenVecRes_UNDEF(N); break;
case ISD::VECTOR_SHUFFLE:
Res = WidenVecRes_VECTOR_SHUFFLE(cast<ShuffleVectorSDNode>(N));
case ISD::FDIV:
case ISD::FMUL:
case ISD::FPOW:
- case ISD::FPOWI:
case ISD::FREM:
case ISD::FSUB:
case ISD::MUL:
Res = WidenVecRes_Binary(N);
break;
+ case ISD::FPOWI:
+ Res = WidenVecRes_POWI(N);
+ break;
+
case ISD::SHL:
case ISD::SRA:
case ISD::SRL:
case ISD::FNEG:
case ISD::FSIN:
case ISD::FSQRT:
+ case ISD::FEXP:
+ case ISD::FEXP2:
+ case ISD::FLOG:
+ case ISD::FLOG2:
+ case ISD::FLOG10:
Res = WidenVecRes_Unary(N);
break;
}
SDValue DAGTypeLegalizer::WidenVecRes_Binary(SDNode *N) {
// Binary op widening.
+ unsigned Opcode = N->getOpcode();
+ DebugLoc dl = N->getDebugLoc();
EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
- SDValue InOp1 = GetWidenedVector(N->getOperand(0));
- SDValue InOp2 = GetWidenedVector(N->getOperand(1));
- return DAG.getNode(N->getOpcode(), N->getDebugLoc(), WidenVT, InOp1, InOp2);
+ EVT WidenEltVT = WidenVT.getVectorElementType();
+ EVT VT = WidenVT;
+ unsigned NumElts = VT.getVectorNumElements();
+ while (!TLI.isTypeSynthesizable(VT) && NumElts != 1) {
+ NumElts = NumElts / 2;
+ VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts);
+ }
+
+ if (NumElts != 1 && !TLI.canOpTrap(N->getOpcode(), VT)) {
+ // Operation doesn't trap so just widen as normal.
+ SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+ return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2);
+ } else if (NumElts == 1) {
+ // No legal vector version so unroll the vector operation and then widen.
+ return DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements());
+ } else {
+ // Since the operation can trap, apply operation on the original vector.
+ EVT MaxVT = VT;
+ SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+ unsigned CurNumElts = N->getValueType(0).getVectorNumElements();
+
+ SmallVector<SDValue, 16> ConcatOps(CurNumElts);
+ unsigned ConcatEnd = 0; // Current ConcatOps index.
+ int Idx = 0; // Current Idx into input vectors.
+
+ // NumElts := greatest synthesizable vector size (at most WidenVT)
+ // while (orig. vector has unhandled elements) {
+ // take munches of size NumElts from the beginning and add to ConcatOps
+ // NumElts := next smaller supported vector size or 1
+ // }
+ while (CurNumElts != 0) {
+ while (CurNumElts >= NumElts) {
+ SDValue EOp1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp1,
+ DAG.getIntPtrConstant(Idx));
+ SDValue EOp2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp2,
+ DAG.getIntPtrConstant(Idx));
+ ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, VT, EOp1, EOp2);
+ Idx += NumElts;
+ CurNumElts -= NumElts;
+ }
+ do {
+ NumElts = NumElts / 2;
+ VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts);
+ } while (!TLI.isTypeSynthesizable(VT) && NumElts != 1);
+
+ if (NumElts == 1) {
+ for (unsigned i = 0; i != CurNumElts; ++i, ++Idx) {
+ SDValue EOp1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT,
+ InOp1, DAG.getIntPtrConstant(Idx));
+ SDValue EOp2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT,
+ InOp2, DAG.getIntPtrConstant(Idx));
+ ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, WidenEltVT,
+ EOp1, EOp2);
+ }
+ CurNumElts = 0;
+ }
+ }
+
+ // Check to see if we have a single operation with the widen type.
+ if (ConcatEnd == 1) {
+ VT = ConcatOps[0].getValueType();
+ if (VT == WidenVT)
+ return ConcatOps[0];
+ }
+
+ // while (Some element of ConcatOps is not of type MaxVT) {
+ // From the end of ConcatOps, collect elements of the same type and put
+ // them into an op of the next larger supported type
+ // }
+ while (ConcatOps[ConcatEnd-1].getValueType() != MaxVT) {
+ Idx = ConcatEnd - 1;
+ VT = ConcatOps[Idx--].getValueType();
+ while (Idx >= 0 && ConcatOps[Idx].getValueType() == VT)
+ Idx--;
+
+ int NextSize = VT.isVector() ? VT.getVectorNumElements() : 1;
+ EVT NextVT;
+ do {
+ NextSize *= 2;
+ NextVT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NextSize);
+ } while (!TLI.isTypeSynthesizable(NextVT));
+
+ if (!VT.isVector()) {
+ // Scalar type, create an INSERT_VECTOR_ELEMENT of type NextVT
+ SDValue VecOp = DAG.getUNDEF(NextVT);
+ unsigned NumToInsert = ConcatEnd - Idx - 1;
+ for (unsigned i = 0, OpIdx = Idx+1; i < NumToInsert; i++, OpIdx++) {
+ VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NextVT, VecOp,
+ ConcatOps[OpIdx], DAG.getIntPtrConstant(i));
+ }
+ ConcatOps[Idx+1] = VecOp;
+ ConcatEnd = Idx + 2;
+ }
+ else {
+ // Vector type, create a CONCAT_VECTORS of type NextVT
+ SDValue undefVec = DAG.getUNDEF(VT);
+ unsigned OpsToConcat = NextSize/VT.getVectorNumElements();
+ SmallVector<SDValue, 16> SubConcatOps(OpsToConcat);
+ unsigned RealVals = ConcatEnd - Idx - 1;
+ unsigned SubConcatEnd = 0;
+ unsigned SubConcatIdx = Idx + 1;
+ while (SubConcatEnd < RealVals)
+ SubConcatOps[SubConcatEnd++] = ConcatOps[++Idx];
+ while (SubConcatEnd < OpsToConcat)
+ SubConcatOps[SubConcatEnd++] = undefVec;
+ ConcatOps[SubConcatIdx] = DAG.getNode(ISD::CONCAT_VECTORS, dl,
+ NextVT, &SubConcatOps[0],
+ OpsToConcat);
+ ConcatEnd = SubConcatIdx + 1;
+ }
+ }
+
+ // Check to see if we have a single operation with the widen type.
+ if (ConcatEnd == 1) {
+ VT = ConcatOps[0].getValueType();
+ if (VT == WidenVT)
+ return ConcatOps[0];
+ }
+
+ // add undefs of size MaxVT until ConcatOps grows to length of WidenVT
+ unsigned NumOps =
+ WidenVT.getVectorNumElements()/MaxVT.getVectorNumElements();
+ if (NumOps != ConcatEnd ) {
+ SDValue UndefVal = DAG.getUNDEF(MaxVT);
+ for (unsigned j = ConcatEnd; j < NumOps; ++j)
+ ConcatOps[j] = UndefVal;
+ }
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &ConcatOps[0], NumOps);
+ }
}
SDValue DAGTypeLegalizer::WidenVecRes_Convert(SDNode *N) {
return DAG.getNode(Opcode, dl, WidenVT, InOp);
}
- if (TLI.isTypeLegal(InWidenVT)) {
+ if (TLI.isTypeSynthesizable(InWidenVT)) {
// Because the result and the input are different vector types, widening
// the result could create a legal type but widening the input might make
// it an illegal type that might lead to repeatedly splitting the input
return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, &Ops[0], WidenNumElts);
}
+SDValue DAGTypeLegalizer::WidenVecRes_POWI(SDNode *N) {
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ SDValue InOp = GetWidenedVector(N->getOperand(0));
+ SDValue ShOp = N->getOperand(1);
+ return DAG.getNode(N->getOpcode(), N->getDebugLoc(), WidenVT, InOp, ShOp);
+}
+
SDValue DAGTypeLegalizer::WidenVecRes_Shift(SDNode *N) {
EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
SDValue InOp = GetWidenedVector(N->getOperand(0));
ShOp = GetWidenedVector(ShOp);
ShVT = ShOp.getValueType();
}
- EVT ShWidenVT = EVT::getVectorVT(*DAG.getContext(), ShVT.getVectorElementType(),
+ EVT ShWidenVT = EVT::getVectorVT(*DAG.getContext(),
+ ShVT.getVectorElementType(),
WidenVT.getVectorNumElements());
if (ShVT != ShWidenVT)
ShOp = ModifyToType(ShOp, ShWidenVT);
return DAG.getNode(N->getOpcode(), N->getDebugLoc(), WidenVT, InOp);
}
+SDValue DAGTypeLegalizer::WidenVecRes_InregOp(SDNode *N) {
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ EVT ExtVT = EVT::getVectorVT(*DAG.getContext(),
+ cast<VTSDNode>(N->getOperand(1))->getVT()
+ .getVectorElementType(),
+ WidenVT.getVectorNumElements());
+ SDValue WidenLHS = GetWidenedVector(N->getOperand(0));
+ return DAG.getNode(N->getOpcode(), N->getDebugLoc(),
+ WidenVT, WidenLHS, DAG.getValueType(ExtVT));
+}
+
SDValue DAGTypeLegalizer::WidenVecRes_BIT_CONVERT(SDNode *N) {
SDValue InOp = N->getOperand(0);
EVT InVT = InOp.getValueType();
unsigned NewNumElts = WidenSize / InSize;
if (InVT.isVector()) {
EVT InEltVT = InVT.getVectorElementType();
- NewInVT= EVT::getVectorVT(*DAG.getContext(), InEltVT, WidenSize / InEltVT.getSizeInBits());
+ NewInVT= EVT::getVectorVT(*DAG.getContext(), InEltVT,
+ WidenSize / InEltVT.getSizeInBits());
} else {
NewInVT = EVT::getVectorVT(*DAG.getContext(), InVT, NewNumElts);
}
- if (TLI.isTypeLegal(NewInVT)) {
+ if (TLI.isTypeSynthesizable(NewInVT)) {
// Because the result and the input are different vector types, widening
// the result could create a legal type but widening the input might make
// it an illegal type that might lead to repeatedly splitting the input
SDValue RndOp = N->getOperand(3);
SDValue SatOp = N->getOperand(4);
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(),
+ N->getValueType(0));
unsigned WidenNumElts = WidenVT.getVectorNumElements();
EVT InVT = InOp.getValueType();
SatOp, CvtCode);
}
- if (TLI.isTypeLegal(InWidenVT)) {
+ if (TLI.isTypeSynthesizable(InWidenVT)) {
// Because the result and the input are different vector types, widening
// the result could create a legal type but widening the input might make
// it an illegal type that might lead to repeatedly splitting the input
SDValue DAGTypeLegalizer::WidenVecRes_LOAD(SDNode *N) {
LoadSDNode *LD = cast<LoadSDNode>(N);
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), LD->getValueType(0));
- EVT LdVT = LD->getMemoryVT();
- DebugLoc dl = N->getDebugLoc();
- assert(LdVT.isVector() && WidenVT.isVector());
-
- // Load information
- SDValue Chain = LD->getChain();
- SDValue BasePtr = LD->getBasePtr();
- int SVOffset = LD->getSrcValueOffset();
- unsigned Align = LD->getAlignment();
- bool isVolatile = LD->isVolatile();
- const Value *SV = LD->getSrcValue();
ISD::LoadExtType ExtType = LD->getExtensionType();
SDValue Result;
SmallVector<SDValue, 16> LdChain; // Chain for the series of load
- if (ExtType != ISD::NON_EXTLOAD) {
- // For extension loads, we can not play the tricks of chopping legal
- // vector types and bit cast it to the right type. Instead, we unroll
- // the load and build a vector.
- EVT EltVT = WidenVT.getVectorElementType();
- EVT LdEltVT = LdVT.getVectorElementType();
- unsigned NumElts = LdVT.getVectorNumElements();
-
- // Load each element and widen
- unsigned WidenNumElts = WidenVT.getVectorNumElements();
- SmallVector<SDValue, 16> Ops(WidenNumElts);
- unsigned Increment = LdEltVT.getSizeInBits() / 8;
- Ops[0] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, BasePtr, SV, SVOffset,
- LdEltVT, isVolatile, Align);
- LdChain.push_back(Ops[0].getValue(1));
- unsigned i = 0, Offset = Increment;
- for (i=1; i < NumElts; ++i, Offset += Increment) {
- SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),
- BasePtr, DAG.getIntPtrConstant(Offset));
- Ops[i] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, NewBasePtr, SV,
- SVOffset + Offset, LdEltVT, isVolatile, Align);
- LdChain.push_back(Ops[i].getValue(1));
- }
-
- // Fill the rest with undefs
- SDValue UndefVal = DAG.getUNDEF(EltVT);
- for (; i != WidenNumElts; ++i)
- Ops[i] = UndefVal;
-
- Result = DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, &Ops[0], Ops.size());
- } else {
- assert(LdVT.getVectorElementType() == WidenVT.getVectorElementType());
- unsigned int LdWidth = LdVT.getSizeInBits();
- Result = GenWidenVectorLoads(LdChain, Chain, BasePtr, SV, SVOffset,
- Align, isVolatile, LdWidth, WidenVT, dl);
- }
-
- // If we generate a single load, we can use that for the chain. Otherwise,
- // build a factor node to remember the multiple loads are independent and
- // chain to that.
- SDValue NewChain;
- if (LdChain.size() == 1)
- NewChain = LdChain[0];
- else
- NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &LdChain[0],
- LdChain.size());
+ if (ExtType != ISD::NON_EXTLOAD)
+ Result = GenWidenVectorExtLoads(LdChain, LD, ExtType);
+ else
+ Result = GenWidenVectorLoads(LdChain, LD);
+
+ // If we generate a single load, we can use that for the chain. Otherwise,
+ // build a factor node to remember the multiple loads are independent and
+ // chain to that.
+ SDValue NewChain;
+ if (LdChain.size() == 1)
+ NewChain = LdChain[0];
+ else
+ NewChain = DAG.getNode(ISD::TokenFactor, LD->getDebugLoc(), MVT::Other,
+ &LdChain[0], LdChain.size());
// Modified the chain - switch anything that used the old chain to use
// the new one.
EVT CondVT = Cond1.getValueType();
if (CondVT.isVector()) {
EVT CondEltVT = CondVT.getVectorElementType();
- EVT CondWidenVT = EVT::getVectorVT(*DAG.getContext(), CondEltVT, WidenNumElts);
+ EVT CondWidenVT = EVT::getVectorVT(*DAG.getContext(),
+ CondEltVT, WidenNumElts);
if (getTypeAction(CondVT) == WidenVector)
Cond1 = GetWidenedVector(Cond1);
N->getOperand(1), InOp1, InOp2, N->getOperand(4));
}
+SDValue DAGTypeLegalizer::WidenVecRes_SETCC(SDNode *N) {
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+ return DAG.getNode(ISD::SETCC, N->getDebugLoc(), WidenVT,
+ InOp1, InOp2, N->getOperand(2));
+}
+
SDValue DAGTypeLegalizer::WidenVecRes_UNDEF(SDNode *N) {
EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
return DAG.getUNDEF(WidenVT);
SDValue InOp1 = N->getOperand(0);
EVT InVT = InOp1.getValueType();
assert(InVT.isVector() && "can not widen non vector type");
- EVT WidenInVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(), WidenNumElts);
+ EVT WidenInVT = EVT::getVectorVT(*DAG.getContext(),
+ InVT.getVectorElementType(), WidenNumElts);
InOp1 = GetWidenedVector(InOp1);
SDValue InOp2 = GetWidenedVector(N->getOperand(1));
// Widen Vector Operand
//===----------------------------------------------------------------------===//
bool DAGTypeLegalizer::WidenVectorOperand(SDNode *N, unsigned ResNo) {
- DEBUG(errs() << "Widen node operand " << ResNo << ": ";
+ DEBUG(dbgs() << "Widen node operand " << ResNo << ": ";
N->dump(&DAG);
- errs() << "\n");
+ dbgs() << "\n");
SDValue Res = SDValue();
switch (N->getOpcode()) {
default:
#ifndef NDEBUG
- errs() << "WidenVectorOperand op #" << ResNo << ": ";
+ dbgs() << "WidenVectorOperand op #" << ResNo << ": ";
N->dump(&DAG);
- errs() << "\n";
+ dbgs() << "\n";
#endif
llvm_unreachable("Do not know how to widen this operator's operand!");
if (InWidenSize % Size == 0 && !VT.isVector()) {
unsigned NewNumElts = InWidenSize / Size;
EVT NewVT = EVT::getVectorVT(*DAG.getContext(), VT, NewNumElts);
- if (TLI.isTypeLegal(NewVT)) {
+ if (TLI.isTypeSynthesizable(NewVT)) {
SDValue BitOp = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, InOp);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, BitOp,
DAG.getIntPtrConstant(0));
// We have to widen the value but we want only to store the original
// vector type.
StoreSDNode *ST = cast<StoreSDNode>(N);
- SDValue Chain = ST->getChain();
- SDValue BasePtr = ST->getBasePtr();
- const Value *SV = ST->getSrcValue();
- int SVOffset = ST->getSrcValueOffset();
- unsigned Align = ST->getAlignment();
- bool isVolatile = ST->isVolatile();
- SDValue ValOp = GetWidenedVector(ST->getValue());
- DebugLoc dl = N->getDebugLoc();
-
- EVT StVT = ST->getMemoryVT();
- EVT ValVT = ValOp.getValueType();
- // It must be true that we the widen vector type is bigger than where
- // we need to store.
- assert(StVT.isVector() && ValOp.getValueType().isVector());
- assert(StVT.bitsLT(ValOp.getValueType()));
SmallVector<SDValue, 16> StChain;
- if (ST->isTruncatingStore()) {
- // For truncating stores, we can not play the tricks of chopping legal
- // vector types and bit cast it to the right type. Instead, we unroll
- // the store.
- EVT StEltVT = StVT.getVectorElementType();
- EVT ValEltVT = ValVT.getVectorElementType();
- unsigned Increment = ValEltVT.getSizeInBits() / 8;
- unsigned NumElts = StVT.getVectorNumElements();
- SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp,
- DAG.getIntPtrConstant(0));
- StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, BasePtr, SV,
- SVOffset, StEltVT,
- isVolatile, Align));
- unsigned Offset = Increment;
- for (unsigned i=1; i < NumElts; ++i, Offset += Increment) {
- SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),
- BasePtr, DAG.getIntPtrConstant(Offset));
- SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp,
- DAG.getIntPtrConstant(0));
- StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, NewBasePtr, SV,
- SVOffset + Offset, StEltVT,
- isVolatile, MinAlign(Align, Offset)));
- }
- }
- else {
- assert(StVT.getVectorElementType() == ValVT.getVectorElementType());
- // Store value
- GenWidenVectorStores(StChain, Chain, BasePtr, SV, SVOffset,
- Align, isVolatile, ValOp, StVT.getSizeInBits(), dl);
- }
+ if (ST->isTruncatingStore())
+ GenWidenVectorTruncStores(StChain, ST);
+ else
+ GenWidenVectorStores(StChain, ST);
+
if (StChain.size() == 1)
return StChain[0];
else
- return DAG.getNode(ISD::TokenFactor, dl,
+ return DAG.getNode(ISD::TokenFactor, ST->getDebugLoc(),
MVT::Other,&StChain[0],StChain.size());
}
// Vector Widening Utilities
//===----------------------------------------------------------------------===//
+// Utility function to find the type to chop up a widen vector for load/store
+// TLI: Target lowering used to determine legal types.
+// Width: Width left need to load/store.
+// WidenVT: The widen vector type to load to/store from
+// Align: If 0, don't allow use of a wider type
+// WidenEx: If Align is not 0, the amount additional we can load/store from.
+
+static EVT FindMemType(SelectionDAG& DAG, const TargetLowering &TLI,
+ unsigned Width, EVT WidenVT,
+ unsigned Align = 0, unsigned WidenEx = 0) {
+ EVT WidenEltVT = WidenVT.getVectorElementType();
+ unsigned WidenWidth = WidenVT.getSizeInBits();
+ unsigned WidenEltWidth = WidenEltVT.getSizeInBits();
+ unsigned AlignInBits = Align*8;
+
+ // If we have one element to load/store, return it.
+ EVT RetVT = WidenEltVT;
+ if (Width == WidenEltWidth)
+ return RetVT;
+
+ // See if there is larger legal integer than the element type to load/store
+ unsigned VT;
+ for (VT = (unsigned)MVT::LAST_INTEGER_VALUETYPE;
+ VT >= (unsigned)MVT::FIRST_INTEGER_VALUETYPE; --VT) {
+ EVT MemVT((MVT::SimpleValueType) VT);
+ unsigned MemVTWidth = MemVT.getSizeInBits();
+ if (MemVT.getSizeInBits() <= WidenEltWidth)
+ break;
+ if (TLI.isTypeSynthesizable(MemVT) && (WidenWidth % MemVTWidth) == 0 &&
+ (MemVTWidth <= Width ||
+ (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) {
+ RetVT = MemVT;
+ break;
+ }
+ }
+
+ // See if there is a larger vector type to load/store that has the same vector
+ // element type and is evenly divisible with the WidenVT.
+ for (VT = (unsigned)MVT::LAST_VECTOR_VALUETYPE;
+ VT >= (unsigned)MVT::FIRST_VECTOR_VALUETYPE; --VT) {
+ EVT MemVT = (MVT::SimpleValueType) VT;
+ unsigned MemVTWidth = MemVT.getSizeInBits();
+ if (TLI.isTypeSynthesizable(MemVT) && WidenEltVT == MemVT.getVectorElementType() &&
+ (WidenWidth % MemVTWidth) == 0 &&
+ (MemVTWidth <= Width ||
+ (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) {
+ if (RetVT.getSizeInBits() < MemVTWidth || MemVT == WidenVT)
+ return MemVT;
+ }
+ }
-// Utility function to find a vector type and its associated element
-// type from a preferred width and whose vector type must be the same size
-// as the VecVT.
-// TLI: Target lowering used to determine legal types.
-// Width: Preferred width to store.
-// VecVT: Vector value type whose size we must match.
-// Returns NewVecVT and NewEltVT - the vector type and its associated
-// element type.
-static void FindAssocWidenVecType(SelectionDAG& DAG,
- const TargetLowering &TLI, unsigned Width,
- EVT VecVT,
- EVT& NewEltVT, EVT& NewVecVT) {
- unsigned EltWidth = Width + 1;
- if (TLI.isTypeLegal(VecVT)) {
- // We start with the preferred with, making it a power of 2 and find a
- // legal vector type of that width. If not, we reduce it by another of 2.
- // For incoming type is legal, this process will end as a vector of the
- // smallest loadable type should always be legal.
- do {
- assert(EltWidth > 0);
- EltWidth = 1 << Log2_32(EltWidth - 1);
- NewEltVT = EVT::getIntegerVT(*DAG.getContext(), EltWidth);
- unsigned NumElts = VecVT.getSizeInBits() / EltWidth;
- NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewEltVT, NumElts);
- } while (!TLI.isTypeLegal(NewVecVT) ||
- VecVT.getSizeInBits() != NewVecVT.getSizeInBits());
- } else {
- // The incoming vector type is illegal and is the result of widening
- // a vector to a power of 2. In this case, we will use the preferred
- // with as long as it is a multiple of the incoming vector length.
- // The legalization process will eventually make this into a legal type
- // and remove the illegal bit converts (which would turn to stack converts
- // if they are allow to exist).
- do {
- assert(EltWidth > 0);
- EltWidth = 1 << Log2_32(EltWidth - 1);
- NewEltVT = EVT::getIntegerVT(*DAG.getContext(), EltWidth);
- unsigned NumElts = VecVT.getSizeInBits() / EltWidth;
- NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewEltVT, NumElts);
- } while (!TLI.isTypeLegal(NewEltVT) ||
- VecVT.getSizeInBits() != NewVecVT.getSizeInBits());
+ return RetVT;
+}
+
+// Builds a vector type from scalar loads
+// VecTy: Resulting Vector type
+// LDOps: Load operators to build a vector type
+// [Start,End) the list of loads to use.
+static SDValue BuildVectorFromScalar(SelectionDAG& DAG, EVT VecTy,
+ SmallVector<SDValue, 16>& LdOps,
+ unsigned Start, unsigned End) {
+ DebugLoc dl = LdOps[Start].getDebugLoc();
+ EVT LdTy = LdOps[Start].getValueType();
+ unsigned Width = VecTy.getSizeInBits();
+ unsigned NumElts = Width / LdTy.getSizeInBits();
+ EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), LdTy, NumElts);
+
+ unsigned Idx = 1;
+ SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT,LdOps[Start]);
+
+ for (unsigned i = Start + 1; i != End; ++i) {
+ EVT NewLdTy = LdOps[i].getValueType();
+ if (NewLdTy != LdTy) {
+ NumElts = Width / NewLdTy.getSizeInBits();
+ NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewLdTy, NumElts);
+ VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, NewVecVT, VecOp);
+ // Readjust position and vector position based on new load type
+ Idx = Idx * LdTy.getSizeInBits() / NewLdTy.getSizeInBits();
+ LdTy = NewLdTy;
+ }
+ VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NewVecVT, VecOp, LdOps[i],
+ DAG.getIntPtrConstant(Idx++));
}
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VecTy, VecOp);
}
SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVector<SDValue, 16>& LdChain,
- SDValue Chain,
- SDValue BasePtr,
- const Value *SV,
- int SVOffset,
- unsigned Alignment,
- bool isVolatile,
- unsigned LdWidth,
- EVT ResType,
- DebugLoc dl) {
+ LoadSDNode * LD) {
// The strategy assumes that we can efficiently load powers of two widths.
- // The routines chops the vector into the largest power of 2 load and
- // can be inserted into a legal vector and then cast the result into the
- // vector type we want. This avoids unnecessary stack converts.
+ // The routines chops the vector into the largest vector loads with the same
+ // element type or scalar loads and then recombines it to the widen vector
+ // type.
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(),LD->getValueType(0));
+ unsigned WidenWidth = WidenVT.getSizeInBits();
+ EVT LdVT = LD->getMemoryVT();
+ DebugLoc dl = LD->getDebugLoc();
+ assert(LdVT.isVector() && WidenVT.isVector());
+ assert(LdVT.getVectorElementType() == WidenVT.getVectorElementType());
- // TODO: If the Ldwidth is legal, alignment is the same as the LdWidth, and
- // the load is nonvolatile, we an use a wider load for the value.
+ // Load information
+ SDValue Chain = LD->getChain();
+ SDValue BasePtr = LD->getBasePtr();
+ int SVOffset = LD->getSrcValueOffset();
+ unsigned Align = LD->getAlignment();
+ bool isVolatile = LD->isVolatile();
+ bool isNonTemporal = LD->isNonTemporal();
+ const Value *SV = LD->getSrcValue();
+
+ int LdWidth = LdVT.getSizeInBits();
+ int WidthDiff = WidenWidth - LdWidth; // Difference
+ unsigned LdAlign = (isVolatile) ? 0 : Align; // Allow wider loads
// Find the vector type that can load from.
- EVT NewEltVT, NewVecVT;
- unsigned NewEltVTWidth;
- FindAssocWidenVecType(DAG, TLI, LdWidth, ResType, NewEltVT, NewVecVT);
- NewEltVTWidth = NewEltVT.getSizeInBits();
-
- SDValue LdOp = DAG.getLoad(NewEltVT, dl, Chain, BasePtr, SV, SVOffset,
- isVolatile, Alignment);
- SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT, LdOp);
+ EVT NewVT = FindMemType(DAG, TLI, LdWidth, WidenVT, LdAlign, WidthDiff);
+ int NewVTWidth = NewVT.getSizeInBits();
+ SDValue LdOp = DAG.getLoad(NewVT, dl, Chain, BasePtr, SV, SVOffset,
+ isVolatile, isNonTemporal, Align);
LdChain.push_back(LdOp.getValue(1));
// Check if we can load the element with one instruction
- if (LdWidth == NewEltVTWidth) {
- return DAG.getNode(ISD::BIT_CONVERT, dl, ResType, VecOp);
+ if (LdWidth <= NewVTWidth) {
+ if (NewVT.isVector()) {
+ if (NewVT != WidenVT) {
+ assert(WidenWidth % NewVTWidth == 0);
+ unsigned NumConcat = WidenWidth / NewVTWidth;
+ SmallVector<SDValue, 16> ConcatOps(NumConcat);
+ SDValue UndefVal = DAG.getUNDEF(NewVT);
+ ConcatOps[0] = LdOp;
+ for (unsigned i = 1; i != NumConcat; ++i)
+ ConcatOps[i] = UndefVal;
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &ConcatOps[0],
+ NumConcat);
+ } else
+ return LdOp;
+ } else {
+ unsigned NumElts = WidenWidth / NewVTWidth;
+ EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewVT, NumElts);
+ SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT, LdOp);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, WidenVT, VecOp);
+ }
}
- unsigned Idx = 1;
- LdWidth -= NewEltVTWidth;
+ // Load vector by using multiple loads from largest vector to scalar
+ SmallVector<SDValue, 16> LdOps;
+ LdOps.push_back(LdOp);
+
+ LdWidth -= NewVTWidth;
unsigned Offset = 0;
while (LdWidth > 0) {
- unsigned Increment = NewEltVTWidth / 8;
+ unsigned Increment = NewVTWidth / 8;
Offset += Increment;
BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
DAG.getIntPtrConstant(Increment));
- if (LdWidth < NewEltVTWidth) {
- // Our current type we are using is too large, use a smaller size by
- // using a smaller power of 2
- unsigned oNewEltVTWidth = NewEltVTWidth;
- FindAssocWidenVecType(DAG, TLI, LdWidth, ResType, NewEltVT, NewVecVT);
- NewEltVTWidth = NewEltVT.getSizeInBits();
- // Readjust position and vector position based on new load type
- Idx = Idx * (oNewEltVTWidth/NewEltVTWidth);
- VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, NewVecVT, VecOp);
+ if (LdWidth < NewVTWidth) {
+ // Our current type we are using is too large, find a better size
+ NewVT = FindMemType(DAG, TLI, LdWidth, WidenVT, LdAlign, WidthDiff);
+ NewVTWidth = NewVT.getSizeInBits();
}
- SDValue LdOp = DAG.getLoad(NewEltVT, dl, Chain, BasePtr, SV,
- SVOffset+Offset, isVolatile,
- MinAlign(Alignment, Offset));
+ SDValue LdOp = DAG.getLoad(NewVT, dl, Chain, BasePtr, SV,
+ SVOffset+Offset, isVolatile,
+ isNonTemporal, MinAlign(Align, Increment));
LdChain.push_back(LdOp.getValue(1));
- VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NewVecVT, VecOp, LdOp,
- DAG.getIntPtrConstant(Idx++));
+ LdOps.push_back(LdOp);
- LdWidth -= NewEltVTWidth;
+ LdWidth -= NewVTWidth;
}
- return DAG.getNode(ISD::BIT_CONVERT, dl, ResType, VecOp);
-}
+ // Build the vector from the loads operations
+ unsigned End = LdOps.size();
+ if (LdOps[0].getValueType().isVector()) {
+ // If the load contains vectors, build the vector using concat vector.
+ // All of the vectors used to loads are power of 2 and the scalars load
+ // can be combined to make a power of 2 vector.
+ SmallVector<SDValue, 16> ConcatOps(End);
+ int i = End - 1;
+ int Idx = End;
+ EVT LdTy = LdOps[i].getValueType();
+ // First combine the scalar loads to a vector
+ if (!LdTy.isVector()) {
+ for (--i; i >= 0; --i) {
+ LdTy = LdOps[i].getValueType();
+ if (LdTy.isVector())
+ break;
+ }
+ ConcatOps[--Idx] = BuildVectorFromScalar(DAG, LdTy, LdOps, i+1, End);
+ }
+ ConcatOps[--Idx] = LdOps[i];
+ for (--i; i >= 0; --i) {
+ EVT NewLdTy = LdOps[i].getValueType();
+ if (NewLdTy != LdTy) {
+ // Create a larger vector
+ ConcatOps[End-1] = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewLdTy,
+ &ConcatOps[Idx], End - Idx);
+ Idx = End - 1;
+ LdTy = NewLdTy;
+ }
+ ConcatOps[--Idx] = LdOps[i];
+ }
-void DAGTypeLegalizer::GenWidenVectorStores(SmallVector<SDValue, 16>& StChain,
- SDValue Chain,
- SDValue BasePtr,
- const Value *SV,
- int SVOffset,
- unsigned Alignment,
- bool isVolatile,
- SDValue ValOp,
- unsigned StWidth,
- DebugLoc dl) {
- // Breaks the stores into a series of power of 2 width stores. For any
- // width, we convert the vector to the vector of element size that we
- // want to store. This avoids requiring a stack convert.
-
- // Find a width of the element type we can store with
- EVT WidenVT = ValOp.getValueType();
- EVT NewEltVT, NewVecVT;
-
- FindAssocWidenVecType(DAG, TLI, StWidth, WidenVT, NewEltVT, NewVecVT);
- unsigned NewEltVTWidth = NewEltVT.getSizeInBits();
-
- SDValue VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, NewVecVT, ValOp);
- SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NewEltVT, VecOp,
- DAG.getIntPtrConstant(0));
- SDValue StOp = DAG.getStore(Chain, dl, EOp, BasePtr, SV, SVOffset,
- isVolatile, Alignment);
- StChain.push_back(StOp);
+ if (WidenWidth != LdTy.getSizeInBits()*(End - Idx)) {
+ // We need to fill the rest with undefs to build the vector
+ unsigned NumOps = WidenWidth / LdTy.getSizeInBits();
+ SmallVector<SDValue, 16> WidenOps(NumOps);
+ SDValue UndefVal = DAG.getUNDEF(LdTy);
+ unsigned i = 0;
+ for (; i != End-Idx; ++i)
+ WidenOps[i] = ConcatOps[Idx+i];
+ for (; i != NumOps; ++i)
+ WidenOps[i] = UndefVal;
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &WidenOps[0],NumOps);
+ } else
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT,
+ &ConcatOps[Idx], End - Idx);
+ } else // All the loads are scalar loads.
+ return BuildVectorFromScalar(DAG, WidenVT, LdOps, 0, End);
+}
+
+SDValue
+DAGTypeLegalizer::GenWidenVectorExtLoads(SmallVector<SDValue, 16>& LdChain,
+ LoadSDNode * LD,
+ ISD::LoadExtType ExtType) {
+ // For extension loads, it may not be more efficient to chop up the vector
+ // and then extended it. Instead, we unroll the load and build a new vector.
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(),LD->getValueType(0));
+ EVT LdVT = LD->getMemoryVT();
+ DebugLoc dl = LD->getDebugLoc();
+ assert(LdVT.isVector() && WidenVT.isVector());
- // Check if we are done
- if (StWidth == NewEltVTWidth) {
- return;
+ // Load information
+ SDValue Chain = LD->getChain();
+ SDValue BasePtr = LD->getBasePtr();
+ int SVOffset = LD->getSrcValueOffset();
+ unsigned Align = LD->getAlignment();
+ bool isVolatile = LD->isVolatile();
+ bool isNonTemporal = LD->isNonTemporal();
+ const Value *SV = LD->getSrcValue();
+
+ EVT EltVT = WidenVT.getVectorElementType();
+ EVT LdEltVT = LdVT.getVectorElementType();
+ unsigned NumElts = LdVT.getVectorNumElements();
+
+ // Load each element and widen
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+ SmallVector<SDValue, 16> Ops(WidenNumElts);
+ unsigned Increment = LdEltVT.getSizeInBits() / 8;
+ Ops[0] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, BasePtr, SV, SVOffset,
+ LdEltVT, isVolatile, isNonTemporal, Align);
+ LdChain.push_back(Ops[0].getValue(1));
+ unsigned i = 0, Offset = Increment;
+ for (i=1; i < NumElts; ++i, Offset += Increment) {
+ SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),
+ BasePtr, DAG.getIntPtrConstant(Offset));
+ Ops[i] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, NewBasePtr, SV,
+ SVOffset + Offset, LdEltVT, isVolatile,
+ isNonTemporal, Align);
+ LdChain.push_back(Ops[i].getValue(1));
}
- unsigned Idx = 1;
- StWidth -= NewEltVTWidth;
- unsigned Offset = 0;
+ // Fill the rest with undefs
+ SDValue UndefVal = DAG.getUNDEF(EltVT);
+ for (; i != WidenNumElts; ++i)
+ Ops[i] = UndefVal;
- while (StWidth > 0) {
- unsigned Increment = NewEltVTWidth / 8;
- Offset += Increment;
- BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
- DAG.getIntPtrConstant(Increment));
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, &Ops[0], Ops.size());
+}
- if (StWidth < NewEltVTWidth) {
- // Our current type we are using is too large, use a smaller size by
- // using a smaller power of 2
- unsigned oNewEltVTWidth = NewEltVTWidth;
- FindAssocWidenVecType(DAG, TLI, StWidth, WidenVT, NewEltVT, NewVecVT);
- NewEltVTWidth = NewEltVT.getSizeInBits();
- // Readjust position and vector position based on new load type
- Idx = Idx * (oNewEltVTWidth/NewEltVTWidth);
- VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, NewVecVT, VecOp);
- }
- EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NewEltVT, VecOp,
+void DAGTypeLegalizer::GenWidenVectorStores(SmallVector<SDValue, 16>& StChain,
+ StoreSDNode *ST) {
+ // The strategy assumes that we can efficiently store powers of two widths.
+ // The routines chops the vector into the largest vector stores with the same
+ // element type or scalar stores.
+ SDValue Chain = ST->getChain();
+ SDValue BasePtr = ST->getBasePtr();
+ const Value *SV = ST->getSrcValue();
+ int SVOffset = ST->getSrcValueOffset();
+ unsigned Align = ST->getAlignment();
+ bool isVolatile = ST->isVolatile();
+ bool isNonTemporal = ST->isNonTemporal();
+ SDValue ValOp = GetWidenedVector(ST->getValue());
+ DebugLoc dl = ST->getDebugLoc();
+
+ EVT StVT = ST->getMemoryVT();
+ unsigned StWidth = StVT.getSizeInBits();
+ EVT ValVT = ValOp.getValueType();
+ unsigned ValWidth = ValVT.getSizeInBits();
+ EVT ValEltVT = ValVT.getVectorElementType();
+ unsigned ValEltWidth = ValEltVT.getSizeInBits();
+ assert(StVT.getVectorElementType() == ValEltVT);
+
+ int Idx = 0; // current index to store
+ unsigned Offset = 0; // offset from base to store
+ while (StWidth != 0) {
+ // Find the largest vector type we can store with
+ EVT NewVT = FindMemType(DAG, TLI, StWidth, ValVT);
+ unsigned NewVTWidth = NewVT.getSizeInBits();
+ unsigned Increment = NewVTWidth / 8;
+ if (NewVT.isVector()) {
+ unsigned NumVTElts = NewVT.getVectorNumElements();
+ do {
+ SDValue EOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NewVT, ValOp,
+ DAG.getIntPtrConstant(Idx));
+ StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr, SV,
+ SVOffset + Offset, isVolatile,
+ isNonTemporal,
+ MinAlign(Align, Offset)));
+ StWidth -= NewVTWidth;
+ Offset += Increment;
+ Idx += NumVTElts;
+ BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
+ DAG.getIntPtrConstant(Increment));
+ } while (StWidth != 0 && StWidth >= NewVTWidth);
+ } else {
+ // Cast the vector to the scalar type we can store
+ unsigned NumElts = ValWidth / NewVTWidth;
+ EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewVT, NumElts);
+ SDValue VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, NewVecVT, ValOp);
+ // Readjust index position based on new vector type
+ Idx = Idx * ValEltWidth / NewVTWidth;
+ do {
+ SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NewVT, VecOp,
DAG.getIntPtrConstant(Idx++));
- StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr, SV,
- SVOffset + Offset, isVolatile,
- MinAlign(Alignment, Offset)));
- StWidth -= NewEltVTWidth;
+ StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr, SV,
+ SVOffset + Offset, isVolatile,
+ isNonTemporal, MinAlign(Align, Offset)));
+ StWidth -= NewVTWidth;
+ Offset += Increment;
+ BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
+ DAG.getIntPtrConstant(Increment));
+ } while (StWidth != 0 && StWidth >= NewVTWidth);
+ // Restore index back to be relative to the original widen element type
+ Idx = Idx * NewVTWidth / ValEltWidth;
+ }
+ }
+}
+
+void
+DAGTypeLegalizer::GenWidenVectorTruncStores(SmallVector<SDValue, 16>& StChain,
+ StoreSDNode *ST) {
+ // For extension loads, it may not be more efficient to truncate the vector
+ // and then store it. Instead, we extract each element and then store it.
+ SDValue Chain = ST->getChain();
+ SDValue BasePtr = ST->getBasePtr();
+ const Value *SV = ST->getSrcValue();
+ int SVOffset = ST->getSrcValueOffset();
+ unsigned Align = ST->getAlignment();
+ bool isVolatile = ST->isVolatile();
+ bool isNonTemporal = ST->isNonTemporal();
+ SDValue ValOp = GetWidenedVector(ST->getValue());
+ DebugLoc dl = ST->getDebugLoc();
+
+ EVT StVT = ST->getMemoryVT();
+ EVT ValVT = ValOp.getValueType();
+
+ // It must be true that we the widen vector type is bigger than where
+ // we need to store.
+ assert(StVT.isVector() && ValOp.getValueType().isVector());
+ assert(StVT.bitsLT(ValOp.getValueType()));
+
+ // For truncating stores, we can not play the tricks of chopping legal
+ // vector types and bit cast it to the right type. Instead, we unroll
+ // the store.
+ EVT StEltVT = StVT.getVectorElementType();
+ EVT ValEltVT = ValVT.getVectorElementType();
+ unsigned Increment = ValEltVT.getSizeInBits() / 8;
+ unsigned NumElts = StVT.getVectorNumElements();
+ SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp,
+ DAG.getIntPtrConstant(0));
+ StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, BasePtr, SV,
+ SVOffset, StEltVT,
+ isVolatile, isNonTemporal, Align));
+ unsigned Offset = Increment;
+ for (unsigned i=1; i < NumElts; ++i, Offset += Increment) {
+ SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),
+ BasePtr, DAG.getIntPtrConstant(Offset));
+ SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp,
+ DAG.getIntPtrConstant(0));
+ StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, NewBasePtr, SV,
+ SVOffset + Offset, StEltVT,
+ isVolatile, isNonTemporal,
+ MinAlign(Align, Offset)));
}
}
projects / oota-llvm.git / blobdiff