setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
+ setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
// Neon does not support some operations on v1i64 and v2i64 types.
setOperationAction(ISD::MUL, MVT::v1i64, Expand);
}
}
+ // ARM and Thumb2 support UMLAL/SMLAL.
+ if (!Subtarget->isThumb1Only())
+ setTargetDAGCombine(ISD::ADDC);
+
+
computeRegisterProperties();
// ARM does not have f32 extending load.
if (!Subtarget->hasV6Ops())
setOperationAction(ISD::BSWAP, MVT::i32, Expand);
- // These are expanded into libcalls.
- if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
- // v7M has a hardware divider
+ if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
+ !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
+ // These are expanded into libcalls if the cpu doesn't have HW divider.
setOperationAction(ISD::SDIV, MVT::i32, Expand);
setOperationAction(ISD::UDIV, MVT::i32, Expand);
}
setTargetDAGCombine(ISD::ADD);
setTargetDAGCombine(ISD::SUB);
setTargetDAGCombine(ISD::MUL);
-
- if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) {
- setTargetDAGCombine(ISD::AND);
- setTargetDAGCombine(ISD::OR);
- setTargetDAGCombine(ISD::XOR);
- }
+ setTargetDAGCombine(ISD::AND);
+ setTargetDAGCombine(ISD::OR);
+ setTargetDAGCombine(ISD::XOR);
if (Subtarget->hasV6Ops())
setTargetDAGCombine(ISD::SRL);
benefitFromCodePlacementOpt = true;
// Prefer likely predicted branches to selects on out-of-order cores.
- predictableSelectIsExpensive = Subtarget->isCortexA9();
+ predictableSelectIsExpensive = Subtarget->isLikeA9();
setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
}
case ARMISD::VTBL2: return "ARMISD::VTBL2";
case ARMISD::VMULLs: return "ARMISD::VMULLs";
case ARMISD::VMULLu: return "ARMISD::VMULLu";
+ case ARMISD::UMLAL: return "ARMISD::UMLAL";
+ case ARMISD::SMLAL: return "ARMISD::SMLAL";
case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
case ARMISD::FMAX: return "ARMISD::FMAX";
case ARMISD::FMIN: return "ARMISD::FMIN";
void
ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
DebugLoc dl, SDValue &Chain,
+ const Value *OrigArg,
+ unsigned OffsetFromOrigArg,
unsigned ArgOffset) const {
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo *MFI = MF.getFrameInfo();
getPointerTy());
SmallVector<SDValue, 4> MemOps;
- for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
+ for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
const TargetRegisterClass *RC;
if (AFI->isThumb1OnlyFunction())
RC = &ARM::tGPRRegClass;
SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
SDValue Store =
DAG.getStore(Val.getValue(1), dl, Val, FIN,
- MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
+ MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
false, false, 0);
MemOps.push_back(Store);
FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
CCInfo.AnalyzeFormalArguments(Ins,
CCAssignFnForNode(CallConv, /* Return*/ false,
isVarArg));
-
+
SmallVector<SDValue, 16> ArgValues;
int lastInsIndex = -1;
-
SDValue ArgValue;
+ Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
+ unsigned CurArgIdx = 0;
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
-
+ std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
+ CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
// Arguments stored in registers.
if (VA.isRegLoc()) {
EVT RegVT = VA.getLocVT();
if (Flags.isByVal()) {
unsigned VARegSize, VARegSaveSize;
computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
- VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
+ VarArgStyleRegisters(CCInfo, DAG,
+ dl, Chain, CurOrigArg, Ins[VA.getValNo()].PartOffset, 0);
unsigned Bytes = Flags.getByValSize() - VARegSize;
if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
int FI = MFI->CreateFixedObject(Bytes,
// varargs
if (isVarArg)
- VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
+ VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
+ CCInfo.getNextStackOffset());
return Chain;
}
}
// Scan through the operands to see if only one value is used.
+ //
+ // As an optimisation, even if more than one value is used it may be more
+ // profitable to splat with one value then change some lanes.
+ //
+ // Heuristically we decide to do this if the vector has a "dominant" value,
+ // defined as splatted to more than half of the lanes.
unsigned NumElts = VT.getVectorNumElements();
bool isOnlyLowElement = true;
bool usesOnlyOneValue = true;
+ bool hasDominantValue = false;
bool isConstant = true;
+
+ // Map of the number of times a particular SDValue appears in the
+ // element list.
+ DenseMap<SDValue, unsigned> ValueCounts;
SDValue Value;
for (unsigned i = 0; i < NumElts; ++i) {
SDValue V = Op.getOperand(i);
if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
isConstant = false;
- if (!Value.getNode())
+ ValueCounts.insert(std::make_pair(V, 0));
+ unsigned &Count = ValueCounts[V];
+
+ // Is this value dominant? (takes up more than half of the lanes)
+ if (++Count > (NumElts / 2)) {
+ hasDominantValue = true;
Value = V;
- else if (V != Value)
- usesOnlyOneValue = false;
+ }
}
+ if (ValueCounts.size() != 1)
+ usesOnlyOneValue = false;
+ if (!Value.getNode() && ValueCounts.size() > 0)
+ Value = ValueCounts.begin()->first;
- if (!Value.getNode())
+ if (ValueCounts.size() == 0)
return DAG.getUNDEF(VT);
if (isOnlyLowElement)
// Use VDUP for non-constant splats. For f32 constant splats, reduce to
// i32 and try again.
- if (usesOnlyOneValue && EltSize <= 32) {
- if (!isConstant)
- return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
+ if (hasDominantValue && EltSize <= 32) {
+ if (!isConstant) {
+ SDValue N;
+
+ // If we are VDUPing a value that comes directly from a vector, that will
+ // cause an unnecessary move to and from a GPR, where instead we could
+ // just use VDUPLANE.
+ if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT)
+ N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
+ Value->getOperand(0), Value->getOperand(1));
+ else
+ N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
+
+ if (!usesOnlyOneValue) {
+ // The dominant value was splatted as 'N', but we now have to insert
+ // all differing elements.
+ for (unsigned I = 0; I < NumElts; ++I) {
+ if (Op.getOperand(I) == Value)
+ continue;
+ SmallVector<SDValue, 3> Ops;
+ Ops.push_back(N);
+ Ops.push_back(Op.getOperand(I));
+ Ops.push_back(DAG.getConstant(I, MVT::i32));
+ N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
+ }
+ }
+ return N;
+ }
if (VT.getVectorElementType().isFloatingPoint()) {
SmallVector<SDValue, 8> Ops;
for (unsigned i = 0; i < NumElts; ++i)
if (Val.getNode())
return DAG.getNode(ISD::BITCAST, dl, VT, Val);
}
- SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
- if (Val.getNode())
- return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
+ if (usesOnlyOneValue) {
+ SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
+ if (isConstant && Val.getNode())
+ return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
+ }
}
// If all elements are constants and the case above didn't get hit, fall back
exitMBB->transferSuccessorsAndUpdatePHIs(BB);
const TargetRegisterClass *TRC = isThumb2 ?
- (const TargetRegisterClass*)&ARM::tGPRRegClass :
+ (const TargetRegisterClass*)&ARM::rGPRRegClass :
(const TargetRegisterClass*)&ARM::GPRRegClass;
unsigned scratch = MRI.createVirtualRegister(TRC);
unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
exitMBB->transferSuccessorsAndUpdatePHIs(BB);
const TargetRegisterClass *TRC = isThumb2 ?
- (const TargetRegisterClass*)&ARM::tGPRRegClass :
+ (const TargetRegisterClass*)&ARM::rGPRRegClass :
(const TargetRegisterClass*)&ARM::GPRRegClass;
unsigned scratch = MRI.createVirtualRegister(TRC);
unsigned scratch2 = MRI.createVirtualRegister(TRC);
// ldrex dest, ptr
// (sign extend dest, if required)
// cmp dest, incr
- // cmov.cond scratch2, dest, incr
+ // cmov.cond scratch2, incr, dest
// strex scratch, scratch2, ptr
// cmp scratch, #0
// bne- loopMBB
AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
.addReg(oldval).addReg(incr));
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
- .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
+ .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
if (strOpc == ARM::t2STREX)
const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
// MachineConstantPool wants an explicit alignment.
- unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
+ unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
if (Align == 0)
- Align = getTargetData()->getTypeAllocSize(C->getType());
+ Align = getDataLayout()->getTypeAllocSize(C->getType());
unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
unsigned VReg1 = MRI->createVirtualRegister(TRC);
const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
// MachineConstantPool wants an explicit alignment.
- unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
+ unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
if (Align == 0)
- Align = getTargetData()->getTypeAllocSize(C->getType());
+ Align = getDataLayout()->getTypeAllocSize(C->getType());
unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
unsigned VReg1 = MRI->createVirtualRegister(TRC);
UnitSize = 2;
} else {
// Check whether we can use NEON instructions.
- if (!MF->getFunction()->hasFnAttr(Attribute::NoImplicitFloat) &&
+ if (!MF->getFunction()->getFnAttributes().
+ hasAttribute(Attributes::NoImplicitFloat) &&
Subtarget->hasNEON()) {
if ((Align % 16 == 0) && SizeVal >= 16) {
ldrOpc = ARM::VLD1q32wb_fixed;
} else {
AddDefaultPred(BuildMI(*BB, MI, dl,
TII->get(ldrOpc),scratch)
- .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
+ .addReg(srcOut, RegState::Define).addReg(srcIn)
+ .addReg(0).addImm(1));
AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
.addReg(scratch).addReg(destIn)
const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
// MachineConstantPool wants an explicit alignment.
- unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
+ unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
if (Align == 0)
- Align = getTargetData()->getTypeAllocSize(C->getType());
+ Align = getDataLayout()->getTypeAllocSize(C->getType());
unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
}
+static SDValue findMUL_LOHI(SDValue V) {
+ if (V->getOpcode() == ISD::UMUL_LOHI ||
+ V->getOpcode() == ISD::SMUL_LOHI)
+ return V;
+ return SDValue();
+}
+
+static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const ARMSubtarget *Subtarget) {
+
+ if (Subtarget->isThumb1Only()) return SDValue();
+
+ // Only perform the checks after legalize when the pattern is available.
+ if (DCI.isBeforeLegalize()) return SDValue();
+
+ // Look for multiply add opportunities.
+ // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
+ // each add nodes consumes a value from ISD::UMUL_LOHI and there is
+ // a glue link from the first add to the second add.
+ // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
+ // a S/UMLAL instruction.
+ // loAdd UMUL_LOHI
+ // \ / :lo \ :hi
+ // \ / \ [no multiline comment]
+ // ADDC | hiAdd
+ // \ :glue / /
+ // \ / /
+ // ADDE
+ //
+ assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
+ SDValue AddcOp0 = AddcNode->getOperand(0);
+ SDValue AddcOp1 = AddcNode->getOperand(1);
+
+ // Check if the two operands are from the same mul_lohi node.
+ if (AddcOp0.getNode() == AddcOp1.getNode())
+ return SDValue();
+
+ assert(AddcNode->getNumValues() == 2 &&
+ AddcNode->getValueType(0) == MVT::i32 &&
+ AddcNode->getValueType(1) == MVT::Glue &&
+ "Expect ADDC with two result values: i32, glue");
+
+ // Check that the ADDC adds the low result of the S/UMUL_LOHI.
+ if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
+ AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
+ AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
+ AddcOp1->getOpcode() != ISD::SMUL_LOHI)
+ return SDValue();
+
+ // Look for the glued ADDE.
+ SDNode* AddeNode = AddcNode->getGluedUser();
+ if (AddeNode == NULL)
+ return SDValue();
+
+ // Make sure it is really an ADDE.
+ if (AddeNode->getOpcode() != ISD::ADDE)
+ return SDValue();
+
+ assert(AddeNode->getNumOperands() == 3 &&
+ AddeNode->getOperand(2).getValueType() == MVT::Glue &&
+ "ADDE node has the wrong inputs");
+
+ // Check for the triangle shape.
+ SDValue AddeOp0 = AddeNode->getOperand(0);
+ SDValue AddeOp1 = AddeNode->getOperand(1);
+
+ // Make sure that the ADDE operands are not coming from the same node.
+ if (AddeOp0.getNode() == AddeOp1.getNode())
+ return SDValue();
+
+ // Find the MUL_LOHI node walking up ADDE's operands.
+ bool IsLeftOperandMUL = false;
+ SDValue MULOp = findMUL_LOHI(AddeOp0);
+ if (MULOp == SDValue())
+ MULOp = findMUL_LOHI(AddeOp1);
+ else
+ IsLeftOperandMUL = true;
+ if (MULOp == SDValue())
+ return SDValue();
+
+ // Figure out the right opcode.
+ unsigned Opc = MULOp->getOpcode();
+ unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
+
+ // Figure out the high and low input values to the MLAL node.
+ SDValue* HiMul = &MULOp;
+ SDValue* HiAdd = NULL;
+ SDValue* LoMul = NULL;
+ SDValue* LowAdd = NULL;
+
+ if (IsLeftOperandMUL)
+ HiAdd = &AddeOp1;
+ else
+ HiAdd = &AddeOp0;
+
+
+ if (AddcOp0->getOpcode() == Opc) {
+ LoMul = &AddcOp0;
+ LowAdd = &AddcOp1;
+ }
+ if (AddcOp1->getOpcode() == Opc) {
+ LoMul = &AddcOp1;
+ LowAdd = &AddcOp0;
+ }
+
+ if (LoMul == NULL)
+ return SDValue();
+
+ if (LoMul->getNode() != HiMul->getNode())
+ return SDValue();
+
+ // Create the merged node.
+ SelectionDAG &DAG = DCI.DAG;
+
+ // Build operand list.
+ SmallVector<SDValue, 8> Ops;
+ Ops.push_back(LoMul->getOperand(0));
+ Ops.push_back(LoMul->getOperand(1));
+ Ops.push_back(*LowAdd);
+ Ops.push_back(*HiAdd);
+
+ SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
+ DAG.getVTList(MVT::i32, MVT::i32),
+ &Ops[0], Ops.size());
+
+ // Replace the ADDs' nodes uses by the MLA node's values.
+ SDValue HiMLALResult(MLALNode.getNode(), 1);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
+
+ SDValue LoMLALResult(MLALNode.getNode(), 0);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
+
+ // Return original node to notify the driver to stop replacing.
+ SDValue resNode(AddcNode, 0);
+ return resNode;
+}
+
+/// PerformADDCCombine - Target-specific dag combine transform from
+/// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
+static SDValue PerformADDCCombine(SDNode *N,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const ARMSubtarget *Subtarget) {
+
+ return AddCombineTo64bitMLAL(N, DCI, Subtarget);
+
+}
+
/// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
/// operands N0 and N1. This is a helper for PerformADDCombine that is
/// called with the default operands, and if that fails, with commuted
DAGCombinerInfo &DCI) const {
switch (N->getOpcode()) {
default: break;
+ case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
case ISD::SUB: return PerformSUBCombine(N, DCI);
case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
}
bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
- if (!Subtarget->allowsUnalignedMem())
- return false;
+ // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
+ bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
switch (VT.getSimpleVT().SimpleTy) {
default:
case MVT::i8:
case MVT::i16:
case MVT::i32:
- return true;
+ // Unaligned access can use (for example) LRDB, LRDH, LDR
+ return AllowsUnaligned;
case MVT::f64:
- return Subtarget->hasNEON();
- // FIXME: VLD1 etc with standard alignment is legal.
+ case MVT::v2f64:
+ // For any little-endian targets with neon, we can support unaligned ld/st
+ // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
+ // A big-endian target may also explictly support unaligned accesses
+ return Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian());
}
}
// See if we can use NEON instructions for this...
if (IsZeroVal &&
- !F->hasFnAttr(Attribute::NoImplicitFloat) &&
+ !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat) &&
Subtarget->hasNEON()) {
if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
return MVT::v4i32;
case Intrinsic::arm_neon_vld4lane: {
Info.opc = ISD::INTRINSIC_W_CHAIN;
// Conservatively set memVT to the entire set of vectors loaded.
- uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
+ uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
Info.ptrVal = I.getArgOperand(0);
Info.offset = 0;
Type *ArgTy = I.getArgOperand(ArgI)->getType();
if (!ArgTy->isVectorTy())
break;
- NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
+ NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
}
Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
Info.ptrVal = I.getArgOperand(0);