using namespace llvm;
STATISTIC(NumTailCalls, "Number of tail calls");
+STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
// This option should go away when tail calls fully work.
static cl::opt<bool>
setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
}
+ if (HasDivModLibcall) {
+ setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
+ setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
+ }
+
if (Subtarget->isThumb1Only())
addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
else
setOperationAction(ISD::MUL, MVT::v8i16, Custom);
setOperationAction(ISD::MUL, MVT::v4i32, Custom);
setOperationAction(ISD::MUL, MVT::v2i64, Custom);
+ // Custom handling for some vector types to avoid expensive expansions
+ setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
+ setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
+ setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
+ setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
+ // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
+ // a destination type that is wider than the source.
+ setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
+ setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
+ setTargetDAGCombine(ISD::INTRINSIC_VOID);
+ setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
setTargetDAGCombine(ISD::SHL);
setTargetDAGCombine(ISD::SRL);
setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
- // FIXME: Shouldn't need this, since no register is used, but the legalizer
- // doesn't yet know how to not do that for SjLj.
- setExceptionSelectorRegister(ARM::R0);
+ setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
+ setExceptionPointerRegister(ARM::R0);
+ setExceptionSelectorRegister(ARM::R1);
+
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
// ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
// the default expansion.
benefitFromCodePlacementOpt = true;
}
+// FIXME: It might make sense to define the representative register class as the
+// nearest super-register that has a non-null superset. For example, DPR_VFP2 is
+// a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
+// SPR's representative would be DPR_VFP2. This should work well if register
+// pressure tracking were modified such that a register use would increment the
+// pressure of the register class's representative and all of it's super
+// classes' representatives transitively. We have not implemented this because
+// of the difficulty prior to coalescing of modeling operand register classes
+// due to the common occurrence of cross class copies and subregister insertions
+// and extractions.
std::pair<const TargetRegisterClass*, uint8_t>
ARMTargetLowering::findRepresentativeClass(EVT VT) const{
const TargetRegisterClass *RRC = 0;
case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
RRC = ARM::DPRRegisterClass;
+ // When NEON is used for SP, only half of the register file is available
+ // because operations that define both SP and DP results will be constrained
+ // to the VFP2 class (D0-D15). We currently model this constraint prior to
+ // coalescing by double-counting the SP regs. See the FIXME above.
+ if (Subtarget->useNEONForSinglePrecisionFP())
+ Cost = 2;
break;
case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
case MVT::v4f32: case MVT::v2f64:
switch (Opcode) {
default: return 0;
case ARMISD::Wrapper: return "ARMISD::Wrapper";
+ case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
+ case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
case ARMISD::CALL: return "ARMISD::CALL";
case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
case ARMISD::CMOV: return "ARMISD::CMOV";
- case ARMISD::CNEG: return "ARMISD::CNEG";
case ARMISD::RBIT: return "ARMISD::RBIT";
case ARMISD::VZIP: return "ARMISD::VZIP";
case ARMISD::VUZP: return "ARMISD::VUZP";
case ARMISD::VTRN: return "ARMISD::VTRN";
+ case ARMISD::VTBL1: return "ARMISD::VTBL1";
+ case ARMISD::VTBL2: return "ARMISD::VTBL2";
case ARMISD::VMULLs: return "ARMISD::VMULLs";
case ARMISD::VMULLu: return "ARMISD::VMULLu";
case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
case ARMISD::BFI: return "ARMISD::BFI";
case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
+ case ARMISD::VBSL: return "ARMISD::VBSL";
case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
+ case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
+ case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
+ case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
+ case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
+ case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
+ case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
+ case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
+ case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
+ case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
+ case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
+ case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
+ case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
+ case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
+ case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
+ case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
+ case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
+ case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
}
}
return Sched::RegPressure;
}
-unsigned
-ARMTargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
- MachineFunction &MF) const {
- const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo();
-
- switch (RC->getID()) {
- default:
- return 0;
- case ARM::tGPRRegClassID:
- return TFI->hasFP(MF) ? 4 : 5;
- case ARM::GPRRegClassID: {
- unsigned FP = TFI->hasFP(MF) ? 1 : 0;
- return 10 - FP - (Subtarget->isR9Reserved() ? 1 : 0);
- }
- case ARM::SPRRegClassID: // Currently not used as 'rep' register class.
- case ARM::DPRRegClassID:
- return 32 - 10;
- }
-}
-
//===----------------------------------------------------------------------===//
// Lowering Code
//===----------------------------------------------------------------------===//
CCValAssign &VA = ArgLocs[i];
SDValue Arg = OutVals[realArgIdx];
ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
+ bool isByVal = Flags.isByVal();
// Promote the value if needed.
switch (VA.getLocInfo()) {
}
} else if (VA.isRegLoc()) {
RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
- } else if (!IsSibCall) {
+ } else if (!IsSibCall || isByVal) {
assert(VA.isMemLoc());
MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
// than necessary, because it means that each store effectively depends
// on every argument instead of just those arguments it would clobber.
- // Do not flag preceeding copytoreg stuff together with the following stuff.
+ // Do not flag preceding copytoreg stuff together with the following stuff.
InFlag = SDValue();
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
const GlobalValue *GV = G->getGlobal();
// Create a constant pool entry for the callee address
- unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
ARMPCLabelIndex,
ARMCP::CPValue, 0);
const char *Sym = S->getSymbol();
// Create a constant pool entry for the callee address
- unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
Sym, ARMPCLabelIndex, 0);
// Get the address of the callee into a register
isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
// tBX takes a register source operand.
if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
- unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
ARMPCLabelIndex,
ARMCP::CPValue, 4);
// tBX takes a register source operand.
const char *Sym = S->getSymbol();
if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
- unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
Sym, ARMPCLabelIndex, 4);
SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
dl, DAG, InVals);
}
+/// HandleByVal - Every parameter *after* a byval parameter is passed
+/// on the stack. Confiscate all the parameter registers to insure
+/// this.
+void
+llvm::ARMTargetLowering::HandleByVal(CCState *State) const {
+ static const unsigned RegList1[] = {
+ ARM::R0, ARM::R1, ARM::R2, ARM::R3
+ };
+ do {} while (State->AllocateReg(RegList1, 4));
+}
+
/// MatchingStackOffset - Return true if the given stack call argument is
/// already available in the same position (relatively) of the caller's
/// incoming argument stack.
return HasRet;
}
+bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
+ if (!EnableARMTailCalls)
+ return false;
+
+ if (!CI->isTailCall())
+ return false;
+
+ return !Subtarget->isThumb1Only();
+}
+
// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
// their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
// one of the above mentioned nodes. It has to be wrapped because otherwise
CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
} else {
unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
- ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ ARMPCLabelIndex = AFI->createPICLabelUId();
ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
ARMCP::CPBlockAddress,
PCAdj);
unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
MachineFunction &MF = DAG.getMachineFunction();
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
- unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
ARMConstantPoolValue *CPV =
new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
if (GV->isDeclaration()) {
MachineFunction &MF = DAG.getMachineFunction();
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
- unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
// Initial exec model.
unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
ARMConstantPoolValue *CPV =
Result = DAG.getLoad(PtrVT, dl, Chain, Result,
MachinePointerInfo::getGOT(), false, false, 0);
return Result;
+ }
+
+ // If we have T2 ops, we can materialize the address directly via movt/movw
+ // pair. This is always cheaper.
+ if (Subtarget->useMovt()) {
+ ++NumMovwMovt;
+ // FIXME: Once remat is capable of dealing with instructions with register
+ // operands, expand this into two nodes.
+ return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
+ DAG.getTargetGlobalAddress(GV, dl, PtrVT));
} else {
- // If we have T2 ops, we can materialize the address directly via movt/movw
- // pair. This is always cheaper.
- if (Subtarget->useMovt()) {
- return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
- DAG.getTargetGlobalAddress(GV, dl, PtrVT));
- } else {
- SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
- CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
- return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
- MachinePointerInfo::getConstantPool(),
- false, false, 0);
- }
+ SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
+ CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
+ return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
+ MachinePointerInfo::getConstantPool(),
+ false, false, 0);
}
}
SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
SelectionDAG &DAG) const {
- MachineFunction &MF = DAG.getMachineFunction();
- ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
- unsigned ARMPCLabelIndex = 0;
EVT PtrVT = getPointerTy();
DebugLoc dl = Op.getDebugLoc();
const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
Reloc::Model RelocM = getTargetMachine().getRelocationModel();
+ MachineFunction &MF = DAG.getMachineFunction();
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+
+ if (Subtarget->useMovt()) {
+ ++NumMovwMovt;
+ // FIXME: Once remat is capable of dealing with instructions with register
+ // operands, expand this into two nodes.
+ if (RelocM == Reloc::Static)
+ return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
+ DAG.getTargetGlobalAddress(GV, dl, PtrVT));
+
+ unsigned Wrapper = (RelocM == Reloc::PIC_)
+ ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
+ SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
+ DAG.getTargetGlobalAddress(GV, dl, PtrVT));
+ if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
+ Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
+ MachinePointerInfo::getGOT(), false, false, 0);
+ return Result;
+ }
+
+ unsigned ARMPCLabelIndex = 0;
SDValue CPAddr;
- if (RelocM == Reloc::Static)
+ if (RelocM == Reloc::Static) {
CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
- else {
- ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ } else {
+ ARMPCLabelIndex = AFI->createPICLabelUId();
unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
ARMConstantPoolValue *CPV =
new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
"GLOBAL OFFSET TABLE not implemented for non-ELF targets");
MachineFunction &MF = DAG.getMachineFunction();
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
- unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
EVT PtrVT = getPointerTy();
DebugLoc dl = Op.getDebugLoc();
unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
const {
DebugLoc dl = Op.getDebugLoc();
return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
- Op.getOperand(0), Op.getOperand(1));
+ Op.getOperand(0));
}
SDValue
case Intrinsic::eh_sjlj_lsda: {
MachineFunction &MF = DAG.getMachineFunction();
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
- unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
EVT PtrVT = getPointerTy();
DebugLoc dl = Op.getDebugLoc();
Reloc::Model RelocM = getTargetMachine().getRelocationModel();
}
return Result;
}
+ case Intrinsic::arm_neon_vmulls:
+ case Intrinsic::arm_neon_vmullu: {
+ unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
+ ? ARMISD::VMULLs : ARMISD::VMULLu;
+ return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
+ Op.getOperand(1), Op.getOperand(2));
+ }
}
}
isVarArg));
SmallVector<SDValue, 16> ArgValues;
+ int lastInsIndex = -1;
+ SDValue ArgValue;
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
if (VA.isRegLoc()) {
EVT RegVT = VA.getLocVT();
- SDValue ArgValue;
if (VA.needsCustom()) {
// f64 and vector types are split up into multiple registers or
// combinations of registers and stack slots.
assert(VA.isMemLoc());
assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
- unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
- int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(), true);
+ int index = ArgLocs[i].getValNo();
+
+ // Some Ins[] entries become multiple ArgLoc[] entries.
+ // Process them only once.
+ if (index != lastInsIndex)
+ {
+ ISD::ArgFlagsTy Flags = Ins[index].Flags;
+ // FIXME: For now, all byval parameter objects are marked mutable. This can be
+ // changed with more analysis.
+ // In case of tail call optimization mark all arguments mutable. Since they
+ // could be overwritten by lowering of arguments in case of a tail call.
+ if (Flags.isByVal()) {
+ unsigned Bytes = Flags.getByValSize();
+ if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
+ int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), false);
+ InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
+ } else {
+ int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
+ VA.getLocMemOffset(), true);
- // Create load nodes to retrieve arguments from the stack.
- SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
- InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
- MachinePointerInfo::getFixedStack(FI),
- false, false, 0));
+ // Create load nodes to retrieve arguments from the stack.
+ SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
+ InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
+ MachinePointerInfo::getFixedStack(FI),
+ false, false, 0));
+ }
+ lastInsIndex = index;
+ }
}
}
unsigned NumGPRs = CCInfo.getFirstUnallocated
(GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
- unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
+ unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
unsigned VARegSize = (4 - NumGPRs) * 4;
unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
unsigned ArgOffset = CCInfo.getNextStackOffset();
return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
}
+/// duplicateCmp - Glue values can have only one use, so this function
+/// duplicates a comparison node.
+SDValue
+ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
+ unsigned Opc = Cmp.getOpcode();
+ DebugLoc DL = Cmp.getDebugLoc();
+ if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
+ return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
+
+ assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
+ Cmp = Cmp.getOperand(0);
+ Opc = Cmp.getOpcode();
+ if (Opc == ARMISD::CMPFP)
+ Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
+ else {
+ assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
+ Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
+ }
+ return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
+}
+
SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
SDValue Cond = Op.getOperand(0);
SDValue SelectTrue = Op.getOperand(1);
EVT VT = Cond.getValueType();
SDValue ARMcc = Cond.getOperand(2);
SDValue CCR = Cond.getOperand(3);
- SDValue Cmp = Cond.getOperand(4);
+ SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
}
}
// If one of the operand is zero, it's safe to ignore the NaN case since
// we only care about equality comparisons.
(SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
- // If unsafe fp math optimization is enabled and there are no othter uses of
- // the CMP operands, and the condition code is EQ oe NE, we can optimize it
+ // If unsafe fp math optimization is enabled and there are no other uses of
+ // the CMP operands, and the condition code is EQ or NE, we can optimize it
// to an integer comparison.
if (CC == ISD::SETOEQ)
CC = ISD::SETEQ;
return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
}
+static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
+ EVT VT = Op.getValueType();
+ DebugLoc dl = Op.getDebugLoc();
+
+ EVT OperandVT = Op.getOperand(0).getValueType();
+ assert(OperandVT == MVT::v4i16 && "Invalid type for custom lowering!");
+ if (VT != MVT::v4f32)
+ return DAG.UnrollVectorOp(Op.getNode());
+
+ unsigned CastOpc;
+ unsigned Opc;
+ switch (Op.getOpcode()) {
+ default:
+ assert(0 && "Invalid opcode!");
+ case ISD::SINT_TO_FP:
+ CastOpc = ISD::SIGN_EXTEND;
+ Opc = ISD::SINT_TO_FP;
+ break;
+ case ISD::UINT_TO_FP:
+ CastOpc = ISD::ZERO_EXTEND;
+ Opc = ISD::UINT_TO_FP;
+ break;
+ }
+
+ Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
+ return DAG.getNode(Opc, dl, VT, Op);
+}
+
static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
EVT VT = Op.getValueType();
+ if (VT.isVector())
+ return LowerVectorINT_TO_FP(Op, DAG);
+
DebugLoc dl = Op.getDebugLoc();
unsigned Opc;
DebugLoc dl = Op.getDebugLoc();
EVT VT = Op.getValueType();
EVT SrcVT = Tmp1.getValueType();
- SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
- SDValue ARMcc = DAG.getConstant(ARMCC::LT, MVT::i32);
- SDValue FP0 = DAG.getConstantFP(0.0, SrcVT);
- SDValue Cmp = getVFPCmp(Tmp1, FP0, DAG, dl);
- SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
- return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMcc, CCR, Cmp);
+ bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
+ Tmp0.getOpcode() == ARMISD::VMOVDRR;
+ bool UseNEON = !InGPR && Subtarget->hasNEON();
+
+ if (UseNEON) {
+ // Use VBSL to copy the sign bit.
+ unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
+ SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
+ DAG.getTargetConstant(EncodedVal, MVT::i32));
+ EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
+ if (VT == MVT::f64)
+ Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
+ DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
+ DAG.getConstant(32, MVT::i32));
+ else /*if (VT == MVT::f32)*/
+ Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
+ if (SrcVT == MVT::f32) {
+ Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
+ if (VT == MVT::f64)
+ Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
+ DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
+ DAG.getConstant(32, MVT::i32));
+ } else if (VT == MVT::f32)
+ Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
+ DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
+ DAG.getConstant(32, MVT::i32));
+ Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
+ Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
+
+ SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
+ MVT::i32);
+ AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
+ SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
+ DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
+
+ SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
+ DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
+ DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
+ if (VT == MVT::f32) {
+ Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
+ Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
+ DAG.getConstant(0, MVT::i32));
+ } else {
+ Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
+ }
+
+ return Res;
+ }
+
+ // Bitcast operand 1 to i32.
+ if (SrcVT == MVT::f64)
+ Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
+ &Tmp1, 1).getValue(1);
+ Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
+
+ // Or in the signbit with integer operations.
+ SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
+ SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
+ Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
+ if (VT == MVT::f32) {
+ Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
+ DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
+ return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
+ DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
+ }
+
+ // f64: Or the high part with signbit and then combine two parts.
+ Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
+ &Tmp0, 1);
+ SDValue Lo = Tmp0.getValue(0);
+ SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
+ Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
+ return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
}
SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
return true;
}
+static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) {
+ // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
+ // range, then 0 is placed into the resulting vector. So pretty much any mask
+ // of 8 elements can work here.
+ return VT == MVT::v8i8 && M.size() == 8;
+}
+
static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
unsigned &WhichResult) {
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
}
// Gather data to see if the operation can be modelled as a
-// shuffle in combination with VEXTs.
-SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const {
+// shuffle in combination with VEXTs.
+SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
+ SelectionDAG &DAG) const {
DebugLoc dl = Op.getDebugLoc();
EVT VT = Op.getValueType();
unsigned NumElts = VT.getVectorNumElements();
SmallVector<SDValue, 2> SourceVecs;
SmallVector<unsigned, 2> MinElts;
SmallVector<unsigned, 2> MaxElts;
-
+
for (unsigned i = 0; i < NumElts; ++i) {
SDValue V = Op.getOperand(i);
if (V.getOpcode() == ISD::UNDEF)
// elements of other vectors.
return SDValue();
}
-
+
// Record this extraction against the appropriate vector if possible...
SDValue SourceVec = V.getOperand(0);
unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
break;
}
}
-
+
// Or record a new source if not...
if (!FoundSource) {
SourceVecs.push_back(SourceVec);
MaxElts.push_back(EltNo);
}
}
-
+
// Currently only do something sane when at most two source vectors
// involved.
if (SourceVecs.size() > 2)
SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
int VEXTOffsets[2] = {0, 0};
-
+
// This loop extracts the usage patterns of the source vectors
// and prepares appropriate SDValues for a shuffle if possible.
for (unsigned i = 0; i < SourceVecs.size(); ++i) {
// break it down again in a shuffle.
return SDValue();
}
-
+
// Since only 64-bit and 128-bit vectors are legal on ARM and
// we've eliminated the other cases...
assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
"unexpected vector sizes in ReconstructShuffle");
-
+
if (MaxElts[i] - MinElts[i] >= NumElts) {
// Span too large for a VEXT to cope
return SDValue();
- }
-
+ }
+
if (MinElts[i] >= NumElts) {
// The extraction can just take the second half
VEXTOffsets[i] = NumElts;
- ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
+ ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
+ SourceVecs[i],
DAG.getIntPtrConstant(NumElts));
} else if (MaxElts[i] < NumElts) {
// The extraction can just take the first half
VEXTOffsets[i] = 0;
- ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
+ ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
+ SourceVecs[i],
DAG.getIntPtrConstant(0));
} else {
// An actual VEXT is needed
VEXTOffsets[i] = MinElts[i];
- SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
+ SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
+ SourceVecs[i],
DAG.getIntPtrConstant(0));
- SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
+ SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
+ SourceVecs[i],
DAG.getIntPtrConstant(NumElts));
ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
DAG.getConstant(VEXTOffsets[i], MVT::i32));
}
}
-
+
SmallVector<int, 8> Mask;
-
+
for (unsigned i = 0; i < NumElts; ++i) {
SDValue Entry = Op.getOperand(i);
if (Entry.getOpcode() == ISD::UNDEF) {
Mask.push_back(-1);
continue;
}
-
+
SDValue ExtractVec = Entry.getOperand(0);
- int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
+ int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
+ .getOperand(1))->getSExtValue();
if (ExtractVec == SourceVecs[0]) {
Mask.push_back(ExtractElt - VEXTOffsets[0]);
} else {
Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
}
}
-
+
// Final check before we try to produce nonsense...
if (isShuffleMaskLegal(Mask, VT))
- return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1], &Mask[0]);
-
+ return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
+ &Mask[0]);
+
return SDValue();
}
isVREVMask(M, VT, 32) ||
isVREVMask(M, VT, 16) ||
isVEXTMask(M, VT, ReverseVEXT, Imm) ||
+ isVTBLMask(M, VT) ||
isVTRNMask(M, VT, WhichResult) ||
isVUZPMask(M, VT, WhichResult) ||
isVZIPMask(M, VT, WhichResult) ||
}
}
+static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
+ SmallVectorImpl<int> &ShuffleMask,
+ SelectionDAG &DAG) {
+ // Check to see if we can use the VTBL instruction.
+ SDValue V1 = Op.getOperand(0);
+ SDValue V2 = Op.getOperand(1);
+ DebugLoc DL = Op.getDebugLoc();
+
+ SmallVector<SDValue, 8> VTBLMask;
+ for (SmallVectorImpl<int>::iterator
+ I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
+ VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
+
+ if (V2.getNode()->getOpcode() == ISD::UNDEF)
+ return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
+ &VTBLMask[0], 8));
+
+ return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
+ &VTBLMask[0], 8));
+}
+
static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
return DAG.getNode(ISD::BITCAST, dl, VT, Val);
}
+ if (VT == MVT::v8i8) {
+ SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
+ if (NewOp.getNode())
+ return NewOp;
+ }
+
return SDValue();
}
MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
}
+static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
+ unsigned Opcode = N->getOpcode();
+ if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
+ SDNode *N0 = N->getOperand(0).getNode();
+ SDNode *N1 = N->getOperand(1).getNode();
+ return N0->hasOneUse() && N1->hasOneUse() &&
+ isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
+ }
+ return false;
+}
+
+static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
+ unsigned Opcode = N->getOpcode();
+ if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
+ SDNode *N0 = N->getOperand(0).getNode();
+ SDNode *N1 = N->getOperand(1).getNode();
+ return N0->hasOneUse() && N1->hasOneUse() &&
+ isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
+ }
+ return false;
+}
+
static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
// Multiplications are only custom-lowered for 128-bit vectors so that
// VMULL can be detected. Otherwise v2i64 multiplications are not legal.
SDNode *N0 = Op.getOperand(0).getNode();
SDNode *N1 = Op.getOperand(1).getNode();
unsigned NewOpc = 0;
- if (isSignExtended(N0, DAG) && isSignExtended(N1, DAG))
+ bool isMLA = false;
+ bool isN0SExt = isSignExtended(N0, DAG);
+ bool isN1SExt = isSignExtended(N1, DAG);
+ if (isN0SExt && isN1SExt)
NewOpc = ARMISD::VMULLs;
- else if (isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG))
- NewOpc = ARMISD::VMULLu;
- else if (VT == MVT::v2i64)
- // Fall through to expand this. It is not legal.
- return SDValue();
- else
- // Other vector multiplications are legal.
- return Op;
+ else {
+ bool isN0ZExt = isZeroExtended(N0, DAG);
+ bool isN1ZExt = isZeroExtended(N1, DAG);
+ if (isN0ZExt && isN1ZExt)
+ NewOpc = ARMISD::VMULLu;
+ else if (isN1SExt || isN1ZExt) {
+ // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
+ // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
+ if (isN1SExt && isAddSubSExt(N0, DAG)) {
+ NewOpc = ARMISD::VMULLs;
+ isMLA = true;
+ } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
+ NewOpc = ARMISD::VMULLu;
+ isMLA = true;
+ } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
+ std::swap(N0, N1);
+ NewOpc = ARMISD::VMULLu;
+ isMLA = true;
+ }
+ }
+
+ if (!NewOpc) {
+ if (VT == MVT::v2i64)
+ // Fall through to expand this. It is not legal.
+ return SDValue();
+ else
+ // Other vector multiplications are legal.
+ return Op;
+ }
+ }
// Legalize to a VMULL instruction.
DebugLoc DL = Op.getDebugLoc();
- SDValue Op0 = SkipExtension(N0, DAG);
+ SDValue Op0;
SDValue Op1 = SkipExtension(N1, DAG);
+ if (!isMLA) {
+ Op0 = SkipExtension(N0, DAG);
+ assert(Op0.getValueType().is64BitVector() &&
+ Op1.getValueType().is64BitVector() &&
+ "unexpected types for extended operands to VMULL");
+ return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
+ }
+
+ // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
+ // isel lowering to take advantage of no-stall back to back vmul + vmla.
+ // vmull q0, d4, d6
+ // vmlal q0, d5, d6
+ // is faster than
+ // vaddl q0, d4, d5
+ // vmovl q1, d6
+ // vmul q0, q0, q1
+ SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
+ SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
+ EVT Op1VT = Op1.getValueType();
+ return DAG.getNode(N0->getOpcode(), DL, VT,
+ DAG.getNode(NewOpc, DL, VT,
+ DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
+ DAG.getNode(NewOpc, DL, VT,
+ DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
+}
+
+static SDValue
+LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
+ // Convert to float
+ // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
+ // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
+ X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
+ Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
+ X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
+ Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
+ // Get reciprocal estimate.
+ // float4 recip = vrecpeq_f32(yf);
+ Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
+ // Because char has a smaller range than uchar, we can actually get away
+ // without any newton steps. This requires that we use a weird bias
+ // of 0xb000, however (again, this has been exhaustively tested).
+ // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
+ X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
+ X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
+ Y = DAG.getConstant(0xb000, MVT::i32);
+ Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
+ X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
+ X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
+ // Convert back to short.
+ X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
+ X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
+ return X;
+}
+
+static SDValue
+LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
+ SDValue N2;
+ // Convert to float.
+ // float4 yf = vcvt_f32_s32(vmovl_s16(y));
+ // float4 xf = vcvt_f32_s32(vmovl_s16(x));
+ N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
+ N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
+ N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
+ N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
+
+ // Use reciprocal estimate and one refinement step.
+ // float4 recip = vrecpeq_f32(yf);
+ // recip *= vrecpsq_f32(yf, recip);
+ N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
+ N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
+ N1, N2);
+ N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
+ // Because short has a smaller range than ushort, we can actually get away
+ // with only a single newton step. This requires that we use a weird bias
+ // of 89, however (again, this has been exhaustively tested).
+ // float4 result = as_float4(as_int4(xf*recip) + 89);
+ N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
+ N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
+ N1 = DAG.getConstant(89, MVT::i32);
+ N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
+ N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
+ N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
+ // Convert back to integer and return.
+ // return vmovn_s32(vcvt_s32_f32(result));
+ N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
+ N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
+ return N0;
+}
+
+static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
+ EVT VT = Op.getValueType();
+ assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
+ "unexpected type for custom-lowering ISD::SDIV");
+
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue N0 = Op.getOperand(0);
+ SDValue N1 = Op.getOperand(1);
+ SDValue N2, N3;
+
+ if (VT == MVT::v8i8) {
+ N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
+ N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
+
+ N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
+ DAG.getIntPtrConstant(4));
+ N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
+ DAG.getIntPtrConstant(4));
+ N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
+ DAG.getIntPtrConstant(0));
+ N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
+ DAG.getIntPtrConstant(0));
- assert(Op0.getValueType().is64BitVector() &&
- Op1.getValueType().is64BitVector() &&
- "unexpected types for extended operands to VMULL");
- return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
+ N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
+ N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
+
+ N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
+ N0 = LowerCONCAT_VECTORS(N0, DAG);
+
+ N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
+ return N0;
+ }
+ return LowerSDIV_v4i16(N0, N1, dl, DAG);
+}
+
+static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
+ EVT VT = Op.getValueType();
+ assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
+ "unexpected type for custom-lowering ISD::UDIV");
+
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue N0 = Op.getOperand(0);
+ SDValue N1 = Op.getOperand(1);
+ SDValue N2, N3;
+
+ if (VT == MVT::v8i8) {
+ N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
+ N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
+
+ N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
+ DAG.getIntPtrConstant(4));
+ N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
+ DAG.getIntPtrConstant(4));
+ N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
+ DAG.getIntPtrConstant(0));
+ N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
+ DAG.getIntPtrConstant(0));
+
+ N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
+ N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
+
+ N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
+ N0 = LowerCONCAT_VECTORS(N0, DAG);
+
+ N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
+ DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
+ N0);
+ return N0;
+ }
+
+ // v4i16 sdiv ... Convert to float.
+ // float4 yf = vcvt_f32_s32(vmovl_u16(y));
+ // float4 xf = vcvt_f32_s32(vmovl_u16(x));
+ N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
+ N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
+ N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
+ N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
+
+ // Use reciprocal estimate and two refinement steps.
+ // float4 recip = vrecpeq_f32(yf);
+ // recip *= vrecpsq_f32(yf, recip);
+ // recip *= vrecpsq_f32(yf, recip);
+ N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
+ N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
+ N1, N2);
+ N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
+ N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
+ N1, N2);
+ N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
+ // Simply multiplying by the reciprocal estimate can leave us a few ulps
+ // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
+ // and that it will never cause us to return an answer too large).
+ // float4 result = as_float4(as_int4(xf*recip) + 89);
+ N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
+ N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
+ N1 = DAG.getConstant(2, MVT::i32);
+ N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
+ N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
+ N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
+ // Convert back to integer and return.
+ // return vmovn_u32(vcvt_s32_f32(result));
+ N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
+ N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
+ return N0;
}
SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
case ISD::GlobalAddress:
return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
LowerGlobalAddressELF(Op, DAG);
- case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
+ case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
case ISD::SELECT: return LowerSELECT(Op, DAG);
case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
case ISD::BR_CC: return LowerBR_CC(Op, DAG);
case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
Subtarget);
- case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
+ case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
case ISD::SHL:
case ISD::SRL:
case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
case ISD::MUL: return LowerMUL(Op, DAG);
+ case ISD::SDIV: return LowerSDIV(Op, DAG);
+ case ISD::UDIV: return LowerUDIV(Op, DAG);
}
return SDValue();
}
default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
case 1:
ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
- strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
+ strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
break;
case 2:
ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
+ case ARM::ADCSSri:
+ case ARM::ADCSSrr:
+ case ARM::ADCSSrs:
+ case ARM::SBCSSri:
+ case ARM::SBCSSrr:
+ case ARM::SBCSSrs:
+ case ARM::RSBSri:
+ case ARM::RSBSrr:
+ case ARM::RSBSrs:
+ case ARM::RSCSri:
+ case ARM::RSCSrs: {
+ unsigned OldOpc = MI->getOpcode();
+ unsigned Opc = 0;
+ switch (OldOpc) {
+ case ARM::ADCSSrr:
+ Opc = ARM::ADCrr;
+ break;
+ case ARM::ADCSSri:
+ Opc = ARM::ADCri;
+ break;
+ case ARM::ADCSSrs:
+ Opc = ARM::ADCrs;
+ break;
+ case ARM::SBCSSrr:
+ Opc = ARM::SBCrr;
+ break;
+ case ARM::SBCSSri:
+ Opc = ARM::SBCri;
+ break;
+ case ARM::SBCSSrs:
+ Opc = ARM::SBCrs;
+ break;
+ case ARM::RSBSri:
+ Opc = ARM::RSBri;
+ break;
+ case ARM::RSBSrr:
+ Opc = ARM::RSBrr;
+ break;
+ case ARM::RSBSrs:
+ Opc = ARM::RSBrs;
+ break;
+ case ARM::RSCSri:
+ Opc = ARM::RSCri;
+ break;
+ case ARM::RSCSrs:
+ Opc = ARM::RSCrs;
+ break;
+ default:
+ llvm_unreachable("Unknown opcode?");
+ }
+
+ MachineInstrBuilder MIB =
+ BuildMI(*BB, MI, MI->getDebugLoc(), TII->get(Opc));
+ for (unsigned i = 0; i < MI->getNumOperands(); ++i)
+ MIB.addOperand(MI->getOperand(i));
+ AddDefaultPred(MIB);
+ MIB.addReg(ARM::CPSR, RegState::Define); // S bit
+ MI->eraseFromParent();
+ return BB;
+ }
+
+
case ARM::tMOVCCr_pseudo: {
// To "insert" a SELECT_CC instruction, we actually have to insert the
// diamond control-flow pattern. The incoming instruction knows the
case ARM::BCCZi64: {
// If there is an unconditional branch to the other successor, remove it.
BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
-
+
// Compare both parts that make up the double comparison separately for
// equality.
bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
return SDValue();
}
+/// PerformVMULCombine
+/// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
+/// special multiplier accumulator forwarding.
+/// vmul d3, d0, d2
+/// vmla d3, d1, d2
+/// is faster than
+/// vadd d3, d0, d1
+/// vmul d3, d3, d2
+static SDValue PerformVMULCombine(SDNode *N,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const ARMSubtarget *Subtarget) {
+ if (!Subtarget->hasVMLxForwarding())
+ return SDValue();
+
+ SelectionDAG &DAG = DCI.DAG;
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ unsigned Opcode = N0.getOpcode();
+ if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
+ Opcode != ISD::FADD && Opcode != ISD::FSUB) {
+ Opcode = N0.getOpcode();
+ if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
+ Opcode != ISD::FADD && Opcode != ISD::FSUB)
+ return SDValue();
+ std::swap(N0, N1);
+ }
+
+ EVT VT = N->getValueType(0);
+ DebugLoc DL = N->getDebugLoc();
+ SDValue N00 = N0->getOperand(0);
+ SDValue N01 = N0->getOperand(1);
+ return DAG.getNode(Opcode, DL, VT,
+ DAG.getNode(ISD::MUL, DL, VT, N00, N1),
+ DAG.getNode(ISD::MUL, DL, VT, N01, N1));
+}
+
static SDValue PerformMULCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI,
const ARMSubtarget *Subtarget) {
return SDValue();
EVT VT = N->getValueType(0);
+ if (VT.is64BitVector() || VT.is128BitVector())
+ return PerformVMULCombine(N, DCI, Subtarget);
if (VT != MVT::i32)
return SDValue();
static SDValue PerformANDCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI) {
+
// Attempt to use immediate-form VBIC
BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
DebugLoc dl = N->getDebugLoc();
EVT VT = N->getValueType(0);
SelectionDAG &DAG = DCI.DAG;
+ if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
+ return SDValue();
+
APInt SplatBits, SplatUndef;
unsigned SplatBitSize;
bool HasAnyUndefs;
EVT VT = N->getValueType(0);
SelectionDAG &DAG = DCI.DAG;
+ if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
+ return SDValue();
+
APInt SplatBits, SplatUndef;
unsigned SplatBitSize;
bool HasAnyUndefs;
}
}
+ SDValue N0 = N->getOperand(0);
+ if (N0.getOpcode() != ISD::AND)
+ return SDValue();
+ SDValue N1 = N->getOperand(1);
+
+ // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
+ if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
+ DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
+ APInt SplatUndef;
+ unsigned SplatBitSize;
+ bool HasAnyUndefs;
+
+ BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
+ APInt SplatBits0;
+ if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
+ HasAnyUndefs) && !HasAnyUndefs) {
+ BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
+ APInt SplatBits1;
+ if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
+ HasAnyUndefs) && !HasAnyUndefs &&
+ SplatBits0 == ~SplatBits1) {
+ // Canonicalize the vector type to make instruction selection simpler.
+ EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
+ SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
+ N0->getOperand(1), N0->getOperand(0),
+ N1->getOperand(0));
+ return DAG.getNode(ISD::BITCAST, dl, VT, Result);
+ }
+ }
+ }
+
// Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
// reasonable.
if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
return SDValue();
- SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
DebugLoc DL = N->getDebugLoc();
// 1) or (and A, mask), val => ARMbfi A, val, mask
// iff (val & mask) == val
//
// 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
// 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
- // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
+ // && mask == ~mask2
// 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
- // && CountPopulation_32(mask) == CountPopulation_32(~mask2)
+ // && ~mask == mask2
// (i.e., copy a bitfield value into another bitfield of the same width)
- if (N0.getOpcode() != ISD::AND)
- return SDValue();
if (VT != MVT::i32)
return SDValue();
return SDValue();
unsigned Mask2 = N11C->getZExtValue();
+ // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
+ // as is to match.
if (ARM::isBitFieldInvertedMask(Mask) &&
- ARM::isBitFieldInvertedMask(~Mask2) &&
- (CountPopulation_32(Mask) == CountPopulation_32(~Mask2))) {
+ (Mask == ~Mask2)) {
// The pack halfword instruction works better for masks that fit it,
// so use that when it's available.
if (Subtarget->hasT2ExtractPack() &&
(Mask == 0xffff || Mask == 0xffff0000))
return SDValue();
// 2a
- unsigned lsb = CountTrailingZeros_32(Mask2);
+ unsigned amt = CountTrailingZeros_32(Mask2);
Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
- DAG.getConstant(lsb, MVT::i32));
+ DAG.getConstant(amt, MVT::i32));
Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
DAG.getConstant(Mask, MVT::i32));
// Do not add new nodes to DAG combiner worklist.
DCI.CombineTo(N, Res, false);
return SDValue();
} else if (ARM::isBitFieldInvertedMask(~Mask) &&
- ARM::isBitFieldInvertedMask(Mask2) &&
- (CountPopulation_32(~Mask) == CountPopulation_32(Mask2))) {
+ (~Mask == Mask2)) {
// The pack halfword instruction works better for masks that fit it,
// so use that when it's available.
if (Subtarget->hasT2ExtractPack() &&
Res = DAG.getNode(ISD::SRL, DL, VT, N00,
DAG.getConstant(lsb, MVT::i32));
Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
- DAG.getConstant(Mask2, MVT::i32));
+ DAG.getConstant(Mask2, MVT::i32));
// Do not add new nodes to DAG combiner worklist.
DCI.CombineTo(N, Res, false);
return SDValue();
SDValue InDouble = N->getOperand(0);
if (InDouble.getOpcode() == ARMISD::VMOVDRR)
return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
+
+ // vmovrrd(load f64) -> (load i32), (load i32)
+ SDNode *InNode = InDouble.getNode();
+ if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
+ InNode->getValueType(0) == MVT::f64 &&
+ InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
+ !cast<LoadSDNode>(InNode)->isVolatile()) {
+ // TODO: Should this be done for non-FrameIndex operands?
+ LoadSDNode *LD = cast<LoadSDNode>(InNode);
+
+ SelectionDAG &DAG = DCI.DAG;
+ DebugLoc DL = LD->getDebugLoc();
+ SDValue BasePtr = LD->getBasePtr();
+ SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
+ LD->getPointerInfo(), LD->isVolatile(),
+ LD->isNonTemporal(), LD->getAlignment());
+
+ SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
+ DAG.getConstant(4, MVT::i32));
+ SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
+ LD->getPointerInfo(), LD->isVolatile(),
+ LD->isNonTemporal(),
+ std::min(4U, LD->getAlignment() / 2));
+
+ DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
+ SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
+ DCI.RemoveFromWorklist(LD);
+ DAG.DeleteNode(LD);
+ return Result;
+ }
+
return SDValue();
}
// Otherwise, the i64 value will be legalized to a pair of i32 values.
StoreSDNode *St = cast<StoreSDNode>(N);
SDValue StVal = St->getValue();
- if (!ISD::isNormalStore(St) || St->isVolatile() ||
- StVal.getValueType() != MVT::i64 ||
+ if (!ISD::isNormalStore(St) || St->isVolatile())
+ return SDValue();
+
+ if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
+ StVal.getNode()->hasOneUse() && !St->isVolatile()) {
+ SelectionDAG &DAG = DCI.DAG;
+ DebugLoc DL = St->getDebugLoc();
+ SDValue BasePtr = St->getBasePtr();
+ SDValue NewST1 = DAG.getStore(St->getChain(), DL,
+ StVal.getNode()->getOperand(0), BasePtr,
+ St->getPointerInfo(), St->isVolatile(),
+ St->isNonTemporal(), St->getAlignment());
+
+ SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
+ DAG.getConstant(4, MVT::i32));
+ return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
+ OffsetPtr, St->getPointerInfo(), St->isVolatile(),
+ St->isNonTemporal(),
+ std::min(4U, St->getAlignment() / 2));
+ }
+
+ if (StVal.getValueType() != MVT::i64 ||
StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
return SDValue();
DAG.getUNDEF(VT), NewMask.data());
}
+/// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
+/// NEON load/store intrinsics to merge base address updates.
+static SDValue CombineBaseUpdate(SDNode *N,
+ TargetLowering::DAGCombinerInfo &DCI) {
+ if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
+ return SDValue();
+
+ SelectionDAG &DAG = DCI.DAG;
+ bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
+ N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
+ unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
+ SDValue Addr = N->getOperand(AddrOpIdx);
+
+ // Search for a use of the address operand that is an increment.
+ for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
+ UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
+ SDNode *User = *UI;
+ if (User->getOpcode() != ISD::ADD ||
+ UI.getUse().getResNo() != Addr.getResNo())
+ continue;
+
+ // Check that the add is independent of the load/store. Otherwise, folding
+ // it would create a cycle.
+ if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
+ continue;
+
+ // Find the new opcode for the updating load/store.
+ bool isLoad = true;
+ bool isLaneOp = false;
+ unsigned NewOpc = 0;
+ unsigned NumVecs = 0;
+ if (isIntrinsic) {
+ unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
+ switch (IntNo) {
+ default: assert(0 && "unexpected intrinsic for Neon base update");
+ case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
+ NumVecs = 1; break;
+ case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
+ NumVecs = 2; break;
+ case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
+ NumVecs = 3; break;
+ case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
+ NumVecs = 4; break;
+ case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
+ NumVecs = 2; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
+ NumVecs = 3; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
+ NumVecs = 4; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
+ NumVecs = 1; isLoad = false; break;
+ case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
+ NumVecs = 2; isLoad = false; break;
+ case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
+ NumVecs = 3; isLoad = false; break;
+ case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
+ NumVecs = 4; isLoad = false; break;
+ case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
+ NumVecs = 2; isLoad = false; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
+ NumVecs = 3; isLoad = false; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
+ NumVecs = 4; isLoad = false; isLaneOp = true; break;
+ }
+ } else {
+ isLaneOp = true;
+ switch (N->getOpcode()) {
+ default: assert(0 && "unexpected opcode for Neon base update");
+ case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
+ case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
+ case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
+ }
+ }
+
+ // Find the size of memory referenced by the load/store.
+ EVT VecTy;
+ if (isLoad)
+ VecTy = N->getValueType(0);
+ else
+ VecTy = N->getOperand(AddrOpIdx+1).getValueType();
+ unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
+ if (isLaneOp)
+ NumBytes /= VecTy.getVectorNumElements();
+
+ // If the increment is a constant, it must match the memory ref size.
+ SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
+ if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
+ uint64_t IncVal = CInc->getZExtValue();
+ if (IncVal != NumBytes)
+ continue;
+ } else if (NumBytes >= 3 * 16) {
+ // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
+ // separate instructions that make it harder to use a non-constant update.
+ continue;
+ }
+
+ // Create the new updating load/store node.
+ EVT Tys[6];
+ unsigned NumResultVecs = (isLoad ? NumVecs : 0);
+ unsigned n;
+ for (n = 0; n < NumResultVecs; ++n)
+ Tys[n] = VecTy;
+ Tys[n++] = MVT::i32;
+ Tys[n] = MVT::Other;
+ SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
+ SmallVector<SDValue, 8> Ops;
+ Ops.push_back(N->getOperand(0)); // incoming chain
+ Ops.push_back(N->getOperand(AddrOpIdx));
+ Ops.push_back(Inc);
+ for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
+ Ops.push_back(N->getOperand(i));
+ }
+ MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
+ SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
+ Ops.data(), Ops.size(),
+ MemInt->getMemoryVT(),
+ MemInt->getMemOperand());
+
+ // Update the uses.
+ std::vector<SDValue> NewResults;
+ for (unsigned i = 0; i < NumResultVecs; ++i) {
+ NewResults.push_back(SDValue(UpdN.getNode(), i));
+ }
+ NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
+ DCI.CombineTo(N, NewResults);
+ DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
+
+ break;
+ }
+ return SDValue();
+}
+
/// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
/// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
/// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
case ISD::ZERO_EXTEND:
case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
+ case ARMISD::VLD2DUP:
+ case ARMISD::VLD3DUP:
+ case ARMISD::VLD4DUP:
+ return CombineBaseUpdate(N, DCI);
+ case ISD::INTRINSIC_VOID:
+ case ISD::INTRINSIC_W_CHAIN:
+ switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
+ case Intrinsic::arm_neon_vld1:
+ case Intrinsic::arm_neon_vld2:
+ case Intrinsic::arm_neon_vld3:
+ case Intrinsic::arm_neon_vld4:
+ case Intrinsic::arm_neon_vld2lane:
+ case Intrinsic::arm_neon_vld3lane:
+ case Intrinsic::arm_neon_vld4lane:
+ case Intrinsic::arm_neon_vst1:
+ case Intrinsic::arm_neon_vst2:
+ case Intrinsic::arm_neon_vst3:
+ case Intrinsic::arm_neon_vst4:
+ case Intrinsic::arm_neon_vst2lane:
+ case Intrinsic::arm_neon_vst3lane:
+ case Intrinsic::arm_neon_vst4lane:
+ return CombineBaseUpdate(N, DCI);
+ default: break;
+ }
+ break;
}
return SDValue();
}
+bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
+ EVT VT) const {
+ return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
+}
+
bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
if (!Subtarget->allowsUnalignedMem())
return false;