#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
#include "llvm/IR/CallingConv.h"
+#include "llvm/Support/MathExtras.h"
using namespace llvm;
setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
+ // i128 shift operation support
+ setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
+ setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
+ setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
+
// This prevents LLVM trying to compress double constants into a floating
// constant-pool entry and trying to load from there. It's of doubtful benefit
// for A64: we'd need LDR followed by FCVT, I believe.
setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Legal);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Legal);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i8, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i16, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
+
setOperationAction(ISD::SETCC, MVT::v8i8, Custom);
setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
setOperationAction(ISD::SETCC, MVT::v4i16, Custom);
setOperationAction(ISD::FP_TO_UINT, MVT::v2i32, Custom);
setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Custom);
+ // Neon does not support vector divide/remainder operations except
+ // floating-point divide.
+ setOperationAction(ISD::SDIV, MVT::v1i8, Expand);
+ setOperationAction(ISD::SDIV, MVT::v8i8, Expand);
+ setOperationAction(ISD::SDIV, MVT::v16i8, Expand);
+ setOperationAction(ISD::SDIV, MVT::v1i16, Expand);
+ setOperationAction(ISD::SDIV, MVT::v4i16, Expand);
+ setOperationAction(ISD::SDIV, MVT::v8i16, Expand);
+ setOperationAction(ISD::SDIV, MVT::v1i32, Expand);
+ setOperationAction(ISD::SDIV, MVT::v2i32, Expand);
+ setOperationAction(ISD::SDIV, MVT::v4i32, Expand);
+ setOperationAction(ISD::SDIV, MVT::v1i64, Expand);
+ setOperationAction(ISD::SDIV, MVT::v2i64, Expand);
+
+ setOperationAction(ISD::UDIV, MVT::v1i8, Expand);
+ setOperationAction(ISD::UDIV, MVT::v8i8, Expand);
+ setOperationAction(ISD::UDIV, MVT::v16i8, Expand);
+ setOperationAction(ISD::UDIV, MVT::v1i16, Expand);
+ setOperationAction(ISD::UDIV, MVT::v4i16, Expand);
+ setOperationAction(ISD::UDIV, MVT::v8i16, Expand);
+ setOperationAction(ISD::UDIV, MVT::v1i32, Expand);
+ setOperationAction(ISD::UDIV, MVT::v2i32, Expand);
+ setOperationAction(ISD::UDIV, MVT::v4i32, Expand);
+ setOperationAction(ISD::UDIV, MVT::v1i64, Expand);
+ setOperationAction(ISD::UDIV, MVT::v2i64, Expand);
+
+ setOperationAction(ISD::SREM, MVT::v1i8, Expand);
+ setOperationAction(ISD::SREM, MVT::v8i8, Expand);
+ setOperationAction(ISD::SREM, MVT::v16i8, Expand);
+ setOperationAction(ISD::SREM, MVT::v1i16, Expand);
+ setOperationAction(ISD::SREM, MVT::v4i16, Expand);
+ setOperationAction(ISD::SREM, MVT::v8i16, Expand);
+ setOperationAction(ISD::SREM, MVT::v1i32, Expand);
+ setOperationAction(ISD::SREM, MVT::v2i32, Expand);
+ setOperationAction(ISD::SREM, MVT::v4i32, Expand);
+ setOperationAction(ISD::SREM, MVT::v1i64, Expand);
+ setOperationAction(ISD::SREM, MVT::v2i64, Expand);
+
+ setOperationAction(ISD::UREM, MVT::v1i8, Expand);
+ setOperationAction(ISD::UREM, MVT::v8i8, Expand);
+ setOperationAction(ISD::UREM, MVT::v16i8, Expand);
+ setOperationAction(ISD::UREM, MVT::v1i16, Expand);
+ setOperationAction(ISD::UREM, MVT::v4i16, Expand);
+ setOperationAction(ISD::UREM, MVT::v8i16, Expand);
+ setOperationAction(ISD::UREM, MVT::v1i32, Expand);
+ setOperationAction(ISD::UREM, MVT::v2i32, Expand);
+ setOperationAction(ISD::UREM, MVT::v4i32, Expand);
+ setOperationAction(ISD::UREM, MVT::v1i64, Expand);
+ setOperationAction(ISD::UREM, MVT::v2i64, Expand);
+
+ setOperationAction(ISD::FREM, MVT::v2f32, Expand);
+ setOperationAction(ISD::FREM, MVT::v4f32, Expand);
+ setOperationAction(ISD::FREM, MVT::v1f64, Expand);
+ setOperationAction(ISD::FREM, MVT::v2f64, Expand);
+
+ setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
+ setOperationAction(ISD::SELECT, MVT::v16i8, Expand);
+ setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
+ setOperationAction(ISD::SELECT, MVT::v8i16, Expand);
+ setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
+ setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
+ setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
+ setOperationAction(ISD::SELECT, MVT::v2i64, Expand);
+ setOperationAction(ISD::SELECT, MVT::v2f32, Expand);
+ setOperationAction(ISD::SELECT, MVT::v4f32, Expand);
+ setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
+ setOperationAction(ISD::SELECT, MVT::v2f64, Expand);
+
+ setOperationAction(ISD::SELECT_CC, MVT::v8i8, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v16i8, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v4i16, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v8i16, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v2i32, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v4i32, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v1i64, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v2i64, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v2f32, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v4f32, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v1f64, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::v2f64, Custom);
+
// Vector ExtLoad and TruncStore are expanded.
for (unsigned I = MVT::FIRST_VECTOR_VALUETYPE;
I <= MVT::LAST_VECTOR_VALUETYPE; ++I) {
// do such optimization in the future.
setOperationAction(ISD::MUL, MVT::v1i64, Expand);
setOperationAction(ISD::MUL, MVT::v2i64, Expand);
+
+ setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
+ setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
+ setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
+ setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
+ setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
+ setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
+ setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
+ setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
+ setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
}
+
+ setTargetDAGCombine(ISD::SIGN_EXTEND);
+ setTargetDAGCombine(ISD::VSELECT);
}
EVT AArch64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
// Transfer the remainder of BB and its successor edges to exitMBB.
exitMBB->splice(exitMBB->begin(), BB,
- llvm::next(MachineBasicBlock::iterator(MI)),
- BB->end());
+ std::next(MachineBasicBlock::iterator(MI)), BB->end());
exitMBB->transferSuccessorsAndUpdatePHIs(BB);
const TargetRegisterClass *TRC
// Transfer the remainder of BB and its successor edges to exitMBB.
exitMBB->splice(exitMBB->begin(), BB,
- llvm::next(MachineBasicBlock::iterator(MI)),
- BB->end());
+ std::next(MachineBasicBlock::iterator(MI)), BB->end());
exitMBB->transferSuccessorsAndUpdatePHIs(BB);
unsigned scratch = MRI.createVirtualRegister(TRC);
// Transfer the remainder of BB and its successor edges to exitMBB.
exitMBB->splice(exitMBB->begin(), BB,
- llvm::next(MachineBasicBlock::iterator(MI)),
- BB->end());
+ std::next(MachineBasicBlock::iterator(MI)), BB->end());
exitMBB->transferSuccessorsAndUpdatePHIs(BB);
// thisMBB:
MF->insert(It, EndBB);
// Transfer rest of current basic-block to EndBB
- EndBB->splice(EndBB->begin(), MBB,
- llvm::next(MachineBasicBlock::iterator(MI)),
+ EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
MBB->end());
EndBB->transferSuccessorsAndUpdatePHIs(MBB);
}
}
-static const uint16_t AArch64FPRArgRegs[] = {
+static const MCPhysReg AArch64FPRArgRegs[] = {
AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7
};
static const unsigned NumFPRArgRegs = llvm::array_lengthof(AArch64FPRArgRegs);
-static const uint16_t AArch64ArgRegs[] = {
+static const MCPhysReg AArch64ArgRegs[] = {
AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3,
AArch64::X4, AArch64::X5, AArch64::X6, AArch64::X7
};
FuncInfo->setVariadicFPRSize(FPRSaveSize);
}
- int StackIdx = MFI->CreateFixedObject(8, CCInfo.getNextStackOffset(), true);
+ unsigned StackOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), 8);
+ int StackIdx = MFI->CreateFixedObject(8, StackOffset, true);
FuncInfo->setVariadicStackIdx(StackIdx);
FuncInfo->setVariadicGPRIdx(GPRIdx);
int Size = Flags.getByValSize();
unsigned NumRegs = (Size + 7) / 8;
+ uint32_t BEAlign = 0;
+ if (Size < 8 && !getSubtarget()->isLittle())
+ BEAlign = 8-Size;
unsigned FrameIdx = MFI->CreateFixedObject(8 * NumRegs,
- VA.getLocMemOffset(),
+ VA.getLocMemOffset() + BEAlign,
false);
SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
InVals.push_back(FrameIdxN);
// loaded before this eventual operation. Otherwise they'll be clobbered.
Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
} else {
- SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset());
+ uint32_t OpSize = Flags.isByVal() ? Flags.getByValSize()*8 :
+ VA.getLocVT().getSizeInBits();
+ OpSize = (OpSize + 7) / 8;
+ uint32_t BEAlign = 0;
+ if (OpSize < 8 && !getSubtarget()->isLittle())
+ BEAlign = 8-OpSize;
+ SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset() + BEAlign);
DstAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
DstInfo = MachinePointerInfo::getStack(VA.getLocMemOffset());
Flags.getByValAlign(),
/*isVolatile = */ false,
/*alwaysInline = */ false,
- DstInfo, MachinePointerInfo(0));
+ DstInfo, MachinePointerInfo());
MemOpChains.push_back(Cpy);
} else {
// Normal stack argument, put it where it's needed.
return DAG.getNode(Opc, dl, VT, Vec);
}
+static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
+ // We custom lower concat_vectors with 4, 8, or 16 operands that are all the
+ // same operand and of type v1* using the DUP instruction.
+ unsigned NumOps = Op->getNumOperands();
+ if (NumOps == 2) {
+ assert(Op.getValueType().getSizeInBits() == 128 && "unexpected concat");
+ return Op;
+ }
+
+ if (NumOps != 4 && NumOps != 8 && NumOps != 16)
+ return SDValue();
+
+ // Must be a single value for VDUP.
+ SDValue Op0 = Op.getOperand(0);
+ for (unsigned i = 1; i < NumOps; ++i) {
+ SDValue OpN = Op.getOperand(i);
+ if (Op0 != OpN)
+ return SDValue();
+ }
+
+ // Verify the value type.
+ EVT EltVT = Op0.getValueType();
+ switch (NumOps) {
+ default: llvm_unreachable("Unexpected number of operands");
+ case 4:
+ if (EltVT != MVT::v1i16 && EltVT != MVT::v1i32)
+ return SDValue();
+ break;
+ case 8:
+ if (EltVT != MVT::v1i8 && EltVT != MVT::v1i16)
+ return SDValue();
+ break;
+ case 16:
+ if (EltVT != MVT::v1i8)
+ return SDValue();
+ break;
+ }
+
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ // VDUP produces better code for constants.
+ if (Op0->getOpcode() == ISD::BUILD_VECTOR)
+ return DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Op0->getOperand(0));
+ return DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, Op0,
+ DAG.getConstant(0, MVT::i64));
+}
+
SDValue
AArch64TargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
bool IsSigned) const {
}
}
-// (SELECT_CC lhs, rhs, iftrue, iffalse, condcode)
-SDValue
-AArch64TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
- SDLoc dl(Op);
- SDValue LHS = Op.getOperand(0);
- SDValue RHS = Op.getOperand(1);
- SDValue IfTrue = Op.getOperand(2);
- SDValue IfFalse = Op.getOperand(3);
- ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
-
- if (LHS.getValueType() == MVT::f128) {
- // f128 comparisons are lowered to libcalls, but slot in nicely here
- // afterwards.
- softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
-
- // If softenSetCCOperands returned a scalar, we need to compare the result
- // against zero to select between true and false values.
- if (RHS.getNode() == 0) {
- RHS = DAG.getConstant(0, LHS.getValueType());
- CC = ISD::SETNE;
- }
- }
-
- if (LHS.getValueType().isInteger()) {
- SDValue A64cc;
-
- // Integers are handled in a separate function because the combinations of
- // immediates and tests can get hairy and we may want to fiddle things.
- SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
-
- return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
- CmpOp, IfTrue, IfFalse, A64cc);
- }
-
- // Note that some LLVM floating-point CondCodes can't be lowered to a single
- // conditional branch, hence FPCCToA64CC can set a second test, where either
- // passing is sufficient.
- A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
- CondCode = FPCCToA64CC(CC, Alternative);
- SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
- SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
- DAG.getCondCode(CC));
- SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl,
- Op.getValueType(),
- SetCC, IfTrue, IfFalse, A64cc);
-
- if (Alternative != A64CC::Invalid) {
- A64cc = DAG.getConstant(Alternative, MVT::i32);
- A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
- SetCC, IfTrue, A64SELECT_CC, A64cc);
-
- }
-
- return A64SELECT_CC;
-}
-
// (SELECT testbit, iftrue, iffalse)
SDValue
AArch64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
return A64SELECT_CC;
}
+static SDValue LowerVectorSELECT_CC(SDValue Op, SelectionDAG &DAG) {
+ SDLoc dl(Op);
+ SDValue LHS = Op.getOperand(0);
+ SDValue RHS = Op.getOperand(1);
+ SDValue IfTrue = Op.getOperand(2);
+ SDValue IfFalse = Op.getOperand(3);
+ EVT IfTrueVT = IfTrue.getValueType();
+ EVT CondVT = IfTrueVT.changeVectorElementTypeToInteger();
+ ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
+
+ // If LHS & RHS are floating point and IfTrue & IfFalse are vectors, we will
+ // use NEON compare.
+ if ((LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64)) {
+ EVT EltVT = LHS.getValueType();
+ unsigned EltNum = 128 / EltVT.getSizeInBits();
+ EVT VT = EVT::getVectorVT(*DAG.getContext(), EltVT, EltNum);
+ unsigned SubConstant =
+ (LHS.getValueType() == MVT::f32) ? AArch64::sub_32 :AArch64::sub_64;
+ EVT CEltT = (LHS.getValueType() == MVT::f32) ? MVT::i32 : MVT::i64;
+ EVT CVT = EVT::getVectorVT(*DAG.getContext(), CEltT, EltNum);
+
+ LHS
+ = SDValue(DAG.getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
+ VT, DAG.getTargetConstant(0, MVT::i32), LHS,
+ DAG.getTargetConstant(SubConstant, MVT::i32)), 0);
+ RHS
+ = SDValue(DAG.getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
+ VT, DAG.getTargetConstant(0, MVT::i32), RHS,
+ DAG.getTargetConstant(SubConstant, MVT::i32)), 0);
+
+ SDValue VSetCC = DAG.getSetCC(dl, CVT, LHS, RHS, CC);
+ SDValue ResCC = LowerVectorSETCC(VSetCC, DAG);
+ if (CEltT.getSizeInBits() < IfTrueVT.getSizeInBits()) {
+ EVT DUPVT =
+ EVT::getVectorVT(*DAG.getContext(), CEltT,
+ IfTrueVT.getSizeInBits() / CEltT.getSizeInBits());
+ ResCC = DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, DUPVT, ResCC,
+ DAG.getConstant(0, MVT::i64, false));
+
+ ResCC = DAG.getNode(ISD::BITCAST, dl, CondVT, ResCC);
+ } else {
+ // FIXME: If IfTrue & IfFalse hold v1i8, v1i16 or v1i32, this function
+ // can't handle them and will hit this assert.
+ assert(CEltT.getSizeInBits() == IfTrueVT.getSizeInBits() &&
+ "Vector of IfTrue & IfFalse is too small.");
+
+ unsigned ExEltNum =
+ EltNum * IfTrueVT.getSizeInBits() / ResCC.getValueSizeInBits();
+ EVT ExVT = EVT::getVectorVT(*DAG.getContext(), CEltT, ExEltNum);
+ ResCC = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ExVT, ResCC,
+ DAG.getConstant(0, MVT::i64, false));
+ ResCC = DAG.getNode(ISD::BITCAST, dl, CondVT, ResCC);
+ }
+ SDValue VSelect = DAG.getNode(ISD::VSELECT, dl, IfTrue.getValueType(),
+ ResCC, IfTrue, IfFalse);
+ return VSelect;
+ }
+
+ // Here we handle the case that LHS & RHS are integer and IfTrue & IfFalse are
+ // vectors.
+ A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
+ CondCode = FPCCToA64CC(CC, Alternative);
+ SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
+ SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
+ DAG.getCondCode(CC));
+ EVT SEVT = MVT::i32;
+ if (IfTrue.getValueType().getVectorElementType().getSizeInBits() > 32)
+ SEVT = MVT::i64;
+ SDValue AllOne = DAG.getConstant(-1, SEVT);
+ SDValue AllZero = DAG.getConstant(0, SEVT);
+ SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, SEVT, SetCC,
+ AllOne, AllZero, A64cc);
+
+ if (Alternative != A64CC::Invalid) {
+ A64cc = DAG.getConstant(Alternative, MVT::i32);
+ A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
+ SetCC, AllOne, A64SELECT_CC, A64cc);
+ }
+ SDValue VDup;
+ if (IfTrue.getValueType().getVectorNumElements() == 1)
+ VDup = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, CondVT, A64SELECT_CC);
+ else
+ VDup = DAG.getNode(AArch64ISD::NEON_VDUP, dl, CondVT, A64SELECT_CC);
+ SDValue VSelect = DAG.getNode(ISD::VSELECT, dl, IfTrue.getValueType(),
+ VDup, IfTrue, IfFalse);
+ return VSelect;
+}
+
+// (SELECT_CC lhs, rhs, iftrue, iffalse, condcode)
+SDValue
+AArch64TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
+ SDLoc dl(Op);
+ SDValue LHS = Op.getOperand(0);
+ SDValue RHS = Op.getOperand(1);
+ SDValue IfTrue = Op.getOperand(2);
+ SDValue IfFalse = Op.getOperand(3);
+ ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
+
+ if (IfTrue.getValueType().isVector())
+ return LowerVectorSELECT_CC(Op, DAG);
+
+ if (LHS.getValueType() == MVT::f128) {
+ // f128 comparisons are lowered to libcalls, but slot in nicely here
+ // afterwards.
+ softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
+
+ // If softenSetCCOperands returned a scalar, we need to compare the result
+ // against zero to select between true and false values.
+ if (RHS.getNode() == 0) {
+ RHS = DAG.getConstant(0, LHS.getValueType());
+ CC = ISD::SETNE;
+ }
+ }
+
+ if (LHS.getValueType().isInteger()) {
+ SDValue A64cc;
+
+ // Integers are handled in a separate function because the combinations of
+ // immediates and tests can get hairy and we may want to fiddle things.
+ SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
+
+ return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(), CmpOp,
+ IfTrue, IfFalse, A64cc);
+ }
+
+ // Note that some LLVM floating-point CondCodes can't be lowered to a single
+ // conditional branch, hence FPCCToA64CC can set a second test, where either
+ // passing is sufficient.
+ A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
+ CondCode = FPCCToA64CC(CC, Alternative);
+ SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
+ SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
+ DAG.getCondCode(CC));
+ SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl,
+ Op.getValueType(),
+ SetCC, IfTrue, IfFalse, A64cc);
+
+ if (Alternative != A64CC::Invalid) {
+ A64cc = DAG.getConstant(Alternative, MVT::i32);
+ A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
+ SetCC, IfTrue, A64SELECT_CC, A64cc);
+
+ }
+
+ return A64SELECT_CC;
+}
+
SDValue
AArch64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
- const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
+ const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
// We have to make sure we copy the entire structure: 8+8+8+4+4 = 32 bytes
// rather than just 8.
case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
+ case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
+ case ISD::SRL_PARTS:
+ case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
+
case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
case ISD::BRCOND: return LowerBRCOND(Op, DAG);
case ISD::BR_CC: return LowerBR_CC(Op, DAG);
case ISD::VASTART: return LowerVASTART(Op, DAG);
case ISD::BUILD_VECTOR:
return LowerBUILD_VECTOR(Op, DAG, getSubtarget());
+ case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
}
Tys[n] = VecTy;
Tys[n++] = MVT::i64;
Tys[n] = MVT::Other;
- SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs + 2);
+ SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs + 2));
SmallVector<SDValue, 8> Ops;
Ops.push_back(N->getOperand(0)); // incoming chain
Ops.push_back(N->getOperand(AddrOpIdx));
for (n = 0; n < NumVecs; ++n)
Tys[n] = VT;
Tys[n] = MVT::Other;
- SDVTList SDTys = DAG.getVTList(Tys, NumVecs + 1);
+ SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumVecs + 1));
SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys, Ops, 2,
return SDValue(N, 0);
}
+// vselect (v1i1 setcc) ->
+// vselect (v1iXX setcc) (XX is the size of the compared operand type)
+// FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
+// condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
+// such VSELECT.
+static SDValue PerformVSelectCombine(SDNode *N, SelectionDAG &DAG) {
+ SDValue N0 = N->getOperand(0);
+ EVT CCVT = N0.getValueType();
+
+ if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
+ CCVT.getVectorElementType() != MVT::i1)
+ return SDValue();
+
+ EVT ResVT = N->getValueType(0);
+ EVT CmpVT = N0.getOperand(0).getValueType();
+ // Only combine when the result type is of the same size as the compared
+ // operands.
+ if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
+ return SDValue();
+
+ SDValue IfTrue = N->getOperand(1);
+ SDValue IfFalse = N->getOperand(2);
+ SDValue SetCC =
+ DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
+ N0.getOperand(0), N0.getOperand(1),
+ cast<CondCodeSDNode>(N0.getOperand(2))->get());
+ return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
+ IfTrue, IfFalse);
+}
+
+// sign_extend (extract_vector_elt (v1i1 setcc)) ->
+// extract_vector_elt (v1iXX setcc)
+// (XX is the size of the compared operand type)
+static SDValue PerformSignExtendCombine(SDNode *N, SelectionDAG &DAG) {
+ SDValue N0 = N->getOperand(0);
+ SDValue Vec = N0.getOperand(0);
+
+ if (N0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+ Vec.getOpcode() != ISD::SETCC)
+ return SDValue();
+
+ EVT ResVT = N->getValueType(0);
+ EVT CmpVT = Vec.getOperand(0).getValueType();
+ // Only optimize when the result type is of the same size as the element
+ // type of the compared operand.
+ if (ResVT.getSizeInBits() != CmpVT.getVectorElementType().getSizeInBits())
+ return SDValue();
+
+ SDValue Lane = N0.getOperand(1);
+ SDValue SetCC =
+ DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
+ Vec.getOperand(0), Vec.getOperand(1),
+ cast<CondCodeSDNode>(Vec.getOperand(2))->get());
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), ResVT,
+ SetCC, Lane);
+}
+
SDValue
AArch64TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
case ISD::SRA:
case ISD::SRL:
return PerformShiftCombine(N, DCI, getSubtarget());
+ case ISD::VSELECT: return PerformVSelectCombine(N, DCI.DAG);
+ case ISD::SIGN_EXTEND: return PerformSignExtendCombine(N, DCI.DAG);
case ISD::INTRINSIC_WO_CHAIN:
return PerformIntrinsicCombine(N, DCI.DAG);
case AArch64ISD::NEON_VDUPLANE:
return false;
}
-// Check whether a Build Vector could be presented as Shuffle Vector. If yes,
-// try to call LowerVECTOR_SHUFFLE to lower it.
+bool AArch64TargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
+ unsigned AddrSpace,
+ bool *Fast) const {
+ const AArch64Subtarget *Subtarget = getSubtarget();
+ // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
+ bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
+
+ switch (VT.getSimpleVT().SimpleTy) {
+ default:
+ return false;
+ // Scalar types
+ case MVT::i8: case MVT::i16:
+ case MVT::i32: case MVT::i64:
+ case MVT::f32: case MVT::f64: {
+ // Unaligned access can use (for example) LRDB, LRDH, LDRW
+ if (AllowsUnaligned) {
+ if (Fast)
+ *Fast = true;
+ return true;
+ }
+ return false;
+ }
+ // 64-bit vector types
+ case MVT::v8i8: case MVT::v4i16:
+ case MVT::v2i32: case MVT::v1i64:
+ case MVT::v2f32: case MVT::v1f64:
+ // 128-bit vector types
+ case MVT::v16i8: case MVT::v8i16:
+ case MVT::v4i32: case MVT::v2i64:
+ case MVT::v4f32: case MVT::v2f64: {
+ // For any little-endian targets with neon, we can support unaligned
+ // load/store of V registers using ld1/st1.
+ // A big-endian target may also explicitly support unaligned accesses
+ if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
+ if (Fast)
+ *Fast = true;
+ return true;
+ }
+ return false;
+ }
+ }
+}
+
+// Check whether a shuffle_vector could be presented as concat_vector.
+bool AArch64TargetLowering::isConcatVector(SDValue Op, SelectionDAG &DAG,
+ SDValue V0, SDValue V1,
+ const int *Mask,
+ SDValue &Res) const {
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ if (VT.getSizeInBits() != 128)
+ return false;
+ if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
+ VT.getVectorElementType() != V1.getValueType().getVectorElementType())
+ return false;
+
+ unsigned NumElts = VT.getVectorNumElements();
+ bool isContactVector = true;
+ bool splitV0 = false;
+ if (V0.getValueType().getSizeInBits() == 128)
+ splitV0 = true;
+
+ for (int I = 0, E = NumElts / 2; I != E; I++) {
+ if (Mask[I] != I) {
+ isContactVector = false;
+ break;
+ }
+ }
+
+ if (isContactVector) {
+ int offset = NumElts / 2;
+ for (int I = NumElts / 2, E = NumElts; I != E; I++) {
+ if (Mask[I] != I + splitV0 * offset) {
+ isContactVector = false;
+ break;
+ }
+ }
+ }
+
+ if (isContactVector) {
+ EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
+ NumElts / 2);
+ if (splitV0) {
+ V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
+ DAG.getConstant(0, MVT::i64));
+ }
+ if (V1.getValueType().getSizeInBits() == 128) {
+ V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
+ DAG.getConstant(0, MVT::i64));
+ }
+ Res = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
+ return true;
+ }
+ return false;
+}
+
+// Check whether a Build Vector could be presented as Shuffle Vector.
+// This Shuffle Vector maybe not legalized, so the length of its operand and
+// the length of result may not equal.
bool AArch64TargetLowering::isKnownShuffleVector(SDValue Op, SelectionDAG &DAG,
- SDValue &Res) const {
+ SDValue &V0, SDValue &V1,
+ int *Mask) const {
SDLoc DL(Op);
EVT VT = Op.getValueType();
unsigned NumElts = VT.getVectorNumElements();
unsigned V0NumElts = 0;
- int Mask[16];
- SDValue V0, V1;
// Check if all elements are extracted from less than 3 vectors.
for (unsigned i = 0; i < NumElts; ++i) {
SDValue Elt = Op.getOperand(i);
- if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
+ if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+ Elt.getOperand(0).getValueType().getVectorElementType() !=
+ VT.getVectorElementType())
return false;
if (V0.getNode() == 0) {
return false;
}
}
+ return true;
+}
- if (!V1.getNode() && V0NumElts == NumElts * 2) {
- V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
- DAG.getConstant(NumElts, MVT::i64));
- V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
- DAG.getConstant(0, MVT::i64));
- V0NumElts = V0.getValueType().getVectorNumElements();
- }
+// LowerShiftRightParts - Lower SRL_PARTS and SRA_PARTS, which returns two
+/// i64 values and take a 2 x i64 value to shift plus a shift amount.
+SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
+ SelectionDAG &DAG) const {
+ assert(Op.getNumOperands() == 3 && "Not a quad-shift!");
+ EVT VT = Op.getValueType();
+ unsigned VTBits = VT.getSizeInBits();
+ SDLoc dl(Op);
+ SDValue ShOpLo = Op.getOperand(0);
+ SDValue ShOpHi = Op.getOperand(1);
+ SDValue ShAmt = Op.getOperand(2);
+ unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
+
+ assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
+ SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
+ DAG.getConstant(VTBits, MVT::i64), ShAmt);
+ SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
+ SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
+ DAG.getConstant(VTBits, MVT::i64));
+ SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
+ SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
+ SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
+ SDValue Tmp3 = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
+
+ SDValue A64cc;
+ SDValue CmpOp = getSelectableIntSetCC(ExtraShAmt,
+ DAG.getConstant(0, MVT::i64),
+ ISD::SETGE, A64cc,
+ DAG, dl);
+
+ SDValue Hi = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
+ DAG.getConstant(0, Tmp3.getValueType()), Tmp3,
+ A64cc);
+ SDValue Lo = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
+ TrueVal, FalseVal, A64cc);
+
+ SDValue Ops[2] = { Lo, Hi };
+ return DAG.getMergeValues(Ops, 2, dl);
+}
- if (V1.getNode() && NumElts == V0NumElts &&
- V0NumElts == V1.getValueType().getVectorNumElements()) {
- SDValue Shuffle = DAG.getVectorShuffle(VT, DL, V0, V1, Mask);
- if(Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
- Res = Shuffle;
- else
- Res = LowerVECTOR_SHUFFLE(Shuffle, DAG);
- return true;
- } else
- return false;
+/// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
+/// i64 values and take a 2 x i64 value to shift plus a shift amount.
+SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
+ SelectionDAG &DAG) const {
+ assert(Op.getNumOperands() == 3 && "Not a quad-shift!");
+ EVT VT = Op.getValueType();
+ unsigned VTBits = VT.getSizeInBits();
+ SDLoc dl(Op);
+ SDValue ShOpLo = Op.getOperand(0);
+ SDValue ShOpHi = Op.getOperand(1);
+ SDValue ShAmt = Op.getOperand(2);
+
+ assert(Op.getOpcode() == ISD::SHL_PARTS);
+ SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
+ DAG.getConstant(VTBits, MVT::i64), ShAmt);
+ SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
+ SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
+ DAG.getConstant(VTBits, MVT::i64));
+ SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
+ SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
+ SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
+ SDValue Tmp4 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
+
+ SDValue A64cc;
+ SDValue CmpOp = getSelectableIntSetCC(ExtraShAmt,
+ DAG.getConstant(0, MVT::i64),
+ ISD::SETGE, A64cc,
+ DAG, dl);
+
+ SDValue Lo = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
+ DAG.getConstant(0, Tmp4.getValueType()), Tmp4,
+ A64cc);
+ SDValue Hi = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
+ Tmp3, FalseVal, A64cc);
+
+ SDValue Ops[2] = { Lo, Hi };
+ return DAG.getMergeValues(Ops, 2, dl);
}
// If this is a case we can't handle, return null and let the default
return SDValue();
// Try to lower this in lowering ShuffleVector way.
- SDValue Shuf;
- if (isKnownShuffleVector(Op, DAG, Shuf))
- return Shuf;
+ SDValue V0, V1;
+ int Mask[16];
+ if (isKnownShuffleVector(Op, DAG, V0, V1, Mask)) {
+ unsigned V0NumElts = V0.getValueType().getVectorNumElements();
+ if (!V1.getNode() && V0NumElts == NumElts * 2) {
+ V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
+ DAG.getConstant(NumElts, MVT::i64));
+ V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
+ DAG.getConstant(0, MVT::i64));
+ V0NumElts = V0.getValueType().getVectorNumElements();
+ }
+
+ if (V1.getNode() && NumElts == V0NumElts &&
+ V0NumElts == V1.getValueType().getVectorNumElements()) {
+ SDValue Shuffle = DAG.getVectorShuffle(VT, DL, V0, V1, Mask);
+ if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
+ return Shuffle;
+ else
+ return LowerVECTOR_SHUFFLE(Shuffle, DAG);
+ } else {
+ SDValue Res;
+ if (isConcatVector(Op, DAG, V0, V1, Mask, Res))
+ return Res;
+ }
+ }
// If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
// know the default expansion would otherwise fall back on something even
return DAG.getNode(ISDNo, dl, VT, V1, V2);
}
+ SDValue Res;
+ if (isConcatVector(Op, DAG, V1, V2, &ShuffleMask[0], Res))
+ return Res;
+
// If the element of shuffle mask are all the same constant, we can
// transform it into either NEON_VDUP or NEON_VDUPLANE
if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
// it into NEON_VEXTRACT.
if (V1EltNum == Length) {
// Check if the shuffle mask is sequential.
- bool IsSequential = true;
- int CurMask = ShuffleMask[0];
- for (int I = 0; I < Length; ++I) {
- if (ShuffleMask[I] != CurMask) {
- IsSequential = false;
- break;
- }
- CurMask++;
+ int SkipUndef = 0;
+ while (ShuffleMask[SkipUndef] == -1) {
+ SkipUndef++;
}
- if (IsSequential) {
- assert((EltSize % 8 == 0) && "Bitsize of vector element is incorrect");
- unsigned VecSize = EltSize * V1EltNum;
- unsigned Index = (EltSize/8) * ShuffleMask[0];
- if (VecSize == 64 || VecSize == 128)
- return DAG.getNode(AArch64ISD::NEON_VEXTRACT, dl, VT, V1, V2,
- DAG.getConstant(Index, MVT::i64));
+ int CurMask = ShuffleMask[SkipUndef];
+ if (CurMask >= SkipUndef) {
+ bool IsSequential = true;
+ for (int I = SkipUndef; I < Length; ++I) {
+ if (ShuffleMask[I] != -1 && ShuffleMask[I] != CurMask) {
+ IsSequential = false;
+ break;
+ }
+ CurMask++;
+ }
+ if (IsSequential) {
+ assert((EltSize % 8 == 0) && "Bitsize of vector element is incorrect");
+ unsigned VecSize = EltSize * V1EltNum;
+ unsigned Index = (EltSize / 8) * (ShuffleMask[SkipUndef] - SkipUndef);
+ if (VecSize == 64 || VecSize == 128)
+ return DAG.getNode(AArch64ISD::NEON_VEXTRACT, dl, VT, V1, V2,
+ DAG.getConstant(Index, MVT::i64));
+ }
}
}
return false;
}
+
+// Truncations from 64-bit GPR to 32-bit GPR is free.
+bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
+ if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
+ return false;
+ unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
+ unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
+ if (NumBits1 <= NumBits2)
+ return false;
+ return true;
+}
+
+bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
+ if (!VT1.isInteger() || !VT2.isInteger())
+ return false;
+ unsigned NumBits1 = VT1.getSizeInBits();
+ unsigned NumBits2 = VT2.getSizeInBits();
+ if (NumBits1 <= NumBits2)
+ return false;
+ return true;
+}
+
+// All 32-bit GPR operations implicitly zero the high-half of the corresponding
+// 64-bit GPR.
+bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
+ if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
+ return false;
+ unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
+ unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
+ if (NumBits1 == 32 && NumBits2 == 64)
+ return true;
+ return false;
+}
+
+bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
+ if (!VT1.isInteger() || !VT2.isInteger())
+ return false;
+ unsigned NumBits1 = VT1.getSizeInBits();
+ unsigned NumBits2 = VT2.getSizeInBits();
+ if (NumBits1 == 32 && NumBits2 == 64)
+ return true;
+ return false;
+}
+
+bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
+ EVT VT1 = Val.getValueType();
+ if (isZExtFree(VT1, VT2)) {
+ return true;
+ }
+
+ if (Val.getOpcode() != ISD::LOAD)
+ return false;
+
+ // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
+ return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
+ VT2.isInteger() && VT1.getSizeInBits() <= 32);
+}
+
+// isLegalAddressingMode - Return true if the addressing mode represented
+/// by AM is legal for this target, for a load/store of the specified type.
+bool AArch64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
+ Type *Ty) const {
+ // AArch64 has five basic addressing modes:
+ // reg
+ // reg + 9-bit signed offset
+ // reg + SIZE_IN_BYTES * 12-bit unsigned offset
+ // reg1 + reg2
+ // reg + SIZE_IN_BYTES * reg
+
+ // No global is ever allowed as a base.
+ if (AM.BaseGV)
+ return false;
+
+ // No reg+reg+imm addressing.
+ if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
+ return false;
+
+ // check reg + imm case:
+ // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
+ uint64_t NumBytes = 0;
+ if (Ty->isSized()) {
+ uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
+ NumBytes = NumBits / 8;
+ if (!isPowerOf2_64(NumBits))
+ NumBytes = 0;
+ }
+
+ if (!AM.Scale) {
+ int64_t Offset = AM.BaseOffs;
+
+ // 9-bit signed offset
+ if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
+ return true;
+
+ // 12-bit unsigned offset
+ unsigned shift = Log2_64(NumBytes);
+ if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
+ // Must be a multiple of NumBytes (NumBytes is a power of 2)
+ (Offset >> shift) << shift == Offset)
+ return true;
+ return false;
+ }
+ if (!AM.Scale || AM.Scale == 1 ||
+ (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
+ return true;
+ return false;
+}
+
+int AArch64TargetLowering::getScalingFactorCost(const AddrMode &AM,
+ Type *Ty) const {
+ // Scaling factors are not free at all.
+ // Operands | Rt Latency
+ // -------------------------------------------
+ // Rt, [Xn, Xm] | 4
+ // -------------------------------------------
+ // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
+ // Rt, [Xn, Wm, <extend> #imm] |
+ if (isLegalAddressingMode(AM, Ty))
+ // Scale represents reg2 * scale, thus account for 1 if
+ // it is not equal to 0 or 1.
+ return AM.Scale != 0 && AM.Scale != 1;
+ return -1;
+}