llvm_unreachable("unknown subtarget type");
}
-
AArch64TargetLowering::AArch64TargetLowering(AArch64TargetMachine &TM)
- : TargetLowering(TM, createTLOF(TM)),
- Subtarget(&TM.getSubtarget<AArch64Subtarget>()),
- RegInfo(TM.getRegisterInfo()),
- Itins(TM.getInstrItineraryData()) {
+ : TargetLowering(TM, createTLOF(TM)), Itins(TM.getInstrItineraryData()) {
// SIMD compares set the entire lane's bits to 1
setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
- // And the vectors
- addRegisterClass(MVT::v8i8, &AArch64::VPR64RegClass);
- addRegisterClass(MVT::v4i16, &AArch64::VPR64RegClass);
- addRegisterClass(MVT::v2i32, &AArch64::VPR64RegClass);
- addRegisterClass(MVT::v2f32, &AArch64::VPR64RegClass);
- addRegisterClass(MVT::v16i8, &AArch64::VPR128RegClass);
- addRegisterClass(MVT::v8i16, &AArch64::VPR128RegClass);
- addRegisterClass(MVT::v4i32, &AArch64::VPR128RegClass);
- addRegisterClass(MVT::v4f32, &AArch64::VPR128RegClass);
- addRegisterClass(MVT::v2f64, &AArch64::VPR128RegClass);
-
computeRegisterProperties();
- // Some atomic operations can be folded into load-acquire or store-release
- // instructions on AArch64. It's marginally simpler to let LLVM expand
- // everything out to a barrier and then recombine the (few) barriers we can.
- setInsertFencesForAtomic(true);
- setTargetDAGCombine(ISD::ATOMIC_FENCE);
- setTargetDAGCombine(ISD::ATOMIC_STORE);
-
// We combine OR nodes for bitfield and NEON BSL operations.
setTargetDAGCombine(ISD::OR);
setOperationAction(ISD::FSIN, MVT::f32, Expand);
setOperationAction(ISD::FSIN, MVT::f64, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
// Virtually no operation on f128 is legal, but LLVM can't expand them when
// there's a valid register class, so we need custom operations in most cases.
setOperationAction(ISD::FREM, MVT::f128, Expand);
setOperationAction(ISD::FRINT, MVT::f128, Expand);
setOperationAction(ISD::FSIN, MVT::f128, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
setOperationAction(ISD::FSQRT, MVT::f128, Expand);
setOperationAction(ISD::FSUB, MVT::f128, Custom);
setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
setExceptionSelectorRegister(AArch64::X1);
}
-EVT AArch64TargetLowering::getSetCCResultType(EVT VT) const {
+EVT AArch64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
// It's reasonably important that this value matches the "natural" legal
// promotion from i1 for scalar types. Otherwise LegalizeTypes can get itself
// in a twist (e.g. inserting an any_extend which then becomes i64 -> i64).
return VT.changeVectorElementTypeToInteger();
}
-static void getExclusiveOperation(unsigned Size, unsigned &ldrOpc,
- unsigned &strOpc) {
- switch (Size) {
- default: llvm_unreachable("unsupported size for atomic binary op!");
- case 1:
- ldrOpc = AArch64::LDXR_byte;
- strOpc = AArch64::STXR_byte;
- break;
- case 2:
- ldrOpc = AArch64::LDXR_hword;
- strOpc = AArch64::STXR_hword;
- break;
- case 4:
- ldrOpc = AArch64::LDXR_word;
- strOpc = AArch64::STXR_word;
- break;
- case 8:
- ldrOpc = AArch64::LDXR_dword;
- strOpc = AArch64::STXR_dword;
- break;
- }
+static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
+ unsigned &LdrOpc,
+ unsigned &StrOpc) {
+ static unsigned LoadBares[] = {AArch64::LDXR_byte, AArch64::LDXR_hword,
+ AArch64::LDXR_word, AArch64::LDXR_dword};
+ static unsigned LoadAcqs[] = {AArch64::LDAXR_byte, AArch64::LDAXR_hword,
+ AArch64::LDAXR_word, AArch64::LDAXR_dword};
+ static unsigned StoreBares[] = {AArch64::STXR_byte, AArch64::STXR_hword,
+ AArch64::STXR_word, AArch64::STXR_dword};
+ static unsigned StoreRels[] = {AArch64::STLXR_byte, AArch64::STLXR_hword,
+ AArch64::STLXR_word, AArch64::STLXR_dword};
+
+ unsigned *LoadOps, *StoreOps;
+ if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
+ LoadOps = LoadAcqs;
+ else
+ LoadOps = LoadBares;
+
+ if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
+ StoreOps = StoreRels;
+ else
+ StoreOps = StoreBares;
+
+ assert(isPowerOf2_32(Size) && Size <= 8 &&
+ "unsupported size for atomic binary op!");
+
+ LdrOpc = LoadOps[Log2_32(Size)];
+ StrOpc = StoreOps[Log2_32(Size)];
}
MachineBasicBlock *
unsigned dest = MI->getOperand(0).getReg();
unsigned ptr = MI->getOperand(1).getReg();
unsigned incr = MI->getOperand(2).getReg();
+ AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
DebugLoc dl = MI->getDebugLoc();
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
unsigned ldrOpc, strOpc;
- getExclusiveOperation(Size, ldrOpc, strOpc);
+ getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
// ldxr dest, ptr
// <binop> scratch, dest, incr
// stxr stxr_status, scratch, ptr
- // cmp stxr_status, #0
- // b.ne loopMBB
+ // cbnz stxr_status, loopMBB
// fallthrough --> exitMBB
BB = loopMBB;
BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(scratch).addReg(ptr);
- BuildMI(BB, dl, TII->get(AArch64::SUBwwi_lsl0_cmp))
- .addReg(stxr_status).addImm(0);
- BuildMI(BB, dl, TII->get(AArch64::Bcc))
- .addImm(A64CC::NE).addMBB(loopMBB);
+ BuildMI(BB, dl, TII->get(AArch64::CBNZw))
+ .addReg(stxr_status).addMBB(loopMBB);
BB->addSuccessor(loopMBB);
BB->addSuccessor(exitMBB);
unsigned dest = MI->getOperand(0).getReg();
unsigned ptr = MI->getOperand(1).getReg();
unsigned incr = MI->getOperand(2).getReg();
+ AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
+
unsigned oldval = dest;
DebugLoc dl = MI->getDebugLoc();
}
unsigned ldrOpc, strOpc;
- getExclusiveOperation(Size, ldrOpc, strOpc);
+ getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
// cmp incr, dest (, sign extend if necessary)
// csel scratch, dest, incr, cond
// stxr stxr_status, scratch, ptr
- // cmp stxr_status, #0
- // b.ne loopMBB
+ // cbnz stxr_status, loopMBB
// fallthrough --> exitMBB
BB = loopMBB;
BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
BuildMI(BB, dl, TII->get(strOpc), stxr_status)
.addReg(scratch).addReg(ptr);
- BuildMI(BB, dl, TII->get(AArch64::SUBwwi_lsl0_cmp))
- .addReg(stxr_status).addImm(0);
- BuildMI(BB, dl, TII->get(AArch64::Bcc))
- .addImm(A64CC::NE).addMBB(loopMBB);
+ BuildMI(BB, dl, TII->get(AArch64::CBNZw))
+ .addReg(stxr_status).addMBB(loopMBB);
BB->addSuccessor(loopMBB);
BB->addSuccessor(exitMBB);
unsigned ptr = MI->getOperand(1).getReg();
unsigned oldval = MI->getOperand(2).getReg();
unsigned newval = MI->getOperand(3).getReg();
+ AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
DebugLoc dl = MI->getDebugLoc();
TRCsp = Size == 8 ? &AArch64::GPR64xspRegClass : &AArch64::GPR32wspRegClass;
unsigned ldrOpc, strOpc;
- getExclusiveOperation(Size, ldrOpc, strOpc);
+ getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
MachineFunction *MF = BB->getParent();
const BasicBlock *LLVM_BB = BB->getBasicBlock();
// loop2MBB:
// strex stxr_status, newval, [ptr]
- // cmp stxr_status, #0
- // b.ne loop1MBB
+ // cbnz stxr_status, loop1MBB
BB = loop2MBB;
unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(newval).addReg(ptr);
- BuildMI(BB, dl, TII->get(AArch64::SUBwwi_lsl0_cmp))
- .addReg(stxr_status).addImm(0);
- BuildMI(BB, dl, TII->get(AArch64::Bcc))
- .addImm(A64CC::NE).addMBB(loop1MBB);
+ BuildMI(BB, dl, TII->get(AArch64::CBNZw))
+ .addReg(stxr_status).addMBB(loop1MBB);
BB->addSuccessor(loop1MBB);
BB->addSuccessor(exitMBB);
case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
case AArch64ISD::TLSDESCCALL: return "AArch64ISD::TLSDESCCALL";
+ case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
case AArch64ISD::WrapperSmall: return "AArch64ISD::WrapperSmall";
default: return NULL;
void
AArch64TargetLowering::SaveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG,
- DebugLoc DL, SDValue &Chain) const {
+ SDLoc DL, SDValue &Chain) const {
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo *MFI = MF.getFrameInfo();
AArch64MachineFunctionInfo *FuncInfo
AArch64TargetLowering::LowerFormalArguments(SDValue Chain,
CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
- DebugLoc dl, SelectionDAG &DAG,
+ SDLoc dl, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const {
MachineFunction &MF = DAG.getMachineFunction();
AArch64MachineFunctionInfo *FuncInfo
CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
- DebugLoc dl, SelectionDAG &DAG) const {
+ SDLoc dl, SelectionDAG &DAG) const {
// CCValAssign - represent the assignment of the return value to a location.
SmallVector<CCValAssign, 16> RVLocs;
AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
SmallVectorImpl<SDValue> &InVals) const {
SelectionDAG &DAG = CLI.DAG;
- DebugLoc &dl = CLI.DL;
+ SDLoc &dl = CLI.DL;
SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
}
if (!IsSibCall)
- Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
+ Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
+ dl);
SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, AArch64::XSP,
getPointerTy());
// in the correct location.
if (IsTailCall && !IsSibCall) {
Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
- DAG.getIntPtrConstant(0, true), InFlag);
+ DAG.getIntPtrConstant(0, true), InFlag, dl);
InFlag = Chain.getValue(1);
}
Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
DAG.getIntPtrConstant(CalleePopBytes, true),
- InFlag);
+ InFlag, dl);
InFlag = Chain.getValue(1);
}
AArch64TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
CallingConv::ID CallConv, bool IsVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
- DebugLoc dl, SelectionDAG &DAG,
+ SDLoc dl, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const {
// Assign locations to each value returned by this call.
SmallVector<CCValAssign, 16> RVLocs;
}
// Build a tokenfactor for all the chains.
- return DAG.getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other,
+ return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other,
&ArgChains[0], ArgChains.size());
}
SDValue AArch64TargetLowering::getSelectableIntSetCC(SDValue LHS, SDValue RHS,
ISD::CondCode CC, SDValue &A64cc,
- SelectionDAG &DAG, DebugLoc &dl) const {
+ SelectionDAG &DAG, SDLoc &dl) const {
if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
int64_t C = 0;
EVT VT = RHSC->getValueType(0);
SDValue
AArch64TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
- DebugLoc DL = Op.getDebugLoc();
+ SDLoc DL(Op);
EVT PtrVT = getPointerTy();
const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
- assert(getTargetMachine().getCodeModel() == CodeModel::Small
- && "Only small code model supported at the moment");
-
- // The most efficient code is PC-relative anyway for the small memory model,
- // so we don't need to worry about relocation model.
- return DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
- DAG.getTargetBlockAddress(BA, PtrVT, 0,
- AArch64II::MO_NO_FLAG),
- DAG.getTargetBlockAddress(BA, PtrVT, 0,
- AArch64II::MO_LO12),
- DAG.getConstant(/*Alignment=*/ 4, MVT::i32));
+ switch(getTargetMachine().getCodeModel()) {
+ case CodeModel::Small:
+ // The most efficient code is PC-relative anyway for the small memory model,
+ // so we don't need to worry about relocation model.
+ return DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
+ DAG.getTargetBlockAddress(BA, PtrVT, 0,
+ AArch64II::MO_NO_FLAG),
+ DAG.getTargetBlockAddress(BA, PtrVT, 0,
+ AArch64II::MO_LO12),
+ DAG.getConstant(/*Alignment=*/ 4, MVT::i32));
+ case CodeModel::Large:
+ return DAG.getNode(
+ AArch64ISD::WrapperLarge, DL, PtrVT,
+ DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G3),
+ DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
+ DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
+ DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
+ default:
+ llvm_unreachable("Only small and large code models supported now");
+ }
}
// (BRCOND chain, val, dest)
SDValue
AArch64TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
SDValue Chain = Op.getOperand(0);
SDValue TheBit = Op.getOperand(1);
SDValue DestBB = Op.getOperand(2);
// (BR_CC chain, condcode, lhs, rhs, dest)
SDValue
AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
SDValue Chain = Op.getOperand(0);
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
SDValue LHS = Op.getOperand(2);
CallLoweringInfo CLI(InChain, RetTy, false, false, false, false,
0, getLibcallCallingConv(Call), isTailCall,
/*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
- Callee, Args, DAG, Op->getDebugLoc());
+ Callee, Args, DAG, SDLoc(Op));
std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
if (!CallInfo.second.getNode())
SDValue SrcVal = Op.getOperand(0);
return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
- /*isSigned*/ false, Op.getDebugLoc());
+ /*isSigned*/ false, SDLoc(Op));
}
SDValue
}
SDValue
-AArch64TargetLowering::LowerGlobalAddressELF(SDValue Op,
- SelectionDAG &DAG) const {
- // TableGen doesn't have easy access to the CodeModel or RelocationModel, so
- // we make that distinction here.
+AArch64TargetLowering::LowerGlobalAddressELFLarge(SDValue Op,
+ SelectionDAG &DAG) const {
+ assert(getTargetMachine().getCodeModel() == CodeModel::Large);
+ assert(getTargetMachine().getRelocationModel() == Reloc::Static);
+
+ EVT PtrVT = getPointerTy();
+ SDLoc dl(Op);
+ const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
+ const GlobalValue *GV = GN->getGlobal();
+
+ SDValue GlobalAddr = DAG.getNode(
+ AArch64ISD::WrapperLarge, dl, PtrVT,
+ DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G3),
+ DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
+ DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
+ DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
+
+ if (GN->getOffset() != 0)
+ return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
+ DAG.getConstant(GN->getOffset(), PtrVT));
+
+ return GlobalAddr;
+}
- // We support the static, small memory model for now.
+SDValue
+AArch64TargetLowering::LowerGlobalAddressELFSmall(SDValue Op,
+ SelectionDAG &DAG) const {
assert(getTargetMachine().getCodeModel() == CodeModel::Small);
EVT PtrVT = getPointerTy();
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
const GlobalValue *GV = GN->getGlobal();
unsigned Alignment = GV->getAlignment();
+ Reloc::Model RelocM = getTargetMachine().getRelocationModel();
+ if (GV->isWeakForLinker() && GV->isDeclaration() && RelocM == Reloc::Static) {
+ // Weak undefined symbols can't use ADRP/ADD pair since they should evaluate
+ // to zero when they remain undefined. In PIC mode the GOT can take care of
+ // this, but in absolute mode we use a constant pool load.
+ SDValue PoolAddr;
+ PoolAddr = DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
+ DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
+ AArch64II::MO_NO_FLAG),
+ DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
+ AArch64II::MO_LO12),
+ DAG.getConstant(8, MVT::i32));
+ SDValue GlobalAddr = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), PoolAddr,
+ MachinePointerInfo::getConstantPool(),
+ /*isVolatile=*/ false,
+ /*isNonTemporal=*/ true,
+ /*isInvariant=*/ true, 8);
+ if (GN->getOffset() != 0)
+ return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
+ DAG.getConstant(GN->getOffset(), PtrVT));
+
+ return GlobalAddr;
+ }
if (Alignment == 0) {
const PointerType *GVPtrTy = cast<PointerType>(GV->getType());
}
unsigned char HiFixup, LoFixup;
- Reloc::Model RelocM = getTargetMachine().getRelocationModel();
- bool UseGOT = Subtarget->GVIsIndirectSymbol(GV, RelocM);
+ bool UseGOT = getSubtarget()->GVIsIndirectSymbol(GV, RelocM);
if (UseGOT) {
HiFixup = AArch64II::MO_GOT;
return GlobalRef;
}
+SDValue
+AArch64TargetLowering::LowerGlobalAddressELF(SDValue Op,
+ SelectionDAG &DAG) const {
+ // TableGen doesn't have easy access to the CodeModel or RelocationModel, so
+ // we make those distinctions here.
+
+ switch (getTargetMachine().getCodeModel()) {
+ case CodeModel::Small:
+ return LowerGlobalAddressELFSmall(Op, DAG);
+ case CodeModel::Large:
+ return LowerGlobalAddressELFLarge(Op, DAG);
+ default:
+ llvm_unreachable("Only small and large code models supported now");
+ }
+}
+
SDValue AArch64TargetLowering::LowerTLSDescCall(SDValue SymAddr,
SDValue DescAddr,
- DebugLoc DL,
+ SDLoc DL,
SelectionDAG &DAG) const {
EVT PtrVT = getPointerTy();
SDValue
AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
SelectionDAG &DAG) const {
- assert(Subtarget->isTargetELF() &&
+ assert(getSubtarget()->isTargetELF() &&
"TLS not implemented for non-ELF targets");
+ assert(getTargetMachine().getCodeModel() == CodeModel::Small
+ && "TLS only supported in small memory model");
const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
SDValue TPOff;
EVT PtrVT = getPointerTy();
- DebugLoc DL = Op.getDebugLoc();
+ SDLoc DL(Op);
const GlobalValue *GV = GA->getGlobal();
SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
SDValue
AArch64TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
- DebugLoc dl = JT->getDebugLoc();
+ SDLoc dl(JT);
+ EVT PtrVT = getPointerTy();
// When compiling PIC, jump tables get put in the code section so a static
// relocation-style is acceptable for both cases.
- return DAG.getNode(AArch64ISD::WrapperSmall, dl, getPointerTy(),
- DAG.getTargetJumpTable(JT->getIndex(), getPointerTy()),
- DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
- AArch64II::MO_LO12),
- DAG.getConstant(1, MVT::i32));
+ switch (getTargetMachine().getCodeModel()) {
+ case CodeModel::Small:
+ return DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
+ DAG.getTargetJumpTable(JT->getIndex(), PtrVT),
+ DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
+ AArch64II::MO_LO12),
+ DAG.getConstant(1, MVT::i32));
+ case CodeModel::Large:
+ return DAG.getNode(
+ AArch64ISD::WrapperLarge, dl, PtrVT,
+ DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G3),
+ DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G2_NC),
+ DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G1_NC),
+ DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G0_NC));
+ default:
+ llvm_unreachable("Only small and large code models supported now");
+ }
}
// (SELECT_CC lhs, rhs, iftrue, iffalse, condcode)
SDValue
AArch64TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
SDValue IfTrue = Op.getOperand(2);
// (SELECT testbit, iftrue, iffalse)
SDValue
AArch64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
SDValue TheBit = Op.getOperand(0);
SDValue IfTrue = Op.getOperand(1);
SDValue IfFalse = Op.getOperand(2);
// (SETCC lhs, rhs, condcode)
SDValue
AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
// We have to make sure we copy the entire structure: 8+8+8+4+4 = 32 bytes
// rather than just 8.
- return DAG.getMemcpy(Op.getOperand(0), Op.getDebugLoc(),
+ return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op),
Op.getOperand(1), Op.getOperand(2),
DAG.getConstant(32, MVT::i32), 8, false, false,
MachinePointerInfo(DestSV), MachinePointerInfo(SrcSV));
MachineFunction &MF = DAG.getMachineFunction();
AArch64MachineFunctionInfo *FuncInfo
= MF.getInfo<AArch64MachineFunctionInfo>();
- DebugLoc DL = Op.getDebugLoc();
+ SDLoc DL(Op);
SDValue Chain = Op.getOperand(0);
SDValue VAList = Op.getOperand(1);
TargetLowering::DAGCombinerInfo &DCI) {
SelectionDAG &DAG = DCI.DAG;
- DebugLoc DL = N->getDebugLoc();
+ SDLoc DL(N);
EVT VT = N->getValueType(0);
// We're looking for an SRA/SHL pair which form an SBFX.
DAG.getConstant(LSB + Width - 1, MVT::i64));
}
-static SDValue PerformATOMIC_FENCECombine(SDNode *FenceNode,
- TargetLowering::DAGCombinerInfo &DCI) {
- // An atomic operation followed by an acquiring atomic fence can be reduced to
- // an acquiring load. The atomic operation provides a convenient pointer to
- // load from. If the original operation was a load anyway we can actually
- // combine the two operations into an acquiring load.
- SelectionDAG &DAG = DCI.DAG;
- SDValue AtomicOp = FenceNode->getOperand(0);
- AtomicSDNode *AtomicNode = dyn_cast<AtomicSDNode>(AtomicOp);
-
- // A fence on its own can't be optimised
- if (!AtomicNode)
- return SDValue();
-
- AtomicOrdering FenceOrder
- = static_cast<AtomicOrdering>(FenceNode->getConstantOperandVal(1));
- SynchronizationScope FenceScope
- = static_cast<SynchronizationScope>(FenceNode->getConstantOperandVal(2));
-
- if (FenceOrder != Acquire || FenceScope != AtomicNode->getSynchScope())
- return SDValue();
-
- // If the original operation was an ATOMIC_LOAD then we'll be replacing it, so
- // the chain we use should be its input, otherwise we'll put our store after
- // it so we use its output chain.
- SDValue Chain = AtomicNode->getOpcode() == ISD::ATOMIC_LOAD ?
- AtomicNode->getChain() : AtomicOp;
-
- // We have an acquire fence with a handy atomic operation nearby, we can
- // convert the fence into a load-acquire, discarding the result.
- DebugLoc DL = FenceNode->getDebugLoc();
- SDValue Op = DAG.getAtomic(ISD::ATOMIC_LOAD, DL, AtomicNode->getMemoryVT(),
- AtomicNode->getValueType(0),
- Chain, // Chain
- AtomicOp.getOperand(1), // Pointer
- AtomicNode->getMemOperand(), Acquire,
- FenceScope);
-
- if (AtomicNode->getOpcode() == ISD::ATOMIC_LOAD)
- DAG.ReplaceAllUsesWith(AtomicNode, Op.getNode());
-
- return Op.getValue(1);
-}
-
-static SDValue PerformATOMIC_STORECombine(SDNode *N,
- TargetLowering::DAGCombinerInfo &DCI) {
- // A releasing atomic fence followed by an atomic store can be combined into a
- // single store operation.
- SelectionDAG &DAG = DCI.DAG;
- AtomicSDNode *AtomicNode = cast<AtomicSDNode>(N);
- SDValue FenceOp = AtomicNode->getOperand(0);
-
- if (FenceOp.getOpcode() != ISD::ATOMIC_FENCE)
- return SDValue();
-
- AtomicOrdering FenceOrder
- = static_cast<AtomicOrdering>(FenceOp->getConstantOperandVal(1));
- SynchronizationScope FenceScope
- = static_cast<SynchronizationScope>(FenceOp->getConstantOperandVal(2));
-
- if (FenceOrder != Release || FenceScope != AtomicNode->getSynchScope())
- return SDValue();
-
- DebugLoc DL = AtomicNode->getDebugLoc();
- return DAG.getAtomic(ISD::ATOMIC_STORE, DL, AtomicNode->getMemoryVT(),
- FenceOp.getOperand(0), // Chain
- AtomicNode->getOperand(1), // Pointer
- AtomicNode->getOperand(2), // Value
- AtomicNode->getMemOperand(), Release,
- FenceScope);
-}
-
/// For a true bitfield insert, the bits getting into that contiguous mask
/// should come from the low part of an existing value: they must be formed from
/// a compatible SHL operation (unless they're already low). This function
/// checks that condition and returns the least-significant bit that's
/// intended. If the operation not a field preparation, -1 is returned.
-static int32_t getLSBForBFI(SelectionDAG &DAG, DebugLoc DL, EVT VT,
+static int32_t getLSBForBFI(SelectionDAG &DAG, SDLoc DL, EVT VT,
SDValue &MaskedVal, uint64_t Mask) {
if (!isShiftedMask_64(Mask))
return -1;
// cases (e.g. bitfield to bitfield copy) may still need a real shift before
// the BFI.
- uint64_t LSB = CountTrailingZeros_64(Mask);
+ uint64_t LSB = countTrailingZeros(Mask);
int64_t ShiftRightRequired = LSB;
if (MaskedVal.getOpcode() == ISD::SHL &&
isa<ConstantSDNode>(MaskedVal.getOperand(1))) {
N = N.getOperand(0);
} else {
// Mask is the whole width.
- Mask = (1ULL << N.getValueType().getSizeInBits()) - 1;
+ Mask = -1ULL >> (64 - N.getValueType().getSizeInBits());
}
if (N.getOpcode() == AArch64ISD::BFI) {
TargetLowering::DAGCombinerInfo &DCI,
const AArch64Subtarget *Subtarget) {
SelectionDAG &DAG = DCI.DAG;
- DebugLoc DL = N->getDebugLoc();
+ SDLoc DL(N);
EVT VT = N->getValueType(0);
assert(N->getOpcode() == ISD::OR && "Unexpected root");
DAG.getConstant(Width, MVT::i64));
// Mask is trivial
- if ((LHSMask | RHSMask) == (1ULL << VT.getSizeInBits()) - 1)
+ if ((LHSMask | RHSMask) == (-1ULL >> (64 - VT.getSizeInBits())))
return BFI;
return DAG.getNode(ISD::AND, DL, VT, BFI,
TargetLowering::DAGCombinerInfo &DCI,
const AArch64Subtarget *Subtarget) {
SelectionDAG &DAG = DCI.DAG;
- DebugLoc DL = N->getDebugLoc();
+ SDLoc DL(N);
EVT VT = N->getValueType(0);
// First job is to hunt for a MaskedBFI on either the left or right. Swap
BFI.getOperand(2), BFI.getOperand(3));
// If the masking is trivial, we don't need to create it.
- if ((ExtraMask | ExistingMask) == (1ULL << VT.getSizeInBits()) - 1)
+ if ((ExtraMask | ExistingMask) == (-1ULL >> (64 - VT.getSizeInBits())))
return BFI;
return DAG.getNode(ISD::AND, DL, VT, BFI,
return true;
}
-/// EXTR instruciton extracts a contiguous chunk of bits from two existing
+/// EXTR instruction extracts a contiguous chunk of bits from two existing
/// registers viewed as a high/low pair. This function looks for the pattern:
/// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
/// EXTR. Can't quite be done in TableGen because the two immediates aren't
static SDValue tryCombineToEXTR(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI) {
SelectionDAG &DAG = DCI.DAG;
- DebugLoc DL = N->getDebugLoc();
+ SDLoc DL(N);
EVT VT = N->getValueType(0);
assert(N->getOpcode() == ISD::OR && "Unexpected root");
TargetLowering::DAGCombinerInfo &DCI) {
SelectionDAG &DAG = DCI.DAG;
- DebugLoc DL = N->getDebugLoc();
+ SDLoc DL(N);
EVT VT = N->getValueType(0);
// We're looking for an SRA/SHL pair which form an SBFX.
switch (N->getOpcode()) {
default: break;
case ISD::AND: return PerformANDCombine(N, DCI);
- case ISD::ATOMIC_FENCE: return PerformATOMIC_FENCECombine(N, DCI);
- case ISD::ATOMIC_STORE: return PerformATOMIC_STORECombine(N, DCI);
- case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
+ case ISD::OR: return PerformORCombine(N, DCI, getSubtarget());
case ISD::SRA: return PerformSRACombine(N, DCI);
}
return SDValue();
case 'S': {
// An absolute symbolic address or label reference.
if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
- Result = DAG.getTargetGlobalAddress(GA->getGlobal(), Op.getDebugLoc(),
+ Result = DAG.getTargetGlobalAddress(GA->getGlobal(), SDLoc(Op),
GA->getValueType(0));
} else if (const BlockAddressSDNode *BA
= dyn_cast<BlockAddressSDNode>(Op)) {
std::pair<unsigned, const TargetRegisterClass*>
AArch64TargetLowering::getRegForInlineAsmConstraint(
const std::string &Constraint,
- EVT VT) const {
+ MVT VT) const {
if (Constraint.size() == 1) {
switch (Constraint[0]) {
case 'r':