X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTarget%2FAArch64%2FAArch64ISelLowering.cpp;h=4ecfbe9e228091335dc720281446e4668be1a672;hp=be8696b25a9fc0f9f6afc957a27e4471af8deca4;hb=19e5c2eef38b9a79fa835d52ca51d5ef7abb0f61;hpb=dc73dc09f135cb1d36fd18f1918395ddf68cec63 diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp index be8696b25a9..4ecfbe9e228 100644 --- a/lib/Target/AArch64/AArch64ISelLowering.cpp +++ b/lib/Target/AArch64/AArch64ISelLowering.cpp @@ -40,23 +40,6 @@ using namespace llvm; STATISTIC(NumTailCalls, "Number of tail calls"); STATISTIC(NumShiftInserts, "Number of vector shift inserts"); -namespace { -enum AlignMode { - StrictAlign, - NoStrictAlign -}; -} - -static cl::opt -Align(cl::desc("Load/store alignment support"), - cl::Hidden, cl::init(NoStrictAlign), - cl::values( - clEnumValN(StrictAlign, "aarch64-strict-align", - "Disallow all unaligned memory accesses"), - clEnumValN(NoStrictAlign, "aarch64-no-strict-align", - "Allow unaligned memory accesses"), - clEnumValEnd)); - // Place holder until extr generation is tested fully. static cl::opt EnableAArch64ExtrGeneration("aarch64-extr-generation", cl::Hidden, @@ -213,11 +196,6 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); - // Exception handling. - // FIXME: These are guesses. Has this been defined yet? - setExceptionPointerRegister(AArch64::X0); - setExceptionSelectorRegister(AArch64::X1); - // Constant pool entries setOperationAction(ISD::ConstantPool, MVT::i64, Custom); @@ -237,6 +215,10 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, // AArch64 lacks both left-rotate and popcount instructions. setOperationAction(ISD::ROTL, MVT::i32, Expand); setOperationAction(ISD::ROTL, MVT::i64, Expand); + for (MVT VT : MVT::vector_valuetypes()) { + setOperationAction(ISD::ROTL, VT, Expand); + setOperationAction(ISD::ROTR, VT, Expand); + } // AArch64 doesn't have {U|S}MUL_LOHI. setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand); @@ -255,6 +237,10 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, setOperationAction(ISD::SDIVREM, MVT::i32, Expand); setOperationAction(ISD::SDIVREM, MVT::i64, Expand); + for (MVT VT : MVT::vector_valuetypes()) { + setOperationAction(ISD::SDIVREM, VT, Expand); + setOperationAction(ISD::UDIVREM, VT, Expand); + } setOperationAction(ISD::SREM, MVT::i32, Expand); setOperationAction(ISD::SREM, MVT::i64, Expand); setOperationAction(ISD::UDIVREM, MVT::i32, Expand); @@ -318,6 +304,8 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, setOperationAction(ISD::FTRUNC, MVT::f16, Promote); setOperationAction(ISD::FMINNUM, MVT::f16, Promote); setOperationAction(ISD::FMAXNUM, MVT::f16, Promote); + setOperationAction(ISD::FMINNAN, MVT::f16, Promote); + setOperationAction(ISD::FMAXNAN, MVT::f16, Promote); // v4f16 is also a storage-only type, so promote it to v4f32 when that is // known to be safe. @@ -406,10 +394,19 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, setOperationAction(ISD::FRINT, Ty, Legal); setOperationAction(ISD::FTRUNC, Ty, Legal); setOperationAction(ISD::FROUND, Ty, Legal); + setOperationAction(ISD::FMINNUM, Ty, Legal); + setOperationAction(ISD::FMAXNUM, Ty, Legal); + setOperationAction(ISD::FMINNAN, Ty, Legal); + setOperationAction(ISD::FMAXNAN, Ty, Legal); } setOperationAction(ISD::PREFETCH, MVT::Other, Custom); + // Lower READCYCLECOUNTER using an mrs from PMCCNTR_EL0. + // This requires the Performance Monitors extension. + if (Subtarget->hasPerfMon()) + setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal); + if (Subtarget->isTargetMachO()) { // For iOS, we don't want to the normal expansion of a libcall to // sincos. We want to issue a libcall to __sincos_stret to avoid memory @@ -459,12 +456,14 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, setIndexedLoadAction(im, MVT::i64, Legal); setIndexedLoadAction(im, MVT::f64, Legal); setIndexedLoadAction(im, MVT::f32, Legal); + setIndexedLoadAction(im, MVT::f16, Legal); setIndexedStoreAction(im, MVT::i8, Legal); setIndexedStoreAction(im, MVT::i16, Legal); setIndexedStoreAction(im, MVT::i32, Legal); setIndexedStoreAction(im, MVT::i64, Legal); setIndexedStoreAction(im, MVT::f64, Legal); setIndexedStoreAction(im, MVT::f32, Legal); + setIndexedStoreAction(im, MVT::f16, Legal); } // Trap. @@ -482,6 +481,10 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, setTargetDAGCombine(ISD::SINT_TO_FP); setTargetDAGCombine(ISD::UINT_TO_FP); + setTargetDAGCombine(ISD::FP_TO_SINT); + setTargetDAGCombine(ISD::FP_TO_UINT); + setTargetDAGCombine(ISD::FDIV); + setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); setTargetDAGCombine(ISD::ANY_EXTEND); @@ -490,16 +493,18 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, setTargetDAGCombine(ISD::BITCAST); setTargetDAGCombine(ISD::CONCAT_VECTORS); setTargetDAGCombine(ISD::STORE); + if (Subtarget->supportsAddressTopByteIgnored()) + setTargetDAGCombine(ISD::LOAD); setTargetDAGCombine(ISD::MUL); setTargetDAGCombine(ISD::SELECT); setTargetDAGCombine(ISD::VSELECT); - setTargetDAGCombine(ISD::SELECT_CC); setTargetDAGCombine(ISD::INTRINSIC_VOID); setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); setTargetDAGCombine(ISD::INSERT_VECTOR_ELT); + setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT); MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8; MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4; @@ -515,8 +520,6 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, setMinFunctionAlignment(2); - RequireStrictAlign = (Align == StrictAlign); - setHasExtractBitsInsn(true); setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); @@ -651,6 +654,9 @@ void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) { setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand); setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand); setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand); + + // But we do support custom-lowering for FCOPYSIGN. + setOperationAction(ISD::FCOPYSIGN, VT.getSimpleVT(), Custom); } setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom); @@ -685,12 +691,16 @@ void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) { setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom); setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom); - // [SU][MIN|MAX] and [SU]ABSDIFF are available for all NEON types apart from - // i64. + // [SU][MIN|MAX] are available for all NEON types apart from i64. if (!VT.isFloatingPoint() && VT.getSimpleVT() != MVT::v2i64 && VT.getSimpleVT() != MVT::v1i64) - for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX, - ISD::SABSDIFF, ISD::UABSDIFF}) + for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}) + setOperationAction(Opcode, VT.getSimpleVT(), Legal); + + // F[MIN|MAX][NUM|NAN] are available for all FP NEON types (not f16 though!). + if (VT.isFloatingPoint() && VT.getVectorElementType() != MVT::f16) + for (unsigned Opcode : {ISD::FMINNAN, ISD::FMAXNAN, + ISD::FMINNUM, ISD::FMAXNUM}) setOperationAction(Opcode, VT.getSimpleVT(), Legal); if (Subtarget->isLittleEndian()) { @@ -737,7 +747,7 @@ void AArch64TargetLowering::computeKnownBitsForTargetNode( break; } case ISD::INTRINSIC_W_CHAIN: { - ConstantSDNode *CN = cast(Op->getOperand(1)); + ConstantSDNode *CN = cast(Op->getOperand(1)); Intrinsic::ID IntID = static_cast(CN->getZExtValue()); switch (IntID) { default: return; @@ -787,6 +797,34 @@ MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL, return MVT::i64; } +bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(EVT VT, + unsigned AddrSpace, + unsigned Align, + bool *Fast) const { + if (Subtarget->requiresStrictAlign()) + return false; + + // FIXME: This is mostly true for Cyclone, but not necessarily others. + if (Fast) { + // FIXME: Define an attribute for slow unaligned accesses instead of + // relying on the CPU type as a proxy. + // On Cyclone, unaligned 128-bit stores are slow. + *Fast = !Subtarget->isCyclone() || VT.getStoreSize() != 16 || + // See comments in performSTORECombine() for more details about + // these conditions. + + // Code that uses clang vector extensions can mark that it + // wants unaligned accesses to be treated as fast by + // underspecifying alignment to be 1 or 2. + Align <= 2 || + + // Disregard v2i64. Memcpy lowering produces those and splitting + // them regresses performance on micro-benchmarks and olden/bh. + VT == MVT::v2i64; + } + return true; +} + FastISel * AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, const TargetLibraryInfo *libInfo) const { @@ -820,8 +858,6 @@ const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const { case AArch64ISD::CCMN: return "AArch64ISD::CCMN"; case AArch64ISD::FCCMP: return "AArch64ISD::FCCMP"; case AArch64ISD::FCMP: return "AArch64ISD::FCMP"; - case AArch64ISD::FMIN: return "AArch64ISD::FMIN"; - case AArch64ISD::FMAX: return "AArch64ISD::FMAX"; case AArch64ISD::DUP: return "AArch64ISD::DUP"; case AArch64ISD::DUPLANE8: return "AArch64ISD::DUPLANE8"; case AArch64ISD::DUPLANE16: return "AArch64ISD::DUPLANE16"; @@ -941,8 +977,7 @@ AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI, const TargetInstrInfo *TII = Subtarget->getInstrInfo(); const BasicBlock *LLVM_BB = MBB->getBasicBlock(); DebugLoc DL = MI->getDebugLoc(); - MachineFunction::iterator It = MBB; - ++It; + MachineFunction::iterator It = ++MBB->getIterator(); unsigned DestReg = MI->getOperand(0).getReg(); unsigned IfTrueReg = MI->getOperand(1).getReg(); @@ -1151,8 +1186,7 @@ static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC, // register to WZR/XZR if it ends up being unused. unsigned Opcode = AArch64ISD::SUBS; - if (RHS.getOpcode() == ISD::SUB && isa(RHS.getOperand(0)) && - cast(RHS.getOperand(0))->getZExtValue() == 0 && + if (RHS.getOpcode() == ISD::SUB && isNullConstant(RHS.getOperand(0)) && (CC == ISD::SETEQ || CC == ISD::SETNE)) { // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags @@ -1166,8 +1200,7 @@ static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC, // the absence of information about op2. Opcode = AArch64ISD::ADDS; RHS = RHS.getOperand(1); - } else if (LHS.getOpcode() == ISD::AND && isa(RHS) && - cast(RHS)->getZExtValue() == 0 && + } else if (LHS.getOpcode() == ISD::AND && isNullConstant(RHS) && !isUnsignedIntSetCC(CC)) { // Similarly, (CMP (and X, Y), 0) can be implemented with a TST // (a.k.a. ANDS) except that the flags are only guaranteed to work for one @@ -1232,8 +1265,7 @@ static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS, Opcode = AArch64ISD::FCCMP; else if (RHS.getOpcode() == ISD::SUB) { SDValue SubOp0 = RHS.getOperand(0); - if (const ConstantSDNode *SubOp0C = dyn_cast(SubOp0)) - if (SubOp0C->isNullValue() && (CC == ISD::SETEQ || CC == ISD::SETNE)) { + if (isNullConstant(SubOp0) && (CC == ISD::SETEQ || CC == ISD::SETNE)) { // See emitComparison() on why we can only do this for SETEQ and SETNE. Opcode = AArch64ISD::CCMN; RHS = RHS.getOperand(1); @@ -1342,7 +1374,7 @@ static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val, unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC); return emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp, NZCV, DL, DAG); - } else if (Opcode != ISD::AND && Opcode != ISD::OR) + } else if ((Opcode != ISD::AND && Opcode != ISD::OR) || !Val->hasOneUse()) return SDValue(); assert((Opcode == ISD::OR || !PushNegate) @@ -1367,10 +1399,17 @@ static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val, if (!CanPushNegateL && !CanPushNegateR) return SDValue(); // Order the side where we can push the negate through to LHS. - if (!CanPushNegateL && CanPushNegateR) { + if (!CanPushNegateL && CanPushNegateR) + std::swap(LHS, RHS); + } else { + bool NeedsNegOutL = LHS->getOpcode() == ISD::OR; + bool NeedsNegOutR = RHS->getOpcode() == ISD::OR; + if (NeedsNegOutL && NeedsNegOutR) + return SDValue(); + // Order the side where we need to negate the output flags to RHS so it + // gets emitted first. + if (NeedsNegOutL) std::swap(LHS, RHS); - CanPushNegateL = true; - } } // Emit RHS. If we want to negate the tree we only need to push a negate @@ -1468,7 +1507,7 @@ static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, // cmn w0, #1 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS) // if and only if (sext LHS) == (sext RHS). The checks are in place to - // ensure both the LHS and RHS are truely zero extended and to make sure the + // ensure both the LHS and RHS are truly zero extended and to make sure the // transformation is profitable. if ((RHSC->getZExtValue() >> 16 == 0) && isa(LHS) && cast(LHS)->getExtensionType() == ISD::ZEXTLOAD && @@ -1620,8 +1659,7 @@ getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) { SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG, RTLIB::Libcall Call) const { SmallVector Ops(Op->op_begin(), Op->op_end()); - return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false, - SDLoc(Op)).first; + return makeLibCall(DAG, Call, MVT::f128, Ops, false, SDLoc(Op)).first; } static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) { @@ -1800,8 +1838,8 @@ SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op, // precise. That doesn't take part in the LibCall so we can't directly use // LowerF128Call. SDValue SrcVal = Op.getOperand(0); - return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1, - /*isSigned*/ false, SDLoc(Op)).first; + return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false, + SDLoc(Op)).first; } static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) { @@ -1810,6 +1848,16 @@ static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) { // in the cost tables. EVT InVT = Op.getOperand(0).getValueType(); EVT VT = Op.getValueType(); + unsigned NumElts = InVT.getVectorNumElements(); + + // f16 vectors are promoted to f32 before a conversion. + if (InVT.getVectorElementType() == MVT::f16) { + MVT NewVT = MVT::getVectorVT(MVT::f32, NumElts); + SDLoc dl(Op); + return DAG.getNode( + Op.getOpcode(), dl, Op.getValueType(), + DAG.getNode(ISD::FP_EXTEND, dl, NewVT, Op.getOperand(0))); + } if (VT.getSizeInBits() < InVT.getSizeInBits()) { SDLoc dl(Op); @@ -1857,8 +1905,7 @@ SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op, LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType()); SmallVector Ops(Op->op_begin(), Op->op_end()); - return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false, - SDLoc(Op)).first; + return makeLibCall(DAG, LC, Op.getValueType(), Ops, false, SDLoc(Op)).first; } static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) { @@ -2170,6 +2217,18 @@ SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, EVT PtrVT = getPointerTy(DAG.getDataLayout()); return DAG.getNode(AArch64ISD::THREAD_POINTER, dl, PtrVT); } + case Intrinsic::aarch64_neon_smax: + return DAG.getNode(ISD::SMAX, dl, Op.getValueType(), + Op.getOperand(1), Op.getOperand(2)); + case Intrinsic::aarch64_neon_umax: + return DAG.getNode(ISD::UMAX, dl, Op.getValueType(), + Op.getOperand(1), Op.getOperand(2)); + case Intrinsic::aarch64_neon_smin: + return DAG.getNode(ISD::SMIN, dl, Op.getValueType(), + Op.getOperand(1), Op.getOperand(2)); + case Intrinsic::aarch64_neon_umin: + return DAG.getNode(ISD::UMIN, dl, Op.getValueType(), + Op.getOperand(1), Op.getOperand(2)); } } @@ -2279,11 +2338,6 @@ SDValue AArch64TargetLowering::LowerOperation(SDValue Op, } } -/// getFunctionAlignment - Return the Log2 alignment of this function. -unsigned AArch64TargetLowering::getFunctionAlignment(const Function *F) const { - return 2; -} - //===----------------------------------------------------------------------===// // Calling Convention Implementation //===----------------------------------------------------------------------===// @@ -2372,7 +2426,7 @@ SDValue AArch64TargetLowering::LowerFormalArguments( continue; } - + if (VA.isRegLoc()) { // Arguments stored in registers. EVT RegVT = VA.getLocVT(); @@ -2458,9 +2512,10 @@ SDValue AArch64TargetLowering::LowerFormalArguments( break; } - ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN, - MachinePointerInfo::getFixedStack(FI), - MemVT, false, false, false, 0); + ArgValue = DAG.getExtLoad( + ExtType, DL, VA.getLocVT(), Chain, FIN, + MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), + MemVT, false, false, false, 0); InVals.push_back(ArgValue); } @@ -2533,9 +2588,10 @@ void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo, for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) { unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass); SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64); - SDValue Store = - DAG.getStore(Val.getValue(1), DL, Val, FIN, - MachinePointerInfo::getStack(i * 8), false, false, 0); + SDValue Store = DAG.getStore( + Val.getValue(1), DL, Val, FIN, + MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 8), false, + false, 0); MemOps.push_back(Store); FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT)); @@ -2562,9 +2618,10 @@ void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo, unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass); SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128); - SDValue Store = - DAG.getStore(Val.getValue(1), DL, Val, FIN, - MachinePointerInfo::getStack(i * 16), false, false, 0); + SDValue Store = DAG.getStore( + Val.getValue(1), DL, Val, FIN, + MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 16), + false, false, 0); MemOps.push_back(Store); FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(16, DL, PtrVT)); @@ -2697,8 +2754,8 @@ bool AArch64TargetLowering::isEligibleForTailCallOptimization( *DAG.getContext()); CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true)); - for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) - if (!ArgLocs[i].isRegLoc()) + for (const CCValAssign &ArgLoc : ArgLocs) + if (!ArgLoc.isRegLoc()) return false; } @@ -3002,7 +3059,8 @@ AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI, int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true); DstAddr = DAG.getFrameIndex(FI, PtrVT); - DstInfo = MachinePointerInfo::getFixedStack(FI); + DstInfo = + MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI); // Make sure any stack arguments overlapping with where we're storing // are loaded before this eventual operation. Otherwise they'll be @@ -3012,7 +3070,8 @@ AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI, SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL); DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff); - DstInfo = MachinePointerInfo::getStack(LocMemOffset); + DstInfo = MachinePointerInfo::getStack(DAG.getMachineFunction(), + LocMemOffset); } if (Outs[i].Flags.isByVal()) { @@ -3046,9 +3105,9 @@ AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI, // Build a sequence of copy-to-reg nodes chained together with token chain // and flag operands which copy the outgoing args into the appropriate regs. SDValue InFlag; - for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { - Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first, - RegsToPass[i].second, InFlag); + for (auto &RegToPass : RegsToPass) { + Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first, + RegToPass.second, InFlag); InFlag = Chain.getValue(1); } @@ -3104,9 +3163,9 @@ AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI, // Add argument registers to the end of the list so that they are known live // into the call. - for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) - Ops.push_back(DAG.getRegister(RegsToPass[i].first, - RegsToPass[i].second.getValueType())); + for (auto &RegToPass : RegsToPass) + Ops.push_back(DAG.getRegister(RegToPass.first, + RegToPass.second.getValueType())); // Add a register mask operand representing the call-preserved registers. const uint32_t *Mask; @@ -3212,6 +3271,19 @@ AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, Flag = Chain.getValue(1); RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); } + const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo(); + const MCPhysReg *I = + TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction()); + if (I) { + for (; *I; ++I) { + if (AArch64::GPR64RegClass.contains(*I)) + RetOps.push_back(DAG.getRegister(*I, MVT::i64)); + else if (AArch64::FPR64RegClass.contains(*I)) + RetOps.push_back(DAG.getRegister(*I, MVT::getFloatingPointVT(64))); + else + llvm_unreachable("Unexpected register class in CSRsViaCopy!"); + } + } RetOps[0] = Chain; // Update chain. @@ -3254,11 +3326,12 @@ SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op, unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC; SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags); SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo); - SDValue GlobalAddr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), PoolAddr, - MachinePointerInfo::getConstantPool(), - /*isVolatile=*/ false, - /*isNonTemporal=*/ true, - /*isInvariant=*/ true, 8); + SDValue GlobalAddr = DAG.getLoad( + PtrVT, DL, DAG.getEntryNode(), PoolAddr, + MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), + /*isVolatile=*/false, + /*isNonTemporal=*/true, + /*isInvariant=*/true, 8); if (GN->getOffset() != 0) return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr, DAG.getConstant(GN->getOffset(), DL, PtrVT)); @@ -3331,8 +3404,9 @@ AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op, // to obtain the address of the variable. SDValue Chain = DAG.getEntryNode(); SDValue FuncTLVGet = - DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(), - false, true, true, 8); + DAG.getLoad(MVT::i64, DL, Chain, DescAddr, + MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, + true, true, 8); Chain = FuncTLVGet.getValue(1); MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); @@ -3525,8 +3599,7 @@ SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const { // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch // instruction. unsigned Opc = LHS.getOpcode(); - if (LHS.getResNo() == 1 && isa(RHS) && - cast(RHS)->isOne() && + if (LHS.getResNo() == 1 && isOneConstant(RHS) && (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO || Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) { assert((CC == ISD::SETEQ || CC == ISD::SETNE) && @@ -3640,17 +3713,11 @@ SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op, SDValue In1 = Op.getOperand(0); SDValue In2 = Op.getOperand(1); EVT SrcVT = In2.getValueType(); - if (SrcVT != VT) { - if (SrcVT == MVT::f32 && VT == MVT::f64) - In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2); - else if (SrcVT == MVT::f64 && VT == MVT::f32) - In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, - DAG.getIntPtrConstant(0, DL)); - else - // FIXME: Src type is different, bail out for now. Can VT really be a - // vector type? - return SDValue(); - } + + if (SrcVT.bitsLT(VT)) + In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2); + else if (SrcVT.bitsGT(VT)) + In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0, DL)); EVT VecVT; EVT EltVT; @@ -3658,7 +3725,7 @@ SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op, SDValue VecVal1, VecVal2; if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) { EltVT = MVT::i32; - VecVT = MVT::v4i32; + VecVT = (VT == MVT::v2f32 ? MVT::v2i32 : MVT::v4i32); EltMask = 0x80000000ULL; if (!VT.isVector()) { @@ -3819,32 +3886,6 @@ SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const { } } -/// A SELECT_CC operation is really some kind of max or min if both values being -/// compared are, in some sense, equal to the results in either case. However, -/// it is permissible to compare f32 values and produce directly extended f64 -/// values. -/// -/// Extending the comparison operands would also be allowed, but is less likely -/// to happen in practice since their use is right here. Note that truncate -/// operations would *not* be semantically equivalent. -static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) { - if (Cmp == Result) - return (Cmp.getValueType() == MVT::f32 || - Cmp.getValueType() == MVT::f64); - - ConstantFPSDNode *CCmp = dyn_cast(Cmp); - ConstantFPSDNode *CResult = dyn_cast(Result); - if (CCmp && CResult && Cmp.getValueType() == MVT::f32 && - Result.getValueType() == MVT::f64) { - bool Lossy; - APFloat CmpVal = CCmp->getValueAPF(); - CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy); - return CResult->getValueAPF().bitwiseIsEqual(CmpVal); - } - - return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp; -} - SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS, SDValue RHS, SDValue TVal, SDValue FVal, SDLoc dl, @@ -3862,7 +3903,13 @@ SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS, } } - // Handle integers first. + // Also handle f16, for which we need to do a f32 comparison. + if (LHS.getValueType() == MVT::f16) { + LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, LHS); + RHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, RHS); + } + + // Next, handle integers. if (LHS.getValueType().isInteger()) { assert((LHS.getValueType() == RHS.getValueType()) && (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64)); @@ -3885,9 +3932,7 @@ SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS, } else if (TVal.getOpcode() == ISD::XOR) { // If TVal is a NOT we want to swap TVal and FVal so that we can match // with a CSINV rather than a CSEL. - ConstantSDNode *CVal = dyn_cast(TVal.getOperand(1)); - - if (CVal && CVal->isAllOnesValue()) { + if (isAllOnesConstant(TVal.getOperand(1))) { std::swap(TVal, FVal); std::swap(CTVal, CFVal); CC = ISD::getSetCCInverse(CC, true); @@ -3895,9 +3940,7 @@ SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS, } else if (TVal.getOpcode() == ISD::SUB) { // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so // that we can match with a CSNEG rather than a CSEL. - ConstantSDNode *CVal = dyn_cast(TVal.getOperand(0)); - - if (CVal && CVal->isNullValue()) { + if (isNullConstant(TVal.getOperand(0))) { std::swap(TVal, FVal); std::swap(CTVal, CFVal); CC = ISD::getSetCCInverse(CC, true); @@ -4357,46 +4400,57 @@ SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op, SDValue ShOpLo = Op.getOperand(0); SDValue ShOpHi = Op.getOperand(1); SDValue ShAmt = Op.getOperand(2); - SDValue ARMcc; 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, dl, MVT::i64), ShAmt); - SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt); + SDValue HiBitsForLo = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt); + + // Unfortunately, if ShAmt == 0, we just calculated "(SHL ShOpHi, 64)" which + // is "undef". We wanted 0, so CSEL it directly. + SDValue Cmp = emitComparison(ShAmt, DAG.getConstant(0, dl, MVT::i64), + ISD::SETEQ, dl, DAG); + SDValue CCVal = DAG.getConstant(AArch64CC::EQ, dl, MVT::i32); + HiBitsForLo = + DAG.getNode(AArch64ISD::CSEL, dl, VT, DAG.getConstant(0, dl, MVT::i64), + HiBitsForLo, CCVal, Cmp); + SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt, DAG.getConstant(VTBits, dl, MVT::i64)); - SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt); - SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64), - ISD::SETGE, dl, DAG); - SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32); + SDValue LoBitsForLo = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt); + SDValue LoForNormalShift = + DAG.getNode(ISD::OR, dl, VT, LoBitsForLo, HiBitsForLo); - SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); - SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt); - SDValue Lo = - DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp); + Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64), ISD::SETGE, + dl, DAG); + CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32); + SDValue LoForBigShift = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt); + SDValue Lo = DAG.getNode(AArch64ISD::CSEL, dl, VT, LoForBigShift, + LoForNormalShift, CCVal, Cmp); // AArch64 shifts larger than the register width are wrapped rather than // clamped, so we can't just emit "hi >> x". - SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt); - SDValue TrueValHi = Opc == ISD::SRA - ? DAG.getNode(Opc, dl, VT, ShOpHi, - DAG.getConstant(VTBits - 1, dl, - MVT::i64)) - : DAG.getConstant(0, dl, VT); - SDValue Hi = - DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp); + SDValue HiForNormalShift = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt); + SDValue HiForBigShift = + Opc == ISD::SRA + ? DAG.getNode(Opc, dl, VT, ShOpHi, + DAG.getConstant(VTBits - 1, dl, MVT::i64)) + : DAG.getConstant(0, dl, VT); + SDValue Hi = DAG.getNode(AArch64ISD::CSEL, dl, VT, HiForBigShift, + HiForNormalShift, CCVal, Cmp); SDValue Ops[2] = { Lo, Hi }; return DAG.getMergeValues(Ops, dl); } + /// 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 { + SelectionDAG &DAG) const { assert(Op.getNumOperands() == 3 && "Not a double-shift!"); EVT VT = Op.getValueType(); unsigned VTBits = VT.getSizeInBits(); @@ -4404,31 +4458,41 @@ SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op, SDValue ShOpLo = Op.getOperand(0); SDValue ShOpHi = Op.getOperand(1); SDValue ShAmt = Op.getOperand(2); - SDValue ARMcc; assert(Op.getOpcode() == ISD::SHL_PARTS); SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, DAG.getConstant(VTBits, dl, MVT::i64), ShAmt); - SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt); + SDValue LoBitsForHi = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt); + + // Unfortunately, if ShAmt == 0, we just calculated "(SRL ShOpLo, 64)" which + // is "undef". We wanted 0, so CSEL it directly. + SDValue Cmp = emitComparison(ShAmt, DAG.getConstant(0, dl, MVT::i64), + ISD::SETEQ, dl, DAG); + SDValue CCVal = DAG.getConstant(AArch64CC::EQ, dl, MVT::i32); + LoBitsForHi = + DAG.getNode(AArch64ISD::CSEL, dl, VT, DAG.getConstant(0, dl, MVT::i64), + LoBitsForHi, CCVal, Cmp); + SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt, DAG.getConstant(VTBits, dl, MVT::i64)); - SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt); - SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt); + SDValue HiBitsForHi = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt); + SDValue HiForNormalShift = + DAG.getNode(ISD::OR, dl, VT, LoBitsForHi, HiBitsForHi); - SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); + SDValue HiForBigShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt); - SDValue Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64), - ISD::SETGE, dl, DAG); - SDValue CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32); - SDValue Hi = - DAG.getNode(AArch64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp); + Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64), ISD::SETGE, + dl, DAG); + CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32); + SDValue Hi = DAG.getNode(AArch64ISD::CSEL, dl, VT, HiForBigShift, + HiForNormalShift, CCVal, Cmp); // AArch64 shifts of larger than register sizes are wrapped rather than // clamped, so we can't just emit "lo << a" if a is too big. - SDValue TrueValLo = DAG.getConstant(0, dl, VT); - SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt); - SDValue Lo = - DAG.getNode(AArch64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp); + SDValue LoForBigShift = DAG.getConstant(0, dl, VT); + SDValue LoForNormalShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt); + SDValue Lo = DAG.getNode(AArch64ISD::CSEL, dl, VT, LoForBigShift, + LoForNormalShift, CCVal, Cmp); SDValue Ops[2] = { Lo, Hi }; return DAG.getMergeValues(Ops, dl); @@ -4610,8 +4674,7 @@ void AArch64TargetLowering::LowerAsmOperandForConstraint( // Validate and return a target constant for them if we can. case 'z': { // 'z' maps to xzr or wzr so it needs an input of 0. - ConstantSDNode *C = dyn_cast(Op); - if (!C || C->getZExtValue() != 0) + if (!isNullConstant(Op)) return; if (Op.getValueType() == MVT::i64) @@ -5011,7 +5074,7 @@ static bool isEXTMask(ArrayRef M, EVT VT, bool &ReverseEXT, // The index of an EXT is the first element if it is not UNDEF. // Watch out for the beginning UNDEFs. The EXT index should be the expected - // value of the first element. E.g. + // value of the first element. E.g. // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>. // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>. // ExpectedElt is the last mask index plus 1. @@ -5901,11 +5964,10 @@ static SDValue NormalizeBuildVector(SDValue Op, return Op; SmallVector Ops; - for (unsigned I = 0, E = VT.getVectorNumElements(); I != E; ++I) { - SDValue Lane = Op.getOperand(I); - if (Lane.getOpcode() == ISD::Constant) { + for (SDValue Lane : Op->ops()) { + if (auto *CstLane = dyn_cast(Lane)) { APInt LowBits(EltTy.getSizeInBits(), - cast(Lane)->getZExtValue()); + CstLane->getZExtValue()); Lane = DAG.getConstant(LowBits.getZExtValue(), dl, MVT::i32); } Ops.push_back(Lane); @@ -6245,8 +6307,7 @@ FailedModImm: // Empirical tests suggest this is rarely worth it for vectors of length <= 2. if (NumElts >= 4) { - SDValue shuffle = ReconstructShuffle(Op, DAG); - if (shuffle != SDValue()) + if (SDValue shuffle = ReconstructShuffle(Op, DAG)) return shuffle; } @@ -6265,7 +6326,10 @@ FailedModImm: // a) Avoid a RMW dependency on the full vector register, and // b) Allow the register coalescer to fold away the copy if the // value is already in an S or D register. - if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) { + // Do not do this for UNDEF/LOAD nodes because we have better patterns + // for those avoiding the SCALAR_TO_VECTOR/BUILD_VECTOR. + if (Op0.getOpcode() != ISD::UNDEF && Op0.getOpcode() != ISD::LOAD && + (ElemSize == 32 || ElemSize == 64)) { unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub; MachineSDNode *N = DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0, @@ -6371,24 +6435,11 @@ SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, unsigned Val = Cst->getZExtValue(); unsigned Size = Op.getValueType().getSizeInBits(); - if (Val == 0) { - switch (Size) { - case 8: - return DAG.getTargetExtractSubreg(AArch64::bsub, dl, Op.getValueType(), - Op.getOperand(0)); - case 16: - return DAG.getTargetExtractSubreg(AArch64::hsub, dl, Op.getValueType(), - Op.getOperand(0)); - case 32: - return DAG.getTargetExtractSubreg(AArch64::ssub, dl, Op.getValueType(), - Op.getOperand(0)); - case 64: - return DAG.getTargetExtractSubreg(AArch64::dsub, dl, Op.getValueType(), - Op.getOperand(0)); - default: - llvm_unreachable("Unexpected vector type in extract_subvector!"); - } - } + + // This will get lowered to an appropriate EXTRACT_SUBREG in ISel. + if (Val == 0) + return Op; + // If this is extracting the upper 64-bits of a 128-bit vector, we match // that directly. if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64) @@ -6692,7 +6743,7 @@ bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, case Intrinsic::aarch64_neon_ld4r: { Info.opc = ISD::INTRINSIC_W_CHAIN; // Conservatively set memVT to the entire set of vectors loaded. - uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8; + uint64_t NumElts = DL.getTypeSizeInBits(I.getType()) / 64; Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts); Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1); Info.offset = 0; @@ -6718,7 +6769,7 @@ bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, Type *ArgTy = I.getArgOperand(ArgI)->getType(); if (!ArgTy->isVectorTy()) break; - NumElts += DL.getTypeAllocSize(ArgTy) / 8; + NumElts += DL.getTypeSizeInBits(ArgTy) / 64; } Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts); Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1); @@ -6961,10 +7012,10 @@ bool AArch64TargetLowering::lowerInterleavedLoad( const DataLayout &DL = LI->getModule()->getDataLayout(); VectorType *VecTy = Shuffles[0]->getType(); - unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy); + unsigned VecSize = DL.getTypeSizeInBits(VecTy); - // Skip illegal vector types. - if (VecSize != 64 && VecSize != 128) + // Skip if we do not have NEON and skip illegal vector types. + if (!Subtarget->hasNEON() || (VecSize != 64 && VecSize != 128)) return false; // A pointer vector can not be the return type of the ldN intrinsics. Need to @@ -7047,10 +7098,10 @@ bool AArch64TargetLowering::lowerInterleavedStore(StoreInst *SI, VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts); const DataLayout &DL = SI->getModule()->getDataLayout(); - unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy); + unsigned SubVecSize = DL.getTypeSizeInBits(SubVecTy); - // Skip illegal vector types. - if (SubVecSize != 64 && SubVecSize != 128) + // Skip if we do not have NEON and skip illegal vector types. + if (!Subtarget->hasNEON() || (SubVecSize != 64 && SubVecSize != 128)) return false; Value *Op0 = SVI->getOperand(0); @@ -7469,8 +7520,7 @@ static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG, const AArch64Subtarget *Subtarget) { // First try to optimize away the conversion when it's conditionally from // a constant. Vectors only. - SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG); - if (Res != SDValue()) + if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG)) return Res; EVT VT = N->getValueType(0); @@ -7483,7 +7533,7 @@ static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG, // If the result of an integer load is only used by an integer-to-float // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead. - // This eliminates an "integer-to-vector-move UOP and improve throughput. + // This eliminates an "integer-to-vector-move" UOP and improves throughput. SDValue N0 = N->getOperand(0); if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && // Do not change the width of a volatile load. @@ -7506,6 +7556,134 @@ static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG, return SDValue(); } +/// Fold a floating-point multiply by power of two into floating-point to +/// fixed-point conversion. +static SDValue performFpToIntCombine(SDNode *N, SelectionDAG &DAG, + const AArch64Subtarget *Subtarget) { + if (!Subtarget->hasNEON()) + return SDValue(); + + SDValue Op = N->getOperand(0); + if (!Op.getValueType().isVector() || Op.getOpcode() != ISD::FMUL) + return SDValue(); + + SDValue ConstVec = Op->getOperand(1); + if (!isa(ConstVec)) + return SDValue(); + + MVT FloatTy = Op.getSimpleValueType().getVectorElementType(); + uint32_t FloatBits = FloatTy.getSizeInBits(); + if (FloatBits != 32 && FloatBits != 64) + return SDValue(); + + MVT IntTy = N->getSimpleValueType(0).getVectorElementType(); + uint32_t IntBits = IntTy.getSizeInBits(); + if (IntBits != 16 && IntBits != 32 && IntBits != 64) + return SDValue(); + + // Avoid conversions where iN is larger than the float (e.g., float -> i64). + if (IntBits > FloatBits) + return SDValue(); + + BitVector UndefElements; + BuildVectorSDNode *BV = cast(ConstVec); + int32_t Bits = IntBits == 64 ? 64 : 32; + int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, Bits + 1); + if (C == -1 || C == 0 || C > Bits) + return SDValue(); + + MVT ResTy; + unsigned NumLanes = Op.getValueType().getVectorNumElements(); + switch (NumLanes) { + default: + return SDValue(); + case 2: + ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64; + break; + case 4: + ResTy = MVT::v4i32; + break; + } + + SDLoc DL(N); + bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT; + unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfp2fxs + : Intrinsic::aarch64_neon_vcvtfp2fxu; + SDValue FixConv = + DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ResTy, + DAG.getConstant(IntrinsicOpcode, DL, MVT::i32), + Op->getOperand(0), DAG.getConstant(C, DL, MVT::i32)); + // We can handle smaller integers by generating an extra trunc. + if (IntBits < FloatBits) + FixConv = DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), FixConv); + + return FixConv; +} + +/// Fold a floating-point divide by power of two into fixed-point to +/// floating-point conversion. +static SDValue performFDivCombine(SDNode *N, SelectionDAG &DAG, + const AArch64Subtarget *Subtarget) { + if (!Subtarget->hasNEON()) + return SDValue(); + + SDValue Op = N->getOperand(0); + unsigned Opc = Op->getOpcode(); + if (!Op.getValueType().isVector() || + (Opc != ISD::SINT_TO_FP && Opc != ISD::UINT_TO_FP)) + return SDValue(); + + SDValue ConstVec = N->getOperand(1); + if (!isa(ConstVec)) + return SDValue(); + + MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType(); + int32_t IntBits = IntTy.getSizeInBits(); + if (IntBits != 16 && IntBits != 32 && IntBits != 64) + return SDValue(); + + MVT FloatTy = N->getSimpleValueType(0).getVectorElementType(); + int32_t FloatBits = FloatTy.getSizeInBits(); + if (FloatBits != 32 && FloatBits != 64) + return SDValue(); + + // Avoid conversions where iN is larger than the float (e.g., i64 -> float). + if (IntBits > FloatBits) + return SDValue(); + + BitVector UndefElements; + BuildVectorSDNode *BV = cast(ConstVec); + int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, FloatBits + 1); + if (C == -1 || C == 0 || C > FloatBits) + return SDValue(); + + MVT ResTy; + unsigned NumLanes = Op.getValueType().getVectorNumElements(); + switch (NumLanes) { + default: + return SDValue(); + case 2: + ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64; + break; + case 4: + ResTy = MVT::v4i32; + break; + } + + SDLoc DL(N); + SDValue ConvInput = Op.getOperand(0); + bool IsSigned = Opc == ISD::SINT_TO_FP; + if (IntBits < FloatBits) + ConvInput = DAG.getNode(IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, DL, + ResTy, ConvInput); + + unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfxs2fp + : Intrinsic::aarch64_neon_vcvtfxu2fp; + return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(), + DAG.getConstant(IntrinsicOpcode, DL, MVT::i32), ConvInput, + DAG.getConstant(C, DL, MVT::i32)); +} + /// An EXTR instruction is made up of two shifts, ORed together. This helper /// searches for and classifies those shifts. static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount, @@ -8079,15 +8257,14 @@ static SDValue performAddSubLongCombine(SDNode *N, // (aarch64_neon_umull (extract_high (v2i64 vec))) // (extract_high (v2i64 (dup128 scalar))))) // -static SDValue tryCombineLongOpWithDup(SDNode *N, +static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N, TargetLowering::DAGCombinerInfo &DCI, SelectionDAG &DAG) { if (DCI.isBeforeLegalizeOps()) return SDValue(); - bool IsIntrinsic = N->getOpcode() == ISD::INTRINSIC_WO_CHAIN; - SDValue LHS = N->getOperand(IsIntrinsic ? 1 : 0); - SDValue RHS = N->getOperand(IsIntrinsic ? 2 : 1); + SDValue LHS = N->getOperand(1); + SDValue RHS = N->getOperand(2); assert(LHS.getValueType().is64BitVector() && RHS.getValueType().is64BitVector() && "unexpected shape for long operation"); @@ -8105,13 +8282,8 @@ static SDValue tryCombineLongOpWithDup(SDNode *N, return SDValue(); } - // N could either be an intrinsic or a sabsdiff/uabsdiff node. - if (IsIntrinsic) - return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0), - N->getOperand(0), LHS, RHS); - else - return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0), - LHS, RHS); + return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0), + N->getOperand(0), LHS, RHS); } static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) { @@ -8211,7 +8383,6 @@ static SDValue performIntrinsicCombine(SDNode *N, case Intrinsic::aarch64_neon_vcvtfxs2fp: case Intrinsic::aarch64_neon_vcvtfxu2fp: return tryCombineFixedPointConvert(N, DCI, DAG); - break; case Intrinsic::aarch64_neon_saddv: return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG); case Intrinsic::aarch64_neon_uaddv: @@ -8225,22 +8396,22 @@ static SDValue performIntrinsicCombine(SDNode *N, case Intrinsic::aarch64_neon_umaxv: return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG); case Intrinsic::aarch64_neon_fmax: - return DAG.getNode(AArch64ISD::FMAX, SDLoc(N), N->getValueType(0), + return DAG.getNode(ISD::FMAXNAN, SDLoc(N), N->getValueType(0), N->getOperand(1), N->getOperand(2)); case Intrinsic::aarch64_neon_fmin: - return DAG.getNode(AArch64ISD::FMIN, SDLoc(N), N->getValueType(0), + return DAG.getNode(ISD::FMINNAN, SDLoc(N), N->getValueType(0), N->getOperand(1), N->getOperand(2)); - case Intrinsic::aarch64_neon_sabd: - return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0), + case Intrinsic::aarch64_neon_fmaxnm: + return DAG.getNode(ISD::FMAXNUM, SDLoc(N), N->getValueType(0), N->getOperand(1), N->getOperand(2)); - case Intrinsic::aarch64_neon_uabd: - return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0), + case Intrinsic::aarch64_neon_fminnm: + return DAG.getNode(ISD::FMINNUM, SDLoc(N), N->getValueType(0), N->getOperand(1), N->getOperand(2)); case Intrinsic::aarch64_neon_smull: case Intrinsic::aarch64_neon_umull: case Intrinsic::aarch64_neon_pmull: case Intrinsic::aarch64_neon_sqdmull: - return tryCombineLongOpWithDup(N, DCI, DAG); + return tryCombineLongOpWithDup(IID, N, DCI, DAG); case Intrinsic::aarch64_neon_sqshl: case Intrinsic::aarch64_neon_uqshl: case Intrinsic::aarch64_neon_sqshlu: @@ -8265,15 +8436,18 @@ static SDValue performExtendCombine(SDNode *N, // helps the backend to decide that an sabdl2 would be useful, saving a real // extract_high operation. if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND && - (N->getOperand(0).getOpcode() == ISD::SABSDIFF || - N->getOperand(0).getOpcode() == ISD::UABSDIFF)) { + N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) { SDNode *ABDNode = N->getOperand(0).getNode(); - SDValue NewABD = tryCombineLongOpWithDup(ABDNode, DCI, DAG); - if (!NewABD.getNode()) - return SDValue(); + unsigned IID = getIntrinsicID(ABDNode); + if (IID == Intrinsic::aarch64_neon_sabd || + IID == Intrinsic::aarch64_neon_uabd) { + SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG); + if (!NewABD.getNode()) + return SDValue(); - return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0), - NewABD); + return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0), + NewABD); + } } // This is effectively a custom type legalization for AArch64. @@ -8391,7 +8565,7 @@ static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) { unsigned Alignment = std::min(OrigAlignment, EltOffset); // Create scalar stores. This is at least as good as the code sequence for a - // split unaligned store wich is a dup.s, ext.b, and two stores. + // split unaligned store which is a dup.s, ext.b, and two stores. // Most of the time the three stores should be replaced by store pair // instructions (stp). SDLoc DL(St); @@ -8412,10 +8586,9 @@ static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) { return NewST1; } -static SDValue performSTORECombine(SDNode *N, - TargetLowering::DAGCombinerInfo &DCI, - SelectionDAG &DAG, - const AArch64Subtarget *Subtarget) { +static SDValue split16BStores(SDNode *N, TargetLowering::DAGCombinerInfo &DCI, + SelectionDAG &DAG, + const AArch64Subtarget *Subtarget) { if (!DCI.isBeforeLegalize()) return SDValue(); @@ -8423,15 +8596,17 @@ static SDValue performSTORECombine(SDNode *N, if (S->isVolatile()) return SDValue(); + // FIXME: The logic for deciding if an unaligned store should be split should + // be included in TLI.allowsMisalignedMemoryAccesses(), and there should be + // a call to that function here. + // Cyclone has bad performance on unaligned 16B stores when crossing line and // page boundaries. We want to split such stores. if (!Subtarget->isCyclone()) return SDValue(); - // Don't split at Oz. - MachineFunction &MF = DAG.getMachineFunction(); - bool IsMinSize = MF.getFunction()->hasFnAttribute(Attribute::MinSize); - if (IsMinSize) + // Don't split at -Oz. + if (DAG.getMachineFunction().getFunction()->optForMinSize()) return SDValue(); SDValue StVal = S->getValue(); @@ -8454,8 +8629,7 @@ static SDValue performSTORECombine(SDNode *N, // If we get a splat of a scalar convert this vector store to a store of // scalars. They will be merged into store pairs thereby removing two // instructions. - SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S); - if (ReplacedSplat != SDValue()) + if (SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S)) return ReplacedSplat; SDLoc DL(S); @@ -8576,6 +8750,299 @@ static SDValue performPostLD1Combine(SDNode *N, return SDValue(); } +/// Simplify \Addr given that the top byte of it is ignored by HW during +/// address translation. +static bool performTBISimplification(SDValue Addr, + TargetLowering::DAGCombinerInfo &DCI, + SelectionDAG &DAG) { + APInt DemandedMask = APInt::getLowBitsSet(64, 56); + APInt KnownZero, KnownOne; + TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(), + DCI.isBeforeLegalizeOps()); + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); + if (TLI.SimplifyDemandedBits(Addr, DemandedMask, KnownZero, KnownOne, TLO)) { + DCI.CommitTargetLoweringOpt(TLO); + return true; + } + return false; +} + +static SDValue performSTORECombine(SDNode *N, + TargetLowering::DAGCombinerInfo &DCI, + SelectionDAG &DAG, + const AArch64Subtarget *Subtarget) { + SDValue Split = split16BStores(N, DCI, DAG, Subtarget); + if (Split.getNode()) + return Split; + + if (Subtarget->supportsAddressTopByteIgnored() && + performTBISimplification(N->getOperand(2), DCI, DAG)) + return SDValue(N, 0); + + return SDValue(); +} + + /// This function handles the log2-shuffle pattern produced by the +/// LoopVectorizer for the across vector reduction. It consists of +/// log2(NumVectorElements) steps and, in each step, 2^(s) elements +/// are reduced, where s is an induction variable from 0 to +/// log2(NumVectorElements). +static SDValue tryMatchAcrossLaneShuffleForReduction(SDNode *N, SDValue OpV, + unsigned Op, + SelectionDAG &DAG) { + EVT VTy = OpV->getOperand(0).getValueType(); + if (!VTy.isVector()) + return SDValue(); + + int NumVecElts = VTy.getVectorNumElements(); + if (Op == ISD::FMAXNUM || Op == ISD::FMINNUM) { + if (NumVecElts != 4) + return SDValue(); + } else { + if (NumVecElts != 4 && NumVecElts != 8 && NumVecElts != 16) + return SDValue(); + } + + int NumExpectedSteps = APInt(8, NumVecElts).logBase2(); + SDValue PreOp = OpV; + // Iterate over each step of the across vector reduction. + for (int CurStep = 0; CurStep != NumExpectedSteps; ++CurStep) { + SDValue CurOp = PreOp.getOperand(0); + SDValue Shuffle = PreOp.getOperand(1); + if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE) { + // Try to swap the 1st and 2nd operand as add and min/max instructions + // are commutative. + CurOp = PreOp.getOperand(1); + Shuffle = PreOp.getOperand(0); + if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE) + return SDValue(); + } + + // Check if the input vector is fed by the operator we want to handle, + // except the last step; the very first input vector is not necessarily + // the same operator we are handling. + if (CurOp.getOpcode() != Op && (CurStep != (NumExpectedSteps - 1))) + return SDValue(); + + // Check if it forms one step of the across vector reduction. + // E.g., + // %cur = add %1, %0 + // %shuffle = vector_shuffle %cur, <2, 3, u, u> + // %pre = add %cur, %shuffle + if (Shuffle.getOperand(0) != CurOp) + return SDValue(); + + int NumMaskElts = 1 << CurStep; + ArrayRef Mask = cast(Shuffle)->getMask(); + // Check mask values in each step. + // We expect the shuffle mask in each step follows a specific pattern + // denoted here by the form, where M is a sequence of integers + // starting from NumMaskElts, increasing by 1, and the number integers + // in M should be NumMaskElts. U is a sequence of UNDEFs and the number + // of undef in U should be NumVecElts - NumMaskElts. + // E.g., for <8 x i16>, mask values in each step should be : + // step 0 : <1,u,u,u,u,u,u,u> + // step 1 : <2,3,u,u,u,u,u,u> + // step 2 : <4,5,6,7,u,u,u,u> + for (int i = 0; i < NumVecElts; ++i) + if ((i < NumMaskElts && Mask[i] != (NumMaskElts + i)) || + (i >= NumMaskElts && !(Mask[i] < 0))) + return SDValue(); + + PreOp = CurOp; + } + unsigned Opcode; + bool IsIntrinsic = false; + + switch (Op) { + default: + llvm_unreachable("Unexpected operator for across vector reduction"); + case ISD::ADD: + Opcode = AArch64ISD::UADDV; + break; + case ISD::SMAX: + Opcode = AArch64ISD::SMAXV; + break; + case ISD::UMAX: + Opcode = AArch64ISD::UMAXV; + break; + case ISD::SMIN: + Opcode = AArch64ISD::SMINV; + break; + case ISD::UMIN: + Opcode = AArch64ISD::UMINV; + break; + case ISD::FMAXNUM: + Opcode = Intrinsic::aarch64_neon_fmaxnmv; + IsIntrinsic = true; + break; + case ISD::FMINNUM: + Opcode = Intrinsic::aarch64_neon_fminnmv; + IsIntrinsic = true; + break; + } + SDLoc DL(N); + + return IsIntrinsic + ? DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, N->getValueType(0), + DAG.getConstant(Opcode, DL, MVT::i32), PreOp) + : DAG.getNode( + ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0), + DAG.getNode(Opcode, DL, PreOp.getSimpleValueType(), PreOp), + DAG.getConstant(0, DL, MVT::i64)); +} + +/// Target-specific DAG combine for the across vector min/max reductions. +/// This function specifically handles the final clean-up step of the vector +/// min/max reductions produced by the LoopVectorizer. It is the log2-shuffle +/// pattern, which narrows down and finds the final min/max value from all +/// elements of the vector. +/// For example, for a <16 x i8> vector : +/// svn0 = vector_shuffle %0, undef<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> +/// %smax0 = smax %arr, svn0 +/// %svn1 = vector_shuffle %smax0, undef<4,5,6,7,u,u,u,u,u,u,u,u,u,u,u,u> +/// %smax1 = smax %smax0, %svn1 +/// %svn2 = vector_shuffle %smax1, undef<2,3,u,u,u,u,u,u,u,u,u,u,u,u,u,u> +/// %smax2 = smax %smax1, svn2 +/// %svn3 = vector_shuffle %smax2, undef<1,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u> +/// %sc = setcc %smax2, %svn3, gt +/// %n0 = extract_vector_elt %sc, #0 +/// %n1 = extract_vector_elt %smax2, #0 +/// %n2 = extract_vector_elt $smax2, #1 +/// %result = select %n0, %n1, n2 +/// becomes : +/// %1 = smaxv %0 +/// %result = extract_vector_elt %1, 0 +static SDValue +performAcrossLaneMinMaxReductionCombine(SDNode *N, SelectionDAG &DAG, + const AArch64Subtarget *Subtarget) { + if (!Subtarget->hasNEON()) + return SDValue(); + + SDValue N0 = N->getOperand(0); + SDValue IfTrue = N->getOperand(1); + SDValue IfFalse = N->getOperand(2); + + // Check if the SELECT merges up the final result of the min/max + // from a vector. + if (N0.getOpcode() != ISD::EXTRACT_VECTOR_ELT || + IfTrue.getOpcode() != ISD::EXTRACT_VECTOR_ELT || + IfFalse.getOpcode() != ISD::EXTRACT_VECTOR_ELT) + return SDValue(); + + // Expect N0 is fed by SETCC. + SDValue SetCC = N0.getOperand(0); + EVT SetCCVT = SetCC.getValueType(); + if (SetCC.getOpcode() != ISD::SETCC || !SetCCVT.isVector() || + SetCCVT.getVectorElementType() != MVT::i1) + return SDValue(); + + SDValue VectorOp = SetCC.getOperand(0); + unsigned Op = VectorOp->getOpcode(); + // Check if the input vector is fed by the operator we want to handle. + if (Op != ISD::SMAX && Op != ISD::UMAX && Op != ISD::SMIN && + Op != ISD::UMIN && Op != ISD::FMAXNUM && Op != ISD::FMINNUM) + return SDValue(); + + EVT VTy = VectorOp.getValueType(); + if (!VTy.isVector()) + return SDValue(); + + if (VTy.getSizeInBits() < 64) + return SDValue(); + + EVT EltTy = VTy.getVectorElementType(); + if (Op == ISD::FMAXNUM || Op == ISD::FMINNUM) { + if (EltTy != MVT::f32) + return SDValue(); + } else { + if (EltTy != MVT::i32 && EltTy != MVT::i16 && EltTy != MVT::i8) + return SDValue(); + } + + // Check if extracting from the same vector. + // For example, + // %sc = setcc %vector, %svn1, gt + // %n0 = extract_vector_elt %sc, #0 + // %n1 = extract_vector_elt %vector, #0 + // %n2 = extract_vector_elt $vector, #1 + if (!(VectorOp == IfTrue->getOperand(0) && + VectorOp == IfFalse->getOperand(0))) + return SDValue(); + + // Check if the condition code is matched with the operator type. + ISD::CondCode CC = cast(SetCC->getOperand(2))->get(); + if ((Op == ISD::SMAX && CC != ISD::SETGT && CC != ISD::SETGE) || + (Op == ISD::UMAX && CC != ISD::SETUGT && CC != ISD::SETUGE) || + (Op == ISD::SMIN && CC != ISD::SETLT && CC != ISD::SETLE) || + (Op == ISD::UMIN && CC != ISD::SETULT && CC != ISD::SETULE) || + (Op == ISD::FMAXNUM && CC != ISD::SETOGT && CC != ISD::SETOGE && + CC != ISD::SETUGT && CC != ISD::SETUGE && CC != ISD::SETGT && + CC != ISD::SETGE) || + (Op == ISD::FMINNUM && CC != ISD::SETOLT && CC != ISD::SETOLE && + CC != ISD::SETULT && CC != ISD::SETULE && CC != ISD::SETLT && + CC != ISD::SETLE)) + return SDValue(); + + // Expect to check only lane 0 from the vector SETCC. + if (!isNullConstant(N0.getOperand(1))) + return SDValue(); + + // Expect to extract the true value from lane 0. + if (!isNullConstant(IfTrue.getOperand(1))) + return SDValue(); + + // Expect to extract the false value from lane 1. + if (!isOneConstant(IfFalse.getOperand(1))) + return SDValue(); + + return tryMatchAcrossLaneShuffleForReduction(N, SetCC, Op, DAG); +} + +/// Target-specific DAG combine for the across vector add reduction. +/// This function specifically handles the final clean-up step of the vector +/// add reduction produced by the LoopVectorizer. It is the log2-shuffle +/// pattern, which adds all elements of a vector together. +/// For example, for a <4 x i32> vector : +/// %1 = vector_shuffle %0, <2,3,u,u> +/// %2 = add %0, %1 +/// %3 = vector_shuffle %2, <1,u,u,u> +/// %4 = add %2, %3 +/// %result = extract_vector_elt %4, 0 +/// becomes : +/// %0 = uaddv %0 +/// %result = extract_vector_elt %0, 0 +static SDValue +performAcrossLaneAddReductionCombine(SDNode *N, SelectionDAG &DAG, + const AArch64Subtarget *Subtarget) { + if (!Subtarget->hasNEON()) + return SDValue(); + SDValue N0 = N->getOperand(0); + SDValue N1 = N->getOperand(1); + + // Check if the input vector is fed by the ADD. + if (N0->getOpcode() != ISD::ADD) + return SDValue(); + + // The vector extract idx must constant zero because we only expect the final + // result of the reduction is placed in lane 0. + if (!isNullConstant(N1)) + return SDValue(); + + EVT VTy = N0.getValueType(); + if (!VTy.isVector()) + return SDValue(); + + EVT EltTy = VTy.getVectorElementType(); + if (EltTy != MVT::i32 && EltTy != MVT::i16 && EltTy != MVT::i8) + return SDValue(); + + if (VTy.getSizeInBits() < 64) + return SDValue(); + + return tryMatchAcrossLaneShuffleForReduction(N, N0, ISD::ADD, DAG); +} + /// Target-specific DAG combine function for NEON load/store intrinsics /// to merge base address updates. static SDValue performNEONPostLDSTCombine(SDNode *N, @@ -9001,10 +9468,10 @@ static SDValue performBRCONDCombine(SDNode *N, if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64) return SDValue(); - if (isa(LHS) && cast(LHS)->isNullValue()) + if (isNullConstant(LHS)) std::swap(LHS, RHS); - if (!isa(RHS) || !cast(RHS)->isNullValue()) + if (!isNullConstant(RHS)) return SDValue(); if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA || @@ -9024,6 +9491,103 @@ static SDValue performBRCONDCombine(SDNode *N, return SDValue(); } +// Optimize some simple tbz/tbnz cases. Returns the new operand and bit to test +// as well as whether the test should be inverted. This code is required to +// catch these cases (as opposed to standard dag combines) because +// AArch64ISD::TBZ is matched during legalization. +static SDValue getTestBitOperand(SDValue Op, unsigned &Bit, bool &Invert, + SelectionDAG &DAG) { + + if (!Op->hasOneUse()) + return Op; + + // We don't handle undef/constant-fold cases below, as they should have + // already been taken care of (e.g. and of 0, test of undefined shifted bits, + // etc.) + + // (tbz (trunc x), b) -> (tbz x, b) + // This case is just here to enable more of the below cases to be caught. + if (Op->getOpcode() == ISD::TRUNCATE && + Bit < Op->getValueType(0).getSizeInBits()) { + return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); + } + + if (Op->getNumOperands() != 2) + return Op; + + auto *C = dyn_cast(Op->getOperand(1)); + if (!C) + return Op; + + switch (Op->getOpcode()) { + default: + return Op; + + // (tbz (and x, m), b) -> (tbz x, b) + case ISD::AND: + if ((C->getZExtValue() >> Bit) & 1) + return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); + return Op; + + // (tbz (shl x, c), b) -> (tbz x, b-c) + case ISD::SHL: + if (C->getZExtValue() <= Bit && + (Bit - C->getZExtValue()) < Op->getValueType(0).getSizeInBits()) { + Bit = Bit - C->getZExtValue(); + return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); + } + return Op; + + // (tbz (sra x, c), b) -> (tbz x, b+c) or (tbz x, msb) if b+c is > # bits in x + case ISD::SRA: + Bit = Bit + C->getZExtValue(); + if (Bit >= Op->getValueType(0).getSizeInBits()) + Bit = Op->getValueType(0).getSizeInBits() - 1; + return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); + + // (tbz (srl x, c), b) -> (tbz x, b+c) + case ISD::SRL: + if ((Bit + C->getZExtValue()) < Op->getValueType(0).getSizeInBits()) { + Bit = Bit + C->getZExtValue(); + return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); + } + return Op; + + // (tbz (xor x, -1), b) -> (tbnz x, b) + case ISD::XOR: + if ((C->getZExtValue() >> Bit) & 1) + Invert = !Invert; + return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); + } +} + +// Optimize test single bit zero/non-zero and branch. +static SDValue performTBZCombine(SDNode *N, + TargetLowering::DAGCombinerInfo &DCI, + SelectionDAG &DAG) { + unsigned Bit = cast(N->getOperand(2))->getZExtValue(); + bool Invert = false; + SDValue TestSrc = N->getOperand(1); + SDValue NewTestSrc = getTestBitOperand(TestSrc, Bit, Invert, DAG); + + if (TestSrc == NewTestSrc) + return SDValue(); + + unsigned NewOpc = N->getOpcode(); + if (Invert) { + if (NewOpc == AArch64ISD::TBZ) + NewOpc = AArch64ISD::TBNZ; + else { + assert(NewOpc == AArch64ISD::TBNZ); + NewOpc = AArch64ISD::TBZ; + } + } + + SDLoc DL(N); + return DAG.getNode(NewOpc, DL, MVT::Other, N->getOperand(0), NewTestSrc, + DAG.getConstant(Bit, DL, MVT::i64), N->getOperand(3)); +} + // 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 @@ -9118,75 +9682,6 @@ static SDValue performSelectCombine(SDNode *N, return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2)); } -/// performSelectCCCombine - Target-specific DAG combining for ISD::SELECT_CC -/// to match FMIN/FMAX patterns. -static SDValue performSelectCCCombine(SDNode *N, SelectionDAG &DAG) { - // Try to use FMIN/FMAX instructions for FP selects like "x < y ? x : y". - // Unless the NoNaNsFPMath option is set, be careful about NaNs: - // vmax/vmin return NaN if either operand is a NaN; - // only do the transformation when it matches that behavior. - - SDValue CondLHS = N->getOperand(0); - SDValue CondRHS = N->getOperand(1); - SDValue LHS = N->getOperand(2); - SDValue RHS = N->getOperand(3); - ISD::CondCode CC = cast(N->getOperand(4))->get(); - - unsigned Opcode; - bool IsReversed; - if (selectCCOpsAreFMaxCompatible(CondLHS, LHS) && - selectCCOpsAreFMaxCompatible(CondRHS, RHS)) { - IsReversed = false; // x CC y ? x : y - } else if (selectCCOpsAreFMaxCompatible(CondRHS, LHS) && - selectCCOpsAreFMaxCompatible(CondLHS, RHS)) { - IsReversed = true ; // x CC y ? y : x - } else { - return SDValue(); - } - - bool IsUnordered = false, IsOrEqual; - switch (CC) { - default: - return SDValue(); - case ISD::SETULT: - case ISD::SETULE: - IsUnordered = true; - case ISD::SETOLT: - case ISD::SETOLE: - case ISD::SETLT: - case ISD::SETLE: - IsOrEqual = (CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE); - Opcode = IsReversed ? AArch64ISD::FMAX : AArch64ISD::FMIN; - break; - - case ISD::SETUGT: - case ISD::SETUGE: - IsUnordered = true; - case ISD::SETOGT: - case ISD::SETOGE: - case ISD::SETGT: - case ISD::SETGE: - IsOrEqual = (CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE); - Opcode = IsReversed ? AArch64ISD::FMIN : AArch64ISD::FMAX; - break; - } - - // If LHS is NaN, an ordered comparison will be false and the result will be - // the RHS, but FMIN(NaN, RHS) = FMAX(NaN, RHS) = NaN. Avoid this by checking - // that LHS != NaN. Likewise, for unordered comparisons, check for RHS != NaN. - if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS)) - return SDValue(); - - // For xxx-or-equal comparisons, "+0 <= -0" and "-0 >= +0" will both be true, - // but FMIN will return -0, and FMAX will return +0. So FMIN/FMAX can only be - // used for unsafe math or if one of the operands is known to be nonzero. - if (IsOrEqual && !DAG.getTarget().Options.UnsafeFPMath && - !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) - return SDValue(); - - return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS); -} - /// Get rid of unnecessary NVCASTs (that don't change the type). static SDValue performNVCASTCombine(SDNode *N) { if (N->getValueType(0) == N->getOperand(0).getValueType()) @@ -9211,6 +9706,11 @@ SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N, case ISD::SINT_TO_FP: case ISD::UINT_TO_FP: return performIntToFpCombine(N, DAG, Subtarget); + case ISD::FP_TO_SINT: + case ISD::FP_TO_UINT: + return performFpToIntCombine(N, DAG, Subtarget); + case ISD::FDIV: + return performFDivCombine(N, DAG, Subtarget); case ISD::OR: return performORCombine(N, DCI, Subtarget); case ISD::INTRINSIC_WO_CHAIN: @@ -9223,16 +9723,25 @@ SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N, return performBitcastCombine(N, DCI, DAG); case ISD::CONCAT_VECTORS: return performConcatVectorsCombine(N, DCI, DAG); - case ISD::SELECT: - return performSelectCombine(N, DCI); + case ISD::SELECT: { + SDValue RV = performSelectCombine(N, DCI); + if (!RV.getNode()) + RV = performAcrossLaneMinMaxReductionCombine(N, DAG, Subtarget); + return RV; + } case ISD::VSELECT: return performVSelectCombine(N, DCI.DAG); - case ISD::SELECT_CC: - return performSelectCCCombine(N, DCI.DAG); + case ISD::LOAD: + if (performTBISimplification(N->getOperand(1), DCI, DAG)) + return SDValue(N, 0); + break; case ISD::STORE: return performSTORECombine(N, DCI, DAG, Subtarget); case AArch64ISD::BRCOND: return performBRCONDCombine(N, DCI, DAG); + case AArch64ISD::TBNZ: + case AArch64ISD::TBZ: + return performTBZCombine(N, DCI, DAG); case AArch64ISD::CSEL: return performCONDCombine(N, DCI, DAG, 2, 3); case AArch64ISD::DUP: @@ -9241,6 +9750,8 @@ SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N, return performNVCASTCombine(N); case ISD::INSERT_VECTOR_ELT: return performPostLD1Combine(N, DCI, true); + case ISD::EXTRACT_VECTOR_ELT: + return performAcrossLaneAddReductionCombine(N, DAG, Subtarget); case ISD::INTRINSIC_VOID: case ISD::INTRINSIC_W_CHAIN: switch (cast(N->getOperand(1))->getZExtValue()) { @@ -9407,6 +9918,20 @@ static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl &Results, Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op)); } +static void ReplaceReductionResults(SDNode *N, + SmallVectorImpl &Results, + SelectionDAG &DAG, unsigned InterOp, + unsigned AcrossOp) { + EVT LoVT, HiVT; + SDValue Lo, Hi; + SDLoc dl(N); + std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0)); + std::tie(Lo, Hi) = DAG.SplitVectorOperand(N, 0); + SDValue InterVal = DAG.getNode(InterOp, dl, LoVT, Lo, Hi); + SDValue SplitVal = DAG.getNode(AcrossOp, dl, LoVT, InterVal); + Results.push_back(SplitVal); +} + void AArch64TargetLowering::ReplaceNodeResults( SDNode *N, SmallVectorImpl &Results, SelectionDAG &DAG) const { switch (N->getOpcode()) { @@ -9415,6 +9940,24 @@ void AArch64TargetLowering::ReplaceNodeResults( case ISD::BITCAST: ReplaceBITCASTResults(N, Results, DAG); return; + case AArch64ISD::SADDV: + ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::SADDV); + return; + case AArch64ISD::UADDV: + ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::UADDV); + return; + case AArch64ISD::SMINV: + ReplaceReductionResults(N, Results, DAG, ISD::SMIN, AArch64ISD::SMINV); + return; + case AArch64ISD::UMINV: + ReplaceReductionResults(N, Results, DAG, ISD::UMIN, AArch64ISD::UMINV); + return; + case AArch64ISD::SMAXV: + ReplaceReductionResults(N, Results, DAG, ISD::SMAX, AArch64ISD::SMAXV); + return; + case AArch64ISD::UMAXV: + ReplaceReductionResults(N, Results, DAG, ISD::UMAX, AArch64ISD::UMAXV); + return; case ISD::FP_TO_UINT: case ISD::FP_TO_SINT: assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion"); @@ -9427,10 +9970,10 @@ bool AArch64TargetLowering::useLoadStackGuardNode() const { return true; } -bool AArch64TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const { +unsigned AArch64TargetLowering::combineRepeatedFPDivisors() const { // Combine multiple FDIVs with the same divisor into multiple FMULs by the // reciprocal if there are three or more FDIVs. - return NumUsers > 2; + return 3; } TargetLoweringBase::LegalizeTypeAction @@ -9456,20 +9999,21 @@ bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const { // Loads and stores less than 128-bits are already atomic; ones above that // are doomed anyway, so defer to the default libcall and blame the OS when // things go wrong. -bool AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const { +TargetLowering::AtomicExpansionKind +AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const { unsigned Size = LI->getType()->getPrimitiveSizeInBits(); - return Size == 128; + return Size == 128 ? AtomicExpansionKind::LLSC : AtomicExpansionKind::None; } // For the real atomic operations, we have ldxr/stxr up to 128 bits, -TargetLoweringBase::AtomicRMWExpansionKind +TargetLowering::AtomicExpansionKind AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const { unsigned Size = AI->getType()->getPrimitiveSizeInBits(); - return Size <= 128 ? AtomicRMWExpansionKind::LLSC - : AtomicRMWExpansionKind::None; + return Size <= 128 ? AtomicExpansionKind::LLSC : AtomicExpansionKind::None; } -bool AArch64TargetLowering::hasLoadLinkedStoreConditional() const { +bool AArch64TargetLowering::shouldExpandAtomicCmpXchgInIR( + AtomicCmpXchgInst *AI) const { return true; } @@ -9508,6 +10052,13 @@ Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr, cast(Addr->getType())->getElementType()); } +void AArch64TargetLowering::emitAtomicCmpXchgNoStoreLLBalance( + IRBuilder<> &Builder) const { + Module *M = Builder.GetInsertBlock()->getParent()->getParent(); + Builder.CreateCall( + llvm::Intrinsic::getDeclaration(M, Intrinsic::aarch64_clrex)); +} + Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val, Value *Addr, AtomicOrdering Ord) const { @@ -9549,3 +10100,65 @@ bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &, EVT) const { return false; } + +Value *AArch64TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const { + if (!Subtarget->isTargetAndroid()) + return TargetLowering::getSafeStackPointerLocation(IRB); + + // Android provides a fixed TLS slot for the SafeStack pointer. See the + // definition of TLS_SLOT_SAFESTACK in + // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h + const unsigned TlsOffset = 0x48; + Module *M = IRB.GetInsertBlock()->getParent()->getParent(); + Function *ThreadPointerFunc = + Intrinsic::getDeclaration(M, Intrinsic::aarch64_thread_pointer); + return IRB.CreatePointerCast( + IRB.CreateConstGEP1_32(IRB.CreateCall(ThreadPointerFunc), TlsOffset), + Type::getInt8PtrTy(IRB.getContext())->getPointerTo(0)); +} + +void AArch64TargetLowering::initializeSplitCSR(MachineBasicBlock *Entry) const { + // Update IsSplitCSR in AArch64unctionInfo. + AArch64FunctionInfo *AFI = Entry->getParent()->getInfo(); + AFI->setIsSplitCSR(true); +} + +void AArch64TargetLowering::insertCopiesSplitCSR( + MachineBasicBlock *Entry, + const SmallVectorImpl &Exits) const { + const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo(); + const MCPhysReg *IStart = TRI->getCalleeSavedRegsViaCopy(Entry->getParent()); + if (!IStart) + return; + + const TargetInstrInfo *TII = Subtarget->getInstrInfo(); + MachineRegisterInfo *MRI = &Entry->getParent()->getRegInfo(); + for (const MCPhysReg *I = IStart; *I; ++I) { + const TargetRegisterClass *RC = nullptr; + if (AArch64::GPR64RegClass.contains(*I)) + RC = &AArch64::GPR64RegClass; + else if (AArch64::FPR64RegClass.contains(*I)) + RC = &AArch64::FPR64RegClass; + else + llvm_unreachable("Unexpected register class in CSRsViaCopy!"); + + unsigned NewVR = MRI->createVirtualRegister(RC); + // Create copy from CSR to a virtual register. + // FIXME: this currently does not emit CFI pseudo-instructions, it works + // fine for CXX_FAST_TLS since the C++-style TLS access functions should be + // nounwind. If we want to generalize this later, we may need to emit + // CFI pseudo-instructions. + assert(Entry->getParent()->getFunction()->hasFnAttribute( + Attribute::NoUnwind) && + "Function should be nounwind in insertCopiesSplitCSR!"); + Entry->addLiveIn(*I); + BuildMI(*Entry, Entry->begin(), DebugLoc(), TII->get(TargetOpcode::COPY), + NewVR) + .addReg(*I); + + for (auto *Exit : Exits) + BuildMI(*Exit, Exit->begin(), DebugLoc(), TII->get(TargetOpcode::COPY), + *I) + .addReg(NewVR); + } +}