1 //==-- AArch64ISelLowering.h - AArch64 DAG Lowering Interface ----*- C++ -*-==//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that AArch64 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_TARGET_AARCH64_ISELLOWERING_H
16 #define LLVM_TARGET_AARCH64_ISELLOWERING_H
18 #include "Utils/AArch64BaseInfo.h"
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/Target/TargetLowering.h"
22 #include "llvm/IR/Intrinsics.h"
25 namespace AArch64ISD {
27 // Start the numbering from where ISD NodeType finishes.
28 FIRST_NUMBER = ISD::BUILTIN_OP_END,
30 // This is a conditional branch which also notes the flag needed
31 // (eq/sgt/...). A64 puts this information on the branches rather than
32 // compares as LLVM does.
35 // A node to be selected to an actual call operation: either BL or BLR in
36 // the absence of tail calls.
39 // Indicates a floating-point immediate which fits into the format required
40 // by the FMOV instructions. First (and only) operand is the 8-bit encoded
41 // value of that immediate.
44 // Corresponds directly to an EXTR instruction. Operands are an LHS an RHS
48 // Wraps a load from the GOT, which should always be performed with a 64-bit
49 // load instruction. This prevents the DAG combiner folding a truncate to
50 // form a smaller memory access.
53 // Performs a bitfield insert. Arguments are: the value being inserted into;
54 // the value being inserted; least significant bit changed; width of the
58 // Simply a convenient node inserted during ISelLowering to represent
59 // procedure return. Will almost certainly be selected to "RET".
62 /// Extracts a field of contiguous bits from the source and sign extends
63 /// them into a single register. Arguments are: source; immr; imms. Note
64 /// these are pre-encoded since DAG matching can't cope with combining LSB
65 /// and Width into these values itself.
68 /// This is an A64-ification of the standard LLVM SELECT_CC operation. The
69 /// main difference is that it only has the values and an A64 condition,
70 /// which will be produced by a setcc instruction.
73 /// This serves most of the functions of the LLVM SETCC instruction, for two
74 /// purposes. First, it prevents optimisations from fiddling with the
75 /// compare after we've moved the CondCode information onto the SELECT_CC or
76 /// BR_CC instructions. Second, it gives a legal instruction for the actual
79 /// It keeps a record of the condition flags asked for because certain
80 /// instructions are only valid for a subset of condition codes.
83 // Designates a node which is a tail call: both a call and a return
84 // instruction as far as selction is concerned. It should be selected to an
85 // unconditional branch. Has the usual plethora of call operands, but: 1st
86 // is callee, 2nd is stack adjustment required immediately before branch.
89 // Designates a call used to support the TLS descriptor ABI. The call itself
90 // will be indirect ("BLR xN") but a relocation-specifier (".tlsdesccall
91 // var") must be attached somehow during code generation. It takes two
92 // operands: the callee and the symbol to be relocated against.
95 // Leaf node which will be lowered to an appropriate MRS to obtain the
96 // thread pointer: TPIDR_EL0.
99 /// Extracts a field of contiguous bits from the source and zero extends
100 /// them into a single register. Arguments are: source; immr; imms. Note
101 /// these are pre-encoded since DAG matching can't cope with combining LSB
102 /// and Width into these values itself.
105 // Wraps an address which the ISelLowering phase has decided should be
106 // created using the large memory model style: i.e. a sequence of four
107 // movz/movk instructions.
110 // Wraps an address which the ISelLowering phase has decided should be
111 // created using the small memory model style: i.e. adrp/add or
112 // adrp/mem-op. This exists to prevent bare TargetAddresses which may never
116 // Vector bitwise select
119 // Vector move immediate
122 // Vector Move Inverted Immediate
125 // Vector FP move immediate
136 // Vector Element reverse
144 // Vector compare zero
147 // Vector compare bitwise test
150 // Vector saturating shift
157 // Vector dup by lane
163 // NEON duplicate lane loads
164 NEON_LD2DUP = ISD::FIRST_TARGET_MEMORY_OPCODE,
168 // NEON loads with post-increment base updates:
177 // NEON stores with post-increment base updates:
186 // NEON duplicate lane loads with post-increment base updates:
191 // NEON lane loads with post-increment base updates:
196 // NEON lane store with post-increment base updates:
204 class AArch64Subtarget;
205 class AArch64TargetMachine;
207 class AArch64TargetLowering : public TargetLowering {
209 explicit AArch64TargetLowering(AArch64TargetMachine &TM);
211 const char *getTargetNodeName(unsigned Opcode) const;
213 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC) const;
215 SDValue LowerFormalArguments(SDValue Chain,
216 CallingConv::ID CallConv, bool isVarArg,
217 const SmallVectorImpl<ISD::InputArg> &Ins,
218 SDLoc dl, SelectionDAG &DAG,
219 SmallVectorImpl<SDValue> &InVals) const;
221 SDValue LowerReturn(SDValue Chain,
222 CallingConv::ID CallConv, bool isVarArg,
223 const SmallVectorImpl<ISD::OutputArg> &Outs,
224 const SmallVectorImpl<SDValue> &OutVals,
225 SDLoc dl, SelectionDAG &DAG) const;
227 SDValue LowerCall(CallLoweringInfo &CLI,
228 SmallVectorImpl<SDValue> &InVals) const;
230 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
231 CallingConv::ID CallConv, bool IsVarArg,
232 const SmallVectorImpl<ISD::InputArg> &Ins,
233 SDLoc dl, SelectionDAG &DAG,
234 SmallVectorImpl<SDValue> &InVals) const;
236 bool isKnownShuffleVector(SDValue Op, SelectionDAG &DAG, SDValue &Res) const;
238 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
239 const AArch64Subtarget *ST) const;
241 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
243 void SaveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG, SDLoc DL,
244 SDValue &Chain) const;
246 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
247 /// for tail call optimization. Targets which want to do tail call
248 /// optimization should implement this function.
249 bool IsEligibleForTailCallOptimization(SDValue Callee,
250 CallingConv::ID CalleeCC,
252 bool IsCalleeStructRet,
253 bool IsCallerStructRet,
254 const SmallVectorImpl<ISD::OutputArg> &Outs,
255 const SmallVectorImpl<SDValue> &OutVals,
256 const SmallVectorImpl<ISD::InputArg> &Ins,
257 SelectionDAG& DAG) const;
259 /// Finds the incoming stack arguments which overlap the given fixed stack
260 /// object and incorporates their load into the current chain. This prevents
261 /// an upcoming store from clobbering the stack argument before it's used.
262 SDValue addTokenForArgument(SDValue Chain, SelectionDAG &DAG,
263 MachineFrameInfo *MFI, int ClobberedFI) const;
265 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
267 bool DoesCalleeRestoreStack(CallingConv::ID CallCC, bool TailCallOpt) const;
269 bool IsTailCallConvention(CallingConv::ID CallCC) const;
271 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
273 bool isLegalICmpImmediate(int64_t Val) const;
274 SDValue getSelectableIntSetCC(SDValue LHS, SDValue RHS, ISD::CondCode CC,
275 SDValue &A64cc, SelectionDAG &DAG, SDLoc &dl) const;
277 virtual MachineBasicBlock *
278 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const;
281 emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *MBB,
282 unsigned Size, unsigned Opcode) const;
285 emitAtomicBinaryMinMax(MachineInstr *MI, MachineBasicBlock *BB,
286 unsigned Size, unsigned CmpOp,
287 A64CC::CondCodes Cond) const;
289 emitAtomicCmpSwap(MachineInstr *MI, MachineBasicBlock *BB,
290 unsigned Size) const;
293 EmitF128CSEL(MachineInstr *MI, MachineBasicBlock *MBB) const;
295 SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) const;
296 SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) const;
297 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
298 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
299 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
300 SDValue LowerF128ToCall(SDValue Op, SelectionDAG &DAG,
301 RTLIB::Libcall Call) const;
302 SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
303 SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
304 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG, bool IsSigned) const;
305 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
306 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
308 SDValue LowerGlobalAddressELFSmall(SDValue Op, SelectionDAG &DAG) const;
309 SDValue LowerGlobalAddressELFLarge(SDValue Op, SelectionDAG &DAG) const;
310 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const;
312 SDValue LowerTLSDescCall(SDValue SymAddr, SDValue DescAddr, SDLoc DL,
313 SelectionDAG &DAG) const;
314 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
315 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG, bool IsSigned) const;
316 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
317 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
318 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
319 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
320 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
321 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
323 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
325 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
326 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
327 /// expanded to FMAs when this method returns true, otherwise fmuladd is
328 /// expanded to fmul + fadd.
329 virtual bool isFMAFasterThanFMulAndFAdd(EVT VT) const;
331 ConstraintType getConstraintType(const std::string &Constraint) const;
333 ConstraintWeight getSingleConstraintMatchWeight(AsmOperandInfo &Info,
334 const char *Constraint) const;
335 void LowerAsmOperandForConstraint(SDValue Op,
336 std::string &Constraint,
337 std::vector<SDValue> &Ops,
338 SelectionDAG &DAG) const;
340 std::pair<unsigned, const TargetRegisterClass*>
341 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const;
343 virtual bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
344 unsigned Intrinsic) const LLVM_OVERRIDE;
347 const InstrItineraryData *Itins;
349 const AArch64Subtarget *getSubtarget() const {
350 return &getTargetMachine().getSubtarget<AArch64Subtarget>();
353 enum NeonModImmType {
358 extern SDValue ScanBUILD_VECTOR(SDValue Op, bool &isOnlyLowElement,
359 bool &usesOnlyOneValue, bool &hasDominantValue,
360 bool &isConstant, bool &isUNDEF);
363 #endif // LLVM_TARGET_AARCH64_ISELLOWERING_H