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_LIB_TARGET_AARCH64_AARCH64ISELLOWERING_H
16 #define LLVM_LIB_TARGET_AARCH64_AARCH64ISELLOWERING_H
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/IR/CallingConv.h"
22 #include "llvm/IR/Instruction.h"
23 #include "llvm/Target/TargetLowering.h"
27 namespace AArch64ISD {
29 enum NodeType : unsigned {
30 FIRST_NUMBER = ISD::BUILTIN_OP_END,
31 WrapperLarge, // 4-instruction MOVZ/MOVK sequence for 64-bit addresses.
32 CALL, // Function call.
34 // Produces the full sequence of instructions for getting the thread pointer
35 // offset of a variable into X0, using the TLSDesc model.
37 ADRP, // Page address of a TargetGlobalAddress operand.
38 ADDlow, // Add the low 12 bits of a TargetGlobalAddress operand.
39 LOADgot, // Load from automatically generated descriptor (e.g. Global
40 // Offset Table, TLS record).
41 RET_FLAG, // Return with a flag operand. Operand 0 is the chain operand.
42 BRCOND, // Conditional branch instruction; "b.cond".
44 FCSEL, // Conditional move instruction.
45 CSINV, // Conditional select invert.
46 CSNEG, // Conditional select negate.
47 CSINC, // Conditional select increment.
49 // Pointer to the thread's local storage area. Materialised from TPIDR_EL0 on
53 SBC, // adc, sbc instructions
55 // Arithmetic instructions which write flags.
62 // Conditional compares. Operands: left,right,falsecc,cc,flags
67 // Floating point comparison
73 // Scalar-to-vector duplication
80 // Vector immedate moves
89 // Vector immediate ops
93 // Vector bit select: similar to ISD::VSELECT but not all bits within an
94 // element must be identical.
97 // Vector arithmetic negation
112 // Vector shift by scalar
117 // Vector shift by scalar (again)
124 // Vector comparisons
134 // Vector zero comparisons
146 // Vector across-lanes addition
147 // Only the lower result lane is defined.
151 // Vector across-lanes min/max
152 // Only the lower result lane is defined.
158 // Vector bitwise negation
161 // Vector bitwise selection
164 // Compare-and-branch
173 // Custom prefetch handling
176 // {s|u}int to FP within a FP register.
180 /// Natural vector cast. ISD::BITCAST is not natural in the big-endian
181 /// world w.r.t vectors; which causes additional REV instructions to be
182 /// generated to compensate for the byte-swapping. But sometimes we do
183 /// need to re-interpret the data in SIMD vector registers in big-endian
184 /// mode without emitting such REV instructions.
190 // Reciprocal estimates and steps.
194 // NEON Load/Store with post-increment base updates
195 LD2post = ISD::FIRST_TARGET_MEMORY_OPCODE,
220 } // end namespace AArch64ISD
224 // Any instruction that defines a 32-bit result zeros out the high half of the
225 // register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
226 // be copying from a truncate. But any other 32-bit operation will zero-extend
228 // FIXME: X86 also checks for CMOV here. Do we need something similar?
229 static inline bool isDef32(const SDNode &N) {
230 unsigned Opc = N.getOpcode();
231 return Opc != ISD::TRUNCATE && Opc != TargetOpcode::EXTRACT_SUBREG &&
232 Opc != ISD::CopyFromReg;
235 } // end anonymous namespace
237 class AArch64Subtarget;
238 class AArch64TargetMachine;
240 class AArch64TargetLowering : public TargetLowering {
242 explicit AArch64TargetLowering(const TargetMachine &TM,
243 const AArch64Subtarget &STI);
245 /// Selects the correct CCAssignFn for a given CallingConvention value.
246 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool IsVarArg) const;
248 /// Selects the correct CCAssignFn for a given CallingConvention value.
249 CCAssignFn *CCAssignFnForReturn(CallingConv::ID CC) const;
251 /// Determine which of the bits specified in Mask are known to be either zero
252 /// or one and return them in the KnownZero/KnownOne bitsets.
253 void computeKnownBitsForTargetNode(const SDValue Op, APInt &KnownZero,
254 APInt &KnownOne, const APInt &DemandedElts,
255 const SelectionDAG &DAG,
256 unsigned Depth = 0) const override;
258 bool targetShrinkDemandedConstant(SDValue Op, const APInt &Demanded,
259 TargetLoweringOpt &TLO) const override;
261 MVT getScalarShiftAmountTy(const DataLayout &DL, EVT) const override;
263 /// Returns true if the target allows unaligned memory accesses of the
265 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AddrSpace = 0,
267 bool *Fast = nullptr) const override;
269 /// Provide custom lowering hooks for some operations.
270 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
272 const char *getTargetNodeName(unsigned Opcode) const override;
274 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
276 /// Returns true if a cast between SrcAS and DestAS is a noop.
277 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
278 // Addrspacecasts are always noops.
282 /// This method returns a target specific FastISel object, or null if the
283 /// target does not support "fast" ISel.
284 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
285 const TargetLibraryInfo *libInfo) const override;
287 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
289 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
291 /// Return true if the given shuffle mask can be codegen'd directly, or if it
292 /// should be stack expanded.
293 bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const override;
295 /// Return the ISD::SETCC ValueType.
296 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
297 EVT VT) const override;
299 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
301 MachineBasicBlock *EmitF128CSEL(MachineInstr &MI,
302 MachineBasicBlock *BB) const;
305 EmitInstrWithCustomInserter(MachineInstr &MI,
306 MachineBasicBlock *MBB) const override;
308 bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
309 unsigned Intrinsic) const override;
311 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
312 bool isTruncateFree(EVT VT1, EVT VT2) const override;
314 bool isProfitableToHoist(Instruction *I) const override;
316 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
317 bool isZExtFree(EVT VT1, EVT VT2) const override;
318 bool isZExtFree(SDValue Val, EVT VT2) const override;
320 bool hasPairedLoad(EVT LoadedType, unsigned &RequiredAligment) const override;
322 unsigned getMaxSupportedInterleaveFactor() const override { return 4; }
324 bool lowerInterleavedLoad(LoadInst *LI,
325 ArrayRef<ShuffleVectorInst *> Shuffles,
326 ArrayRef<unsigned> Indices,
327 unsigned Factor) const override;
328 bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI,
329 unsigned Factor) const override;
331 bool isLegalAddImmediate(int64_t) const override;
332 bool isLegalICmpImmediate(int64_t) const override;
334 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
335 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
336 MachineFunction &MF) const override;
338 /// Return true if the addressing mode represented by AM is legal for this
339 /// target, for a load/store of the specified type.
340 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
341 unsigned AS) const override;
343 /// \brief Return the cost of the scaling factor used in the addressing
344 /// mode represented by AM for this target, for a load/store
345 /// of the specified type.
346 /// If the AM is supported, the return value must be >= 0.
347 /// If the AM is not supported, it returns a negative value.
348 int getScalingFactorCost(const DataLayout &DL, const AddrMode &AM, Type *Ty,
349 unsigned AS) const override;
351 /// Return true if an FMA operation is faster than a pair of fmul and fadd
352 /// instructions. fmuladd intrinsics will be expanded to FMAs when this method
353 /// returns true, otherwise fmuladd is expanded to fmul + fadd.
354 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
356 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
358 /// \brief Returns false if N is a bit extraction pattern of (X >> C) & Mask.
359 bool isDesirableToCommuteWithShift(const SDNode *N) const override;
361 /// \brief Returns true if it is beneficial to convert a load of a constant
362 /// to just the constant itself.
363 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
364 Type *Ty) const override;
366 Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
367 AtomicOrdering Ord) const override;
368 Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val,
369 Value *Addr, AtomicOrdering Ord) const override;
371 void emitAtomicCmpXchgNoStoreLLBalance(IRBuilder<> &Builder) const override;
373 TargetLoweringBase::AtomicExpansionKind
374 shouldExpandAtomicLoadInIR(LoadInst *LI) const override;
375 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
376 TargetLoweringBase::AtomicExpansionKind
377 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
379 bool shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override;
381 bool useLoadStackGuardNode() const override;
382 TargetLoweringBase::LegalizeTypeAction
383 getPreferredVectorAction(EVT VT) const override;
385 /// If the target has a standard location for the stack protector cookie,
386 /// returns the address of that location. Otherwise, returns nullptr.
387 Value *getIRStackGuard(IRBuilder<> &IRB) const override;
389 /// If the target has a standard location for the unsafe stack pointer,
390 /// returns the address of that location. Otherwise, returns nullptr.
391 Value *getSafeStackPointerLocation(IRBuilder<> &IRB) const override;
393 /// If a physical register, this returns the register that receives the
394 /// exception address on entry to an EH pad.
396 getExceptionPointerRegister(const Constant *PersonalityFn) const override {
397 // FIXME: This is a guess. Has this been defined yet?
401 /// If a physical register, this returns the register that receives the
402 /// exception typeid on entry to a landing pad.
404 getExceptionSelectorRegister(const Constant *PersonalityFn) const override {
405 // FIXME: This is a guess. Has this been defined yet?
409 bool isIntDivCheap(EVT VT, AttributeList Attr) const override;
411 bool isCheapToSpeculateCttz() const override {
415 bool isCheapToSpeculateCtlz() const override {
419 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
421 bool hasAndNotCompare(SDValue) const override {
426 bool hasBitPreservingFPLogic(EVT VT) const override {
427 // FIXME: Is this always true? It should be true for vectors at least.
428 return VT == MVT::f32 || VT == MVT::f64;
431 bool supportSplitCSR(MachineFunction *MF) const override {
432 return MF->getFunction()->getCallingConv() == CallingConv::CXX_FAST_TLS &&
433 MF->getFunction()->hasFnAttribute(Attribute::NoUnwind);
435 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
436 void insertCopiesSplitCSR(
437 MachineBasicBlock *Entry,
438 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
440 bool supportSwiftError() const override {
444 /// Returns the size of the platform's va_list object.
445 unsigned getVaListSizeInBits(const DataLayout &DL) const override;
447 /// Returns true if \p VecTy is a legal interleaved access type. This
448 /// function checks the vector element type and the overall width of the
450 bool isLegalInterleavedAccessType(VectorType *VecTy,
451 const DataLayout &DL) const;
453 /// Returns the number of interleaved accesses that will be generated when
454 /// lowering accesses of the given type.
455 unsigned getNumInterleavedAccesses(VectorType *VecTy,
456 const DataLayout &DL) const;
459 bool isExtFreeImpl(const Instruction *Ext) const override;
461 /// Keep a pointer to the AArch64Subtarget around so that we can
462 /// make the right decision when generating code for different targets.
463 const AArch64Subtarget *Subtarget;
465 void addTypeForNEON(MVT VT, MVT PromotedBitwiseVT);
466 void addDRTypeForNEON(MVT VT);
467 void addQRTypeForNEON(MVT VT);
469 SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
471 const SmallVectorImpl<ISD::InputArg> &Ins,
472 const SDLoc &DL, SelectionDAG &DAG,
473 SmallVectorImpl<SDValue> &InVals) const override;
475 SDValue LowerCall(CallLoweringInfo & /*CLI*/,
476 SmallVectorImpl<SDValue> &InVals) const override;
478 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
479 CallingConv::ID CallConv, bool isVarArg,
480 const SmallVectorImpl<ISD::InputArg> &Ins,
481 const SDLoc &DL, SelectionDAG &DAG,
482 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
483 SDValue ThisVal) const;
485 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
487 bool isEligibleForTailCallOptimization(
488 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
489 const SmallVectorImpl<ISD::OutputArg> &Outs,
490 const SmallVectorImpl<SDValue> &OutVals,
491 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const;
493 /// Finds the incoming stack arguments which overlap the given fixed stack
494 /// object and incorporates their load into the current chain. This prevents
495 /// an upcoming store from clobbering the stack argument before it's used.
496 SDValue addTokenForArgument(SDValue Chain, SelectionDAG &DAG,
497 MachineFrameInfo &MFI, int ClobberedFI) const;
499 bool DoesCalleeRestoreStack(CallingConv::ID CallCC, bool TailCallOpt) const;
501 void saveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG, const SDLoc &DL,
502 SDValue &Chain) const;
504 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
506 const SmallVectorImpl<ISD::OutputArg> &Outs,
507 LLVMContext &Context) const override;
509 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
510 const SmallVectorImpl<ISD::OutputArg> &Outs,
511 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
512 SelectionDAG &DAG) const override;
514 SDValue getTargetNode(GlobalAddressSDNode *N, EVT Ty, SelectionDAG &DAG,
515 unsigned Flag) const;
516 SDValue getTargetNode(JumpTableSDNode *N, EVT Ty, SelectionDAG &DAG,
517 unsigned Flag) const;
518 SDValue getTargetNode(ConstantPoolSDNode *N, EVT Ty, SelectionDAG &DAG,
519 unsigned Flag) const;
520 SDValue getTargetNode(BlockAddressSDNode *N, EVT Ty, SelectionDAG &DAG,
521 unsigned Flag) const;
522 template <class NodeTy> SDValue getGOT(NodeTy *N, SelectionDAG &DAG) const;
523 template <class NodeTy>
524 SDValue getAddrLarge(NodeTy *N, SelectionDAG &DAG) const;
525 template <class NodeTy> SDValue getAddr(NodeTy *N, SelectionDAG &DAG) const;
526 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
527 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
528 SDValue LowerDarwinGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
529 SDValue LowerELFGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
530 SDValue LowerELFTLSDescCallSeq(SDValue SymAddr, const SDLoc &DL,
531 SelectionDAG &DAG) const;
532 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
533 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
534 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
535 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
536 SDValue LowerSELECT_CC(ISD::CondCode CC, SDValue LHS, SDValue RHS,
537 SDValue TVal, SDValue FVal, const SDLoc &dl,
538 SelectionDAG &DAG) const;
539 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
540 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
541 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
542 SDValue LowerAAPCS_VASTART(SDValue Op, SelectionDAG &DAG) const;
543 SDValue LowerDarwin_VASTART(SDValue Op, SelectionDAG &DAG) const;
544 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
545 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
546 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
547 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
548 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
549 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
550 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
551 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
552 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
553 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
554 SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
555 SDValue LowerVectorSRA_SRL_SHL(SDValue Op, SelectionDAG &DAG) const;
556 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
557 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
558 SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) const;
559 SDValue LowerCTPOP(SDValue Op, SelectionDAG &DAG) const;
560 SDValue LowerF128Call(SDValue Op, SelectionDAG &DAG,
561 RTLIB::Libcall Call) const;
562 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
563 SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
564 SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
565 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
566 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
567 SDValue LowerVectorAND(SDValue Op, SelectionDAG &DAG) const;
568 SDValue LowerVectorOR(SDValue Op, SelectionDAG &DAG) const;
569 SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const;
570 SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const;
572 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
573 std::vector<SDNode *> *Created) const override;
574 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
575 int &ExtraSteps, bool &UseOneConst,
576 bool Reciprocal) const override;
577 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
578 int &ExtraSteps) const override;
579 unsigned combineRepeatedFPDivisors() const override;
581 ConstraintType getConstraintType(StringRef Constraint) const override;
582 unsigned getRegisterByName(const char* RegName, EVT VT,
583 SelectionDAG &DAG) const override;
585 /// Examine constraint string and operand type and determine a weight value.
586 /// The operand object must already have been set up with the operand type.
588 getSingleConstraintMatchWeight(AsmOperandInfo &info,
589 const char *constraint) const override;
591 std::pair<unsigned, const TargetRegisterClass *>
592 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
593 StringRef Constraint, MVT VT) const override;
595 const char *LowerXConstraint(EVT ConstraintVT) const override;
597 void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
598 std::vector<SDValue> &Ops,
599 SelectionDAG &DAG) const override;
601 unsigned getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
602 if (ConstraintCode == "Q")
603 return InlineAsm::Constraint_Q;
604 // FIXME: clang has code for 'Ump', 'Utf', 'Usa', and 'Ush' but these are
605 // followed by llvm_unreachable so we'll leave them unimplemented in
606 // the backend for now.
607 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
610 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
611 bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
612 bool getIndexedAddressParts(SDNode *Op, SDValue &Base, SDValue &Offset,
613 ISD::MemIndexedMode &AM, bool &IsInc,
614 SelectionDAG &DAG) const;
615 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, SDValue &Offset,
616 ISD::MemIndexedMode &AM,
617 SelectionDAG &DAG) const override;
618 bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base,
619 SDValue &Offset, ISD::MemIndexedMode &AM,
620 SelectionDAG &DAG) const override;
622 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
623 SelectionDAG &DAG) const override;
625 bool functionArgumentNeedsConsecutiveRegisters(Type *Ty,
626 CallingConv::ID CallConv,
627 bool isVarArg) const override;
629 bool shouldNormalizeToSelectSequence(LLVMContext &, EVT) const override;
633 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
634 const TargetLibraryInfo *libInfo);
635 } // end namespace AArch64
637 } // end namespace llvm