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 MVT getScalarShiftAmountTy(const DataLayout &DL, EVT) const override;
260 /// Returns true if the target allows unaligned memory accesses of the
262 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AddrSpace = 0,
264 bool *Fast = nullptr) const override;
266 /// Provide custom lowering hooks for some operations.
267 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
269 const char *getTargetNodeName(unsigned Opcode) const override;
271 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
273 /// Returns true if a cast between SrcAS and DestAS is a noop.
274 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
275 // Addrspacecasts are always noops.
279 /// This method returns a target specific FastISel object, or null if the
280 /// target does not support "fast" ISel.
281 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
282 const TargetLibraryInfo *libInfo) const override;
284 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
286 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
288 /// Return true if the given shuffle mask can be codegen'd directly, or if it
289 /// should be stack expanded.
290 bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const override;
292 /// Return the ISD::SETCC ValueType.
293 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
294 EVT VT) const override;
296 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
298 MachineBasicBlock *EmitF128CSEL(MachineInstr &MI,
299 MachineBasicBlock *BB) const;
302 EmitInstrWithCustomInserter(MachineInstr &MI,
303 MachineBasicBlock *MBB) const override;
305 bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
306 unsigned Intrinsic) const override;
308 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
309 bool isTruncateFree(EVT VT1, EVT VT2) const override;
311 bool isProfitableToHoist(Instruction *I) const override;
313 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
314 bool isZExtFree(EVT VT1, EVT VT2) const override;
315 bool isZExtFree(SDValue Val, EVT VT2) const override;
317 bool hasPairedLoad(EVT LoadedType, unsigned &RequiredAligment) const override;
319 unsigned getMaxSupportedInterleaveFactor() const override { return 4; }
321 bool lowerInterleavedLoad(LoadInst *LI,
322 ArrayRef<ShuffleVectorInst *> Shuffles,
323 ArrayRef<unsigned> Indices,
324 unsigned Factor) const override;
325 bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI,
326 unsigned Factor) const override;
328 bool isLegalAddImmediate(int64_t) const override;
329 bool isLegalICmpImmediate(int64_t) const override;
331 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
332 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
333 MachineFunction &MF) const override;
335 /// Return true if the addressing mode represented by AM is legal for this
336 /// target, for a load/store of the specified type.
337 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
338 unsigned AS) const override;
340 /// \brief Return the cost of the scaling factor used in the addressing
341 /// mode represented by AM for this target, for a load/store
342 /// of the specified type.
343 /// If the AM is supported, the return value must be >= 0.
344 /// If the AM is not supported, it returns a negative value.
345 int getScalingFactorCost(const DataLayout &DL, const AddrMode &AM, Type *Ty,
346 unsigned AS) const override;
348 /// Return true if an FMA operation is faster than a pair of fmul and fadd
349 /// instructions. fmuladd intrinsics will be expanded to FMAs when this method
350 /// returns true, otherwise fmuladd is expanded to fmul + fadd.
351 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
353 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
355 /// \brief Returns false if N is a bit extraction pattern of (X >> C) & Mask.
356 bool isDesirableToCommuteWithShift(const SDNode *N) const override;
358 /// \brief Returns true if it is beneficial to convert a load of a constant
359 /// to just the constant itself.
360 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
361 Type *Ty) const override;
363 Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
364 AtomicOrdering Ord) const override;
365 Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val,
366 Value *Addr, AtomicOrdering Ord) const override;
368 void emitAtomicCmpXchgNoStoreLLBalance(IRBuilder<> &Builder) const override;
370 TargetLoweringBase::AtomicExpansionKind
371 shouldExpandAtomicLoadInIR(LoadInst *LI) const override;
372 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
373 TargetLoweringBase::AtomicExpansionKind
374 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
376 bool shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override;
378 bool useLoadStackGuardNode() const override;
379 TargetLoweringBase::LegalizeTypeAction
380 getPreferredVectorAction(EVT VT) const override;
382 /// If the target has a standard location for the stack protector cookie,
383 /// returns the address of that location. Otherwise, returns nullptr.
384 Value *getIRStackGuard(IRBuilder<> &IRB) const override;
386 /// If the target has a standard location for the unsafe stack pointer,
387 /// returns the address of that location. Otherwise, returns nullptr.
388 Value *getSafeStackPointerLocation(IRBuilder<> &IRB) const override;
390 /// If a physical register, this returns the register that receives the
391 /// exception address on entry to an EH pad.
393 getExceptionPointerRegister(const Constant *PersonalityFn) const override {
394 // FIXME: This is a guess. Has this been defined yet?
398 /// If a physical register, this returns the register that receives the
399 /// exception typeid on entry to a landing pad.
401 getExceptionSelectorRegister(const Constant *PersonalityFn) const override {
402 // FIXME: This is a guess. Has this been defined yet?
406 bool isIntDivCheap(EVT VT, AttributeList Attr) const override;
408 bool isCheapToSpeculateCttz() const override {
412 bool isCheapToSpeculateCtlz() const override {
416 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
418 bool hasAndNotCompare(SDValue) const override {
423 bool hasBitPreservingFPLogic(EVT VT) const override {
424 // FIXME: Is this always true? It should be true for vectors at least.
425 return VT == MVT::f32 || VT == MVT::f64;
428 bool supportSplitCSR(MachineFunction *MF) const override {
429 return MF->getFunction()->getCallingConv() == CallingConv::CXX_FAST_TLS &&
430 MF->getFunction()->hasFnAttribute(Attribute::NoUnwind);
432 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
433 void insertCopiesSplitCSR(
434 MachineBasicBlock *Entry,
435 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
437 bool supportSwiftError() const override {
441 /// Returns the size of the platform's va_list object.
442 unsigned getVaListSizeInBits(const DataLayout &DL) const override;
444 /// Returns true if \p VecTy is a legal interleaved access type. This
445 /// function checks the vector element type and the overall width of the
447 bool isLegalInterleavedAccessType(VectorType *VecTy,
448 const DataLayout &DL) const;
450 /// Returns the number of interleaved accesses that will be generated when
451 /// lowering accesses of the given type.
452 unsigned getNumInterleavedAccesses(VectorType *VecTy,
453 const DataLayout &DL) const;
456 bool isExtFreeImpl(const Instruction *Ext) const override;
458 /// Keep a pointer to the AArch64Subtarget around so that we can
459 /// make the right decision when generating code for different targets.
460 const AArch64Subtarget *Subtarget;
462 void addTypeForNEON(MVT VT, MVT PromotedBitwiseVT);
463 void addDRTypeForNEON(MVT VT);
464 void addQRTypeForNEON(MVT VT);
466 SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
468 const SmallVectorImpl<ISD::InputArg> &Ins,
469 const SDLoc &DL, SelectionDAG &DAG,
470 SmallVectorImpl<SDValue> &InVals) const override;
472 SDValue LowerCall(CallLoweringInfo & /*CLI*/,
473 SmallVectorImpl<SDValue> &InVals) const override;
475 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
476 CallingConv::ID CallConv, bool isVarArg,
477 const SmallVectorImpl<ISD::InputArg> &Ins,
478 const SDLoc &DL, SelectionDAG &DAG,
479 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
480 SDValue ThisVal) const;
482 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
484 bool isEligibleForTailCallOptimization(
485 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
486 const SmallVectorImpl<ISD::OutputArg> &Outs,
487 const SmallVectorImpl<SDValue> &OutVals,
488 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const;
490 /// Finds the incoming stack arguments which overlap the given fixed stack
491 /// object and incorporates their load into the current chain. This prevents
492 /// an upcoming store from clobbering the stack argument before it's used.
493 SDValue addTokenForArgument(SDValue Chain, SelectionDAG &DAG,
494 MachineFrameInfo &MFI, int ClobberedFI) const;
496 bool DoesCalleeRestoreStack(CallingConv::ID CallCC, bool TailCallOpt) const;
498 void saveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG, const SDLoc &DL,
499 SDValue &Chain) const;
501 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
503 const SmallVectorImpl<ISD::OutputArg> &Outs,
504 LLVMContext &Context) const override;
506 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
507 const SmallVectorImpl<ISD::OutputArg> &Outs,
508 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
509 SelectionDAG &DAG) const override;
511 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
512 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
513 SDValue LowerDarwinGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
514 SDValue LowerELFGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
515 SDValue LowerELFTLSDescCallSeq(SDValue SymAddr, const SDLoc &DL,
516 SelectionDAG &DAG) const;
517 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
518 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
519 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
520 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
521 SDValue LowerSELECT_CC(ISD::CondCode CC, SDValue LHS, SDValue RHS,
522 SDValue TVal, SDValue FVal, const SDLoc &dl,
523 SelectionDAG &DAG) const;
524 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
525 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
526 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
527 SDValue LowerAAPCS_VASTART(SDValue Op, SelectionDAG &DAG) const;
528 SDValue LowerDarwin_VASTART(SDValue Op, SelectionDAG &DAG) const;
529 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
530 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
531 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
532 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
533 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
534 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
535 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
536 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
537 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
538 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
539 SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
540 SDValue LowerVectorSRA_SRL_SHL(SDValue Op, SelectionDAG &DAG) const;
541 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
542 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
543 SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) const;
544 SDValue LowerCTPOP(SDValue Op, SelectionDAG &DAG) const;
545 SDValue LowerF128Call(SDValue Op, SelectionDAG &DAG,
546 RTLIB::Libcall Call) const;
547 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
548 SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
549 SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
550 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
551 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
552 SDValue LowerVectorAND(SDValue Op, SelectionDAG &DAG) const;
553 SDValue LowerVectorOR(SDValue Op, SelectionDAG &DAG) const;
554 SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const;
555 SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const;
557 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
558 std::vector<SDNode *> *Created) const override;
559 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
560 int &ExtraSteps, bool &UseOneConst,
561 bool Reciprocal) const override;
562 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
563 int &ExtraSteps) const override;
564 unsigned combineRepeatedFPDivisors() const override;
566 ConstraintType getConstraintType(StringRef Constraint) const override;
567 unsigned getRegisterByName(const char* RegName, EVT VT,
568 SelectionDAG &DAG) const override;
570 /// Examine constraint string and operand type and determine a weight value.
571 /// The operand object must already have been set up with the operand type.
573 getSingleConstraintMatchWeight(AsmOperandInfo &info,
574 const char *constraint) const override;
576 std::pair<unsigned, const TargetRegisterClass *>
577 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
578 StringRef Constraint, MVT VT) const override;
580 const char *LowerXConstraint(EVT ConstraintVT) const override;
582 void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
583 std::vector<SDValue> &Ops,
584 SelectionDAG &DAG) const override;
586 unsigned getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
587 if (ConstraintCode == "Q")
588 return InlineAsm::Constraint_Q;
589 // FIXME: clang has code for 'Ump', 'Utf', 'Usa', and 'Ush' but these are
590 // followed by llvm_unreachable so we'll leave them unimplemented in
591 // the backend for now.
592 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
595 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
596 bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
597 bool getIndexedAddressParts(SDNode *Op, SDValue &Base, SDValue &Offset,
598 ISD::MemIndexedMode &AM, bool &IsInc,
599 SelectionDAG &DAG) const;
600 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, SDValue &Offset,
601 ISD::MemIndexedMode &AM,
602 SelectionDAG &DAG) const override;
603 bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base,
604 SDValue &Offset, ISD::MemIndexedMode &AM,
605 SelectionDAG &DAG) const override;
607 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
608 SelectionDAG &DAG) const override;
610 bool functionArgumentNeedsConsecutiveRegisters(Type *Ty,
611 CallingConv::ID CallConv,
612 bool isVarArg) const override;
614 bool shouldNormalizeToSelectSequence(LLVMContext &, EVT) const override;
618 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
619 const TargetLibraryInfo *libInfo);
620 } // end namespace AArch64
622 } // end namespace llvm