1 //==-- ARM64ISelLowering.h - ARM64 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 ARM64 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_TARGET_ARM64_ISELLOWERING_H
16 #define LLVM_TARGET_ARM64_ISELLOWERING_H
18 #include "llvm/CodeGen/CallingConvLower.h"
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/IR/CallingConv.h"
21 #include "llvm/Target/TargetLowering.h"
28 FIRST_NUMBER = ISD::BUILTIN_OP_END,
29 WrapperLarge, // 4-instruction MOVZ/MOVK sequence for 64-bit addresses.
30 CALL, // Function call.
32 // Almost the same as a normal call node, except that a TLSDesc relocation is
33 // needed so the linker can relax it correctly if possible.
35 ADRP, // Page address of a TargetGlobalAddress operand.
36 ADDlow, // Add the low 12 bits of a TargetGlobalAddress operand.
37 LOADgot, // Load from automatically generated descriptor (e.g. Global
38 // Offset Table, TLS record).
39 RET_FLAG, // Return with a flag operand. Operand 0 is the chain operand.
40 BRCOND, // Conditional branch instruction; "b.cond".
42 FCSEL, // Conditional move instruction.
43 CSINV, // Conditional select invert.
44 CSNEG, // Conditional select negate.
45 CSINC, // Conditional select increment.
47 // Pointer to the thread's local storage area. Materialised from TPIDR_EL0 on
51 SBC, // adc, sbc instructions
53 // Arithmetic instructions which write flags.
60 // Floating point comparison
63 // Floating point max and min instructions.
70 // Scalar-to-vector duplication
77 // Vector immedate moves
86 // Vector immediate ops
90 // Vector bit select: similar to ISD::VSELECT but not all bits within an
91 // element must be identical.
94 // Vector arithmetic negation
109 // Vector shift by scalar
114 // Vector shift by scalar (again)
121 // Vector comparisons
131 // Vector zero comparisons
143 // Vector bitwise negation
146 // Vector bitwise selection
149 // Compare-and-branch
158 // Custom prefetch handling
161 // {s|u}int to FP within a FP register.
165 // NEON Load/Store with post-increment base updates
166 LD2post = ISD::FIRST_TARGET_MEMORY_OPCODE,
189 } // end namespace ARM64ISD
191 class ARM64Subtarget;
192 class ARM64TargetMachine;
194 class ARM64TargetLowering : public TargetLowering {
195 bool RequireStrictAlign;
198 explicit ARM64TargetLowering(ARM64TargetMachine &TM);
200 /// Selects the correct CCAssignFn for a the given CallingConvention
202 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool IsVarArg) const;
204 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
205 /// Mask are known to be either zero or one and return them in the
206 /// KnownZero/KnownOne bitsets.
207 void computeMaskedBitsForTargetNode(const SDValue Op, APInt &KnownZero,
208 APInt &KnownOne, const SelectionDAG &DAG,
209 unsigned Depth = 0) const override;
211 MVT getScalarShiftAmountTy(EVT LHSTy) const override;
213 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
214 /// unaligned memory accesses. of the specified type.
215 bool allowsUnalignedMemoryAccesses(EVT VT, unsigned AddrSpace = 0,
216 bool *Fast = nullptr) const override {
217 if (RequireStrictAlign)
219 // FIXME: True for Cyclone, but not necessary others.
225 /// LowerOperation - Provide custom lowering hooks for some operations.
226 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
228 const char *getTargetNodeName(unsigned Opcode) const override;
230 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
232 /// getFunctionAlignment - Return the Log2 alignment of this function.
233 unsigned getFunctionAlignment(const Function *F) const;
235 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
236 /// be used for loads / stores from the global.
237 unsigned getMaximalGlobalOffset() const override;
239 /// Returns true if a cast between SrcAS and DestAS is a noop.
240 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
241 // Addrspacecasts are always noops.
245 /// createFastISel - This method returns a target specific FastISel object,
246 /// or null if the target does not support "fast" ISel.
247 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
248 const TargetLibraryInfo *libInfo) const override;
250 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
252 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
254 /// isShuffleMaskLegal - Return true if the given shuffle mask can be
255 /// codegen'd directly, or if it should be stack expanded.
256 bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const override;
258 /// getSetCCResultType - Return the ISD::SETCC ValueType
259 EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
261 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
263 MachineBasicBlock *EmitF128CSEL(MachineInstr *MI,
264 MachineBasicBlock *BB) const;
267 EmitInstrWithCustomInserter(MachineInstr *MI,
268 MachineBasicBlock *MBB) const override;
270 bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
271 unsigned Intrinsic) const override;
273 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
274 bool isTruncateFree(EVT VT1, EVT VT2) const override;
276 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
277 bool isZExtFree(EVT VT1, EVT VT2) const override;
278 bool isZExtFree(SDValue Val, EVT VT2) const override;
280 bool hasPairedLoad(Type *LoadedType,
281 unsigned &RequiredAligment) const override;
282 bool hasPairedLoad(EVT LoadedType, unsigned &RequiredAligment) const override;
284 bool isLegalAddImmediate(int64_t) const override;
285 bool isLegalICmpImmediate(int64_t) const override;
287 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
288 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
289 MachineFunction &MF) const override;
291 /// isLegalAddressingMode - Return true if the addressing mode represented
292 /// by AM is legal for this target, for a load/store of the specified type.
293 bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const override;
295 /// \brief Return the cost of the scaling factor used in the addressing
296 /// mode represented by AM for this target, for a load/store
297 /// of the specified type.
298 /// If the AM is supported, the return value must be >= 0.
299 /// If the AM is not supported, it returns a negative value.
300 int getScalingFactorCost(const AddrMode &AM, Type *Ty) const override;
302 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
303 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
304 /// expanded to FMAs when this method returns true, otherwise fmuladd is
305 /// expanded to fmul + fadd.
306 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
308 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
310 /// \brief Returns false if N is a bit extraction pattern of (X >> C) & Mask.
311 bool isDesirableToCommuteWithShift(const SDNode *N) const override;
313 /// \brief Returns true if it is beneficial to convert a load of a constant
314 /// to just the constant itself.
315 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
316 Type *Ty) const override;
318 Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
319 AtomicOrdering Ord) const override;
320 Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val,
321 Value *Addr, AtomicOrdering Ord) const override;
323 bool shouldExpandAtomicInIR(Instruction *Inst) const override;
326 /// Subtarget - Keep a pointer to the ARM64Subtarget around so that we can
327 /// make the right decision when generating code for different targets.
328 const ARM64Subtarget *Subtarget;
330 void addTypeForNEON(EVT VT, EVT PromotedBitwiseVT);
331 void addDRTypeForNEON(MVT VT);
332 void addQRTypeForNEON(MVT VT);
335 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
336 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL,
338 SmallVectorImpl<SDValue> &InVals) const override;
340 SDValue LowerCall(CallLoweringInfo & /*CLI*/,
341 SmallVectorImpl<SDValue> &InVals) const override;
343 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
344 CallingConv::ID CallConv, bool isVarArg,
345 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL,
346 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
347 bool isThisReturn, SDValue ThisVal) const;
349 bool isEligibleForTailCallOptimization(
350 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
351 bool isCalleeStructRet, bool isCallerStructRet,
352 const SmallVectorImpl<ISD::OutputArg> &Outs,
353 const SmallVectorImpl<SDValue> &OutVals,
354 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const;
356 void saveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG, SDLoc DL,
357 SDValue &Chain) const;
359 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
361 const SmallVectorImpl<ISD::OutputArg> &Outs,
362 LLVMContext &Context) const override;
364 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
365 const SmallVectorImpl<ISD::OutputArg> &Outs,
366 const SmallVectorImpl<SDValue> &OutVals, SDLoc DL,
367 SelectionDAG &DAG) const override;
369 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
370 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
371 SDValue LowerDarwinGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
372 SDValue LowerELFGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
373 SDValue LowerELFTLSDescCall(SDValue SymAddr, SDValue DescAddr, SDLoc DL,
374 SelectionDAG &DAG) const;
375 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
376 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
377 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
378 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
379 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
380 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
381 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
382 SDValue LowerAAPCS_VASTART(SDValue Op, SelectionDAG &DAG) const;
383 SDValue LowerDarwin_VASTART(SDValue Op, SelectionDAG &DAG) const;
384 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
385 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
386 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
387 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
388 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
389 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
390 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
391 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
392 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
393 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
394 SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
395 SDValue LowerVectorSRA_SRL_SHL(SDValue Op, SelectionDAG &DAG) const;
396 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
397 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
398 SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) const;
399 SDValue LowerCTPOP(SDValue Op, SelectionDAG &DAG) const;
400 SDValue LowerF128Call(SDValue Op, SelectionDAG &DAG,
401 RTLIB::Libcall Call) const;
402 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
403 SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
404 SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
405 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
406 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
407 SDValue LowerVectorAND(SDValue Op, SelectionDAG &DAG) const;
408 SDValue LowerVectorOR(SDValue Op, SelectionDAG &DAG) const;
409 SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const;
410 SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const;
413 getConstraintType(const std::string &Constraint) const override;
414 unsigned getRegisterByName(const char* RegName) const;
416 /// Examine constraint string and operand type and determine a weight value.
417 /// The operand object must already have been set up with the operand type.
419 getSingleConstraintMatchWeight(AsmOperandInfo &info,
420 const char *constraint) const override;
422 std::pair<unsigned, const TargetRegisterClass *>
423 getRegForInlineAsmConstraint(const std::string &Constraint,
424 MVT VT) const override;
425 void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
426 std::vector<SDValue> &Ops,
427 SelectionDAG &DAG) const override;
429 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
430 bool mayBeEmittedAsTailCall(CallInst *CI) const override;
431 bool getIndexedAddressParts(SDNode *Op, SDValue &Base, SDValue &Offset,
432 ISD::MemIndexedMode &AM, bool &IsInc,
433 SelectionDAG &DAG) const;
434 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, SDValue &Offset,
435 ISD::MemIndexedMode &AM,
436 SelectionDAG &DAG) const override;
437 bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base,
438 SDValue &Offset, ISD::MemIndexedMode &AM,
439 SelectionDAG &DAG) const override;
441 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
442 SelectionDAG &DAG) const override;
446 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
447 const TargetLibraryInfo *libInfo);
448 } // end namespace ARM64
450 } // end namespace llvm
452 #endif // LLVM_TARGET_ARM64_ISELLOWERING_H