1 //===-- RISCVISelLowering.h - RISCV DAG Lowering Interface ------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interfaces that RISCV uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_LIB_TARGET_RISCV_RISCVISELLOWERING_H
15 #define LLVM_LIB_TARGET_RISCV_RISCVISELLOWERING_H
16 
17 #include "RISCV.h"
18 #include "llvm/CodeGen/CallingConvLower.h"
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/CodeGen/TargetLowering.h"
21 
22 namespace llvm {
23 class RISCVSubtarget;
24 struct RISCVRegisterInfo;
25 namespace RISCVISD {
26 enum NodeType : unsigned {
27   FIRST_NUMBER = ISD::BUILTIN_OP_END,
28   RET_FLAG,
29   URET_FLAG,
30   SRET_FLAG,
31   MRET_FLAG,
32   CALL,
33   /// Select with condition operator - This selects between a true value and
34   /// a false value (ops #3 and #4) based on the boolean result of comparing
35   /// the lhs and rhs (ops #0 and #1) of a conditional expression with the
36   /// condition code in op #2, a XLenVT constant from the ISD::CondCode enum.
37   /// The lhs and rhs are XLenVT integers. The true and false values can be
38   /// integer or floating point.
39   SELECT_CC,
40   BR_CC,
41   BuildPairF64,
42   SplitF64,
43   TAIL,
44   // Multiply high for signedxunsigned.
45   MULHSU,
46   // RV64I shifts, directly matching the semantics of the named RISC-V
47   // instructions.
48   SLLW,
49   SRAW,
50   SRLW,
51   // 32-bit operations from RV64M that can't be simply matched with a pattern
52   // at instruction selection time. These have undefined behavior for division
53   // by 0 or overflow (divw) like their target independent counterparts.
54   DIVW,
55   DIVUW,
56   REMUW,
57   // RV64IB rotates, directly matching the semantics of the named RISC-V
58   // instructions.
59   ROLW,
60   RORW,
61   // RV64IZbb bit counting instructions directly matching the semantics of the
62   // named RISC-V instructions.
63   CLZW,
64   CTZW,
65   // RV64IB/RV32IB funnel shifts, with the semantics of the named RISC-V
66   // instructions, but the same operand order as fshl/fshr intrinsics.
67   FSR,
68   FSL,
69   // RV64IB funnel shifts, with the semantics of the named RISC-V instructions,
70   // but the same operand order as fshl/fshr intrinsics.
71   FSRW,
72   FSLW,
73   // FPR<->GPR transfer operations when the FPR is smaller than XLEN, needed as
74   // XLEN is the only legal integer width.
75   //
76   // FMV_H_X matches the semantics of the FMV.H.X.
77   // FMV_X_ANYEXTH is similar to FMV.X.H but has an any-extended result.
78   // FMV_W_X_RV64 matches the semantics of the FMV.W.X.
79   // FMV_X_ANYEXTW_RV64 is similar to FMV.X.W but has an any-extended result.
80   //
81   // This is a more convenient semantic for producing dagcombines that remove
82   // unnecessary GPR->FPR->GPR moves.
83   FMV_H_X,
84   FMV_X_ANYEXTH,
85   FMV_W_X_RV64,
86   FMV_X_ANYEXTW_RV64,
87   // READ_CYCLE_WIDE - A read of the 64-bit cycle CSR on a 32-bit target
88   // (returns (Lo, Hi)). It takes a chain operand.
89   READ_CYCLE_WIDE,
90   // Generalized Reverse and Generalized Or-Combine - directly matching the
91   // semantics of the named RISC-V instructions. Lowered as custom nodes as
92   // TableGen chokes when faced with commutative permutations in deeply-nested
93   // DAGs. Each node takes an input operand and a control operand and outputs a
94   // bit-manipulated version of input. All operands are i32 or XLenVT.
95   GREV,
96   GREVW,
97   GORC,
98   GORCW,
99   SHFL,
100   SHFLW,
101   UNSHFL,
102   UNSHFLW,
103   // Bit Compress/Decompress implement the generic bit extract and bit deposit
104   // functions. This operation is also referred to as bit gather/scatter, bit
105   // pack/unpack, parallel extract/deposit, compress/expand, or right
106   // compress/right expand.
107   BCOMPRESS,
108   BCOMPRESSW,
109   BDECOMPRESS,
110   BDECOMPRESSW,
111   // Vector Extension
112   // VMV_V_X_VL matches the semantics of vmv.v.x but includes an extra operand
113   // for the VL value to be used for the operation.
114   VMV_V_X_VL,
115   // VFMV_V_F_VL matches the semantics of vfmv.v.f but includes an extra operand
116   // for the VL value to be used for the operation.
117   VFMV_V_F_VL,
118   // VMV_X_S matches the semantics of vmv.x.s. The result is always XLenVT sign
119   // extended from the vector element size.
120   VMV_X_S,
121   // VMV_S_X_VL matches the semantics of vmv.s.x. It carries a VL operand.
122   VMV_S_X_VL,
123   // VFMV_S_F_VL matches the semantics of vfmv.s.f. It carries a VL operand.
124   VFMV_S_F_VL,
125   // Splats an i64 scalar to a vector type (with element type i64) where the
126   // scalar is a sign-extended i32.
127   SPLAT_VECTOR_I64,
128   // Splats an 64-bit value that has been split into two i32 parts. This is
129   // expanded late to two scalar stores and a stride 0 vector load.
130   SPLAT_VECTOR_SPLIT_I64_VL,
131   // Read VLENB CSR
132   READ_VLENB,
133   // Truncates a RVV integer vector by one power-of-two. Carries both an extra
134   // mask and VL operand.
135   TRUNCATE_VECTOR_VL,
136   // Matches the semantics of vslideup/vslidedown. The first operand is the
137   // pass-thru operand, the second is the source vector, the third is the
138   // XLenVT index (either constant or non-constant), the fourth is the mask
139   // and the fifth the VL.
140   VSLIDEUP_VL,
141   VSLIDEDOWN_VL,
142   // Matches the semantics of vslide1up/slide1down. The first operand is the
143   // source vector, the second is the XLenVT scalar value. The third and fourth
144   // operands are the mask and VL operands.
145   VSLIDE1UP_VL,
146   VSLIDE1DOWN_VL,
147   // Matches the semantics of the vid.v instruction, with a mask and VL
148   // operand.
149   VID_VL,
150   // Matches the semantics of the vfcnvt.rod function (Convert double-width
151   // float to single-width float, rounding towards odd). Takes a double-width
152   // float vector and produces a single-width float vector. Also has a mask and
153   // VL operand.
154   VFNCVT_ROD_VL,
155   // These nodes match the semantics of the corresponding RVV vector reduction
156   // instructions. They produce a vector result which is the reduction
157   // performed over the first vector operand plus the first element of the
158   // second vector operand. The first operand is an unconstrained vector type,
159   // and the result and second operand's types are expected to be the
160   // corresponding full-width LMUL=1 type for the first operand:
161   //   nxv8i8 = vecreduce_add nxv32i8, nxv8i8
162   //   nxv2i32 = vecreduce_add nxv8i32, nxv2i32
163   // The different in types does introduce extra vsetvli instructions but
164   // similarly it reduces the number of registers consumed per reduction.
165   // Also has a mask and VL operand.
166   VECREDUCE_ADD_VL,
167   VECREDUCE_UMAX_VL,
168   VECREDUCE_SMAX_VL,
169   VECREDUCE_UMIN_VL,
170   VECREDUCE_SMIN_VL,
171   VECREDUCE_AND_VL,
172   VECREDUCE_OR_VL,
173   VECREDUCE_XOR_VL,
174   VECREDUCE_FADD_VL,
175   VECREDUCE_SEQ_FADD_VL,
176   VECREDUCE_FMIN_VL,
177   VECREDUCE_FMAX_VL,
178 
179   // Vector binary and unary ops with a mask as a third operand, and VL as a
180   // fourth operand.
181   // FIXME: Can we replace these with ISD::VP_*?
182   ADD_VL,
183   AND_VL,
184   MUL_VL,
185   OR_VL,
186   SDIV_VL,
187   SHL_VL,
188   SREM_VL,
189   SRA_VL,
190   SRL_VL,
191   SUB_VL,
192   UDIV_VL,
193   UREM_VL,
194   XOR_VL,
195   FADD_VL,
196   FSUB_VL,
197   FMUL_VL,
198   FDIV_VL,
199   FNEG_VL,
200   FABS_VL,
201   FSQRT_VL,
202   FMA_VL,
203   FCOPYSIGN_VL,
204   SMIN_VL,
205   SMAX_VL,
206   UMIN_VL,
207   UMAX_VL,
208   FMINNUM_VL,
209   FMAXNUM_VL,
210   MULHS_VL,
211   MULHU_VL,
212   FP_TO_SINT_VL,
213   FP_TO_UINT_VL,
214   SINT_TO_FP_VL,
215   UINT_TO_FP_VL,
216   FP_ROUND_VL,
217   FP_EXTEND_VL,
218 
219   // Vector compare producing a mask. Fourth operand is input mask. Fifth
220   // operand is VL.
221   SETCC_VL,
222 
223   // Vector select with an additional VL operand. This operation is unmasked.
224   VSELECT_VL,
225 
226   // Mask binary operators.
227   VMAND_VL,
228   VMOR_VL,
229   VMXOR_VL,
230 
231   // Set mask vector to all zeros or ones.
232   VMCLR_VL,
233   VMSET_VL,
234 
235   // Matches the semantics of vrgather.vx and vrgather.vv with an extra operand
236   // for VL.
237   VRGATHER_VX_VL,
238   VRGATHER_VV_VL,
239   VRGATHEREI16_VV_VL,
240 
241   // Vector sign/zero extend with additional mask & VL operands.
242   VSEXT_VL,
243   VZEXT_VL,
244   //  vpopc.m with additional mask and VL operands.
245   VPOPC_VL,
246 
247   // Reads value of CSR.
248   // The first operand is a chain pointer. The second specifies address of the
249   // required CSR. Two results are produced, the read value and the new chain
250   // pointer.
251   READ_CSR,
252   // Write value to CSR.
253   // The first operand is a chain pointer, the second specifies address of the
254   // required CSR and the third is the value to write. The result is the new
255   // chain pointer.
256   WRITE_CSR,
257   // Read and write value of CSR.
258   // The first operand is a chain pointer, the second specifies address of the
259   // required CSR and the third is the value to write. Two results are produced,
260   // the value read before the modification and the new chain pointer.
261   SWAP_CSR,
262 
263   // Memory opcodes start here.
264   VLE_VL = ISD::FIRST_TARGET_MEMORY_OPCODE,
265   VSE_VL,
266 
267   // WARNING: Do not add anything in the end unless you want the node to
268   // have memop! In fact, starting from FIRST_TARGET_MEMORY_OPCODE all
269   // opcodes will be thought as target memory ops!
270 };
271 } // namespace RISCVISD
272 
273 class RISCVTargetLowering : public TargetLowering {
274   const RISCVSubtarget &Subtarget;
275 
276 public:
277   explicit RISCVTargetLowering(const TargetMachine &TM,
278                                const RISCVSubtarget &STI);
279 
280   const RISCVSubtarget &getSubtarget() const { return Subtarget; }
281 
282   bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
283                           MachineFunction &MF,
284                           unsigned Intrinsic) const override;
285   bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
286                              unsigned AS,
287                              Instruction *I = nullptr) const override;
288   bool isLegalICmpImmediate(int64_t Imm) const override;
289   bool isLegalAddImmediate(int64_t Imm) const override;
290   bool isTruncateFree(Type *SrcTy, Type *DstTy) const override;
291   bool isTruncateFree(EVT SrcVT, EVT DstVT) const override;
292   bool isZExtFree(SDValue Val, EVT VT2) const override;
293   bool isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const override;
294   bool isCheapToSpeculateCttz() const override;
295   bool isCheapToSpeculateCtlz() const override;
296   bool isFPImmLegal(const APFloat &Imm, EVT VT,
297                     bool ForCodeSize) const override;
298 
299   bool softPromoteHalfType() const override { return true; }
300 
301   /// Return the register type for a given MVT, ensuring vectors are treated
302   /// as a series of gpr sized integers.
303   MVT getRegisterTypeForCallingConv(LLVMContext &Context, CallingConv::ID CC,
304                                     EVT VT) const override;
305 
306   /// Return the number of registers for a given MVT, ensuring vectors are
307   /// treated as a series of gpr sized integers.
308   unsigned getNumRegistersForCallingConv(LLVMContext &Context,
309                                          CallingConv::ID CC,
310                                          EVT VT) const override;
311 
312   /// Return true if the given shuffle mask can be codegen'd directly, or if it
313   /// should be stack expanded.
314   bool isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const override;
315 
316   bool hasBitPreservingFPLogic(EVT VT) const override;
317   bool
318   shouldExpandBuildVectorWithShuffles(EVT VT,
319                                       unsigned DefinedValues) const override;
320 
321   // Provide custom lowering hooks for some operations.
322   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
323   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
324                           SelectionDAG &DAG) const override;
325 
326   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
327 
328   bool targetShrinkDemandedConstant(SDValue Op, const APInt &DemandedBits,
329                                     const APInt &DemandedElts,
330                                     TargetLoweringOpt &TLO) const override;
331 
332   void computeKnownBitsForTargetNode(const SDValue Op,
333                                      KnownBits &Known,
334                                      const APInt &DemandedElts,
335                                      const SelectionDAG &DAG,
336                                      unsigned Depth) const override;
337   unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
338                                            const APInt &DemandedElts,
339                                            const SelectionDAG &DAG,
340                                            unsigned Depth) const override;
341 
342   // This method returns the name of a target specific DAG node.
343   const char *getTargetNodeName(unsigned Opcode) const override;
344 
345   ConstraintType getConstraintType(StringRef Constraint) const override;
346 
347   unsigned getInlineAsmMemConstraint(StringRef ConstraintCode) const override;
348 
349   std::pair<unsigned, const TargetRegisterClass *>
350   getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
351                                StringRef Constraint, MVT VT) const override;
352 
353   void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
354                                     std::vector<SDValue> &Ops,
355                                     SelectionDAG &DAG) const override;
356 
357   MachineBasicBlock *
358   EmitInstrWithCustomInserter(MachineInstr &MI,
359                               MachineBasicBlock *BB) const override;
360 
361   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
362                          EVT VT) const override;
363 
364   bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
365     return VT.isScalarInteger();
366   }
367   bool convertSelectOfConstantsToMath(EVT VT) const override { return true; }
368 
369   bool shouldInsertFencesForAtomic(const Instruction *I) const override {
370     return isa<LoadInst>(I) || isa<StoreInst>(I);
371   }
372   Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst,
373                                 AtomicOrdering Ord) const override;
374   Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst,
375                                  AtomicOrdering Ord) const override;
376 
377   bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
378                                   EVT VT) const override;
379 
380   ISD::NodeType getExtendForAtomicOps() const override {
381     return ISD::SIGN_EXTEND;
382   }
383 
384   ISD::NodeType getExtendForAtomicCmpSwapArg() const override {
385     return ISD::SIGN_EXTEND;
386   }
387 
388   bool shouldExpandShift(SelectionDAG &DAG, SDNode *N) const override {
389     if (DAG.getMachineFunction().getFunction().hasMinSize())
390       return false;
391     return true;
392   }
393   bool isDesirableToCommuteWithShift(const SDNode *N,
394                                      CombineLevel Level) const override;
395 
396   /// If a physical register, this returns the register that receives the
397   /// exception address on entry to an EH pad.
398   Register
399   getExceptionPointerRegister(const Constant *PersonalityFn) const override;
400 
401   /// If a physical register, this returns the register that receives the
402   /// exception typeid on entry to a landing pad.
403   Register
404   getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
405 
406   bool shouldExtendTypeInLibCall(EVT Type) const override;
407   bool shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const override;
408 
409   /// Returns the register with the specified architectural or ABI name. This
410   /// method is necessary to lower the llvm.read_register.* and
411   /// llvm.write_register.* intrinsics. Allocatable registers must be reserved
412   /// with the clang -ffixed-xX flag for access to be allowed.
413   Register getRegisterByName(const char *RegName, LLT VT,
414                              const MachineFunction &MF) const override;
415 
416   // Lower incoming arguments, copy physregs into vregs
417   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
418                                bool IsVarArg,
419                                const SmallVectorImpl<ISD::InputArg> &Ins,
420                                const SDLoc &DL, SelectionDAG &DAG,
421                                SmallVectorImpl<SDValue> &InVals) const override;
422   bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
423                       bool IsVarArg,
424                       const SmallVectorImpl<ISD::OutputArg> &Outs,
425                       LLVMContext &Context) const override;
426   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
427                       const SmallVectorImpl<ISD::OutputArg> &Outs,
428                       const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
429                       SelectionDAG &DAG) const override;
430   SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
431                     SmallVectorImpl<SDValue> &InVals) const override;
432 
433   bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
434                                          Type *Ty) const override {
435     return true;
436   }
437   bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
438   bool shouldConsiderGEPOffsetSplit() const override { return true; }
439 
440   bool decomposeMulByConstant(LLVMContext &Context, EVT VT,
441                               SDValue C) const override;
442 
443   TargetLowering::AtomicExpansionKind
444   shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
445   Value *emitMaskedAtomicRMWIntrinsic(IRBuilderBase &Builder, AtomicRMWInst *AI,
446                                       Value *AlignedAddr, Value *Incr,
447                                       Value *Mask, Value *ShiftAmt,
448                                       AtomicOrdering Ord) const override;
449   TargetLowering::AtomicExpansionKind
450   shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *CI) const override;
451   Value *emitMaskedAtomicCmpXchgIntrinsic(IRBuilderBase &Builder,
452                                           AtomicCmpXchgInst *CI,
453                                           Value *AlignedAddr, Value *CmpVal,
454                                           Value *NewVal, Value *Mask,
455                                           AtomicOrdering Ord) const override;
456 
457   /// Returns true if the target allows unaligned memory accesses of the
458   /// specified type.
459   bool allowsMisalignedMemoryAccesses(
460       EVT VT, unsigned AddrSpace = 0, Align Alignment = Align(1),
461       MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
462       bool *Fast = nullptr) const override;
463 
464   bool splitValueIntoRegisterParts(SelectionDAG &DAG, const SDLoc &DL,
465                                    SDValue Val, SDValue *Parts,
466                                    unsigned NumParts, MVT PartVT,
467                                    Optional<CallingConv::ID> CC) const override;
468 
469   SDValue
470   joinRegisterPartsIntoValue(SelectionDAG &DAG, const SDLoc &DL,
471                              const SDValue *Parts, unsigned NumParts,
472                              MVT PartVT, EVT ValueVT,
473                              Optional<CallingConv::ID> CC) const override;
474 
475   static RISCVII::VLMUL getLMUL(MVT VT);
476   static unsigned getRegClassIDForLMUL(RISCVII::VLMUL LMul);
477   static unsigned getSubregIndexByMVT(MVT VT, unsigned Index);
478   static unsigned getRegClassIDForVecVT(MVT VT);
479   static std::pair<unsigned, unsigned>
480   decomposeSubvectorInsertExtractToSubRegs(MVT VecVT, MVT SubVecVT,
481                                            unsigned InsertExtractIdx,
482                                            const RISCVRegisterInfo *TRI);
483   MVT getContainerForFixedLengthVector(MVT VT) const;
484 
485   bool shouldRemoveExtendFromGSIndex(EVT VT) const override;
486 
487 private:
488   /// RISCVCCAssignFn - This target-specific function extends the default
489   /// CCValAssign with additional information used to lower RISC-V calling
490   /// conventions.
491   typedef bool RISCVCCAssignFn(const DataLayout &DL, RISCVABI::ABI,
492                                unsigned ValNo, MVT ValVT, MVT LocVT,
493                                CCValAssign::LocInfo LocInfo,
494                                ISD::ArgFlagsTy ArgFlags, CCState &State,
495                                bool IsFixed, bool IsRet, Type *OrigTy,
496                                const RISCVTargetLowering &TLI,
497                                Optional<unsigned> FirstMaskArgument);
498 
499   void analyzeInputArgs(MachineFunction &MF, CCState &CCInfo,
500                         const SmallVectorImpl<ISD::InputArg> &Ins, bool IsRet,
501                         RISCVCCAssignFn Fn) const;
502   void analyzeOutputArgs(MachineFunction &MF, CCState &CCInfo,
503                          const SmallVectorImpl<ISD::OutputArg> &Outs,
504                          bool IsRet, CallLoweringInfo *CLI,
505                          RISCVCCAssignFn Fn) const;
506 
507   template <class NodeTy>
508   SDValue getAddr(NodeTy *N, SelectionDAG &DAG, bool IsLocal = true) const;
509 
510   SDValue getStaticTLSAddr(GlobalAddressSDNode *N, SelectionDAG &DAG,
511                            bool UseGOT) const;
512   SDValue getDynamicTLSAddr(GlobalAddressSDNode *N, SelectionDAG &DAG) const;
513 
514   SDValue lowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
515   SDValue lowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
516   SDValue lowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
517   SDValue lowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
518   SDValue lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
519   SDValue lowerSELECT(SDValue Op, SelectionDAG &DAG) const;
520   SDValue lowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
521   SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const;
522   SDValue lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
523   SDValue lowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
524   SDValue lowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
525   SDValue lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, bool IsSRA) const;
526   SDValue lowerSPLAT_VECTOR_PARTS(SDValue Op, SelectionDAG &DAG) const;
527   SDValue lowerVectorMaskSplat(SDValue Op, SelectionDAG &DAG) const;
528   SDValue lowerVectorMaskExt(SDValue Op, SelectionDAG &DAG,
529                              int64_t ExtTrueVal) const;
530   SDValue lowerVectorMaskTrunc(SDValue Op, SelectionDAG &DAG) const;
531   SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
532   SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
533   SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
534   SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const;
535   SDValue lowerVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
536   SDValue lowerVectorMaskVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
537   SDValue lowerFPVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
538   SDValue lowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
539   SDValue lowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
540   SDValue lowerSTEP_VECTOR(SDValue Op, SelectionDAG &DAG) const;
541   SDValue lowerVECTOR_REVERSE(SDValue Op, SelectionDAG &DAG) const;
542   SDValue lowerABS(SDValue Op, SelectionDAG &DAG) const;
543   SDValue lowerMLOAD(SDValue Op, SelectionDAG &DAG) const;
544   SDValue lowerMSTORE(SDValue Op, SelectionDAG &DAG) const;
545   SDValue lowerFixedLengthVectorFCOPYSIGNToRVV(SDValue Op,
546                                                SelectionDAG &DAG) const;
547   SDValue lowerMGATHER(SDValue Op, SelectionDAG &DAG) const;
548   SDValue lowerMSCATTER(SDValue Op, SelectionDAG &DAG) const;
549   SDValue lowerFixedLengthVectorLoadToRVV(SDValue Op, SelectionDAG &DAG) const;
550   SDValue lowerFixedLengthVectorStoreToRVV(SDValue Op, SelectionDAG &DAG) const;
551   SDValue lowerFixedLengthVectorSetccToRVV(SDValue Op, SelectionDAG &DAG) const;
552   SDValue lowerFixedLengthVectorLogicOpToRVV(SDValue Op, SelectionDAG &DAG,
553                                              unsigned MaskOpc,
554                                              unsigned VecOpc) const;
555   SDValue lowerFixedLengthVectorShiftToRVV(SDValue Op, SelectionDAG &DAG) const;
556   SDValue lowerFixedLengthVectorSelectToRVV(SDValue Op,
557                                             SelectionDAG &DAG) const;
558   SDValue lowerToScalableOp(SDValue Op, SelectionDAG &DAG, unsigned NewOpc,
559                             bool HasMask = true) const;
560   SDValue lowerVPOp(SDValue Op, SelectionDAG &DAG, unsigned RISCVISDOpc) const;
561   SDValue lowerFixedLengthVectorExtendToRVV(SDValue Op, SelectionDAG &DAG,
562                                             unsigned ExtendOpc) const;
563   SDValue lowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
564   SDValue lowerSET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
565 
566   SDValue expandUnalignedRVVLoad(SDValue Op, SelectionDAG &DAG) const;
567   SDValue expandUnalignedRVVStore(SDValue Op, SelectionDAG &DAG) const;
568 
569   bool isEligibleForTailCallOptimization(
570       CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF,
571       const SmallVector<CCValAssign, 16> &ArgLocs) const;
572 
573   /// Generate error diagnostics if any register used by CC has been marked
574   /// reserved.
575   void validateCCReservedRegs(
576       const SmallVectorImpl<std::pair<llvm::Register, llvm::SDValue>> &Regs,
577       MachineFunction &MF) const;
578 
579   bool useRVVForFixedLengthVectorVT(MVT VT) const;
580 
581   MVT getVPExplicitVectorLengthTy() const override;
582 
583   /// RVV code generation for fixed length vectors does not lower all
584   /// BUILD_VECTORs. This makes BUILD_VECTOR legalisation a source of stores to
585   /// merge. However, merging them creates a BUILD_VECTOR that is just as
586   /// illegal as the original, thus leading to an infinite legalisation loop.
587   /// NOTE: Once BUILD_VECTOR can be custom lowered for all legal vector types,
588   /// this override can be removed.
589   bool mergeStoresAfterLegalization(EVT VT) const override;
590 };
591 
592 namespace RISCV {
593 // We use 64 bits as the known part in the scalable vector types.
594 static constexpr unsigned RVVBitsPerBlock = 64;
595 } // namespace RISCV
596 
597 namespace RISCVVIntrinsicsTable {
598 
599 struct RISCVVIntrinsicInfo {
600   unsigned IntrinsicID;
601   uint8_t SplatOperand;
602 };
603 
604 using namespace RISCV;
605 
606 #define GET_RISCVVIntrinsicsTable_DECL
607 #include "RISCVGenSearchableTables.inc"
608 
609 } // end namespace RISCVVIntrinsicsTable
610 
611 } // end namespace llvm
612 
613 #endif
614