1//===-- RISCVISelLowering.h - RISC-V 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 RISC-V 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 "RISCVCallingConv.h"
19#include "llvm/CodeGen/CallingConvLower.h"
20#include "llvm/CodeGen/SelectionDAG.h"
21#include "llvm/CodeGen/TargetLowering.h"
22#include <optional>
23
24namespace llvm {
25class InstructionCost;
26class RISCVSubtarget;
27struct RISCVRegisterInfo;
28
29class RISCVTargetLowering : public TargetLowering {
30 const RISCVSubtarget &Subtarget;
31
32public:
33 explicit RISCVTargetLowering(const TargetMachine &TM,
34 const RISCVSubtarget &STI);
35
36 const RISCVSubtarget &getSubtarget() const { return Subtarget; }
37
38 void getTgtMemIntrinsic(SmallVectorImpl<IntrinsicInfo> &Infos,
39 const CallBase &I, MachineFunction &MF,
40 unsigned Intrinsic) const override;
41 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
42 unsigned AS,
43 Instruction *I = nullptr) const override;
44 bool isLegalICmpImmediate(int64_t Imm) const override;
45 bool isLegalAddImmediate(int64_t Imm) const override;
46 bool isTruncateFree(Type *SrcTy, Type *DstTy) const override;
47 bool isTruncateFree(EVT SrcVT, EVT DstVT) const override;
48 bool isTruncateFree(SDValue Val, EVT VT2) const override;
49 bool isZExtFree(SDValue Val, EVT VT2) const override;
50 bool isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const override;
51 bool signExtendConstant(const ConstantInt *CI) const override;
52 bool isCheapToSpeculateCttz(Type *Ty) const override;
53 bool isCheapToSpeculateCtlz(Type *Ty) const override;
54 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
55 bool hasAndNotCompare(SDValue Y) const override;
56 bool hasAndNot(SDValue Y) const override;
57 bool hasBitTest(SDValue X, SDValue Y) const override;
58 bool shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
59 SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y,
60 unsigned OldShiftOpcode, unsigned NewShiftOpcode,
61 SelectionDAG &DAG) const override;
62 bool shouldScalarizeBinop(SDValue VecOp) const override;
63 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
64 int getLegalZfaFPImm(const APFloat &Imm, EVT VT) const;
65 bool isFPImmLegal(const APFloat &Imm, EVT VT,
66 bool ForCodeSize) const override;
67 bool isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
68 unsigned Index) const override;
69
70 bool isIntDivCheap(EVT VT, AttributeList Attr) const override;
71
72 bool preferScalarizeSplat(SDNode *N) const override;
73
74 void finalizeLowering(MachineFunction &MF) const override;
75
76 /// Customize the preferred legalization strategy for certain types.
77 LegalizeTypeAction getPreferredVectorAction(MVT VT) const override;
78
79 /// Return the register type for a given MVT, ensuring vectors are treated
80 /// as a series of gpr sized integers.
81 MVT getRegisterTypeForCallingConv(LLVMContext &Context, CallingConv::ID CC,
82 EVT VT) const override;
83
84 /// Return the number of registers for a given MVT, for inline assembly
85 unsigned
86 getNumRegisters(LLVMContext &Context, EVT VT,
87 std::optional<MVT> RegisterVT = std::nullopt) const override;
88
89 /// Return the number of registers for a given MVT, ensuring vectors are
90 /// treated as a series of gpr sized integers.
91 unsigned getNumRegistersForCallingConv(LLVMContext &Context,
92 CallingConv::ID CC,
93 EVT VT) const override;
94
95 bool shouldFoldSelectWithIdentityConstant(unsigned BinOpcode, EVT VT,
96 unsigned SelectOpcode, SDValue X,
97 SDValue Y) const override;
98
99 /// Return true if the given shuffle mask can be codegen'd directly, or if it
100 /// should be stack expanded.
101 bool isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const override;
102
103 bool isMultiStoresCheaperThanBitsMerge(EVT LTy, EVT HTy) const override {
104 // If the pair to store is a mixture of float and int values, we will
105 // save two bitwise instructions and one float-to-int instruction and
106 // increase one store instruction. There is potentially a more
107 // significant benefit because it avoids the float->int domain switch
108 // for input value. So It is more likely a win.
109 if ((LTy.isFloatingPoint() && HTy.isInteger()) ||
110 (LTy.isInteger() && HTy.isFloatingPoint()))
111 return true;
112 // If the pair only contains int values, we will save two bitwise
113 // instructions and increase one store instruction (costing one more
114 // store buffer). Since the benefit is more blurred we leave such a pair
115 // out until we get testcase to prove it is a win.
116 return false;
117 }
118
119 bool
120 shouldExpandBuildVectorWithShuffles(EVT VT,
121 unsigned DefinedValues) const override;
122
123 bool shouldExpandCttzElements(EVT VT) const override;
124
125 /// Return the cost of LMUL for linear operations.
126 InstructionCost getLMULCost(MVT VT) const;
127
128 InstructionCost getVRGatherVVCost(MVT VT) const;
129 InstructionCost getVRGatherVICost(MVT VT) const;
130 InstructionCost getVSlideVXCost(MVT VT) const;
131 InstructionCost getVSlideVICost(MVT VT) const;
132
133 // Provide custom lowering hooks for some operations.
134 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
135 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
136 SelectionDAG &DAG) const override;
137
138 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
139
140 bool targetShrinkDemandedConstant(SDValue Op, const APInt &DemandedBits,
141 const APInt &DemandedElts,
142 TargetLoweringOpt &TLO) const override;
143
144 void computeKnownBitsForTargetNode(const SDValue Op,
145 KnownBits &Known,
146 const APInt &DemandedElts,
147 const SelectionDAG &DAG,
148 unsigned Depth) const override;
149
150 void computeKnownBitsForTargetInstr(GISelValueTracking &Analysis, Register R,
151 KnownBits &Known,
152 const APInt &DemandedElts,
153 const MachineRegisterInfo &MRI,
154 unsigned Depth = 0) const override;
155 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
156 const APInt &DemandedElts,
157 const SelectionDAG &DAG,
158 unsigned Depth) const override;
159
160 bool SimplifyDemandedBitsForTargetNode(SDValue Op, const APInt &DemandedBits,
161 const APInt &DemandedElts,
162 KnownBits &Known,
163 TargetLoweringOpt &TLO,
164 unsigned Depth) const override;
165
166 bool canCreateUndefOrPoisonForTargetNode(
167 SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
168 UndefPoisonKind Kind, bool ConsiderFlags, unsigned Depth) const override;
169
170 const Constant *getTargetConstantFromLoad(LoadSDNode *LD) const override;
171
172 MachineMemOperand::Flags
173 getTargetMMOFlags(const Instruction &I) const override;
174
175 MachineMemOperand::Flags
176 getTargetMMOFlags(const MemSDNode &Node) const override;
177
178 bool
179 areTwoSDNodeTargetMMOFlagsMergeable(const MemSDNode &NodeX,
180 const MemSDNode &NodeY) const override;
181
182 ConstraintType getConstraintType(StringRef Constraint) const override;
183
184 InlineAsm::ConstraintCode
185 getInlineAsmMemConstraint(StringRef ConstraintCode) const override;
186
187 std::pair<unsigned, const TargetRegisterClass *>
188 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
189 StringRef Constraint, MVT VT) const override;
190
191 void LowerAsmOperandForConstraint(SDValue Op, StringRef Constraint,
192 std::vector<SDValue> &Ops,
193 SelectionDAG &DAG) const override;
194
195 MachineBasicBlock *
196 EmitInstrWithCustomInserter(MachineInstr &MI,
197 MachineBasicBlock *BB) const override;
198
199 void AdjustInstrPostInstrSelection(MachineInstr &MI,
200 SDNode *Node) const override;
201
202 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
203 EVT VT) const override;
204
205 CondMergingParams
206 getJumpConditionMergingParams(Instruction::BinaryOps Opc, const Value *LHS,
207 const Value *RHS,
208 const Function *F) const override;
209
210 bool shouldFormOverflowOp(unsigned Opcode, EVT VT,
211 bool MathUsed) const override {
212 if (VT == MVT::i8 || VT == MVT::i16)
213 return false;
214
215 return TargetLowering::shouldFormOverflowOp(Opcode, VT, MathUsed);
216 }
217
218 bool storeOfVectorConstantIsCheap(bool IsZero, EVT MemVT, unsigned NumElem,
219 unsigned AddrSpace) const override {
220 // If we can replace 4 or more scalar stores, there will be a reduction
221 // in instructions even after we add a vector constant load.
222 return NumElem >= 4;
223 }
224
225 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
226 return VT.isScalarInteger();
227 }
228 bool convertSelectOfConstantsToMath(EVT VT) const override { return true; }
229
230 bool isCtpopFast(EVT VT) const override;
231
232 unsigned getCustomCtpopCost(EVT VT, ISD::CondCode Cond) const override;
233
234 bool preferZeroCompareBranch() const override { return true; }
235
236 // Note that one specific case requires fence insertion for an
237 // AtomicCmpXchgInst but is handled via the RISCVZacasABIFix pass rather
238 // than this hook due to limitations in the interface here.
239 bool shouldInsertFencesForAtomic(const Instruction *I) const override;
240
241 Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst,
242 AtomicOrdering Ord) const override;
243 Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst,
244 AtomicOrdering Ord) const override;
245
246 bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
247 EVT VT) const override;
248
249 ISD::NodeType getExtendForAtomicOps() const override {
250 return ISD::SIGN_EXTEND;
251 }
252
253 ISD::NodeType getExtendForAtomicCmpSwapArg() const override;
254 ISD::NodeType getExtendForAtomicRMWArg(unsigned Op) const override;
255
256 bool shouldTransformSignedTruncationCheck(EVT XVT,
257 unsigned KeptBits) const override;
258
259 TargetLowering::ShiftLegalizationStrategy
260 preferredShiftLegalizationStrategy(SelectionDAG &DAG, SDNode *N,
261 unsigned ExpansionFactor) const override {
262 if (DAG.getMachineFunction().getFunction().hasMinSize())
263 return ShiftLegalizationStrategy::LowerToLibcall;
264 return TargetLowering::preferredShiftLegalizationStrategy(DAG, N,
265 ExpansionFactor);
266 }
267
268 bool isDesirableToCommuteWithShift(const SDNode *N,
269 CombineLevel Level) const override;
270
271 /// If a physical register, this returns the register that receives the
272 /// exception address on entry to an EH pad.
273 Register
274 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
275
276 /// If a physical register, this returns the register that receives the
277 /// exception typeid on entry to a landing pad.
278 Register
279 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
280
281 bool shouldExtendTypeInLibCall(EVT Type) const override;
282 bool shouldSignExtendTypeInLibCall(Type *Ty, bool IsSigned) const override;
283
284 /// Returns the register with the specified architectural or ABI name. This
285 /// method is necessary to lower the llvm.read_register.* and
286 /// llvm.write_register.* intrinsics. Allocatable registers must be reserved
287 /// with the clang -ffixed-xX flag for access to be allowed.
288 Register getRegisterByName(const char *RegName, LLT VT,
289 const MachineFunction &MF) const override;
290
291 // Lower incoming arguments, copy physregs into vregs
292 SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
293 bool IsVarArg,
294 const SmallVectorImpl<ISD::InputArg> &Ins,
295 const SDLoc &DL, SelectionDAG &DAG,
296 SmallVectorImpl<SDValue> &InVals) const override;
297 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
298 bool IsVarArg,
299 const SmallVectorImpl<ISD::OutputArg> &Outs,
300 LLVMContext &Context, const Type *RetTy) const override;
301 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
302 const SmallVectorImpl<ISD::OutputArg> &Outs,
303 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
304 SelectionDAG &DAG) const override;
305 SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
306 SmallVectorImpl<SDValue> &InVals) const override;
307
308 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
309 Type *Ty) const override;
310 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
311 bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
312 bool shouldConsiderGEPOffsetSplit() const override { return true; }
313
314 bool decomposeMulByConstant(LLVMContext &Context, EVT VT,
315 SDValue C) const override;
316
317 bool isMulAddWithConstProfitable(SDValue AddNode,
318 SDValue ConstNode) const override;
319
320 TargetLowering::AtomicExpansionKind
321 shouldExpandAtomicRMWInIR(const AtomicRMWInst *AI) const override;
322 Value *emitMaskedAtomicRMWIntrinsic(IRBuilderBase &Builder, AtomicRMWInst *AI,
323 Value *AlignedAddr, Value *Incr,
324 Value *Mask, Value *ShiftAmt,
325 AtomicOrdering Ord) const override;
326 TargetLowering::AtomicExpansionKind
327 shouldExpandAtomicCmpXchgInIR(const AtomicCmpXchgInst *CI) const override;
328 Value *emitMaskedAtomicCmpXchgIntrinsic(IRBuilderBase &Builder,
329 AtomicCmpXchgInst *CI,
330 Value *AlignedAddr, Value *CmpVal,
331 Value *NewVal, Value *Mask,
332 AtomicOrdering Ord) const override;
333
334 /// Returns true if the target allows unaligned memory accesses of the
335 /// specified type.
336 bool allowsMisalignedMemoryAccesses(
337 EVT VT, unsigned AddrSpace = 0, Align Alignment = Align(1),
338 MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
339 unsigned *Fast = nullptr) const override;
340
341 EVT getOptimalMemOpType(LLVMContext &Context, const MemOp &Op,
342 const AttributeList &FuncAttributes) const override;
343
344 bool splitValueIntoRegisterParts(
345 SelectionDAG & DAG, const SDLoc &DL, SDValue Val, SDValue *Parts,
346 unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC)
347 const override;
348
349 SDValue joinRegisterPartsIntoValue(
350 SelectionDAG & DAG, const SDLoc &DL, const SDValue *Parts,
351 unsigned NumParts, MVT PartVT, EVT ValueVT,
352 std::optional<CallingConv::ID> CC) const override;
353
354 // Return the value of VLMax for the given vector type (i.e. SEW and LMUL)
355 SDValue computeVLMax(MVT VecVT, const SDLoc &DL, SelectionDAG &DAG) const;
356
357 static RISCVVType::VLMUL getLMUL(MVT VT);
358 inline static unsigned computeVLMAX(unsigned VectorBits, unsigned EltSize,
359 unsigned MinSize) {
360 // Original equation:
361 // VLMAX = (VectorBits / EltSize) * LMUL
362 // where LMUL = MinSize / RISCV::RVVBitsPerBlock
363 // The following equations have been reordered to prevent loss of precision
364 // when calculating fractional LMUL.
365 return ((VectorBits / EltSize) * MinSize) / RISCV::RVVBitsPerBlock;
366 }
367
368 // Return inclusive (low, high) bounds on the value of VLMAX for the
369 // given scalable container type given known bounds on VLEN.
370 static std::pair<unsigned, unsigned>
371 computeVLMAXBounds(MVT ContainerVT, const RISCVSubtarget &Subtarget);
372
373 /// Given a vector (either fixed or scalable), return the scalable vector
374 /// corresponding to a vector register (i.e. an m1 register group).
375 static MVT getM1VT(MVT VT) {
376 unsigned EltSizeInBits = VT.getVectorElementType().getSizeInBits();
377 assert(EltSizeInBits <= RISCV::RVVBitsPerBlock && "Unexpected vector MVT");
378 return MVT::getScalableVectorVT(VT: VT.getVectorElementType(),
379 NumElements: RISCV::RVVBitsPerBlock / EltSizeInBits);
380 }
381
382 static unsigned getRegClassIDForLMUL(RISCVVType::VLMUL LMul);
383 static unsigned getSubregIndexByMVT(MVT VT, unsigned Index);
384 static unsigned getRegClassIDForVecVT(MVT VT);
385 static std::pair<unsigned, unsigned>
386 decomposeSubvectorInsertExtractToSubRegs(MVT VecVT, MVT SubVecVT,
387 unsigned InsertExtractIdx,
388 const RISCVRegisterInfo *TRI);
389 MVT getContainerForFixedLengthVector(MVT VT) const;
390
391 bool shouldRemoveExtendFromGSIndex(SDValue Extend, EVT DataVT) const override;
392
393 bool isLegalElementTypeForRVV(EVT ScalarTy) const;
394
395 bool shouldConvertFpToSat(unsigned Op, EVT FPVT, EVT VT) const override;
396
397 unsigned getJumpTableEncoding() const override;
398
399 const MCExpr *LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
400 const MachineBasicBlock *MBB,
401 unsigned uid,
402 MCContext &Ctx) const override;
403
404 bool getIndexedAddressParts(SDNode *Op, SDValue &Base, SDValue &Offset,
405 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const;
406 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, SDValue &Offset,
407 ISD::MemIndexedMode &AM,
408 SelectionDAG &DAG) const override;
409 bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base,
410 SDValue &Offset, ISD::MemIndexedMode &AM,
411 SelectionDAG &DAG) const override;
412
413 bool isLegalScaleForGatherScatter(uint64_t Scale,
414 uint64_t ElemSize) const override {
415 // Scaled addressing not supported on indexed load/stores
416 return Scale == 1;
417 }
418
419 /// If the target has a standard location for the stack protector cookie,
420 /// returns the address of that location. Otherwise, returns nullptr.
421 Value *getIRStackGuard(IRBuilderBase &IRB,
422 const LibcallLoweringInfo &Libcalls) const override;
423
424 /// Returns whether or not generating a interleaved load/store intrinsic for
425 /// this type will be legal.
426 bool isLegalInterleavedAccessType(VectorType *VTy, unsigned Factor,
427 Align Alignment, unsigned AddrSpace,
428 const DataLayout &) const;
429
430 /// Return true if a stride load store of the given result type and
431 /// alignment is legal.
432 bool isLegalStridedLoadStore(EVT DataType, Align Alignment) const;
433
434 /// Return true if a fault-only-first load of the given result type and
435 /// alignment is legal.
436 bool isLegalFirstFaultLoad(EVT DataType, Align Alignment) const;
437
438 unsigned getMaxSupportedInterleaveFactor() const override { return 8; }
439
440 bool fallBackToDAGISel(const Instruction &Inst) const override;
441
442 bool lowerInterleavedLoad(Instruction *Load, Value *Mask,
443 ArrayRef<ShuffleVectorInst *> Shuffles,
444 ArrayRef<unsigned> Indices, unsigned Factor,
445 const APInt &GapMask) const override;
446
447 bool lowerInterleavedStore(Instruction *Store, Value *Mask,
448 ShuffleVectorInst *SVI, unsigned Factor,
449 const APInt &GapMask) const override;
450
451 bool lowerDeinterleaveIntrinsicToLoad(Instruction *Load, Value *Mask,
452 IntrinsicInst *DI,
453 const APInt &GapMask) const override;
454
455 bool lowerInterleaveIntrinsicToStore(
456 Instruction *Store, Value *Mask,
457 ArrayRef<Value *> InterleaveValues) const override;
458
459 bool supportKCFIBundles() const override { return true; }
460
461 SDValue expandIndirectJTBranch(const SDLoc &dl, SDValue Value, SDValue Addr,
462 int JTI, SelectionDAG &DAG) const override;
463
464 MachineInstr *EmitKCFICheck(MachineBasicBlock &MBB,
465 MachineBasicBlock::instr_iterator &MBBI,
466 const TargetInstrInfo *TII) const override;
467
468 /// True if stack clash protection is enabled for this functions.
469 bool hasInlineStackProbe(const MachineFunction &MF) const override;
470
471 unsigned getStackProbeSize(const MachineFunction &MF, Align StackAlign) const;
472
473 MachineBasicBlock *emitDynamicProbedAlloc(MachineInstr &MI,
474 MachineBasicBlock *MBB) const;
475
476 ArrayRef<MCPhysReg> getRoundingControlRegisters() const override;
477
478 bool shouldFoldMaskToVariableShiftPair(SDValue Y) const override;
479
480 /// Control the following reassociation of operands: (op (op x, c1), y) -> (op
481 /// (op x, y), c1) where N0 is (op x, c1) and N1 is y.
482 bool isReassocProfitable(SelectionDAG &DAG, SDValue N0,
483 SDValue N1) const override;
484
485 /// Match a mask which "spreads" the leading elements of a vector evenly
486 /// across the result. Factor is the spread amount, and Index is the
487 /// offset applied.
488 static bool isSpreadMask(ArrayRef<int> Mask, unsigned Factor,
489 unsigned &Index);
490
491private:
492 template <class NodeTy>
493 SDValue getAddr(NodeTy *N, SelectionDAG &DAG, bool IsLocal = true,
494 bool IsExternWeak = false) const;
495 SDValue getStaticTLSAddr(GlobalAddressSDNode *N, SelectionDAG &DAG,
496 bool UseGOT) const;
497 SDValue getDynamicTLSAddr(GlobalAddressSDNode *N, SelectionDAG &DAG) const;
498 SDValue getTLSDescAddr(GlobalAddressSDNode *N, SelectionDAG &DAG) const;
499
500 SDValue lowerConstantFP(SDValue Op, SelectionDAG &DAG) const;
501 SDValue lowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
502 SDValue lowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
503 SDValue lowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
504 SDValue lowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
505 SDValue lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
506 SDValue lowerSELECT(SDValue Op, SelectionDAG &DAG) const;
507 SDValue lowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
508 SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const;
509 SDValue lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
510 SDValue lowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
511 SDValue lowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
512 SDValue lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, bool IsSRA) const;
513 SDValue lowerSPLAT_VECTOR_PARTS(SDValue Op, SelectionDAG &DAG) const;
514 SDValue lowerVectorMaskSplat(SDValue Op, SelectionDAG &DAG) const;
515 SDValue lowerVectorMaskExt(SDValue Op, SelectionDAG &DAG,
516 int64_t ExtTrueVal) const;
517 SDValue lowerVectorMaskTrunc(SDValue Op, SelectionDAG &DAG) const;
518 SDValue lowerVectorTrunc(SDValue Op, SelectionDAG &DAG) const;
519 SDValue lowerVectorFPExtendOrRound(SDValue Op, SelectionDAG &DAG) const;
520 SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
521 SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
522 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
523 SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const;
524 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
525 SDValue lowerVPREDUCE(SDValue Op, SelectionDAG &DAG) const;
526 SDValue lowerVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
527 SDValue lowerVectorMaskVecReduction(SDValue Op, SelectionDAG &DAG,
528 bool IsVP) const;
529 SDValue lowerFPVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
530 SDValue lowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
531 SDValue lowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
532 SDValue lowerVECTOR_DEINTERLEAVE(SDValue Op, SelectionDAG &DAG) const;
533 SDValue lowerVECTOR_INTERLEAVE(SDValue Op, SelectionDAG &DAG) const;
534 SDValue lowerSTEP_VECTOR(SDValue Op, SelectionDAG &DAG) const;
535 SDValue lowerVECTOR_REVERSE(SDValue Op, SelectionDAG &DAG) const;
536 SDValue lowerVECTOR_SPLICE(SDValue Op, SelectionDAG &DAG) const;
537 SDValue lowerABS(SDValue Op, SelectionDAG &DAG) const;
538 SDValue lowerMaskedLoad(SDValue Op, SelectionDAG &DAG) const;
539 SDValue lowerLoadFF(SDValue Op, SelectionDAG &DAG) const;
540 SDValue lowerMaskedStore(SDValue Op, SelectionDAG &DAG) const;
541 SDValue lowerVectorCompress(SDValue Op, SelectionDAG &DAG) const;
542 SDValue lowerFixedLengthVectorFCOPYSIGNToRVV(SDValue Op,
543 SelectionDAG &DAG) const;
544 SDValue lowerMaskedGather(SDValue Op, SelectionDAG &DAG) const;
545 SDValue lowerMaskedScatter(SDValue Op, SelectionDAG &DAG) const;
546 SDValue lowerFixedLengthVectorLoadToRVV(SDValue Op, SelectionDAG &DAG) const;
547 SDValue lowerFixedLengthVectorStoreToRVV(SDValue Op, SelectionDAG &DAG) const;
548 SDValue lowerToScalableOp(SDValue Op, SelectionDAG &DAG) const;
549 SDValue LowerIS_FPCLASS(SDValue Op, SelectionDAG &DAG) const;
550 SDValue lowerVPOp(SDValue Op, SelectionDAG &DAG) const;
551 SDValue lowerVPMergeMask(SDValue Op, SelectionDAG &DAG) const;
552 SDValue lowerVPSpliceExperimental(SDValue Op, SelectionDAG &DAG) const;
553 SDValue lowerVPReverseExperimental(SDValue Op, SelectionDAG &DAG) const;
554 SDValue lowerVPStridedLoad(SDValue Op, SelectionDAG &DAG) const;
555 SDValue lowerVPStridedStore(SDValue Op, SelectionDAG &DAG) const;
556 SDValue lowerVPCttzElements(SDValue Op, SelectionDAG &DAG) const;
557 SDValue lowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
558 SDValue lowerSET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
559 SDValue lowerGET_FPENV(SDValue Op, SelectionDAG &DAG) const;
560 SDValue lowerSET_FPENV(SDValue Op, SelectionDAG &DAG) const;
561 SDValue lowerRESET_FPENV(SDValue Op, SelectionDAG &DAG) const;
562 SDValue lowerGET_FPMODE(SDValue Op, SelectionDAG &DAG) const;
563 SDValue lowerSET_FPMODE(SDValue Op, SelectionDAG &DAG) const;
564 SDValue lowerRESET_FPMODE(SDValue Op, SelectionDAG &DAG) const;
565
566 SDValue lowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
567 SDValue lowerCTLZ_CTTZ_ZERO_POISON(SDValue Op, SelectionDAG &DAG) const;
568
569 SDValue lowerStrictFPExtendOrRoundLike(SDValue Op, SelectionDAG &DAG) const;
570
571 SDValue lowerVectorStrictFSetcc(SDValue Op, SelectionDAG &DAG) const;
572
573 SDValue lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
574
575 SDValue expandUnalignedRVVLoad(SDValue Op, SelectionDAG &DAG) const;
576 SDValue expandUnalignedRVVStore(SDValue Op, SelectionDAG &DAG) const;
577
578 SDValue expandUnalignedVPLoad(SDValue Op, SelectionDAG &DAG) const;
579 SDValue expandUnalignedVPStore(SDValue Op, SelectionDAG &DAG) const;
580
581 SDValue lowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
582 SDValue lowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
583 SDValue lowerPARTIAL_REDUCE_MLA(SDValue Op, SelectionDAG &DAG) const;
584 SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
585
586 SDValue lowerXAndesBfHCvtBFloat16Load(SDValue Op, SelectionDAG &DAG) const;
587 SDValue lowerXAndesBfHCvtBFloat16Store(SDValue Op, SelectionDAG &DAG) const;
588
589 bool isEligibleForTailCallOptimization(
590 CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF,
591 const SmallVector<CCValAssign, 16> &ArgLocs) const;
592
593 /// Generate error diagnostics if any register used by CC has been marked
594 /// reserved.
595 void validateCCReservedRegs(
596 const SmallVectorImpl<std::pair<llvm::Register, llvm::SDValue>> &Regs,
597 MachineFunction &MF) const;
598
599 bool useRVVForFixedLengthVectorVT(MVT VT) const;
600
601 MVT getVPExplicitVectorLengthTy() const override;
602
603 bool shouldExpandGetVectorLength(EVT TripCountVT, unsigned VF,
604 bool IsScalable) const override;
605
606 /// RVV code generation for fixed length vectors does not lower all
607 /// BUILD_VECTORs. This makes BUILD_VECTOR legalisation a source of stores to
608 /// merge. However, merging them creates a BUILD_VECTOR that is just as
609 /// illegal as the original, thus leading to an infinite legalisation loop.
610 /// NOTE: Once BUILD_VECTOR can be custom lowered for all legal vector types,
611 /// this override can be removed.
612 bool mergeStoresAfterLegalization(EVT VT) const override;
613
614 /// Disable normalizing
615 /// select(N0&N1, X, Y) => select(N0, select(N1, X, Y), Y) and
616 /// select(N0|N1, X, Y) => select(N0, select(N1, X, Y, Y))
617 /// RISC-V doesn't have flags so it's better to perform the and/or in a GPR.
618 bool shouldNormalizeToSelectSequence(LLVMContext &, EVT, EVT) const override {
619 return false;
620 }
621
622 /// Disables storing and loading vectors by default when there are function
623 /// calls between the load and store, since these are more expensive than just
624 /// using scalars
625 bool shouldMergeStoreOfLoadsOverCall(EVT SrcVT, EVT MergedVT) const override {
626 return !MergedVT.isVector() || SrcVT.isVector();
627 }
628
629 /// For available scheduling models FDIV + two independent FMULs are much
630 /// faster than two FDIVs.
631 unsigned combineRepeatedFPDivisors() const override;
632
633 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
634 SmallVectorImpl<SDNode *> &Created) const override;
635
636 bool shouldFoldSelectWithSingleBitTest(EVT VT,
637 const APInt &AndMask) const override;
638
639 unsigned getMinimumJumpTableEntries() const override;
640
641 SDValue emitFlushICache(SelectionDAG &DAG, SDValue InChain, SDValue Start,
642 SDValue End, SDValue Flags, SDLoc DL) const;
643
644 std::pair<const TargetRegisterClass *, uint8_t>
645 findRepresentativeClass(const TargetRegisterInfo *TRI, MVT VT) const override;
646};
647
648namespace RISCVVIntrinsicsTable {
649
650struct RISCVVIntrinsicInfo {
651 unsigned IntrinsicID;
652 uint8_t ScalarOperand;
653 uint8_t VLOperand;
654 bool IsFPIntrinsic;
655 bool hasScalarOperand() const {
656 // 0xF is not valid. See NoScalarOperand in IntrinsicsRISCV.td.
657 return ScalarOperand != 0xF;
658 }
659 bool hasVLOperand() const {
660 // 0x1F is not valid. See NoVLOperand in IntrinsicsRISCV.td.
661 return VLOperand != 0x1F;
662 }
663};
664
665using namespace RISCV;
666
667#define GET_RISCVVIntrinsicsTable_DECL
668#include "RISCVGenSearchableTables.inc"
669#undef GET_RISCVVIntrinsicsTable_DECL
670
671} // end namespace RISCVVIntrinsicsTable
672
673} // end namespace llvm
674
675#endif
676