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/SelectionDAG.h"
19 #include "llvm/CodeGen/TargetLowering.h"
20 
21 namespace llvm {
22 class RISCVSubtarget;
23 namespace RISCVISD {
24 enum NodeType : unsigned {
25   FIRST_NUMBER = ISD::BUILTIN_OP_END,
26   RET_FLAG,
27   URET_FLAG,
28   SRET_FLAG,
29   MRET_FLAG,
30   CALL,
31   /// Select with condition operator - This selects between a true value and
32   /// a false value (ops #3 and #4) based on the boolean result of comparing
33   /// the lhs and rhs (ops #0 and #1) of a conditional expression with the
34   /// condition code in op #2, a XLenVT constant from the ISD::CondCode enum.
35   /// The lhs and rhs are XLenVT integers. The true and false values can be
36   /// integer or floating point.
37   SELECT_CC,
38   BuildPairF64,
39   SplitF64,
40   TAIL,
41   // RV64I shifts, directly matching the semantics of the named RISC-V
42   // instructions.
43   SLLW,
44   SRAW,
45   SRLW,
46   // 32-bit operations from RV64M that can't be simply matched with a pattern
47   // at instruction selection time. These have undefined behavior for division
48   // by 0 or overflow (divw) like their target independent counterparts.
49   DIVW,
50   DIVUW,
51   REMUW,
52   // RV64IB rotates, directly matching the semantics of the named RISC-V
53   // instructions.
54   ROLW,
55   RORW,
56   // RV64IB/RV32IB funnel shifts, with the semantics of the named RISC-V
57   // instructions, but the same operand order as fshl/fshr intrinsics.
58   FSR,
59   FSL,
60   // RV64IB funnel shifts, with the semantics of the named RISC-V instructions,
61   // but the same operand order as fshl/fshr intrinsics.
62   FSRW,
63   FSLW,
64   // FPR<->GPR transfer operations when the FPR is smaller than XLEN, needed as
65   // XLEN is the only legal integer width.
66   //
67   // FMV_H_X matches the semantics of the FMV.H.X.
68   // FMV_X_ANYEXTH is similar to FMV.X.H but has an any-extended result.
69   // FMV_W_X_RV64 matches the semantics of the FMV.W.X.
70   // FMV_X_ANYEXTW_RV64 is similar to FMV.X.W but has an any-extended result.
71   //
72   // This is a more convenient semantic for producing dagcombines that remove
73   // unnecessary GPR->FPR->GPR moves.
74   FMV_H_X,
75   FMV_X_ANYEXTH,
76   FMV_W_X_RV64,
77   FMV_X_ANYEXTW_RV64,
78   // READ_CYCLE_WIDE - A read of the 64-bit cycle CSR on a 32-bit target
79   // (returns (Lo, Hi)). It takes a chain operand.
80   READ_CYCLE_WIDE,
81   // Generalized Reverse and Generalized Or-Combine - directly matching the
82   // semantics of the named RISC-V instructions. Lowered as custom nodes as
83   // TableGen chokes when faced with commutative permutations in deeply-nested
84   // DAGs. Each node takes an input operand and a TargetConstant immediate
85   // shift amount, and outputs a bit-manipulated version of input. All operands
86   // are of type XLenVT.
87   GREVI,
88   GREVIW,
89   GORCI,
90   GORCIW,
91   // Vector Extension
92   // VMV_V_X_VL matches the semantics of vmv.v.x but includes an extra operand
93   // for the VL value to be used for the operation.
94   VMV_V_X_VL,
95   // VFMV_V_F_VL matches the semantics of vfmv.v.f but includes an extra operand
96   // for the VL value to be used for the operation.
97   VFMV_V_F_VL,
98   // VMV_X_S matches the semantics of vmv.x.s. The result is always XLenVT sign
99   // extended from the vector element size.
100   VMV_X_S,
101   // Splats an i64 scalar to a vector type (with element type i64) where the
102   // scalar is a sign-extended i32.
103   SPLAT_VECTOR_I64,
104   // Read VLENB CSR
105   READ_VLENB,
106   // Truncates a RVV integer vector by one power-of-two.
107   TRUNCATE_VECTOR,
108   // Unit-stride fault-only-first load
109   VLEFF,
110   VLEFF_MASK,
111   // Matches the semantics of vslideup/vslidedown. The first operand is the
112   // pass-thru operand, the second is the source vector, the third is the
113   // XLenVT index (either constant or non-constant), the fourth is the mask
114   // and the fifth the VL.
115   VSLIDEUP_VL,
116   VSLIDEDOWN_VL,
117   // Matches the semantics of the vid.v instruction, with a mask and VL
118   // operand.
119   VID_VL,
120   // Matches the semantics of the vfcnvt.rod function (Convert double-width
121   // float to single-width float, rounding towards odd). Takes a double-width
122   // float vector and produces a single-width float vector.
123   VFNCVT_ROD,
124   // These nodes match the semantics of the corresponding RVV vector reduction
125   // instructions. They produce a vector result which is the reduction
126   // performed over the first vector operand plus the first element of the
127   // second vector operand. The first operand is an unconstrained vector type,
128   // and the result and second operand's types are expected to be the
129   // corresponding full-width LMUL=1 type for the first operand:
130   //   nxv8i8 = vecreduce_add nxv32i8, nxv8i8
131   //   nxv2i32 = vecreduce_add nxv8i32, nxv2i32
132   // The different in types does introduce extra vsetvli instructions but
133   // similarly it reduces the number of registers consumed per reduction.
134   VECREDUCE_ADD,
135   VECREDUCE_UMAX,
136   VECREDUCE_SMAX,
137   VECREDUCE_UMIN,
138   VECREDUCE_SMIN,
139   VECREDUCE_AND,
140   VECREDUCE_OR,
141   VECREDUCE_XOR,
142   VECREDUCE_FADD,
143   VECREDUCE_SEQ_FADD,
144 
145   // Vector binary and unary ops with a mask as a third operand, and VL as a
146   // fourth operand.
147   // FIXME: Can we replace these with ISD::VP_*?
148   ADD_VL,
149   AND_VL,
150   MUL_VL,
151   OR_VL,
152   SDIV_VL,
153   SHL_VL,
154   SREM_VL,
155   SRA_VL,
156   SRL_VL,
157   SUB_VL,
158   UDIV_VL,
159   UREM_VL,
160   XOR_VL,
161   FADD_VL,
162   FSUB_VL,
163   FMUL_VL,
164   FDIV_VL,
165   FNEG_VL,
166   FABS_VL,
167   FSQRT_VL,
168   FMA_VL,
169   SMIN_VL,
170   SMAX_VL,
171   UMIN_VL,
172   UMAX_VL,
173 
174   // Vector compare producing a mask. Fourth operand is input mask. Fifth
175   // operand is VL.
176   SETCC_VL,
177 
178   // Vector select with an additional VL operand. This operation is unmasked.
179   VSELECT_VL,
180 
181   // Mask binary operators.
182   VMAND_VL,
183   VMOR_VL,
184   VMXOR_VL,
185 
186   // Set mask vector to all zeros or ones.
187   VMCLR_VL,
188   VMSET_VL,
189 
190   // Matches the semantics of vrgather.vx with an extra operand for VL.
191   VRGATHER_VX_VL,
192 
193   // Memory opcodes start here.
194   VLE_VL = ISD::FIRST_TARGET_MEMORY_OPCODE,
195   VSE_VL,
196 
197   // WARNING: Do not add anything in the end unless you want the node to
198   // have memop! In fact, starting from FIRST_TARGET_MEMORY_OPCODE all
199   // opcodes will be thought as target memory ops!
200 };
201 } // namespace RISCVISD
202 
203 class RISCVTargetLowering : public TargetLowering {
204   const RISCVSubtarget &Subtarget;
205 
206 public:
207   explicit RISCVTargetLowering(const TargetMachine &TM,
208                                const RISCVSubtarget &STI);
209 
210   const RISCVSubtarget &getSubtarget() const { return Subtarget; }
211 
212   bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
213                           MachineFunction &MF,
214                           unsigned Intrinsic) const override;
215   bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
216                              unsigned AS,
217                              Instruction *I = nullptr) const override;
218   bool isLegalICmpImmediate(int64_t Imm) const override;
219   bool isLegalAddImmediate(int64_t Imm) const override;
220   bool isTruncateFree(Type *SrcTy, Type *DstTy) const override;
221   bool isTruncateFree(EVT SrcVT, EVT DstVT) const override;
222   bool isZExtFree(SDValue Val, EVT VT2) const override;
223   bool isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const override;
224   bool isCheapToSpeculateCttz() const override;
225   bool isCheapToSpeculateCtlz() const override;
226   bool isFPImmLegal(const APFloat &Imm, EVT VT,
227                     bool ForCodeSize) const override;
228 
229   bool hasBitPreservingFPLogic(EVT VT) const override;
230 
231   // Provide custom lowering hooks for some operations.
232   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
233   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
234                           SelectionDAG &DAG) const override;
235 
236   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
237 
238   bool targetShrinkDemandedConstant(SDValue Op, const APInt &DemandedBits,
239                                     const APInt &DemandedElts,
240                                     TargetLoweringOpt &TLO) const override;
241 
242   void computeKnownBitsForTargetNode(const SDValue Op,
243                                      KnownBits &Known,
244                                      const APInt &DemandedElts,
245                                      const SelectionDAG &DAG,
246                                      unsigned Depth) const override;
247   unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
248                                            const APInt &DemandedElts,
249                                            const SelectionDAG &DAG,
250                                            unsigned Depth) const override;
251 
252   // This method returns the name of a target specific DAG node.
253   const char *getTargetNodeName(unsigned Opcode) const override;
254 
255   ConstraintType getConstraintType(StringRef Constraint) const override;
256 
257   unsigned getInlineAsmMemConstraint(StringRef ConstraintCode) const override;
258 
259   std::pair<unsigned, const TargetRegisterClass *>
260   getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
261                                StringRef Constraint, MVT VT) const override;
262 
263   void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
264                                     std::vector<SDValue> &Ops,
265                                     SelectionDAG &DAG) const override;
266 
267   MachineBasicBlock *
268   EmitInstrWithCustomInserter(MachineInstr &MI,
269                               MachineBasicBlock *BB) const override;
270 
271   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
272                          EVT VT) const override;
273 
274   bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
275     return VT.isScalarInteger();
276   }
277   bool convertSelectOfConstantsToMath(EVT VT) const override { return true; }
278 
279   bool shouldInsertFencesForAtomic(const Instruction *I) const override {
280     return isa<LoadInst>(I) || isa<StoreInst>(I);
281   }
282   Instruction *emitLeadingFence(IRBuilder<> &Builder, Instruction *Inst,
283                                 AtomicOrdering Ord) const override;
284   Instruction *emitTrailingFence(IRBuilder<> &Builder, Instruction *Inst,
285                                  AtomicOrdering Ord) const override;
286 
287   bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
288                                   EVT VT) const override;
289 
290   ISD::NodeType getExtendForAtomicOps() const override {
291     return ISD::SIGN_EXTEND;
292   }
293 
294   ISD::NodeType getExtendForAtomicCmpSwapArg() const override {
295     return ISD::SIGN_EXTEND;
296   }
297 
298   bool shouldExpandShift(SelectionDAG &DAG, SDNode *N) const override {
299     if (DAG.getMachineFunction().getFunction().hasMinSize())
300       return false;
301     return true;
302   }
303   bool isDesirableToCommuteWithShift(const SDNode *N,
304                                      CombineLevel Level) const override;
305 
306   /// If a physical register, this returns the register that receives the
307   /// exception address on entry to an EH pad.
308   Register
309   getExceptionPointerRegister(const Constant *PersonalityFn) const override;
310 
311   /// If a physical register, this returns the register that receives the
312   /// exception typeid on entry to a landing pad.
313   Register
314   getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
315 
316   bool shouldExtendTypeInLibCall(EVT Type) const override;
317   bool shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const override;
318 
319   /// Returns the register with the specified architectural or ABI name. This
320   /// method is necessary to lower the llvm.read_register.* and
321   /// llvm.write_register.* intrinsics. Allocatable registers must be reserved
322   /// with the clang -ffixed-xX flag for access to be allowed.
323   Register getRegisterByName(const char *RegName, LLT VT,
324                              const MachineFunction &MF) const override;
325 
326   // Lower incoming arguments, copy physregs into vregs
327   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
328                                bool IsVarArg,
329                                const SmallVectorImpl<ISD::InputArg> &Ins,
330                                const SDLoc &DL, SelectionDAG &DAG,
331                                SmallVectorImpl<SDValue> &InVals) const override;
332   bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
333                       bool IsVarArg,
334                       const SmallVectorImpl<ISD::OutputArg> &Outs,
335                       LLVMContext &Context) const override;
336   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
337                       const SmallVectorImpl<ISD::OutputArg> &Outs,
338                       const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
339                       SelectionDAG &DAG) const override;
340   SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
341                     SmallVectorImpl<SDValue> &InVals) const override;
342 
343   bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
344                                          Type *Ty) const override {
345     return true;
346   }
347   bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
348   bool shouldConsiderGEPOffsetSplit() const override { return true; }
349 
350   bool decomposeMulByConstant(LLVMContext &Context, EVT VT,
351                               SDValue C) const override;
352 
353   TargetLowering::AtomicExpansionKind
354   shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
355   Value *emitMaskedAtomicRMWIntrinsic(IRBuilder<> &Builder, AtomicRMWInst *AI,
356                                       Value *AlignedAddr, Value *Incr,
357                                       Value *Mask, Value *ShiftAmt,
358                                       AtomicOrdering Ord) const override;
359   TargetLowering::AtomicExpansionKind
360   shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *CI) const override;
361   Value *emitMaskedAtomicCmpXchgIntrinsic(IRBuilder<> &Builder,
362                                           AtomicCmpXchgInst *CI,
363                                           Value *AlignedAddr, Value *CmpVal,
364                                           Value *NewVal, Value *Mask,
365                                           AtomicOrdering Ord) const override;
366 
367   /// Returns true if the target allows unaligned memory accesses of the
368   /// specified type.
369   bool allowsMisalignedMemoryAccesses(
370       EVT VT, unsigned AddrSpace = 0, Align Alignment = Align(1),
371       MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
372       bool *Fast = nullptr) const override;
373 
374 private:
375   void analyzeInputArgs(MachineFunction &MF, CCState &CCInfo,
376                         const SmallVectorImpl<ISD::InputArg> &Ins,
377                         bool IsRet) const;
378   void analyzeOutputArgs(MachineFunction &MF, CCState &CCInfo,
379                          const SmallVectorImpl<ISD::OutputArg> &Outs,
380                          bool IsRet, CallLoweringInfo *CLI) const;
381 
382   template <class NodeTy>
383   SDValue getAddr(NodeTy *N, SelectionDAG &DAG, bool IsLocal = true) const;
384 
385   SDValue getStaticTLSAddr(GlobalAddressSDNode *N, SelectionDAG &DAG,
386                            bool UseGOT) const;
387   SDValue getDynamicTLSAddr(GlobalAddressSDNode *N, SelectionDAG &DAG) const;
388 
389   SDValue lowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
390   SDValue lowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
391   SDValue lowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
392   SDValue lowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
393   SDValue lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
394   SDValue lowerSELECT(SDValue Op, SelectionDAG &DAG) const;
395   SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const;
396   SDValue lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
397   SDValue lowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
398   SDValue lowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
399   SDValue lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, bool IsSRA) const;
400   SDValue lowerSPLATVECTOR(SDValue Op, SelectionDAG &DAG) const;
401   SDValue lowerVectorMaskExt(SDValue Op, SelectionDAG &DAG,
402                              int64_t ExtTrueVal) const;
403   SDValue lowerVectorMaskTrunc(SDValue Op, SelectionDAG &DAG) const;
404   SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
405   SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
406   SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
407   SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const;
408   SDValue lowerVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
409   SDValue lowerFPVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
410   SDValue lowerFixedLengthVectorLoadToRVV(SDValue Op, SelectionDAG &DAG) const;
411   SDValue lowerFixedLengthVectorStoreToRVV(SDValue Op, SelectionDAG &DAG) const;
412   SDValue lowerFixedLengthVectorSetccToRVV(SDValue Op, SelectionDAG &DAG) const;
413   SDValue lowerFixedLengthVectorLogicOpToRVV(SDValue Op, SelectionDAG &DAG,
414                                              unsigned MaskOpc,
415                                              unsigned VecOpc) const;
416   SDValue lowerFixedLengthVectorSelectToRVV(SDValue Op,
417                                             SelectionDAG &DAG) const;
418   SDValue lowerToScalableOp(SDValue Op, SelectionDAG &DAG, unsigned NewOpc,
419                             bool HasMask = true) const;
420 
421   bool isEligibleForTailCallOptimization(
422       CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF,
423       const SmallVector<CCValAssign, 16> &ArgLocs) const;
424 
425   /// Generate error diagnostics if any register used by CC has been marked
426   /// reserved.
427   void validateCCReservedRegs(
428       const SmallVectorImpl<std::pair<llvm::Register, llvm::SDValue>> &Regs,
429       MachineFunction &MF) const;
430 
431   bool useRVVForFixedLengthVectorVT(MVT VT) const;
432 };
433 
434 namespace RISCV {
435 // We use 64 bits as the known part in the scalable vector types.
436 static constexpr unsigned RVVBitsPerBlock = 64;
437 } // namespace RISCV
438 
439 namespace RISCVVIntrinsicsTable {
440 
441 struct RISCVVIntrinsicInfo {
442   unsigned IntrinsicID;
443   uint8_t ExtendedOperand;
444 };
445 
446 using namespace RISCV;
447 
448 #define GET_RISCVVIntrinsicsTable_DECL
449 #include "RISCVGenSearchableTables.inc"
450 
451 } // end namespace RISCVVIntrinsicsTable
452 
453 } // end namespace llvm
454 
455 #endif
456