xref: /freebsd/contrib/llvm-project/llvm/lib/Target/VE/VEISelLowering.h (revision d9a42747950146bf03cda7f6e25d219253f8a57a)
1 //===-- VEISelLowering.h - VE 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 VE uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_LIB_TARGET_VE_VEISELLOWERING_H
15 #define LLVM_LIB_TARGET_VE_VEISELLOWERING_H
16 
17 #include "VE.h"
18 #include "llvm/CodeGen/TargetLowering.h"
19 
20 namespace llvm {
21 class VESubtarget;
22 
23 namespace VEISD {
24 enum NodeType : unsigned {
25   FIRST_NUMBER = ISD::BUILTIN_OP_END,
26 
27   CALL,                   // A call instruction.
28   EH_SJLJ_LONGJMP,        // SjLj exception handling longjmp.
29   EH_SJLJ_SETJMP,         // SjLj exception handling setjmp.
30   EH_SJLJ_SETUP_DISPATCH, // SjLj exception handling setup_dispatch.
31   GETFUNPLT,              // Load function address through %plt insturction.
32   GETTLSADDR,             // Load address for TLS access.
33   GETSTACKTOP,            // Retrieve address of stack top (first address of
34                           // locals and temporaries).
35   GLOBAL_BASE_REG,        // Global base reg for PIC.
36   Hi,                     // Hi/Lo operations, typically on a global address.
37   Lo,                     // Hi/Lo operations, typically on a global address.
38   MEMBARRIER,             // Compiler barrier only; generate a no-op.
39   RET_FLAG,               // Return with a flag operand.
40   TS1AM,                  // A TS1AM instruction used for 1/2 bytes swap.
41   VEC_UNPACK_LO,          // unpack the lo v256 slice of a packed v512 vector.
42   VEC_UNPACK_HI,          // unpack the hi v256 slice of a packed v512 vector.
43                           //    0: v512 vector, 1: AVL
44   VEC_PACK,               // pack a lo and a hi vector into one v512 vector
45                           //    0: v256 lo vector, 1: v256 hi vector, 2: AVL
46 
47   VEC_BROADCAST, // A vector broadcast instruction.
48                  //   0: scalar value, 1: VL
49   REPL_I32,
50   REPL_F32, // Replicate subregister to other half.
51 
52   // Annotation as a wrapper. LEGALAVL(VL) means that VL refers to 64bit of
53   // data, whereas the raw EVL coming in from VP nodes always refers to number
54   // of elements, regardless of their size.
55   LEGALAVL,
56 
57 // VVP_* nodes.
58 #define ADD_VVP_OP(VVP_NAME, ...) VVP_NAME,
59 #include "VVPNodes.def"
60 };
61 }
62 
63 class VECustomDAG;
64 
65 class VETargetLowering : public TargetLowering {
66   const VESubtarget *Subtarget;
67 
68   void initRegisterClasses();
69   void initSPUActions();
70   void initVPUActions();
71 
72 public:
73   VETargetLowering(const TargetMachine &TM, const VESubtarget &STI);
74 
75   const char *getTargetNodeName(unsigned Opcode) const override;
76   MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
77     return MVT::i32;
78   }
79 
80   Register getRegisterByName(const char *RegName, LLT VT,
81                              const MachineFunction &MF) const override;
82 
83   /// getSetCCResultType - Return the ISD::SETCC ValueType
84   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
85                          EVT VT) const override;
86 
87   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
88                                bool isVarArg,
89                                const SmallVectorImpl<ISD::InputArg> &Ins,
90                                const SDLoc &dl, SelectionDAG &DAG,
91                                SmallVectorImpl<SDValue> &InVals) const override;
92 
93   SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
94                     SmallVectorImpl<SDValue> &InVals) const override;
95 
96   bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
97                       bool isVarArg,
98                       const SmallVectorImpl<ISD::OutputArg> &ArgsFlags,
99                       LLVMContext &Context) const override;
100   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
101                       const SmallVectorImpl<ISD::OutputArg> &Outs,
102                       const SmallVectorImpl<SDValue> &OutVals, const SDLoc &dl,
103                       SelectionDAG &DAG) const override;
104 
105   /// Helper functions for atomic operations.
106   bool shouldInsertFencesForAtomic(const Instruction *I) const override {
107     // VE uses release consistency, so need fence for each atomics.
108     return true;
109   }
110   Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst,
111                                 AtomicOrdering Ord) const override;
112   Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst,
113                                  AtomicOrdering Ord) const override;
114   TargetLoweringBase::AtomicExpansionKind
115   shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
116   ISD::NodeType getExtendForAtomicOps() const override {
117     return ISD::ANY_EXTEND;
118   }
119 
120   /// Custom Lower {
121   TargetLoweringBase::LegalizeAction
122   getCustomOperationAction(SDNode &) const override;
123 
124   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
125   unsigned getJumpTableEncoding() const override;
126   const MCExpr *LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
127                                           const MachineBasicBlock *MBB,
128                                           unsigned Uid,
129                                           MCContext &Ctx) const override;
130   SDValue getPICJumpTableRelocBase(SDValue Table,
131                                    SelectionDAG &DAG) const override;
132   // VE doesn't need getPICJumpTableRelocBaseExpr since it is used for only
133   // EK_LabelDifference32.
134 
135   SDValue lowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) const;
136   SDValue lowerATOMIC_SWAP(SDValue Op, SelectionDAG &DAG) const;
137   SDValue lowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
138   SDValue lowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
139   SDValue lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
140   SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
141   SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
142   SDValue lowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const;
143   SDValue lowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
144   SDValue lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
145   SDValue lowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
146   SDValue lowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
147   SDValue lowerLOAD(SDValue Op, SelectionDAG &DAG) const;
148   SDValue lowerSTORE(SDValue Op, SelectionDAG &DAG) const;
149   SDValue lowerToTLSGeneralDynamicModel(SDValue Op, SelectionDAG &DAG) const;
150   SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const;
151   SDValue lowerVAARG(SDValue Op, SelectionDAG &DAG) const;
152 
153   SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
154   SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
155   SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
156   /// } Custom Lower
157 
158   /// Replace the results of node with an illegal result
159   /// type with new values built out of custom code.
160   ///
161   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
162                           SelectionDAG &DAG) const override;
163 
164   /// Custom Inserter {
165   MachineBasicBlock *
166   EmitInstrWithCustomInserter(MachineInstr &MI,
167                               MachineBasicBlock *MBB) const override;
168   MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
169                                        MachineBasicBlock *MBB) const;
170   MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
171                                       MachineBasicBlock *MBB) const;
172   MachineBasicBlock *emitSjLjDispatchBlock(MachineInstr &MI,
173                                            MachineBasicBlock *BB) const;
174 
175   void setupEntryBlockForSjLj(MachineInstr &MI, MachineBasicBlock *MBB,
176                               MachineBasicBlock *DispatchBB, int FI,
177                               int Offset) const;
178   // Setup basic block address.
179   Register prepareMBB(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
180                       MachineBasicBlock *TargetBB, const DebugLoc &DL) const;
181   // Prepare function/variable address.
182   Register prepareSymbol(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
183                          StringRef Symbol, const DebugLoc &DL, bool IsLocal,
184                          bool IsCall) const;
185   /// } Custom Inserter
186 
187   /// VVP Lowering {
188   SDValue lowerToVVP(SDValue Op, SelectionDAG &DAG) const;
189   SDValue lowerVVP_LOAD_STORE(SDValue Op, VECustomDAG &) const;
190   SDValue lowerVVP_GATHER_SCATTER(SDValue Op, VECustomDAG &) const;
191 
192   SDValue legalizeInternalVectorOp(SDValue Op, SelectionDAG &DAG) const;
193   SDValue legalizeInternalLoadStoreOp(SDValue Op, VECustomDAG &CDAG) const;
194   SDValue splitVectorOp(SDValue Op, VECustomDAG &CDAG) const;
195   SDValue splitPackedLoadStore(SDValue Op, VECustomDAG &CDAG) const;
196   SDValue legalizePackedAVL(SDValue Op, VECustomDAG &CDAG) const;
197   SDValue splitMaskArithmetic(SDValue Op, SelectionDAG &DAG) const;
198   /// } VVPLowering
199 
200   /// Custom DAGCombine {
201   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
202 
203   SDValue combineTRUNCATE(SDNode *N, DAGCombinerInfo &DCI) const;
204   /// } Custom DAGCombine
205 
206   SDValue withTargetFlags(SDValue Op, unsigned TF, SelectionDAG &DAG) const;
207   SDValue makeHiLoPair(SDValue Op, unsigned HiTF, unsigned LoTF,
208                        SelectionDAG &DAG) const;
209   SDValue makeAddress(SDValue Op, SelectionDAG &DAG) const;
210 
211   bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
212   bool isFPImmLegal(const APFloat &Imm, EVT VT,
213                     bool ForCodeSize) const override;
214   /// Returns true if the target allows unaligned memory accesses of the
215   /// specified type.
216   bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, Align A,
217                                       MachineMemOperand::Flags Flags,
218                                       bool *Fast) const override;
219 
220   /// Inline Assembly {
221 
222   ConstraintType getConstraintType(StringRef Constraint) const override;
223   std::pair<unsigned, const TargetRegisterClass *>
224   getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
225                                StringRef Constraint, MVT VT) const override;
226 
227   /// } Inline Assembly
228 
229   /// Target Optimization {
230 
231   // Return lower limit for number of blocks in a jump table.
232   unsigned getMinimumJumpTableEntries() const override;
233 
234   // SX-Aurora VE's s/udiv is 5-9 times slower than multiply.
235   bool isIntDivCheap(EVT, AttributeList) const override { return false; }
236   // VE doesn't have rem.
237   bool hasStandaloneRem(EVT) const override { return false; }
238   // VE LDZ instruction returns 64 if the input is zero.
239   bool isCheapToSpeculateCtlz() const override { return true; }
240   // VE LDZ instruction is fast.
241   bool isCtlzFast() const override { return true; }
242   // VE has NND instruction.
243   bool hasAndNot(SDValue Y) const override;
244 
245   /// } Target Optimization
246 };
247 } // namespace llvm
248 
249 #endif // LLVM_LIB_TARGET_VE_VEISELLOWERING_H
250