xref: /freebsd/contrib/llvm-project/llvm/lib/Target/VE/VEISelLowering.h (revision 2e3f49888ec8851bafb22011533217487764fdb0)
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   CMPI, // Compare between two signed integer values.
28   CMPU, // Compare between two unsigned integer values.
29   CMPF, // Compare between two floating-point values.
30   CMPQ, // Compare between two quad floating-point values.
31   CMOV, // Select between two values using the result of comparison.
32 
33   CALL,                   // A call instruction.
34   EH_SJLJ_LONGJMP,        // SjLj exception handling longjmp.
35   EH_SJLJ_SETJMP,         // SjLj exception handling setjmp.
36   EH_SJLJ_SETUP_DISPATCH, // SjLj exception handling setup_dispatch.
37   GETFUNPLT,              // Load function address through %plt insturction.
38   GETTLSADDR,             // Load address for TLS access.
39   GETSTACKTOP,            // Retrieve address of stack top (first address of
40                           // locals and temporaries).
41   GLOBAL_BASE_REG,        // Global base reg for PIC.
42   Hi,                     // Hi/Lo operations, typically on a global address.
43   Lo,                     // Hi/Lo operations, typically on a global address.
44   RET_GLUE,               // Return with a flag operand.
45   TS1AM,                  // A TS1AM instruction used for 1/2 bytes swap.
46   VEC_UNPACK_LO,          // unpack the lo v256 slice of a packed v512 vector.
47   VEC_UNPACK_HI,          // unpack the hi v256 slice of a packed v512 vector.
48                           //    0: v512 vector, 1: AVL
49   VEC_PACK,               // pack a lo and a hi vector into one v512 vector
50                           //    0: v256 lo vector, 1: v256 hi vector, 2: AVL
51 
52   VEC_BROADCAST, // A vector broadcast instruction.
53                  //   0: scalar value, 1: VL
54   REPL_I32,
55   REPL_F32, // Replicate subregister to other half.
56 
57   // Annotation as a wrapper. LEGALAVL(VL) means that VL refers to 64bit of
58   // data, whereas the raw EVL coming in from VP nodes always refers to number
59   // of elements, regardless of their size.
60   LEGALAVL,
61 
62 // VVP_* nodes.
63 #define ADD_VVP_OP(VVP_NAME, ...) VVP_NAME,
64 #include "VVPNodes.def"
65 };
66 }
67 
68 class VECustomDAG;
69 
70 class VETargetLowering : public TargetLowering {
71   const VESubtarget *Subtarget;
72 
73   void initRegisterClasses();
74   void initSPUActions();
75   void initVPUActions();
76 
77 public:
78   VETargetLowering(const TargetMachine &TM, const VESubtarget &STI);
79 
80   const char *getTargetNodeName(unsigned Opcode) const override;
81   MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
82     return MVT::i32;
83   }
84 
85   Register getRegisterByName(const char *RegName, LLT VT,
86                              const MachineFunction &MF) const override;
87 
88   /// getSetCCResultType - Return the ISD::SETCC ValueType
89   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
90                          EVT VT) const override;
91 
92   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
93                                bool isVarArg,
94                                const SmallVectorImpl<ISD::InputArg> &Ins,
95                                const SDLoc &dl, SelectionDAG &DAG,
96                                SmallVectorImpl<SDValue> &InVals) const override;
97 
98   SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
99                     SmallVectorImpl<SDValue> &InVals) const override;
100 
101   bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
102                       bool isVarArg,
103                       const SmallVectorImpl<ISD::OutputArg> &ArgsFlags,
104                       LLVMContext &Context) const override;
105   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
106                       const SmallVectorImpl<ISD::OutputArg> &Outs,
107                       const SmallVectorImpl<SDValue> &OutVals, const SDLoc &dl,
108                       SelectionDAG &DAG) const override;
109 
110   /// Helper functions for atomic operations.
111   bool shouldInsertFencesForAtomic(const Instruction *I) const override {
112     // VE uses release consistency, so need fence for each atomics.
113     return true;
114   }
115   Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst,
116                                 AtomicOrdering Ord) const override;
117   Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst,
118                                  AtomicOrdering Ord) const override;
119   TargetLoweringBase::AtomicExpansionKind
120   shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
121   ISD::NodeType getExtendForAtomicOps() const override {
122     return ISD::ANY_EXTEND;
123   }
124 
125   /// Custom Lower {
126   TargetLoweringBase::LegalizeAction
127   getCustomOperationAction(SDNode &) const override;
128 
129   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
130   unsigned getJumpTableEncoding() const override;
131   const MCExpr *LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
132                                           const MachineBasicBlock *MBB,
133                                           unsigned Uid,
134                                           MCContext &Ctx) const override;
135   SDValue getPICJumpTableRelocBase(SDValue Table,
136                                    SelectionDAG &DAG) const override;
137   // VE doesn't need getPICJumpTableRelocBaseExpr since it is used for only
138   // EK_LabelDifference32.
139 
140   SDValue lowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) const;
141   SDValue lowerATOMIC_SWAP(SDValue Op, SelectionDAG &DAG) const;
142   SDValue lowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
143   SDValue lowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
144   SDValue lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
145   SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
146   SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
147   SDValue lowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const;
148   SDValue lowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
149   SDValue lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
150   SDValue lowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
151   SDValue lowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
152   SDValue lowerLOAD(SDValue Op, SelectionDAG &DAG) const;
153   SDValue lowerSTORE(SDValue Op, SelectionDAG &DAG) const;
154   SDValue lowerToTLSGeneralDynamicModel(SDValue Op, SelectionDAG &DAG) const;
155   SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const;
156   SDValue lowerVAARG(SDValue Op, SelectionDAG &DAG) const;
157 
158   SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
159   SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
160   SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
161   /// } Custom Lower
162 
163   /// Replace the results of node with an illegal result
164   /// type with new values built out of custom code.
165   ///
166   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
167                           SelectionDAG &DAG) const override;
168 
169   /// Custom Inserter {
170   MachineBasicBlock *
171   EmitInstrWithCustomInserter(MachineInstr &MI,
172                               MachineBasicBlock *MBB) const override;
173   MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
174                                        MachineBasicBlock *MBB) const;
175   MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
176                                       MachineBasicBlock *MBB) const;
177   MachineBasicBlock *emitSjLjDispatchBlock(MachineInstr &MI,
178                                            MachineBasicBlock *BB) const;
179 
180   void setupEntryBlockForSjLj(MachineInstr &MI, MachineBasicBlock *MBB,
181                               MachineBasicBlock *DispatchBB, int FI,
182                               int Offset) const;
183   // Setup basic block address.
184   Register prepareMBB(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
185                       MachineBasicBlock *TargetBB, const DebugLoc &DL) const;
186   // Prepare function/variable address.
187   Register prepareSymbol(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
188                          StringRef Symbol, const DebugLoc &DL, bool IsLocal,
189                          bool IsCall) const;
190   /// } Custom Inserter
191 
192   /// VVP Lowering {
193   SDValue lowerToVVP(SDValue Op, SelectionDAG &DAG) const;
194   SDValue lowerVVP_LOAD_STORE(SDValue Op, VECustomDAG &) const;
195   SDValue lowerVVP_GATHER_SCATTER(SDValue Op, VECustomDAG &) const;
196 
197   SDValue legalizeInternalVectorOp(SDValue Op, SelectionDAG &DAG) const;
198   SDValue legalizeInternalLoadStoreOp(SDValue Op, VECustomDAG &CDAG) const;
199   SDValue splitVectorOp(SDValue Op, VECustomDAG &CDAG) const;
200   SDValue splitPackedLoadStore(SDValue Op, VECustomDAG &CDAG) const;
201   SDValue legalizePackedAVL(SDValue Op, VECustomDAG &CDAG) const;
202   SDValue splitMaskArithmetic(SDValue Op, SelectionDAG &DAG) const;
203   /// } VVPLowering
204 
205   /// Custom DAGCombine {
206   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
207 
208   SDValue combineSelect(SDNode *N, DAGCombinerInfo &DCI) const;
209   SDValue combineSelectCC(SDNode *N, DAGCombinerInfo &DCI) const;
210   SDValue combineTRUNCATE(SDNode *N, DAGCombinerInfo &DCI) const;
211   /// } Custom DAGCombine
212 
213   SDValue withTargetFlags(SDValue Op, unsigned TF, SelectionDAG &DAG) const;
214   SDValue makeHiLoPair(SDValue Op, unsigned HiTF, unsigned LoTF,
215                        SelectionDAG &DAG) const;
216   SDValue makeAddress(SDValue Op, SelectionDAG &DAG) const;
217 
218   bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
219   bool isFPImmLegal(const APFloat &Imm, EVT VT,
220                     bool ForCodeSize) const override;
221   /// Returns true if the target allows unaligned memory accesses of the
222   /// specified type.
223   bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, Align A,
224                                       MachineMemOperand::Flags Flags,
225                                       unsigned *Fast) const override;
226 
227   /// Inline Assembly {
228 
229   ConstraintType getConstraintType(StringRef Constraint) const override;
230   std::pair<unsigned, const TargetRegisterClass *>
231   getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
232                                StringRef Constraint, MVT VT) const override;
233 
234   /// } Inline Assembly
235 
236   /// Target Optimization {
237 
238   // Return lower limit for number of blocks in a jump table.
239   unsigned getMinimumJumpTableEntries() const override;
240 
241   // SX-Aurora VE's s/udiv is 5-9 times slower than multiply.
242   bool isIntDivCheap(EVT, AttributeList) const override { return false; }
243   // VE doesn't have rem.
244   bool hasStandaloneRem(EVT) const override { return false; }
245   // VE LDZ instruction returns 64 if the input is zero.
246   bool isCheapToSpeculateCtlz(Type *) const override { return true; }
247   // VE LDZ instruction is fast.
248   bool isCtlzFast() const override { return true; }
249   // VE has NND instruction.
250   bool hasAndNot(SDValue Y) const override;
251 
252   /// } Target Optimization
253 };
254 } // namespace llvm
255 
256 #endif // LLVM_LIB_TARGET_VE_VEISELLOWERING_H
257