1 //===-- AMDGPUISelLowering.h - AMDGPU Lowering Interface --------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief Interface definition of the TargetLowering class that is common
14 //===----------------------------------------------------------------------===//
16 #ifndef AMDGPUISELLOWERING_H
17 #define AMDGPUISELLOWERING_H
19 #include "llvm/Target/TargetLowering.h"
23 class AMDGPUMachineFunction;
24 class AMDGPUSubtarget;
25 class MachineRegisterInfo;
27 class AMDGPUTargetLowering : public TargetLowering {
29 const AMDGPUSubtarget *Subtarget;
32 SDValue LowerConstantInitializer(const Constant* Init, const GlobalValue *GV,
33 const SDValue &InitPtr,
35 SelectionDAG &DAG) const;
36 SDValue LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const;
37 SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
38 SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const;
39 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
40 /// \brief Lower vector stores by merging the vector elements into an integer
41 /// of the same bitwidth.
42 SDValue MergeVectorStore(const SDValue &Op, SelectionDAG &DAG) const;
43 /// \brief Split a vector store into multiple scalar stores.
44 /// \returns The resulting chain.
46 SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) const;
47 SDValue LowerSDIV24(SDValue Op, SelectionDAG &DAG) const;
48 SDValue LowerSDIV32(SDValue Op, SelectionDAG &DAG) const;
49 SDValue LowerSDIV64(SDValue Op, SelectionDAG &DAG) const;
50 SDValue LowerSREM(SDValue Op, SelectionDAG &DAG) const;
51 SDValue LowerSREM32(SDValue Op, SelectionDAG &DAG) const;
52 SDValue LowerSREM64(SDValue Op, SelectionDAG &DAG) const;
53 SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
54 SDValue LowerFRINT(SDValue Op, SelectionDAG &DAG) const;
55 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
57 SDValue ExpandSIGN_EXTEND_INREG(SDValue Op,
59 SelectionDAG &DAG) const;
60 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
63 static EVT getEquivalentMemType(LLVMContext &Context, EVT VT);
64 static EVT getEquivalentLoadRegType(LLVMContext &Context, EVT VT);
66 /// \brief Helper function that adds Reg to the LiveIn list of the DAG's
69 /// \returns a RegisterSDNode representing Reg.
70 virtual SDValue CreateLiveInRegister(SelectionDAG &DAG,
71 const TargetRegisterClass *RC,
72 unsigned Reg, EVT VT) const;
73 SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op,
74 SelectionDAG &DAG) const;
75 /// \brief Split a vector load into multiple scalar loads.
76 SDValue SplitVectorLoad(const SDValue &Op, SelectionDAG &DAG) const;
77 SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const;
78 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
79 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
80 bool isHWTrueValue(SDValue Op) const;
81 bool isHWFalseValue(SDValue Op) const;
83 /// The SelectionDAGBuilder will automatically promote function arguments
84 /// with illegal types. However, this does not work for the AMDGPU targets
85 /// since the function arguments are stored in memory as these illegal types.
86 /// In order to handle this properly we need to get the origianl types sizes
87 /// from the LLVM IR Function and fixup the ISD:InputArg values before
88 /// passing them to AnalyzeFormalArguments()
89 void getOriginalFunctionArgs(SelectionDAG &DAG,
91 const SmallVectorImpl<ISD::InputArg> &Ins,
92 SmallVectorImpl<ISD::InputArg> &OrigIns) const;
93 void AnalyzeFormalArguments(CCState &State,
94 const SmallVectorImpl<ISD::InputArg> &Ins) const;
97 AMDGPUTargetLowering(TargetMachine &TM);
99 bool isFAbsFree(EVT VT) const override;
100 bool isFNegFree(EVT VT) const override;
101 bool isTruncateFree(EVT Src, EVT Dest) const override;
102 bool isTruncateFree(Type *Src, Type *Dest) const override;
104 bool isZExtFree(Type *Src, Type *Dest) const override;
105 bool isZExtFree(EVT Src, EVT Dest) const override;
107 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
109 MVT getVectorIdxTy() const override;
111 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
112 bool ShouldShrinkFPConstant(EVT VT) const override;
114 bool isLoadBitCastBeneficial(EVT, EVT) const override;
115 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv,
117 const SmallVectorImpl<ISD::OutputArg> &Outs,
118 const SmallVectorImpl<SDValue> &OutVals,
119 SDLoc DL, SelectionDAG &DAG) const override;
120 SDValue LowerCall(CallLoweringInfo &CLI,
121 SmallVectorImpl<SDValue> &InVals) const override;
123 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
124 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
125 void ReplaceNodeResults(SDNode * N,
126 SmallVectorImpl<SDValue> &Results,
127 SelectionDAG &DAG) const override;
129 SDValue LowerIntrinsicIABS(SDValue Op, SelectionDAG &DAG) const;
130 SDValue LowerIntrinsicLRP(SDValue Op, SelectionDAG &DAG) const;
131 SDValue CombineMinMax(SDNode *N, SelectionDAG &DAG) const;
132 const char* getTargetNodeName(unsigned Opcode) const override;
134 virtual SDNode *PostISelFolding(MachineSDNode *N,
135 SelectionDAG &DAG) const {
139 /// \brief Determine which of the bits specified in \p Mask are known to be
140 /// either zero or one and return them in the \p KnownZero and \p KnownOne
142 void computeKnownBitsForTargetNode(const SDValue Op,
145 const SelectionDAG &DAG,
146 unsigned Depth = 0) const override;
148 virtual unsigned ComputeNumSignBitsForTargetNode(
150 const SelectionDAG &DAG,
151 unsigned Depth = 0) const override;
154 // Functions defined in AMDILISelLowering.cpp
155 void InitAMDILLowering();
156 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
159 namespace AMDGPUISD {
163 FIRST_NUMBER = ISD::BUILTIN_OP_END,
164 CALL, // Function call based on a single integer
165 UMUL, // 32bit unsigned multiplication
166 DIV_INF, // Divide with infinity returned on zero divisor
169 // End AMDIL ISD Opcodes
183 BFE_U32, // Extract range of bits with zero extension to 32-bits.
184 BFE_I32, // Extract range of bits with sign extension to 32-bits.
185 BFI, // (src0 & src1) | (~src0 & src2)
186 BFM, // Insert a range of bits into a 32-bit word.
202 // These cvt_f32_ubyte* nodes need to remain consecutive and in order.
207 /// This node is for VLIW targets and it is used to represent a vector
208 /// that is stored in consecutive registers with the same channel.
215 BUILD_VERTICAL_VECTOR,
216 FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE,
219 TBUFFER_STORE_FORMAT,
220 LAST_AMDGPU_ISD_NUMBER
224 } // End namespace AMDGPUISD
226 } // End namespace llvm
228 #endif // AMDGPUISELLOWERING_H