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.
45 SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
46 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
49 static EVT getEquivalentMemType(LLVMContext &Context, EVT VT);
50 static EVT getEquivalentLoadRegType(LLVMContext &Context, EVT VT);
52 /// \brief Helper function that adds Reg to the LiveIn list of the DAG's
55 /// \returns a RegisterSDNode representing Reg.
56 virtual SDValue CreateLiveInRegister(SelectionDAG &DAG,
57 const TargetRegisterClass *RC,
58 unsigned Reg, EVT VT) const;
59 SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op,
60 SelectionDAG &DAG) const;
61 /// \brief Split a vector load into multiple scalar loads.
62 SDValue SplitVectorLoad(const SDValue &Op, SelectionDAG &DAG) const;
63 SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const;
64 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
65 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
66 bool isHWTrueValue(SDValue Op) const;
67 bool isHWFalseValue(SDValue Op) const;
69 /// The SelectionDAGBuilder will automatically promote function arguments
70 /// with illegal types. However, this does not work for the AMDGPU targets
71 /// since the function arguments are stored in memory as these illegal types.
72 /// In order to handle this properly we need to get the origianl types sizes
73 /// from the LLVM IR Function and fixup the ISD:InputArg values before
74 /// passing them to AnalyzeFormalArguments()
75 void getOriginalFunctionArgs(SelectionDAG &DAG,
77 const SmallVectorImpl<ISD::InputArg> &Ins,
78 SmallVectorImpl<ISD::InputArg> &OrigIns) const;
79 void AnalyzeFormalArguments(CCState &State,
80 const SmallVectorImpl<ISD::InputArg> &Ins) const;
83 AMDGPUTargetLowering(TargetMachine &TM);
85 bool isFAbsFree(EVT VT) const override;
86 bool isFNegFree(EVT VT) const override;
87 bool isTruncateFree(EVT Src, EVT Dest) const override;
88 bool isTruncateFree(Type *Src, Type *Dest) const override;
90 bool isZExtFree(Type *Src, Type *Dest) const override;
91 bool isZExtFree(EVT Src, EVT Dest) const override;
93 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
95 MVT getVectorIdxTy() const override;
96 bool isLoadBitCastBeneficial(EVT, EVT) const override;
97 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv,
99 const SmallVectorImpl<ISD::OutputArg> &Outs,
100 const SmallVectorImpl<SDValue> &OutVals,
101 SDLoc DL, SelectionDAG &DAG) const override;
102 SDValue LowerCall(CallLoweringInfo &CLI,
103 SmallVectorImpl<SDValue> &InVals) const override;
105 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
106 void ReplaceNodeResults(SDNode * N,
107 SmallVectorImpl<SDValue> &Results,
108 SelectionDAG &DAG) const override;
110 SDValue LowerIntrinsicIABS(SDValue Op, SelectionDAG &DAG) const;
111 SDValue LowerIntrinsicLRP(SDValue Op, SelectionDAG &DAG) const;
112 SDValue CombineMinMax(SDNode *N, SelectionDAG &DAG) const;
113 const char* getTargetNodeName(unsigned Opcode) const override;
115 virtual SDNode *PostISelFolding(MachineSDNode *N,
116 SelectionDAG &DAG) const {
120 /// \brief Determine which of the bits specified in \p Mask are known to be
121 /// either zero or one and return them in the \p KnownZero and \p KnownOne
123 void computeKnownBitsForTargetNode(const SDValue Op,
126 const SelectionDAG &DAG,
127 unsigned Depth = 0) const override;
129 virtual unsigned ComputeNumSignBitsForTargetNode(
131 const SelectionDAG &DAG,
132 unsigned Depth = 0) const override;
134 // Functions defined in AMDILISelLowering.cpp
136 bool getTgtMemIntrinsic(IntrinsicInfo &Info,
137 const CallInst &I, unsigned Intrinsic) const override;
139 /// We want to mark f32/f64 floating point values as legal.
140 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
142 /// We don't want to shrink f64/f32 constants.
143 bool ShouldShrinkFPConstant(EVT VT) const override;
145 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
148 void InitAMDILLowering();
149 SDValue LowerSREM(SDValue Op, SelectionDAG &DAG) const;
150 SDValue LowerSREM32(SDValue Op, SelectionDAG &DAG) const;
151 SDValue LowerSREM64(SDValue Op, SelectionDAG &DAG) const;
152 SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) const;
153 SDValue LowerSDIV24(SDValue Op, SelectionDAG &DAG) const;
154 SDValue LowerSDIV32(SDValue Op, SelectionDAG &DAG) const;
155 SDValue LowerSDIV64(SDValue Op, SelectionDAG &DAG) const;
157 SDValue ExpandSIGN_EXTEND_INREG(SDValue Op,
159 SelectionDAG &DAG) const;
160 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
161 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
164 namespace AMDGPUISD {
168 FIRST_NUMBER = ISD::BUILTIN_OP_END,
169 CALL, // Function call based on a single integer
170 UMUL, // 32bit unsigned multiplication
171 DIV_INF, // Divide with infinity returned on zero divisor
174 // End AMDIL ISD Opcodes
188 BFE_U32, // Extract range of bits with zero extension to 32-bits.
189 BFE_I32, // Extract range of bits with sign extension to 32-bits.
190 BFI, // (src0 & src1) | (~src0 & src2)
191 BFM, // Insert a range of bits into a 32-bit word.
207 // These cvt_f32_ubyte* nodes need to remain consecutive and in order.
212 FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE,
215 TBUFFER_STORE_FORMAT,
216 LAST_AMDGPU_ISD_NUMBER
220 } // End namespace AMDGPUISD
222 } // End namespace llvm
224 #endif // AMDGPUISELLOWERING_H