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 LLVM_LIB_TARGET_R600_AMDGPUISELLOWERING_H
17 #define LLVM_LIB_TARGET_R600_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 LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
47 SDValue LowerFREM(SDValue Op, SelectionDAG &DAG) const;
48 SDValue LowerFCEIL(SDValue Op, SelectionDAG &DAG) const;
49 SDValue LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const;
50 SDValue LowerFRINT(SDValue Op, SelectionDAG &DAG) const;
51 SDValue LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const;
52 SDValue LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const;
54 SDValue LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG, bool Signed) const;
55 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
56 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
58 SDValue LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG, bool Signed) const;
59 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
60 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
62 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
64 SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const;
65 SDValue performMulCombine(SDNode *N, DAGCombinerInfo &DCI) const;
68 static EVT getEquivalentMemType(LLVMContext &Context, EVT VT);
69 static EVT getEquivalentLoadRegType(LLVMContext &Context, EVT VT);
71 virtual SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op,
72 SelectionDAG &DAG) const;
74 /// \brief Split a vector load into a scalar load of each component.
75 SDValue ScalarizeVectorLoad(SDValue Op, SelectionDAG &DAG) const;
77 /// \brief Split a vector load into 2 loads of half the vector.
78 SDValue SplitVectorLoad(SDValue Op, SelectionDAG &DAG) const;
80 /// \brief Split a vector store into a scalar store of each component.
81 SDValue ScalarizeVectorStore(SDValue Op, SelectionDAG &DAG) const;
83 /// \brief Split a vector store into 2 stores of half the vector.
84 SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const;
86 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
87 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
88 SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const;
89 SDValue LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const;
90 void LowerUDIVREM64(SDValue Op, SelectionDAG &DAG,
91 SmallVectorImpl<SDValue> &Results) const;
92 bool isHWTrueValue(SDValue Op) const;
93 bool isHWFalseValue(SDValue Op) const;
95 /// The SelectionDAGBuilder will automatically promote function arguments
96 /// with illegal types. However, this does not work for the AMDGPU targets
97 /// since the function arguments are stored in memory as these illegal types.
98 /// In order to handle this properly we need to get the origianl types sizes
99 /// from the LLVM IR Function and fixup the ISD:InputArg values before
100 /// passing them to AnalyzeFormalArguments()
101 void getOriginalFunctionArgs(SelectionDAG &DAG,
103 const SmallVectorImpl<ISD::InputArg> &Ins,
104 SmallVectorImpl<ISD::InputArg> &OrigIns) const;
105 void AnalyzeFormalArguments(CCState &State,
106 const SmallVectorImpl<ISD::InputArg> &Ins) const;
109 AMDGPUTargetLowering(TargetMachine &TM);
111 bool isFAbsFree(EVT VT) const override;
112 bool isFNegFree(EVT VT) const override;
113 bool isTruncateFree(EVT Src, EVT Dest) const override;
114 bool isTruncateFree(Type *Src, Type *Dest) const override;
116 bool isZExtFree(Type *Src, Type *Dest) const override;
117 bool isZExtFree(EVT Src, EVT Dest) const override;
118 bool isZExtFree(SDValue Val, EVT VT2) const override;
120 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
122 MVT getVectorIdxTy() const override;
123 bool isSelectSupported(SelectSupportKind) const override;
125 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
126 bool ShouldShrinkFPConstant(EVT VT) const override;
127 bool shouldReduceLoadWidth(SDNode *Load,
128 ISD::LoadExtType ExtType,
129 EVT ExtVT) const override;
131 bool isLoadBitCastBeneficial(EVT, EVT) const override;
132 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv,
134 const SmallVectorImpl<ISD::OutputArg> &Outs,
135 const SmallVectorImpl<SDValue> &OutVals,
136 SDLoc DL, SelectionDAG &DAG) const override;
137 SDValue LowerCall(CallLoweringInfo &CLI,
138 SmallVectorImpl<SDValue> &InVals) const override;
140 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
141 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
142 void ReplaceNodeResults(SDNode * N,
143 SmallVectorImpl<SDValue> &Results,
144 SelectionDAG &DAG) const override;
146 SDValue LowerIntrinsicIABS(SDValue Op, SelectionDAG &DAG) const;
147 SDValue LowerIntrinsicLRP(SDValue Op, SelectionDAG &DAG) const;
148 SDValue CombineFMinMaxLegacy(SDLoc DL,
155 DAGCombinerInfo &DCI) const;
156 SDValue CombineIMinMax(SDLoc DL,
163 SelectionDAG &DAG) const;
165 const char* getTargetNodeName(unsigned Opcode) const override;
167 virtual SDNode *PostISelFolding(MachineSDNode *N,
168 SelectionDAG &DAG) const {
172 /// \brief Determine which of the bits specified in \p Mask are known to be
173 /// either zero or one and return them in the \p KnownZero and \p KnownOne
175 void computeKnownBitsForTargetNode(const SDValue Op,
178 const SelectionDAG &DAG,
179 unsigned Depth = 0) const override;
181 unsigned ComputeNumSignBitsForTargetNode(SDValue Op, const SelectionDAG &DAG,
182 unsigned Depth = 0) const override;
184 /// \brief Helper function that adds Reg to the LiveIn list of the DAG's
187 /// \returns a RegisterSDNode representing Reg.
188 virtual SDValue CreateLiveInRegister(SelectionDAG &DAG,
189 const TargetRegisterClass *RC,
190 unsigned Reg, EVT VT) const;
193 namespace AMDGPUISD {
197 FIRST_NUMBER = ISD::BUILTIN_OP_END,
198 CALL, // Function call based on a single integer
199 UMUL, // 32bit unsigned multiplication
202 // End AMDIL ISD Opcodes
206 MAD, // Multiply + add with same result as the separate operations.
208 // SIN_HW, COS_HW - f32 for SI, 1 ULP max error, valid from -100 pi to 100 pi.
209 // Denormals handled on some parts.
228 TRIG_PREOP, // 1 ULP max error for f64
230 // RCP, RSQ - For f32, 1 ULP max error, no denormal handling.
231 // For f64, max error 2^29 ULP, handles denormals.
238 BFE_U32, // Extract range of bits with zero extension to 32-bits.
239 BFE_I32, // Extract range of bits with sign extension to 32-bits.
240 BFI, // (src0 & src1) | (~src0 & src2)
241 BFM, // Insert a range of bits into a 32-bit word.
242 BREV, // Reverse bits.
258 // These cvt_f32_ubyte* nodes need to remain consecutive and in order.
263 /// This node is for VLIW targets and it is used to represent a vector
264 /// that is stored in consecutive registers with the same channel.
271 BUILD_VERTICAL_VECTOR,
272 /// Pointer to the start of the shader's constant data.
274 FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE,
277 TBUFFER_STORE_FORMAT,
278 LAST_AMDGPU_ISD_NUMBER
282 } // End namespace AMDGPUISD
284 } // End namespace llvm