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 LowerFREM(SDValue Op, SelectionDAG &DAG) const;
47 SDValue LowerFCEIL(SDValue Op, SelectionDAG &DAG) const;
48 SDValue LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const;
49 SDValue LowerFRINT(SDValue Op, SelectionDAG &DAG) const;
50 SDValue LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const;
52 SDValue LowerFROUND32(SDValue Op, SelectionDAG &DAG) const;
53 SDValue LowerFROUND64(SDValue Op, SelectionDAG &DAG) const;
54 SDValue LowerFROUND(SDValue Op, SelectionDAG &DAG) const;
55 SDValue LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const;
57 SDValue LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG, bool Signed) const;
58 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
59 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
61 SDValue LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG, bool Signed) const;
62 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
63 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
65 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
67 SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const;
68 SDValue performShlCombine(SDNode *N, DAGCombinerInfo &DCI) const;
69 SDValue performMulCombine(SDNode *N, DAGCombinerInfo &DCI) const;
72 static EVT getEquivalentMemType(LLVMContext &Context, EVT VT);
73 static EVT getEquivalentLoadRegType(LLVMContext &Context, EVT VT);
75 virtual SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op,
76 SelectionDAG &DAG) const;
78 /// \brief Split a vector load into a scalar load of each component.
79 SDValue ScalarizeVectorLoad(SDValue Op, SelectionDAG &DAG) const;
81 /// \brief Split a vector load into 2 loads of half the vector.
82 SDValue SplitVectorLoad(SDValue Op, SelectionDAG &DAG) const;
84 /// \brief Split a vector store into a scalar store of each component.
85 SDValue ScalarizeVectorStore(SDValue Op, SelectionDAG &DAG) const;
87 /// \brief Split a vector store into 2 stores of half the vector.
88 SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const;
90 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
91 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
92 SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const;
93 SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
94 SDValue LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const;
95 void LowerUDIVREM64(SDValue Op, SelectionDAG &DAG,
96 SmallVectorImpl<SDValue> &Results) const;
97 bool isHWTrueValue(SDValue Op) const;
98 bool isHWFalseValue(SDValue Op) const;
100 /// The SelectionDAGBuilder will automatically promote function arguments
101 /// with illegal types. However, this does not work for the AMDGPU targets
102 /// since the function arguments are stored in memory as these illegal types.
103 /// In order to handle this properly we need to get the origianl types sizes
104 /// from the LLVM IR Function and fixup the ISD:InputArg values before
105 /// passing them to AnalyzeFormalArguments()
106 void getOriginalFunctionArgs(SelectionDAG &DAG,
108 const SmallVectorImpl<ISD::InputArg> &Ins,
109 SmallVectorImpl<ISD::InputArg> &OrigIns) const;
110 void AnalyzeFormalArguments(CCState &State,
111 const SmallVectorImpl<ISD::InputArg> &Ins) const;
114 AMDGPUTargetLowering(TargetMachine &TM, const AMDGPUSubtarget &STI);
116 bool isFAbsFree(EVT VT) const override;
117 bool isFNegFree(EVT VT) const override;
118 bool isTruncateFree(EVT Src, EVT Dest) const override;
119 bool isTruncateFree(Type *Src, Type *Dest) const override;
121 bool isZExtFree(Type *Src, Type *Dest) const override;
122 bool isZExtFree(EVT Src, EVT Dest) const override;
123 bool isZExtFree(SDValue Val, EVT VT2) const override;
125 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
127 MVT getVectorIdxTy(const DataLayout &) const override;
128 bool isSelectSupported(SelectSupportKind) const override;
130 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
131 bool ShouldShrinkFPConstant(EVT VT) const override;
132 bool shouldReduceLoadWidth(SDNode *Load,
133 ISD::LoadExtType ExtType,
134 EVT ExtVT) const override;
136 bool isLoadBitCastBeneficial(EVT, EVT) const override;
138 bool storeOfVectorConstantIsCheap(EVT MemVT,
140 unsigned AS) const override;
141 bool aggressivelyPreferBuildVectorSources(EVT VecVT) const override;
142 bool isCheapToSpeculateCttz() const override;
143 bool isCheapToSpeculateCtlz() const override;
145 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv,
147 const SmallVectorImpl<ISD::OutputArg> &Outs,
148 const SmallVectorImpl<SDValue> &OutVals,
149 SDLoc DL, SelectionDAG &DAG) const override;
150 SDValue LowerCall(CallLoweringInfo &CLI,
151 SmallVectorImpl<SDValue> &InVals) const override;
153 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op,
154 SelectionDAG &DAG) const;
156 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
157 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
158 void ReplaceNodeResults(SDNode * N,
159 SmallVectorImpl<SDValue> &Results,
160 SelectionDAG &DAG) const override;
162 SDValue LowerIntrinsicIABS(SDValue Op, SelectionDAG &DAG) const;
163 SDValue LowerIntrinsicLRP(SDValue Op, SelectionDAG &DAG) const;
164 SDValue CombineFMinMaxLegacy(SDLoc DL,
171 DAGCombinerInfo &DCI) const;
173 const char* getTargetNodeName(unsigned Opcode) const override;
175 SDValue getRsqrtEstimate(SDValue Operand,
176 DAGCombinerInfo &DCI,
177 unsigned &RefinementSteps,
178 bool &UseOneConstNR) const override;
179 SDValue getRecipEstimate(SDValue Operand,
180 DAGCombinerInfo &DCI,
181 unsigned &RefinementSteps) const override;
183 virtual SDNode *PostISelFolding(MachineSDNode *N,
184 SelectionDAG &DAG) const {
188 /// \brief Determine which of the bits specified in \p Mask are known to be
189 /// either zero or one and return them in the \p KnownZero and \p KnownOne
191 void computeKnownBitsForTargetNode(const SDValue Op,
194 const SelectionDAG &DAG,
195 unsigned Depth = 0) const override;
197 unsigned ComputeNumSignBitsForTargetNode(SDValue Op, const SelectionDAG &DAG,
198 unsigned Depth = 0) const override;
200 /// \brief Helper function that adds Reg to the LiveIn list of the DAG's
203 /// \returns a RegisterSDNode representing Reg.
204 virtual SDValue CreateLiveInRegister(SelectionDAG &DAG,
205 const TargetRegisterClass *RC,
206 unsigned Reg, EVT VT) const;
208 enum ImplicitParameter {
213 /// \brief Helper function that returns the byte offset of the given
214 /// type of implicit parameter.
215 uint32_t getImplicitParameterOffset(const AMDGPUMachineFunction *MFI,
216 const ImplicitParameter Param) const;
219 namespace AMDGPUISD {
221 enum NodeType : unsigned {
223 FIRST_NUMBER = ISD::BUILTIN_OP_END,
224 CALL, // Function call based on a single integer
225 UMUL, // 32bit unsigned multiplication
228 // End AMDIL ISD Opcodes
233 // SIN_HW, COS_HW - f32 for SI, 1 ULP max error, valid from -100 pi to 100 pi.
234 // Denormals handled on some parts.
249 TRIG_PREOP, // 1 ULP max error for f64
251 // RCP, RSQ - For f32, 1 ULP max error, no denormal handling.
252 // For f64, max error 2^29 ULP, handles denormals.
262 BFE_U32, // Extract range of bits with zero extension to 32-bits.
263 BFE_I32, // Extract range of bits with sign extension to 32-bits.
264 BFI, // (src0 & src1) | (~src0 & src2)
265 BFM, // Insert a range of bits into a 32-bit word.
281 // These cvt_f32_ubyte* nodes need to remain consecutive and in order.
286 /// This node is for VLIW targets and it is used to represent a vector
287 /// that is stored in consecutive registers with the same channel.
294 BUILD_VERTICAL_VECTOR,
295 /// Pointer to the start of the shader's constant data.
301 FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE,
304 TBUFFER_STORE_FORMAT,
305 LAST_AMDGPU_ISD_NUMBER
309 } // End namespace AMDGPUISD
311 } // End namespace llvm