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 LowerCTLZ(SDValue Op, SelectionDAG &DAG) const;
59 SDValue LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG, bool Signed) const;
60 SDValue LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG, bool Signed) const;
61 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
62 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
64 SDValue LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG, bool Signed) const;
65 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
66 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
68 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
70 SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const;
71 SDValue performShlCombine(SDNode *N, DAGCombinerInfo &DCI) const;
72 SDValue performMulCombine(SDNode *N, DAGCombinerInfo &DCI) const;
73 SDValue performCtlzCombine(SDLoc SL, SDValue Cond, SDValue LHS, SDValue RHS,
74 DAGCombinerInfo &DCI) const;
75 SDValue performSelectCombine(SDNode *N, DAGCombinerInfo &DCI) const;
78 static EVT getEquivalentMemType(LLVMContext &Context, EVT VT);
79 static EVT getEquivalentLoadRegType(LLVMContext &Context, EVT VT);
81 virtual SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op,
82 SelectionDAG &DAG) const;
84 /// \brief Split a vector load into a scalar load of each component.
85 SDValue ScalarizeVectorLoad(SDValue Op, SelectionDAG &DAG) const;
87 /// \brief Split a vector load into 2 loads of half the vector.
88 SDValue SplitVectorLoad(SDValue Op, SelectionDAG &DAG) const;
90 /// \brief Split a vector store into a scalar store of each component.
91 SDValue ScalarizeVectorStore(SDValue Op, SelectionDAG &DAG) const;
93 /// \brief Split a vector store into 2 stores of half the vector.
94 SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const;
96 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
97 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
98 SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const;
99 SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
100 SDValue LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const;
101 void LowerUDIVREM64(SDValue Op, SelectionDAG &DAG,
102 SmallVectorImpl<SDValue> &Results) const;
103 bool isHWTrueValue(SDValue Op) const;
104 bool isHWFalseValue(SDValue Op) const;
106 /// The SelectionDAGBuilder will automatically promote function arguments
107 /// with illegal types. However, this does not work for the AMDGPU targets
108 /// since the function arguments are stored in memory as these illegal types.
109 /// In order to handle this properly we need to get the origianl types sizes
110 /// from the LLVM IR Function and fixup the ISD:InputArg values before
111 /// passing them to AnalyzeFormalArguments()
112 void getOriginalFunctionArgs(SelectionDAG &DAG,
114 const SmallVectorImpl<ISD::InputArg> &Ins,
115 SmallVectorImpl<ISD::InputArg> &OrigIns) const;
116 void AnalyzeFormalArguments(CCState &State,
117 const SmallVectorImpl<ISD::InputArg> &Ins) const;
120 AMDGPUTargetLowering(TargetMachine &TM, const AMDGPUSubtarget &STI);
122 bool isFAbsFree(EVT VT) const override;
123 bool isFNegFree(EVT VT) const override;
124 bool isTruncateFree(EVT Src, EVT Dest) const override;
125 bool isTruncateFree(Type *Src, Type *Dest) const override;
127 bool isZExtFree(Type *Src, Type *Dest) const override;
128 bool isZExtFree(EVT Src, EVT Dest) const override;
129 bool isZExtFree(SDValue Val, EVT VT2) const override;
131 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
133 MVT getVectorIdxTy(const DataLayout &) const override;
134 bool isSelectSupported(SelectSupportKind) const override;
136 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
137 bool ShouldShrinkFPConstant(EVT VT) const override;
138 bool shouldReduceLoadWidth(SDNode *Load,
139 ISD::LoadExtType ExtType,
140 EVT ExtVT) const override;
142 bool isLoadBitCastBeneficial(EVT, EVT) const override;
144 bool storeOfVectorConstantIsCheap(EVT MemVT,
146 unsigned AS) const override;
147 bool aggressivelyPreferBuildVectorSources(EVT VecVT) const override;
148 bool isCheapToSpeculateCttz() const override;
149 bool isCheapToSpeculateCtlz() const override;
151 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv,
153 const SmallVectorImpl<ISD::OutputArg> &Outs,
154 const SmallVectorImpl<SDValue> &OutVals,
155 SDLoc DL, SelectionDAG &DAG) const override;
156 SDValue LowerCall(CallLoweringInfo &CLI,
157 SmallVectorImpl<SDValue> &InVals) const override;
159 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op,
160 SelectionDAG &DAG) const;
162 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
163 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
164 void ReplaceNodeResults(SDNode * N,
165 SmallVectorImpl<SDValue> &Results,
166 SelectionDAG &DAG) const override;
168 SDValue LowerIntrinsicIABS(SDValue Op, SelectionDAG &DAG) const;
169 SDValue LowerIntrinsicLRP(SDValue Op, SelectionDAG &DAG) const;
170 SDValue CombineFMinMaxLegacy(SDLoc DL,
177 DAGCombinerInfo &DCI) const;
179 const char* getTargetNodeName(unsigned Opcode) const override;
181 SDValue getRsqrtEstimate(SDValue Operand,
182 DAGCombinerInfo &DCI,
183 unsigned &RefinementSteps,
184 bool &UseOneConstNR) const override;
185 SDValue getRecipEstimate(SDValue Operand,
186 DAGCombinerInfo &DCI,
187 unsigned &RefinementSteps) const override;
189 virtual SDNode *PostISelFolding(MachineSDNode *N,
190 SelectionDAG &DAG) const {
194 /// \brief Determine which of the bits specified in \p Mask are known to be
195 /// either zero or one and return them in the \p KnownZero and \p KnownOne
197 void computeKnownBitsForTargetNode(const SDValue Op,
200 const SelectionDAG &DAG,
201 unsigned Depth = 0) const override;
203 unsigned ComputeNumSignBitsForTargetNode(SDValue Op, const SelectionDAG &DAG,
204 unsigned Depth = 0) const override;
206 /// \brief Helper function that adds Reg to the LiveIn list of the DAG's
209 /// \returns a RegisterSDNode representing Reg.
210 virtual SDValue CreateLiveInRegister(SelectionDAG &DAG,
211 const TargetRegisterClass *RC,
212 unsigned Reg, EVT VT) const;
214 enum ImplicitParameter {
219 /// \brief Helper function that returns the byte offset of the given
220 /// type of implicit parameter.
221 uint32_t getImplicitParameterOffset(const AMDGPUMachineFunction *MFI,
222 const ImplicitParameter Param) const;
225 namespace AMDGPUISD {
227 enum NodeType : unsigned {
229 FIRST_NUMBER = ISD::BUILTIN_OP_END,
230 CALL, // Function call based on a single integer
231 UMUL, // 32bit unsigned multiplication
234 // End AMDIL ISD Opcodes
239 // SIN_HW, COS_HW - f32 for SI, 1 ULP max error, valid from -100 pi to 100 pi.
240 // Denormals handled on some parts.
255 TRIG_PREOP, // 1 ULP max error for f64
257 // RCP, RSQ - For f32, 1 ULP max error, no denormal handling.
258 // For f64, max error 2^29 ULP, handles denormals.
268 BFE_U32, // Extract range of bits with zero extension to 32-bits.
269 BFE_I32, // Extract range of bits with sign extension to 32-bits.
270 BFI, // (src0 & src1) | (~src0 & src2)
271 BFM, // Insert a range of bits into a 32-bit word.
272 FFBH_U32, // ctlz with -1 if input is zero.
288 // These cvt_f32_ubyte* nodes need to remain consecutive and in order.
293 /// This node is for VLIW targets and it is used to represent a vector
294 /// that is stored in consecutive registers with the same channel.
301 BUILD_VERTICAL_VECTOR,
302 /// Pointer to the start of the shader's constant data.
308 FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE,
311 TBUFFER_STORE_FORMAT,
312 LAST_AMDGPU_ISD_NUMBER
316 } // End namespace AMDGPUISD
318 } // End namespace llvm