xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/SIISelLowering.h (revision 8c2f6c3be0125142d3c1782e4b0ee0634c584b9e)
1 //===-- SIISelLowering.h - SI DAG Lowering Interface ------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// SI DAG Lowering interface definition
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_LIB_TARGET_AMDGPU_SIISELLOWERING_H
15 #define LLVM_LIB_TARGET_AMDGPU_SIISELLOWERING_H
16 
17 #include "AMDGPUISelLowering.h"
18 #include "AMDGPUArgumentUsageInfo.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 
21 namespace llvm {
22 
23 class GCNSubtarget;
24 class SIMachineFunctionInfo;
25 class SIRegisterInfo;
26 
27 namespace AMDGPU {
28 struct ImageDimIntrinsicInfo;
29 }
30 
31 class SITargetLowering final : public AMDGPUTargetLowering {
32 private:
33   const GCNSubtarget *Subtarget;
34 
35 public:
36   MVT getRegisterTypeForCallingConv(LLVMContext &Context,
37                                     CallingConv::ID CC,
38                                     EVT VT) const override;
39   unsigned getNumRegistersForCallingConv(LLVMContext &Context,
40                                          CallingConv::ID CC,
41                                          EVT VT) const override;
42 
43   unsigned getVectorTypeBreakdownForCallingConv(
44     LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT,
45     unsigned &NumIntermediates, MVT &RegisterVT) const override;
46 
47 private:
48   SDValue lowerKernArgParameterPtr(SelectionDAG &DAG, const SDLoc &SL,
49                                    SDValue Chain, uint64_t Offset) const;
50   SDValue getImplicitArgPtr(SelectionDAG &DAG, const SDLoc &SL) const;
51   SDValue getLDSKernelId(SelectionDAG &DAG, const SDLoc &SL) const;
52   SDValue lowerKernargMemParameter(SelectionDAG &DAG, EVT VT, EVT MemVT,
53                                    const SDLoc &SL, SDValue Chain,
54                                    uint64_t Offset, Align Alignment,
55                                    bool Signed,
56                                    const ISD::InputArg *Arg = nullptr) const;
57   SDValue loadImplicitKernelArgument(SelectionDAG &DAG, MVT VT, const SDLoc &DL,
58                                      Align Alignment,
59                                      ImplicitParameter Param) const;
60 
61   SDValue lowerStackParameter(SelectionDAG &DAG, CCValAssign &VA,
62                               const SDLoc &SL, SDValue Chain,
63                               const ISD::InputArg &Arg) const;
64   SDValue getPreloadedValue(SelectionDAG &DAG,
65                             const SIMachineFunctionInfo &MFI,
66                             EVT VT,
67                             AMDGPUFunctionArgInfo::PreloadedValue) const;
68 
69   SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op,
70                              SelectionDAG &DAG) const override;
71   SDValue lowerImplicitZextParam(SelectionDAG &DAG, SDValue Op,
72                                  MVT VT, unsigned Offset) const;
73   SDValue lowerImage(SDValue Op, const AMDGPU::ImageDimIntrinsicInfo *Intr,
74                      SelectionDAG &DAG, bool WithChain) const;
75   SDValue lowerSBuffer(EVT VT, SDLoc DL, SDValue Rsrc, SDValue Offset,
76                        SDValue CachePolicy, SelectionDAG &DAG) const;
77 
78   SDValue lowerRawBufferAtomicIntrin(SDValue Op, SelectionDAG &DAG,
79                                      unsigned NewOpcode) const;
80   SDValue lowerStructBufferAtomicIntrin(SDValue Op, SelectionDAG &DAG,
81                                         unsigned NewOpcode) const;
82 
83   SDValue lowerWaveID(SelectionDAG &DAG, SDValue Op) const;
84   SDValue lowerWorkitemID(SelectionDAG &DAG, SDValue Op, unsigned Dim,
85                           const ArgDescriptor &ArgDesc) const;
86 
87   SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
88   SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const;
89   SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
90 
91   // The raw.tbuffer and struct.tbuffer intrinsics have two offset args: offset
92   // (the offset that is included in bounds checking and swizzling, to be split
93   // between the instruction's voffset and immoffset fields) and soffset (the
94   // offset that is excluded from bounds checking and swizzling, to go in the
95   // instruction's soffset field).  This function takes the first kind of
96   // offset and figures out how to split it between voffset and immoffset.
97   std::pair<SDValue, SDValue> splitBufferOffsets(SDValue Offset,
98                                                  SelectionDAG &DAG) const;
99 
100   SDValue widenLoad(LoadSDNode *Ld, DAGCombinerInfo &DCI) const;
101   SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
102   SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
103   SDValue lowerFastUnsafeFDIV(SDValue Op, SelectionDAG &DAG) const;
104   SDValue lowerFastUnsafeFDIV64(SDValue Op, SelectionDAG &DAG) const;
105   SDValue lowerFDIV_FAST(SDValue Op, SelectionDAG &DAG) const;
106   SDValue LowerFDIV16(SDValue Op, SelectionDAG &DAG) const;
107   SDValue LowerFDIV32(SDValue Op, SelectionDAG &DAG) const;
108   SDValue LowerFDIV64(SDValue Op, SelectionDAG &DAG) const;
109   SDValue LowerFDIV(SDValue Op, SelectionDAG &DAG) const;
110   SDValue LowerFFREXP(SDValue Op, SelectionDAG &DAG) const;
111   SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
112   SDValue LowerTrig(SDValue Op, SelectionDAG &DAG) const;
113   SDValue lowerFSQRTF16(SDValue Op, SelectionDAG &DAG) const;
114   SDValue lowerFSQRTF32(SDValue Op, SelectionDAG &DAG) const;
115   SDValue lowerFSQRTF64(SDValue Op, SelectionDAG &DAG) const;
116   SDValue LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const;
117   SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
118   SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
119   SDValue adjustLoadValueType(unsigned Opcode, MemSDNode *M,
120                               SelectionDAG &DAG, ArrayRef<SDValue> Ops,
121                               bool IsIntrinsic = false) const;
122 
123   SDValue lowerIntrinsicLoad(MemSDNode *M, bool IsFormat, SelectionDAG &DAG,
124                              ArrayRef<SDValue> Ops) const;
125 
126   // Call DAG.getMemIntrinsicNode for a load, but first widen a dwordx3 type to
127   // dwordx4 if on SI.
128   SDValue getMemIntrinsicNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
129                               ArrayRef<SDValue> Ops, EVT MemVT,
130                               MachineMemOperand *MMO, SelectionDAG &DAG) const;
131 
132   SDValue handleD16VData(SDValue VData, SelectionDAG &DAG,
133                          bool ImageStore = false) const;
134 
135   /// Converts \p Op, which must be of floating point type, to the
136   /// floating point type \p VT, by either extending or truncating it.
137   SDValue getFPExtOrFPRound(SelectionDAG &DAG,
138                             SDValue Op,
139                             const SDLoc &DL,
140                             EVT VT) const;
141 
142   SDValue convertArgType(
143     SelectionDAG &DAG, EVT VT, EVT MemVT, const SDLoc &SL, SDValue Val,
144     bool Signed, const ISD::InputArg *Arg = nullptr) const;
145 
146   /// Custom lowering for ISD::FP_ROUND for MVT::f16.
147   SDValue lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
148   SDValue lowerFMINNUM_FMAXNUM(SDValue Op, SelectionDAG &DAG) const;
149   SDValue lowerFLDEXP(SDValue Op, SelectionDAG &DAG) const;
150   SDValue lowerMUL(SDValue Op, SelectionDAG &DAG) const;
151   SDValue lowerXMULO(SDValue Op, SelectionDAG &DAG) const;
152   SDValue lowerXMUL_LOHI(SDValue Op, SelectionDAG &DAG) const;
153 
154   SDValue getSegmentAperture(unsigned AS, const SDLoc &DL,
155                              SelectionDAG &DAG) const;
156 
157   SDValue lowerADDRSPACECAST(SDValue Op, SelectionDAG &DAG) const;
158   SDValue lowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
159   SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
160   SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
161   SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
162   SDValue lowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
163   SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
164 
165   SDValue lowerTRAP(SDValue Op, SelectionDAG &DAG) const;
166   SDValue lowerTrapEndpgm(SDValue Op, SelectionDAG &DAG) const;
167   SDValue lowerTrapHsaQueuePtr(SDValue Op, SelectionDAG &DAG) const;
168   SDValue lowerTrapHsa(SDValue Op, SelectionDAG &DAG) const;
169   SDValue lowerDEBUGTRAP(SDValue Op, SelectionDAG &DAG) const;
170 
171   SDNode *adjustWritemask(MachineSDNode *&N, SelectionDAG &DAG) const;
172 
173   SDValue performUCharToFloatCombine(SDNode *N,
174                                      DAGCombinerInfo &DCI) const;
175   SDValue performFCopySignCombine(SDNode *N, DAGCombinerInfo &DCI) const;
176 
177   SDValue performSHLPtrCombine(SDNode *N,
178                                unsigned AS,
179                                EVT MemVT,
180                                DAGCombinerInfo &DCI) const;
181 
182   SDValue performMemSDNodeCombine(MemSDNode *N, DAGCombinerInfo &DCI) const;
183 
184   SDValue splitBinaryBitConstantOp(DAGCombinerInfo &DCI, const SDLoc &SL,
185                                    unsigned Opc, SDValue LHS,
186                                    const ConstantSDNode *CRHS) const;
187 
188   SDValue performAndCombine(SDNode *N, DAGCombinerInfo &DCI) const;
189   SDValue performOrCombine(SDNode *N, DAGCombinerInfo &DCI) const;
190   SDValue performXorCombine(SDNode *N, DAGCombinerInfo &DCI) const;
191   SDValue performZeroExtendCombine(SDNode *N, DAGCombinerInfo &DCI) const;
192   SDValue performSignExtendInRegCombine(SDNode *N, DAGCombinerInfo &DCI) const;
193   SDValue performClassCombine(SDNode *N, DAGCombinerInfo &DCI) const;
194   SDValue getCanonicalConstantFP(SelectionDAG &DAG, const SDLoc &SL, EVT VT,
195                                  const APFloat &C) const;
196   SDValue performFCanonicalizeCombine(SDNode *N, DAGCombinerInfo &DCI) const;
197 
198   SDValue performFPMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL,
199                                   SDValue Op0, SDValue Op1) const;
200   SDValue performIntMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL,
201                                    SDValue Src, SDValue MinVal, SDValue MaxVal,
202                                    bool Signed) const;
203   SDValue performMinMaxCombine(SDNode *N, DAGCombinerInfo &DCI) const;
204   SDValue performFMed3Combine(SDNode *N, DAGCombinerInfo &DCI) const;
205   SDValue performCvtPkRTZCombine(SDNode *N, DAGCombinerInfo &DCI) const;
206   SDValue performExtractVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const;
207   SDValue performInsertVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const;
208   SDValue performFPRoundCombine(SDNode *N, DAGCombinerInfo &DCI) const;
209 
210   SDValue reassociateScalarOps(SDNode *N, SelectionDAG &DAG) const;
211   unsigned getFusedOpcode(const SelectionDAG &DAG,
212                           const SDNode *N0, const SDNode *N1) const;
213   SDValue tryFoldToMad64_32(SDNode *N, DAGCombinerInfo &DCI) const;
214   SDValue performAddCombine(SDNode *N, DAGCombinerInfo &DCI) const;
215   SDValue performAddCarrySubCarryCombine(SDNode *N, DAGCombinerInfo &DCI) const;
216   SDValue performSubCombine(SDNode *N, DAGCombinerInfo &DCI) const;
217   SDValue performFAddCombine(SDNode *N, DAGCombinerInfo &DCI) const;
218   SDValue performFSubCombine(SDNode *N, DAGCombinerInfo &DCI) const;
219   SDValue performFDivCombine(SDNode *N, DAGCombinerInfo &DCI) const;
220   SDValue performFMACombine(SDNode *N, DAGCombinerInfo &DCI) const;
221   SDValue performSetCCCombine(SDNode *N, DAGCombinerInfo &DCI) const;
222   SDValue performCvtF32UByteNCombine(SDNode *N, DAGCombinerInfo &DCI) const;
223   SDValue performClampCombine(SDNode *N, DAGCombinerInfo &DCI) const;
224   SDValue performRcpCombine(SDNode *N, DAGCombinerInfo &DCI) const;
225 
226   bool isLegalFlatAddressingMode(const AddrMode &AM, unsigned AddrSpace,
227                                  uint64_t FlatVariant) const;
228   bool isLegalMUBUFAddressingMode(const AddrMode &AM) const;
229 
230   unsigned isCFIntrinsic(const SDNode *Intr) const;
231 
232 public:
233   /// \returns True if fixup needs to be emitted for given global value \p GV,
234   /// false otherwise.
235   bool shouldEmitFixup(const GlobalValue *GV) const;
236 
237   /// \returns True if GOT relocation needs to be emitted for given global value
238   /// \p GV, false otherwise.
239   bool shouldEmitGOTReloc(const GlobalValue *GV) const;
240 
241   /// \returns True if PC-relative relocation needs to be emitted for given
242   /// global value \p GV, false otherwise.
243   bool shouldEmitPCReloc(const GlobalValue *GV) const;
244 
245   /// \returns true if this should use a literal constant for an LDS address,
246   /// and not emit a relocation for an LDS global.
247   bool shouldUseLDSConstAddress(const GlobalValue *GV) const;
248 
249   /// Check if EXTRACT_VECTOR_ELT/INSERT_VECTOR_ELT (<n x e>, var-idx) should be
250   /// expanded into a set of cmp/select instructions.
251   static bool shouldExpandVectorDynExt(unsigned EltSize, unsigned NumElem,
252                                        bool IsDivergentIdx,
253                                        const GCNSubtarget *Subtarget);
254 
255   bool shouldExpandVectorDynExt(SDNode *N) const;
256 
257 private:
258   // Analyze a combined offset from an amdgcn_buffer_ intrinsic and store the
259   // three offsets (voffset, soffset and instoffset) into the SDValue[3] array
260   // pointed to by Offsets.
261   void setBufferOffsets(SDValue CombinedOffset, SelectionDAG &DAG,
262                         SDValue *Offsets, Align Alignment = Align(4)) const;
263 
264   // Convert the i128 that an addrspace(8) pointer is natively represented as
265   // into the v4i32 that all the buffer intrinsics expect to receive. We can't
266   // add register classes for i128 on pain of the promotion logic going haywire,
267   // so this slightly ugly hack is what we've got. If passed a non-pointer
268   // argument (as would be seen in older buffer intrinsics), does nothing.
269   SDValue bufferRsrcPtrToVector(SDValue MaybePointer, SelectionDAG &DAG) const;
270 
271   // Wrap a 64-bit pointer into a v4i32 (which is how all SelectionDAG code
272   // represents ptr addrspace(8)) using the flags specified in the intrinsic.
273   SDValue lowerPointerAsRsrcIntrin(SDNode *Op, SelectionDAG &DAG) const;
274 
275   // Handle 8 bit and 16 bit buffer loads
276   SDValue handleByteShortBufferLoads(SelectionDAG &DAG, EVT LoadVT, SDLoc DL,
277                                      ArrayRef<SDValue> Ops,
278                                      MachineMemOperand *MMO) const;
279 
280   // Handle 8 bit and 16 bit buffer stores
281   SDValue handleByteShortBufferStores(SelectionDAG &DAG, EVT VDataType,
282                                       SDLoc DL, SDValue Ops[],
283                                       MemSDNode *M) const;
284 
285 public:
286   SITargetLowering(const TargetMachine &tm, const GCNSubtarget &STI);
287 
288   const GCNSubtarget *getSubtarget() const;
289 
290   bool isFPExtFoldable(const SelectionDAG &DAG, unsigned Opcode, EVT DestVT,
291                        EVT SrcVT) const override;
292 
293   bool isFPExtFoldable(const MachineInstr &MI, unsigned Opcode, LLT DestTy,
294                        LLT SrcTy) const override;
295 
296   bool isShuffleMaskLegal(ArrayRef<int> /*Mask*/, EVT /*VT*/) const override;
297 
298   // While address space 7 should never make it to codegen, it still needs to
299   // have a MVT to prevent some analyses that query this function from breaking,
300   // so, to work around the lack of i160, map it to v5i32.
301   MVT getPointerTy(const DataLayout &DL, unsigned AS) const override;
302   MVT getPointerMemTy(const DataLayout &DL, unsigned AS) const override;
303 
304   bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &,
305                           MachineFunction &MF,
306                           unsigned IntrinsicID) const override;
307 
308   bool getAddrModeArguments(IntrinsicInst * /*I*/,
309                             SmallVectorImpl<Value*> &/*Ops*/,
310                             Type *&/*AccessTy*/) const override;
311 
312   bool isLegalGlobalAddressingMode(const AddrMode &AM) const;
313   bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
314                              unsigned AS,
315                              Instruction *I = nullptr) const override;
316 
317   bool canMergeStoresTo(unsigned AS, EVT MemVT,
318                         const MachineFunction &MF) const override;
319 
320   bool allowsMisalignedMemoryAccessesImpl(
321       unsigned Size, unsigned AddrSpace, Align Alignment,
322       MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
323       unsigned *IsFast = nullptr) const;
324 
325   bool allowsMisalignedMemoryAccesses(
326       LLT Ty, unsigned AddrSpace, Align Alignment,
327       MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
328       unsigned *IsFast = nullptr) const override {
329     if (IsFast)
330       *IsFast = 0;
331     return allowsMisalignedMemoryAccessesImpl(Ty.getSizeInBits(), AddrSpace,
332                                               Alignment, Flags, IsFast);
333   }
334 
335   bool allowsMisalignedMemoryAccesses(
336       EVT VT, unsigned AS, Align Alignment,
337       MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
338       unsigned *IsFast = nullptr) const override;
339 
340   EVT getOptimalMemOpType(const MemOp &Op,
341                           const AttributeList &FuncAttributes) const override;
342 
343   bool isMemOpUniform(const SDNode *N) const;
344   bool isMemOpHasNoClobberedMemOperand(const SDNode *N) const;
345 
346   static bool isNonGlobalAddrSpace(unsigned AS);
347 
348   bool isFreeAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
349 
350   TargetLoweringBase::LegalizeTypeAction
351   getPreferredVectorAction(MVT VT) const override;
352 
353   bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
354                                         Type *Ty) const override;
355 
356   bool isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
357                                unsigned Index) const override;
358 
359   bool isTypeDesirableForOp(unsigned Op, EVT VT) const override;
360 
361   bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
362 
363   unsigned combineRepeatedFPDivisors() const override {
364     // Combine multiple FDIVs with the same divisor into multiple FMULs by the
365     // reciprocal.
366     return 2;
367   }
368 
369   bool supportSplitCSR(MachineFunction *MF) const override;
370   void initializeSplitCSR(MachineBasicBlock *Entry) const override;
371   void insertCopiesSplitCSR(
372     MachineBasicBlock *Entry,
373     const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
374 
375   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
376                                bool isVarArg,
377                                const SmallVectorImpl<ISD::InputArg> &Ins,
378                                const SDLoc &DL, SelectionDAG &DAG,
379                                SmallVectorImpl<SDValue> &InVals) const override;
380 
381   bool CanLowerReturn(CallingConv::ID CallConv,
382                       MachineFunction &MF, bool isVarArg,
383                       const SmallVectorImpl<ISD::OutputArg> &Outs,
384                       LLVMContext &Context) const override;
385 
386   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
387                       const SmallVectorImpl<ISD::OutputArg> &Outs,
388                       const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
389                       SelectionDAG &DAG) const override;
390 
391   void passSpecialInputs(
392     CallLoweringInfo &CLI,
393     CCState &CCInfo,
394     const SIMachineFunctionInfo &Info,
395     SmallVectorImpl<std::pair<unsigned, SDValue>> &RegsToPass,
396     SmallVectorImpl<SDValue> &MemOpChains,
397     SDValue Chain) const;
398 
399   SDValue LowerCallResult(SDValue Chain, SDValue InGlue,
400                           CallingConv::ID CallConv, bool isVarArg,
401                           const SmallVectorImpl<ISD::InputArg> &Ins,
402                           const SDLoc &DL, SelectionDAG &DAG,
403                           SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
404                           SDValue ThisVal) const;
405 
406   bool mayBeEmittedAsTailCall(const CallInst *) const override;
407 
408   bool isEligibleForTailCallOptimization(
409     SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
410     const SmallVectorImpl<ISD::OutputArg> &Outs,
411     const SmallVectorImpl<SDValue> &OutVals,
412     const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const;
413 
414   SDValue LowerCall(CallLoweringInfo &CLI,
415                     SmallVectorImpl<SDValue> &InVals) const override;
416 
417   SDValue lowerDYNAMIC_STACKALLOCImpl(SDValue Op, SelectionDAG &DAG) const;
418   SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
419   SDValue LowerSTACKSAVE(SDValue Op, SelectionDAG &DAG) const;
420   SDValue lowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
421 
422   SDValue lowerPREFETCH(SDValue Op, SelectionDAG &DAG) const;
423   SDValue lowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
424 
425   Register getRegisterByName(const char* RegName, LLT VT,
426                              const MachineFunction &MF) const override;
427 
428   MachineBasicBlock *splitKillBlock(MachineInstr &MI,
429                                     MachineBasicBlock *BB) const;
430 
431   void bundleInstWithWaitcnt(MachineInstr &MI) const;
432   MachineBasicBlock *emitGWSMemViolTestLoop(MachineInstr &MI,
433                                             MachineBasicBlock *BB) const;
434 
435   MachineBasicBlock *
436   EmitInstrWithCustomInserter(MachineInstr &MI,
437                               MachineBasicBlock *BB) const override;
438 
439   bool hasAtomicFaddRtnForTy(SDValue &Op) const;
440   bool enableAggressiveFMAFusion(EVT VT) const override;
441   bool enableAggressiveFMAFusion(LLT Ty) const override;
442   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
443                          EVT VT) const override;
444   MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override;
445   LLT getPreferredShiftAmountTy(LLT Ty) const override;
446 
447   bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
448                                   EVT VT) const override;
449   bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
450                                   const LLT Ty) const override;
451   bool isFMADLegal(const SelectionDAG &DAG, const SDNode *N) const override;
452   bool isFMADLegal(const MachineInstr &MI, const LLT Ty) const override;
453 
454   SDValue splitUnaryVectorOp(SDValue Op, SelectionDAG &DAG) const;
455   SDValue splitBinaryVectorOp(SDValue Op, SelectionDAG &DAG) const;
456   SDValue splitTernaryVectorOp(SDValue Op, SelectionDAG &DAG) const;
457   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
458 
459   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
460                           SelectionDAG &DAG) const override;
461 
462   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
463   SDNode *PostISelFolding(MachineSDNode *N, SelectionDAG &DAG) const override;
464   void AddIMGInit(MachineInstr &MI) const;
465   void AdjustInstrPostInstrSelection(MachineInstr &MI,
466                                      SDNode *Node) const override;
467 
468   SDNode *legalizeTargetIndependentNode(SDNode *Node, SelectionDAG &DAG) const;
469 
470   MachineSDNode *wrapAddr64Rsrc(SelectionDAG &DAG, const SDLoc &DL,
471                                 SDValue Ptr) const;
472   MachineSDNode *buildRSRC(SelectionDAG &DAG, const SDLoc &DL, SDValue Ptr,
473                            uint32_t RsrcDword1, uint64_t RsrcDword2And3) const;
474   std::pair<unsigned, const TargetRegisterClass *>
475   getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
476                                StringRef Constraint, MVT VT) const override;
477   ConstraintType getConstraintType(StringRef Constraint) const override;
478   void LowerAsmOperandForConstraint(SDValue Op, StringRef Constraint,
479                                     std::vector<SDValue> &Ops,
480                                     SelectionDAG &DAG) const override;
481   bool getAsmOperandConstVal(SDValue Op, uint64_t &Val) const;
482   bool checkAsmConstraintVal(SDValue Op, StringRef Constraint,
483                              uint64_t Val) const;
484   bool checkAsmConstraintValA(SDValue Op,
485                               uint64_t Val,
486                               unsigned MaxSize = 64) const;
487   SDValue copyToM0(SelectionDAG &DAG, SDValue Chain, const SDLoc &DL,
488                    SDValue V) const;
489 
490   void finalizeLowering(MachineFunction &MF) const override;
491 
492   void computeKnownBitsForTargetNode(const SDValue Op, KnownBits &Known,
493                                      const APInt &DemandedElts,
494                                      const SelectionDAG &DAG,
495                                      unsigned Depth = 0) const override;
496   void computeKnownBitsForFrameIndex(int FrameIdx,
497                                      KnownBits &Known,
498                                      const MachineFunction &MF) const override;
499   void computeKnownBitsForTargetInstr(GISelKnownBits &Analysis, Register R,
500                                       KnownBits &Known,
501                                       const APInt &DemandedElts,
502                                       const MachineRegisterInfo &MRI,
503                                       unsigned Depth = 0) const override;
504 
505   Align computeKnownAlignForTargetInstr(GISelKnownBits &Analysis, Register R,
506                                         const MachineRegisterInfo &MRI,
507                                         unsigned Depth = 0) const override;
508   bool isSDNodeSourceOfDivergence(const SDNode *N, FunctionLoweringInfo *FLI,
509                                   UniformityInfo *UA) const override;
510 
511   bool hasMemSDNodeUser(SDNode *N) const;
512 
513   bool isReassocProfitable(SelectionDAG &DAG, SDValue N0,
514                            SDValue N1) const override;
515 
516   bool isReassocProfitable(MachineRegisterInfo &MRI, Register N0,
517                            Register N1) const override;
518 
519   bool isCanonicalized(SelectionDAG &DAG, SDValue Op,
520                        unsigned MaxDepth = 5) const;
521   bool isCanonicalized(Register Reg, MachineFunction &MF,
522                        unsigned MaxDepth = 5) const;
523   bool denormalsEnabledForType(const SelectionDAG &DAG, EVT VT) const;
524   bool denormalsEnabledForType(LLT Ty, MachineFunction &MF) const;
525 
526   bool checkForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
527                                  const TargetRegisterInfo *TRI,
528                                  const TargetInstrInfo *TII, unsigned &PhysReg,
529                                  int &Cost) const override;
530 
531   bool isKnownNeverNaNForTargetNode(SDValue Op,
532                                     const SelectionDAG &DAG,
533                                     bool SNaN = false,
534                                     unsigned Depth = 0) const override;
535   AtomicExpansionKind shouldExpandAtomicRMWInIR(AtomicRMWInst *) const override;
536   AtomicExpansionKind shouldExpandAtomicLoadInIR(LoadInst *LI) const override;
537   AtomicExpansionKind shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
538   AtomicExpansionKind
539   shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override;
540   void emitExpandAtomicRMW(AtomicRMWInst *AI) const override;
541 
542   LoadInst *
543   lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
544 
545   const TargetRegisterClass *getRegClassFor(MVT VT,
546                                             bool isDivergent) const override;
547   bool requiresUniformRegister(MachineFunction &MF,
548                                const Value *V) const override;
549   Align getPrefLoopAlignment(MachineLoop *ML) const override;
550 
551   void allocateHSAUserSGPRs(CCState &CCInfo,
552                             MachineFunction &MF,
553                             const SIRegisterInfo &TRI,
554                             SIMachineFunctionInfo &Info) const;
555 
556   void allocatePreloadKernArgSGPRs(CCState &CCInfo,
557                                    SmallVectorImpl<CCValAssign> &ArgLocs,
558                                    const SmallVectorImpl<ISD::InputArg> &Ins,
559                                    MachineFunction &MF,
560                                    const SIRegisterInfo &TRI,
561                                    SIMachineFunctionInfo &Info) const;
562 
563   void allocateLDSKernelId(CCState &CCInfo, MachineFunction &MF,
564                            const SIRegisterInfo &TRI,
565                            SIMachineFunctionInfo &Info) const;
566 
567   void allocateSystemSGPRs(CCState &CCInfo,
568                            MachineFunction &MF,
569                            SIMachineFunctionInfo &Info,
570                            CallingConv::ID CallConv,
571                            bool IsShader) const;
572 
573   void allocateSpecialEntryInputVGPRs(CCState &CCInfo,
574                                       MachineFunction &MF,
575                                       const SIRegisterInfo &TRI,
576                                       SIMachineFunctionInfo &Info) const;
577   void allocateSpecialInputSGPRs(
578     CCState &CCInfo,
579     MachineFunction &MF,
580     const SIRegisterInfo &TRI,
581     SIMachineFunctionInfo &Info) const;
582 
583   void allocateSpecialInputVGPRs(CCState &CCInfo,
584                                  MachineFunction &MF,
585                                  const SIRegisterInfo &TRI,
586                                  SIMachineFunctionInfo &Info) const;
587   void allocateSpecialInputVGPRsFixed(CCState &CCInfo,
588                                       MachineFunction &MF,
589                                       const SIRegisterInfo &TRI,
590                                       SIMachineFunctionInfo &Info) const;
591 
592   MachineMemOperand::Flags
593   getTargetMMOFlags(const Instruction &I) const override;
594 };
595 
596 // Returns true if argument is a boolean value which is not serialized into
597 // memory or argument and does not require v_cndmask_b32 to be deserialized.
598 bool isBoolSGPR(SDValue V);
599 
600 } // End namespace llvm
601 
602 #endif
603