1 //===- SIInstrInfo.h - SI Instruction Info 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 /// Interface definition for SIInstrInfo.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H
15 #define LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H
16 
17 #include "AMDGPUMIRFormatter.h"
18 #include "SIRegisterInfo.h"
19 #include "Utils/AMDGPUBaseInfo.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/CodeGen/TargetInstrInfo.h"
22 #include "llvm/CodeGen/TargetSchedule.h"
23 
24 #define GET_INSTRINFO_HEADER
25 #include "AMDGPUGenInstrInfo.inc"
26 
27 namespace llvm {
28 
29 class APInt;
30 class GCNSubtarget;
31 class LiveVariables;
32 class MachineDominatorTree;
33 class MachineRegisterInfo;
34 class RegScavenger;
35 class TargetRegisterClass;
36 class ScheduleHazardRecognizer;
37 
38 class SIInstrInfo final : public AMDGPUGenInstrInfo {
39 private:
40   const SIRegisterInfo RI;
41   const GCNSubtarget &ST;
42   TargetSchedModel SchedModel;
43   mutable std::unique_ptr<AMDGPUMIRFormatter> Formatter;
44 
45   // The inverse predicate should have the negative value.
46   enum BranchPredicate {
47     INVALID_BR = 0,
48     SCC_TRUE = 1,
49     SCC_FALSE = -1,
50     VCCNZ = 2,
51     VCCZ = -2,
52     EXECNZ = -3,
53     EXECZ = 3
54   };
55 
56   using SetVectorType = SmallSetVector<MachineInstr *, 32>;
57 
58   static unsigned getBranchOpcode(BranchPredicate Cond);
59   static BranchPredicate getBranchPredicate(unsigned Opcode);
60 
61 public:
62   unsigned buildExtractSubReg(MachineBasicBlock::iterator MI,
63                               MachineRegisterInfo &MRI,
64                               MachineOperand &SuperReg,
65                               const TargetRegisterClass *SuperRC,
66                               unsigned SubIdx,
67                               const TargetRegisterClass *SubRC) const;
68   MachineOperand buildExtractSubRegOrImm(MachineBasicBlock::iterator MI,
69                                          MachineRegisterInfo &MRI,
70                                          MachineOperand &SuperReg,
71                                          const TargetRegisterClass *SuperRC,
72                                          unsigned SubIdx,
73                                          const TargetRegisterClass *SubRC) const;
74 private:
75   void swapOperands(MachineInstr &Inst) const;
76 
77   std::pair<bool, MachineBasicBlock *>
78   moveScalarAddSub(SetVectorType &Worklist, MachineInstr &Inst,
79                    MachineDominatorTree *MDT = nullptr) const;
80 
81   void lowerSelect(SetVectorType &Worklist, MachineInstr &Inst,
82                    MachineDominatorTree *MDT = nullptr) const;
83 
84   void lowerScalarAbs(SetVectorType &Worklist,
85                       MachineInstr &Inst) const;
86 
87   void lowerScalarXnor(SetVectorType &Worklist,
88                        MachineInstr &Inst) const;
89 
90   void splitScalarNotBinop(SetVectorType &Worklist,
91                            MachineInstr &Inst,
92                            unsigned Opcode) const;
93 
94   void splitScalarBinOpN2(SetVectorType &Worklist,
95                           MachineInstr &Inst,
96                           unsigned Opcode) const;
97 
98   void splitScalar64BitUnaryOp(SetVectorType &Worklist,
99                                MachineInstr &Inst, unsigned Opcode,
100                                bool Swap = false) const;
101 
102   void splitScalar64BitAddSub(SetVectorType &Worklist, MachineInstr &Inst,
103                               MachineDominatorTree *MDT = nullptr) const;
104 
105   void splitScalar64BitBinaryOp(SetVectorType &Worklist, MachineInstr &Inst,
106                                 unsigned Opcode,
107                                 MachineDominatorTree *MDT = nullptr) const;
108 
109   void splitScalar64BitXnor(SetVectorType &Worklist, MachineInstr &Inst,
110                                 MachineDominatorTree *MDT = nullptr) const;
111 
112   void splitScalar64BitBCNT(SetVectorType &Worklist,
113                             MachineInstr &Inst) const;
114   void splitScalar64BitBFE(SetVectorType &Worklist,
115                            MachineInstr &Inst) const;
116   void movePackToVALU(SetVectorType &Worklist,
117                       MachineRegisterInfo &MRI,
118                       MachineInstr &Inst) const;
119 
120   void addUsersToMoveToVALUWorklist(Register Reg, MachineRegisterInfo &MRI,
121                                     SetVectorType &Worklist) const;
122 
123   void addSCCDefUsersToVALUWorklist(MachineOperand &Op,
124                                     MachineInstr &SCCDefInst,
125                                     SetVectorType &Worklist) const;
126   void addSCCDefsToVALUWorklist(MachineOperand &Op,
127                                 SetVectorType &Worklist) const;
128 
129   const TargetRegisterClass *
130   getDestEquivalentVGPRClass(const MachineInstr &Inst) const;
131 
132   bool checkInstOffsetsDoNotOverlap(const MachineInstr &MIa,
133                                     const MachineInstr &MIb) const;
134 
135   Register findUsedSGPR(const MachineInstr &MI, int OpIndices[3]) const;
136 
137 protected:
138   bool swapSourceModifiers(MachineInstr &MI,
139                            MachineOperand &Src0, unsigned Src0OpName,
140                            MachineOperand &Src1, unsigned Src1OpName) const;
141 
142   MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
143                                        unsigned OpIdx0,
144                                        unsigned OpIdx1) const override;
145 
146 public:
147   enum TargetOperandFlags {
148     MO_MASK = 0xf,
149 
150     MO_NONE = 0,
151     // MO_GOTPCREL -> symbol@GOTPCREL -> R_AMDGPU_GOTPCREL.
152     MO_GOTPCREL = 1,
153     // MO_GOTPCREL32_LO -> symbol@gotpcrel32@lo -> R_AMDGPU_GOTPCREL32_LO.
154     MO_GOTPCREL32 = 2,
155     MO_GOTPCREL32_LO = 2,
156     // MO_GOTPCREL32_HI -> symbol@gotpcrel32@hi -> R_AMDGPU_GOTPCREL32_HI.
157     MO_GOTPCREL32_HI = 3,
158     // MO_REL32_LO -> symbol@rel32@lo -> R_AMDGPU_REL32_LO.
159     MO_REL32 = 4,
160     MO_REL32_LO = 4,
161     // MO_REL32_HI -> symbol@rel32@hi -> R_AMDGPU_REL32_HI.
162     MO_REL32_HI = 5,
163 
164     MO_FAR_BRANCH_OFFSET = 6,
165 
166     MO_ABS32_LO = 8,
167     MO_ABS32_HI = 9,
168   };
169 
170   explicit SIInstrInfo(const GCNSubtarget &ST);
171 
getRegisterInfo()172   const SIRegisterInfo &getRegisterInfo() const {
173     return RI;
174   }
175 
getSubtarget()176   const GCNSubtarget &getSubtarget() const {
177     return ST;
178   }
179 
180   bool isReallyTriviallyReMaterializable(const MachineInstr &MI,
181                                          AAResults *AA) const override;
182 
183   bool isIgnorableUse(const MachineOperand &MO) const override;
184 
185   bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
186                                int64_t &Offset1,
187                                int64_t &Offset2) const override;
188 
189   bool getMemOperandsWithOffsetWidth(
190       const MachineInstr &LdSt,
191       SmallVectorImpl<const MachineOperand *> &BaseOps, int64_t &Offset,
192       bool &OffsetIsScalable, unsigned &Width,
193       const TargetRegisterInfo *TRI) const final;
194 
195   bool shouldClusterMemOps(ArrayRef<const MachineOperand *> BaseOps1,
196                            ArrayRef<const MachineOperand *> BaseOps2,
197                            unsigned NumLoads, unsigned NumBytes) const override;
198 
199   bool shouldScheduleLoadsNear(SDNode *Load0, SDNode *Load1, int64_t Offset0,
200                                int64_t Offset1, unsigned NumLoads) const override;
201 
202   void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
203                    const DebugLoc &DL, MCRegister DestReg, MCRegister SrcReg,
204                    bool KillSrc) const override;
205 
206   void materializeImmediate(MachineBasicBlock &MBB,
207                             MachineBasicBlock::iterator MI,
208                             const DebugLoc &DL,
209                             unsigned DestReg,
210                             int64_t Value) const;
211 
212   const TargetRegisterClass *getPreferredSelectRegClass(
213                                unsigned Size) const;
214 
215   Register insertNE(MachineBasicBlock *MBB,
216                     MachineBasicBlock::iterator I, const DebugLoc &DL,
217                     Register SrcReg, int Value) const;
218 
219   Register insertEQ(MachineBasicBlock *MBB,
220                     MachineBasicBlock::iterator I, const DebugLoc &DL,
221                     Register SrcReg, int Value)  const;
222 
223   void storeRegToStackSlot(MachineBasicBlock &MBB,
224                            MachineBasicBlock::iterator MI, Register SrcReg,
225                            bool isKill, int FrameIndex,
226                            const TargetRegisterClass *RC,
227                            const TargetRegisterInfo *TRI) const override;
228 
229   void loadRegFromStackSlot(MachineBasicBlock &MBB,
230                             MachineBasicBlock::iterator MI, Register DestReg,
231                             int FrameIndex, const TargetRegisterClass *RC,
232                             const TargetRegisterInfo *TRI) const override;
233 
234   bool expandPostRAPseudo(MachineInstr &MI) const override;
235 
236   // Splits a V_MOV_B64_DPP_PSEUDO opcode into a pair of v_mov_b32_dpp
237   // instructions. Returns a pair of generated instructions.
238   // Can split either post-RA with physical registers or pre-RA with
239   // virtual registers. In latter case IR needs to be in SSA form and
240   // and a REG_SEQUENCE is produced to define original register.
241   std::pair<MachineInstr*, MachineInstr*>
242   expandMovDPP64(MachineInstr &MI) const;
243 
244   // Returns an opcode that can be used to move a value to a \p DstRC
245   // register.  If there is no hardware instruction that can store to \p
246   // DstRC, then AMDGPU::COPY is returned.
247   unsigned getMovOpcode(const TargetRegisterClass *DstRC) const;
248 
249   const MCInstrDesc &getIndirectRegWriteMovRelPseudo(unsigned VecSize,
250                                                      unsigned EltSize,
251                                                      bool IsSGPR) const;
252 
253   const MCInstrDesc &getIndirectGPRIDXPseudo(unsigned VecSize,
254                                              bool IsIndirectSrc) const;
255   LLVM_READONLY
256   int commuteOpcode(unsigned Opc) const;
257 
258   LLVM_READONLY
commuteOpcode(const MachineInstr & MI)259   inline int commuteOpcode(const MachineInstr &MI) const {
260     return commuteOpcode(MI.getOpcode());
261   }
262 
263   bool findCommutedOpIndices(const MachineInstr &MI, unsigned &SrcOpIdx1,
264                              unsigned &SrcOpIdx2) const override;
265 
266   bool findCommutedOpIndices(MCInstrDesc Desc, unsigned & SrcOpIdx0,
267    unsigned & SrcOpIdx1) const;
268 
269   bool isBranchOffsetInRange(unsigned BranchOpc,
270                              int64_t BrOffset) const override;
271 
272   MachineBasicBlock *getBranchDestBlock(const MachineInstr &MI) const override;
273 
274   unsigned insertIndirectBranch(MachineBasicBlock &MBB,
275                                 MachineBasicBlock &NewDestBB,
276                                 const DebugLoc &DL,
277                                 int64_t BrOffset,
278                                 RegScavenger *RS = nullptr) const override;
279 
280   bool analyzeBranchImpl(MachineBasicBlock &MBB,
281                          MachineBasicBlock::iterator I,
282                          MachineBasicBlock *&TBB,
283                          MachineBasicBlock *&FBB,
284                          SmallVectorImpl<MachineOperand> &Cond,
285                          bool AllowModify) const;
286 
287   bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
288                      MachineBasicBlock *&FBB,
289                      SmallVectorImpl<MachineOperand> &Cond,
290                      bool AllowModify = false) const override;
291 
292   unsigned removeBranch(MachineBasicBlock &MBB,
293                         int *BytesRemoved = nullptr) const override;
294 
295   unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
296                         MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond,
297                         const DebugLoc &DL,
298                         int *BytesAdded = nullptr) const override;
299 
300   bool reverseBranchCondition(
301     SmallVectorImpl<MachineOperand> &Cond) const override;
302 
303   bool canInsertSelect(const MachineBasicBlock &MBB,
304                        ArrayRef<MachineOperand> Cond, Register DstReg,
305                        Register TrueReg, Register FalseReg, int &CondCycles,
306                        int &TrueCycles, int &FalseCycles) const override;
307 
308   void insertSelect(MachineBasicBlock &MBB,
309                     MachineBasicBlock::iterator I, const DebugLoc &DL,
310                     Register DstReg, ArrayRef<MachineOperand> Cond,
311                     Register TrueReg, Register FalseReg) const override;
312 
313   void insertVectorSelect(MachineBasicBlock &MBB,
314                           MachineBasicBlock::iterator I, const DebugLoc &DL,
315                           Register DstReg, ArrayRef<MachineOperand> Cond,
316                           Register TrueReg, Register FalseReg) const;
317 
318   unsigned getAddressSpaceForPseudoSourceKind(
319              unsigned Kind) const override;
320 
321   bool
322   areMemAccessesTriviallyDisjoint(const MachineInstr &MIa,
323                                   const MachineInstr &MIb) const override;
324 
325   bool isFoldableCopy(const MachineInstr &MI) const;
326 
327   bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, Register Reg,
328                      MachineRegisterInfo *MRI) const final;
329 
getMachineCSELookAheadLimit()330   unsigned getMachineCSELookAheadLimit() const override { return 500; }
331 
332   MachineInstr *convertToThreeAddress(MachineFunction::iterator &MBB,
333                                       MachineInstr &MI,
334                                       LiveVariables *LV) const override;
335 
336   bool isSchedulingBoundary(const MachineInstr &MI,
337                             const MachineBasicBlock *MBB,
338                             const MachineFunction &MF) const override;
339 
isSALU(const MachineInstr & MI)340   static bool isSALU(const MachineInstr &MI) {
341     return MI.getDesc().TSFlags & SIInstrFlags::SALU;
342   }
343 
isSALU(uint16_t Opcode)344   bool isSALU(uint16_t Opcode) const {
345     return get(Opcode).TSFlags & SIInstrFlags::SALU;
346   }
347 
isVALU(const MachineInstr & MI)348   static bool isVALU(const MachineInstr &MI) {
349     return MI.getDesc().TSFlags & SIInstrFlags::VALU;
350   }
351 
isVALU(uint16_t Opcode)352   bool isVALU(uint16_t Opcode) const {
353     return get(Opcode).TSFlags & SIInstrFlags::VALU;
354   }
355 
isVMEM(const MachineInstr & MI)356   static bool isVMEM(const MachineInstr &MI) {
357     return isMUBUF(MI) || isMTBUF(MI) || isMIMG(MI);
358   }
359 
isVMEM(uint16_t Opcode)360   bool isVMEM(uint16_t Opcode) const {
361     return isMUBUF(Opcode) || isMTBUF(Opcode) || isMIMG(Opcode);
362   }
363 
isSOP1(const MachineInstr & MI)364   static bool isSOP1(const MachineInstr &MI) {
365     return MI.getDesc().TSFlags & SIInstrFlags::SOP1;
366   }
367 
isSOP1(uint16_t Opcode)368   bool isSOP1(uint16_t Opcode) const {
369     return get(Opcode).TSFlags & SIInstrFlags::SOP1;
370   }
371 
isSOP2(const MachineInstr & MI)372   static bool isSOP2(const MachineInstr &MI) {
373     return MI.getDesc().TSFlags & SIInstrFlags::SOP2;
374   }
375 
isSOP2(uint16_t Opcode)376   bool isSOP2(uint16_t Opcode) const {
377     return get(Opcode).TSFlags & SIInstrFlags::SOP2;
378   }
379 
isSOPC(const MachineInstr & MI)380   static bool isSOPC(const MachineInstr &MI) {
381     return MI.getDesc().TSFlags & SIInstrFlags::SOPC;
382   }
383 
isSOPC(uint16_t Opcode)384   bool isSOPC(uint16_t Opcode) const {
385     return get(Opcode).TSFlags & SIInstrFlags::SOPC;
386   }
387 
isSOPK(const MachineInstr & MI)388   static bool isSOPK(const MachineInstr &MI) {
389     return MI.getDesc().TSFlags & SIInstrFlags::SOPK;
390   }
391 
isSOPK(uint16_t Opcode)392   bool isSOPK(uint16_t Opcode) const {
393     return get(Opcode).TSFlags & SIInstrFlags::SOPK;
394   }
395 
isSOPP(const MachineInstr & MI)396   static bool isSOPP(const MachineInstr &MI) {
397     return MI.getDesc().TSFlags & SIInstrFlags::SOPP;
398   }
399 
isSOPP(uint16_t Opcode)400   bool isSOPP(uint16_t Opcode) const {
401     return get(Opcode).TSFlags & SIInstrFlags::SOPP;
402   }
403 
isPacked(const MachineInstr & MI)404   static bool isPacked(const MachineInstr &MI) {
405     return MI.getDesc().TSFlags & SIInstrFlags::IsPacked;
406   }
407 
isPacked(uint16_t Opcode)408   bool isPacked(uint16_t Opcode) const {
409     return get(Opcode).TSFlags & SIInstrFlags::IsPacked;
410   }
411 
isVOP1(const MachineInstr & MI)412   static bool isVOP1(const MachineInstr &MI) {
413     return MI.getDesc().TSFlags & SIInstrFlags::VOP1;
414   }
415 
isVOP1(uint16_t Opcode)416   bool isVOP1(uint16_t Opcode) const {
417     return get(Opcode).TSFlags & SIInstrFlags::VOP1;
418   }
419 
isVOP2(const MachineInstr & MI)420   static bool isVOP2(const MachineInstr &MI) {
421     return MI.getDesc().TSFlags & SIInstrFlags::VOP2;
422   }
423 
isVOP2(uint16_t Opcode)424   bool isVOP2(uint16_t Opcode) const {
425     return get(Opcode).TSFlags & SIInstrFlags::VOP2;
426   }
427 
isVOP3(const MachineInstr & MI)428   static bool isVOP3(const MachineInstr &MI) {
429     return MI.getDesc().TSFlags & SIInstrFlags::VOP3;
430   }
431 
isVOP3(uint16_t Opcode)432   bool isVOP3(uint16_t Opcode) const {
433     return get(Opcode).TSFlags & SIInstrFlags::VOP3;
434   }
435 
isSDWA(const MachineInstr & MI)436   static bool isSDWA(const MachineInstr &MI) {
437     return MI.getDesc().TSFlags & SIInstrFlags::SDWA;
438   }
439 
isSDWA(uint16_t Opcode)440   bool isSDWA(uint16_t Opcode) const {
441     return get(Opcode).TSFlags & SIInstrFlags::SDWA;
442   }
443 
isVOPC(const MachineInstr & MI)444   static bool isVOPC(const MachineInstr &MI) {
445     return MI.getDesc().TSFlags & SIInstrFlags::VOPC;
446   }
447 
isVOPC(uint16_t Opcode)448   bool isVOPC(uint16_t Opcode) const {
449     return get(Opcode).TSFlags & SIInstrFlags::VOPC;
450   }
451 
isMUBUF(const MachineInstr & MI)452   static bool isMUBUF(const MachineInstr &MI) {
453     return MI.getDesc().TSFlags & SIInstrFlags::MUBUF;
454   }
455 
isMUBUF(uint16_t Opcode)456   bool isMUBUF(uint16_t Opcode) const {
457     return get(Opcode).TSFlags & SIInstrFlags::MUBUF;
458   }
459 
isMTBUF(const MachineInstr & MI)460   static bool isMTBUF(const MachineInstr &MI) {
461     return MI.getDesc().TSFlags & SIInstrFlags::MTBUF;
462   }
463 
isMTBUF(uint16_t Opcode)464   bool isMTBUF(uint16_t Opcode) const {
465     return get(Opcode).TSFlags & SIInstrFlags::MTBUF;
466   }
467 
isSMRD(const MachineInstr & MI)468   static bool isSMRD(const MachineInstr &MI) {
469     return MI.getDesc().TSFlags & SIInstrFlags::SMRD;
470   }
471 
isSMRD(uint16_t Opcode)472   bool isSMRD(uint16_t Opcode) const {
473     return get(Opcode).TSFlags & SIInstrFlags::SMRD;
474   }
475 
476   bool isBufferSMRD(const MachineInstr &MI) const;
477 
isDS(const MachineInstr & MI)478   static bool isDS(const MachineInstr &MI) {
479     return MI.getDesc().TSFlags & SIInstrFlags::DS;
480   }
481 
isDS(uint16_t Opcode)482   bool isDS(uint16_t Opcode) const {
483     return get(Opcode).TSFlags & SIInstrFlags::DS;
484   }
485 
486   bool isAlwaysGDS(uint16_t Opcode) const;
487 
isMIMG(const MachineInstr & MI)488   static bool isMIMG(const MachineInstr &MI) {
489     return MI.getDesc().TSFlags & SIInstrFlags::MIMG;
490   }
491 
isMIMG(uint16_t Opcode)492   bool isMIMG(uint16_t Opcode) const {
493     return get(Opcode).TSFlags & SIInstrFlags::MIMG;
494   }
495 
isGather4(const MachineInstr & MI)496   static bool isGather4(const MachineInstr &MI) {
497     return MI.getDesc().TSFlags & SIInstrFlags::Gather4;
498   }
499 
isGather4(uint16_t Opcode)500   bool isGather4(uint16_t Opcode) const {
501     return get(Opcode).TSFlags & SIInstrFlags::Gather4;
502   }
503 
isFLAT(const MachineInstr & MI)504   static bool isFLAT(const MachineInstr &MI) {
505     return MI.getDesc().TSFlags & SIInstrFlags::FLAT;
506   }
507 
508   // Is a FLAT encoded instruction which accesses a specific segment,
509   // i.e. global_* or scratch_*.
isSegmentSpecificFLAT(const MachineInstr & MI)510   static bool isSegmentSpecificFLAT(const MachineInstr &MI) {
511     auto Flags = MI.getDesc().TSFlags;
512     return Flags & (SIInstrFlags::FlatGlobal | SIInstrFlags::FlatScratch);
513   }
514 
isSegmentSpecificFLAT(uint16_t Opcode)515   bool isSegmentSpecificFLAT(uint16_t Opcode) const {
516     auto Flags = get(Opcode).TSFlags;
517     return Flags & (SIInstrFlags::FlatGlobal | SIInstrFlags::FlatScratch);
518   }
519 
isFLATGlobal(const MachineInstr & MI)520   static bool isFLATGlobal(const MachineInstr &MI) {
521     return MI.getDesc().TSFlags & SIInstrFlags::FlatGlobal;
522   }
523 
isFLATGlobal(uint16_t Opcode)524   bool isFLATGlobal(uint16_t Opcode) const {
525     return get(Opcode).TSFlags & SIInstrFlags::FlatGlobal;
526   }
527 
isFLATScratch(const MachineInstr & MI)528   static bool isFLATScratch(const MachineInstr &MI) {
529     return MI.getDesc().TSFlags & SIInstrFlags::FlatScratch;
530   }
531 
isFLATScratch(uint16_t Opcode)532   bool isFLATScratch(uint16_t Opcode) const {
533     return get(Opcode).TSFlags & SIInstrFlags::FlatScratch;
534   }
535 
536   // Any FLAT encoded instruction, including global_* and scratch_*.
isFLAT(uint16_t Opcode)537   bool isFLAT(uint16_t Opcode) const {
538     return get(Opcode).TSFlags & SIInstrFlags::FLAT;
539   }
540 
isEXP(const MachineInstr & MI)541   static bool isEXP(const MachineInstr &MI) {
542     return MI.getDesc().TSFlags & SIInstrFlags::EXP;
543   }
544 
isEXP(uint16_t Opcode)545   bool isEXP(uint16_t Opcode) const {
546     return get(Opcode).TSFlags & SIInstrFlags::EXP;
547   }
548 
isAtomicNoRet(const MachineInstr & MI)549   static bool isAtomicNoRet(const MachineInstr &MI) {
550     return MI.getDesc().TSFlags & SIInstrFlags::IsAtomicNoRet;
551   }
552 
isAtomicNoRet(uint16_t Opcode)553   bool isAtomicNoRet(uint16_t Opcode) const {
554     return get(Opcode).TSFlags & SIInstrFlags::IsAtomicNoRet;
555   }
556 
isAtomicRet(const MachineInstr & MI)557   static bool isAtomicRet(const MachineInstr &MI) {
558     return MI.getDesc().TSFlags & SIInstrFlags::IsAtomicRet;
559   }
560 
isAtomicRet(uint16_t Opcode)561   bool isAtomicRet(uint16_t Opcode) const {
562     return get(Opcode).TSFlags & SIInstrFlags::IsAtomicRet;
563   }
564 
isAtomic(const MachineInstr & MI)565   static bool isAtomic(const MachineInstr &MI) {
566     return MI.getDesc().TSFlags & (SIInstrFlags::IsAtomicRet |
567                                    SIInstrFlags::IsAtomicNoRet);
568   }
569 
isAtomic(uint16_t Opcode)570   bool isAtomic(uint16_t Opcode) const {
571     return get(Opcode).TSFlags & (SIInstrFlags::IsAtomicRet |
572                                   SIInstrFlags::IsAtomicNoRet);
573   }
574 
isWQM(const MachineInstr & MI)575   static bool isWQM(const MachineInstr &MI) {
576     return MI.getDesc().TSFlags & SIInstrFlags::WQM;
577   }
578 
isWQM(uint16_t Opcode)579   bool isWQM(uint16_t Opcode) const {
580     return get(Opcode).TSFlags & SIInstrFlags::WQM;
581   }
582 
isDisableWQM(const MachineInstr & MI)583   static bool isDisableWQM(const MachineInstr &MI) {
584     return MI.getDesc().TSFlags & SIInstrFlags::DisableWQM;
585   }
586 
isDisableWQM(uint16_t Opcode)587   bool isDisableWQM(uint16_t Opcode) const {
588     return get(Opcode).TSFlags & SIInstrFlags::DisableWQM;
589   }
590 
isVGPRSpill(const MachineInstr & MI)591   static bool isVGPRSpill(const MachineInstr &MI) {
592     return MI.getDesc().TSFlags & SIInstrFlags::VGPRSpill;
593   }
594 
isVGPRSpill(uint16_t Opcode)595   bool isVGPRSpill(uint16_t Opcode) const {
596     return get(Opcode).TSFlags & SIInstrFlags::VGPRSpill;
597   }
598 
isSGPRSpill(const MachineInstr & MI)599   static bool isSGPRSpill(const MachineInstr &MI) {
600     return MI.getDesc().TSFlags & SIInstrFlags::SGPRSpill;
601   }
602 
isSGPRSpill(uint16_t Opcode)603   bool isSGPRSpill(uint16_t Opcode) const {
604     return get(Opcode).TSFlags & SIInstrFlags::SGPRSpill;
605   }
606 
isDPP(const MachineInstr & MI)607   static bool isDPP(const MachineInstr &MI) {
608     return MI.getDesc().TSFlags & SIInstrFlags::DPP;
609   }
610 
isDPP(uint16_t Opcode)611   bool isDPP(uint16_t Opcode) const {
612     return get(Opcode).TSFlags & SIInstrFlags::DPP;
613   }
614 
isTRANS(const MachineInstr & MI)615   static bool isTRANS(const MachineInstr &MI) {
616     return MI.getDesc().TSFlags & SIInstrFlags::TRANS;
617   }
618 
isTRANS(uint16_t Opcode)619   bool isTRANS(uint16_t Opcode) const {
620     return get(Opcode).TSFlags & SIInstrFlags::TRANS;
621   }
622 
isVOP3P(const MachineInstr & MI)623   static bool isVOP3P(const MachineInstr &MI) {
624     return MI.getDesc().TSFlags & SIInstrFlags::VOP3P;
625   }
626 
isVOP3P(uint16_t Opcode)627   bool isVOP3P(uint16_t Opcode) const {
628     return get(Opcode).TSFlags & SIInstrFlags::VOP3P;
629   }
630 
isVINTRP(const MachineInstr & MI)631   static bool isVINTRP(const MachineInstr &MI) {
632     return MI.getDesc().TSFlags & SIInstrFlags::VINTRP;
633   }
634 
isVINTRP(uint16_t Opcode)635   bool isVINTRP(uint16_t Opcode) const {
636     return get(Opcode).TSFlags & SIInstrFlags::VINTRP;
637   }
638 
isMAI(const MachineInstr & MI)639   static bool isMAI(const MachineInstr &MI) {
640     return MI.getDesc().TSFlags & SIInstrFlags::IsMAI;
641   }
642 
isMAI(uint16_t Opcode)643   bool isMAI(uint16_t Opcode) const {
644     return get(Opcode).TSFlags & SIInstrFlags::IsMAI;
645   }
646 
isDOT(const MachineInstr & MI)647   static bool isDOT(const MachineInstr &MI) {
648     return MI.getDesc().TSFlags & SIInstrFlags::IsDOT;
649   }
650 
isDOT(uint16_t Opcode)651   bool isDOT(uint16_t Opcode) const {
652     return get(Opcode).TSFlags & SIInstrFlags::IsDOT;
653   }
654 
isScalarUnit(const MachineInstr & MI)655   static bool isScalarUnit(const MachineInstr &MI) {
656     return MI.getDesc().TSFlags & (SIInstrFlags::SALU | SIInstrFlags::SMRD);
657   }
658 
usesVM_CNT(const MachineInstr & MI)659   static bool usesVM_CNT(const MachineInstr &MI) {
660     return MI.getDesc().TSFlags & SIInstrFlags::VM_CNT;
661   }
662 
usesLGKM_CNT(const MachineInstr & MI)663   static bool usesLGKM_CNT(const MachineInstr &MI) {
664     return MI.getDesc().TSFlags & SIInstrFlags::LGKM_CNT;
665   }
666 
sopkIsZext(const MachineInstr & MI)667   static bool sopkIsZext(const MachineInstr &MI) {
668     return MI.getDesc().TSFlags & SIInstrFlags::SOPK_ZEXT;
669   }
670 
sopkIsZext(uint16_t Opcode)671   bool sopkIsZext(uint16_t Opcode) const {
672     return get(Opcode).TSFlags & SIInstrFlags::SOPK_ZEXT;
673   }
674 
675   /// \returns true if this is an s_store_dword* instruction. This is more
676   /// specific than than isSMEM && mayStore.
isScalarStore(const MachineInstr & MI)677   static bool isScalarStore(const MachineInstr &MI) {
678     return MI.getDesc().TSFlags & SIInstrFlags::SCALAR_STORE;
679   }
680 
isScalarStore(uint16_t Opcode)681   bool isScalarStore(uint16_t Opcode) const {
682     return get(Opcode).TSFlags & SIInstrFlags::SCALAR_STORE;
683   }
684 
isFixedSize(const MachineInstr & MI)685   static bool isFixedSize(const MachineInstr &MI) {
686     return MI.getDesc().TSFlags & SIInstrFlags::FIXED_SIZE;
687   }
688 
isFixedSize(uint16_t Opcode)689   bool isFixedSize(uint16_t Opcode) const {
690     return get(Opcode).TSFlags & SIInstrFlags::FIXED_SIZE;
691   }
692 
hasFPClamp(const MachineInstr & MI)693   static bool hasFPClamp(const MachineInstr &MI) {
694     return MI.getDesc().TSFlags & SIInstrFlags::FPClamp;
695   }
696 
hasFPClamp(uint16_t Opcode)697   bool hasFPClamp(uint16_t Opcode) const {
698     return get(Opcode).TSFlags & SIInstrFlags::FPClamp;
699   }
700 
hasIntClamp(const MachineInstr & MI)701   static bool hasIntClamp(const MachineInstr &MI) {
702     return MI.getDesc().TSFlags & SIInstrFlags::IntClamp;
703   }
704 
getClampMask(const MachineInstr & MI)705   uint64_t getClampMask(const MachineInstr &MI) const {
706     const uint64_t ClampFlags = SIInstrFlags::FPClamp |
707                                 SIInstrFlags::IntClamp |
708                                 SIInstrFlags::ClampLo |
709                                 SIInstrFlags::ClampHi;
710       return MI.getDesc().TSFlags & ClampFlags;
711   }
712 
usesFPDPRounding(const MachineInstr & MI)713   static bool usesFPDPRounding(const MachineInstr &MI) {
714     return MI.getDesc().TSFlags & SIInstrFlags::FPDPRounding;
715   }
716 
usesFPDPRounding(uint16_t Opcode)717   bool usesFPDPRounding(uint16_t Opcode) const {
718     return get(Opcode).TSFlags & SIInstrFlags::FPDPRounding;
719   }
720 
isFPAtomic(const MachineInstr & MI)721   static bool isFPAtomic(const MachineInstr &MI) {
722     return MI.getDesc().TSFlags & SIInstrFlags::FPAtomic;
723   }
724 
isFPAtomic(uint16_t Opcode)725   bool isFPAtomic(uint16_t Opcode) const {
726     return get(Opcode).TSFlags & SIInstrFlags::FPAtomic;
727   }
728 
isVGPRCopy(const MachineInstr & MI)729   bool isVGPRCopy(const MachineInstr &MI) const {
730     assert(MI.isCopy());
731     Register Dest = MI.getOperand(0).getReg();
732     const MachineFunction &MF = *MI.getParent()->getParent();
733     const MachineRegisterInfo &MRI = MF.getRegInfo();
734     return !RI.isSGPRReg(MRI, Dest);
735   }
736 
hasVGPRUses(const MachineInstr & MI)737   bool hasVGPRUses(const MachineInstr &MI) const {
738     const MachineFunction &MF = *MI.getParent()->getParent();
739     const MachineRegisterInfo &MRI = MF.getRegInfo();
740     return llvm::any_of(MI.explicit_uses(),
741                         [&MRI, this](const MachineOperand &MO) {
742       return MO.isReg() && RI.isVGPR(MRI, MO.getReg());});
743   }
744 
745   /// Return true if the instruction modifies the mode register.q
746   static bool modifiesModeRegister(const MachineInstr &MI);
747 
748   /// Whether we must prevent this instruction from executing with EXEC = 0.
749   bool hasUnwantedEffectsWhenEXECEmpty(const MachineInstr &MI) const;
750 
751   /// Returns true if the instruction could potentially depend on the value of
752   /// exec. If false, exec dependencies may safely be ignored.
753   bool mayReadEXEC(const MachineRegisterInfo &MRI, const MachineInstr &MI) const;
754 
755   bool isInlineConstant(const APInt &Imm) const;
756 
isInlineConstant(const APFloat & Imm)757   bool isInlineConstant(const APFloat &Imm) const {
758     return isInlineConstant(Imm.bitcastToAPInt());
759   }
760 
761   bool isInlineConstant(const MachineOperand &MO, uint8_t OperandType) const;
762 
isInlineConstant(const MachineOperand & MO,const MCOperandInfo & OpInfo)763   bool isInlineConstant(const MachineOperand &MO,
764                         const MCOperandInfo &OpInfo) const {
765     return isInlineConstant(MO, OpInfo.OperandType);
766   }
767 
768   /// \p returns true if \p UseMO is substituted with \p DefMO in \p MI it would
769   /// be an inline immediate.
isInlineConstant(const MachineInstr & MI,const MachineOperand & UseMO,const MachineOperand & DefMO)770   bool isInlineConstant(const MachineInstr &MI,
771                         const MachineOperand &UseMO,
772                         const MachineOperand &DefMO) const {
773     assert(UseMO.getParent() == &MI);
774     int OpIdx = MI.getOperandNo(&UseMO);
775     if (!MI.getDesc().OpInfo || OpIdx >= MI.getDesc().NumOperands) {
776       return false;
777     }
778 
779     return isInlineConstant(DefMO, MI.getDesc().OpInfo[OpIdx]);
780   }
781 
782   /// \p returns true if the operand \p OpIdx in \p MI is a valid inline
783   /// immediate.
isInlineConstant(const MachineInstr & MI,unsigned OpIdx)784   bool isInlineConstant(const MachineInstr &MI, unsigned OpIdx) const {
785     const MachineOperand &MO = MI.getOperand(OpIdx);
786     return isInlineConstant(MO, MI.getDesc().OpInfo[OpIdx].OperandType);
787   }
788 
isInlineConstant(const MachineInstr & MI,unsigned OpIdx,const MachineOperand & MO)789   bool isInlineConstant(const MachineInstr &MI, unsigned OpIdx,
790                         const MachineOperand &MO) const {
791     if (!MI.getDesc().OpInfo || OpIdx >= MI.getDesc().NumOperands)
792       return false;
793 
794     if (MI.isCopy()) {
795       unsigned Size = getOpSize(MI, OpIdx);
796       assert(Size == 8 || Size == 4);
797 
798       uint8_t OpType = (Size == 8) ?
799         AMDGPU::OPERAND_REG_IMM_INT64 : AMDGPU::OPERAND_REG_IMM_INT32;
800       return isInlineConstant(MO, OpType);
801     }
802 
803     return isInlineConstant(MO, MI.getDesc().OpInfo[OpIdx].OperandType);
804   }
805 
isInlineConstant(const MachineOperand & MO)806   bool isInlineConstant(const MachineOperand &MO) const {
807     const MachineInstr *Parent = MO.getParent();
808     return isInlineConstant(*Parent, Parent->getOperandNo(&MO));
809   }
810 
isLiteralConstant(const MachineOperand & MO,const MCOperandInfo & OpInfo)811   bool isLiteralConstant(const MachineOperand &MO,
812                          const MCOperandInfo &OpInfo) const {
813     return MO.isImm() && !isInlineConstant(MO, OpInfo.OperandType);
814   }
815 
isLiteralConstant(const MachineInstr & MI,int OpIdx)816   bool isLiteralConstant(const MachineInstr &MI, int OpIdx) const {
817     const MachineOperand &MO = MI.getOperand(OpIdx);
818     return MO.isImm() && !isInlineConstant(MI, OpIdx);
819   }
820 
821   // Returns true if this operand could potentially require a 32-bit literal
822   // operand, but not necessarily. A FrameIndex for example could resolve to an
823   // inline immediate value that will not require an additional 4-bytes; this
824   // assumes that it will.
825   bool isLiteralConstantLike(const MachineOperand &MO,
826                              const MCOperandInfo &OpInfo) const;
827 
828   bool isImmOperandLegal(const MachineInstr &MI, unsigned OpNo,
829                          const MachineOperand &MO) const;
830 
831   /// Return true if this 64-bit VALU instruction has a 32-bit encoding.
832   /// This function will return false if you pass it a 32-bit instruction.
833   bool hasVALU32BitEncoding(unsigned Opcode) const;
834 
835   /// Returns true if this operand uses the constant bus.
836   bool usesConstantBus(const MachineRegisterInfo &MRI,
837                        const MachineOperand &MO,
838                        const MCOperandInfo &OpInfo) const;
839 
840   /// Return true if this instruction has any modifiers.
841   ///  e.g. src[012]_mod, omod, clamp.
842   bool hasModifiers(unsigned Opcode) const;
843 
844   bool hasModifiersSet(const MachineInstr &MI,
845                        unsigned OpName) const;
846   bool hasAnyModifiersSet(const MachineInstr &MI) const;
847 
848   bool canShrink(const MachineInstr &MI,
849                  const MachineRegisterInfo &MRI) const;
850 
851   MachineInstr *buildShrunkInst(MachineInstr &MI,
852                                 unsigned NewOpcode) const;
853 
854   bool verifyInstruction(const MachineInstr &MI,
855                          StringRef &ErrInfo) const override;
856 
857   unsigned getVALUOp(const MachineInstr &MI) const;
858 
859   /// Return the correct register class for \p OpNo.  For target-specific
860   /// instructions, this will return the register class that has been defined
861   /// in tablegen.  For generic instructions, like REG_SEQUENCE it will return
862   /// the register class of its machine operand.
863   /// to infer the correct register class base on the other operands.
864   const TargetRegisterClass *getOpRegClass(const MachineInstr &MI,
865                                            unsigned OpNo) const;
866 
867   /// Return the size in bytes of the operand OpNo on the given
868   // instruction opcode.
getOpSize(uint16_t Opcode,unsigned OpNo)869   unsigned getOpSize(uint16_t Opcode, unsigned OpNo) const {
870     const MCOperandInfo &OpInfo = get(Opcode).OpInfo[OpNo];
871 
872     if (OpInfo.RegClass == -1) {
873       // If this is an immediate operand, this must be a 32-bit literal.
874       assert(OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE);
875       return 4;
876     }
877 
878     return RI.getRegSizeInBits(*RI.getRegClass(OpInfo.RegClass)) / 8;
879   }
880 
881   /// This form should usually be preferred since it handles operands
882   /// with unknown register classes.
getOpSize(const MachineInstr & MI,unsigned OpNo)883   unsigned getOpSize(const MachineInstr &MI, unsigned OpNo) const {
884     const MachineOperand &MO = MI.getOperand(OpNo);
885     if (MO.isReg()) {
886       if (unsigned SubReg = MO.getSubReg()) {
887         return RI.getSubRegIdxSize(SubReg) / 8;
888       }
889     }
890     return RI.getRegSizeInBits(*getOpRegClass(MI, OpNo)) / 8;
891   }
892 
893   /// Legalize the \p OpIndex operand of this instruction by inserting
894   /// a MOV.  For example:
895   /// ADD_I32_e32 VGPR0, 15
896   /// to
897   /// MOV VGPR1, 15
898   /// ADD_I32_e32 VGPR0, VGPR1
899   ///
900   /// If the operand being legalized is a register, then a COPY will be used
901   /// instead of MOV.
902   void legalizeOpWithMove(MachineInstr &MI, unsigned OpIdx) const;
903 
904   /// Check if \p MO is a legal operand if it was the \p OpIdx Operand
905   /// for \p MI.
906   bool isOperandLegal(const MachineInstr &MI, unsigned OpIdx,
907                       const MachineOperand *MO = nullptr) const;
908 
909   /// Check if \p MO would be a valid operand for the given operand
910   /// definition \p OpInfo. Note this does not attempt to validate constant bus
911   /// restrictions (e.g. literal constant usage).
912   bool isLegalVSrcOperand(const MachineRegisterInfo &MRI,
913                           const MCOperandInfo &OpInfo,
914                           const MachineOperand &MO) const;
915 
916   /// Check if \p MO (a register operand) is a legal register for the
917   /// given operand description.
918   bool isLegalRegOperand(const MachineRegisterInfo &MRI,
919                          const MCOperandInfo &OpInfo,
920                          const MachineOperand &MO) const;
921 
922   /// Legalize operands in \p MI by either commuting it or inserting a
923   /// copy of src1.
924   void legalizeOperandsVOP2(MachineRegisterInfo &MRI, MachineInstr &MI) const;
925 
926   /// Fix operands in \p MI to satisfy constant bus requirements.
927   void legalizeOperandsVOP3(MachineRegisterInfo &MRI, MachineInstr &MI) const;
928 
929   /// Copy a value from a VGPR (\p SrcReg) to SGPR.  This function can only
930   /// be used when it is know that the value in SrcReg is same across all
931   /// threads in the wave.
932   /// \returns The SGPR register that \p SrcReg was copied to.
933   Register readlaneVGPRToSGPR(Register SrcReg, MachineInstr &UseMI,
934                               MachineRegisterInfo &MRI) const;
935 
936   void legalizeOperandsSMRD(MachineRegisterInfo &MRI, MachineInstr &MI) const;
937   void legalizeOperandsFLAT(MachineRegisterInfo &MRI, MachineInstr &MI) const;
938 
939   void legalizeGenericOperand(MachineBasicBlock &InsertMBB,
940                               MachineBasicBlock::iterator I,
941                               const TargetRegisterClass *DstRC,
942                               MachineOperand &Op, MachineRegisterInfo &MRI,
943                               const DebugLoc &DL) const;
944 
945   /// Legalize all operands in this instruction.  This function may create new
946   /// instructions and control-flow around \p MI.  If present, \p MDT is
947   /// updated.
948   /// \returns A new basic block that contains \p MI if new blocks were created.
949   MachineBasicBlock *
950   legalizeOperands(MachineInstr &MI, MachineDominatorTree *MDT = nullptr) const;
951 
952   /// Change SADDR form of a FLAT \p Inst to its VADDR form if saddr operand
953   /// was moved to VGPR. \returns true if succeeded.
954   bool moveFlatAddrToVGPR(MachineInstr &Inst) const;
955 
956   /// Replace this instruction's opcode with the equivalent VALU
957   /// opcode.  This function will also move the users of \p MI to the
958   /// VALU if necessary. If present, \p MDT is updated.
959   MachineBasicBlock *moveToVALU(MachineInstr &MI,
960                                 MachineDominatorTree *MDT = nullptr) const;
961 
962   void insertNoop(MachineBasicBlock &MBB,
963                   MachineBasicBlock::iterator MI) const override;
964 
965   void insertNoops(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
966                    unsigned Quantity) const override;
967 
968   void insertReturn(MachineBasicBlock &MBB) const;
969   /// Return the number of wait states that result from executing this
970   /// instruction.
971   static unsigned getNumWaitStates(const MachineInstr &MI);
972 
973   /// Returns the operand named \p Op.  If \p MI does not have an
974   /// operand named \c Op, this function returns nullptr.
975   LLVM_READONLY
976   MachineOperand *getNamedOperand(MachineInstr &MI, unsigned OperandName) const;
977 
978   LLVM_READONLY
getNamedOperand(const MachineInstr & MI,unsigned OpName)979   const MachineOperand *getNamedOperand(const MachineInstr &MI,
980                                         unsigned OpName) const {
981     return getNamedOperand(const_cast<MachineInstr &>(MI), OpName);
982   }
983 
984   /// Get required immediate operand
getNamedImmOperand(const MachineInstr & MI,unsigned OpName)985   int64_t getNamedImmOperand(const MachineInstr &MI, unsigned OpName) const {
986     int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OpName);
987     return MI.getOperand(Idx).getImm();
988   }
989 
990   uint64_t getDefaultRsrcDataFormat() const;
991   uint64_t getScratchRsrcWords23() const;
992 
993   bool isLowLatencyInstruction(const MachineInstr &MI) const;
994   bool isHighLatencyDef(int Opc) const override;
995 
996   /// Return the descriptor of the target-specific machine instruction
997   /// that corresponds to the specified pseudo or native opcode.
getMCOpcodeFromPseudo(unsigned Opcode)998   const MCInstrDesc &getMCOpcodeFromPseudo(unsigned Opcode) const {
999     return get(pseudoToMCOpcode(Opcode));
1000   }
1001 
1002   unsigned isStackAccess(const MachineInstr &MI, int &FrameIndex) const;
1003   unsigned isSGPRStackAccess(const MachineInstr &MI, int &FrameIndex) const;
1004 
1005   unsigned isLoadFromStackSlot(const MachineInstr &MI,
1006                                int &FrameIndex) const override;
1007   unsigned isStoreToStackSlot(const MachineInstr &MI,
1008                               int &FrameIndex) const override;
1009 
1010   unsigned getInstBundleSize(const MachineInstr &MI) const;
1011   unsigned getInstSizeInBytes(const MachineInstr &MI) const override;
1012 
1013   bool mayAccessFlatAddressSpace(const MachineInstr &MI) const;
1014 
1015   bool isNonUniformBranchInstr(MachineInstr &Instr) const;
1016 
1017   void convertNonUniformIfRegion(MachineBasicBlock *IfEntry,
1018                                  MachineBasicBlock *IfEnd) const;
1019 
1020   void convertNonUniformLoopRegion(MachineBasicBlock *LoopEntry,
1021                                    MachineBasicBlock *LoopEnd) const;
1022 
1023   std::pair<unsigned, unsigned>
1024   decomposeMachineOperandsTargetFlags(unsigned TF) const override;
1025 
1026   ArrayRef<std::pair<int, const char *>>
1027   getSerializableTargetIndices() const override;
1028 
1029   ArrayRef<std::pair<unsigned, const char *>>
1030   getSerializableDirectMachineOperandTargetFlags() const override;
1031 
1032   ScheduleHazardRecognizer *
1033   CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
1034                                  const ScheduleDAG *DAG) const override;
1035 
1036   ScheduleHazardRecognizer *
1037   CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const override;
1038 
1039   bool isBasicBlockPrologue(const MachineInstr &MI) const override;
1040 
1041   MachineInstr *createPHIDestinationCopy(MachineBasicBlock &MBB,
1042                                          MachineBasicBlock::iterator InsPt,
1043                                          const DebugLoc &DL, Register Src,
1044                                          Register Dst) const override;
1045 
1046   MachineInstr *createPHISourceCopy(MachineBasicBlock &MBB,
1047                                     MachineBasicBlock::iterator InsPt,
1048                                     const DebugLoc &DL, Register Src,
1049                                     unsigned SrcSubReg,
1050                                     Register Dst) const override;
1051 
1052   bool isWave32() const;
1053 
1054   /// Return a partially built integer add instruction without carry.
1055   /// Caller must add source operands.
1056   /// For pre-GFX9 it will generate unused carry destination operand.
1057   /// TODO: After GFX9 it should return a no-carry operation.
1058   MachineInstrBuilder getAddNoCarry(MachineBasicBlock &MBB,
1059                                     MachineBasicBlock::iterator I,
1060                                     const DebugLoc &DL,
1061                                     Register DestReg) const;
1062 
1063   MachineInstrBuilder getAddNoCarry(MachineBasicBlock &MBB,
1064                                     MachineBasicBlock::iterator I,
1065                                     const DebugLoc &DL,
1066                                     Register DestReg,
1067                                     RegScavenger &RS) const;
1068 
1069   static bool isKillTerminator(unsigned Opcode);
1070   const MCInstrDesc &getKillTerminatorFromPseudo(unsigned Opcode) const;
1071 
isLegalMUBUFImmOffset(unsigned Imm)1072   static bool isLegalMUBUFImmOffset(unsigned Imm) {
1073     return isUInt<12>(Imm);
1074   }
1075 
1076   /// Returns if \p Offset is legal for the subtarget as the offset to a FLAT
1077   /// encoded instruction. If \p Signed, this is for an instruction that
1078   /// interprets the offset as signed.
1079   bool isLegalFLATOffset(int64_t Offset, unsigned AddrSpace,
1080                          uint64_t FlatVariant) const;
1081 
1082   /// Split \p COffsetVal into {immediate offset field, remainder offset}
1083   /// values.
1084   std::pair<int64_t, int64_t> splitFlatOffset(int64_t COffsetVal,
1085                                               unsigned AddrSpace,
1086                                               uint64_t FlatVariant) const;
1087 
1088   /// \brief Return a target-specific opcode if Opcode is a pseudo instruction.
1089   /// Return -1 if the target-specific opcode for the pseudo instruction does
1090   /// not exist. If Opcode is not a pseudo instruction, this is identity.
1091   int pseudoToMCOpcode(int Opcode) const;
1092 
1093   /// \brief Check if this instruction should only be used by assembler.
1094   /// Return true if this opcode should not be used by codegen.
1095   bool isAsmOnlyOpcode(int MCOp) const;
1096 
1097   const TargetRegisterClass *getRegClass(const MCInstrDesc &TID, unsigned OpNum,
1098                                          const TargetRegisterInfo *TRI,
1099                                          const MachineFunction &MF)
1100     const override;
1101 
1102   void fixImplicitOperands(MachineInstr &MI) const;
1103 
1104   MachineInstr *foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI,
1105                                       ArrayRef<unsigned> Ops,
1106                                       MachineBasicBlock::iterator InsertPt,
1107                                       int FrameIndex,
1108                                       LiveIntervals *LIS = nullptr,
1109                                       VirtRegMap *VRM = nullptr) const override;
1110 
1111   unsigned getInstrLatency(const InstrItineraryData *ItinData,
1112                            const MachineInstr &MI,
1113                            unsigned *PredCost = nullptr) const override;
1114 
getMIRFormatter()1115   const MIRFormatter *getMIRFormatter() const override {
1116     if (!Formatter.get())
1117       Formatter = std::make_unique<AMDGPUMIRFormatter>();
1118     return Formatter.get();
1119   }
1120 
1121   static unsigned getDSShaderTypeValue(const MachineFunction &MF);
1122 };
1123 
1124 /// \brief Returns true if a reg:subreg pair P has a TRC class
isOfRegClass(const TargetInstrInfo::RegSubRegPair & P,const TargetRegisterClass & TRC,MachineRegisterInfo & MRI)1125 inline bool isOfRegClass(const TargetInstrInfo::RegSubRegPair &P,
1126                          const TargetRegisterClass &TRC,
1127                          MachineRegisterInfo &MRI) {
1128   auto *RC = MRI.getRegClass(P.Reg);
1129   if (!P.SubReg)
1130     return RC == &TRC;
1131   auto *TRI = MRI.getTargetRegisterInfo();
1132   return RC == TRI->getMatchingSuperRegClass(RC, &TRC, P.SubReg);
1133 }
1134 
1135 /// \brief Create RegSubRegPair from a register MachineOperand
1136 inline
getRegSubRegPair(const MachineOperand & O)1137 TargetInstrInfo::RegSubRegPair getRegSubRegPair(const MachineOperand &O) {
1138   assert(O.isReg());
1139   return TargetInstrInfo::RegSubRegPair(O.getReg(), O.getSubReg());
1140 }
1141 
1142 /// \brief Return the SubReg component from REG_SEQUENCE
1143 TargetInstrInfo::RegSubRegPair getRegSequenceSubReg(MachineInstr &MI,
1144                                                     unsigned SubReg);
1145 
1146 /// \brief Return the defining instruction for a given reg:subreg pair
1147 /// skipping copy like instructions and subreg-manipulation pseudos.
1148 /// Following another subreg of a reg:subreg isn't supported.
1149 MachineInstr *getVRegSubRegDef(const TargetInstrInfo::RegSubRegPair &P,
1150                                MachineRegisterInfo &MRI);
1151 
1152 /// \brief Return false if EXEC is not changed between the def of \p VReg at \p
1153 /// DefMI and the use at \p UseMI. Should be run on SSA. Currently does not
1154 /// attempt to track between blocks.
1155 bool execMayBeModifiedBeforeUse(const MachineRegisterInfo &MRI,
1156                                 Register VReg,
1157                                 const MachineInstr &DefMI,
1158                                 const MachineInstr &UseMI);
1159 
1160 /// \brief Return false if EXEC is not changed between the def of \p VReg at \p
1161 /// DefMI and all its uses. Should be run on SSA. Currently does not attempt to
1162 /// track between blocks.
1163 bool execMayBeModifiedBeforeAnyUse(const MachineRegisterInfo &MRI,
1164                                    Register VReg,
1165                                    const MachineInstr &DefMI);
1166 
1167 namespace AMDGPU {
1168 
1169   LLVM_READONLY
1170   int getVOPe64(uint16_t Opcode);
1171 
1172   LLVM_READONLY
1173   int getVOPe32(uint16_t Opcode);
1174 
1175   LLVM_READONLY
1176   int getSDWAOp(uint16_t Opcode);
1177 
1178   LLVM_READONLY
1179   int getDPPOp32(uint16_t Opcode);
1180 
1181   LLVM_READONLY
1182   int getBasicFromSDWAOp(uint16_t Opcode);
1183 
1184   LLVM_READONLY
1185   int getCommuteRev(uint16_t Opcode);
1186 
1187   LLVM_READONLY
1188   int getCommuteOrig(uint16_t Opcode);
1189 
1190   LLVM_READONLY
1191   int getAddr64Inst(uint16_t Opcode);
1192 
1193   /// Check if \p Opcode is an Addr64 opcode.
1194   ///
1195   /// \returns \p Opcode if it is an Addr64 opcode, otherwise -1.
1196   LLVM_READONLY
1197   int getIfAddr64Inst(uint16_t Opcode);
1198 
1199   LLVM_READONLY
1200   int getMUBUFNoLdsInst(uint16_t Opcode);
1201 
1202   LLVM_READONLY
1203   int getAtomicNoRetOp(uint16_t Opcode);
1204 
1205   LLVM_READONLY
1206   int getSOPKOp(uint16_t Opcode);
1207 
1208   /// \returns SADDR form of a FLAT Global instruction given an \p Opcode
1209   /// of a VADDR form.
1210   LLVM_READONLY
1211   int getGlobalSaddrOp(uint16_t Opcode);
1212 
1213   /// \returns VADDR form of a FLAT Global instruction given an \p Opcode
1214   /// of a SADDR form.
1215   LLVM_READONLY
1216   int getGlobalVaddrOp(uint16_t Opcode);
1217 
1218   LLVM_READONLY
1219   int getVCMPXNoSDstOp(uint16_t Opcode);
1220 
1221   /// \returns ST form with only immediate offset of a FLAT Scratch instruction
1222   /// given an \p Opcode of an SS (SADDR) form.
1223   LLVM_READONLY
1224   int getFlatScratchInstSTfromSS(uint16_t Opcode);
1225 
1226   /// \returns SS (SADDR) form of a FLAT Scratch instruction given an \p Opcode
1227   /// of an SV (VADDR) form.
1228   LLVM_READONLY
1229   int getFlatScratchInstSSfromSV(uint16_t Opcode);
1230 
1231   /// \returns SV (VADDR) form of a FLAT Scratch instruction given an \p Opcode
1232   /// of an SS (SADDR) form.
1233   LLVM_READONLY
1234   int getFlatScratchInstSVfromSS(uint16_t Opcode);
1235 
1236   const uint64_t RSRC_DATA_FORMAT = 0xf00000000000LL;
1237   const uint64_t RSRC_ELEMENT_SIZE_SHIFT = (32 + 19);
1238   const uint64_t RSRC_INDEX_STRIDE_SHIFT = (32 + 21);
1239   const uint64_t RSRC_TID_ENABLE = UINT64_C(1) << (32 + 23);
1240 
1241 } // end namespace AMDGPU
1242 
1243 namespace SI {
1244 namespace KernelInputOffsets {
1245 
1246 /// Offsets in bytes from the start of the input buffer
1247 enum Offsets {
1248   NGROUPS_X = 0,
1249   NGROUPS_Y = 4,
1250   NGROUPS_Z = 8,
1251   GLOBAL_SIZE_X = 12,
1252   GLOBAL_SIZE_Y = 16,
1253   GLOBAL_SIZE_Z = 20,
1254   LOCAL_SIZE_X = 24,
1255   LOCAL_SIZE_Y = 28,
1256   LOCAL_SIZE_Z = 32
1257 };
1258 
1259 } // end namespace KernelInputOffsets
1260 } // end namespace SI
1261 
1262 } // end namespace llvm
1263 
1264 #endif // LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H
1265