AMDGPU: Factor switch into separate function
[oota-llvm.git] / lib / Target / AMDGPU / SIInstrInfo.h
index 7b58e0f5df21ace386ce2340b953a9ac92ce26e8..f4b2f2da2a01ff598c29660a89e72067ef390629 100644 (file)
@@ -1,4 +1,4 @@
-//===-- SIInstrInfo.h - SI Instruction Info Interface ---------------------===//
+//===-- SIInstrInfo.h - SI Instruction Info Interface -----------*- C++ -*-===//
 //
 //                     The LLVM Compiler Infrastructure
 //
@@ -7,15 +7,17 @@
 //
 //===----------------------------------------------------------------------===//
 //
-// Interface definition for SIInstrInfo.
+/// \file
+/// \brief Interface definition for SIInstrInfo.
 //
 //===----------------------------------------------------------------------===//
 
 
-#ifndef SIINSTRINFO_H
-#define SIINSTRINFO_H
+#ifndef LLVM_LIB_TARGET_R600_SIINSTRINFO_H
+#define LLVM_LIB_TARGET_R600_SIINSTRINFO_H
 
 #include "AMDGPUInstrInfo.h"
+#include "SIDefines.h"
 #include "SIRegisterInfo.h"
 
 namespace llvm {
@@ -24,66 +26,385 @@ class SIInstrInfo : public AMDGPUInstrInfo {
 private:
   const SIRegisterInfo RI;
 
+  unsigned buildExtractSubReg(MachineBasicBlock::iterator MI,
+                              MachineRegisterInfo &MRI,
+                              MachineOperand &SuperReg,
+                              const TargetRegisterClass *SuperRC,
+                              unsigned SubIdx,
+                              const TargetRegisterClass *SubRC) const;
+  MachineOperand buildExtractSubRegOrImm(MachineBasicBlock::iterator MI,
+                                         MachineRegisterInfo &MRI,
+                                         MachineOperand &SuperReg,
+                                         const TargetRegisterClass *SuperRC,
+                                         unsigned SubIdx,
+                                         const TargetRegisterClass *SubRC) const;
+
+  void swapOperands(MachineBasicBlock::iterator Inst) const;
+
+  void splitScalar64BitUnaryOp(SmallVectorImpl<MachineInstr *> &Worklist,
+                               MachineInstr *Inst, unsigned Opcode) const;
+
+  void splitScalar64BitBinaryOp(SmallVectorImpl<MachineInstr *> &Worklist,
+                                MachineInstr *Inst, unsigned Opcode) const;
+
+  void splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
+                            MachineInstr *Inst) const;
+  void splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
+                           MachineInstr *Inst) const;
+
+  void addUsersToMoveToVALUWorklist(
+    unsigned Reg, MachineRegisterInfo &MRI,
+    SmallVectorImpl<MachineInstr *> &Worklist) const;
+
+  const TargetRegisterClass *
+  getDestEquivalentVGPRClass(const MachineInstr &Inst) const;
+
+  bool checkInstOffsetsDoNotOverlap(MachineInstr *MIa,
+                                    MachineInstr *MIb) const;
+
+  unsigned findUsedSGPR(const MachineInstr *MI, int OpIndices[3]) const;
+
+protected:
+  MachineInstr *commuteInstructionImpl(MachineInstr *MI,
+                                       bool NewMI,
+                                       unsigned OpIdx0,
+                                       unsigned OpIdx1) const override;
+
 public:
-  explicit SIInstrInfo(AMDGPUTargetMachine &tm);
+  explicit SIInstrInfo(const AMDGPUSubtarget &st);
+
+  const SIRegisterInfo &getRegisterInfo() const override {
+    return RI;
+  }
+
+  bool isReallyTriviallyReMaterializable(const MachineInstr *MI,
+                                         AliasAnalysis *AA) const override;
+
+  bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
+                               int64_t &Offset1,
+                               int64_t &Offset2) const override;
+
+  bool getMemOpBaseRegImmOfs(MachineInstr *LdSt, unsigned &BaseReg,
+                             unsigned &Offset,
+                             const TargetRegisterInfo *TRI) const final;
+
+  bool shouldClusterLoads(MachineInstr *FirstLdSt,
+                          MachineInstr *SecondLdSt,
+                          unsigned NumLoads) const final;
+
+  void copyPhysReg(MachineBasicBlock &MBB,
+                   MachineBasicBlock::iterator MI, DebugLoc DL,
+                   unsigned DestReg, unsigned SrcReg,
+                   bool KillSrc) const override;
+
+  unsigned calculateLDSSpillAddress(MachineBasicBlock &MBB,
+                                    MachineBasicBlock::iterator MI,
+                                    RegScavenger *RS,
+                                    unsigned TmpReg,
+                                    unsigned Offset,
+                                    unsigned Size) const;
+
+  void storeRegToStackSlot(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator MI,
+                           unsigned SrcReg, bool isKill, int FrameIndex,
+                           const TargetRegisterClass *RC,
+                           const TargetRegisterInfo *TRI) const override;
+
+  void loadRegFromStackSlot(MachineBasicBlock &MBB,
+                            MachineBasicBlock::iterator MI,
+                            unsigned DestReg, int FrameIndex,
+                            const TargetRegisterClass *RC,
+                            const TargetRegisterInfo *TRI) const override;
+
+  bool expandPostRAPseudo(MachineBasicBlock::iterator MI) const override;
+
+  // \brief Returns an opcode that can be used to move a value to a \p DstRC
+  // register.  If there is no hardware instruction that can store to \p
+  // DstRC, then AMDGPU::COPY is returned.
+  unsigned getMovOpcode(const TargetRegisterClass *DstRC) const;
+
+  LLVM_READONLY
+  int commuteOpcode(const MachineInstr &MI) const;
+
+  bool findCommutedOpIndices(MachineInstr *MI,
+                             unsigned &SrcOpIdx1,
+                             unsigned &SrcOpIdx2) const override;
+
+  bool areMemAccessesTriviallyDisjoint(
+    MachineInstr *MIa, MachineInstr *MIb,
+    AliasAnalysis *AA = nullptr) const override;
+
+  MachineInstr *buildMovInstr(MachineBasicBlock *MBB,
+                              MachineBasicBlock::iterator I,
+                              unsigned DstReg, unsigned SrcReg) const override;
+  bool isMov(unsigned Opcode) const override;
+
+  bool FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
+                     unsigned Reg, MachineRegisterInfo *MRI) const final;
+
+  unsigned getMachineCSELookAheadLimit() const override { return 500; }
+
+  MachineInstr *convertToThreeAddress(MachineFunction::iterator &MBB,
+                                      MachineBasicBlock::iterator &MI,
+                                      LiveVariables *LV) const override;
+
+  bool isSALU(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::SALU;
+  }
+
+  bool isVALU(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::VALU;
+  }
+
+  bool isSOP1(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::SOP1;
+  }
+
+  bool isSOP2(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::SOP2;
+  }
+
+  bool isSOPC(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::SOPC;
+  }
+
+  bool isSOPK(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::SOPK;
+  }
+
+  bool isSOPP(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::SOPP;
+  }
+
+  bool isVOP1(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::VOP1;
+  }
+
+  bool isVOP2(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::VOP2;
+  }
+
+  bool isVOP3(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::VOP3;
+  }
+
+  bool isVOPC(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::VOPC;
+  }
+
+  bool isMUBUF(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::MUBUF;
+  }
+
+  bool isMTBUF(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::MTBUF;
+  }
+
+  bool isSMRD(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::SMRD;
+  }
 
-  const SIRegisterInfo &getRegisterInfo() const;
+  bool isDS(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::DS;
+  }
 
-  virtual void copyPhysReg(MachineBasicBlock &MBB,
-                           MachineBasicBlock::iterator MI, DebugLoc DL,
-                           unsigned DestReg, unsigned SrcReg,
-                           bool KillSrc) const;
+  bool isMIMG(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::MIMG;
+  }
 
-  /// getEncodingType - Returns the encoding type of this instruction.  
-  unsigned getEncodingType(const MachineInstr &MI) const;
+  bool isFLAT(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::FLAT;
+  }
 
-  /// getEncodingBytes - Returns the size of this instructions encoding in
-  /// number of bytes.
-  unsigned getEncodingBytes(const MachineInstr &MI) const;
+  bool isWQM(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::WQM;
+  }
 
-  virtual MachineInstr * getMovImmInstr(MachineFunction *MF, unsigned DstReg,
-                                        int64_t Imm) const;
+  bool isVGPRSpill(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::VGPRSpill;
+  }
 
-  virtual unsigned getIEQOpcode() const { assert(!"Implement"); return 0;}
-  virtual bool isMov(unsigned Opcode) const;
+  bool isInlineConstant(const APInt &Imm) const;
+  bool isInlineConstant(const MachineOperand &MO, unsigned OpSize) const;
+  bool isLiteralConstant(const MachineOperand &MO, unsigned OpSize) const;
 
-  };
+  bool isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
+                         const MachineOperand &MO) const;
+
+  /// \brief Return true if this 64-bit VALU instruction has a 32-bit encoding.
+  /// This function will return false if you pass it a 32-bit instruction.
+  bool hasVALU32BitEncoding(unsigned Opcode) const;
+
+  /// \brief Returns true if this operand uses the constant bus.
+  bool usesConstantBus(const MachineRegisterInfo &MRI,
+                       const MachineOperand &MO,
+                       unsigned OpSize) const;
+
+  /// \brief Return true if this instruction has any modifiers.
+  ///  e.g. src[012]_mod, omod, clamp.
+  bool hasModifiers(unsigned Opcode) const;
+
+  bool hasModifiersSet(const MachineInstr &MI,
+                       unsigned OpName) const;
+
+  bool verifyInstruction(const MachineInstr *MI,
+                         StringRef &ErrInfo) const override;
+
+  static unsigned getVALUOp(const MachineInstr &MI);
+
+  bool isSALUOpSupportedOnVALU(const MachineInstr &MI) const;
+
+  /// \brief Return the correct register class for \p OpNo.  For target-specific
+  /// instructions, this will return the register class that has been defined
+  /// in tablegen.  For generic instructions, like REG_SEQUENCE it will return
+  /// the register class of its machine operand.
+  /// to infer the correct register class base on the other operands.
+  const TargetRegisterClass *getOpRegClass(const MachineInstr &MI,
+                                           unsigned OpNo) const;
+
+  /// \brief Return the size in bytes of the operand OpNo on the given
+  // instruction opcode.
+  unsigned getOpSize(uint16_t Opcode, unsigned OpNo) const {
+    const MCOperandInfo &OpInfo = get(Opcode).OpInfo[OpNo];
+
+    if (OpInfo.RegClass == -1) {
+      // If this is an immediate operand, this must be a 32-bit literal.
+      assert(OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE);
+      return 4;
+    }
+
+    return RI.getRegClass(OpInfo.RegClass)->getSize();
+  }
+
+  /// \brief This form should usually be preferred since it handles operands
+  /// with unknown register classes.
+  unsigned getOpSize(const MachineInstr &MI, unsigned OpNo) const {
+    return getOpRegClass(MI, OpNo)->getSize();
+  }
+
+  /// \returns true if it is legal for the operand at index \p OpNo
+  /// to read a VGPR.
+  bool canReadVGPR(const MachineInstr &MI, unsigned OpNo) const;
+
+  /// \brief Legalize the \p OpIndex operand of this instruction by inserting
+  /// a MOV.  For example:
+  /// ADD_I32_e32 VGPR0, 15
+  /// to
+  /// MOV VGPR1, 15
+  /// ADD_I32_e32 VGPR0, VGPR1
+  ///
+  /// If the operand being legalized is a register, then a COPY will be used
+  /// instead of MOV.
+  void legalizeOpWithMove(MachineInstr *MI, unsigned OpIdx) const;
+
+  /// \brief Check if \p MO is a legal operand if it was the \p OpIdx Operand
+  /// for \p MI.
+  bool isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
+                      const MachineOperand *MO = nullptr) const;
+
+  /// \brief Legalize all operands in this instruction.  This function may
+  /// create new instruction and insert them before \p MI.
+  void legalizeOperands(MachineInstr *MI) const;
+
+  /// \brief Split an SMRD instruction into two smaller loads of half the
+  //  size storing the results in \p Lo and \p Hi.
+  void splitSMRD(MachineInstr *MI, const TargetRegisterClass *HalfRC,
+                 unsigned HalfImmOp, unsigned HalfSGPROp,
+                 MachineInstr *&Lo, MachineInstr *&Hi) const;
+
+  void moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI,
+                      SmallVectorImpl<MachineInstr *> &Worklist) const;
+
+  /// \brief Replace this instruction's opcode with the equivalent VALU
+  /// opcode.  This function will also move the users of \p MI to the
+  /// VALU if necessary.
+  void moveToVALU(MachineInstr &MI) const;
+
+  unsigned calculateIndirectAddress(unsigned RegIndex,
+                                    unsigned Channel) const override;
+
+  const TargetRegisterClass *getIndirectAddrRegClass() const override;
+
+  MachineInstrBuilder buildIndirectWrite(MachineBasicBlock *MBB,
+                                         MachineBasicBlock::iterator I,
+                                         unsigned ValueReg,
+                                         unsigned Address,
+                                         unsigned OffsetReg) const override;
+
+  MachineInstrBuilder buildIndirectRead(MachineBasicBlock *MBB,
+                                        MachineBasicBlock::iterator I,
+                                        unsigned ValueReg,
+                                        unsigned Address,
+                                        unsigned OffsetReg) const override;
+  void reserveIndirectRegisters(BitVector &Reserved,
+                                const MachineFunction &MF) const;
+
+  void LoadM0(MachineInstr *MoveRel, MachineBasicBlock::iterator I,
+              unsigned SavReg, unsigned IndexReg) const;
+
+  void insertNOPs(MachineBasicBlock::iterator MI, int Count) const;
+
+  /// \brief Returns the operand named \p Op.  If \p MI does not have an
+  /// operand named \c Op, this function returns nullptr.
+  LLVM_READONLY
+  MachineOperand *getNamedOperand(MachineInstr &MI, unsigned OperandName) const;
+
+  LLVM_READONLY
+  const MachineOperand *getNamedOperand(const MachineInstr &MI,
+                                        unsigned OpName) const {
+    return getNamedOperand(const_cast<MachineInstr &>(MI), OpName);
+  }
+
+  uint64_t getDefaultRsrcDataFormat() const;
+
+};
+
+namespace AMDGPU {
+  LLVM_READONLY
+  int getVOPe64(uint16_t Opcode);
+
+  LLVM_READONLY
+  int getVOPe32(uint16_t Opcode);
+
+  LLVM_READONLY
+  int getCommuteRev(uint16_t Opcode);
+
+  LLVM_READONLY
+  int getCommuteOrig(uint16_t Opcode);
+
+  LLVM_READONLY
+  int getAddr64Inst(uint16_t Opcode);
+
+  LLVM_READONLY
+  int getAtomicRetOp(uint16_t Opcode);
+
+  LLVM_READONLY
+  int getAtomicNoRetOp(uint16_t Opcode);
+
+  const uint64_t RSRC_DATA_FORMAT = 0xf00000000000LL;
+  const uint64_t RSRC_TID_ENABLE = 1LL << 55;
+
+} // End namespace AMDGPU
+
+namespace SI {
+namespace KernelInputOffsets {
+
+/// Offsets in bytes from the start of the input buffer
+enum Offsets {
+  NGROUPS_X = 0,
+  NGROUPS_Y = 4,
+  NGROUPS_Z = 8,
+  GLOBAL_SIZE_X = 12,
+  GLOBAL_SIZE_Y = 16,
+  GLOBAL_SIZE_Z = 20,
+  LOCAL_SIZE_X = 24,
+  LOCAL_SIZE_Y = 28,
+  LOCAL_SIZE_Z = 32
+};
+
+} // End namespace KernelInputOffsets
+} // End namespace SI
 
 } // End namespace llvm
 
-// These must be kept in sync with SIInstructions.td and also the
-// InstrEncodingInfo array in SIInstrInfo.cpp.
-//
-// NOTE: This enum is only used to identify the encoding type within LLVM,
-// the actual encoding type that is part of the instruction format is different
-namespace SIInstrEncodingType {
-  enum Encoding {
-    EXP = 0,
-    LDS = 1,
-    MIMG = 2,
-    MTBUF = 3,
-    MUBUF = 4,
-    SMRD = 5,
-    SOP1 = 6,
-    SOP2 = 7,
-    SOPC = 8,
-    SOPK = 9,
-    SOPP = 10,
-    VINTRP = 11,
-    VOP1 = 12,
-    VOP2 = 13,
-    VOP3 = 14,
-    VOPC = 15
-  };
-}
-
-#define SI_INSTR_FLAGS_ENCODING_MASK 0xf
-
-namespace SIInstrFlags {
-  enum Flags {
-    // First 4 bits are the instruction encoding
-    NEED_WAIT = 1 << 4
-  };
-}
-
-#endif //SIINSTRINFO_H
+#endif