PTX: Set PTX 2.0 as the minimum supported version
[oota-llvm.git] / lib / Target / ARM / ARMBaseInstrInfo.h
index c0b13b340b5cbbd87dd6462e88b44a07677fa776..228b6cd9141a3188e3c2d9110922427f1761e934 100644 (file)
@@ -17,6 +17,8 @@
 #include "ARM.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallSet.h"
 
 namespace llvm {
   class ARMSubtarget;
@@ -50,6 +52,7 @@ namespace ARMII {
     AddrModeT2_so   = 13,
     AddrModeT2_pc   = 14, // +/- i12 for pc relative data
     AddrModeT2_i8s4 = 15, // i8 * 4
+    AddrMode_i12    = 16,
 
     // Size* - Flags to keep track of the size of an instruction.
     SizeShift     = 5,
@@ -152,14 +155,20 @@ namespace ARMII {
     //===------------------------------------------------------------------===//
     // Code domain.
     DomainShift   = 18,
-    DomainMask    = 3 << DomainShift,
+    DomainMask    = 7 << DomainShift,
     DomainGeneral = 0 << DomainShift,
     DomainVFP     = 1 << DomainShift,
     DomainNEON    = 2 << DomainShift,
+    DomainNEONA8  = 4 << DomainShift,
 
     //===------------------------------------------------------------------===//
     // Field shifts - such shifts are used to set field while generating
     // machine instructions.
+    //
+    // FIXME: This list will need adjusting/fixing as the MC code emitter
+    // takes shape and the ARMCodeEmitter.cpp bits go away.
+    ShiftTypeShift = 4,
+
     M_BitShift     = 5,
     ShiftImmShift  = 5,
     ShiftShift     = 7,
@@ -185,9 +194,11 @@ namespace ARMII {
 
 class ARMBaseInstrInfo : public TargetInstrInfoImpl {
   const ARMSubtarget &Subtarget;
+
 protected:
   // Can be only subclassed.
   explicit ARMBaseInstrInfo(const ARMSubtarget &STI);
+
 public:
   // Return the non-pre/post incrementing version of 'Opc'. Return 0
   // if there is not such an opcode.
@@ -200,10 +211,13 @@ public:
   virtual const ARMBaseRegisterInfo &getRegisterInfo() const =0;
   const ARMSubtarget &getSubtarget() const { return Subtarget; }
 
-  bool spillCalleeSavedRegisters(MachineBasicBlock &MBB,
-                                 MachineBasicBlock::iterator MI,
-                                 const std::vector<CalleeSavedInfo> &CSI,
-                                 const TargetRegisterInfo *TRI) const;
+  ScheduleHazardRecognizer *
+  CreateTargetHazardRecognizer(const TargetMachine *TM,
+                               const ScheduleDAG *DAG) const;
+
+  ScheduleHazardRecognizer *
+  CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
+                                     const ScheduleDAG *DAG) const;
 
   // Branch analysis.
   virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
@@ -285,7 +299,8 @@ public:
   MachineInstr *duplicate(MachineInstr *Orig, MachineFunction &MF) const;
 
   virtual bool produceSameValue(const MachineInstr *MI0,
-                                const MachineInstr *MI1) const;
+                                const MachineInstr *MI1,
+                                const MachineRegisterInfo *MRI) const;
 
   /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
   /// determine if two loads are loading from the same base address. It should
@@ -312,18 +327,20 @@ public:
                                     const MachineFunction &MF) const;
 
   virtual bool isProfitableToIfCvt(MachineBasicBlock &MBB,
-                                   unsigned NumInstrs,
+                                   unsigned NumCyles, unsigned ExtraPredCycles,
                                    float Prob, float Confidence) const;
 
-  virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB,unsigned NumT,
-                                   MachineBasicBlock &FMBB,unsigned NumF,
+  virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB,
+                                   unsigned NumT, unsigned ExtraT,
+                                   MachineBasicBlock &FMBB,
+                                   unsigned NumF, unsigned ExtraF,
                                    float Probability, float Confidence) const;
 
   virtual bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
-                                         unsigned NumInstrs,
+                                         unsigned NumCyles,
                                          float Probability,
                                          float Confidence) const {
-    return NumInstrs && NumInstrs == 1;
+    return NumCyles == 1;
   }
 
   /// AnalyzeCompare - For a comparison instruction, return the source register
@@ -336,10 +353,91 @@ public:
   /// that we can remove a "comparison with zero".
   virtual bool OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg,
                                     int CmpMask, int CmpValue,
-                                    MachineBasicBlock::iterator &MII) const;
+                                    const MachineRegisterInfo *MRI) const;
+
+  /// FoldImmediate - 'Reg' is known to be defined by a move immediate
+  /// instruction, try to fold the immediate into the use instruction.
+  virtual bool FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
+                             unsigned Reg, MachineRegisterInfo *MRI) const;
+
+  virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData,
+                                  const MachineInstr *MI) const;
 
-  virtual unsigned getNumMicroOps(const MachineInstr *MI,
-                                  const InstrItineraryData *ItinData) const;
+  virtual
+  int getOperandLatency(const InstrItineraryData *ItinData,
+                        const MachineInstr *DefMI, unsigned DefIdx,
+                        const MachineInstr *UseMI, unsigned UseIdx) const;
+  virtual
+  int getOperandLatency(const InstrItineraryData *ItinData,
+                        SDNode *DefNode, unsigned DefIdx,
+                        SDNode *UseNode, unsigned UseIdx) const;
+private:
+  int getVLDMDefCycle(const InstrItineraryData *ItinData,
+                      const TargetInstrDesc &DefTID,
+                      unsigned DefClass,
+                      unsigned DefIdx, unsigned DefAlign) const;
+  int getLDMDefCycle(const InstrItineraryData *ItinData,
+                     const TargetInstrDesc &DefTID,
+                     unsigned DefClass,
+                     unsigned DefIdx, unsigned DefAlign) const;
+  int getVSTMUseCycle(const InstrItineraryData *ItinData,
+                      const TargetInstrDesc &UseTID,
+                      unsigned UseClass,
+                      unsigned UseIdx, unsigned UseAlign) const;
+  int getSTMUseCycle(const InstrItineraryData *ItinData,
+                     const TargetInstrDesc &UseTID,
+                     unsigned UseClass,
+                     unsigned UseIdx, unsigned UseAlign) const;
+  int getOperandLatency(const InstrItineraryData *ItinData,
+                        const TargetInstrDesc &DefTID,
+                        unsigned DefIdx, unsigned DefAlign,
+                        const TargetInstrDesc &UseTID,
+                        unsigned UseIdx, unsigned UseAlign) const;
+
+  int getInstrLatency(const InstrItineraryData *ItinData,
+                      const MachineInstr *MI, unsigned *PredCost = 0) const;
+
+  int getInstrLatency(const InstrItineraryData *ItinData,
+                      SDNode *Node) const;
+
+  bool hasHighOperandLatency(const InstrItineraryData *ItinData,
+                             const MachineRegisterInfo *MRI,
+                             const MachineInstr *DefMI, unsigned DefIdx,
+                             const MachineInstr *UseMI, unsigned UseIdx) const;
+  bool hasLowDefLatency(const InstrItineraryData *ItinData,
+                        const MachineInstr *DefMI, unsigned DefIdx) const;
+
+private:
+  /// Modeling special VFP / NEON fp MLA / MLS hazards.
+
+  /// MLxEntryMap - Map fp MLA / MLS to the corresponding entry in the internal
+  /// MLx table.
+  DenseMap<unsigned, unsigned> MLxEntryMap;
+
+  /// MLxHazardOpcodes - Set of add / sub and multiply opcodes that would cause
+  /// stalls when scheduled together with fp MLA / MLS opcodes.
+  SmallSet<unsigned, 16> MLxHazardOpcodes;
+
+public:
+  /// isFpMLxInstruction - Return true if the specified opcode is a fp MLA / MLS
+  /// instruction.
+  bool isFpMLxInstruction(unsigned Opcode) const {
+    return MLxEntryMap.count(Opcode);
+  }
+
+  /// isFpMLxInstruction - This version also returns the multiply opcode and the
+  /// addition / subtraction opcode to expand to. Return true for 'HasLane' for
+  /// the MLX instructions with an extra lane operand.
+  bool isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
+                          unsigned &AddSubOpc, bool &NegAcc,
+                          bool &HasLane) const;
+
+  /// canCauseFpMLxStall - Return true if an instruction of the specified opcode
+  /// will cause stalls when scheduled after (within 4-cycle window) a fp
+  /// MLA / MLS instruction.
+  bool canCauseFpMLxStall(unsigned Opcode) const {
+    return MLxHazardOpcodes.count(Opcode);
+  }
 };
 
 static inline
@@ -381,7 +479,7 @@ bool isJumpTableBranchOpcode(int Opc) {
 
 static inline
 bool isIndirectBranchOpcode(int Opc) {
-  return Opc == ARM::BRIND || Opc == ARM::MOVPCRX || Opc == ARM::tBRIND;
+  return Opc == ARM::BX || Opc == ARM::MOVPCRX || Opc == ARM::tBRIND;
 }
 
 /// getInstrPredicate - If instruction is predicated, returns its predicate
@@ -398,13 +496,19 @@ void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
                              MachineBasicBlock::iterator &MBBI, DebugLoc dl,
                              unsigned DestReg, unsigned BaseReg, int NumBytes,
                              ARMCC::CondCodes Pred, unsigned PredReg,
-                             const ARMBaseInstrInfo &TII);
+                             const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
 
 void emitT2RegPlusImmediate(MachineBasicBlock &MBB,
                             MachineBasicBlock::iterator &MBBI, DebugLoc dl,
                             unsigned DestReg, unsigned BaseReg, int NumBytes,
                             ARMCC::CondCodes Pred, unsigned PredReg,
-                            const ARMBaseInstrInfo &TII);
+                            const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
+void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
+                               MachineBasicBlock::iterator &MBBI, DebugLoc dl,
+                               unsigned DestReg, unsigned BaseReg,
+                               int NumBytes, const TargetInstrInfo &TII,
+                               const ARMBaseRegisterInfo& MRI,
+                               unsigned MIFlags = 0);
 
 
 /// rewriteARMFrameIndex / rewriteT2FrameIndex -