Home | History | Annotate | Download | only in SystemZ
      1 //===-- SystemZISelLowering.h - SystemZ DAG lowering interface --*- C++ -*-===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 // This file defines the interfaces that SystemZ uses to lower LLVM code into a
     11 // selection DAG.
     12 //
     13 //===----------------------------------------------------------------------===//
     14 
     15 #ifndef LLVM_TARGET_SystemZ_ISELLOWERING_H
     16 #define LLVM_TARGET_SystemZ_ISELLOWERING_H
     17 
     18 #include "SystemZ.h"
     19 #include "llvm/CodeGen/MachineBasicBlock.h"
     20 #include "llvm/CodeGen/SelectionDAG.h"
     21 #include "llvm/Target/TargetLowering.h"
     22 
     23 namespace llvm {
     24 namespace SystemZISD {
     25 enum {
     26   FIRST_NUMBER = ISD::BUILTIN_OP_END,
     27 
     28   // Return with a flag operand.  Operand 0 is the chain operand.
     29   RET_FLAG,
     30 
     31   // Calls a function.  Operand 0 is the chain operand and operand 1
     32   // is the target address.  The arguments start at operand 2.
     33   // There is an optional glue operand at the end.
     34   CALL,
     35   SIBCALL,
     36 
     37   // Wraps a TargetGlobalAddress that should be loaded using PC-relative
     38   // accesses (LARL).  Operand 0 is the address.
     39   PCREL_WRAPPER,
     40 
     41   // Used in cases where an offset is applied to a TargetGlobalAddress.
     42   // Operand 0 is the full TargetGlobalAddress and operand 1 is a
     43   // PCREL_WRAPPER for an anchor point.  This is used so that we can
     44   // cheaply refer to either the full address or the anchor point
     45   // as a register base.
     46   PCREL_OFFSET,
     47 
     48   // Integer absolute.
     49   IABS,
     50 
     51   // Integer comparisons.  There are three operands: the two values
     52   // to compare, and an integer of type SystemZICMP.
     53   ICMP,
     54 
     55   // Floating-point comparisons.  The two operands are the values to compare.
     56   FCMP,
     57 
     58   // Test under mask.  The first operand is ANDed with the second operand
     59   // and the condition codes are set on the result.  The third operand is
     60   // a boolean that is true if the condition codes need to distinguish
     61   // between CCMASK_TM_MIXED_MSB_0 and CCMASK_TM_MIXED_MSB_1 (which the
     62   // register forms do but the memory forms don't).
     63   TM,
     64 
     65   // Branches if a condition is true.  Operand 0 is the chain operand;
     66   // operand 1 is the 4-bit condition-code mask, with bit N in
     67   // big-endian order meaning "branch if CC=N"; operand 2 is the
     68   // target block and operand 3 is the flag operand.
     69   BR_CCMASK,
     70 
     71   // Selects between operand 0 and operand 1.  Operand 2 is the
     72   // mask of condition-code values for which operand 0 should be
     73   // chosen over operand 1; it has the same form as BR_CCMASK.
     74   // Operand 3 is the flag operand.
     75   SELECT_CCMASK,
     76 
     77   // Evaluates to the gap between the stack pointer and the
     78   // base of the dynamically-allocatable area.
     79   ADJDYNALLOC,
     80 
     81   // Extracts the value of a 32-bit access register.  Operand 0 is
     82   // the number of the register.
     83   EXTRACT_ACCESS,
     84 
     85   // Wrappers around the ISD opcodes of the same name.  The output and
     86   // first input operands are GR128s.  The trailing numbers are the
     87   // widths of the second operand in bits.
     88   UMUL_LOHI64,
     89   SDIVREM32,
     90   SDIVREM64,
     91   UDIVREM32,
     92   UDIVREM64,
     93 
     94   // Use a series of MVCs to copy bytes from one memory location to another.
     95   // The operands are:
     96   // - the target address
     97   // - the source address
     98   // - the constant length
     99   //
    100   // This isn't a memory opcode because we'd need to attach two
    101   // MachineMemOperands rather than one.
    102   MVC,
    103 
    104   // Like MVC, but implemented as a loop that handles X*256 bytes
    105   // followed by straight-line code to handle the rest (if any).
    106   // The value of X is passed as an additional operand.
    107   MVC_LOOP,
    108 
    109   // Similar to MVC and MVC_LOOP, but for logic operations (AND, OR, XOR).
    110   NC,
    111   NC_LOOP,
    112   OC,
    113   OC_LOOP,
    114   XC,
    115   XC_LOOP,
    116 
    117   // Use CLC to compare two blocks of memory, with the same comments
    118   // as for MVC and MVC_LOOP.
    119   CLC,
    120   CLC_LOOP,
    121 
    122   // Use an MVST-based sequence to implement stpcpy().
    123   STPCPY,
    124 
    125   // Use a CLST-based sequence to implement strcmp().  The two input operands
    126   // are the addresses of the strings to compare.
    127   STRCMP,
    128 
    129   // Use an SRST-based sequence to search a block of memory.  The first
    130   // operand is the end address, the second is the start, and the third
    131   // is the character to search for.  CC is set to 1 on success and 2
    132   // on failure.
    133   SEARCH_STRING,
    134 
    135   // Store the CC value in bits 29 and 28 of an integer.
    136   IPM,
    137 
    138   // Perform a serialization operation.  (BCR 15,0 or BCR 14,0.)
    139   SERIALIZE,
    140 
    141   // Wrappers around the inner loop of an 8- or 16-bit ATOMIC_SWAP or
    142   // ATOMIC_LOAD_<op>.
    143   //
    144   // Operand 0: the address of the containing 32-bit-aligned field
    145   // Operand 1: the second operand of <op>, in the high bits of an i32
    146   //            for everything except ATOMIC_SWAPW
    147   // Operand 2: how many bits to rotate the i32 left to bring the first
    148   //            operand into the high bits
    149   // Operand 3: the negative of operand 2, for rotating the other way
    150   // Operand 4: the width of the field in bits (8 or 16)
    151   ATOMIC_SWAPW = ISD::FIRST_TARGET_MEMORY_OPCODE,
    152   ATOMIC_LOADW_ADD,
    153   ATOMIC_LOADW_SUB,
    154   ATOMIC_LOADW_AND,
    155   ATOMIC_LOADW_OR,
    156   ATOMIC_LOADW_XOR,
    157   ATOMIC_LOADW_NAND,
    158   ATOMIC_LOADW_MIN,
    159   ATOMIC_LOADW_MAX,
    160   ATOMIC_LOADW_UMIN,
    161   ATOMIC_LOADW_UMAX,
    162 
    163   // A wrapper around the inner loop of an ATOMIC_CMP_SWAP.
    164   //
    165   // Operand 0: the address of the containing 32-bit-aligned field
    166   // Operand 1: the compare value, in the low bits of an i32
    167   // Operand 2: the swap value, in the low bits of an i32
    168   // Operand 3: how many bits to rotate the i32 left to bring the first
    169   //            operand into the high bits
    170   // Operand 4: the negative of operand 2, for rotating the other way
    171   // Operand 5: the width of the field in bits (8 or 16)
    172   ATOMIC_CMP_SWAPW,
    173 
    174   // Prefetch from the second operand using the 4-bit control code in
    175   // the first operand.  The code is 1 for a load prefetch and 2 for
    176   // a store prefetch.
    177   PREFETCH
    178 };
    179 
    180 // Return true if OPCODE is some kind of PC-relative address.
    181 inline bool isPCREL(unsigned Opcode) {
    182   return Opcode == PCREL_WRAPPER || Opcode == PCREL_OFFSET;
    183 }
    184 } // end namespace SystemZISD
    185 
    186 namespace SystemZICMP {
    187 // Describes whether an integer comparison needs to be signed or unsigned,
    188 // or whether either type is OK.
    189 enum {
    190   Any,
    191   UnsignedOnly,
    192   SignedOnly
    193 };
    194 } // end namespace SystemZICMP
    195 
    196 class SystemZSubtarget;
    197 class SystemZTargetMachine;
    198 
    199 class SystemZTargetLowering : public TargetLowering {
    200 public:
    201   explicit SystemZTargetLowering(const TargetMachine &TM);
    202 
    203   // Override TargetLowering.
    204   MVT getScalarShiftAmountTy(EVT LHSTy) const override {
    205     return MVT::i32;
    206   }
    207   EVT getSetCCResultType(LLVMContext &, EVT) const override;
    208   bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
    209   bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
    210   bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const override;
    211   bool allowsUnalignedMemoryAccesses(EVT VT, unsigned AS,
    212                                      bool *Fast) const override;
    213   bool isTruncateFree(Type *, Type *) const override;
    214   bool isTruncateFree(EVT, EVT) const override;
    215   const char *getTargetNodeName(unsigned Opcode) const override;
    216   std::pair<unsigned, const TargetRegisterClass *>
    217     getRegForInlineAsmConstraint(const std::string &Constraint,
    218                                  MVT VT) const override;
    219   TargetLowering::ConstraintType
    220     getConstraintType(const std::string &Constraint) const override;
    221   TargetLowering::ConstraintWeight
    222     getSingleConstraintMatchWeight(AsmOperandInfo &info,
    223                                    const char *constraint) const override;
    224   void LowerAsmOperandForConstraint(SDValue Op,
    225                                     std::string &Constraint,
    226                                     std::vector<SDValue> &Ops,
    227                                     SelectionDAG &DAG) const override;
    228   MachineBasicBlock *EmitInstrWithCustomInserter(MachineInstr *MI,
    229                                                  MachineBasicBlock *BB) const
    230     override;
    231   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
    232   bool allowTruncateForTailCall(Type *, Type *) const override;
    233   bool mayBeEmittedAsTailCall(CallInst *CI) const override;
    234   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
    235                                bool isVarArg,
    236                                const SmallVectorImpl<ISD::InputArg> &Ins,
    237                                SDLoc DL, SelectionDAG &DAG,
    238                                SmallVectorImpl<SDValue> &InVals) const override;
    239   SDValue LowerCall(CallLoweringInfo &CLI,
    240                     SmallVectorImpl<SDValue> &InVals) const override;
    241 
    242   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
    243                       const SmallVectorImpl<ISD::OutputArg> &Outs,
    244                       const SmallVectorImpl<SDValue> &OutVals,
    245                       SDLoc DL, SelectionDAG &DAG) const override;
    246   SDValue prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL,
    247                                       SelectionDAG &DAG) const override;
    248   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
    249 
    250 private:
    251   const SystemZSubtarget &Subtarget;
    252 
    253   // Implement LowerOperation for individual opcodes.
    254   SDValue lowerSETCC(SDValue Op, SelectionDAG &DAG) const;
    255   SDValue lowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
    256   SDValue lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
    257   SDValue lowerGlobalAddress(GlobalAddressSDNode *Node,
    258                              SelectionDAG &DAG) const;
    259   SDValue lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
    260                                 SelectionDAG &DAG) const;
    261   SDValue lowerBlockAddress(BlockAddressSDNode *Node,
    262                             SelectionDAG &DAG) const;
    263   SDValue lowerJumpTable(JumpTableSDNode *JT, SelectionDAG &DAG) const;
    264   SDValue lowerConstantPool(ConstantPoolSDNode *CP, SelectionDAG &DAG) const;
    265   SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const;
    266   SDValue lowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
    267   SDValue lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
    268   SDValue lowerSMUL_LOHI(SDValue Op, SelectionDAG &DAG) const;
    269   SDValue lowerUMUL_LOHI(SDValue Op, SelectionDAG &DAG) const;
    270   SDValue lowerSDIVREM(SDValue Op, SelectionDAG &DAG) const;
    271   SDValue lowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
    272   SDValue lowerBITCAST(SDValue Op, SelectionDAG &DAG) const;
    273   SDValue lowerOR(SDValue Op, SelectionDAG &DAG) const;
    274   SDValue lowerATOMIC_LOAD(SDValue Op, SelectionDAG &DAG) const;
    275   SDValue lowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) const;
    276   SDValue lowerATOMIC_LOAD_OP(SDValue Op, SelectionDAG &DAG,
    277                               unsigned Opcode) const;
    278   SDValue lowerATOMIC_LOAD_SUB(SDValue Op, SelectionDAG &DAG) const;
    279   SDValue lowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const;
    280   SDValue lowerLOAD_SEQUENCE_POINT(SDValue Op, SelectionDAG &DAG) const;
    281   SDValue lowerSTACKSAVE(SDValue Op, SelectionDAG &DAG) const;
    282   SDValue lowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const;
    283   SDValue lowerPREFETCH(SDValue Op, SelectionDAG &DAG) const;
    284 
    285   // If the last instruction before MBBI in MBB was some form of COMPARE,
    286   // try to replace it with a COMPARE AND BRANCH just before MBBI.
    287   // CCMask and Target are the BRC-like operands for the branch.
    288   // Return true if the change was made.
    289   bool convertPrevCompareToBranch(MachineBasicBlock *MBB,
    290                                   MachineBasicBlock::iterator MBBI,
    291                                   unsigned CCMask,
    292                                   MachineBasicBlock *Target) const;
    293 
    294   // Implement EmitInstrWithCustomInserter for individual operation types.
    295   MachineBasicBlock *emitSelect(MachineInstr *MI,
    296                                 MachineBasicBlock *BB) const;
    297   MachineBasicBlock *emitCondStore(MachineInstr *MI,
    298                                    MachineBasicBlock *BB,
    299                                    unsigned StoreOpcode, unsigned STOCOpcode,
    300                                    bool Invert) const;
    301   MachineBasicBlock *emitExt128(MachineInstr *MI,
    302                                 MachineBasicBlock *MBB,
    303                                 bool ClearEven, unsigned SubReg) const;
    304   MachineBasicBlock *emitAtomicLoadBinary(MachineInstr *MI,
    305                                           MachineBasicBlock *BB,
    306                                           unsigned BinOpcode, unsigned BitSize,
    307                                           bool Invert = false) const;
    308   MachineBasicBlock *emitAtomicLoadMinMax(MachineInstr *MI,
    309                                           MachineBasicBlock *MBB,
    310                                           unsigned CompareOpcode,
    311                                           unsigned KeepOldMask,
    312                                           unsigned BitSize) const;
    313   MachineBasicBlock *emitAtomicCmpSwapW(MachineInstr *MI,
    314                                         MachineBasicBlock *BB) const;
    315   MachineBasicBlock *emitMemMemWrapper(MachineInstr *MI,
    316                                        MachineBasicBlock *BB,
    317                                        unsigned Opcode) const;
    318   MachineBasicBlock *emitStringWrapper(MachineInstr *MI,
    319                                        MachineBasicBlock *BB,
    320                                        unsigned Opcode) const;
    321 };
    322 } // end namespace llvm
    323 
    324 #endif // LLVM_TARGET_SystemZ_ISELLOWERING_H
    325