PPCISelLowering.h revision 61e729e2e9517ab2d8887bab86fb377900fa1081
130d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun//===-- PPCISelLowering.h - PPC32 DAG Lowering Interface --------*- C++ -*-===//
230d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun//
330d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun//                     The LLVM Compiler Infrastructure
430d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun//
530d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun// This file was developed by Chris Lattner and is distributed under
630d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun// the University of Illinois Open Source License. See LICENSE.TXT for details.
730d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun//
830d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun//===----------------------------------------------------------------------===//
930d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun//
1030d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun// This file defines the interfaces that PPC uses to lower LLVM code into a
1130d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun// selection DAG.
1230d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun//
1330d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun//===----------------------------------------------------------------------===//
1430d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
1530d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun#ifndef LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H
1630d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun#define LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H
1730d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
1830d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun#include "llvm/Target/TargetLowering.h"
1930d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun#include "llvm/CodeGen/SelectionDAG.h"
2030d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun#include "PPC.h"
2130d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun#include "PPCSubtarget.h"
2230d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
2330d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurunnamespace llvm {
2430d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun  namespace PPCISD {
2530d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun    enum NodeType {
2630d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      // Start the numbering where the builtin ops and target ops leave off.
2730d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      FIRST_NUMBER = ISD::BUILTIN_OP_END+PPC::INSTRUCTION_LIST_END,
2830d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
2930d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// FSEL - Traditional three-operand fsel node.
3030d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      ///
3130d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      FSEL,
3230d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
3330d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// FCFID - The FCFID instruction, taking an f64 operand and producing
3430d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// and f64 value containing the FP representation of the integer that
3530d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// was temporarily in the f64 operand.
3630d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      FCFID,
3730d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
3830d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// FCTI[D,W]Z - The FCTIDZ and FCTIWZ instructions, taking an f32 or f64
3930d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// operand, producing an f64 value containing the integer representation
4030d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// of that FP value.
4130d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      FCTIDZ, FCTIWZ,
4230d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
4330d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// STFIWX - The STFIWX instruction.  The first operand is an input token
4430d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// chain, then an f64 value to store, then an address to store it to,
4530d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// then a SRCVALUE for the address.
4630d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      STFIWX,
4730d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
4830d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      // VMADDFP, VNMSUBFP - The VMADDFP and VNMSUBFP instructions, taking
4930d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      // three v4f32 operands and producing a v4f32 result.
5030d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      VMADDFP, VNMSUBFP,
5130d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
5230d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// VPERM - The PPC VPERM Instruction.
5330d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      ///
5430d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      VPERM,
5530d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
5630d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// Hi/Lo - These represent the high and low 16-bit parts of a global
5730d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// address respectively.  These nodes have two operands, the first of
5830d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// which must be a TargetGlobalAddress, and the second of which must be a
5930d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// Constant.  Selected naively, these turn into 'lis G+C' and 'li G+C',
6030d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// though these are usually folded into other nodes.
6130d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      Hi, Lo,
6230d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
6330d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
6430d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
6530d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// compute an allocation on the stack.
6630d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      DYNALLOC,
6730d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun
6830d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// GlobalBaseReg - On Darwin, this node represents the result of the mflr
6930d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      /// at function entry, used for PIC code.
7030d4e1f3d81ad9f7a1aa14ce6d2ceb5df56c15cdSelim Gurun      GlobalBaseReg,
71
72      /// These nodes represent the 32-bit PPC shifts that operate on 6-bit
73      /// shift amounts.  These nodes are generated by the multi-precision shift
74      /// code.
75      SRL, SRA, SHL,
76
77      /// EXTSW_32 - This is the EXTSW instruction for use with "32-bit"
78      /// registers.
79      EXTSW_32,
80
81      /// STD_32 - This is the STD instruction for use with "32-bit" registers.
82      STD_32,
83
84      /// CALL - A direct function call.
85      CALL_Macho, CALL_ELF,
86
87      /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
88      /// MTCTR instruction.
89      MTCTR,
90
91      /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
92      /// BCTRL instruction.
93      BCTRL_Macho, BCTRL_ELF,
94
95      /// Return with a flag operand, matched by 'blr'
96      RET_FLAG,
97
98      /// R32 = MFCR(CRREG, INFLAG) - Represents the MFCR/MFOCRF instructions.
99      /// This copies the bits corresponding to the specified CRREG into the
100      /// resultant GPR.  Bits corresponding to other CR regs are undefined.
101      MFCR,
102
103      /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
104      /// instructions.  For lack of better number, we use the opcode number
105      /// encoding for the OPC field to identify the compare.  For example, 838
106      /// is VCMPGTSH.
107      VCMP,
108
109      /// RESVEC, OUTFLAG = VCMPo(LHS, RHS, OPC) - Represents one of the
110      /// altivec VCMP*o instructions.  For lack of better number, we use the
111      /// opcode number encoding for the OPC field to identify the compare.  For
112      /// example, 838 is VCMPGTSH.
113      VCMPo,
114
115      /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
116      /// corresponds to the COND_BRANCH pseudo instruction.  CRRC is the
117      /// condition register to branch on, OPC is the branch opcode to use (e.g.
118      /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
119      /// an optional input flag argument.
120      COND_BRANCH,
121
122      /// CHAIN = STBRX CHAIN, GPRC, Ptr, SRCVALUE, Type - This is a
123      /// byte-swapping store instruction.  It byte-swaps the low "Type" bits of
124      /// the GPRC input, then stores it through Ptr.  Type can be either i16 or
125      /// i32.
126      STBRX,
127
128      /// GPRC, CHAIN = LBRX CHAIN, Ptr, SRCVALUE, Type - This is a
129      /// byte-swapping load instruction.  It loads "Type" bits, byte swaps it,
130      /// then puts it in the bottom bits of the GPRC.  TYPE can be either i16
131      /// or i32.
132      LBRX
133    };
134  }
135
136  /// Define some predicates that are used for node matching.
137  namespace PPC {
138    /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
139    /// VPKUHUM instruction.
140    bool isVPKUHUMShuffleMask(SDNode *N, bool isUnary);
141
142    /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
143    /// VPKUWUM instruction.
144    bool isVPKUWUMShuffleMask(SDNode *N, bool isUnary);
145
146    /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
147    /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
148    bool isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary);
149
150    /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
151    /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
152    bool isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary);
153
154    /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
155    /// amount, otherwise return -1.
156    int isVSLDOIShuffleMask(SDNode *N, bool isUnary);
157
158    /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
159    /// specifies a splat of a single element that is suitable for input to
160    /// VSPLTB/VSPLTH/VSPLTW.
161    bool isSplatShuffleMask(SDNode *N, unsigned EltSize);
162
163    /// isAllNegativeZeroVector - Returns true if all elements of build_vector
164    /// are -0.0.
165    bool isAllNegativeZeroVector(SDNode *N);
166
167    /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
168    /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
169    unsigned getVSPLTImmediate(SDNode *N, unsigned EltSize);
170
171    /// get_VSPLTI_elt - If this is a build_vector of constants which can be
172    /// formed by using a vspltis[bhw] instruction of the specified element
173    /// size, return the constant being splatted.  The ByteSize field indicates
174    /// the number of bytes of each element [124] -> [bhw].
175    SDOperand get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
176  }
177
178  class PPCTargetLowering : public TargetLowering {
179    int VarArgsFrameIndex;            // FrameIndex for start of varargs area.
180    int VarArgsStackOffset;           // StackOffset for start of stack
181                                      // arguments.
182    unsigned VarArgsNumGPR;           // Index of the first unused integer
183                                      // register for parameter passing.
184    unsigned VarArgsNumFPR;           // Index of the first unused double
185                                      // register for parameter passing.
186    int ReturnAddrIndex;              // FrameIndex for return slot.
187    const PPCSubtarget &PPCSubTarget;
188  public:
189    explicit PPCTargetLowering(PPCTargetMachine &TM);
190
191    /// getTargetNodeName() - This method returns the name of a target specific
192    /// DAG node.
193    virtual const char *getTargetNodeName(unsigned Opcode) const;
194
195    /// getPreIndexedAddressParts - returns true by value, base pointer and
196    /// offset pointer and addressing mode by reference if the node's address
197    /// can be legally represented as pre-indexed load / store address.
198    virtual bool getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
199                                           SDOperand &Offset,
200                                           ISD::MemIndexedMode &AM,
201                                           SelectionDAG &DAG);
202
203    /// SelectAddressRegReg - Given the specified addressed, check to see if it
204    /// can be represented as an indexed [r+r] operation.  Returns false if it
205    /// can be more efficiently represented with [r+imm].
206    bool SelectAddressRegReg(SDOperand N, SDOperand &Base, SDOperand &Index,
207                             SelectionDAG &DAG);
208
209    /// SelectAddressRegImm - Returns true if the address N can be represented
210    /// by a base register plus a signed 16-bit displacement [r+imm], and if it
211    /// is not better represented as reg+reg.
212    bool SelectAddressRegImm(SDOperand N, SDOperand &Disp, SDOperand &Base,
213                             SelectionDAG &DAG);
214
215    /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
216    /// represented as an indexed [r+r] operation.
217    bool SelectAddressRegRegOnly(SDOperand N, SDOperand &Base, SDOperand &Index,
218                                 SelectionDAG &DAG);
219
220    /// SelectAddressRegImmShift - Returns true if the address N can be
221    /// represented by a base register plus a signed 14-bit displacement
222    /// [r+imm*4].  Suitable for use by STD and friends.
223    bool SelectAddressRegImmShift(SDOperand N, SDOperand &Disp, SDOperand &Base,
224                                  SelectionDAG &DAG);
225
226
227    /// LowerOperation - Provide custom lowering hooks for some operations.
228    ///
229    virtual SDOperand LowerOperation(SDOperand Op, SelectionDAG &DAG);
230
231    virtual SDOperand PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
232
233    virtual void computeMaskedBitsForTargetNode(const SDOperand Op,
234                                                uint64_t Mask,
235                                                uint64_t &KnownZero,
236                                                uint64_t &KnownOne,
237                                                const SelectionDAG &DAG,
238                                                unsigned Depth = 0) const;
239
240    virtual MachineBasicBlock *InsertAtEndOfBasicBlock(MachineInstr *MI,
241                                                       MachineBasicBlock *MBB);
242
243    ConstraintType getConstraintType(const std::string &Constraint) const;
244    std::pair<unsigned, const TargetRegisterClass*>
245      getRegForInlineAsmConstraint(const std::string &Constraint,
246                                   MVT::ValueType VT) const;
247    SDOperand isOperandValidForConstraint(SDOperand Op, char ConstraintLetter,
248                                          SelectionDAG &DAG);
249
250    /// isLegalAddressingMode - Return true if the addressing mode represented
251    /// by AM is legal for this target, for a load/store of the specified type.
252    virtual bool isLegalAddressingMode(const AddrMode &AM, const Type *Ty)const;
253
254    /// isLegalAddressImmediate - Return true if the integer value can be used
255    /// as the offset of the target addressing mode for load / store of the
256    /// given type.
257    virtual bool isLegalAddressImmediate(int64_t V, const Type *Ty) const;
258
259    /// isLegalAddressImmediate - Return true if the GlobalValue can be used as
260    /// the offset of the target addressing mode.
261    virtual bool isLegalAddressImmediate(GlobalValue *GV) const;
262
263    SDOperand LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG);
264  };
265}
266
267#endif   // LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H
268