TargetLowering.h revision 6586adfc1a52fbafd0733b74cd73651e328ce019
1//===-- llvm/Target/TargetLowering.h - Target Lowering Info -----*- C++ -*-===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by the LLVM research group and is distributed under
6// the University of Illinois Open Source License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file describes how to lower LLVM code to machine code.  This has two
11// main components:
12//
13//  1. Which ValueTypes are natively supported by the target.
14//  2. Which operations are supported for supported ValueTypes.
15//  3. Cost thresholds for alternative implementations of certain operations.
16//
17// In addition it has a few other components, like information about FP
18// immediates.
19//
20//===----------------------------------------------------------------------===//
21
22#ifndef LLVM_TARGET_TARGETLOWERING_H
23#define LLVM_TARGET_TARGETLOWERING_H
24
25#include "llvm/Type.h"
26#include "llvm/CodeGen/SelectionDAGNodes.h"
27#include <map>
28
29namespace llvm {
30  class Value;
31  class Function;
32  class TargetMachine;
33  class TargetData;
34  class TargetRegisterClass;
35  class SDNode;
36  class SDOperand;
37  class SelectionDAG;
38  class MachineBasicBlock;
39  class MachineInstr;
40
41//===----------------------------------------------------------------------===//
42/// TargetLowering - This class defines information used to lower LLVM code to
43/// legal SelectionDAG operators that the target instruction selector can accept
44/// natively.
45///
46/// This class also defines callbacks that targets must implement to lower
47/// target-specific constructs to SelectionDAG operators.
48///
49class TargetLowering {
50public:
51  /// LegalizeAction - This enum indicates whether operations are valid for a
52  /// target, and if not, what action should be used to make them valid.
53  enum LegalizeAction {
54    Legal,      // The target natively supports this operation.
55    Promote,    // This operation should be executed in a larger type.
56    Expand,     // Try to expand this to other ops, otherwise use a libcall.
57    Custom      // Use the LowerOperation hook to implement custom lowering.
58  };
59
60  enum OutOfRangeShiftAmount {
61    Undefined,  // Oversized shift amounts are undefined (default).
62    Mask,       // Shift amounts are auto masked (anded) to value size.
63    Extend      // Oversized shift pulls in zeros or sign bits.
64  };
65
66  enum SetCCResultValue {
67    UndefinedSetCCResult,          // SetCC returns a garbage/unknown extend.
68    ZeroOrOneSetCCResult,          // SetCC returns a zero extended result.
69    ZeroOrNegativeOneSetCCResult   // SetCC returns a sign extended result.
70  };
71
72  enum SchedPreference {
73    SchedulingForLatency,          // Scheduling for shortest total latency.
74    SchedulingForRegPressure       // Scheduling for lowest register pressure.
75  };
76
77  TargetLowering(TargetMachine &TM);
78  virtual ~TargetLowering();
79
80  TargetMachine &getTargetMachine() const { return TM; }
81  const TargetData *getTargetData() const { return TD; }
82
83  bool isLittleEndian() const { return IsLittleEndian; }
84  MVT::ValueType getPointerTy() const { return PointerTy; }
85  MVT::ValueType getShiftAmountTy() const { return ShiftAmountTy; }
86  OutOfRangeShiftAmount getShiftAmountFlavor() const {return ShiftAmtHandling; }
87
88  /// usesGlobalOffsetTable - Return true if this target uses a GOT for PIC
89  /// codegen.
90  bool usesGlobalOffsetTable() const { return UsesGlobalOffsetTable; }
91
92  /// isSetCCExpensive - Return true if the setcc operation is expensive for
93  /// this target.
94  bool isSetCCExpensive() const { return SetCCIsExpensive; }
95
96  /// isIntDivCheap() - Return true if integer divide is usually cheaper than
97  /// a sequence of several shifts, adds, and multiplies for this target.
98  bool isIntDivCheap() const { return IntDivIsCheap; }
99
100  /// isPow2DivCheap() - Return true if pow2 div is cheaper than a chain of
101  /// srl/add/sra.
102  bool isPow2DivCheap() const { return Pow2DivIsCheap; }
103
104  /// getSetCCResultTy - Return the ValueType of the result of setcc operations.
105  ///
106  MVT::ValueType getSetCCResultTy() const { return SetCCResultTy; }
107
108  /// getSetCCResultContents - For targets without boolean registers, this flag
109  /// returns information about the contents of the high-bits in the setcc
110  /// result register.
111  SetCCResultValue getSetCCResultContents() const { return SetCCResultContents;}
112
113  /// getSchedulingPreference - Return target scheduling preference.
114  SchedPreference getSchedulingPreference() const {
115    return SchedPreferenceInfo;
116  }
117
118  /// getRegClassFor - Return the register class that should be used for the
119  /// specified value type.  This may only be called on legal types.
120  TargetRegisterClass *getRegClassFor(MVT::ValueType VT) const {
121    TargetRegisterClass *RC = RegClassForVT[VT];
122    assert(RC && "This value type is not natively supported!");
123    return RC;
124  }
125
126  /// isTypeLegal - Return true if the target has native support for the
127  /// specified value type.  This means that it has a register that directly
128  /// holds it without promotions or expansions.
129  bool isTypeLegal(MVT::ValueType VT) const {
130    return RegClassForVT[VT] != 0;
131  }
132
133  class ValueTypeActionImpl {
134    /// ValueTypeActions - This is a bitvector that contains two bits for each
135    /// value type, where the two bits correspond to the LegalizeAction enum.
136    /// This can be queried with "getTypeAction(VT)".
137    uint32_t ValueTypeActions[2];
138  public:
139    ValueTypeActionImpl() {
140      ValueTypeActions[0] = ValueTypeActions[1] = 0;
141    }
142    ValueTypeActionImpl(const ValueTypeActionImpl &RHS) {
143      ValueTypeActions[0] = RHS.ValueTypeActions[0];
144      ValueTypeActions[1] = RHS.ValueTypeActions[1];
145    }
146
147    LegalizeAction getTypeAction(MVT::ValueType VT) const {
148      return (LegalizeAction)((ValueTypeActions[VT>>4] >> ((2*VT) & 31)) & 3);
149    }
150    void setTypeAction(MVT::ValueType VT, LegalizeAction Action) {
151      assert(unsigned(VT >> 4) <
152             sizeof(ValueTypeActions)/sizeof(ValueTypeActions[0]));
153      ValueTypeActions[VT>>4] |= Action << ((VT*2) & 31);
154    }
155  };
156
157  const ValueTypeActionImpl &getValueTypeActions() const {
158    return ValueTypeActions;
159  }
160
161  /// getTypeAction - Return how we should legalize values of this type, either
162  /// it is already legal (return 'Legal') or we need to promote it to a larger
163  /// type (return 'Promote'), or we need to expand it into multiple registers
164  /// of smaller integer type (return 'Expand').  'Custom' is not an option.
165  LegalizeAction getTypeAction(MVT::ValueType VT) const {
166    return ValueTypeActions.getTypeAction(VT);
167  }
168
169  /// getTypeToTransformTo - For types supported by the target, this is an
170  /// identity function.  For types that must be promoted to larger types, this
171  /// returns the larger type to promote to.  For integer types that are larger
172  /// than the largest integer register, this contains one step in the expansion
173  /// to get to the smaller register. For illegal floating point types, this
174  /// returns the integer type to transform to.
175  MVT::ValueType getTypeToTransformTo(MVT::ValueType VT) const {
176    return TransformToType[VT];
177  }
178
179  /// getPackedTypeBreakdown - Packed types are broken down into some number of
180  /// legal first class types.  For example, <8 x float> maps to 2 MVT::v4f32
181  /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
182  /// Similarly, <2 x long> turns into 4 MVT::i32 values with both PPC and X86.
183  ///
184  /// This method returns the number of registers needed, and the VT for each
185  /// register.  It also returns the VT of the PackedType elements before they
186  /// are promoted/expanded.
187  ///
188  unsigned getPackedTypeBreakdown(const PackedType *PTy,
189                                  MVT::ValueType &PTyElementVT,
190                                  MVT::ValueType &PTyLegalElementVT) const;
191
192  typedef std::vector<double>::const_iterator legal_fpimm_iterator;
193  legal_fpimm_iterator legal_fpimm_begin() const {
194    return LegalFPImmediates.begin();
195  }
196  legal_fpimm_iterator legal_fpimm_end() const {
197    return LegalFPImmediates.end();
198  }
199
200  /// isShuffleMaskLegal - Targets can use this to indicate that they only
201  /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
202  /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
203  /// are assumed to be legal.
204  virtual bool isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
205    return true;
206  }
207
208  /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
209  /// used by Targets can use this to indicate if there is a suitable
210  /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
211  /// pool entry.
212  virtual bool isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
213                                      MVT::ValueType EVT,
214                                      SelectionDAG &DAG) const {
215    return false;
216  }
217
218  /// getOperationAction - Return how this operation should be treated: either
219  /// it is legal, needs to be promoted to a larger size, needs to be
220  /// expanded to some other code sequence, or the target has a custom expander
221  /// for it.
222  LegalizeAction getOperationAction(unsigned Op, MVT::ValueType VT) const {
223    return (LegalizeAction)((OpActions[Op] >> (2*VT)) & 3);
224  }
225
226  /// isOperationLegal - Return true if the specified operation is legal on this
227  /// target.
228  bool isOperationLegal(unsigned Op, MVT::ValueType VT) const {
229    return getOperationAction(Op, VT) == Legal ||
230           getOperationAction(Op, VT) == Custom;
231  }
232
233  /// getLoadXAction - Return how this load with extension should be treated:
234  /// either it is legal, needs to be promoted to a larger size, needs to be
235  /// expanded to some other code sequence, or the target has a custom expander
236  /// for it.
237  LegalizeAction getLoadXAction(unsigned LType, MVT::ValueType VT) const {
238    return (LegalizeAction)((LoadXActions[LType] >> (2*VT)) & 3);
239  }
240
241  /// isLoadXLegal - Return true if the specified load with extension is legal
242  /// on this target.
243  bool isLoadXLegal(unsigned LType, MVT::ValueType VT) const {
244    return getLoadXAction(LType, VT) == Legal ||
245           getLoadXAction(LType, VT) == Custom;
246  }
247
248  /// getStoreXAction - Return how this store with truncation should be treated:
249  /// either it is legal, needs to be promoted to a larger size, needs to be
250  /// expanded to some other code sequence, or the target has a custom expander
251  /// for it.
252  LegalizeAction getStoreXAction(MVT::ValueType VT) const {
253    return (LegalizeAction)((StoreXActions >> (2*VT)) & 3);
254  }
255
256  /// isStoreXLegal - Return true if the specified store with truncation is
257  /// legal on this target.
258  bool isStoreXLegal(MVT::ValueType VT) const {
259    return getStoreXAction(VT) == Legal || getStoreXAction(VT) == Custom;
260  }
261
262  /// getIndexedLoadAction - Return how the indexed load should be treated:
263  /// either it is legal, needs to be promoted to a larger size, needs to be
264  /// expanded to some other code sequence, or the target has a custom expander
265  /// for it.
266  LegalizeAction
267  getIndexedLoadAction(unsigned IdxMode, MVT::ValueType VT) const {
268    return (LegalizeAction)((IndexedModeActions[0][IdxMode] >> (2*VT)) & 3);
269  }
270
271  /// isIndexedLoadLegal - Return true if the specified indexed load is legal
272  /// on this target.
273  bool isIndexedLoadLegal(unsigned IdxMode, MVT::ValueType VT) const {
274    return getIndexedLoadAction(IdxMode, VT) == Legal ||
275           getIndexedLoadAction(IdxMode, VT) == Custom;
276  }
277
278  /// getIndexedStoreAction - Return how the indexed store should be treated:
279  /// either it is legal, needs to be promoted to a larger size, needs to be
280  /// expanded to some other code sequence, or the target has a custom expander
281  /// for it.
282  LegalizeAction
283  getIndexedStoreAction(unsigned IdxMode, MVT::ValueType VT) const {
284    return (LegalizeAction)((IndexedModeActions[1][IdxMode] >> (2*VT)) & 3);
285  }
286
287  /// isIndexedStoreLegal - Return true if the specified indexed load is legal
288  /// on this target.
289  bool isIndexedStoreLegal(unsigned IdxMode, MVT::ValueType VT) const {
290    return getIndexedStoreAction(IdxMode, VT) == Legal ||
291           getIndexedStoreAction(IdxMode, VT) == Custom;
292  }
293
294  /// getTypeToPromoteTo - If the action for this operation is to promote, this
295  /// method returns the ValueType to promote to.
296  MVT::ValueType getTypeToPromoteTo(unsigned Op, MVT::ValueType VT) const {
297    assert(getOperationAction(Op, VT) == Promote &&
298           "This operation isn't promoted!");
299
300    // See if this has an explicit type specified.
301    std::map<std::pair<unsigned, MVT::ValueType>,
302             MVT::ValueType>::const_iterator PTTI =
303      PromoteToType.find(std::make_pair(Op, VT));
304    if (PTTI != PromoteToType.end()) return PTTI->second;
305
306    assert((MVT::isInteger(VT) || MVT::isFloatingPoint(VT)) &&
307           "Cannot autopromote this type, add it with AddPromotedToType.");
308
309    MVT::ValueType NVT = VT;
310    do {
311      NVT = (MVT::ValueType)(NVT+1);
312      assert(MVT::isInteger(NVT) == MVT::isInteger(VT) && NVT != MVT::isVoid &&
313             "Didn't find type to promote to!");
314    } while (!isTypeLegal(NVT) ||
315              getOperationAction(Op, NVT) == Promote);
316    return NVT;
317  }
318
319  /// getValueType - Return the MVT::ValueType corresponding to this LLVM type.
320  /// This is fixed by the LLVM operations except for the pointer size.
321  MVT::ValueType getValueType(const Type *Ty) const {
322    switch (Ty->getTypeID()) {
323    default: assert(0 && "Unknown type!");
324    case Type::VoidTyID:    return MVT::isVoid;
325    case Type::BoolTyID:    return MVT::i1;
326    case Type::UByteTyID:
327    case Type::SByteTyID:   return MVT::i8;
328    case Type::ShortTyID:
329    case Type::UShortTyID:  return MVT::i16;
330    case Type::IntTyID:
331    case Type::UIntTyID:    return MVT::i32;
332    case Type::LongTyID:
333    case Type::ULongTyID:   return MVT::i64;
334    case Type::FloatTyID:   return MVT::f32;
335    case Type::DoubleTyID:  return MVT::f64;
336    case Type::PointerTyID: return PointerTy;
337    case Type::PackedTyID:  return MVT::Vector;
338    }
339  }
340
341  /// getNumElements - Return the number of registers that this ValueType will
342  /// eventually require.  This is one for any types promoted to live in larger
343  /// registers, but may be more than one for types (like i64) that are split
344  /// into pieces.
345  unsigned getNumElements(MVT::ValueType VT) const {
346    return NumElementsForVT[VT];
347  }
348
349  /// hasTargetDAGCombine - If true, the target has custom DAG combine
350  /// transformations that it can perform for the specified node.
351  bool hasTargetDAGCombine(ISD::NodeType NT) const {
352    return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7));
353  }
354
355  /// This function returns the maximum number of store operations permitted
356  /// to replace a call to llvm.memset. The value is set by the target at the
357  /// performance threshold for such a replacement.
358  /// @brief Get maximum # of store operations permitted for llvm.memset
359  unsigned getMaxStoresPerMemset() const { return maxStoresPerMemset; }
360
361  /// This function returns the maximum number of store operations permitted
362  /// to replace a call to llvm.memcpy. The value is set by the target at the
363  /// performance threshold for such a replacement.
364  /// @brief Get maximum # of store operations permitted for llvm.memcpy
365  unsigned getMaxStoresPerMemcpy() const { return maxStoresPerMemcpy; }
366
367  /// This function returns the maximum number of store operations permitted
368  /// to replace a call to llvm.memmove. The value is set by the target at the
369  /// performance threshold for such a replacement.
370  /// @brief Get maximum # of store operations permitted for llvm.memmove
371  unsigned getMaxStoresPerMemmove() const { return maxStoresPerMemmove; }
372
373  /// This function returns true if the target allows unaligned memory accesses.
374  /// This is used, for example, in situations where an array copy/move/set is
375  /// converted to a sequence of store operations. It's use helps to ensure that
376  /// such replacements don't generate code that causes an alignment error
377  /// (trap) on the target machine.
378  /// @brief Determine if the target supports unaligned memory accesses.
379  bool allowsUnalignedMemoryAccesses() const {
380    return allowUnalignedMemoryAccesses;
381  }
382
383  /// usesUnderscoreSetJmp - Determine if we should use _setjmp or setjmp
384  /// to implement llvm.setjmp.
385  bool usesUnderscoreSetJmp() const {
386    return UseUnderscoreSetJmp;
387  }
388
389  /// usesUnderscoreLongJmp - Determine if we should use _longjmp or longjmp
390  /// to implement llvm.longjmp.
391  bool usesUnderscoreLongJmp() const {
392    return UseUnderscoreLongJmp;
393  }
394
395  /// getStackPointerRegisterToSaveRestore - If a physical register, this
396  /// specifies the register that llvm.savestack/llvm.restorestack should save
397  /// and restore.
398  unsigned getStackPointerRegisterToSaveRestore() const {
399    return StackPointerRegisterToSaveRestore;
400  }
401
402  /// getJumpBufSize - returns the target's jmp_buf size in bytes (if never
403  /// set, the default is 200)
404  unsigned getJumpBufSize() const {
405    return JumpBufSize;
406  }
407
408  /// getJumpBufAlignment - returns the target's jmp_buf alignment in bytes
409  /// (if never set, the default is 0)
410  unsigned getJumpBufAlignment() const {
411    return JumpBufAlignment;
412  }
413
414  /// getPreIndexedAddressParts - returns true by value, base pointer and
415  /// offset pointer and addressing mode by reference if the node's address
416  /// can be legally represented as pre-indexed load / store address.
417  virtual bool getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
418                                         SDOperand &Offset,
419                                         ISD::MemIndexedMode &AM,
420                                         SelectionDAG &DAG) {
421    return false;
422  }
423
424  /// getPostIndexedAddressParts - returns true by value, base pointer and
425  /// offset pointer and addressing mode by reference if this node can be
426  /// combined with a load / store to form a post-indexed load / store.
427  virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
428                                          SDOperand &Base, SDOperand &Offset,
429                                          ISD::MemIndexedMode &AM,
430                                          SelectionDAG &DAG) {
431    return false;
432  }
433
434  //===--------------------------------------------------------------------===//
435  // TargetLowering Optimization Methods
436  //
437
438  /// TargetLoweringOpt - A convenience struct that encapsulates a DAG, and two
439  /// SDOperands for returning information from TargetLowering to its clients
440  /// that want to combine
441  struct TargetLoweringOpt {
442    SelectionDAG &DAG;
443    SDOperand Old;
444    SDOperand New;
445
446    TargetLoweringOpt(SelectionDAG &InDAG) : DAG(InDAG) {}
447
448    bool CombineTo(SDOperand O, SDOperand N) {
449      Old = O;
450      New = N;
451      return true;
452    }
453
454    /// ShrinkDemandedConstant - Check to see if the specified operand of the
455    /// specified instruction is a constant integer.  If so, check to see if there
456    /// are any bits set in the constant that are not demanded.  If so, shrink the
457    /// constant and return true.
458    bool ShrinkDemandedConstant(SDOperand Op, uint64_t Demanded);
459  };
460
461  /// MaskedValueIsZero - Return true if 'Op & Mask' is known to be zero.  We
462  /// use this predicate to simplify operations downstream.  Op and Mask are
463  /// known to be the same type.
464  bool MaskedValueIsZero(SDOperand Op, uint64_t Mask, unsigned Depth = 0)
465    const;
466
467  /// ComputeMaskedBits - Determine which of the bits specified in Mask are
468  /// known to be either zero or one and return them in the KnownZero/KnownOne
469  /// bitsets.  This code only analyzes bits in Mask, in order to short-circuit
470  /// processing.  Targets can implement the computeMaskedBitsForTargetNode
471  /// method, to allow target nodes to be understood.
472  void ComputeMaskedBits(SDOperand Op, uint64_t Mask, uint64_t &KnownZero,
473                         uint64_t &KnownOne, unsigned Depth = 0) const;
474
475  /// SimplifyDemandedBits - Look at Op.  At this point, we know that only the
476  /// DemandedMask bits of the result of Op are ever used downstream.  If we can
477  /// use this information to simplify Op, create a new simplified DAG node and
478  /// return true, returning the original and new nodes in Old and New.
479  /// Otherwise, analyze the expression and return a mask of KnownOne and
480  /// KnownZero bits for the expression (used to simplify the caller).
481  /// The KnownZero/One bits may only be accurate for those bits in the
482  /// DemandedMask.
483  bool SimplifyDemandedBits(SDOperand Op, uint64_t DemandedMask,
484                            uint64_t &KnownZero, uint64_t &KnownOne,
485                            TargetLoweringOpt &TLO, unsigned Depth = 0) const;
486
487  /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
488  /// Mask are known to be either zero or one and return them in the
489  /// KnownZero/KnownOne bitsets.
490  virtual void computeMaskedBitsForTargetNode(const SDOperand Op,
491                                              uint64_t Mask,
492                                              uint64_t &KnownZero,
493                                              uint64_t &KnownOne,
494                                              unsigned Depth = 0) const;
495
496  /// ComputeNumSignBits - Return the number of times the sign bit of the
497  /// register is replicated into the other bits.  We know that at least 1 bit
498  /// is always equal to the sign bit (itself), but other cases can give us
499  /// information.  For example, immediately after an "SRA X, 2", we know that
500  /// the top 3 bits are all equal to each other, so we return 3.
501  unsigned ComputeNumSignBits(SDOperand Op, unsigned Depth = 0) const;
502
503  /// ComputeNumSignBitsForTargetNode - This method can be implemented by
504  /// targets that want to expose additional information about sign bits to the
505  /// DAG Combiner.
506  virtual unsigned ComputeNumSignBitsForTargetNode(SDOperand Op,
507                                                   unsigned Depth = 0) const;
508
509  struct DAGCombinerInfo {
510    void *DC;  // The DAG Combiner object.
511    bool BeforeLegalize;
512  public:
513    SelectionDAG &DAG;
514
515    DAGCombinerInfo(SelectionDAG &dag, bool bl, void *dc)
516      : DC(dc), BeforeLegalize(bl), DAG(dag) {}
517
518    bool isBeforeLegalize() const { return BeforeLegalize; }
519
520    void AddToWorklist(SDNode *N);
521    SDOperand CombineTo(SDNode *N, const std::vector<SDOperand> &To);
522    SDOperand CombineTo(SDNode *N, SDOperand Res);
523    SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1);
524  };
525
526  /// PerformDAGCombine - This method will be invoked for all target nodes and
527  /// for any target-independent nodes that the target has registered with
528  /// invoke it for.
529  ///
530  /// The semantics are as follows:
531  /// Return Value:
532  ///   SDOperand.Val == 0   - No change was made
533  ///   SDOperand.Val == N   - N was replaced, is dead, and is already handled.
534  ///   otherwise            - N should be replaced by the returned Operand.
535  ///
536  /// In addition, methods provided by DAGCombinerInfo may be used to perform
537  /// more complex transformations.
538  ///
539  virtual SDOperand PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
540
541  //===--------------------------------------------------------------------===//
542  // TargetLowering Configuration Methods - These methods should be invoked by
543  // the derived class constructor to configure this object for the target.
544  //
545
546protected:
547  /// setUsesGlobalOffsetTable - Specify that this target does or doesn't use a
548  /// GOT for PC-relative code.
549  void setUsesGlobalOffsetTable(bool V) { UsesGlobalOffsetTable = V; }
550
551  /// setShiftAmountType - Describe the type that should be used for shift
552  /// amounts.  This type defaults to the pointer type.
553  void setShiftAmountType(MVT::ValueType VT) { ShiftAmountTy = VT; }
554
555  /// setSetCCResultType - Describe the type that shoudl be used as the result
556  /// of a setcc operation.  This defaults to the pointer type.
557  void setSetCCResultType(MVT::ValueType VT) { SetCCResultTy = VT; }
558
559  /// setSetCCResultContents - Specify how the target extends the result of a
560  /// setcc operation in a register.
561  void setSetCCResultContents(SetCCResultValue Ty) { SetCCResultContents = Ty; }
562
563  /// setSchedulingPreference - Specify the target scheduling preference.
564  void setSchedulingPreference(SchedPreference Pref) {
565    SchedPreferenceInfo = Pref;
566  }
567
568  /// setShiftAmountFlavor - Describe how the target handles out of range shift
569  /// amounts.
570  void setShiftAmountFlavor(OutOfRangeShiftAmount OORSA) {
571    ShiftAmtHandling = OORSA;
572  }
573
574  /// setUseUnderscoreSetJmp - Indicate whether this target prefers to
575  /// use _setjmp to implement llvm.setjmp or the non _ version.
576  /// Defaults to false.
577  void setUseUnderscoreSetJmp(bool Val) {
578    UseUnderscoreSetJmp = Val;
579  }
580
581  /// setUseUnderscoreLongJmp - Indicate whether this target prefers to
582  /// use _longjmp to implement llvm.longjmp or the non _ version.
583  /// Defaults to false.
584  void setUseUnderscoreLongJmp(bool Val) {
585    UseUnderscoreLongJmp = Val;
586  }
587
588  /// setStackPointerRegisterToSaveRestore - If set to a physical register, this
589  /// specifies the register that llvm.savestack/llvm.restorestack should save
590  /// and restore.
591  void setStackPointerRegisterToSaveRestore(unsigned R) {
592    StackPointerRegisterToSaveRestore = R;
593  }
594
595  /// setSetCCIxExpensive - This is a short term hack for targets that codegen
596  /// setcc as a conditional branch.  This encourages the code generator to fold
597  /// setcc operations into other operations if possible.
598  void setSetCCIsExpensive() { SetCCIsExpensive = true; }
599
600  /// setIntDivIsCheap - Tells the code generator that integer divide is
601  /// expensive, and if possible, should be replaced by an alternate sequence
602  /// of instructions not containing an integer divide.
603  void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; }
604
605  /// setPow2DivIsCheap - Tells the code generator that it shouldn't generate
606  /// srl/add/sra for a signed divide by power of two, and let the target handle
607  /// it.
608  void setPow2DivIsCheap(bool isCheap = true) { Pow2DivIsCheap = isCheap; }
609
610  /// addRegisterClass - Add the specified register class as an available
611  /// regclass for the specified value type.  This indicates the selector can
612  /// handle values of that class natively.
613  void addRegisterClass(MVT::ValueType VT, TargetRegisterClass *RC) {
614    AvailableRegClasses.push_back(std::make_pair(VT, RC));
615    RegClassForVT[VT] = RC;
616  }
617
618  /// computeRegisterProperties - Once all of the register classes are added,
619  /// this allows us to compute derived properties we expose.
620  void computeRegisterProperties();
621
622  /// setOperationAction - Indicate that the specified operation does not work
623  /// with the specified type and indicate what to do about it.
624  void setOperationAction(unsigned Op, MVT::ValueType VT,
625                          LegalizeAction Action) {
626    assert(VT < 32 && Op < sizeof(OpActions)/sizeof(OpActions[0]) &&
627           "Table isn't big enough!");
628    OpActions[Op] &= ~(uint64_t(3UL) << VT*2);
629    OpActions[Op] |= (uint64_t)Action << VT*2;
630  }
631
632  /// setLoadXAction - Indicate that the specified load with extension does not
633  /// work with the with specified type and indicate what to do about it.
634  void setLoadXAction(unsigned ExtType, MVT::ValueType VT,
635                      LegalizeAction Action) {
636    assert(VT < 32 && ExtType < sizeof(LoadXActions)/sizeof(LoadXActions[0]) &&
637           "Table isn't big enough!");
638    LoadXActions[ExtType] &= ~(uint64_t(3UL) << VT*2);
639    LoadXActions[ExtType] |= (uint64_t)Action << VT*2;
640  }
641
642  /// setStoreXAction - Indicate that the specified store with truncation does
643  /// not work with the with specified type and indicate what to do about it.
644  void setStoreXAction(MVT::ValueType VT, LegalizeAction Action) {
645    assert(VT < 32 && "Table isn't big enough!");
646    StoreXActions &= ~(uint64_t(3UL) << VT*2);
647    StoreXActions |= (uint64_t)Action << VT*2;
648  }
649
650  /// setIndexedLoadAction - Indicate that the specified indexed load does or
651  /// does not work with the with specified type and indicate what to do abort
652  /// it. NOTE: All indexed mode loads are initialized to Expand in
653  /// TargetLowering.cpp
654  void setIndexedLoadAction(unsigned IdxMode, MVT::ValueType VT,
655                            LegalizeAction Action) {
656    assert(VT < 32 && IdxMode <
657           sizeof(IndexedModeActions[0]) / sizeof(IndexedModeActions[0][0]) &&
658           "Table isn't big enough!");
659    IndexedModeActions[0][IdxMode] &= ~(uint64_t(3UL) << VT*2);
660    IndexedModeActions[0][IdxMode] |= (uint64_t)Action << VT*2;
661  }
662
663  /// setIndexedStoreAction - Indicate that the specified indexed store does or
664  /// does not work with the with specified type and indicate what to do about
665  /// it. NOTE: All indexed mode stores are initialized to Expand in
666  /// TargetLowering.cpp
667  void setIndexedStoreAction(unsigned IdxMode, MVT::ValueType VT,
668                             LegalizeAction Action) {
669    assert(VT < 32 && IdxMode <
670           sizeof(IndexedModeActions[1]) / sizeof(IndexedModeActions[1][0]) &&
671           "Table isn't big enough!");
672    IndexedModeActions[1][IdxMode] &= ~(uint64_t(3UL) << VT*2);
673    IndexedModeActions[1][IdxMode] |= (uint64_t)Action << VT*2;
674  }
675
676  /// AddPromotedToType - If Opc/OrigVT is specified as being promoted, the
677  /// promotion code defaults to trying a larger integer/fp until it can find
678  /// one that works.  If that default is insufficient, this method can be used
679  /// by the target to override the default.
680  void AddPromotedToType(unsigned Opc, MVT::ValueType OrigVT,
681                         MVT::ValueType DestVT) {
682    PromoteToType[std::make_pair(Opc, OrigVT)] = DestVT;
683  }
684
685  /// addLegalFPImmediate - Indicate that this target can instruction select
686  /// the specified FP immediate natively.
687  void addLegalFPImmediate(double Imm) {
688    LegalFPImmediates.push_back(Imm);
689  }
690
691  /// setTargetDAGCombine - Targets should invoke this method for each target
692  /// independent node that they want to provide a custom DAG combiner for by
693  /// implementing the PerformDAGCombine virtual method.
694  void setTargetDAGCombine(ISD::NodeType NT) {
695    TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7);
696  }
697
698  /// setJumpBufSize - Set the target's required jmp_buf buffer size (in
699  /// bytes); default is 200
700  void setJumpBufSize(unsigned Size) {
701    JumpBufSize = Size;
702  }
703
704  /// setJumpBufAlignment - Set the target's required jmp_buf buffer
705  /// alignment (in bytes); default is 0
706  void setJumpBufAlignment(unsigned Align) {
707    JumpBufAlignment = Align;
708  }
709
710public:
711
712  //===--------------------------------------------------------------------===//
713  // Lowering methods - These methods must be implemented by targets so that
714  // the SelectionDAGLowering code knows how to lower these.
715  //
716
717  /// LowerArguments - This hook must be implemented to indicate how we should
718  /// lower the arguments for the specified function, into the specified DAG.
719  virtual std::vector<SDOperand>
720  LowerArguments(Function &F, SelectionDAG &DAG);
721
722  /// LowerCallTo - This hook lowers an abstract call to a function into an
723  /// actual call.  This returns a pair of operands.  The first element is the
724  /// return value for the function (if RetTy is not VoidTy).  The second
725  /// element is the outgoing token chain.
726  typedef std::vector<std::pair<SDOperand, const Type*> > ArgListTy;
727  virtual std::pair<SDOperand, SDOperand>
728  LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg,
729              unsigned CallingConv, bool isTailCall, SDOperand Callee,
730              ArgListTy &Args, SelectionDAG &DAG);
731
732  /// LowerFrameReturnAddress - This hook lowers a call to llvm.returnaddress or
733  /// llvm.frameaddress (depending on the value of the first argument).  The
734  /// return values are the result pointer and the resultant token chain.  If
735  /// not implemented, both of these intrinsics will return null.
736  virtual std::pair<SDOperand, SDOperand>
737  LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
738                          SelectionDAG &DAG);
739
740  /// LowerOperation - This callback is invoked for operations that are
741  /// unsupported by the target, which are registered to use 'custom' lowering,
742  /// and whose defined values are all legal.
743  /// If the target has no operations that require custom lowering, it need not
744  /// implement this.  The default implementation of this aborts.
745  virtual SDOperand LowerOperation(SDOperand Op, SelectionDAG &DAG);
746
747  /// CustomPromoteOperation - This callback is invoked for operations that are
748  /// unsupported by the target, are registered to use 'custom' lowering, and
749  /// whose type needs to be promoted.
750  virtual SDOperand CustomPromoteOperation(SDOperand Op, SelectionDAG &DAG);
751
752  /// getTargetNodeName() - This method returns the name of a target specific
753  /// DAG node.
754  virtual const char *getTargetNodeName(unsigned Opcode) const;
755
756  //===--------------------------------------------------------------------===//
757  // Inline Asm Support hooks
758  //
759
760  enum ConstraintType {
761    C_Register,            // Constraint represents a single register.
762    C_RegisterClass,       // Constraint represents one or more registers.
763    C_Memory,              // Memory constraint.
764    C_Other,               // Something else.
765    C_Unknown              // Unsupported constraint.
766  };
767
768  /// getConstraintType - Given a constraint letter, return the type of
769  /// constraint it is for this target.
770  virtual ConstraintType getConstraintType(char ConstraintLetter) const;
771
772
773  /// getRegClassForInlineAsmConstraint - Given a constraint letter (e.g. "r"),
774  /// return a list of registers that can be used to satisfy the constraint.
775  /// This should only be used for C_RegisterClass constraints.
776  virtual std::vector<unsigned>
777  getRegClassForInlineAsmConstraint(const std::string &Constraint,
778                                    MVT::ValueType VT) const;
779
780  /// getRegForInlineAsmConstraint - Given a physical register constraint (e.g.
781  /// {edx}), return the register number and the register class for the
782  /// register.
783  ///
784  /// Given a register class constraint, like 'r', if this corresponds directly
785  /// to an LLVM register class, return a register of 0 and the register class
786  /// pointer.
787  ///
788  /// This should only be used for C_Register constraints.  On error,
789  /// this returns a register number of 0 and a null register class pointer..
790  virtual std::pair<unsigned, const TargetRegisterClass*>
791    getRegForInlineAsmConstraint(const std::string &Constraint,
792                                 MVT::ValueType VT) const;
793
794
795  /// isOperandValidForConstraint - Return the specified operand (possibly
796  /// modified) if the specified SDOperand is valid for the specified target
797  /// constraint letter, otherwise return null.
798  virtual SDOperand
799    isOperandValidForConstraint(SDOperand Op, char ConstraintLetter,
800                                SelectionDAG &DAG);
801
802  //===--------------------------------------------------------------------===//
803  // Scheduler hooks
804  //
805
806  // InsertAtEndOfBasicBlock - This method should be implemented by targets that
807  // mark instructions with the 'usesCustomDAGSchedInserter' flag.  These
808  // instructions are special in various ways, which require special support to
809  // insert.  The specified MachineInstr is created but not inserted into any
810  // basic blocks, and the scheduler passes ownership of it to this method.
811  virtual MachineBasicBlock *InsertAtEndOfBasicBlock(MachineInstr *MI,
812                                                     MachineBasicBlock *MBB);
813
814  //===--------------------------------------------------------------------===//
815  // Loop Strength Reduction hooks
816  //
817
818  /// isLegalAddressImmediate - Return true if the integer value or GlobalValue
819  /// can be used as the offset of the target addressing mode.
820  virtual bool isLegalAddressImmediate(int64_t V) const;
821  virtual bool isLegalAddressImmediate(GlobalValue *GV) const;
822
823  typedef std::vector<unsigned>::const_iterator legal_am_scale_iterator;
824  legal_am_scale_iterator legal_am_scale_begin() const {
825    return LegalAddressScales.begin();
826  }
827  legal_am_scale_iterator legal_am_scale_end() const {
828    return LegalAddressScales.end();
829  }
830
831  //===--------------------------------------------------------------------===//
832  // Div utility functions
833  //
834  SDOperand BuildSDIV(SDNode *N, SelectionDAG &DAG,
835		      std::vector<SDNode*>* Created) const;
836  SDOperand BuildUDIV(SDNode *N, SelectionDAG &DAG,
837		      std::vector<SDNode*>* Created) const;
838
839
840protected:
841  /// addLegalAddressScale - Add a integer (> 1) value which can be used as
842  /// scale in the target addressing mode. Note: the ordering matters so the
843  /// least efficient ones should be entered first.
844  void addLegalAddressScale(unsigned Scale) {
845    LegalAddressScales.push_back(Scale);
846  }
847
848private:
849  std::vector<unsigned> LegalAddressScales;
850
851  TargetMachine &TM;
852  const TargetData *TD;
853
854  /// IsLittleEndian - True if this is a little endian target.
855  ///
856  bool IsLittleEndian;
857
858  /// PointerTy - The type to use for pointers, usually i32 or i64.
859  ///
860  MVT::ValueType PointerTy;
861
862  /// UsesGlobalOffsetTable - True if this target uses a GOT for PIC codegen.
863  ///
864  bool UsesGlobalOffsetTable;
865
866  /// ShiftAmountTy - The type to use for shift amounts, usually i8 or whatever
867  /// PointerTy is.
868  MVT::ValueType ShiftAmountTy;
869
870  OutOfRangeShiftAmount ShiftAmtHandling;
871
872  /// SetCCIsExpensive - This is a short term hack for targets that codegen
873  /// setcc as a conditional branch.  This encourages the code generator to fold
874  /// setcc operations into other operations if possible.
875  bool SetCCIsExpensive;
876
877  /// IntDivIsCheap - Tells the code generator not to expand integer divides by
878  /// constants into a sequence of muls, adds, and shifts.  This is a hack until
879  /// a real cost model is in place.  If we ever optimize for size, this will be
880  /// set to true unconditionally.
881  bool IntDivIsCheap;
882
883  /// Pow2DivIsCheap - Tells the code generator that it shouldn't generate
884  /// srl/add/sra for a signed divide by power of two, and let the target handle
885  /// it.
886  bool Pow2DivIsCheap;
887
888  /// SetCCResultTy - The type that SetCC operations use.  This defaults to the
889  /// PointerTy.
890  MVT::ValueType SetCCResultTy;
891
892  /// SetCCResultContents - Information about the contents of the high-bits in
893  /// the result of a setcc comparison operation.
894  SetCCResultValue SetCCResultContents;
895
896  /// SchedPreferenceInfo - The target scheduling preference: shortest possible
897  /// total cycles or lowest register usage.
898  SchedPreference SchedPreferenceInfo;
899
900  /// UseUnderscoreSetJmp - This target prefers to use _setjmp to implement
901  /// llvm.setjmp.  Defaults to false.
902  bool UseUnderscoreSetJmp;
903
904  /// UseUnderscoreLongJmp - This target prefers to use _longjmp to implement
905  /// llvm.longjmp.  Defaults to false.
906  bool UseUnderscoreLongJmp;
907
908  /// JumpBufSize - The size, in bytes, of the target's jmp_buf buffers
909  unsigned JumpBufSize;
910
911  /// JumpBufAlignment - The alignment, in bytes, of the target's jmp_buf
912  /// buffers
913  unsigned JumpBufAlignment;
914
915  /// StackPointerRegisterToSaveRestore - If set to a physical register, this
916  /// specifies the register that llvm.savestack/llvm.restorestack should save
917  /// and restore.
918  unsigned StackPointerRegisterToSaveRestore;
919
920  /// RegClassForVT - This indicates the default register class to use for
921  /// each ValueType the target supports natively.
922  TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE];
923  unsigned char NumElementsForVT[MVT::LAST_VALUETYPE];
924
925  /// TransformToType - For any value types we are promoting or expanding, this
926  /// contains the value type that we are changing to.  For Expanded types, this
927  /// contains one step of the expand (e.g. i64 -> i32), even if there are
928  /// multiple steps required (e.g. i64 -> i16).  For types natively supported
929  /// by the system, this holds the same type (e.g. i32 -> i32).
930  MVT::ValueType TransformToType[MVT::LAST_VALUETYPE];
931
932  /// OpActions - For each operation and each value type, keep a LegalizeAction
933  /// that indicates how instruction selection should deal with the operation.
934  /// Most operations are Legal (aka, supported natively by the target), but
935  /// operations that are not should be described.  Note that operations on
936  /// non-legal value types are not described here.
937  uint64_t OpActions[156];
938
939  /// LoadXActions - For each load of load extension type and each value type,
940  /// keep a LegalizeAction that indicates how instruction selection should deal
941  /// with the load.
942  uint64_t LoadXActions[ISD::LAST_LOADX_TYPE];
943
944  /// StoreXActions - For each store with truncation of each value type, keep a
945  /// LegalizeAction that indicates how instruction selection should deal with
946  /// the store.
947  uint64_t StoreXActions;
948
949  /// IndexedModeActions - For each indexed mode and each value type, keep a
950  /// pair of LegalizeAction that indicates how instruction selection should
951  /// deal with the load / store.
952  uint64_t IndexedModeActions[2][ISD::LAST_INDEXED_MODE];
953
954  ValueTypeActionImpl ValueTypeActions;
955
956  std::vector<double> LegalFPImmediates;
957
958  std::vector<std::pair<MVT::ValueType,
959                        TargetRegisterClass*> > AvailableRegClasses;
960
961  /// TargetDAGCombineArray - Targets can specify ISD nodes that they would
962  /// like PerformDAGCombine callbacks for by calling setTargetDAGCombine(),
963  /// which sets a bit in this array.
964  unsigned char TargetDAGCombineArray[156/(sizeof(unsigned char)*8)];
965
966  /// PromoteToType - For operations that must be promoted to a specific type,
967  /// this holds the destination type.  This map should be sparse, so don't hold
968  /// it as an array.
969  ///
970  /// Targets add entries to this map with AddPromotedToType(..), clients access
971  /// this with getTypeToPromoteTo(..).
972  std::map<std::pair<unsigned, MVT::ValueType>, MVT::ValueType> PromoteToType;
973
974protected:
975  /// When lowering %llvm.memset this field specifies the maximum number of
976  /// store operations that may be substituted for the call to memset. Targets
977  /// must set this value based on the cost threshold for that target. Targets
978  /// should assume that the memset will be done using as many of the largest
979  /// store operations first, followed by smaller ones, if necessary, per
980  /// alignment restrictions. For example, storing 9 bytes on a 32-bit machine
981  /// with 16-bit alignment would result in four 2-byte stores and one 1-byte
982  /// store.  This only applies to setting a constant array of a constant size.
983  /// @brief Specify maximum number of store instructions per memset call.
984  unsigned maxStoresPerMemset;
985
986  /// When lowering %llvm.memcpy this field specifies the maximum number of
987  /// store operations that may be substituted for a call to memcpy. Targets
988  /// must set this value based on the cost threshold for that target. Targets
989  /// should assume that the memcpy will be done using as many of the largest
990  /// store operations first, followed by smaller ones, if necessary, per
991  /// alignment restrictions. For example, storing 7 bytes on a 32-bit machine
992  /// with 32-bit alignment would result in one 4-byte store, a one 2-byte store
993  /// and one 1-byte store. This only applies to copying a constant array of
994  /// constant size.
995  /// @brief Specify maximum bytes of store instructions per memcpy call.
996  unsigned maxStoresPerMemcpy;
997
998  /// When lowering %llvm.memmove this field specifies the maximum number of
999  /// store instructions that may be substituted for a call to memmove. Targets
1000  /// must set this value based on the cost threshold for that target. Targets
1001  /// should assume that the memmove will be done using as many of the largest
1002  /// store operations first, followed by smaller ones, if necessary, per
1003  /// alignment restrictions. For example, moving 9 bytes on a 32-bit machine
1004  /// with 8-bit alignment would result in nine 1-byte stores.  This only
1005  /// applies to copying a constant array of constant size.
1006  /// @brief Specify maximum bytes of store instructions per memmove call.
1007  unsigned maxStoresPerMemmove;
1008
1009  /// This field specifies whether the target machine permits unaligned memory
1010  /// accesses.  This is used, for example, to determine the size of store
1011  /// operations when copying small arrays and other similar tasks.
1012  /// @brief Indicate whether the target permits unaligned memory accesses.
1013  bool allowUnalignedMemoryAccesses;
1014};
1015} // end llvm namespace
1016
1017#endif
1018