SelectionDAGBuilder.h revision 207618
1//===-- SelectionDAGBuilder.h - Selection-DAG building --------------------===//
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 implements routines for translating from LLVM IR into SelectionDAG IR.
11//
12//===----------------------------------------------------------------------===//
13
14#ifndef SELECTIONDAGBUILDER_H
15#define SELECTIONDAGBUILDER_H
16
17#include "llvm/Constants.h"
18#include "llvm/CodeGen/SelectionDAG.h"
19#include "llvm/ADT/APInt.h"
20#include "llvm/ADT/DenseMap.h"
21#ifndef NDEBUG
22#include "llvm/ADT/SmallSet.h"
23#endif
24#include "llvm/CodeGen/SelectionDAGNodes.h"
25#include "llvm/CodeGen/ValueTypes.h"
26#include "llvm/Support/CallSite.h"
27#include "llvm/Support/ErrorHandling.h"
28#include <vector>
29#include <set>
30
31namespace llvm {
32
33class AliasAnalysis;
34class AllocaInst;
35class BasicBlock;
36class BitCastInst;
37class BranchInst;
38class CallInst;
39class DbgValueInst;
40class ExtractElementInst;
41class ExtractValueInst;
42class FCmpInst;
43class FPExtInst;
44class FPToSIInst;
45class FPToUIInst;
46class FPTruncInst;
47class Function;
48class FunctionLoweringInfo;
49class GetElementPtrInst;
50class GCFunctionInfo;
51class ICmpInst;
52class IntToPtrInst;
53class IndirectBrInst;
54class InvokeInst;
55class InsertElementInst;
56class InsertValueInst;
57class Instruction;
58class LoadInst;
59class MachineBasicBlock;
60class MachineInstr;
61class MachineRegisterInfo;
62class MDNode;
63class PHINode;
64class PtrToIntInst;
65class ReturnInst;
66class SDISelAsmOperandInfo;
67class SExtInst;
68class SelectInst;
69class ShuffleVectorInst;
70class SIToFPInst;
71class StoreInst;
72class SwitchInst;
73class TargetData;
74class TargetLowering;
75class TruncInst;
76class UIToFPInst;
77class UnreachableInst;
78class UnwindInst;
79class VAArgInst;
80class ZExtInst;
81
82//===----------------------------------------------------------------------===//
83/// SelectionDAGBuilder - This is the common target-independent lowering
84/// implementation that is parameterized by a TargetLowering object.
85///
86class SelectionDAGBuilder {
87  /// CurDebugLoc - current file + line number.  Changes as we build the DAG.
88  DebugLoc CurDebugLoc;
89
90  DenseMap<const Value*, SDValue> NodeMap;
91
92public:
93  /// PendingLoads - Loads are not emitted to the program immediately.  We bunch
94  /// them up and then emit token factor nodes when possible.  This allows us to
95  /// get simple disambiguation between loads without worrying about alias
96  /// analysis.
97  SmallVector<SDValue, 8> PendingLoads;
98private:
99
100  /// PendingExports - CopyToReg nodes that copy values to virtual registers
101  /// for export to other blocks need to be emitted before any terminator
102  /// instruction, but they have no other ordering requirements. We bunch them
103  /// up and the emit a single tokenfactor for them just before terminator
104  /// instructions.
105  SmallVector<SDValue, 8> PendingExports;
106
107  /// SDNodeOrder - A unique monotonically increasing number used to order the
108  /// SDNodes we create.
109  unsigned SDNodeOrder;
110
111  /// Case - A struct to record the Value for a switch case, and the
112  /// case's target basic block.
113  struct Case {
114    Constant* Low;
115    Constant* High;
116    MachineBasicBlock* BB;
117
118    Case() : Low(0), High(0), BB(0) { }
119    Case(Constant* low, Constant* high, MachineBasicBlock* bb) :
120      Low(low), High(high), BB(bb) { }
121    APInt size() const {
122      const APInt &rHigh = cast<ConstantInt>(High)->getValue();
123      const APInt &rLow  = cast<ConstantInt>(Low)->getValue();
124      return (rHigh - rLow + 1ULL);
125    }
126  };
127
128  struct CaseBits {
129    uint64_t Mask;
130    MachineBasicBlock* BB;
131    unsigned Bits;
132
133    CaseBits(uint64_t mask, MachineBasicBlock* bb, unsigned bits):
134      Mask(mask), BB(bb), Bits(bits) { }
135  };
136
137  typedef std::vector<Case>           CaseVector;
138  typedef std::vector<CaseBits>       CaseBitsVector;
139  typedef CaseVector::iterator        CaseItr;
140  typedef std::pair<CaseItr, CaseItr> CaseRange;
141
142  /// CaseRec - A struct with ctor used in lowering switches to a binary tree
143  /// of conditional branches.
144  struct CaseRec {
145    CaseRec(MachineBasicBlock *bb, const Constant *lt, const Constant *ge,
146            CaseRange r) :
147    CaseBB(bb), LT(lt), GE(ge), Range(r) {}
148
149    /// CaseBB - The MBB in which to emit the compare and branch
150    MachineBasicBlock *CaseBB;
151    /// LT, GE - If nonzero, we know the current case value must be less-than or
152    /// greater-than-or-equal-to these Constants.
153    const Constant *LT;
154    const Constant *GE;
155    /// Range - A pair of iterators representing the range of case values to be
156    /// processed at this point in the binary search tree.
157    CaseRange Range;
158  };
159
160  typedef std::vector<CaseRec> CaseRecVector;
161
162  /// The comparison function for sorting the switch case values in the vector.
163  /// WARNING: Case ranges should be disjoint!
164  struct CaseCmp {
165    bool operator()(const Case &C1, const Case &C2) {
166      assert(isa<ConstantInt>(C1.Low) && isa<ConstantInt>(C2.High));
167      const ConstantInt* CI1 = cast<const ConstantInt>(C1.Low);
168      const ConstantInt* CI2 = cast<const ConstantInt>(C2.High);
169      return CI1->getValue().slt(CI2->getValue());
170    }
171  };
172
173  struct CaseBitsCmp {
174    bool operator()(const CaseBits &C1, const CaseBits &C2) {
175      return C1.Bits > C2.Bits;
176    }
177  };
178
179  size_t Clusterify(CaseVector &Cases, const SwitchInst &SI);
180
181  /// CaseBlock - This structure is used to communicate between
182  /// SelectionDAGBuilder and SDISel for the code generation of additional basic
183  /// blocks needed by multi-case switch statements.
184  struct CaseBlock {
185    CaseBlock(ISD::CondCode cc, const Value *cmplhs, const Value *cmprhs,
186              const Value *cmpmiddle,
187              MachineBasicBlock *truebb, MachineBasicBlock *falsebb,
188              MachineBasicBlock *me)
189      : CC(cc), CmpLHS(cmplhs), CmpMHS(cmpmiddle), CmpRHS(cmprhs),
190        TrueBB(truebb), FalseBB(falsebb), ThisBB(me) {}
191    // CC - the condition code to use for the case block's setcc node
192    ISD::CondCode CC;
193    // CmpLHS/CmpRHS/CmpMHS - The LHS/MHS/RHS of the comparison to emit.
194    // Emit by default LHS op RHS. MHS is used for range comparisons:
195    // If MHS is not null: (LHS <= MHS) and (MHS <= RHS).
196    const Value *CmpLHS, *CmpMHS, *CmpRHS;
197    // TrueBB/FalseBB - the block to branch to if the setcc is true/false.
198    MachineBasicBlock *TrueBB, *FalseBB;
199    // ThisBB - the block into which to emit the code for the setcc and branches
200    MachineBasicBlock *ThisBB;
201  };
202  struct JumpTable {
203    JumpTable(unsigned R, unsigned J, MachineBasicBlock *M,
204              MachineBasicBlock *D): Reg(R), JTI(J), MBB(M), Default(D) {}
205
206    /// Reg - the virtual register containing the index of the jump table entry
207    //. to jump to.
208    unsigned Reg;
209    /// JTI - the JumpTableIndex for this jump table in the function.
210    unsigned JTI;
211    /// MBB - the MBB into which to emit the code for the indirect jump.
212    MachineBasicBlock *MBB;
213    /// Default - the MBB of the default bb, which is a successor of the range
214    /// check MBB.  This is when updating PHI nodes in successors.
215    MachineBasicBlock *Default;
216  };
217  struct JumpTableHeader {
218    JumpTableHeader(APInt F, APInt L, const Value *SV, MachineBasicBlock *H,
219                    bool E = false):
220      First(F), Last(L), SValue(SV), HeaderBB(H), Emitted(E) {}
221    APInt First;
222    APInt Last;
223    const Value *SValue;
224    MachineBasicBlock *HeaderBB;
225    bool Emitted;
226  };
227  typedef std::pair<JumpTableHeader, JumpTable> JumpTableBlock;
228
229  struct BitTestCase {
230    BitTestCase(uint64_t M, MachineBasicBlock* T, MachineBasicBlock* Tr):
231      Mask(M), ThisBB(T), TargetBB(Tr) { }
232    uint64_t Mask;
233    MachineBasicBlock *ThisBB;
234    MachineBasicBlock *TargetBB;
235  };
236
237  typedef SmallVector<BitTestCase, 3> BitTestInfo;
238
239  struct BitTestBlock {
240    BitTestBlock(APInt F, APInt R, const Value* SV,
241                 unsigned Rg, bool E,
242                 MachineBasicBlock* P, MachineBasicBlock* D,
243                 const BitTestInfo& C):
244      First(F), Range(R), SValue(SV), Reg(Rg), Emitted(E),
245      Parent(P), Default(D), Cases(C) { }
246    APInt First;
247    APInt Range;
248    const Value *SValue;
249    unsigned Reg;
250    bool Emitted;
251    MachineBasicBlock *Parent;
252    MachineBasicBlock *Default;
253    BitTestInfo Cases;
254  };
255
256public:
257  // TLI - This is information that describes the available target features we
258  // need for lowering.  This indicates when operations are unavailable,
259  // implemented with a libcall, etc.
260  const TargetMachine &TM;
261  const TargetLowering &TLI;
262  SelectionDAG &DAG;
263  const TargetData *TD;
264  AliasAnalysis *AA;
265
266  /// SwitchCases - Vector of CaseBlock structures used to communicate
267  /// SwitchInst code generation information.
268  std::vector<CaseBlock> SwitchCases;
269  /// JTCases - Vector of JumpTable structures used to communicate
270  /// SwitchInst code generation information.
271  std::vector<JumpTableBlock> JTCases;
272  /// BitTestCases - Vector of BitTestBlock structures used to communicate
273  /// SwitchInst code generation information.
274  std::vector<BitTestBlock> BitTestCases;
275
276  // Emit PHI-node-operand constants only once even if used by multiple
277  // PHI nodes.
278  DenseMap<const Constant *, unsigned> ConstantsOut;
279
280  /// FuncInfo - Information about the function as a whole.
281  ///
282  FunctionLoweringInfo &FuncInfo;
283
284  /// OptLevel - What optimization level we're generating code for.
285  ///
286  CodeGenOpt::Level OptLevel;
287
288  /// GFI - Garbage collection metadata for the function.
289  GCFunctionInfo *GFI;
290
291  /// HasTailCall - This is set to true if a call in the current
292  /// block has been translated as a tail call. In this case,
293  /// no subsequent DAG nodes should be created.
294  ///
295  bool HasTailCall;
296
297  LLVMContext *Context;
298
299  SelectionDAGBuilder(SelectionDAG &dag, FunctionLoweringInfo &funcinfo,
300                      CodeGenOpt::Level ol)
301    : SDNodeOrder(0), TM(dag.getTarget()), TLI(dag.getTargetLoweringInfo()),
302      DAG(dag), FuncInfo(funcinfo), OptLevel(ol),
303      HasTailCall(false), Context(dag.getContext()) {
304  }
305
306  void init(GCFunctionInfo *gfi, AliasAnalysis &aa);
307
308  /// clear - Clear out the current SelectionDAG and the associated
309  /// state and prepare this SelectionDAGBuilder object to be used
310  /// for a new block. This doesn't clear out information about
311  /// additional blocks that are needed to complete switch lowering
312  /// or PHI node updating; that information is cleared out as it is
313  /// consumed.
314  void clear();
315
316  /// getRoot - Return the current virtual root of the Selection DAG,
317  /// flushing any PendingLoad items. This must be done before emitting
318  /// a store or any other node that may need to be ordered after any
319  /// prior load instructions.
320  ///
321  SDValue getRoot();
322
323  /// getControlRoot - Similar to getRoot, but instead of flushing all the
324  /// PendingLoad items, flush all the PendingExports items. It is necessary
325  /// to do this before emitting a terminator instruction.
326  ///
327  SDValue getControlRoot();
328
329  DebugLoc getCurDebugLoc() const { return CurDebugLoc; }
330
331  unsigned getSDNodeOrder() const { return SDNodeOrder; }
332
333  void CopyValueToVirtualRegister(const Value *V, unsigned Reg);
334
335  /// AssignOrderingToNode - Assign an ordering to the node. The order is gotten
336  /// from how the code appeared in the source. The ordering is used by the
337  /// scheduler to effectively turn off scheduling.
338  void AssignOrderingToNode(const SDNode *Node);
339
340  void visit(const Instruction &I);
341
342  void visit(unsigned Opcode, const User &I);
343
344  SDValue getValue(const Value *V);
345
346  void setValue(const Value *V, SDValue NewN) {
347    SDValue &N = NodeMap[V];
348    assert(N.getNode() == 0 && "Already set a value for this node!");
349    N = NewN;
350  }
351
352  void GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
353                            std::set<unsigned> &OutputRegs,
354                            std::set<unsigned> &InputRegs);
355
356  void FindMergedConditions(const Value *Cond, MachineBasicBlock *TBB,
357                            MachineBasicBlock *FBB, MachineBasicBlock *CurBB,
358                            MachineBasicBlock *SwitchBB, unsigned Opc);
359  void EmitBranchForMergedCondition(const Value *Cond, MachineBasicBlock *TBB,
360                                    MachineBasicBlock *FBB,
361                                    MachineBasicBlock *CurBB,
362                                    MachineBasicBlock *SwitchBB);
363  bool ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases);
364  bool isExportableFromCurrentBlock(const Value *V, const BasicBlock *FromBB);
365  void CopyToExportRegsIfNeeded(const Value *V);
366  void ExportFromCurrentBlock(const Value *V);
367  void LowerCallTo(ImmutableCallSite CS, SDValue Callee, bool IsTailCall,
368                   MachineBasicBlock *LandingPad = NULL);
369
370private:
371  // Terminator instructions.
372  void visitRet(const ReturnInst &I);
373  void visitBr(const BranchInst &I);
374  void visitSwitch(const SwitchInst &I);
375  void visitIndirectBr(const IndirectBrInst &I);
376  void visitUnreachable(const UnreachableInst &I) { /* noop */ }
377
378  // Helpers for visitSwitch
379  bool handleSmallSwitchRange(CaseRec& CR,
380                              CaseRecVector& WorkList,
381                              const Value* SV,
382                              MachineBasicBlock* Default,
383                              MachineBasicBlock *SwitchBB);
384  bool handleJTSwitchCase(CaseRec& CR,
385                          CaseRecVector& WorkList,
386                          const Value* SV,
387                          MachineBasicBlock* Default,
388                          MachineBasicBlock *SwitchBB);
389  bool handleBTSplitSwitchCase(CaseRec& CR,
390                               CaseRecVector& WorkList,
391                               const Value* SV,
392                               MachineBasicBlock* Default,
393                               MachineBasicBlock *SwitchBB);
394  bool handleBitTestsSwitchCase(CaseRec& CR,
395                                CaseRecVector& WorkList,
396                                const Value* SV,
397                                MachineBasicBlock* Default,
398                                MachineBasicBlock *SwitchBB);
399public:
400  void visitSwitchCase(CaseBlock &CB,
401                       MachineBasicBlock *SwitchBB);
402  void visitBitTestHeader(BitTestBlock &B, MachineBasicBlock *SwitchBB);
403  void visitBitTestCase(MachineBasicBlock* NextMBB,
404                        unsigned Reg,
405                        BitTestCase &B,
406                        MachineBasicBlock *SwitchBB);
407  void visitJumpTable(JumpTable &JT);
408  void visitJumpTableHeader(JumpTable &JT, JumpTableHeader &JTH,
409                            MachineBasicBlock *SwitchBB);
410
411private:
412  // These all get lowered before this pass.
413  void visitInvoke(const InvokeInst &I);
414  void visitUnwind(const UnwindInst &I);
415
416  void visitBinary(const User &I, unsigned OpCode);
417  void visitShift(const User &I, unsigned Opcode);
418  void visitAdd(const User &I)  { visitBinary(I, ISD::ADD); }
419  void visitFAdd(const User &I) { visitBinary(I, ISD::FADD); }
420  void visitSub(const User &I)  { visitBinary(I, ISD::SUB); }
421  void visitFSub(const User &I);
422  void visitMul(const User &I)  { visitBinary(I, ISD::MUL); }
423  void visitFMul(const User &I) { visitBinary(I, ISD::FMUL); }
424  void visitURem(const User &I) { visitBinary(I, ISD::UREM); }
425  void visitSRem(const User &I) { visitBinary(I, ISD::SREM); }
426  void visitFRem(const User &I) { visitBinary(I, ISD::FREM); }
427  void visitUDiv(const User &I) { visitBinary(I, ISD::UDIV); }
428  void visitSDiv(const User &I) { visitBinary(I, ISD::SDIV); }
429  void visitFDiv(const User &I) { visitBinary(I, ISD::FDIV); }
430  void visitAnd (const User &I) { visitBinary(I, ISD::AND); }
431  void visitOr  (const User &I) { visitBinary(I, ISD::OR); }
432  void visitXor (const User &I) { visitBinary(I, ISD::XOR); }
433  void visitShl (const User &I) { visitShift(I, ISD::SHL); }
434  void visitLShr(const User &I) { visitShift(I, ISD::SRL); }
435  void visitAShr(const User &I) { visitShift(I, ISD::SRA); }
436  void visitICmp(const User &I);
437  void visitFCmp(const User &I);
438  // Visit the conversion instructions
439  void visitTrunc(const User &I);
440  void visitZExt(const User &I);
441  void visitSExt(const User &I);
442  void visitFPTrunc(const User &I);
443  void visitFPExt(const User &I);
444  void visitFPToUI(const User &I);
445  void visitFPToSI(const User &I);
446  void visitUIToFP(const User &I);
447  void visitSIToFP(const User &I);
448  void visitPtrToInt(const User &I);
449  void visitIntToPtr(const User &I);
450  void visitBitCast(const User &I);
451
452  void visitExtractElement(const User &I);
453  void visitInsertElement(const User &I);
454  void visitShuffleVector(const User &I);
455
456  void visitExtractValue(const ExtractValueInst &I);
457  void visitInsertValue(const InsertValueInst &I);
458
459  void visitGetElementPtr(const User &I);
460  void visitSelect(const User &I);
461
462  void visitAlloca(const AllocaInst &I);
463  void visitLoad(const LoadInst &I);
464  void visitStore(const StoreInst &I);
465  void visitPHI(const PHINode &I);
466  void visitCall(const CallInst &I);
467  bool visitMemCmpCall(const CallInst &I);
468
469  void visitInlineAsm(ImmutableCallSite CS);
470  const char *visitIntrinsicCall(const CallInst &I, unsigned Intrinsic);
471  void visitTargetIntrinsic(const CallInst &I, unsigned Intrinsic);
472
473  void visitPow(const CallInst &I);
474  void visitExp2(const CallInst &I);
475  void visitExp(const CallInst &I);
476  void visitLog(const CallInst &I);
477  void visitLog2(const CallInst &I);
478  void visitLog10(const CallInst &I);
479
480  void visitVAStart(const CallInst &I);
481  void visitVAArg(const VAArgInst &I);
482  void visitVAEnd(const CallInst &I);
483  void visitVACopy(const CallInst &I);
484
485  void visitUserOp1(const Instruction &I) {
486    llvm_unreachable("UserOp1 should not exist at instruction selection time!");
487  }
488  void visitUserOp2(const Instruction &I) {
489    llvm_unreachable("UserOp2 should not exist at instruction selection time!");
490  }
491
492  const char *implVisitBinaryAtomic(const CallInst& I, ISD::NodeType Op);
493  const char *implVisitAluOverflow(const CallInst &I, ISD::NodeType Op);
494
495  void HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB);
496
497  /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a
498  /// function argument, create the corresponding DBG_VALUE machine instruction
499  /// for it now. At the end of instruction selection, they will be inserted to
500  /// the entry BB.
501  bool EmitFuncArgumentDbgValue(const DbgValueInst &DI,
502                                const Value *V, MDNode *Variable,
503                                uint64_t Offset, const SDValue &N);
504};
505
506} // end namespace llvm
507
508#endif
509