SCCP.cpp revision ecd94c804a563f2a86572dcf1d2e81f397e19daa
1//===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===//
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 implements sparse conditional constant propagation and merging:
11//
12// Specifically, this:
13//   * Assumes values are constant unless proven otherwise
14//   * Assumes BasicBlocks are dead unless proven otherwise
15//   * Proves values to be constant, and replaces them with constants
16//   * Proves conditional branches to be unconditional
17//
18// Notice that:
19//   * This pass has a habit of making definitions be dead.  It is a good idea
20//     to to run a DCE pass sometime after running this pass.
21//
22//===----------------------------------------------------------------------===//
23
24#define DEBUG_TYPE "sccp"
25#include "llvm/Transforms/Scalar.h"
26#include "llvm/Transforms/IPO.h"
27#include "llvm/Constants.h"
28#include "llvm/DerivedTypes.h"
29#include "llvm/Instructions.h"
30#include "llvm/Pass.h"
31#include "llvm/Analysis/ConstantFolding.h"
32#include "llvm/Transforms/Utils/Local.h"
33#include "llvm/Support/CallSite.h"
34#include "llvm/Support/Compiler.h"
35#include "llvm/Support/Debug.h"
36#include "llvm/Support/InstVisitor.h"
37#include "llvm/ADT/DenseMap.h"
38#include "llvm/ADT/SmallSet.h"
39#include "llvm/ADT/SmallVector.h"
40#include "llvm/ADT/Statistic.h"
41#include "llvm/ADT/STLExtras.h"
42#include <algorithm>
43using namespace llvm;
44
45STATISTIC(NumInstRemoved, "Number of instructions removed");
46STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
47
48STATISTIC(IPNumInstRemoved, "Number ofinstructions removed by IPSCCP");
49STATISTIC(IPNumDeadBlocks , "Number of basic blocks unreachable by IPSCCP");
50STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
51STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
52
53namespace {
54/// LatticeVal class - This class represents the different lattice values that
55/// an LLVM value may occupy.  It is a simple class with value semantics.
56///
57class VISIBILITY_HIDDEN LatticeVal {
58  enum {
59    /// undefined - This LLVM Value has no known value yet.
60    undefined,
61
62    /// constant - This LLVM Value has a specific constant value.
63    constant,
64
65    /// forcedconstant - This LLVM Value was thought to be undef until
66    /// ResolvedUndefsIn.  This is treated just like 'constant', but if merged
67    /// with another (different) constant, it goes to overdefined, instead of
68    /// asserting.
69    forcedconstant,
70
71    /// overdefined - This instruction is not known to be constant, and we know
72    /// it has a value.
73    overdefined
74  } LatticeValue;    // The current lattice position
75
76  Constant *ConstantVal; // If Constant value, the current value
77public:
78  inline LatticeVal() : LatticeValue(undefined), ConstantVal(0) {}
79
80  // markOverdefined - Return true if this is a new status to be in...
81  inline bool markOverdefined() {
82    if (LatticeValue != overdefined) {
83      LatticeValue = overdefined;
84      return true;
85    }
86    return false;
87  }
88
89  // markConstant - Return true if this is a new status for us.
90  inline bool markConstant(Constant *V) {
91    if (LatticeValue != constant) {
92      if (LatticeValue == undefined) {
93        LatticeValue = constant;
94        assert(V && "Marking constant with NULL");
95        ConstantVal = V;
96      } else {
97        assert(LatticeValue == forcedconstant &&
98               "Cannot move from overdefined to constant!");
99        // Stay at forcedconstant if the constant is the same.
100        if (V == ConstantVal) return false;
101
102        // Otherwise, we go to overdefined.  Assumptions made based on the
103        // forced value are possibly wrong.  Assuming this is another constant
104        // could expose a contradiction.
105        LatticeValue = overdefined;
106      }
107      return true;
108    } else {
109      assert(ConstantVal == V && "Marking constant with different value");
110    }
111    return false;
112  }
113
114  inline void markForcedConstant(Constant *V) {
115    assert(LatticeValue == undefined && "Can't force a defined value!");
116    LatticeValue = forcedconstant;
117    ConstantVal = V;
118  }
119
120  inline bool isUndefined() const { return LatticeValue == undefined; }
121  inline bool isConstant() const {
122    return LatticeValue == constant || LatticeValue == forcedconstant;
123  }
124  inline bool isOverdefined() const { return LatticeValue == overdefined; }
125
126  inline Constant *getConstant() const {
127    assert(isConstant() && "Cannot get the constant of a non-constant!");
128    return ConstantVal;
129  }
130};
131
132} // end anonymous namespace
133
134
135//===----------------------------------------------------------------------===//
136//
137/// SCCPSolver - This class is a general purpose solver for Sparse Conditional
138/// Constant Propagation.
139///
140class SCCPSolver : public InstVisitor<SCCPSolver> {
141  SmallSet<BasicBlock*, 16> BBExecutable;// The basic blocks that are executable
142  std::map<Value*, LatticeVal> ValueState;  // The state each value is in.
143
144  /// GlobalValue - If we are tracking any values for the contents of a global
145  /// variable, we keep a mapping from the constant accessor to the element of
146  /// the global, to the currently known value.  If the value becomes
147  /// overdefined, it's entry is simply removed from this map.
148  DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
149
150  /// TrackedFunctionRetVals - If we are tracking arguments into and the return
151  /// value out of a function, it will have an entry in this map, indicating
152  /// what the known return value for the function is.
153  DenseMap<Function*, LatticeVal> TrackedFunctionRetVals;
154
155  // The reason for two worklists is that overdefined is the lowest state
156  // on the lattice, and moving things to overdefined as fast as possible
157  // makes SCCP converge much faster.
158  // By having a separate worklist, we accomplish this because everything
159  // possibly overdefined will become overdefined at the soonest possible
160  // point.
161  std::vector<Value*> OverdefinedInstWorkList;
162  std::vector<Value*> InstWorkList;
163
164
165  std::vector<BasicBlock*>  BBWorkList;  // The BasicBlock work list
166
167  /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
168  /// overdefined, despite the fact that the PHI node is overdefined.
169  std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
170
171  /// KnownFeasibleEdges - Entries in this set are edges which have already had
172  /// PHI nodes retriggered.
173  typedef std::pair<BasicBlock*,BasicBlock*> Edge;
174  std::set<Edge> KnownFeasibleEdges;
175public:
176
177  /// MarkBlockExecutable - This method can be used by clients to mark all of
178  /// the blocks that are known to be intrinsically live in the processed unit.
179  void MarkBlockExecutable(BasicBlock *BB) {
180    DOUT << "Marking Block Executable: " << BB->getName() << "\n";
181    BBExecutable.insert(BB);   // Basic block is executable!
182    BBWorkList.push_back(BB);  // Add the block to the work list!
183  }
184
185  /// TrackValueOfGlobalVariable - Clients can use this method to
186  /// inform the SCCPSolver that it should track loads and stores to the
187  /// specified global variable if it can.  This is only legal to call if
188  /// performing Interprocedural SCCP.
189  void TrackValueOfGlobalVariable(GlobalVariable *GV) {
190    const Type *ElTy = GV->getType()->getElementType();
191    if (ElTy->isFirstClassType()) {
192      LatticeVal &IV = TrackedGlobals[GV];
193      if (!isa<UndefValue>(GV->getInitializer()))
194        IV.markConstant(GV->getInitializer());
195    }
196  }
197
198  /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
199  /// and out of the specified function (which cannot have its address taken),
200  /// this method must be called.
201  void AddTrackedFunction(Function *F) {
202    assert(F->hasInternalLinkage() && "Can only track internal functions!");
203    // Add an entry, F -> undef.
204    TrackedFunctionRetVals[F];
205  }
206
207  /// Solve - Solve for constants and executable blocks.
208  ///
209  void Solve();
210
211  /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
212  /// that branches on undef values cannot reach any of their successors.
213  /// However, this is not a safe assumption.  After we solve dataflow, this
214  /// method should be use to handle this.  If this returns true, the solver
215  /// should be rerun.
216  bool ResolvedUndefsIn(Function &F);
217
218  /// getExecutableBlocks - Once we have solved for constants, return the set of
219  /// blocks that is known to be executable.
220  SmallSet<BasicBlock*, 16> &getExecutableBlocks() {
221    return BBExecutable;
222  }
223
224  /// getValueMapping - Once we have solved for constants, return the mapping of
225  /// LLVM values to LatticeVals.
226  std::map<Value*, LatticeVal> &getValueMapping() {
227    return ValueState;
228  }
229
230  /// getTrackedFunctionRetVals - Get the inferred return value map.
231  ///
232  const DenseMap<Function*, LatticeVal> &getTrackedFunctionRetVals() {
233    return TrackedFunctionRetVals;
234  }
235
236  /// getTrackedGlobals - Get and return the set of inferred initializers for
237  /// global variables.
238  const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
239    return TrackedGlobals;
240  }
241
242  inline void markOverdefined(Value *V) {
243    markOverdefined(ValueState[V], V);
244  }
245
246private:
247  // markConstant - Make a value be marked as "constant".  If the value
248  // is not already a constant, add it to the instruction work list so that
249  // the users of the instruction are updated later.
250  //
251  inline void markConstant(LatticeVal &IV, Value *V, Constant *C) {
252    if (IV.markConstant(C)) {
253      DOUT << "markConstant: " << *C << ": " << *V;
254      InstWorkList.push_back(V);
255    }
256  }
257
258  inline void markForcedConstant(LatticeVal &IV, Value *V, Constant *C) {
259    IV.markForcedConstant(C);
260    DOUT << "markForcedConstant: " << *C << ": " << *V;
261    InstWorkList.push_back(V);
262  }
263
264  inline void markConstant(Value *V, Constant *C) {
265    markConstant(ValueState[V], V, C);
266  }
267
268  // markOverdefined - Make a value be marked as "overdefined". If the
269  // value is not already overdefined, add it to the overdefined instruction
270  // work list so that the users of the instruction are updated later.
271
272  inline void markOverdefined(LatticeVal &IV, Value *V) {
273    if (IV.markOverdefined()) {
274      DEBUG(DOUT << "markOverdefined: ";
275            if (Function *F = dyn_cast<Function>(V))
276              DOUT << "Function '" << F->getName() << "'\n";
277            else
278              DOUT << *V);
279      // Only instructions go on the work list
280      OverdefinedInstWorkList.push_back(V);
281    }
282  }
283
284  inline void mergeInValue(LatticeVal &IV, Value *V, LatticeVal &MergeWithV) {
285    if (IV.isOverdefined() || MergeWithV.isUndefined())
286      return;  // Noop.
287    if (MergeWithV.isOverdefined())
288      markOverdefined(IV, V);
289    else if (IV.isUndefined())
290      markConstant(IV, V, MergeWithV.getConstant());
291    else if (IV.getConstant() != MergeWithV.getConstant())
292      markOverdefined(IV, V);
293  }
294
295  inline void mergeInValue(Value *V, LatticeVal &MergeWithV) {
296    return mergeInValue(ValueState[V], V, MergeWithV);
297  }
298
299
300  // getValueState - Return the LatticeVal object that corresponds to the value.
301  // This function is necessary because not all values should start out in the
302  // underdefined state... Argument's should be overdefined, and
303  // constants should be marked as constants.  If a value is not known to be an
304  // Instruction object, then use this accessor to get its value from the map.
305  //
306  inline LatticeVal &getValueState(Value *V) {
307    std::map<Value*, LatticeVal>::iterator I = ValueState.find(V);
308    if (I != ValueState.end()) return I->second;  // Common case, in the map
309
310    if (Constant *C = dyn_cast<Constant>(V)) {
311      if (isa<UndefValue>(V)) {
312        // Nothing to do, remain undefined.
313      } else {
314        LatticeVal &LV = ValueState[C];
315        LV.markConstant(C);          // Constants are constant
316        return LV;
317      }
318    }
319    // All others are underdefined by default...
320    return ValueState[V];
321  }
322
323  // markEdgeExecutable - Mark a basic block as executable, adding it to the BB
324  // work list if it is not already executable...
325  //
326  void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
327    if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
328      return;  // This edge is already known to be executable!
329
330    if (BBExecutable.count(Dest)) {
331      DOUT << "Marking Edge Executable: " << Source->getName()
332           << " -> " << Dest->getName() << "\n";
333
334      // The destination is already executable, but we just made an edge
335      // feasible that wasn't before.  Revisit the PHI nodes in the block
336      // because they have potentially new operands.
337      for (BasicBlock::iterator I = Dest->begin(); isa<PHINode>(I); ++I)
338        visitPHINode(*cast<PHINode>(I));
339
340    } else {
341      MarkBlockExecutable(Dest);
342    }
343  }
344
345  // getFeasibleSuccessors - Return a vector of booleans to indicate which
346  // successors are reachable from a given terminator instruction.
347  //
348  void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
349
350  // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
351  // block to the 'To' basic block is currently feasible...
352  //
353  bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
354
355  // OperandChangedState - This method is invoked on all of the users of an
356  // instruction that was just changed state somehow....  Based on this
357  // information, we need to update the specified user of this instruction.
358  //
359  void OperandChangedState(User *U) {
360    // Only instructions use other variable values!
361    Instruction &I = cast<Instruction>(*U);
362    if (BBExecutable.count(I.getParent()))   // Inst is executable?
363      visit(I);
364  }
365
366private:
367  friend class InstVisitor<SCCPSolver>;
368
369  // visit implementations - Something changed in this instruction... Either an
370  // operand made a transition, or the instruction is newly executable.  Change
371  // the value type of I to reflect these changes if appropriate.
372  //
373  void visitPHINode(PHINode &I);
374
375  // Terminators
376  void visitReturnInst(ReturnInst &I);
377  void visitTerminatorInst(TerminatorInst &TI);
378
379  void visitCastInst(CastInst &I);
380  void visitSelectInst(SelectInst &I);
381  void visitBinaryOperator(Instruction &I);
382  void visitCmpInst(CmpInst &I);
383  void visitExtractElementInst(ExtractElementInst &I);
384  void visitInsertElementInst(InsertElementInst &I);
385  void visitShuffleVectorInst(ShuffleVectorInst &I);
386
387  // Instructions that cannot be folded away...
388  void visitStoreInst     (Instruction &I);
389  void visitLoadInst      (LoadInst &I);
390  void visitGetElementPtrInst(GetElementPtrInst &I);
391  void visitCallInst      (CallInst &I) { visitCallSite(CallSite::get(&I)); }
392  void visitInvokeInst    (InvokeInst &II) {
393    visitCallSite(CallSite::get(&II));
394    visitTerminatorInst(II);
395  }
396  void visitCallSite      (CallSite CS);
397  void visitUnwindInst    (TerminatorInst &I) { /*returns void*/ }
398  void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
399  void visitAllocationInst(Instruction &I) { markOverdefined(&I); }
400  void visitVANextInst    (Instruction &I) { markOverdefined(&I); }
401  void visitVAArgInst     (Instruction &I) { markOverdefined(&I); }
402  void visitFreeInst      (Instruction &I) { /*returns void*/ }
403
404  void visitInstruction(Instruction &I) {
405    // If a new instruction is added to LLVM that we don't handle...
406    cerr << "SCCP: Don't know how to handle: " << I;
407    markOverdefined(&I);   // Just in case
408  }
409};
410
411// getFeasibleSuccessors - Return a vector of booleans to indicate which
412// successors are reachable from a given terminator instruction.
413//
414void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
415                                       SmallVector<bool, 16> &Succs) {
416  Succs.resize(TI.getNumSuccessors());
417  if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
418    if (BI->isUnconditional()) {
419      Succs[0] = true;
420    } else {
421      LatticeVal &BCValue = getValueState(BI->getCondition());
422      if (BCValue.isOverdefined() ||
423          (BCValue.isConstant() && !isa<ConstantInt>(BCValue.getConstant()))) {
424        // Overdefined condition variables, and branches on unfoldable constant
425        // conditions, mean the branch could go either way.
426        Succs[0] = Succs[1] = true;
427      } else if (BCValue.isConstant()) {
428        // Constant condition variables mean the branch can only go a single way
429        Succs[BCValue.getConstant() == ConstantInt::getFalse()] = true;
430      }
431    }
432  } else if (isa<InvokeInst>(&TI)) {
433    // Invoke instructions successors are always executable.
434    Succs[0] = Succs[1] = true;
435  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
436    LatticeVal &SCValue = getValueState(SI->getCondition());
437    if (SCValue.isOverdefined() ||   // Overdefined condition?
438        (SCValue.isConstant() && !isa<ConstantInt>(SCValue.getConstant()))) {
439      // All destinations are executable!
440      Succs.assign(TI.getNumSuccessors(), true);
441    } else if (SCValue.isConstant()) {
442      Constant *CPV = SCValue.getConstant();
443      // Make sure to skip the "default value" which isn't a value
444      for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i) {
445        if (SI->getSuccessorValue(i) == CPV) {// Found the right branch...
446          Succs[i] = true;
447          return;
448        }
449      }
450
451      // Constant value not equal to any of the branches... must execute
452      // default branch then...
453      Succs[0] = true;
454    }
455  } else {
456    assert(0 && "SCCP: Don't know how to handle this terminator!");
457  }
458}
459
460
461// isEdgeFeasible - Return true if the control flow edge from the 'From' basic
462// block to the 'To' basic block is currently feasible...
463//
464bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
465  assert(BBExecutable.count(To) && "Dest should always be alive!");
466
467  // Make sure the source basic block is executable!!
468  if (!BBExecutable.count(From)) return false;
469
470  // Check to make sure this edge itself is actually feasible now...
471  TerminatorInst *TI = From->getTerminator();
472  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
473    if (BI->isUnconditional())
474      return true;
475    else {
476      LatticeVal &BCValue = getValueState(BI->getCondition());
477      if (BCValue.isOverdefined()) {
478        // Overdefined condition variables mean the branch could go either way.
479        return true;
480      } else if (BCValue.isConstant()) {
481        // Not branching on an evaluatable constant?
482        if (!isa<ConstantInt>(BCValue.getConstant())) return true;
483
484        // Constant condition variables mean the branch can only go a single way
485        return BI->getSuccessor(BCValue.getConstant() ==
486                                       ConstantInt::getFalse()) == To;
487      }
488      return false;
489    }
490  } else if (isa<InvokeInst>(TI)) {
491    // Invoke instructions successors are always executable.
492    return true;
493  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
494    LatticeVal &SCValue = getValueState(SI->getCondition());
495    if (SCValue.isOverdefined()) {  // Overdefined condition?
496      // All destinations are executable!
497      return true;
498    } else if (SCValue.isConstant()) {
499      Constant *CPV = SCValue.getConstant();
500      if (!isa<ConstantInt>(CPV))
501        return true;  // not a foldable constant?
502
503      // Make sure to skip the "default value" which isn't a value
504      for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i)
505        if (SI->getSuccessorValue(i) == CPV) // Found the taken branch...
506          return SI->getSuccessor(i) == To;
507
508      // Constant value not equal to any of the branches... must execute
509      // default branch then...
510      return SI->getDefaultDest() == To;
511    }
512    return false;
513  } else {
514    cerr << "Unknown terminator instruction: " << *TI;
515    abort();
516  }
517}
518
519// visit Implementations - Something changed in this instruction... Either an
520// operand made a transition, or the instruction is newly executable.  Change
521// the value type of I to reflect these changes if appropriate.  This method
522// makes sure to do the following actions:
523//
524// 1. If a phi node merges two constants in, and has conflicting value coming
525//    from different branches, or if the PHI node merges in an overdefined
526//    value, then the PHI node becomes overdefined.
527// 2. If a phi node merges only constants in, and they all agree on value, the
528//    PHI node becomes a constant value equal to that.
529// 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
530// 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
531// 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
532// 6. If a conditional branch has a value that is constant, make the selected
533//    destination executable
534// 7. If a conditional branch has a value that is overdefined, make all
535//    successors executable.
536//
537void SCCPSolver::visitPHINode(PHINode &PN) {
538  LatticeVal &PNIV = getValueState(&PN);
539  if (PNIV.isOverdefined()) {
540    // There may be instructions using this PHI node that are not overdefined
541    // themselves.  If so, make sure that they know that the PHI node operand
542    // changed.
543    std::multimap<PHINode*, Instruction*>::iterator I, E;
544    tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN);
545    if (I != E) {
546      SmallVector<Instruction*, 16> Users;
547      for (; I != E; ++I) Users.push_back(I->second);
548      while (!Users.empty()) {
549        visit(Users.back());
550        Users.pop_back();
551      }
552    }
553    return;  // Quick exit
554  }
555
556  // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
557  // and slow us down a lot.  Just mark them overdefined.
558  if (PN.getNumIncomingValues() > 64) {
559    markOverdefined(PNIV, &PN);
560    return;
561  }
562
563  // Look at all of the executable operands of the PHI node.  If any of them
564  // are overdefined, the PHI becomes overdefined as well.  If they are all
565  // constant, and they agree with each other, the PHI becomes the identical
566  // constant.  If they are constant and don't agree, the PHI is overdefined.
567  // If there are no executable operands, the PHI remains undefined.
568  //
569  Constant *OperandVal = 0;
570  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
571    LatticeVal &IV = getValueState(PN.getIncomingValue(i));
572    if (IV.isUndefined()) continue;  // Doesn't influence PHI node.
573
574    if (isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) {
575      if (IV.isOverdefined()) {   // PHI node becomes overdefined!
576        markOverdefined(PNIV, &PN);
577        return;
578      }
579
580      if (OperandVal == 0) {   // Grab the first value...
581        OperandVal = IV.getConstant();
582      } else {                // Another value is being merged in!
583        // There is already a reachable operand.  If we conflict with it,
584        // then the PHI node becomes overdefined.  If we agree with it, we
585        // can continue on.
586
587        // Check to see if there are two different constants merging...
588        if (IV.getConstant() != OperandVal) {
589          // Yes there is.  This means the PHI node is not constant.
590          // You must be overdefined poor PHI.
591          //
592          markOverdefined(PNIV, &PN);    // The PHI node now becomes overdefined
593          return;    // I'm done analyzing you
594        }
595      }
596    }
597  }
598
599  // If we exited the loop, this means that the PHI node only has constant
600  // arguments that agree with each other(and OperandVal is the constant) or
601  // OperandVal is null because there are no defined incoming arguments.  If
602  // this is the case, the PHI remains undefined.
603  //
604  if (OperandVal)
605    markConstant(PNIV, &PN, OperandVal);      // Acquire operand value
606}
607
608void SCCPSolver::visitReturnInst(ReturnInst &I) {
609  if (I.getNumOperands() == 0) return;  // Ret void
610
611  // If we are tracking the return value of this function, merge it in.
612  Function *F = I.getParent()->getParent();
613  if (F->hasInternalLinkage() && !TrackedFunctionRetVals.empty()) {
614    DenseMap<Function*, LatticeVal>::iterator TFRVI =
615      TrackedFunctionRetVals.find(F);
616    if (TFRVI != TrackedFunctionRetVals.end() &&
617        !TFRVI->second.isOverdefined()) {
618      LatticeVal &IV = getValueState(I.getOperand(0));
619      mergeInValue(TFRVI->second, F, IV);
620    }
621  }
622}
623
624
625void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
626  SmallVector<bool, 16> SuccFeasible;
627  getFeasibleSuccessors(TI, SuccFeasible);
628
629  BasicBlock *BB = TI.getParent();
630
631  // Mark all feasible successors executable...
632  for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
633    if (SuccFeasible[i])
634      markEdgeExecutable(BB, TI.getSuccessor(i));
635}
636
637void SCCPSolver::visitCastInst(CastInst &I) {
638  Value *V = I.getOperand(0);
639  LatticeVal &VState = getValueState(V);
640  if (VState.isOverdefined())          // Inherit overdefinedness of operand
641    markOverdefined(&I);
642  else if (VState.isConstant())        // Propagate constant value
643    markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
644                                           VState.getConstant(), I.getType()));
645}
646
647void SCCPSolver::visitSelectInst(SelectInst &I) {
648  LatticeVal &CondValue = getValueState(I.getCondition());
649  if (CondValue.isUndefined())
650    return;
651  if (CondValue.isConstant()) {
652    if (ConstantInt *CondCB = dyn_cast<ConstantInt>(CondValue.getConstant())){
653      mergeInValue(&I, getValueState(CondCB->getZExtValue() ? I.getTrueValue()
654                                                          : I.getFalseValue()));
655      return;
656    }
657  }
658
659  // Otherwise, the condition is overdefined or a constant we can't evaluate.
660  // See if we can produce something better than overdefined based on the T/F
661  // value.
662  LatticeVal &TVal = getValueState(I.getTrueValue());
663  LatticeVal &FVal = getValueState(I.getFalseValue());
664
665  // select ?, C, C -> C.
666  if (TVal.isConstant() && FVal.isConstant() &&
667      TVal.getConstant() == FVal.getConstant()) {
668    markConstant(&I, FVal.getConstant());
669    return;
670  }
671
672  if (TVal.isUndefined()) {  // select ?, undef, X -> X.
673    mergeInValue(&I, FVal);
674  } else if (FVal.isUndefined()) {  // select ?, X, undef -> X.
675    mergeInValue(&I, TVal);
676  } else {
677    markOverdefined(&I);
678  }
679}
680
681// Handle BinaryOperators and Shift Instructions...
682void SCCPSolver::visitBinaryOperator(Instruction &I) {
683  LatticeVal &IV = ValueState[&I];
684  if (IV.isOverdefined()) return;
685
686  LatticeVal &V1State = getValueState(I.getOperand(0));
687  LatticeVal &V2State = getValueState(I.getOperand(1));
688
689  if (V1State.isOverdefined() || V2State.isOverdefined()) {
690    // If this is an AND or OR with 0 or -1, it doesn't matter that the other
691    // operand is overdefined.
692    if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
693      LatticeVal *NonOverdefVal = 0;
694      if (!V1State.isOverdefined()) {
695        NonOverdefVal = &V1State;
696      } else if (!V2State.isOverdefined()) {
697        NonOverdefVal = &V2State;
698      }
699
700      if (NonOverdefVal) {
701        if (NonOverdefVal->isUndefined()) {
702          // Could annihilate value.
703          if (I.getOpcode() == Instruction::And)
704            markConstant(IV, &I, Constant::getNullValue(I.getType()));
705          else if (const VectorType *PT = dyn_cast<VectorType>(I.getType()))
706            markConstant(IV, &I, ConstantVector::getAllOnesValue(PT));
707          else
708            markConstant(IV, &I, ConstantInt::getAllOnesValue(I.getType()));
709          return;
710        } else {
711          if (I.getOpcode() == Instruction::And) {
712            if (NonOverdefVal->getConstant()->isNullValue()) {
713              markConstant(IV, &I, NonOverdefVal->getConstant());
714              return;      // X and 0 = 0
715            }
716          } else {
717            if (ConstantInt *CI =
718                     dyn_cast<ConstantInt>(NonOverdefVal->getConstant()))
719              if (CI->isAllOnesValue()) {
720                markConstant(IV, &I, NonOverdefVal->getConstant());
721                return;    // X or -1 = -1
722              }
723          }
724        }
725      }
726    }
727
728
729    // If both operands are PHI nodes, it is possible that this instruction has
730    // a constant value, despite the fact that the PHI node doesn't.  Check for
731    // this condition now.
732    if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
733      if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
734        if (PN1->getParent() == PN2->getParent()) {
735          // Since the two PHI nodes are in the same basic block, they must have
736          // entries for the same predecessors.  Walk the predecessor list, and
737          // if all of the incoming values are constants, and the result of
738          // evaluating this expression with all incoming value pairs is the
739          // same, then this expression is a constant even though the PHI node
740          // is not a constant!
741          LatticeVal Result;
742          for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
743            LatticeVal &In1 = getValueState(PN1->getIncomingValue(i));
744            BasicBlock *InBlock = PN1->getIncomingBlock(i);
745            LatticeVal &In2 =
746              getValueState(PN2->getIncomingValueForBlock(InBlock));
747
748            if (In1.isOverdefined() || In2.isOverdefined()) {
749              Result.markOverdefined();
750              break;  // Cannot fold this operation over the PHI nodes!
751            } else if (In1.isConstant() && In2.isConstant()) {
752              Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(),
753                                              In2.getConstant());
754              if (Result.isUndefined())
755                Result.markConstant(V);
756              else if (Result.isConstant() && Result.getConstant() != V) {
757                Result.markOverdefined();
758                break;
759              }
760            }
761          }
762
763          // If we found a constant value here, then we know the instruction is
764          // constant despite the fact that the PHI nodes are overdefined.
765          if (Result.isConstant()) {
766            markConstant(IV, &I, Result.getConstant());
767            // Remember that this instruction is virtually using the PHI node
768            // operands.
769            UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
770            UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
771            return;
772          } else if (Result.isUndefined()) {
773            return;
774          }
775
776          // Okay, this really is overdefined now.  Since we might have
777          // speculatively thought that this was not overdefined before, and
778          // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
779          // make sure to clean out any entries that we put there, for
780          // efficiency.
781          std::multimap<PHINode*, Instruction*>::iterator It, E;
782          tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1);
783          while (It != E) {
784            if (It->second == &I) {
785              UsersOfOverdefinedPHIs.erase(It++);
786            } else
787              ++It;
788          }
789          tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2);
790          while (It != E) {
791            if (It->second == &I) {
792              UsersOfOverdefinedPHIs.erase(It++);
793            } else
794              ++It;
795          }
796        }
797
798    markOverdefined(IV, &I);
799  } else if (V1State.isConstant() && V2State.isConstant()) {
800    markConstant(IV, &I, ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
801                                           V2State.getConstant()));
802  }
803}
804
805// Handle ICmpInst instruction...
806void SCCPSolver::visitCmpInst(CmpInst &I) {
807  LatticeVal &IV = ValueState[&I];
808  if (IV.isOverdefined()) return;
809
810  LatticeVal &V1State = getValueState(I.getOperand(0));
811  LatticeVal &V2State = getValueState(I.getOperand(1));
812
813  if (V1State.isOverdefined() || V2State.isOverdefined()) {
814    // If both operands are PHI nodes, it is possible that this instruction has
815    // a constant value, despite the fact that the PHI node doesn't.  Check for
816    // this condition now.
817    if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
818      if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
819        if (PN1->getParent() == PN2->getParent()) {
820          // Since the two PHI nodes are in the same basic block, they must have
821          // entries for the same predecessors.  Walk the predecessor list, and
822          // if all of the incoming values are constants, and the result of
823          // evaluating this expression with all incoming value pairs is the
824          // same, then this expression is a constant even though the PHI node
825          // is not a constant!
826          LatticeVal Result;
827          for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
828            LatticeVal &In1 = getValueState(PN1->getIncomingValue(i));
829            BasicBlock *InBlock = PN1->getIncomingBlock(i);
830            LatticeVal &In2 =
831              getValueState(PN2->getIncomingValueForBlock(InBlock));
832
833            if (In1.isOverdefined() || In2.isOverdefined()) {
834              Result.markOverdefined();
835              break;  // Cannot fold this operation over the PHI nodes!
836            } else if (In1.isConstant() && In2.isConstant()) {
837              Constant *V = ConstantExpr::getCompare(I.getPredicate(),
838                                                     In1.getConstant(),
839                                                     In2.getConstant());
840              if (Result.isUndefined())
841                Result.markConstant(V);
842              else if (Result.isConstant() && Result.getConstant() != V) {
843                Result.markOverdefined();
844                break;
845              }
846            }
847          }
848
849          // If we found a constant value here, then we know the instruction is
850          // constant despite the fact that the PHI nodes are overdefined.
851          if (Result.isConstant()) {
852            markConstant(IV, &I, Result.getConstant());
853            // Remember that this instruction is virtually using the PHI node
854            // operands.
855            UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
856            UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
857            return;
858          } else if (Result.isUndefined()) {
859            return;
860          }
861
862          // Okay, this really is overdefined now.  Since we might have
863          // speculatively thought that this was not overdefined before, and
864          // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
865          // make sure to clean out any entries that we put there, for
866          // efficiency.
867          std::multimap<PHINode*, Instruction*>::iterator It, E;
868          tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1);
869          while (It != E) {
870            if (It->second == &I) {
871              UsersOfOverdefinedPHIs.erase(It++);
872            } else
873              ++It;
874          }
875          tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2);
876          while (It != E) {
877            if (It->second == &I) {
878              UsersOfOverdefinedPHIs.erase(It++);
879            } else
880              ++It;
881          }
882        }
883
884    markOverdefined(IV, &I);
885  } else if (V1State.isConstant() && V2State.isConstant()) {
886    markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
887                                                  V1State.getConstant(),
888                                                  V2State.getConstant()));
889  }
890}
891
892void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
893  // FIXME : SCCP does not handle vectors properly.
894  markOverdefined(&I);
895  return;
896
897#if 0
898  LatticeVal &ValState = getValueState(I.getOperand(0));
899  LatticeVal &IdxState = getValueState(I.getOperand(1));
900
901  if (ValState.isOverdefined() || IdxState.isOverdefined())
902    markOverdefined(&I);
903  else if(ValState.isConstant() && IdxState.isConstant())
904    markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
905                                                     IdxState.getConstant()));
906#endif
907}
908
909void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
910  // FIXME : SCCP does not handle vectors properly.
911  markOverdefined(&I);
912  return;
913#if 0
914  LatticeVal &ValState = getValueState(I.getOperand(0));
915  LatticeVal &EltState = getValueState(I.getOperand(1));
916  LatticeVal &IdxState = getValueState(I.getOperand(2));
917
918  if (ValState.isOverdefined() || EltState.isOverdefined() ||
919      IdxState.isOverdefined())
920    markOverdefined(&I);
921  else if(ValState.isConstant() && EltState.isConstant() &&
922          IdxState.isConstant())
923    markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
924                                                    EltState.getConstant(),
925                                                    IdxState.getConstant()));
926  else if (ValState.isUndefined() && EltState.isConstant() &&
927           IdxState.isConstant())
928    markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
929                                                   EltState.getConstant(),
930                                                   IdxState.getConstant()));
931#endif
932}
933
934void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
935  // FIXME : SCCP does not handle vectors properly.
936  markOverdefined(&I);
937  return;
938#if 0
939  LatticeVal &V1State   = getValueState(I.getOperand(0));
940  LatticeVal &V2State   = getValueState(I.getOperand(1));
941  LatticeVal &MaskState = getValueState(I.getOperand(2));
942
943  if (MaskState.isUndefined() ||
944      (V1State.isUndefined() && V2State.isUndefined()))
945    return;  // Undefined output if mask or both inputs undefined.
946
947  if (V1State.isOverdefined() || V2State.isOverdefined() ||
948      MaskState.isOverdefined()) {
949    markOverdefined(&I);
950  } else {
951    // A mix of constant/undef inputs.
952    Constant *V1 = V1State.isConstant() ?
953        V1State.getConstant() : UndefValue::get(I.getType());
954    Constant *V2 = V2State.isConstant() ?
955        V2State.getConstant() : UndefValue::get(I.getType());
956    Constant *Mask = MaskState.isConstant() ?
957      MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
958    markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
959  }
960#endif
961}
962
963// Handle getelementptr instructions... if all operands are constants then we
964// can turn this into a getelementptr ConstantExpr.
965//
966void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
967  LatticeVal &IV = ValueState[&I];
968  if (IV.isOverdefined()) return;
969
970  SmallVector<Constant*, 8> Operands;
971  Operands.reserve(I.getNumOperands());
972
973  for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
974    LatticeVal &State = getValueState(I.getOperand(i));
975    if (State.isUndefined())
976      return;  // Operands are not resolved yet...
977    else if (State.isOverdefined()) {
978      markOverdefined(IV, &I);
979      return;
980    }
981    assert(State.isConstant() && "Unknown state!");
982    Operands.push_back(State.getConstant());
983  }
984
985  Constant *Ptr = Operands[0];
986  Operands.erase(Operands.begin());  // Erase the pointer from idx list...
987
988  markConstant(IV, &I, ConstantExpr::getGetElementPtr(Ptr, &Operands[0],
989                                                      Operands.size()));
990}
991
992void SCCPSolver::visitStoreInst(Instruction &SI) {
993  if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
994    return;
995  GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
996  DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
997  if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
998
999  // Get the value we are storing into the global.
1000  LatticeVal &PtrVal = getValueState(SI.getOperand(0));
1001
1002  mergeInValue(I->second, GV, PtrVal);
1003  if (I->second.isOverdefined())
1004    TrackedGlobals.erase(I);      // No need to keep tracking this!
1005}
1006
1007
1008// Handle load instructions.  If the operand is a constant pointer to a constant
1009// global, we can replace the load with the loaded constant value!
1010void SCCPSolver::visitLoadInst(LoadInst &I) {
1011  LatticeVal &IV = ValueState[&I];
1012  if (IV.isOverdefined()) return;
1013
1014  LatticeVal &PtrVal = getValueState(I.getOperand(0));
1015  if (PtrVal.isUndefined()) return;   // The pointer is not resolved yet!
1016  if (PtrVal.isConstant() && !I.isVolatile()) {
1017    Value *Ptr = PtrVal.getConstant();
1018    if (isa<ConstantPointerNull>(Ptr)) {
1019      // load null -> null
1020      markConstant(IV, &I, Constant::getNullValue(I.getType()));
1021      return;
1022    }
1023
1024    // Transform load (constant global) into the value loaded.
1025    if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1026      if (GV->isConstant()) {
1027        if (!GV->isDeclaration()) {
1028          markConstant(IV, &I, GV->getInitializer());
1029          return;
1030        }
1031      } else if (!TrackedGlobals.empty()) {
1032        // If we are tracking this global, merge in the known value for it.
1033        DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1034          TrackedGlobals.find(GV);
1035        if (It != TrackedGlobals.end()) {
1036          mergeInValue(IV, &I, It->second);
1037          return;
1038        }
1039      }
1040    }
1041
1042    // Transform load (constantexpr_GEP global, 0, ...) into the value loaded.
1043    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
1044      if (CE->getOpcode() == Instruction::GetElementPtr)
1045    if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
1046      if (GV->isConstant() && !GV->isDeclaration())
1047        if (Constant *V =
1048             ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) {
1049          markConstant(IV, &I, V);
1050          return;
1051        }
1052  }
1053
1054  // Otherwise we cannot say for certain what value this load will produce.
1055  // Bail out.
1056  markOverdefined(IV, &I);
1057}
1058
1059void SCCPSolver::visitCallSite(CallSite CS) {
1060  Function *F = CS.getCalledFunction();
1061
1062  // If we are tracking this function, we must make sure to bind arguments as
1063  // appropriate.
1064  DenseMap<Function*, LatticeVal>::iterator TFRVI =TrackedFunctionRetVals.end();
1065  if (F && F->hasInternalLinkage())
1066    TFRVI = TrackedFunctionRetVals.find(F);
1067
1068  if (TFRVI != TrackedFunctionRetVals.end()) {
1069    // If this is the first call to the function hit, mark its entry block
1070    // executable.
1071    if (!BBExecutable.count(F->begin()))
1072      MarkBlockExecutable(F->begin());
1073
1074    CallSite::arg_iterator CAI = CS.arg_begin();
1075    for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1076         AI != E; ++AI, ++CAI) {
1077      LatticeVal &IV = ValueState[AI];
1078      if (!IV.isOverdefined())
1079        mergeInValue(IV, AI, getValueState(*CAI));
1080    }
1081  }
1082  Instruction *I = CS.getInstruction();
1083  if (I->getType() == Type::VoidTy) return;
1084
1085  LatticeVal &IV = ValueState[I];
1086  if (IV.isOverdefined()) return;
1087
1088  // Propagate the return value of the function to the value of the instruction.
1089  if (TFRVI != TrackedFunctionRetVals.end()) {
1090    mergeInValue(IV, I, TFRVI->second);
1091    return;
1092  }
1093
1094  if (F == 0 || !F->isDeclaration() || !canConstantFoldCallTo(F)) {
1095    markOverdefined(IV, I);
1096    return;
1097  }
1098
1099  SmallVector<Constant*, 8> Operands;
1100  Operands.reserve(I->getNumOperands()-1);
1101
1102  for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1103       AI != E; ++AI) {
1104    LatticeVal &State = getValueState(*AI);
1105    if (State.isUndefined())
1106      return;  // Operands are not resolved yet...
1107    else if (State.isOverdefined()) {
1108      markOverdefined(IV, I);
1109      return;
1110    }
1111    assert(State.isConstant() && "Unknown state!");
1112    Operands.push_back(State.getConstant());
1113  }
1114
1115  if (Constant *C = ConstantFoldCall(F, &Operands[0], Operands.size()))
1116    markConstant(IV, I, C);
1117  else
1118    markOverdefined(IV, I);
1119}
1120
1121
1122void SCCPSolver::Solve() {
1123  // Process the work lists until they are empty!
1124  while (!BBWorkList.empty() || !InstWorkList.empty() ||
1125         !OverdefinedInstWorkList.empty()) {
1126    // Process the instruction work list...
1127    while (!OverdefinedInstWorkList.empty()) {
1128      Value *I = OverdefinedInstWorkList.back();
1129      OverdefinedInstWorkList.pop_back();
1130
1131      DOUT << "\nPopped off OI-WL: " << *I;
1132
1133      // "I" got into the work list because it either made the transition from
1134      // bottom to constant
1135      //
1136      // Anything on this worklist that is overdefined need not be visited
1137      // since all of its users will have already been marked as overdefined
1138      // Update all of the users of this instruction's value...
1139      //
1140      for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1141           UI != E; ++UI)
1142        OperandChangedState(*UI);
1143    }
1144    // Process the instruction work list...
1145    while (!InstWorkList.empty()) {
1146      Value *I = InstWorkList.back();
1147      InstWorkList.pop_back();
1148
1149      DOUT << "\nPopped off I-WL: " << *I;
1150
1151      // "I" got into the work list because it either made the transition from
1152      // bottom to constant
1153      //
1154      // Anything on this worklist that is overdefined need not be visited
1155      // since all of its users will have already been marked as overdefined.
1156      // Update all of the users of this instruction's value...
1157      //
1158      if (!getValueState(I).isOverdefined())
1159        for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1160             UI != E; ++UI)
1161          OperandChangedState(*UI);
1162    }
1163
1164    // Process the basic block work list...
1165    while (!BBWorkList.empty()) {
1166      BasicBlock *BB = BBWorkList.back();
1167      BBWorkList.pop_back();
1168
1169      DOUT << "\nPopped off BBWL: " << *BB;
1170
1171      // Notify all instructions in this basic block that they are newly
1172      // executable.
1173      visit(BB);
1174    }
1175  }
1176}
1177
1178/// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1179/// that branches on undef values cannot reach any of their successors.
1180/// However, this is not a safe assumption.  After we solve dataflow, this
1181/// method should be use to handle this.  If this returns true, the solver
1182/// should be rerun.
1183///
1184/// This method handles this by finding an unresolved branch and marking it one
1185/// of the edges from the block as being feasible, even though the condition
1186/// doesn't say it would otherwise be.  This allows SCCP to find the rest of the
1187/// CFG and only slightly pessimizes the analysis results (by marking one,
1188/// potentially infeasible, edge feasible).  This cannot usefully modify the
1189/// constraints on the condition of the branch, as that would impact other users
1190/// of the value.
1191///
1192/// This scan also checks for values that use undefs, whose results are actually
1193/// defined.  For example, 'zext i8 undef to i32' should produce all zeros
1194/// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1195/// even if X isn't defined.
1196bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1197  for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1198    if (!BBExecutable.count(BB))
1199      continue;
1200
1201    for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1202      // Look for instructions which produce undef values.
1203      if (I->getType() == Type::VoidTy) continue;
1204
1205      LatticeVal &LV = getValueState(I);
1206      if (!LV.isUndefined()) continue;
1207
1208      // Get the lattice values of the first two operands for use below.
1209      LatticeVal &Op0LV = getValueState(I->getOperand(0));
1210      LatticeVal Op1LV;
1211      if (I->getNumOperands() == 2) {
1212        // If this is a two-operand instruction, and if both operands are
1213        // undefs, the result stays undef.
1214        Op1LV = getValueState(I->getOperand(1));
1215        if (Op0LV.isUndefined() && Op1LV.isUndefined())
1216          continue;
1217      }
1218
1219      // If this is an instructions whose result is defined even if the input is
1220      // not fully defined, propagate the information.
1221      const Type *ITy = I->getType();
1222      switch (I->getOpcode()) {
1223      default: break;          // Leave the instruction as an undef.
1224      case Instruction::ZExt:
1225        // After a zero extend, we know the top part is zero.  SExt doesn't have
1226        // to be handled here, because we don't know whether the top part is 1's
1227        // or 0's.
1228        assert(Op0LV.isUndefined());
1229        markForcedConstant(LV, I, Constant::getNullValue(ITy));
1230        return true;
1231      case Instruction::Mul:
1232      case Instruction::And:
1233        // undef * X -> 0.   X could be zero.
1234        // undef & X -> 0.   X could be zero.
1235        markForcedConstant(LV, I, Constant::getNullValue(ITy));
1236        return true;
1237
1238      case Instruction::Or:
1239        // undef | X -> -1.   X could be -1.
1240        if (const VectorType *PTy = dyn_cast<VectorType>(ITy))
1241          markForcedConstant(LV, I, ConstantVector::getAllOnesValue(PTy));
1242        else
1243          markForcedConstant(LV, I, ConstantInt::getAllOnesValue(ITy));
1244        return true;
1245
1246      case Instruction::SDiv:
1247      case Instruction::UDiv:
1248      case Instruction::SRem:
1249      case Instruction::URem:
1250        // X / undef -> undef.  No change.
1251        // X % undef -> undef.  No change.
1252        if (Op1LV.isUndefined()) break;
1253
1254        // undef / X -> 0.   X could be maxint.
1255        // undef % X -> 0.   X could be 1.
1256        markForcedConstant(LV, I, Constant::getNullValue(ITy));
1257        return true;
1258
1259      case Instruction::AShr:
1260        // undef >>s X -> undef.  No change.
1261        if (Op0LV.isUndefined()) break;
1262
1263        // X >>s undef -> X.  X could be 0, X could have the high-bit known set.
1264        if (Op0LV.isConstant())
1265          markForcedConstant(LV, I, Op0LV.getConstant());
1266        else
1267          markOverdefined(LV, I);
1268        return true;
1269      case Instruction::LShr:
1270      case Instruction::Shl:
1271        // undef >> X -> undef.  No change.
1272        // undef << X -> undef.  No change.
1273        if (Op0LV.isUndefined()) break;
1274
1275        // X >> undef -> 0.  X could be 0.
1276        // X << undef -> 0.  X could be 0.
1277        markForcedConstant(LV, I, Constant::getNullValue(ITy));
1278        return true;
1279      case Instruction::Select:
1280        // undef ? X : Y  -> X or Y.  There could be commonality between X/Y.
1281        if (Op0LV.isUndefined()) {
1282          if (!Op1LV.isConstant())  // Pick the constant one if there is any.
1283            Op1LV = getValueState(I->getOperand(2));
1284        } else if (Op1LV.isUndefined()) {
1285          // c ? undef : undef -> undef.  No change.
1286          Op1LV = getValueState(I->getOperand(2));
1287          if (Op1LV.isUndefined())
1288            break;
1289          // Otherwise, c ? undef : x -> x.
1290        } else {
1291          // Leave Op1LV as Operand(1)'s LatticeValue.
1292        }
1293
1294        if (Op1LV.isConstant())
1295          markForcedConstant(LV, I, Op1LV.getConstant());
1296        else
1297          markOverdefined(LV, I);
1298        return true;
1299      }
1300    }
1301
1302    TerminatorInst *TI = BB->getTerminator();
1303    if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1304      if (!BI->isConditional()) continue;
1305      if (!getValueState(BI->getCondition()).isUndefined())
1306        continue;
1307    } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1308      if (!getValueState(SI->getCondition()).isUndefined())
1309        continue;
1310    } else {
1311      continue;
1312    }
1313
1314    // If the edge to the first successor isn't thought to be feasible yet, mark
1315    // it so now.
1316    if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(0))))
1317      continue;
1318
1319    // Otherwise, it isn't already thought to be feasible.  Mark it as such now
1320    // and return.  This will make other blocks reachable, which will allow new
1321    // values to be discovered and existing ones to be moved in the lattice.
1322    markEdgeExecutable(BB, TI->getSuccessor(0));
1323    return true;
1324  }
1325
1326  return false;
1327}
1328
1329
1330namespace {
1331  //===--------------------------------------------------------------------===//
1332  //
1333  /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1334  /// Sparse Conditional Constant Propagator.
1335  ///
1336  struct VISIBILITY_HIDDEN SCCP : public FunctionPass {
1337    static char ID; // Pass identification, replacement for typeid
1338    SCCP() : FunctionPass((intptr_t)&ID) {}
1339
1340    // runOnFunction - Run the Sparse Conditional Constant Propagation
1341    // algorithm, and return true if the function was modified.
1342    //
1343    bool runOnFunction(Function &F);
1344
1345    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1346      AU.setPreservesCFG();
1347    }
1348  };
1349
1350  char SCCP::ID = 0;
1351  RegisterPass<SCCP> X("sccp", "Sparse Conditional Constant Propagation");
1352} // end anonymous namespace
1353
1354
1355// createSCCPPass - This is the public interface to this file...
1356FunctionPass *llvm::createSCCPPass() {
1357  return new SCCP();
1358}
1359
1360
1361// runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1362// and return true if the function was modified.
1363//
1364bool SCCP::runOnFunction(Function &F) {
1365  DOUT << "SCCP on function '" << F.getName() << "'\n";
1366  SCCPSolver Solver;
1367
1368  // Mark the first block of the function as being executable.
1369  Solver.MarkBlockExecutable(F.begin());
1370
1371  // Mark all arguments to the function as being overdefined.
1372  for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1373    Solver.markOverdefined(AI);
1374
1375  // Solve for constants.
1376  bool ResolvedUndefs = true;
1377  while (ResolvedUndefs) {
1378    Solver.Solve();
1379    DOUT << "RESOLVING UNDEFs\n";
1380    ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1381  }
1382
1383  bool MadeChanges = false;
1384
1385  // If we decided that there are basic blocks that are dead in this function,
1386  // delete their contents now.  Note that we cannot actually delete the blocks,
1387  // as we cannot modify the CFG of the function.
1388  //
1389  SmallSet<BasicBlock*, 16> &ExecutableBBs = Solver.getExecutableBlocks();
1390  SmallVector<Instruction*, 32> Insts;
1391  std::map<Value*, LatticeVal> &Values = Solver.getValueMapping();
1392
1393  for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1394    if (!ExecutableBBs.count(BB)) {
1395      DOUT << "  BasicBlock Dead:" << *BB;
1396      ++NumDeadBlocks;
1397
1398      // Delete the instructions backwards, as it has a reduced likelihood of
1399      // having to update as many def-use and use-def chains.
1400      for (BasicBlock::iterator I = BB->begin(), E = BB->getTerminator();
1401           I != E; ++I)
1402        Insts.push_back(I);
1403      while (!Insts.empty()) {
1404        Instruction *I = Insts.back();
1405        Insts.pop_back();
1406        if (!I->use_empty())
1407          I->replaceAllUsesWith(UndefValue::get(I->getType()));
1408        BB->getInstList().erase(I);
1409        MadeChanges = true;
1410        ++NumInstRemoved;
1411      }
1412    } else {
1413      // Iterate over all of the instructions in a function, replacing them with
1414      // constants if we have found them to be of constant values.
1415      //
1416      for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1417        Instruction *Inst = BI++;
1418        if (Inst->getType() != Type::VoidTy) {
1419          LatticeVal &IV = Values[Inst];
1420          if (IV.isConstant() || IV.isUndefined() &&
1421              !isa<TerminatorInst>(Inst)) {
1422            Constant *Const = IV.isConstant()
1423              ? IV.getConstant() : UndefValue::get(Inst->getType());
1424            DOUT << "  Constant: " << *Const << " = " << *Inst;
1425
1426            // Replaces all of the uses of a variable with uses of the constant.
1427            Inst->replaceAllUsesWith(Const);
1428
1429            // Delete the instruction.
1430            BB->getInstList().erase(Inst);
1431
1432            // Hey, we just changed something!
1433            MadeChanges = true;
1434            ++NumInstRemoved;
1435          }
1436        }
1437      }
1438    }
1439
1440  return MadeChanges;
1441}
1442
1443namespace {
1444  //===--------------------------------------------------------------------===//
1445  //
1446  /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1447  /// Constant Propagation.
1448  ///
1449  struct VISIBILITY_HIDDEN IPSCCP : public ModulePass {
1450    static char ID;
1451    IPSCCP() : ModulePass((intptr_t)&ID) {}
1452    bool runOnModule(Module &M);
1453  };
1454
1455  char IPSCCP::ID = 0;
1456  RegisterPass<IPSCCP>
1457  Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation");
1458} // end anonymous namespace
1459
1460// createIPSCCPPass - This is the public interface to this file...
1461ModulePass *llvm::createIPSCCPPass() {
1462  return new IPSCCP();
1463}
1464
1465
1466static bool AddressIsTaken(GlobalValue *GV) {
1467  // Delete any dead constantexpr klingons.
1468  GV->removeDeadConstantUsers();
1469
1470  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
1471       UI != E; ++UI)
1472    if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1473      if (SI->getOperand(0) == GV || SI->isVolatile())
1474        return true;  // Storing addr of GV.
1475    } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) {
1476      // Make sure we are calling the function, not passing the address.
1477      CallSite CS = CallSite::get(cast<Instruction>(*UI));
1478      for (CallSite::arg_iterator AI = CS.arg_begin(),
1479             E = CS.arg_end(); AI != E; ++AI)
1480        if (*AI == GV)
1481          return true;
1482    } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1483      if (LI->isVolatile())
1484        return true;
1485    } else {
1486      return true;
1487    }
1488  return false;
1489}
1490
1491bool IPSCCP::runOnModule(Module &M) {
1492  SCCPSolver Solver;
1493
1494  // Loop over all functions, marking arguments to those with their addresses
1495  // taken or that are external as overdefined.
1496  //
1497  for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1498    if (!F->hasInternalLinkage() || AddressIsTaken(F)) {
1499      if (!F->isDeclaration())
1500        Solver.MarkBlockExecutable(F->begin());
1501      for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1502           AI != E; ++AI)
1503        Solver.markOverdefined(AI);
1504    } else {
1505      Solver.AddTrackedFunction(F);
1506    }
1507
1508  // Loop over global variables.  We inform the solver about any internal global
1509  // variables that do not have their 'addresses taken'.  If they don't have
1510  // their addresses taken, we can propagate constants through them.
1511  for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1512       G != E; ++G)
1513    if (!G->isConstant() && G->hasInternalLinkage() && !AddressIsTaken(G))
1514      Solver.TrackValueOfGlobalVariable(G);
1515
1516  // Solve for constants.
1517  bool ResolvedUndefs = true;
1518  while (ResolvedUndefs) {
1519    Solver.Solve();
1520
1521    DOUT << "RESOLVING UNDEFS\n";
1522    ResolvedUndefs = false;
1523    for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1524      ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1525  }
1526
1527  bool MadeChanges = false;
1528
1529  // Iterate over all of the instructions in the module, replacing them with
1530  // constants if we have found them to be of constant values.
1531  //
1532  SmallSet<BasicBlock*, 16> &ExecutableBBs = Solver.getExecutableBlocks();
1533  SmallVector<Instruction*, 32> Insts;
1534  SmallVector<BasicBlock*, 32> BlocksToErase;
1535  std::map<Value*, LatticeVal> &Values = Solver.getValueMapping();
1536
1537  for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1538    for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1539         AI != E; ++AI)
1540      if (!AI->use_empty()) {
1541        LatticeVal &IV = Values[AI];
1542        if (IV.isConstant() || IV.isUndefined()) {
1543          Constant *CST = IV.isConstant() ?
1544            IV.getConstant() : UndefValue::get(AI->getType());
1545          DOUT << "***  Arg " << *AI << " = " << *CST <<"\n";
1546
1547          // Replaces all of the uses of a variable with uses of the
1548          // constant.
1549          AI->replaceAllUsesWith(CST);
1550          ++IPNumArgsElimed;
1551        }
1552      }
1553
1554    for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1555      if (!ExecutableBBs.count(BB)) {
1556        DOUT << "  BasicBlock Dead:" << *BB;
1557        ++IPNumDeadBlocks;
1558
1559        // Delete the instructions backwards, as it has a reduced likelihood of
1560        // having to update as many def-use and use-def chains.
1561        TerminatorInst *TI = BB->getTerminator();
1562        for (BasicBlock::iterator I = BB->begin(), E = TI; I != E; ++I)
1563          Insts.push_back(I);
1564
1565        while (!Insts.empty()) {
1566          Instruction *I = Insts.back();
1567          Insts.pop_back();
1568          if (!I->use_empty())
1569            I->replaceAllUsesWith(UndefValue::get(I->getType()));
1570          BB->getInstList().erase(I);
1571          MadeChanges = true;
1572          ++IPNumInstRemoved;
1573        }
1574
1575        for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1576          BasicBlock *Succ = TI->getSuccessor(i);
1577          if (Succ->begin() != Succ->end() && isa<PHINode>(Succ->begin()))
1578            TI->getSuccessor(i)->removePredecessor(BB);
1579        }
1580        if (!TI->use_empty())
1581          TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1582        BB->getInstList().erase(TI);
1583
1584        if (&*BB != &F->front())
1585          BlocksToErase.push_back(BB);
1586        else
1587          new UnreachableInst(BB);
1588
1589      } else {
1590        for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1591          Instruction *Inst = BI++;
1592          if (Inst->getType() != Type::VoidTy) {
1593            LatticeVal &IV = Values[Inst];
1594            if (IV.isConstant() || IV.isUndefined() &&
1595                !isa<TerminatorInst>(Inst)) {
1596              Constant *Const = IV.isConstant()
1597                ? IV.getConstant() : UndefValue::get(Inst->getType());
1598              DOUT << "  Constant: " << *Const << " = " << *Inst;
1599
1600              // Replaces all of the uses of a variable with uses of the
1601              // constant.
1602              Inst->replaceAllUsesWith(Const);
1603
1604              // Delete the instruction.
1605              if (!isa<TerminatorInst>(Inst) && !isa<CallInst>(Inst))
1606                BB->getInstList().erase(Inst);
1607
1608              // Hey, we just changed something!
1609              MadeChanges = true;
1610              ++IPNumInstRemoved;
1611            }
1612          }
1613        }
1614      }
1615
1616    // Now that all instructions in the function are constant folded, erase dead
1617    // blocks, because we can now use ConstantFoldTerminator to get rid of
1618    // in-edges.
1619    for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1620      // If there are any PHI nodes in this successor, drop entries for BB now.
1621      BasicBlock *DeadBB = BlocksToErase[i];
1622      while (!DeadBB->use_empty()) {
1623        Instruction *I = cast<Instruction>(DeadBB->use_back());
1624        bool Folded = ConstantFoldTerminator(I->getParent());
1625        if (!Folded) {
1626          // The constant folder may not have been able to fold the terminator
1627          // if this is a branch or switch on undef.  Fold it manually as a
1628          // branch to the first successor.
1629          if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1630            assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1631                   "Branch should be foldable!");
1632          } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1633            assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1634          } else {
1635            assert(0 && "Didn't fold away reference to block!");
1636          }
1637
1638          // Make this an uncond branch to the first successor.
1639          TerminatorInst *TI = I->getParent()->getTerminator();
1640          new BranchInst(TI->getSuccessor(0), TI);
1641
1642          // Remove entries in successor phi nodes to remove edges.
1643          for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1644            TI->getSuccessor(i)->removePredecessor(TI->getParent());
1645
1646          // Remove the old terminator.
1647          TI->eraseFromParent();
1648        }
1649      }
1650
1651      // Finally, delete the basic block.
1652      F->getBasicBlockList().erase(DeadBB);
1653    }
1654    BlocksToErase.clear();
1655  }
1656
1657  // If we inferred constant or undef return values for a function, we replaced
1658  // all call uses with the inferred value.  This means we don't need to bother
1659  // actually returning anything from the function.  Replace all return
1660  // instructions with return undef.
1661  const DenseMap<Function*, LatticeVal> &RV =Solver.getTrackedFunctionRetVals();
1662  for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1663         E = RV.end(); I != E; ++I)
1664    if (!I->second.isOverdefined() &&
1665        I->first->getReturnType() != Type::VoidTy) {
1666      Function *F = I->first;
1667      for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1668        if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1669          if (!isa<UndefValue>(RI->getOperand(0)))
1670            RI->setOperand(0, UndefValue::get(F->getReturnType()));
1671    }
1672
1673  // If we infered constant or undef values for globals variables, we can delete
1674  // the global and any stores that remain to it.
1675  const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1676  for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1677         E = TG.end(); I != E; ++I) {
1678    GlobalVariable *GV = I->first;
1679    assert(!I->second.isOverdefined() &&
1680           "Overdefined values should have been taken out of the map!");
1681    DOUT << "Found that GV '" << GV->getName()<< "' is constant!\n";
1682    while (!GV->use_empty()) {
1683      StoreInst *SI = cast<StoreInst>(GV->use_back());
1684      SI->eraseFromParent();
1685    }
1686    M.getGlobalList().erase(GV);
1687    ++IPNumGlobalConst;
1688  }
1689
1690  return MadeChanges;
1691}
1692