SCCP.cpp revision 05bb789430bab7d8fae1e94fb9aa0bb21e679ebf
1//===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file 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//===----------------------------------------------------------------------===//
133//
134/// SCCPSolver - This class is a general purpose solver for Sparse Conditional
135/// Constant Propagation.
136///
137class SCCPSolver : public InstVisitor<SCCPSolver> {
138  SmallSet<BasicBlock*, 16> BBExecutable;// The basic blocks that are executable
139  std::map<Value*, LatticeVal> ValueState;  // The state each value is in.
140
141  /// GlobalValue - If we are tracking any values for the contents of a global
142  /// variable, we keep a mapping from the constant accessor to the element of
143  /// the global, to the currently known value.  If the value becomes
144  /// overdefined, it's entry is simply removed from this map.
145  DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
146
147  /// TrackedFunctionRetVals - If we are tracking arguments into and the return
148  /// value out of a function, it will have an entry in this map, indicating
149  /// what the known return value for the function is.
150  DenseMap<Function*, LatticeVal> TrackedFunctionRetVals;
151
152  // The reason for two worklists is that overdefined is the lowest state
153  // on the lattice, and moving things to overdefined as fast as possible
154  // makes SCCP converge much faster.
155  // By having a separate worklist, we accomplish this because everything
156  // possibly overdefined will become overdefined at the soonest possible
157  // point.
158  std::vector<Value*> OverdefinedInstWorkList;
159  std::vector<Value*> InstWorkList;
160
161
162  std::vector<BasicBlock*>  BBWorkList;  // The BasicBlock work list
163
164  /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
165  /// overdefined, despite the fact that the PHI node is overdefined.
166  std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
167
168  /// KnownFeasibleEdges - Entries in this set are edges which have already had
169  /// PHI nodes retriggered.
170  typedef std::pair<BasicBlock*,BasicBlock*> Edge;
171  std::set<Edge> KnownFeasibleEdges;
172public:
173
174  /// MarkBlockExecutable - This method can be used by clients to mark all of
175  /// the blocks that are known to be intrinsically live in the processed unit.
176  void MarkBlockExecutable(BasicBlock *BB) {
177    DOUT << "Marking Block Executable: " << BB->getName() << "\n";
178    BBExecutable.insert(BB);   // Basic block is executable!
179    BBWorkList.push_back(BB);  // Add the block to the work list!
180  }
181
182  /// TrackValueOfGlobalVariable - Clients can use this method to
183  /// inform the SCCPSolver that it should track loads and stores to the
184  /// specified global variable if it can.  This is only legal to call if
185  /// performing Interprocedural SCCP.
186  void TrackValueOfGlobalVariable(GlobalVariable *GV) {
187    const Type *ElTy = GV->getType()->getElementType();
188    if (ElTy->isFirstClassType()) {
189      LatticeVal &IV = TrackedGlobals[GV];
190      if (!isa<UndefValue>(GV->getInitializer()))
191        IV.markConstant(GV->getInitializer());
192    }
193  }
194
195  /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
196  /// and out of the specified function (which cannot have its address taken),
197  /// this method must be called.
198  void AddTrackedFunction(Function *F) {
199    assert(F->hasInternalLinkage() && "Can only track internal functions!");
200    // Add an entry, F -> undef.
201    TrackedFunctionRetVals[F];
202  }
203
204  /// Solve - Solve for constants and executable blocks.
205  ///
206  void Solve();
207
208  /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
209  /// that branches on undef values cannot reach any of their successors.
210  /// However, this is not a safe assumption.  After we solve dataflow, this
211  /// method should be use to handle this.  If this returns true, the solver
212  /// should be rerun.
213  bool ResolvedUndefsIn(Function &F);
214
215  /// getExecutableBlocks - Once we have solved for constants, return the set of
216  /// blocks that is known to be executable.
217  SmallSet<BasicBlock*, 16> &getExecutableBlocks() {
218    return BBExecutable;
219  }
220
221  /// getValueMapping - Once we have solved for constants, return the mapping of
222  /// LLVM values to LatticeVals.
223  std::map<Value*, LatticeVal> &getValueMapping() {
224    return ValueState;
225  }
226
227  /// getTrackedFunctionRetVals - Get the inferred return value map.
228  ///
229  const DenseMap<Function*, LatticeVal> &getTrackedFunctionRetVals() {
230    return TrackedFunctionRetVals;
231  }
232
233  /// getTrackedGlobals - Get and return the set of inferred initializers for
234  /// global variables.
235  const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
236    return TrackedGlobals;
237  }
238
239  inline void markOverdefined(Value *V) {
240    markOverdefined(ValueState[V], V);
241  }
242
243private:
244  // markConstant - Make a value be marked as "constant".  If the value
245  // is not already a constant, add it to the instruction work list so that
246  // the users of the instruction are updated later.
247  //
248  inline void markConstant(LatticeVal &IV, Value *V, Constant *C) {
249    if (IV.markConstant(C)) {
250      DOUT << "markConstant: " << *C << ": " << *V;
251      InstWorkList.push_back(V);
252    }
253  }
254
255  inline void markForcedConstant(LatticeVal &IV, Value *V, Constant *C) {
256    IV.markForcedConstant(C);
257    DOUT << "markForcedConstant: " << *C << ": " << *V;
258    InstWorkList.push_back(V);
259  }
260
261  inline void markConstant(Value *V, Constant *C) {
262    markConstant(ValueState[V], V, C);
263  }
264
265  // markOverdefined - Make a value be marked as "overdefined". If the
266  // value is not already overdefined, add it to the overdefined instruction
267  // work list so that the users of the instruction are updated later.
268
269  inline void markOverdefined(LatticeVal &IV, Value *V) {
270    if (IV.markOverdefined()) {
271      DEBUG(DOUT << "markOverdefined: ";
272            if (Function *F = dyn_cast<Function>(V))
273              DOUT << "Function '" << F->getName() << "'\n";
274            else
275              DOUT << *V);
276      // Only instructions go on the work list
277      OverdefinedInstWorkList.push_back(V);
278    }
279  }
280
281  inline void mergeInValue(LatticeVal &IV, Value *V, LatticeVal &MergeWithV) {
282    if (IV.isOverdefined() || MergeWithV.isUndefined())
283      return;  // Noop.
284    if (MergeWithV.isOverdefined())
285      markOverdefined(IV, V);
286    else if (IV.isUndefined())
287      markConstant(IV, V, MergeWithV.getConstant());
288    else if (IV.getConstant() != MergeWithV.getConstant())
289      markOverdefined(IV, V);
290  }
291
292  inline void mergeInValue(Value *V, LatticeVal &MergeWithV) {
293    return mergeInValue(ValueState[V], V, MergeWithV);
294  }
295
296
297  // getValueState - Return the LatticeVal object that corresponds to the value.
298  // This function is necessary because not all values should start out in the
299  // underdefined state... Argument's should be overdefined, and
300  // constants should be marked as constants.  If a value is not known to be an
301  // Instruction object, then use this accessor to get its value from the map.
302  //
303  inline LatticeVal &getValueState(Value *V) {
304    std::map<Value*, LatticeVal>::iterator I = ValueState.find(V);
305    if (I != ValueState.end()) return I->second;  // Common case, in the map
306
307    if (Constant *C = dyn_cast<Constant>(V)) {
308      if (isa<UndefValue>(V)) {
309        // Nothing to do, remain undefined.
310      } else {
311        LatticeVal &LV = ValueState[C];
312        LV.markConstant(C);          // Constants are constant
313        return LV;
314      }
315    }
316    // All others are underdefined by default...
317    return ValueState[V];
318  }
319
320  // markEdgeExecutable - Mark a basic block as executable, adding it to the BB
321  // work list if it is not already executable...
322  //
323  void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
324    if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
325      return;  // This edge is already known to be executable!
326
327    if (BBExecutable.count(Dest)) {
328      DOUT << "Marking Edge Executable: " << Source->getName()
329           << " -> " << Dest->getName() << "\n";
330
331      // The destination is already executable, but we just made an edge
332      // feasible that wasn't before.  Revisit the PHI nodes in the block
333      // because they have potentially new operands.
334      for (BasicBlock::iterator I = Dest->begin(); isa<PHINode>(I); ++I)
335        visitPHINode(*cast<PHINode>(I));
336
337    } else {
338      MarkBlockExecutable(Dest);
339    }
340  }
341
342  // getFeasibleSuccessors - Return a vector of booleans to indicate which
343  // successors are reachable from a given terminator instruction.
344  //
345  void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
346
347  // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
348  // block to the 'To' basic block is currently feasible...
349  //
350  bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
351
352  // OperandChangedState - This method is invoked on all of the users of an
353  // instruction that was just changed state somehow....  Based on this
354  // information, we need to update the specified user of this instruction.
355  //
356  void OperandChangedState(User *U) {
357    // Only instructions use other variable values!
358    Instruction &I = cast<Instruction>(*U);
359    if (BBExecutable.count(I.getParent()))   // Inst is executable?
360      visit(I);
361  }
362
363private:
364  friend class InstVisitor<SCCPSolver>;
365
366  // visit implementations - Something changed in this instruction... Either an
367  // operand made a transition, or the instruction is newly executable.  Change
368  // the value type of I to reflect these changes if appropriate.
369  //
370  void visitPHINode(PHINode &I);
371
372  // Terminators
373  void visitReturnInst(ReturnInst &I);
374  void visitTerminatorInst(TerminatorInst &TI);
375
376  void visitCastInst(CastInst &I);
377  void visitSelectInst(SelectInst &I);
378  void visitBinaryOperator(Instruction &I);
379  void visitCmpInst(CmpInst &I);
380  void visitExtractElementInst(ExtractElementInst &I);
381  void visitInsertElementInst(InsertElementInst &I);
382  void visitShuffleVectorInst(ShuffleVectorInst &I);
383
384  // Instructions that cannot be folded away...
385  void visitStoreInst     (Instruction &I);
386  void visitLoadInst      (LoadInst &I);
387  void visitGetElementPtrInst(GetElementPtrInst &I);
388  void visitCallInst      (CallInst &I) { visitCallSite(CallSite::get(&I)); }
389  void visitInvokeInst    (InvokeInst &II) {
390    visitCallSite(CallSite::get(&II));
391    visitTerminatorInst(II);
392  }
393  void visitCallSite      (CallSite CS);
394  void visitUnwindInst    (TerminatorInst &I) { /*returns void*/ }
395  void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
396  void visitAllocationInst(Instruction &I) { markOverdefined(&I); }
397  void visitVANextInst    (Instruction &I) { markOverdefined(&I); }
398  void visitVAArgInst     (Instruction &I) { markOverdefined(&I); }
399  void visitFreeInst      (Instruction &I) { /*returns void*/ }
400
401  void visitInstruction(Instruction &I) {
402    // If a new instruction is added to LLVM that we don't handle...
403    cerr << "SCCP: Don't know how to handle: " << I;
404    markOverdefined(&I);   // Just in case
405  }
406};
407
408} // end anonymous namespace
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    // TODO: Consider a target hook for valid address spaces for this xform.
1019    if (isa<ConstantPointerNull>(Ptr) &&
1020        cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
1021      // load null -> null
1022      markConstant(IV, &I, Constant::getNullValue(I.getType()));
1023      return;
1024    }
1025
1026    // Transform load (constant global) into the value loaded.
1027    if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1028      if (GV->isConstant()) {
1029        if (!GV->isDeclaration()) {
1030          markConstant(IV, &I, GV->getInitializer());
1031          return;
1032        }
1033      } else if (!TrackedGlobals.empty()) {
1034        // If we are tracking this global, merge in the known value for it.
1035        DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1036          TrackedGlobals.find(GV);
1037        if (It != TrackedGlobals.end()) {
1038          mergeInValue(IV, &I, It->second);
1039          return;
1040        }
1041      }
1042    }
1043
1044    // Transform load (constantexpr_GEP global, 0, ...) into the value loaded.
1045    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
1046      if (CE->getOpcode() == Instruction::GetElementPtr)
1047    if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
1048      if (GV->isConstant() && !GV->isDeclaration())
1049        if (Constant *V =
1050             ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) {
1051          markConstant(IV, &I, V);
1052          return;
1053        }
1054  }
1055
1056  // Otherwise we cannot say for certain what value this load will produce.
1057  // Bail out.
1058  markOverdefined(IV, &I);
1059}
1060
1061void SCCPSolver::visitCallSite(CallSite CS) {
1062  Function *F = CS.getCalledFunction();
1063
1064  // If we are tracking this function, we must make sure to bind arguments as
1065  // appropriate.
1066  DenseMap<Function*, LatticeVal>::iterator TFRVI =TrackedFunctionRetVals.end();
1067  if (F && F->hasInternalLinkage())
1068    TFRVI = TrackedFunctionRetVals.find(F);
1069
1070  if (TFRVI != TrackedFunctionRetVals.end()) {
1071    // If this is the first call to the function hit, mark its entry block
1072    // executable.
1073    if (!BBExecutable.count(F->begin()))
1074      MarkBlockExecutable(F->begin());
1075
1076    CallSite::arg_iterator CAI = CS.arg_begin();
1077    for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1078         AI != E; ++AI, ++CAI) {
1079      LatticeVal &IV = ValueState[AI];
1080      if (!IV.isOverdefined())
1081        mergeInValue(IV, AI, getValueState(*CAI));
1082    }
1083  }
1084  Instruction *I = CS.getInstruction();
1085  if (I->getType() == Type::VoidTy) return;
1086
1087  LatticeVal &IV = ValueState[I];
1088  if (IV.isOverdefined()) return;
1089
1090  // Propagate the return value of the function to the value of the instruction.
1091  if (TFRVI != TrackedFunctionRetVals.end()) {
1092    mergeInValue(IV, I, TFRVI->second);
1093    return;
1094  }
1095
1096  if (F == 0 || !F->isDeclaration() || !canConstantFoldCallTo(F)) {
1097    markOverdefined(IV, I);
1098    return;
1099  }
1100
1101  SmallVector<Constant*, 8> Operands;
1102  Operands.reserve(I->getNumOperands()-1);
1103
1104  for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1105       AI != E; ++AI) {
1106    LatticeVal &State = getValueState(*AI);
1107    if (State.isUndefined())
1108      return;  // Operands are not resolved yet...
1109    else if (State.isOverdefined()) {
1110      markOverdefined(IV, I);
1111      return;
1112    }
1113    assert(State.isConstant() && "Unknown state!");
1114    Operands.push_back(State.getConstant());
1115  }
1116
1117  if (Constant *C = ConstantFoldCall(F, &Operands[0], Operands.size()))
1118    markConstant(IV, I, C);
1119  else
1120    markOverdefined(IV, I);
1121}
1122
1123
1124void SCCPSolver::Solve() {
1125  // Process the work lists until they are empty!
1126  while (!BBWorkList.empty() || !InstWorkList.empty() ||
1127         !OverdefinedInstWorkList.empty()) {
1128    // Process the instruction work list...
1129    while (!OverdefinedInstWorkList.empty()) {
1130      Value *I = OverdefinedInstWorkList.back();
1131      OverdefinedInstWorkList.pop_back();
1132
1133      DOUT << "\nPopped off OI-WL: " << *I;
1134
1135      // "I" got into the work list because it either made the transition from
1136      // bottom to constant
1137      //
1138      // Anything on this worklist that is overdefined need not be visited
1139      // since all of its users will have already been marked as overdefined
1140      // Update all of the users of this instruction's value...
1141      //
1142      for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1143           UI != E; ++UI)
1144        OperandChangedState(*UI);
1145    }
1146    // Process the instruction work list...
1147    while (!InstWorkList.empty()) {
1148      Value *I = InstWorkList.back();
1149      InstWorkList.pop_back();
1150
1151      DOUT << "\nPopped off I-WL: " << *I;
1152
1153      // "I" got into the work list because it either made the transition from
1154      // bottom to constant
1155      //
1156      // Anything on this worklist that is overdefined need not be visited
1157      // since all of its users will have already been marked as overdefined.
1158      // Update all of the users of this instruction's value...
1159      //
1160      if (!getValueState(I).isOverdefined())
1161        for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1162             UI != E; ++UI)
1163          OperandChangedState(*UI);
1164    }
1165
1166    // Process the basic block work list...
1167    while (!BBWorkList.empty()) {
1168      BasicBlock *BB = BBWorkList.back();
1169      BBWorkList.pop_back();
1170
1171      DOUT << "\nPopped off BBWL: " << *BB;
1172
1173      // Notify all instructions in this basic block that they are newly
1174      // executable.
1175      visit(BB);
1176    }
1177  }
1178}
1179
1180/// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1181/// that branches on undef values cannot reach any of their successors.
1182/// However, this is not a safe assumption.  After we solve dataflow, this
1183/// method should be use to handle this.  If this returns true, the solver
1184/// should be rerun.
1185///
1186/// This method handles this by finding an unresolved branch and marking it one
1187/// of the edges from the block as being feasible, even though the condition
1188/// doesn't say it would otherwise be.  This allows SCCP to find the rest of the
1189/// CFG and only slightly pessimizes the analysis results (by marking one,
1190/// potentially infeasible, edge feasible).  This cannot usefully modify the
1191/// constraints on the condition of the branch, as that would impact other users
1192/// of the value.
1193///
1194/// This scan also checks for values that use undefs, whose results are actually
1195/// defined.  For example, 'zext i8 undef to i32' should produce all zeros
1196/// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1197/// even if X isn't defined.
1198bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1199  for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1200    if (!BBExecutable.count(BB))
1201      continue;
1202
1203    for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1204      // Look for instructions which produce undef values.
1205      if (I->getType() == Type::VoidTy) continue;
1206
1207      LatticeVal &LV = getValueState(I);
1208      if (!LV.isUndefined()) continue;
1209
1210      // Get the lattice values of the first two operands for use below.
1211      LatticeVal &Op0LV = getValueState(I->getOperand(0));
1212      LatticeVal Op1LV;
1213      if (I->getNumOperands() == 2) {
1214        // If this is a two-operand instruction, and if both operands are
1215        // undefs, the result stays undef.
1216        Op1LV = getValueState(I->getOperand(1));
1217        if (Op0LV.isUndefined() && Op1LV.isUndefined())
1218          continue;
1219      }
1220
1221      // If this is an instructions whose result is defined even if the input is
1222      // not fully defined, propagate the information.
1223      const Type *ITy = I->getType();
1224      switch (I->getOpcode()) {
1225      default: break;          // Leave the instruction as an undef.
1226      case Instruction::ZExt:
1227        // After a zero extend, we know the top part is zero.  SExt doesn't have
1228        // to be handled here, because we don't know whether the top part is 1's
1229        // or 0's.
1230        assert(Op0LV.isUndefined());
1231        markForcedConstant(LV, I, Constant::getNullValue(ITy));
1232        return true;
1233      case Instruction::Mul:
1234      case Instruction::And:
1235        // undef * X -> 0.   X could be zero.
1236        // undef & X -> 0.   X could be zero.
1237        markForcedConstant(LV, I, Constant::getNullValue(ITy));
1238        return true;
1239
1240      case Instruction::Or:
1241        // undef | X -> -1.   X could be -1.
1242        if (const VectorType *PTy = dyn_cast<VectorType>(ITy))
1243          markForcedConstant(LV, I, ConstantVector::getAllOnesValue(PTy));
1244        else
1245          markForcedConstant(LV, I, ConstantInt::getAllOnesValue(ITy));
1246        return true;
1247
1248      case Instruction::SDiv:
1249      case Instruction::UDiv:
1250      case Instruction::SRem:
1251      case Instruction::URem:
1252        // X / undef -> undef.  No change.
1253        // X % undef -> undef.  No change.
1254        if (Op1LV.isUndefined()) break;
1255
1256        // undef / X -> 0.   X could be maxint.
1257        // undef % X -> 0.   X could be 1.
1258        markForcedConstant(LV, I, Constant::getNullValue(ITy));
1259        return true;
1260
1261      case Instruction::AShr:
1262        // undef >>s X -> undef.  No change.
1263        if (Op0LV.isUndefined()) break;
1264
1265        // X >>s undef -> X.  X could be 0, X could have the high-bit known set.
1266        if (Op0LV.isConstant())
1267          markForcedConstant(LV, I, Op0LV.getConstant());
1268        else
1269          markOverdefined(LV, I);
1270        return true;
1271      case Instruction::LShr:
1272      case Instruction::Shl:
1273        // undef >> X -> undef.  No change.
1274        // undef << X -> undef.  No change.
1275        if (Op0LV.isUndefined()) break;
1276
1277        // X >> undef -> 0.  X could be 0.
1278        // X << undef -> 0.  X could be 0.
1279        markForcedConstant(LV, I, Constant::getNullValue(ITy));
1280        return true;
1281      case Instruction::Select:
1282        // undef ? X : Y  -> X or Y.  There could be commonality between X/Y.
1283        if (Op0LV.isUndefined()) {
1284          if (!Op1LV.isConstant())  // Pick the constant one if there is any.
1285            Op1LV = getValueState(I->getOperand(2));
1286        } else if (Op1LV.isUndefined()) {
1287          // c ? undef : undef -> undef.  No change.
1288          Op1LV = getValueState(I->getOperand(2));
1289          if (Op1LV.isUndefined())
1290            break;
1291          // Otherwise, c ? undef : x -> x.
1292        } else {
1293          // Leave Op1LV as Operand(1)'s LatticeValue.
1294        }
1295
1296        if (Op1LV.isConstant())
1297          markForcedConstant(LV, I, Op1LV.getConstant());
1298        else
1299          markOverdefined(LV, I);
1300        return true;
1301      }
1302    }
1303
1304    TerminatorInst *TI = BB->getTerminator();
1305    if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1306      if (!BI->isConditional()) continue;
1307      if (!getValueState(BI->getCondition()).isUndefined())
1308        continue;
1309    } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1310      if (!getValueState(SI->getCondition()).isUndefined())
1311        continue;
1312    } else {
1313      continue;
1314    }
1315
1316    // If the edge to the second successor isn't thought to be feasible yet,
1317    // mark it so now.  We pick the second one so that this goes to some
1318    // enumerated value in a switch instead of going to the default destination.
1319    if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(1))))
1320      continue;
1321
1322    // Otherwise, it isn't already thought to be feasible.  Mark it as such now
1323    // and return.  This will make other blocks reachable, which will allow new
1324    // values to be discovered and existing ones to be moved in the lattice.
1325    markEdgeExecutable(BB, TI->getSuccessor(1));
1326
1327    // This must be a conditional branch of switch on undef.  At this point,
1328    // force the old terminator to branch to the first successor.  This is
1329    // required because we are now influencing the dataflow of the function with
1330    // the assumption that this edge is taken.  If we leave the branch condition
1331    // as undef, then further analysis could think the undef went another way
1332    // leading to an inconsistent set of conclusions.
1333    if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1334      BI->setCondition(ConstantInt::getFalse());
1335    } else {
1336      SwitchInst *SI = cast<SwitchInst>(TI);
1337      SI->setCondition(SI->getCaseValue(1));
1338    }
1339
1340    return true;
1341  }
1342
1343  return false;
1344}
1345
1346
1347namespace {
1348  //===--------------------------------------------------------------------===//
1349  //
1350  /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1351  /// Sparse Conditional Constant Propagator.
1352  ///
1353  struct VISIBILITY_HIDDEN SCCP : public FunctionPass {
1354    static char ID; // Pass identification, replacement for typeid
1355    SCCP() : FunctionPass((intptr_t)&ID) {}
1356
1357    // runOnFunction - Run the Sparse Conditional Constant Propagation
1358    // algorithm, and return true if the function was modified.
1359    //
1360    bool runOnFunction(Function &F);
1361
1362    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1363      AU.setPreservesCFG();
1364    }
1365  };
1366
1367  char SCCP::ID = 0;
1368  RegisterPass<SCCP> X("sccp", "Sparse Conditional Constant Propagation");
1369} // end anonymous namespace
1370
1371
1372// createSCCPPass - This is the public interface to this file...
1373FunctionPass *llvm::createSCCPPass() {
1374  return new SCCP();
1375}
1376
1377
1378// runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1379// and return true if the function was modified.
1380//
1381bool SCCP::runOnFunction(Function &F) {
1382  DOUT << "SCCP on function '" << F.getName() << "'\n";
1383  SCCPSolver Solver;
1384
1385  // Mark the first block of the function as being executable.
1386  Solver.MarkBlockExecutable(F.begin());
1387
1388  // Mark all arguments to the function as being overdefined.
1389  for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1390    Solver.markOverdefined(AI);
1391
1392  // Solve for constants.
1393  bool ResolvedUndefs = true;
1394  while (ResolvedUndefs) {
1395    Solver.Solve();
1396    DOUT << "RESOLVING UNDEFs\n";
1397    ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1398  }
1399
1400  bool MadeChanges = false;
1401
1402  // If we decided that there are basic blocks that are dead in this function,
1403  // delete their contents now.  Note that we cannot actually delete the blocks,
1404  // as we cannot modify the CFG of the function.
1405  //
1406  SmallSet<BasicBlock*, 16> &ExecutableBBs = Solver.getExecutableBlocks();
1407  SmallVector<Instruction*, 32> Insts;
1408  std::map<Value*, LatticeVal> &Values = Solver.getValueMapping();
1409
1410  for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1411    if (!ExecutableBBs.count(BB)) {
1412      DOUT << "  BasicBlock Dead:" << *BB;
1413      ++NumDeadBlocks;
1414
1415      // Delete the instructions backwards, as it has a reduced likelihood of
1416      // having to update as many def-use and use-def chains.
1417      for (BasicBlock::iterator I = BB->begin(), E = BB->getTerminator();
1418           I != E; ++I)
1419        Insts.push_back(I);
1420      while (!Insts.empty()) {
1421        Instruction *I = Insts.back();
1422        Insts.pop_back();
1423        if (!I->use_empty())
1424          I->replaceAllUsesWith(UndefValue::get(I->getType()));
1425        BB->getInstList().erase(I);
1426        MadeChanges = true;
1427        ++NumInstRemoved;
1428      }
1429    } else {
1430      // Iterate over all of the instructions in a function, replacing them with
1431      // constants if we have found them to be of constant values.
1432      //
1433      for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1434        Instruction *Inst = BI++;
1435        if (Inst->getType() != Type::VoidTy) {
1436          LatticeVal &IV = Values[Inst];
1437          if ((IV.isConstant() || IV.isUndefined()) &&
1438              !isa<TerminatorInst>(Inst)) {
1439            Constant *Const = IV.isConstant()
1440              ? IV.getConstant() : UndefValue::get(Inst->getType());
1441            DOUT << "  Constant: " << *Const << " = " << *Inst;
1442
1443            // Replaces all of the uses of a variable with uses of the constant.
1444            Inst->replaceAllUsesWith(Const);
1445
1446            // Delete the instruction.
1447            BB->getInstList().erase(Inst);
1448
1449            // Hey, we just changed something!
1450            MadeChanges = true;
1451            ++NumInstRemoved;
1452          }
1453        }
1454      }
1455    }
1456
1457  return MadeChanges;
1458}
1459
1460namespace {
1461  //===--------------------------------------------------------------------===//
1462  //
1463  /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1464  /// Constant Propagation.
1465  ///
1466  struct VISIBILITY_HIDDEN IPSCCP : public ModulePass {
1467    static char ID;
1468    IPSCCP() : ModulePass((intptr_t)&ID) {}
1469    bool runOnModule(Module &M);
1470  };
1471
1472  char IPSCCP::ID = 0;
1473  RegisterPass<IPSCCP>
1474  Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation");
1475} // end anonymous namespace
1476
1477// createIPSCCPPass - This is the public interface to this file...
1478ModulePass *llvm::createIPSCCPPass() {
1479  return new IPSCCP();
1480}
1481
1482
1483static bool AddressIsTaken(GlobalValue *GV) {
1484  // Delete any dead constantexpr klingons.
1485  GV->removeDeadConstantUsers();
1486
1487  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
1488       UI != E; ++UI)
1489    if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1490      if (SI->getOperand(0) == GV || SI->isVolatile())
1491        return true;  // Storing addr of GV.
1492    } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) {
1493      // Make sure we are calling the function, not passing the address.
1494      CallSite CS = CallSite::get(cast<Instruction>(*UI));
1495      for (CallSite::arg_iterator AI = CS.arg_begin(),
1496             E = CS.arg_end(); AI != E; ++AI)
1497        if (*AI == GV)
1498          return true;
1499    } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1500      if (LI->isVolatile())
1501        return true;
1502    } else {
1503      return true;
1504    }
1505  return false;
1506}
1507
1508bool IPSCCP::runOnModule(Module &M) {
1509  SCCPSolver Solver;
1510
1511  // Loop over all functions, marking arguments to those with their addresses
1512  // taken or that are external as overdefined.
1513  //
1514  for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1515    if (!F->hasInternalLinkage() || AddressIsTaken(F)) {
1516      if (!F->isDeclaration())
1517        Solver.MarkBlockExecutable(F->begin());
1518      for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1519           AI != E; ++AI)
1520        Solver.markOverdefined(AI);
1521    } else {
1522      Solver.AddTrackedFunction(F);
1523    }
1524
1525  // Loop over global variables.  We inform the solver about any internal global
1526  // variables that do not have their 'addresses taken'.  If they don't have
1527  // their addresses taken, we can propagate constants through them.
1528  for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1529       G != E; ++G)
1530    if (!G->isConstant() && G->hasInternalLinkage() && !AddressIsTaken(G))
1531      Solver.TrackValueOfGlobalVariable(G);
1532
1533  // Solve for constants.
1534  bool ResolvedUndefs = true;
1535  while (ResolvedUndefs) {
1536    Solver.Solve();
1537
1538    DOUT << "RESOLVING UNDEFS\n";
1539    ResolvedUndefs = false;
1540    for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1541      ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1542  }
1543
1544  bool MadeChanges = false;
1545
1546  // Iterate over all of the instructions in the module, replacing them with
1547  // constants if we have found them to be of constant values.
1548  //
1549  SmallSet<BasicBlock*, 16> &ExecutableBBs = Solver.getExecutableBlocks();
1550  SmallVector<Instruction*, 32> Insts;
1551  SmallVector<BasicBlock*, 32> BlocksToErase;
1552  std::map<Value*, LatticeVal> &Values = Solver.getValueMapping();
1553
1554  for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1555    for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1556         AI != E; ++AI)
1557      if (!AI->use_empty()) {
1558        LatticeVal &IV = Values[AI];
1559        if (IV.isConstant() || IV.isUndefined()) {
1560          Constant *CST = IV.isConstant() ?
1561            IV.getConstant() : UndefValue::get(AI->getType());
1562          DOUT << "***  Arg " << *AI << " = " << *CST <<"\n";
1563
1564          // Replaces all of the uses of a variable with uses of the
1565          // constant.
1566          AI->replaceAllUsesWith(CST);
1567          ++IPNumArgsElimed;
1568        }
1569      }
1570
1571    for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1572      if (!ExecutableBBs.count(BB)) {
1573        DOUT << "  BasicBlock Dead:" << *BB;
1574        ++IPNumDeadBlocks;
1575
1576        // Delete the instructions backwards, as it has a reduced likelihood of
1577        // having to update as many def-use and use-def chains.
1578        TerminatorInst *TI = BB->getTerminator();
1579        for (BasicBlock::iterator I = BB->begin(), E = TI; I != E; ++I)
1580          Insts.push_back(I);
1581
1582        while (!Insts.empty()) {
1583          Instruction *I = Insts.back();
1584          Insts.pop_back();
1585          if (!I->use_empty())
1586            I->replaceAllUsesWith(UndefValue::get(I->getType()));
1587          BB->getInstList().erase(I);
1588          MadeChanges = true;
1589          ++IPNumInstRemoved;
1590        }
1591
1592        for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1593          BasicBlock *Succ = TI->getSuccessor(i);
1594          if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1595            TI->getSuccessor(i)->removePredecessor(BB);
1596        }
1597        if (!TI->use_empty())
1598          TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1599        BB->getInstList().erase(TI);
1600
1601        if (&*BB != &F->front())
1602          BlocksToErase.push_back(BB);
1603        else
1604          new UnreachableInst(BB);
1605
1606      } else {
1607        for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1608          Instruction *Inst = BI++;
1609          if (Inst->getType() != Type::VoidTy) {
1610            LatticeVal &IV = Values[Inst];
1611            if (IV.isConstant() || IV.isUndefined() &&
1612                !isa<TerminatorInst>(Inst)) {
1613              Constant *Const = IV.isConstant()
1614                ? IV.getConstant() : UndefValue::get(Inst->getType());
1615              DOUT << "  Constant: " << *Const << " = " << *Inst;
1616
1617              // Replaces all of the uses of a variable with uses of the
1618              // constant.
1619              Inst->replaceAllUsesWith(Const);
1620
1621              // Delete the instruction.
1622              if (!isa<TerminatorInst>(Inst) && !isa<CallInst>(Inst))
1623                BB->getInstList().erase(Inst);
1624
1625              // Hey, we just changed something!
1626              MadeChanges = true;
1627              ++IPNumInstRemoved;
1628            }
1629          }
1630        }
1631      }
1632
1633    // Now that all instructions in the function are constant folded, erase dead
1634    // blocks, because we can now use ConstantFoldTerminator to get rid of
1635    // in-edges.
1636    for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1637      // If there are any PHI nodes in this successor, drop entries for BB now.
1638      BasicBlock *DeadBB = BlocksToErase[i];
1639      while (!DeadBB->use_empty()) {
1640        Instruction *I = cast<Instruction>(DeadBB->use_back());
1641        bool Folded = ConstantFoldTerminator(I->getParent());
1642        if (!Folded) {
1643          // The constant folder may not have been able to fold the terminator
1644          // if this is a branch or switch on undef.  Fold it manually as a
1645          // branch to the first successor.
1646          if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1647            assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1648                   "Branch should be foldable!");
1649          } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1650            assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1651          } else {
1652            assert(0 && "Didn't fold away reference to block!");
1653          }
1654
1655          // Make this an uncond branch to the first successor.
1656          TerminatorInst *TI = I->getParent()->getTerminator();
1657          new BranchInst(TI->getSuccessor(0), TI);
1658
1659          // Remove entries in successor phi nodes to remove edges.
1660          for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1661            TI->getSuccessor(i)->removePredecessor(TI->getParent());
1662
1663          // Remove the old terminator.
1664          TI->eraseFromParent();
1665        }
1666      }
1667
1668      // Finally, delete the basic block.
1669      F->getBasicBlockList().erase(DeadBB);
1670    }
1671    BlocksToErase.clear();
1672  }
1673
1674  // If we inferred constant or undef return values for a function, we replaced
1675  // all call uses with the inferred value.  This means we don't need to bother
1676  // actually returning anything from the function.  Replace all return
1677  // instructions with return undef.
1678  const DenseMap<Function*, LatticeVal> &RV =Solver.getTrackedFunctionRetVals();
1679  for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1680         E = RV.end(); I != E; ++I)
1681    if (!I->second.isOverdefined() &&
1682        I->first->getReturnType() != Type::VoidTy) {
1683      Function *F = I->first;
1684      for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1685        if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1686          if (!isa<UndefValue>(RI->getOperand(0)))
1687            RI->setOperand(0, UndefValue::get(F->getReturnType()));
1688    }
1689
1690  // If we infered constant or undef values for globals variables, we can delete
1691  // the global and any stores that remain to it.
1692  const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1693  for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1694         E = TG.end(); I != E; ++I) {
1695    GlobalVariable *GV = I->first;
1696    assert(!I->second.isOverdefined() &&
1697           "Overdefined values should have been taken out of the map!");
1698    DOUT << "Found that GV '" << GV->getName()<< "' is constant!\n";
1699    while (!GV->use_empty()) {
1700      StoreInst *SI = cast<StoreInst>(GV->use_back());
1701      SI->eraseFromParent();
1702    }
1703    M.getGlobalList().erase(GV);
1704    ++IPNumGlobalConst;
1705  }
1706
1707  return MadeChanges;
1708}
1709