1 //===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements sparse conditional constant propagation and merging:
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
18 //===----------------------------------------------------------------------===//
20 #define DEBUG_TYPE "sccp"
21 #include "llvm/Transforms/Scalar.h"
22 #include "llvm/Transforms/IPO.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "llvm/Target/TargetData.h"
32 #include "llvm/Support/CallSite.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/InstVisitor.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/ADT/DenseMap.h"
38 #include "llvm/ADT/DenseSet.h"
39 #include "llvm/ADT/PointerIntPair.h"
40 #include "llvm/ADT/SmallPtrSet.h"
41 #include "llvm/ADT/SmallVector.h"
42 #include "llvm/ADT/Statistic.h"
43 #include "llvm/ADT/STLExtras.h"
48 STATISTIC(NumInstRemoved, "Number of instructions removed");
49 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
51 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP");
52 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
53 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
56 /// LatticeVal class - This class represents the different lattice values that
57 /// an LLVM value may occupy. It is a simple class with value semantics.
61 /// undefined - This LLVM Value has no known value yet.
64 /// constant - This LLVM Value has a specific constant value.
67 /// forcedconstant - This LLVM Value was thought to be undef until
68 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged
69 /// with another (different) constant, it goes to overdefined, instead of
73 /// overdefined - This instruction is not known to be constant, and we know
78 /// Val: This stores the current lattice value along with the Constant* for
79 /// the constant if this is a 'constant' or 'forcedconstant' value.
80 PointerIntPair<Constant *, 2, LatticeValueTy> Val;
82 LatticeValueTy getLatticeValue() const {
87 LatticeVal() : Val(0, undefined) {}
89 bool isUndefined() const { return getLatticeValue() == undefined; }
90 bool isConstant() const {
91 return getLatticeValue() == constant || getLatticeValue() == forcedconstant;
93 bool isOverdefined() const { return getLatticeValue() == overdefined; }
95 Constant *getConstant() const {
96 assert(isConstant() && "Cannot get the constant of a non-constant!");
97 return Val.getPointer();
100 /// markOverdefined - Return true if this is a change in status.
101 bool markOverdefined() {
105 Val.setInt(overdefined);
109 /// markConstant - Return true if this is a change in status.
110 bool markConstant(Constant *V) {
112 assert(getConstant() == V && "Marking constant with different value");
117 Val.setInt(constant);
118 assert(V && "Marking constant with NULL");
121 assert(getLatticeValue() == forcedconstant &&
122 "Cannot move from overdefined to constant!");
123 // Stay at forcedconstant if the constant is the same.
124 if (V == getConstant()) return false;
126 // Otherwise, we go to overdefined. Assumptions made based on the
127 // forced value are possibly wrong. Assuming this is another constant
128 // could expose a contradiction.
129 Val.setInt(overdefined);
134 /// getConstantInt - If this is a constant with a ConstantInt value, return it
135 /// otherwise return null.
136 ConstantInt *getConstantInt() const {
138 return dyn_cast<ConstantInt>(getConstant());
142 void markForcedConstant(Constant *V) {
143 assert(isUndefined() && "Can't force a defined value!");
144 Val.setInt(forcedconstant);
148 } // end anonymous namespace.
153 //===----------------------------------------------------------------------===//
155 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional
156 /// Constant Propagation.
158 class SCCPSolver : public InstVisitor<SCCPSolver> {
159 const TargetData *TD;
160 SmallPtrSet<BasicBlock*, 8> BBExecutable;// The BBs that are executable.
161 DenseMap<Value*, LatticeVal> ValueState; // The state each value is in.
163 /// GlobalValue - If we are tracking any values for the contents of a global
164 /// variable, we keep a mapping from the constant accessor to the element of
165 /// the global, to the currently known value. If the value becomes
166 /// overdefined, it's entry is simply removed from this map.
167 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
169 /// TrackedRetVals - If we are tracking arguments into and the return
170 /// value out of a function, it will have an entry in this map, indicating
171 /// what the known return value for the function is.
172 DenseMap<Function*, LatticeVal> TrackedRetVals;
174 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
175 /// that return multiple values.
176 DenseMap<std::pair<Function*, unsigned>, LatticeVal> TrackedMultipleRetVals;
178 /// The reason for two worklists is that overdefined is the lowest state
179 /// on the lattice, and moving things to overdefined as fast as possible
180 /// makes SCCP converge much faster.
182 /// By having a separate worklist, we accomplish this because everything
183 /// possibly overdefined will become overdefined at the soonest possible
185 SmallVector<Value*, 64> OverdefinedInstWorkList;
186 SmallVector<Value*, 64> InstWorkList;
189 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list
191 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
192 /// overdefined, despite the fact that the PHI node is overdefined.
193 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
195 /// KnownFeasibleEdges - Entries in this set are edges which have already had
196 /// PHI nodes retriggered.
197 typedef std::pair<BasicBlock*, BasicBlock*> Edge;
198 DenseSet<Edge> KnownFeasibleEdges;
200 SCCPSolver(const TargetData *td) : TD(td) {}
202 /// MarkBlockExecutable - This method can be used by clients to mark all of
203 /// the blocks that are known to be intrinsically live in the processed unit.
205 /// This returns true if the block was not considered live before.
206 bool MarkBlockExecutable(BasicBlock *BB) {
207 if (!BBExecutable.insert(BB)) return false;
208 DEBUG(errs() << "Marking Block Executable: " << BB->getName() << "\n");
209 BBWorkList.push_back(BB); // Add the block to the work list!
213 /// TrackValueOfGlobalVariable - Clients can use this method to
214 /// inform the SCCPSolver that it should track loads and stores to the
215 /// specified global variable if it can. This is only legal to call if
216 /// performing Interprocedural SCCP.
217 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
218 const Type *ElTy = GV->getType()->getElementType();
219 if (ElTy->isFirstClassType()) {
220 LatticeVal &IV = TrackedGlobals[GV];
221 if (!isa<UndefValue>(GV->getInitializer()))
222 IV.markConstant(GV->getInitializer());
226 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
227 /// and out of the specified function (which cannot have its address taken),
228 /// this method must be called.
229 void AddTrackedFunction(Function *F) {
230 assert(F->hasLocalLinkage() && "Can only track internal functions!");
231 // Add an entry, F -> undef.
232 if (const StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
233 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
234 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
237 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
240 /// Solve - Solve for constants and executable blocks.
244 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
245 /// that branches on undef values cannot reach any of their successors.
246 /// However, this is not a safe assumption. After we solve dataflow, this
247 /// method should be use to handle this. If this returns true, the solver
249 bool ResolvedUndefsIn(Function &F);
251 bool isBlockExecutable(BasicBlock *BB) const {
252 return BBExecutable.count(BB);
255 LatticeVal getLatticeValueFor(Value *V) const {
256 DenseMap<Value*, LatticeVal>::const_iterator I = ValueState.find(V);
257 assert(I != ValueState.end() && "V is not in valuemap!");
261 /// getTrackedRetVals - Get the inferred return value map.
263 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
264 return TrackedRetVals;
267 /// getTrackedGlobals - Get and return the set of inferred initializers for
268 /// global variables.
269 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
270 return TrackedGlobals;
273 void markOverdefined(Value *V) {
274 markOverdefined(ValueState[V], V);
278 // markConstant - Make a value be marked as "constant". If the value
279 // is not already a constant, add it to the instruction work list so that
280 // the users of the instruction are updated later.
282 void markConstant(LatticeVal &IV, Value *V, Constant *C) {
283 if (!IV.markConstant(C)) return;
284 DEBUG(errs() << "markConstant: " << *C << ": " << *V << '\n');
285 InstWorkList.push_back(V);
288 void markConstant(Value *V, Constant *C) {
289 markConstant(ValueState[V], V, C);
292 void markForcedConstant(Value *V, Constant *C) {
293 ValueState[V].markForcedConstant(C);
294 DEBUG(errs() << "markForcedConstant: " << *C << ": " << *V << '\n');
295 InstWorkList.push_back(V);
299 // markOverdefined - Make a value be marked as "overdefined". If the
300 // value is not already overdefined, add it to the overdefined instruction
301 // work list so that the users of the instruction are updated later.
302 void markOverdefined(LatticeVal &IV, Value *V) {
303 if (!IV.markOverdefined()) return;
305 DEBUG(errs() << "markOverdefined: ";
306 if (Function *F = dyn_cast<Function>(V))
307 errs() << "Function '" << F->getName() << "'\n";
309 errs() << *V << '\n');
310 // Only instructions go on the work list
311 OverdefinedInstWorkList.push_back(V);
314 void mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) {
315 if (IV.isOverdefined() || MergeWithV.isUndefined())
317 if (MergeWithV.isOverdefined())
318 markOverdefined(IV, V);
319 else if (IV.isUndefined())
320 markConstant(IV, V, MergeWithV.getConstant());
321 else if (IV.getConstant() != MergeWithV.getConstant())
322 markOverdefined(IV, V);
325 void mergeInValue(Value *V, LatticeVal MergeWithV) {
326 mergeInValue(ValueState[V], V, MergeWithV);
330 /// getValueState - Return the LatticeVal object that corresponds to the
331 /// value. This function handles the case when the value hasn't been seen yet
332 /// by properly seeding constants etc.
333 LatticeVal &getValueState(Value *V) {
334 DenseMap<Value*, LatticeVal>::iterator I = ValueState.find(V);
335 if (I != ValueState.end()) return I->second; // Common case, in the map
337 LatticeVal &LV = ValueState[V];
339 if (Constant *C = dyn_cast<Constant>(V)) {
340 // Undef values remain undefined.
341 if (!isa<UndefValue>(V))
342 LV.markConstant(C); // Constants are constant
345 // All others are underdefined by default.
349 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
350 /// work list if it is not already executable.
351 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
352 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
353 return; // This edge is already known to be executable!
355 if (!MarkBlockExecutable(Dest)) {
356 // If the destination is already executable, we just made an *edge*
357 // feasible that wasn't before. Revisit the PHI nodes in the block
358 // because they have potentially new operands.
359 DEBUG(errs() << "Marking Edge Executable: " << Source->getName()
360 << " -> " << Dest->getName() << "\n");
363 for (BasicBlock::iterator I = Dest->begin();
364 (PN = dyn_cast<PHINode>(I)); ++I)
369 // getFeasibleSuccessors - Return a vector of booleans to indicate which
370 // successors are reachable from a given terminator instruction.
372 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
374 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
375 // block to the 'To' basic block is currently feasible.
377 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
379 // OperandChangedState - This method is invoked on all of the users of an
380 // instruction that was just changed state somehow. Based on this
381 // information, we need to update the specified user of this instruction.
383 void OperandChangedState(User *U) {
384 // Only instructions use other variable values!
385 Instruction &I = cast<Instruction>(*U);
386 if (BBExecutable.count(I.getParent())) // Inst is executable?
390 /// RemoveFromOverdefinedPHIs - If I has any entries in the
391 /// UsersOfOverdefinedPHIs map for PN, remove them now.
392 void RemoveFromOverdefinedPHIs(Instruction *I, PHINode *PN) {
393 if (UsersOfOverdefinedPHIs.empty()) return;
394 std::multimap<PHINode*, Instruction*>::iterator It, E;
395 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN);
398 UsersOfOverdefinedPHIs.erase(It++);
405 friend class InstVisitor<SCCPSolver>;
407 // visit implementations - Something changed in this instruction. Either an
408 // operand made a transition, or the instruction is newly executable. Change
409 // the value type of I to reflect these changes if appropriate.
410 void visitPHINode(PHINode &I);
413 void visitReturnInst(ReturnInst &I);
414 void visitTerminatorInst(TerminatorInst &TI);
416 void visitCastInst(CastInst &I);
417 void visitSelectInst(SelectInst &I);
418 void visitBinaryOperator(Instruction &I);
419 void visitCmpInst(CmpInst &I);
420 void visitExtractElementInst(ExtractElementInst &I);
421 void visitInsertElementInst(InsertElementInst &I);
422 void visitShuffleVectorInst(ShuffleVectorInst &I);
423 void visitExtractValueInst(ExtractValueInst &EVI);
424 void visitInsertValueInst(InsertValueInst &IVI);
426 // Instructions that cannot be folded away.
427 void visitStoreInst (StoreInst &I);
428 void visitLoadInst (LoadInst &I);
429 void visitGetElementPtrInst(GetElementPtrInst &I);
430 void visitCallInst (CallInst &I) {
433 visitCallSite(CallSite::get(&I));
435 void visitInvokeInst (InvokeInst &II) {
436 visitCallSite(CallSite::get(&II));
437 visitTerminatorInst(II);
439 void visitCallSite (CallSite CS);
440 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ }
441 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
442 void visitAllocaInst (Instruction &I) { markOverdefined(&I); }
443 void visitVANextInst (Instruction &I) { markOverdefined(&I); }
444 void visitVAArgInst (Instruction &I) { markOverdefined(&I); }
446 void visitInstruction(Instruction &I) {
447 // If a new instruction is added to LLVM that we don't handle.
448 errs() << "SCCP: Don't know how to handle: " << I;
449 markOverdefined(&I); // Just in case
453 } // end anonymous namespace
456 // getFeasibleSuccessors - Return a vector of booleans to indicate which
457 // successors are reachable from a given terminator instruction.
459 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
460 SmallVector<bool, 16> &Succs) {
461 Succs.resize(TI.getNumSuccessors());
462 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
463 if (BI->isUnconditional()) {
468 LatticeVal BCValue = getValueState(BI->getCondition());
469 ConstantInt *CI = BCValue.getConstantInt();
471 // Overdefined condition variables, and branches on unfoldable constant
472 // conditions, mean the branch could go either way.
473 if (!BCValue.isUndefined())
474 Succs[0] = Succs[1] = true;
478 // Constant condition variables mean the branch can only go a single way.
479 Succs[CI->isZero()] = true;
483 if (isa<InvokeInst>(TI)) {
484 // Invoke instructions successors are always executable.
485 Succs[0] = Succs[1] = true;
489 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
490 LatticeVal SCValue = getValueState(SI->getCondition());
491 ConstantInt *CI = SCValue.getConstantInt();
493 if (CI == 0) { // Overdefined or undefined condition?
494 // All destinations are executable!
495 if (!SCValue.isUndefined())
496 Succs.assign(TI.getNumSuccessors(), true);
500 Succs[SI->findCaseValue(CI)] = true;
504 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
505 if (isa<IndirectBrInst>(&TI)) {
506 // Just mark all destinations executable!
507 Succs.assign(TI.getNumSuccessors(), true);
512 errs() << "Unknown terminator instruction: " << TI << '\n';
514 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
518 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
519 // block to the 'To' basic block is currently feasible.
521 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
522 assert(BBExecutable.count(To) && "Dest should always be alive!");
524 // Make sure the source basic block is executable!!
525 if (!BBExecutable.count(From)) return false;
527 // Check to make sure this edge itself is actually feasible now.
528 TerminatorInst *TI = From->getTerminator();
529 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
530 if (BI->isUnconditional())
533 LatticeVal BCValue = getValueState(BI->getCondition());
535 // Overdefined condition variables mean the branch could go either way,
536 // undef conditions mean that neither edge is feasible yet.
537 ConstantInt *CI = BCValue.getConstantInt();
539 return !BCValue.isUndefined();
541 // Constant condition variables mean the branch can only go a single way.
542 return BI->getSuccessor(CI->isZero()) == To;
545 // Invoke instructions successors are always executable.
546 if (isa<InvokeInst>(TI))
549 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
550 LatticeVal SCValue = getValueState(SI->getCondition());
551 ConstantInt *CI = SCValue.getConstantInt();
554 return !SCValue.isUndefined();
556 // Make sure to skip the "default value" which isn't a value
557 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i)
558 if (SI->getSuccessorValue(i) == CI) // Found the taken branch.
559 return SI->getSuccessor(i) == To;
561 // If the constant value is not equal to any of the branches, we must
562 // execute default branch.
563 return SI->getDefaultDest() == To;
566 // Just mark all destinations executable!
567 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
568 if (isa<IndirectBrInst>(&TI))
572 errs() << "Unknown terminator instruction: " << *TI << '\n';
577 // visit Implementations - Something changed in this instruction, either an
578 // operand made a transition, or the instruction is newly executable. Change
579 // the value type of I to reflect these changes if appropriate. This method
580 // makes sure to do the following actions:
582 // 1. If a phi node merges two constants in, and has conflicting value coming
583 // from different branches, or if the PHI node merges in an overdefined
584 // value, then the PHI node becomes overdefined.
585 // 2. If a phi node merges only constants in, and they all agree on value, the
586 // PHI node becomes a constant value equal to that.
587 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
588 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
589 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
590 // 6. If a conditional branch has a value that is constant, make the selected
591 // destination executable
592 // 7. If a conditional branch has a value that is overdefined, make all
593 // successors executable.
595 void SCCPSolver::visitPHINode(PHINode &PN) {
596 if (getValueState(&PN).isOverdefined()) {
597 // There may be instructions using this PHI node that are not overdefined
598 // themselves. If so, make sure that they know that the PHI node operand
600 std::multimap<PHINode*, Instruction*>::iterator I, E;
601 tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN);
605 SmallVector<Instruction*, 16> Users;
607 Users.push_back(I->second);
608 while (!Users.empty())
609 visit(Users.pop_back_val());
610 return; // Quick exit
613 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
614 // and slow us down a lot. Just mark them overdefined.
615 if (PN.getNumIncomingValues() > 64)
616 return markOverdefined(&PN);
618 // Look at all of the executable operands of the PHI node. If any of them
619 // are overdefined, the PHI becomes overdefined as well. If they are all
620 // constant, and they agree with each other, the PHI becomes the identical
621 // constant. If they are constant and don't agree, the PHI is overdefined.
622 // If there are no executable operands, the PHI remains undefined.
624 Constant *OperandVal = 0;
625 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
626 LatticeVal IV = getValueState(PN.getIncomingValue(i));
627 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
629 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
632 if (IV.isOverdefined()) // PHI node becomes overdefined!
633 return markOverdefined(&PN);
635 if (OperandVal == 0) { // Grab the first value.
636 OperandVal = IV.getConstant();
640 // There is already a reachable operand. If we conflict with it,
641 // then the PHI node becomes overdefined. If we agree with it, we
644 // Check to see if there are two different constants merging, if so, the PHI
645 // node is overdefined.
646 if (IV.getConstant() != OperandVal)
647 return markOverdefined(&PN);
650 // If we exited the loop, this means that the PHI node only has constant
651 // arguments that agree with each other(and OperandVal is the constant) or
652 // OperandVal is null because there are no defined incoming arguments. If
653 // this is the case, the PHI remains undefined.
656 markConstant(&PN, OperandVal); // Acquire operand value
659 void SCCPSolver::visitReturnInst(ReturnInst &I) {
660 if (I.getNumOperands() == 0) return; // ret void
662 Function *F = I.getParent()->getParent();
663 // If we are tracking the return value of this function, merge it in.
664 if (!F->hasLocalLinkage())
667 if (!TrackedRetVals.empty()) {
668 DenseMap<Function*, LatticeVal>::iterator TFRVI =
669 TrackedRetVals.find(F);
670 if (TFRVI != TrackedRetVals.end() &&
671 !TFRVI->second.isOverdefined()) {
672 mergeInValue(TFRVI->second, F, getValueState(I.getOperand(0)));
677 // Handle functions that return multiple values.
678 if (!TrackedMultipleRetVals.empty() &&
679 isa<StructType>(I.getOperand(0)->getType())) {
680 for (unsigned i = 0, e = I.getOperand(0)->getType()->getNumContainedTypes();
682 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
683 It = TrackedMultipleRetVals.find(std::make_pair(F, i));
684 if (It == TrackedMultipleRetVals.end()) break;
685 if (Value *Val = FindInsertedValue(I.getOperand(0), i, I.getContext()))
686 mergeInValue(It->second, F, getValueState(Val));
691 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
692 SmallVector<bool, 16> SuccFeasible;
693 getFeasibleSuccessors(TI, SuccFeasible);
695 BasicBlock *BB = TI.getParent();
697 // Mark all feasible successors executable.
698 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
700 markEdgeExecutable(BB, TI.getSuccessor(i));
703 void SCCPSolver::visitCastInst(CastInst &I) {
704 LatticeVal OpSt = getValueState(I.getOperand(0));
705 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand
707 else if (OpSt.isConstant()) // Propagate constant value
708 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
709 OpSt.getConstant(), I.getType()));
712 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
713 Value *Aggr = EVI.getAggregateOperand();
715 // If the operand to the extractvalue is an undef, the result is undef.
716 if (isa<UndefValue>(Aggr))
719 // Currently only handle single-index extractvalues.
720 if (EVI.getNumIndices() != 1)
721 return markOverdefined(&EVI);
724 if (CallInst *CI = dyn_cast<CallInst>(Aggr))
725 F = CI->getCalledFunction();
726 else if (InvokeInst *II = dyn_cast<InvokeInst>(Aggr))
727 F = II->getCalledFunction();
729 // TODO: If IPSCCP resolves the callee of this function, we could propagate a
731 if (F == 0 || TrackedMultipleRetVals.empty())
732 return markOverdefined(&EVI);
734 // See if we are tracking the result of the callee. If not tracking this
735 // function (for example, it is a declaration) just move to overdefined.
736 if (!TrackedMultipleRetVals.count(std::make_pair(F, *EVI.idx_begin())))
737 return markOverdefined(&EVI);
739 // Otherwise, the value will be merged in here as a result of CallSite
743 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
744 Value *Aggr = IVI.getAggregateOperand();
745 Value *Val = IVI.getInsertedValueOperand();
747 // If the operands to the insertvalue are undef, the result is undef.
748 if (isa<UndefValue>(Aggr) && isa<UndefValue>(Val))
751 // Currently only handle single-index insertvalues.
752 if (IVI.getNumIndices() != 1)
753 return markOverdefined(&IVI);
755 // Currently only handle insertvalue instructions that are in a single-use
756 // chain that builds up a return value.
757 for (const InsertValueInst *TmpIVI = &IVI; ; ) {
758 if (!TmpIVI->hasOneUse())
759 return markOverdefined(&IVI);
761 const Value *V = *TmpIVI->use_begin();
762 if (isa<ReturnInst>(V))
764 TmpIVI = dyn_cast<InsertValueInst>(V);
766 return markOverdefined(&IVI);
769 // See if we are tracking the result of the callee.
770 Function *F = IVI.getParent()->getParent();
771 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
772 It = TrackedMultipleRetVals.find(std::make_pair(F, *IVI.idx_begin()));
774 // Merge in the inserted member value.
775 if (It != TrackedMultipleRetVals.end())
776 mergeInValue(It->second, F, getValueState(Val));
778 // Mark the aggregate result of the IVI overdefined; any tracking that we do
779 // will be done on the individual member values.
780 markOverdefined(&IVI);
783 void SCCPSolver::visitSelectInst(SelectInst &I) {
784 LatticeVal CondValue = getValueState(I.getCondition());
785 if (CondValue.isUndefined())
788 if (ConstantInt *CondCB = CondValue.getConstantInt()) {
789 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
790 mergeInValue(&I, getValueState(OpVal));
794 // Otherwise, the condition is overdefined or a constant we can't evaluate.
795 // See if we can produce something better than overdefined based on the T/F
797 LatticeVal TVal = getValueState(I.getTrueValue());
798 LatticeVal FVal = getValueState(I.getFalseValue());
800 // select ?, C, C -> C.
801 if (TVal.isConstant() && FVal.isConstant() &&
802 TVal.getConstant() == FVal.getConstant())
803 return markConstant(&I, FVal.getConstant());
805 if (TVal.isUndefined()) // select ?, undef, X -> X.
806 return mergeInValue(&I, FVal);
807 if (FVal.isUndefined()) // select ?, X, undef -> X.
808 return mergeInValue(&I, TVal);
812 // Handle Binary Operators.
813 void SCCPSolver::visitBinaryOperator(Instruction &I) {
814 LatticeVal V1State = getValueState(I.getOperand(0));
815 LatticeVal V2State = getValueState(I.getOperand(1));
817 LatticeVal &IV = ValueState[&I];
818 if (IV.isOverdefined()) return;
820 if (V1State.isConstant() && V2State.isConstant())
821 return markConstant(IV, &I,
822 ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
823 V2State.getConstant()));
825 // If something is undef, wait for it to resolve.
826 if (!V1State.isOverdefined() && !V2State.isOverdefined())
829 // Otherwise, one of our operands is overdefined. Try to produce something
830 // better than overdefined with some tricks.
832 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
833 // operand is overdefined.
834 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
835 LatticeVal *NonOverdefVal = 0;
836 if (!V1State.isOverdefined())
837 NonOverdefVal = &V1State;
838 else if (!V2State.isOverdefined())
839 NonOverdefVal = &V2State;
842 if (NonOverdefVal->isUndefined()) {
843 // Could annihilate value.
844 if (I.getOpcode() == Instruction::And)
845 markConstant(IV, &I, Constant::getNullValue(I.getType()));
846 else if (const VectorType *PT = dyn_cast<VectorType>(I.getType()))
847 markConstant(IV, &I, Constant::getAllOnesValue(PT));
850 Constant::getAllOnesValue(I.getType()));
854 if (I.getOpcode() == Instruction::And) {
856 if (NonOverdefVal->getConstant()->isNullValue())
857 return markConstant(IV, &I, NonOverdefVal->getConstant());
859 if (ConstantInt *CI = NonOverdefVal->getConstantInt())
860 if (CI->isAllOnesValue()) // X or -1 = -1
861 return markConstant(IV, &I, NonOverdefVal->getConstant());
867 // If both operands are PHI nodes, it is possible that this instruction has
868 // a constant value, despite the fact that the PHI node doesn't. Check for
869 // this condition now.
870 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
871 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
872 if (PN1->getParent() == PN2->getParent()) {
873 // Since the two PHI nodes are in the same basic block, they must have
874 // entries for the same predecessors. Walk the predecessor list, and
875 // if all of the incoming values are constants, and the result of
876 // evaluating this expression with all incoming value pairs is the
877 // same, then this expression is a constant even though the PHI node
878 // is not a constant!
880 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
881 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
882 BasicBlock *InBlock = PN1->getIncomingBlock(i);
883 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
885 if (In1.isOverdefined() || In2.isOverdefined()) {
886 Result.markOverdefined();
887 break; // Cannot fold this operation over the PHI nodes!
890 if (In1.isConstant() && In2.isConstant()) {
891 Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(),
893 if (Result.isUndefined())
894 Result.markConstant(V);
895 else if (Result.isConstant() && Result.getConstant() != V) {
896 Result.markOverdefined();
902 // If we found a constant value here, then we know the instruction is
903 // constant despite the fact that the PHI nodes are overdefined.
904 if (Result.isConstant()) {
905 markConstant(IV, &I, Result.getConstant());
906 // Remember that this instruction is virtually using the PHI node
908 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
909 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
913 if (Result.isUndefined())
916 // Okay, this really is overdefined now. Since we might have
917 // speculatively thought that this was not overdefined before, and
918 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
919 // make sure to clean out any entries that we put there, for
921 RemoveFromOverdefinedPHIs(&I, PN1);
922 RemoveFromOverdefinedPHIs(&I, PN2);
928 // Handle ICmpInst instruction.
929 void SCCPSolver::visitCmpInst(CmpInst &I) {
930 LatticeVal V1State = getValueState(I.getOperand(0));
931 LatticeVal V2State = getValueState(I.getOperand(1));
933 LatticeVal &IV = ValueState[&I];
934 if (IV.isOverdefined()) return;
936 if (V1State.isConstant() && V2State.isConstant())
937 return markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
938 V1State.getConstant(),
939 V2State.getConstant()));
941 // If operands are still undefined, wait for it to resolve.
942 if (!V1State.isOverdefined() && !V2State.isOverdefined())
945 // If something is overdefined, use some tricks to avoid ending up and over
946 // defined if we can.
948 // If both operands are PHI nodes, it is possible that this instruction has
949 // a constant value, despite the fact that the PHI node doesn't. Check for
950 // this condition now.
951 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
952 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
953 if (PN1->getParent() == PN2->getParent()) {
954 // Since the two PHI nodes are in the same basic block, they must have
955 // entries for the same predecessors. Walk the predecessor list, and
956 // if all of the incoming values are constants, and the result of
957 // evaluating this expression with all incoming value pairs is the
958 // same, then this expression is a constant even though the PHI node
959 // is not a constant!
961 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
962 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
963 BasicBlock *InBlock = PN1->getIncomingBlock(i);
964 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
966 if (In1.isOverdefined() || In2.isOverdefined()) {
967 Result.markOverdefined();
968 break; // Cannot fold this operation over the PHI nodes!
971 if (In1.isConstant() && In2.isConstant()) {
972 Constant *V = ConstantExpr::getCompare(I.getPredicate(),
975 if (Result.isUndefined())
976 Result.markConstant(V);
977 else if (Result.isConstant() && Result.getConstant() != V) {
978 Result.markOverdefined();
984 // If we found a constant value here, then we know the instruction is
985 // constant despite the fact that the PHI nodes are overdefined.
986 if (Result.isConstant()) {
987 markConstant(&I, Result.getConstant());
988 // Remember that this instruction is virtually using the PHI node
990 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
991 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
995 if (Result.isUndefined())
998 // Okay, this really is overdefined now. Since we might have
999 // speculatively thought that this was not overdefined before, and
1000 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
1001 // make sure to clean out any entries that we put there, for
1003 RemoveFromOverdefinedPHIs(&I, PN1);
1004 RemoveFromOverdefinedPHIs(&I, PN2);
1007 markOverdefined(&I);
1010 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
1011 // FIXME : SCCP does not handle vectors properly.
1012 return markOverdefined(&I);
1015 LatticeVal &ValState = getValueState(I.getOperand(0));
1016 LatticeVal &IdxState = getValueState(I.getOperand(1));
1018 if (ValState.isOverdefined() || IdxState.isOverdefined())
1019 markOverdefined(&I);
1020 else if(ValState.isConstant() && IdxState.isConstant())
1021 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
1022 IdxState.getConstant()));
1026 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
1027 // FIXME : SCCP does not handle vectors properly.
1028 return markOverdefined(&I);
1030 LatticeVal &ValState = getValueState(I.getOperand(0));
1031 LatticeVal &EltState = getValueState(I.getOperand(1));
1032 LatticeVal &IdxState = getValueState(I.getOperand(2));
1034 if (ValState.isOverdefined() || EltState.isOverdefined() ||
1035 IdxState.isOverdefined())
1036 markOverdefined(&I);
1037 else if(ValState.isConstant() && EltState.isConstant() &&
1038 IdxState.isConstant())
1039 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
1040 EltState.getConstant(),
1041 IdxState.getConstant()));
1042 else if (ValState.isUndefined() && EltState.isConstant() &&
1043 IdxState.isConstant())
1044 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
1045 EltState.getConstant(),
1046 IdxState.getConstant()));
1050 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
1051 // FIXME : SCCP does not handle vectors properly.
1052 return markOverdefined(&I);
1054 LatticeVal &V1State = getValueState(I.getOperand(0));
1055 LatticeVal &V2State = getValueState(I.getOperand(1));
1056 LatticeVal &MaskState = getValueState(I.getOperand(2));
1058 if (MaskState.isUndefined() ||
1059 (V1State.isUndefined() && V2State.isUndefined()))
1060 return; // Undefined output if mask or both inputs undefined.
1062 if (V1State.isOverdefined() || V2State.isOverdefined() ||
1063 MaskState.isOverdefined()) {
1064 markOverdefined(&I);
1066 // A mix of constant/undef inputs.
1067 Constant *V1 = V1State.isConstant() ?
1068 V1State.getConstant() : UndefValue::get(I.getType());
1069 Constant *V2 = V2State.isConstant() ?
1070 V2State.getConstant() : UndefValue::get(I.getType());
1071 Constant *Mask = MaskState.isConstant() ?
1072 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
1073 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
1078 // Handle getelementptr instructions. If all operands are constants then we
1079 // can turn this into a getelementptr ConstantExpr.
1081 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1082 LatticeVal &IV = ValueState[&I];
1083 if (IV.isOverdefined()) return;
1085 SmallVector<Constant*, 8> Operands;
1086 Operands.reserve(I.getNumOperands());
1088 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1089 LatticeVal State = getValueState(I.getOperand(i));
1090 if (State.isUndefined())
1091 return; // Operands are not resolved yet.
1093 if (State.isOverdefined())
1094 return markOverdefined(IV, &I);
1096 assert(State.isConstant() && "Unknown state!");
1097 Operands.push_back(State.getConstant());
1100 Constant *Ptr = Operands[0];
1101 markConstant(&I, ConstantExpr::getGetElementPtr(Ptr, &Operands[0]+1,
1102 Operands.size()-1));
1105 void SCCPSolver::visitStoreInst(StoreInst &SI) {
1106 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1109 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1110 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1111 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1113 // Get the value we are storing into the global, then merge it.
1114 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)));
1115 if (I->second.isOverdefined())
1116 TrackedGlobals.erase(I); // No need to keep tracking this!
1120 // Handle load instructions. If the operand is a constant pointer to a constant
1121 // global, we can replace the load with the loaded constant value!
1122 void SCCPSolver::visitLoadInst(LoadInst &I) {
1123 LatticeVal PtrVal = getValueState(I.getOperand(0));
1124 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1126 LatticeVal &IV = ValueState[&I];
1127 if (IV.isOverdefined()) return;
1129 if (!PtrVal.isConstant() || I.isVolatile())
1130 return markOverdefined(IV, &I);
1132 Constant *Ptr = PtrVal.getConstant();
1134 // load null -> null
1135 if (isa<ConstantPointerNull>(Ptr) && I.getPointerAddressSpace() == 0)
1136 return markConstant(IV, &I, Constant::getNullValue(I.getType()));
1138 // Transform load (constant global) into the value loaded.
1139 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1140 if (!TrackedGlobals.empty()) {
1141 // If we are tracking this global, merge in the known value for it.
1142 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1143 TrackedGlobals.find(GV);
1144 if (It != TrackedGlobals.end()) {
1145 mergeInValue(IV, &I, It->second);
1151 // Transform load from a constant into a constant if possible.
1152 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, TD))
1153 return markConstant(IV, &I, C);
1155 // Otherwise we cannot say for certain what value this load will produce.
1157 markOverdefined(IV, &I);
1160 void SCCPSolver::visitCallSite(CallSite CS) {
1161 Function *F = CS.getCalledFunction();
1162 Instruction *I = CS.getInstruction();
1164 // The common case is that we aren't tracking the callee, either because we
1165 // are not doing interprocedural analysis or the callee is indirect, or is
1166 // external. Handle these cases first.
1167 if (F == 0 || !F->hasLocalLinkage()) {
1169 // Void return and not tracking callee, just bail.
1170 if (I->getType()->isVoidTy()) return;
1172 // Otherwise, if we have a single return value case, and if the function is
1173 // a declaration, maybe we can constant fold it.
1174 if (!isa<StructType>(I->getType()) && F && F->isDeclaration() &&
1175 canConstantFoldCallTo(F)) {
1177 SmallVector<Constant*, 8> Operands;
1178 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1180 LatticeVal State = getValueState(*AI);
1182 if (State.isUndefined())
1183 return; // Operands are not resolved yet.
1184 if (State.isOverdefined())
1185 return markOverdefined(I);
1186 assert(State.isConstant() && "Unknown state!");
1187 Operands.push_back(State.getConstant());
1190 // If we can constant fold this, mark the result of the call as a
1192 if (Constant *C = ConstantFoldCall(F, Operands.data(), Operands.size()))
1193 return markConstant(I, C);
1196 // Otherwise, we don't know anything about this call, mark it overdefined.
1197 return markOverdefined(I);
1200 // If this is a single/zero retval case, see if we're tracking the function.
1201 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1202 if (TFRVI != TrackedRetVals.end()) {
1203 // If so, propagate the return value of the callee into this call result.
1204 mergeInValue(I, TFRVI->second);
1205 } else if (isa<StructType>(I->getType())) {
1206 // Check to see if we're tracking this callee, if not, handle it in the
1207 // common path above.
1208 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
1209 TMRVI = TrackedMultipleRetVals.find(std::make_pair(F, 0));
1210 if (TMRVI == TrackedMultipleRetVals.end())
1211 goto CallOverdefined;
1213 // Need to mark as overdefined, otherwise it stays undefined which
1214 // creates extractvalue undef, <idx>
1217 // If we are tracking this callee, propagate the return values of the call
1218 // into this call site. We do this by walking all the uses. Single-index
1219 // ExtractValueInst uses can be tracked; anything more complicated is
1220 // currently handled conservatively.
1221 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1223 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(*UI)) {
1224 if (EVI->getNumIndices() == 1) {
1226 TrackedMultipleRetVals[std::make_pair(F, *EVI->idx_begin())]);
1230 // The aggregate value is used in a way not handled here. Assume nothing.
1231 markOverdefined(*UI);
1234 // Otherwise we're not tracking this callee, so handle it in the
1235 // common path above.
1236 goto CallOverdefined;
1239 // Finally, if this is the first call to the function hit, mark its entry
1240 // block executable.
1241 MarkBlockExecutable(F->begin());
1243 // Propagate information from this call site into the callee.
1244 CallSite::arg_iterator CAI = CS.arg_begin();
1245 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1246 AI != E; ++AI, ++CAI) {
1247 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1248 markOverdefined(AI);
1252 mergeInValue(AI, getValueState(*CAI));
1256 void SCCPSolver::Solve() {
1257 // Process the work lists until they are empty!
1258 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1259 !OverdefinedInstWorkList.empty()) {
1260 // Process the overdefined instruction's work list first, which drives other
1261 // things to overdefined more quickly.
1262 while (!OverdefinedInstWorkList.empty()) {
1263 Value *I = OverdefinedInstWorkList.pop_back_val();
1265 DEBUG(errs() << "\nPopped off OI-WL: " << *I << '\n');
1267 // "I" got into the work list because it either made the transition from
1268 // bottom to constant
1270 // Anything on this worklist that is overdefined need not be visited
1271 // since all of its users will have already been marked as overdefined
1272 // Update all of the users of this instruction's value.
1274 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1276 OperandChangedState(*UI);
1279 // Process the instruction work list.
1280 while (!InstWorkList.empty()) {
1281 Value *I = InstWorkList.pop_back_val();
1283 DEBUG(errs() << "\nPopped off I-WL: " << *I << '\n');
1285 // "I" got into the work list because it made the transition from undef to
1288 // Anything on this worklist that is overdefined need not be visited
1289 // since all of its users will have already been marked as overdefined.
1290 // Update all of the users of this instruction's value.
1292 if (!getValueState(I).isOverdefined())
1293 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1295 OperandChangedState(*UI);
1298 // Process the basic block work list.
1299 while (!BBWorkList.empty()) {
1300 BasicBlock *BB = BBWorkList.back();
1301 BBWorkList.pop_back();
1303 DEBUG(errs() << "\nPopped off BBWL: " << *BB << '\n');
1305 // Notify all instructions in this basic block that they are newly
1312 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1313 /// that branches on undef values cannot reach any of their successors.
1314 /// However, this is not a safe assumption. After we solve dataflow, this
1315 /// method should be use to handle this. If this returns true, the solver
1316 /// should be rerun.
1318 /// This method handles this by finding an unresolved branch and marking it one
1319 /// of the edges from the block as being feasible, even though the condition
1320 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1321 /// CFG and only slightly pessimizes the analysis results (by marking one,
1322 /// potentially infeasible, edge feasible). This cannot usefully modify the
1323 /// constraints on the condition of the branch, as that would impact other users
1326 /// This scan also checks for values that use undefs, whose results are actually
1327 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1328 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1329 /// even if X isn't defined.
1330 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1331 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1332 if (!BBExecutable.count(BB))
1335 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1336 // Look for instructions which produce undef values.
1337 if (I->getType()->isVoidTy()) continue;
1339 LatticeVal &LV = getValueState(I);
1340 if (!LV.isUndefined()) continue;
1342 // Get the lattice values of the first two operands for use below.
1343 LatticeVal Op0LV = getValueState(I->getOperand(0));
1345 if (I->getNumOperands() == 2) {
1346 // If this is a two-operand instruction, and if both operands are
1347 // undefs, the result stays undef.
1348 Op1LV = getValueState(I->getOperand(1));
1349 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1353 // If this is an instructions whose result is defined even if the input is
1354 // not fully defined, propagate the information.
1355 const Type *ITy = I->getType();
1356 switch (I->getOpcode()) {
1357 default: break; // Leave the instruction as an undef.
1358 case Instruction::ZExt:
1359 // After a zero extend, we know the top part is zero. SExt doesn't have
1360 // to be handled here, because we don't know whether the top part is 1's
1362 markForcedConstant(I, Constant::getNullValue(ITy));
1364 case Instruction::Mul:
1365 case Instruction::And:
1366 // undef * X -> 0. X could be zero.
1367 // undef & X -> 0. X could be zero.
1368 markForcedConstant(I, Constant::getNullValue(ITy));
1371 case Instruction::Or:
1372 // undef | X -> -1. X could be -1.
1373 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1376 case Instruction::SDiv:
1377 case Instruction::UDiv:
1378 case Instruction::SRem:
1379 case Instruction::URem:
1380 // X / undef -> undef. No change.
1381 // X % undef -> undef. No change.
1382 if (Op1LV.isUndefined()) break;
1384 // undef / X -> 0. X could be maxint.
1385 // undef % X -> 0. X could be 1.
1386 markForcedConstant(I, Constant::getNullValue(ITy));
1389 case Instruction::AShr:
1390 // undef >>s X -> undef. No change.
1391 if (Op0LV.isUndefined()) break;
1393 // X >>s undef -> X. X could be 0, X could have the high-bit known set.
1394 if (Op0LV.isConstant())
1395 markForcedConstant(I, Op0LV.getConstant());
1399 case Instruction::LShr:
1400 case Instruction::Shl:
1401 // undef >> X -> undef. No change.
1402 // undef << X -> undef. No change.
1403 if (Op0LV.isUndefined()) break;
1405 // X >> undef -> 0. X could be 0.
1406 // X << undef -> 0. X could be 0.
1407 markForcedConstant(I, Constant::getNullValue(ITy));
1409 case Instruction::Select:
1410 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1411 if (Op0LV.isUndefined()) {
1412 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1413 Op1LV = getValueState(I->getOperand(2));
1414 } else if (Op1LV.isUndefined()) {
1415 // c ? undef : undef -> undef. No change.
1416 Op1LV = getValueState(I->getOperand(2));
1417 if (Op1LV.isUndefined())
1419 // Otherwise, c ? undef : x -> x.
1421 // Leave Op1LV as Operand(1)'s LatticeValue.
1424 if (Op1LV.isConstant())
1425 markForcedConstant(I, Op1LV.getConstant());
1429 case Instruction::Call:
1430 // If a call has an undef result, it is because it is constant foldable
1431 // but one of the inputs was undef. Just force the result to
1438 TerminatorInst *TI = BB->getTerminator();
1439 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1440 if (!BI->isConditional()) continue;
1441 if (!getValueState(BI->getCondition()).isUndefined())
1443 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1444 if (SI->getNumSuccessors() < 2) // no cases
1446 if (!getValueState(SI->getCondition()).isUndefined())
1452 // If the edge to the second successor isn't thought to be feasible yet,
1453 // mark it so now. We pick the second one so that this goes to some
1454 // enumerated value in a switch instead of going to the default destination.
1455 if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(1))))
1458 // Otherwise, it isn't already thought to be feasible. Mark it as such now
1459 // and return. This will make other blocks reachable, which will allow new
1460 // values to be discovered and existing ones to be moved in the lattice.
1461 markEdgeExecutable(BB, TI->getSuccessor(1));
1463 // This must be a conditional branch of switch on undef. At this point,
1464 // force the old terminator to branch to the first successor. This is
1465 // required because we are now influencing the dataflow of the function with
1466 // the assumption that this edge is taken. If we leave the branch condition
1467 // as undef, then further analysis could think the undef went another way
1468 // leading to an inconsistent set of conclusions.
1469 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1470 BI->setCondition(ConstantInt::getFalse(BI->getContext()));
1472 SwitchInst *SI = cast<SwitchInst>(TI);
1473 SI->setCondition(SI->getCaseValue(1));
1484 //===--------------------------------------------------------------------===//
1486 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1487 /// Sparse Conditional Constant Propagator.
1489 struct SCCP : public FunctionPass {
1490 static char ID; // Pass identification, replacement for typeid
1491 SCCP() : FunctionPass(&ID) {}
1493 // runOnFunction - Run the Sparse Conditional Constant Propagation
1494 // algorithm, and return true if the function was modified.
1496 bool runOnFunction(Function &F);
1498 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1499 AU.setPreservesCFG();
1502 } // end anonymous namespace
1505 static RegisterPass<SCCP>
1506 X("sccp", "Sparse Conditional Constant Propagation");
1508 // createSCCPPass - This is the public interface to this file.
1509 FunctionPass *llvm::createSCCPPass() {
1513 static void DeleteInstructionInBlock(BasicBlock *BB) {
1514 DEBUG(errs() << " BasicBlock Dead:" << *BB);
1517 // Delete the instructions backwards, as it has a reduced likelihood of
1518 // having to update as many def-use and use-def chains.
1519 while (!isa<TerminatorInst>(BB->begin())) {
1520 Instruction *I = --BasicBlock::iterator(BB->getTerminator());
1522 if (!I->use_empty())
1523 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1524 BB->getInstList().erase(I);
1529 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1530 // and return true if the function was modified.
1532 bool SCCP::runOnFunction(Function &F) {
1533 DEBUG(errs() << "SCCP on function '" << F.getName() << "'\n");
1534 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1536 // Mark the first block of the function as being executable.
1537 Solver.MarkBlockExecutable(F.begin());
1539 // Mark all arguments to the function as being overdefined.
1540 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1541 Solver.markOverdefined(AI);
1543 // Solve for constants.
1544 bool ResolvedUndefs = true;
1545 while (ResolvedUndefs) {
1547 DEBUG(errs() << "RESOLVING UNDEFs\n");
1548 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1551 bool MadeChanges = false;
1553 // If we decided that there are basic blocks that are dead in this function,
1554 // delete their contents now. Note that we cannot actually delete the blocks,
1555 // as we cannot modify the CFG of the function.
1557 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1558 if (!Solver.isBlockExecutable(BB)) {
1559 DeleteInstructionInBlock(BB);
1564 // Iterate over all of the instructions in a function, replacing them with
1565 // constants if we have found them to be of constant values.
1567 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1568 Instruction *Inst = BI++;
1569 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst))
1572 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1573 if (IV.isOverdefined())
1576 Constant *Const = IV.isConstant()
1577 ? IV.getConstant() : UndefValue::get(Inst->getType());
1578 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1580 // Replaces all of the uses of a variable with uses of the constant.
1581 Inst->replaceAllUsesWith(Const);
1583 // Delete the instruction.
1584 Inst->eraseFromParent();
1586 // Hey, we just changed something!
1596 //===--------------------------------------------------------------------===//
1598 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1599 /// Constant Propagation.
1601 struct IPSCCP : public ModulePass {
1603 IPSCCP() : ModulePass(&ID) {}
1604 bool runOnModule(Module &M);
1606 } // end anonymous namespace
1608 char IPSCCP::ID = 0;
1609 static RegisterPass<IPSCCP>
1610 Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation");
1612 // createIPSCCPPass - This is the public interface to this file.
1613 ModulePass *llvm::createIPSCCPPass() {
1614 return new IPSCCP();
1618 static bool AddressIsTaken(GlobalValue *GV) {
1619 // Delete any dead constantexpr klingons.
1620 GV->removeDeadConstantUsers();
1622 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
1624 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1625 if (SI->getOperand(0) == GV || SI->isVolatile())
1626 return true; // Storing addr of GV.
1627 } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) {
1628 // Make sure we are calling the function, not passing the address.
1629 if (UI.getOperandNo() != 0)
1631 } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1632 if (LI->isVolatile())
1634 } else if (isa<BlockAddress>(*UI)) {
1635 // blockaddress doesn't take the address of the function, it takes addr
1643 bool IPSCCP::runOnModule(Module &M) {
1644 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1646 // Loop over all functions, marking arguments to those with their addresses
1647 // taken or that are external as overdefined.
1649 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1650 if (!F->hasLocalLinkage() || AddressIsTaken(F)) {
1651 if (!F->isDeclaration())
1652 Solver.MarkBlockExecutable(F->begin());
1653 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1655 Solver.markOverdefined(AI);
1657 Solver.AddTrackedFunction(F);
1660 // Loop over global variables. We inform the solver about any internal global
1661 // variables that do not have their 'addresses taken'. If they don't have
1662 // their addresses taken, we can propagate constants through them.
1663 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1665 if (!G->isConstant() && G->hasLocalLinkage() && !AddressIsTaken(G))
1666 Solver.TrackValueOfGlobalVariable(G);
1668 // Solve for constants.
1669 bool ResolvedUndefs = true;
1670 while (ResolvedUndefs) {
1673 DEBUG(errs() << "RESOLVING UNDEFS\n");
1674 ResolvedUndefs = false;
1675 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1676 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1679 bool MadeChanges = false;
1681 // Iterate over all of the instructions in the module, replacing them with
1682 // constants if we have found them to be of constant values.
1684 SmallVector<BasicBlock*, 512> BlocksToErase;
1686 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1687 if (Solver.isBlockExecutable(F->begin())) {
1688 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1690 if (AI->use_empty()) continue;
1692 LatticeVal IV = Solver.getLatticeValueFor(AI);
1693 if (IV.isOverdefined()) continue;
1695 Constant *CST = IV.isConstant() ?
1696 IV.getConstant() : UndefValue::get(AI->getType());
1697 DEBUG(errs() << "*** Arg " << *AI << " = " << *CST <<"\n");
1699 // Replaces all of the uses of a variable with uses of the
1701 AI->replaceAllUsesWith(CST);
1706 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1707 if (!Solver.isBlockExecutable(BB)) {
1708 DeleteInstructionInBlock(BB);
1711 TerminatorInst *TI = BB->getTerminator();
1712 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1713 BasicBlock *Succ = TI->getSuccessor(i);
1714 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1715 TI->getSuccessor(i)->removePredecessor(BB);
1717 if (!TI->use_empty())
1718 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1719 TI->eraseFromParent();
1721 if (&*BB != &F->front())
1722 BlocksToErase.push_back(BB);
1724 new UnreachableInst(M.getContext(), BB);
1728 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1729 Instruction *Inst = BI++;
1730 if (Inst->getType()->isVoidTy())
1733 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1734 if (IV.isOverdefined())
1737 Constant *Const = IV.isConstant()
1738 ? IV.getConstant() : UndefValue::get(Inst->getType());
1739 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1741 // Replaces all of the uses of a variable with uses of the
1743 Inst->replaceAllUsesWith(Const);
1745 // Delete the instruction.
1746 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst))
1747 Inst->eraseFromParent();
1749 // Hey, we just changed something!
1755 // Now that all instructions in the function are constant folded, erase dead
1756 // blocks, because we can now use ConstantFoldTerminator to get rid of
1758 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1759 // If there are any PHI nodes in this successor, drop entries for BB now.
1760 BasicBlock *DeadBB = BlocksToErase[i];
1761 while (!DeadBB->use_empty()) {
1762 Instruction *I = cast<Instruction>(DeadBB->use_back());
1763 bool Folded = ConstantFoldTerminator(I->getParent());
1765 // The constant folder may not have been able to fold the terminator
1766 // if this is a branch or switch on undef. Fold it manually as a
1767 // branch to the first successor.
1769 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1770 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1771 "Branch should be foldable!");
1772 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1773 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1775 llvm_unreachable("Didn't fold away reference to block!");
1779 // Make this an uncond branch to the first successor.
1780 TerminatorInst *TI = I->getParent()->getTerminator();
1781 BranchInst::Create(TI->getSuccessor(0), TI);
1783 // Remove entries in successor phi nodes to remove edges.
1784 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1785 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1787 // Remove the old terminator.
1788 TI->eraseFromParent();
1792 // Finally, delete the basic block.
1793 F->getBasicBlockList().erase(DeadBB);
1795 BlocksToErase.clear();
1798 // If we inferred constant or undef return values for a function, we replaced
1799 // all call uses with the inferred value. This means we don't need to bother
1800 // actually returning anything from the function. Replace all return
1801 // instructions with return undef.
1802 // TODO: Process multiple value ret instructions also.
1803 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1804 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1805 E = RV.end(); I != E; ++I)
1806 if (!I->second.isOverdefined() &&
1807 !I->first->getReturnType()->isVoidTy()) {
1808 Function *F = I->first;
1809 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1810 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1811 if (!isa<UndefValue>(RI->getOperand(0)))
1812 RI->setOperand(0, UndefValue::get(F->getReturnType()));
1815 // If we infered constant or undef values for globals variables, we can delete
1816 // the global and any stores that remain to it.
1817 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1818 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1819 E = TG.end(); I != E; ++I) {
1820 GlobalVariable *GV = I->first;
1821 assert(!I->second.isOverdefined() &&
1822 "Overdefined values should have been taken out of the map!");
1823 DEBUG(errs() << "Found that GV '" << GV->getName() << "' is constant!\n");
1824 while (!GV->use_empty()) {
1825 StoreInst *SI = cast<StoreInst>(GV->use_back());
1826 SI->eraseFromParent();
1828 M.getGlobalList().erase(GV);