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/ValueTracking.h"
29 #include "llvm/Transforms/Utils/Local.h"
30 #include "llvm/Target/TargetData.h"
31 #include "llvm/Support/CallSite.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/InstVisitor.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/ADT/DenseMap.h"
37 #include "llvm/ADT/DenseSet.h"
38 #include "llvm/ADT/PointerIntPair.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/SmallVector.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/ADT/STLExtras.h"
47 STATISTIC(NumInstRemoved, "Number of instructions removed");
48 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
50 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP");
51 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
52 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
55 /// LatticeVal class - This class represents the different lattice values that
56 /// an LLVM value may occupy. It is a simple class with value semantics.
60 /// undefined - This LLVM Value has no known value yet.
63 /// constant - This LLVM Value has a specific constant value.
66 /// forcedconstant - This LLVM Value was thought to be undef until
67 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged
68 /// with another (different) constant, it goes to overdefined, instead of
72 /// overdefined - This instruction is not known to be constant, and we know
77 /// Val: This stores the current lattice value along with the Constant* for
78 /// the constant if this is a 'constant' or 'forcedconstant' value.
79 PointerIntPair<Constant *, 2, LatticeValueTy> Val;
81 LatticeValueTy getLatticeValue() const {
86 LatticeVal() : Val(0, undefined) {}
88 bool isUndefined() const { return getLatticeValue() == undefined; }
89 bool isConstant() const {
90 return getLatticeValue() == constant || getLatticeValue() == forcedconstant;
92 bool isOverdefined() const { return getLatticeValue() == overdefined; }
94 Constant *getConstant() const {
95 assert(isConstant() && "Cannot get the constant of a non-constant!");
96 return Val.getPointer();
99 /// markOverdefined - Return true if this is a change in status.
100 bool markOverdefined() {
104 Val.setInt(overdefined);
108 /// markConstant - Return true if this is a change in status.
109 bool markConstant(Constant *V) {
110 if (getLatticeValue() == constant) { // Constant but not forcedconstant.
111 assert(getConstant() == V && "Marking constant with different value");
116 Val.setInt(constant);
117 assert(V && "Marking constant with NULL");
120 assert(getLatticeValue() == forcedconstant &&
121 "Cannot move from overdefined to constant!");
122 // Stay at forcedconstant if the constant is the same.
123 if (V == getConstant()) return false;
125 // Otherwise, we go to overdefined. Assumptions made based on the
126 // forced value are possibly wrong. Assuming this is another constant
127 // could expose a contradiction.
128 Val.setInt(overdefined);
133 /// getConstantInt - If this is a constant with a ConstantInt value, return it
134 /// otherwise return null.
135 ConstantInt *getConstantInt() const {
137 return dyn_cast<ConstantInt>(getConstant());
141 void markForcedConstant(Constant *V) {
142 assert(isUndefined() && "Can't force a defined value!");
143 Val.setInt(forcedconstant);
147 } // end anonymous namespace.
152 //===----------------------------------------------------------------------===//
154 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional
155 /// Constant Propagation.
157 class SCCPSolver : public InstVisitor<SCCPSolver> {
158 const TargetData *TD;
159 SmallPtrSet<BasicBlock*, 8> BBExecutable;// The BBs that are executable.
160 DenseMap<Value*, LatticeVal> ValueState; // The state each value is in.
162 /// GlobalValue - If we are tracking any values for the contents of a global
163 /// variable, we keep a mapping from the constant accessor to the element of
164 /// the global, to the currently known value. If the value becomes
165 /// overdefined, it's entry is simply removed from this map.
166 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
168 /// TrackedRetVals - If we are tracking arguments into and the return
169 /// value out of a function, it will have an entry in this map, indicating
170 /// what the known return value for the function is.
171 DenseMap<Function*, LatticeVal> TrackedRetVals;
173 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
174 /// that return multiple values.
175 DenseMap<std::pair<Function*, unsigned>, LatticeVal> TrackedMultipleRetVals;
177 /// TrackingIncomingArguments - This is the set of functions that are
178 SmallPtrSet<Function*, 16> TrackingIncomingArguments;
180 /// The reason for two worklists is that overdefined is the lowest state
181 /// on the lattice, and moving things to overdefined as fast as possible
182 /// makes SCCP converge much faster.
184 /// By having a separate worklist, we accomplish this because everything
185 /// possibly overdefined will become overdefined at the soonest possible
187 SmallVector<Value*, 64> OverdefinedInstWorkList;
188 SmallVector<Value*, 64> InstWorkList;
191 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list
193 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
194 /// overdefined, despite the fact that the PHI node is overdefined.
195 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
197 /// KnownFeasibleEdges - Entries in this set are edges which have already had
198 /// PHI nodes retriggered.
199 typedef std::pair<BasicBlock*, BasicBlock*> Edge;
200 DenseSet<Edge> KnownFeasibleEdges;
202 SCCPSolver(const TargetData *td) : TD(td) {}
204 /// MarkBlockExecutable - This method can be used by clients to mark all of
205 /// the blocks that are known to be intrinsically live in the processed unit.
207 /// This returns true if the block was not considered live before.
208 bool MarkBlockExecutable(BasicBlock *BB) {
209 if (!BBExecutable.insert(BB)) return false;
210 DEBUG(errs() << "Marking Block Executable: " << BB->getName() << "\n");
211 BBWorkList.push_back(BB); // Add the block to the work list!
215 /// TrackValueOfGlobalVariable - Clients can use this method to
216 /// inform the SCCPSolver that it should track loads and stores to the
217 /// specified global variable if it can. This is only legal to call if
218 /// performing Interprocedural SCCP.
219 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
220 const Type *ElTy = GV->getType()->getElementType();
221 if (ElTy->isFirstClassType()) {
222 LatticeVal &IV = TrackedGlobals[GV];
223 if (!isa<UndefValue>(GV->getInitializer()))
224 IV.markConstant(GV->getInitializer());
228 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
229 /// and out of the specified function (which cannot have its address taken),
230 /// this method must be called.
231 void AddTrackedFunction(Function *F) {
232 // Add an entry, F -> undef.
233 if (const StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
234 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
235 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
238 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
241 void AddArgumentTrackedFunction(Function *F) {
242 TrackingIncomingArguments.insert(F);
245 /// Solve - Solve for constants and executable blocks.
249 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
250 /// that branches on undef values cannot reach any of their successors.
251 /// However, this is not a safe assumption. After we solve dataflow, this
252 /// method should be use to handle this. If this returns true, the solver
254 bool ResolvedUndefsIn(Function &F);
256 bool isBlockExecutable(BasicBlock *BB) const {
257 return BBExecutable.count(BB);
260 LatticeVal getLatticeValueFor(Value *V) const {
261 DenseMap<Value*, LatticeVal>::const_iterator I = ValueState.find(V);
262 assert(I != ValueState.end() && "V is not in valuemap!");
266 /// getTrackedRetVals - Get the inferred return value map.
268 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
269 return TrackedRetVals;
272 /// getTrackedGlobals - Get and return the set of inferred initializers for
273 /// global variables.
274 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
275 return TrackedGlobals;
278 void markOverdefined(Value *V) {
279 markOverdefined(ValueState[V], V);
283 // markConstant - Make a value be marked as "constant". If the value
284 // is not already a constant, add it to the instruction work list so that
285 // the users of the instruction are updated later.
287 void markConstant(LatticeVal &IV, Value *V, Constant *C) {
288 if (!IV.markConstant(C)) return;
289 DEBUG(errs() << "markConstant: " << *C << ": " << *V << '\n');
290 InstWorkList.push_back(V);
293 void markConstant(Value *V, Constant *C) {
294 markConstant(ValueState[V], V, C);
297 void markForcedConstant(Value *V, Constant *C) {
298 ValueState[V].markForcedConstant(C);
299 DEBUG(errs() << "markForcedConstant: " << *C << ": " << *V << '\n');
300 InstWorkList.push_back(V);
304 // markOverdefined - Make a value be marked as "overdefined". If the
305 // value is not already overdefined, add it to the overdefined instruction
306 // work list so that the users of the instruction are updated later.
307 void markOverdefined(LatticeVal &IV, Value *V) {
308 if (!IV.markOverdefined()) return;
310 DEBUG(errs() << "markOverdefined: ";
311 if (Function *F = dyn_cast<Function>(V))
312 errs() << "Function '" << F->getName() << "'\n";
314 errs() << *V << '\n');
315 // Only instructions go on the work list
316 OverdefinedInstWorkList.push_back(V);
319 void mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) {
320 if (IV.isOverdefined() || MergeWithV.isUndefined())
322 if (MergeWithV.isOverdefined())
323 markOverdefined(IV, V);
324 else if (IV.isUndefined())
325 markConstant(IV, V, MergeWithV.getConstant());
326 else if (IV.getConstant() != MergeWithV.getConstant())
327 markOverdefined(IV, V);
330 void mergeInValue(Value *V, LatticeVal MergeWithV) {
331 mergeInValue(ValueState[V], V, MergeWithV);
335 /// getValueState - Return the LatticeVal object that corresponds to the
336 /// value. This function handles the case when the value hasn't been seen yet
337 /// by properly seeding constants etc.
338 LatticeVal &getValueState(Value *V) {
339 DenseMap<Value*, LatticeVal>::iterator I = ValueState.find(V);
340 if (I != ValueState.end()) return I->second; // Common case, in the map
342 LatticeVal &LV = ValueState[V];
344 if (Constant *C = dyn_cast<Constant>(V)) {
345 // Undef values remain undefined.
346 if (!isa<UndefValue>(V))
347 LV.markConstant(C); // Constants are constant
350 // All others are underdefined by default.
354 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
355 /// work list if it is not already executable.
356 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
357 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
358 return; // This edge is already known to be executable!
360 if (!MarkBlockExecutable(Dest)) {
361 // If the destination is already executable, we just made an *edge*
362 // feasible that wasn't before. Revisit the PHI nodes in the block
363 // because they have potentially new operands.
364 DEBUG(errs() << "Marking Edge Executable: " << Source->getName()
365 << " -> " << Dest->getName() << "\n");
368 for (BasicBlock::iterator I = Dest->begin();
369 (PN = dyn_cast<PHINode>(I)); ++I)
374 // getFeasibleSuccessors - Return a vector of booleans to indicate which
375 // successors are reachable from a given terminator instruction.
377 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
379 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
380 // block to the 'To' basic block is currently feasible.
382 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
384 // OperandChangedState - This method is invoked on all of the users of an
385 // instruction that was just changed state somehow. Based on this
386 // information, we need to update the specified user of this instruction.
388 void OperandChangedState(Instruction *I) {
389 if (BBExecutable.count(I->getParent())) // Inst is executable?
393 /// RemoveFromOverdefinedPHIs - If I has any entries in the
394 /// UsersOfOverdefinedPHIs map for PN, remove them now.
395 void RemoveFromOverdefinedPHIs(Instruction *I, PHINode *PN) {
396 if (UsersOfOverdefinedPHIs.empty()) return;
397 std::multimap<PHINode*, Instruction*>::iterator It, E;
398 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN);
401 UsersOfOverdefinedPHIs.erase(It++);
408 friend class InstVisitor<SCCPSolver>;
410 // visit implementations - Something changed in this instruction. Either an
411 // operand made a transition, or the instruction is newly executable. Change
412 // the value type of I to reflect these changes if appropriate.
413 void visitPHINode(PHINode &I);
416 void visitReturnInst(ReturnInst &I);
417 void visitTerminatorInst(TerminatorInst &TI);
419 void visitCastInst(CastInst &I);
420 void visitSelectInst(SelectInst &I);
421 void visitBinaryOperator(Instruction &I);
422 void visitCmpInst(CmpInst &I);
423 void visitExtractElementInst(ExtractElementInst &I);
424 void visitInsertElementInst(InsertElementInst &I);
425 void visitShuffleVectorInst(ShuffleVectorInst &I);
426 void visitExtractValueInst(ExtractValueInst &EVI);
427 void visitInsertValueInst(InsertValueInst &IVI);
429 // Instructions that cannot be folded away.
430 void visitStoreInst (StoreInst &I);
431 void visitLoadInst (LoadInst &I);
432 void visitGetElementPtrInst(GetElementPtrInst &I);
433 void visitCallInst (CallInst &I) {
434 visitCallSite(CallSite::get(&I));
436 void visitInvokeInst (InvokeInst &II) {
437 visitCallSite(CallSite::get(&II));
438 visitTerminatorInst(II);
440 void visitCallSite (CallSite CS);
441 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ }
442 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
443 void visitAllocaInst (Instruction &I) { markOverdefined(&I); }
444 void visitVANextInst (Instruction &I) { markOverdefined(&I); }
445 void visitVAArgInst (Instruction &I) { markOverdefined(&I); }
447 void visitInstruction(Instruction &I) {
448 // If a new instruction is added to LLVM that we don't handle.
449 errs() << "SCCP: Don't know how to handle: " << I;
450 markOverdefined(&I); // Just in case
454 } // end anonymous namespace
457 // getFeasibleSuccessors - Return a vector of booleans to indicate which
458 // successors are reachable from a given terminator instruction.
460 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
461 SmallVector<bool, 16> &Succs) {
462 Succs.resize(TI.getNumSuccessors());
463 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
464 if (BI->isUnconditional()) {
469 LatticeVal BCValue = getValueState(BI->getCondition());
470 ConstantInt *CI = BCValue.getConstantInt();
472 // Overdefined condition variables, and branches on unfoldable constant
473 // conditions, mean the branch could go either way.
474 if (!BCValue.isUndefined())
475 Succs[0] = Succs[1] = true;
479 // Constant condition variables mean the branch can only go a single way.
480 Succs[CI->isZero()] = true;
484 if (isa<InvokeInst>(TI)) {
485 // Invoke instructions successors are always executable.
486 Succs[0] = Succs[1] = true;
490 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
491 LatticeVal SCValue = getValueState(SI->getCondition());
492 ConstantInt *CI = SCValue.getConstantInt();
494 if (CI == 0) { // Overdefined or undefined condition?
495 // All destinations are executable!
496 if (!SCValue.isUndefined())
497 Succs.assign(TI.getNumSuccessors(), true);
501 Succs[SI->findCaseValue(CI)] = true;
505 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
506 if (isa<IndirectBrInst>(&TI)) {
507 // Just mark all destinations executable!
508 Succs.assign(TI.getNumSuccessors(), true);
513 errs() << "Unknown terminator instruction: " << TI << '\n';
515 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
519 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
520 // block to the 'To' basic block is currently feasible.
522 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
523 assert(BBExecutable.count(To) && "Dest should always be alive!");
525 // Make sure the source basic block is executable!!
526 if (!BBExecutable.count(From)) return false;
528 // Check to make sure this edge itself is actually feasible now.
529 TerminatorInst *TI = From->getTerminator();
530 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
531 if (BI->isUnconditional())
534 LatticeVal BCValue = getValueState(BI->getCondition());
536 // Overdefined condition variables mean the branch could go either way,
537 // undef conditions mean that neither edge is feasible yet.
538 ConstantInt *CI = BCValue.getConstantInt();
540 return !BCValue.isUndefined();
542 // Constant condition variables mean the branch can only go a single way.
543 return BI->getSuccessor(CI->isZero()) == To;
546 // Invoke instructions successors are always executable.
547 if (isa<InvokeInst>(TI))
550 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
551 LatticeVal SCValue = getValueState(SI->getCondition());
552 ConstantInt *CI = SCValue.getConstantInt();
555 return !SCValue.isUndefined();
557 // Make sure to skip the "default value" which isn't a value
558 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i)
559 if (SI->getSuccessorValue(i) == CI) // Found the taken branch.
560 return SI->getSuccessor(i) == To;
562 // If the constant value is not equal to any of the branches, we must
563 // execute default branch.
564 return SI->getDefaultDest() == To;
567 // Just mark all destinations executable!
568 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
569 if (isa<IndirectBrInst>(&TI))
573 errs() << "Unknown terminator instruction: " << *TI << '\n';
578 // visit Implementations - Something changed in this instruction, either an
579 // operand made a transition, or the instruction is newly executable. Change
580 // the value type of I to reflect these changes if appropriate. This method
581 // makes sure to do the following actions:
583 // 1. If a phi node merges two constants in, and has conflicting value coming
584 // from different branches, or if the PHI node merges in an overdefined
585 // value, then the PHI node becomes overdefined.
586 // 2. If a phi node merges only constants in, and they all agree on value, the
587 // PHI node becomes a constant value equal to that.
588 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
589 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
590 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
591 // 6. If a conditional branch has a value that is constant, make the selected
592 // destination executable
593 // 7. If a conditional branch has a value that is overdefined, make all
594 // successors executable.
596 void SCCPSolver::visitPHINode(PHINode &PN) {
597 if (getValueState(&PN).isOverdefined()) {
598 // There may be instructions using this PHI node that are not overdefined
599 // themselves. If so, make sure that they know that the PHI node operand
601 std::multimap<PHINode*, Instruction*>::iterator I, E;
602 tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN);
606 SmallVector<Instruction*, 16> Users;
608 Users.push_back(I->second);
609 while (!Users.empty())
610 visit(Users.pop_back_val());
611 return; // Quick exit
614 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
615 // and slow us down a lot. Just mark them overdefined.
616 if (PN.getNumIncomingValues() > 64)
617 return markOverdefined(&PN);
619 // Look at all of the executable operands of the PHI node. If any of them
620 // are overdefined, the PHI becomes overdefined as well. If they are all
621 // constant, and they agree with each other, the PHI becomes the identical
622 // constant. If they are constant and don't agree, the PHI is overdefined.
623 // If there are no executable operands, the PHI remains undefined.
625 Constant *OperandVal = 0;
626 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
627 LatticeVal IV = getValueState(PN.getIncomingValue(i));
628 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
630 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
633 if (IV.isOverdefined()) // PHI node becomes overdefined!
634 return markOverdefined(&PN);
636 if (OperandVal == 0) { // Grab the first value.
637 OperandVal = IV.getConstant();
641 // There is already a reachable operand. If we conflict with it,
642 // then the PHI node becomes overdefined. If we agree with it, we
645 // Check to see if there are two different constants merging, if so, the PHI
646 // node is overdefined.
647 if (IV.getConstant() != OperandVal)
648 return markOverdefined(&PN);
651 // If we exited the loop, this means that the PHI node only has constant
652 // arguments that agree with each other(and OperandVal is the constant) or
653 // OperandVal is null because there are no defined incoming arguments. If
654 // this is the case, the PHI remains undefined.
657 markConstant(&PN, OperandVal); // Acquire operand value
663 void SCCPSolver::visitReturnInst(ReturnInst &I) {
664 if (I.getNumOperands() == 0) return; // ret void
666 Function *F = I.getParent()->getParent();
668 // If we are tracking the return value of this function, merge it in.
669 if (!TrackedRetVals.empty()) {
670 DenseMap<Function*, LatticeVal>::iterator TFRVI =
671 TrackedRetVals.find(F);
672 if (TFRVI != TrackedRetVals.end()) {
673 mergeInValue(TFRVI->second, F, getValueState(I.getOperand(0)));
678 // Handle functions that return multiple values.
679 if (!TrackedMultipleRetVals.empty() &&
680 isa<StructType>(I.getOperand(0)->getType())) {
681 for (unsigned i = 0, e = I.getOperand(0)->getType()->getNumContainedTypes();
683 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
684 It = TrackedMultipleRetVals.find(std::make_pair(F, i));
685 if (It == TrackedMultipleRetVals.end()) break;
686 if (Value *Val = FindInsertedValue(I.getOperand(0), i, I.getContext()))
687 mergeInValue(It->second, F, getValueState(Val));
692 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
693 SmallVector<bool, 16> SuccFeasible;
694 getFeasibleSuccessors(TI, SuccFeasible);
696 BasicBlock *BB = TI.getParent();
698 // Mark all feasible successors executable.
699 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
701 markEdgeExecutable(BB, TI.getSuccessor(i));
704 void SCCPSolver::visitCastInst(CastInst &I) {
705 LatticeVal OpSt = getValueState(I.getOperand(0));
706 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand
708 else if (OpSt.isConstant()) // Propagate constant value
709 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
710 OpSt.getConstant(), I.getType()));
713 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
714 Value *Aggr = EVI.getAggregateOperand();
716 // If the operand to the extractvalue is an undef, the result is undef.
717 if (isa<UndefValue>(Aggr))
720 // Currently only handle single-index extractvalues.
721 if (EVI.getNumIndices() != 1)
722 return markOverdefined(&EVI);
725 if (CallInst *CI = dyn_cast<CallInst>(Aggr))
726 F = CI->getCalledFunction();
727 else if (InvokeInst *II = dyn_cast<InvokeInst>(Aggr))
728 F = II->getCalledFunction();
730 // TODO: If IPSCCP resolves the callee of this function, we could propagate a
732 if (F == 0 || TrackedMultipleRetVals.empty())
733 return markOverdefined(&EVI);
735 // See if we are tracking the result of the callee. If not tracking this
736 // function (for example, it is a declaration) just move to overdefined.
737 if (!TrackedMultipleRetVals.count(std::make_pair(F, *EVI.idx_begin())))
738 return markOverdefined(&EVI);
740 // Otherwise, the value will be merged in here as a result of CallSite
744 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
745 Value *Aggr = IVI.getAggregateOperand();
746 Value *Val = IVI.getInsertedValueOperand();
748 // If the operands to the insertvalue are undef, the result is undef.
749 if (isa<UndefValue>(Aggr) && isa<UndefValue>(Val))
752 // Currently only handle single-index insertvalues.
753 if (IVI.getNumIndices() != 1)
754 return markOverdefined(&IVI);
756 // Currently only handle insertvalue instructions that are in a single-use
757 // chain that builds up a return value.
758 for (const InsertValueInst *TmpIVI = &IVI; ; ) {
759 if (!TmpIVI->hasOneUse())
760 return markOverdefined(&IVI);
762 const Value *V = *TmpIVI->use_begin();
763 if (isa<ReturnInst>(V))
765 TmpIVI = dyn_cast<InsertValueInst>(V);
767 return markOverdefined(&IVI);
770 // See if we are tracking the result of the callee.
771 Function *F = IVI.getParent()->getParent();
772 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
773 It = TrackedMultipleRetVals.find(std::make_pair(F, *IVI.idx_begin()));
775 // Merge in the inserted member value.
776 if (It != TrackedMultipleRetVals.end())
777 mergeInValue(It->second, F, getValueState(Val));
779 // Mark the aggregate result of the IVI overdefined; any tracking that we do
780 // will be done on the individual member values.
781 markOverdefined(&IVI);
784 void SCCPSolver::visitSelectInst(SelectInst &I) {
785 LatticeVal CondValue = getValueState(I.getCondition());
786 if (CondValue.isUndefined())
789 if (ConstantInt *CondCB = CondValue.getConstantInt()) {
790 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
791 mergeInValue(&I, getValueState(OpVal));
795 // Otherwise, the condition is overdefined or a constant we can't evaluate.
796 // See if we can produce something better than overdefined based on the T/F
798 LatticeVal TVal = getValueState(I.getTrueValue());
799 LatticeVal FVal = getValueState(I.getFalseValue());
801 // select ?, C, C -> C.
802 if (TVal.isConstant() && FVal.isConstant() &&
803 TVal.getConstant() == FVal.getConstant())
804 return markConstant(&I, FVal.getConstant());
806 if (TVal.isUndefined()) // select ?, undef, X -> X.
807 return mergeInValue(&I, FVal);
808 if (FVal.isUndefined()) // select ?, X, undef -> X.
809 return mergeInValue(&I, TVal);
813 // Handle Binary Operators.
814 void SCCPSolver::visitBinaryOperator(Instruction &I) {
815 LatticeVal V1State = getValueState(I.getOperand(0));
816 LatticeVal V2State = getValueState(I.getOperand(1));
818 LatticeVal &IV = ValueState[&I];
819 if (IV.isOverdefined()) return;
821 if (V1State.isConstant() && V2State.isConstant())
822 return markConstant(IV, &I,
823 ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
824 V2State.getConstant()));
826 // If something is undef, wait for it to resolve.
827 if (!V1State.isOverdefined() && !V2State.isOverdefined())
830 // Otherwise, one of our operands is overdefined. Try to produce something
831 // better than overdefined with some tricks.
833 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
834 // operand is overdefined.
835 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
836 LatticeVal *NonOverdefVal = 0;
837 if (!V1State.isOverdefined())
838 NonOverdefVal = &V1State;
839 else if (!V2State.isOverdefined())
840 NonOverdefVal = &V2State;
843 if (NonOverdefVal->isUndefined()) {
844 // Could annihilate value.
845 if (I.getOpcode() == Instruction::And)
846 markConstant(IV, &I, Constant::getNullValue(I.getType()));
847 else if (const VectorType *PT = dyn_cast<VectorType>(I.getType()))
848 markConstant(IV, &I, Constant::getAllOnesValue(PT));
851 Constant::getAllOnesValue(I.getType()));
855 if (I.getOpcode() == Instruction::And) {
857 if (NonOverdefVal->getConstant()->isNullValue())
858 return markConstant(IV, &I, NonOverdefVal->getConstant());
860 if (ConstantInt *CI = NonOverdefVal->getConstantInt())
861 if (CI->isAllOnesValue()) // X or -1 = -1
862 return markConstant(IV, &I, NonOverdefVal->getConstant());
868 // If both operands are PHI nodes, it is possible that this instruction has
869 // a constant value, despite the fact that the PHI node doesn't. Check for
870 // this condition now.
871 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
872 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
873 if (PN1->getParent() == PN2->getParent()) {
874 // Since the two PHI nodes are in the same basic block, they must have
875 // entries for the same predecessors. Walk the predecessor list, and
876 // if all of the incoming values are constants, and the result of
877 // evaluating this expression with all incoming value pairs is the
878 // same, then this expression is a constant even though the PHI node
879 // is not a constant!
881 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
882 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
883 BasicBlock *InBlock = PN1->getIncomingBlock(i);
884 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
886 if (In1.isOverdefined() || In2.isOverdefined()) {
887 Result.markOverdefined();
888 break; // Cannot fold this operation over the PHI nodes!
891 if (In1.isConstant() && In2.isConstant()) {
892 Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(),
894 if (Result.isUndefined())
895 Result.markConstant(V);
896 else if (Result.isConstant() && Result.getConstant() != V) {
897 Result.markOverdefined();
903 // If we found a constant value here, then we know the instruction is
904 // constant despite the fact that the PHI nodes are overdefined.
905 if (Result.isConstant()) {
906 markConstant(IV, &I, Result.getConstant());
907 // Remember that this instruction is virtually using the PHI node
909 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
910 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
914 if (Result.isUndefined())
917 // Okay, this really is overdefined now. Since we might have
918 // speculatively thought that this was not overdefined before, and
919 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
920 // make sure to clean out any entries that we put there, for
922 RemoveFromOverdefinedPHIs(&I, PN1);
923 RemoveFromOverdefinedPHIs(&I, PN2);
929 // Handle ICmpInst instruction.
930 void SCCPSolver::visitCmpInst(CmpInst &I) {
931 LatticeVal V1State = getValueState(I.getOperand(0));
932 LatticeVal V2State = getValueState(I.getOperand(1));
934 LatticeVal &IV = ValueState[&I];
935 if (IV.isOverdefined()) return;
937 if (V1State.isConstant() && V2State.isConstant())
938 return markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
939 V1State.getConstant(),
940 V2State.getConstant()));
942 // If operands are still undefined, wait for it to resolve.
943 if (!V1State.isOverdefined() && !V2State.isOverdefined())
946 // If something is overdefined, use some tricks to avoid ending up and over
947 // defined if we can.
949 // If both operands are PHI nodes, it is possible that this instruction has
950 // a constant value, despite the fact that the PHI node doesn't. Check for
951 // this condition now.
952 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
953 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
954 if (PN1->getParent() == PN2->getParent()) {
955 // Since the two PHI nodes are in the same basic block, they must have
956 // entries for the same predecessors. Walk the predecessor list, and
957 // if all of the incoming values are constants, and the result of
958 // evaluating this expression with all incoming value pairs is the
959 // same, then this expression is a constant even though the PHI node
960 // is not a constant!
962 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
963 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
964 BasicBlock *InBlock = PN1->getIncomingBlock(i);
965 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
967 if (In1.isOverdefined() || In2.isOverdefined()) {
968 Result.markOverdefined();
969 break; // Cannot fold this operation over the PHI nodes!
972 if (In1.isConstant() && In2.isConstant()) {
973 Constant *V = ConstantExpr::getCompare(I.getPredicate(),
976 if (Result.isUndefined())
977 Result.markConstant(V);
978 else if (Result.isConstant() && Result.getConstant() != V) {
979 Result.markOverdefined();
985 // If we found a constant value here, then we know the instruction is
986 // constant despite the fact that the PHI nodes are overdefined.
987 if (Result.isConstant()) {
988 markConstant(&I, Result.getConstant());
989 // Remember that this instruction is virtually using the PHI node
991 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
992 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
996 if (Result.isUndefined())
999 // Okay, this really is overdefined now. Since we might have
1000 // speculatively thought that this was not overdefined before, and
1001 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
1002 // make sure to clean out any entries that we put there, for
1004 RemoveFromOverdefinedPHIs(&I, PN1);
1005 RemoveFromOverdefinedPHIs(&I, PN2);
1008 markOverdefined(&I);
1011 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
1012 // FIXME : SCCP does not handle vectors properly.
1013 return markOverdefined(&I);
1016 LatticeVal &ValState = getValueState(I.getOperand(0));
1017 LatticeVal &IdxState = getValueState(I.getOperand(1));
1019 if (ValState.isOverdefined() || IdxState.isOverdefined())
1020 markOverdefined(&I);
1021 else if(ValState.isConstant() && IdxState.isConstant())
1022 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
1023 IdxState.getConstant()));
1027 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
1028 // FIXME : SCCP does not handle vectors properly.
1029 return markOverdefined(&I);
1031 LatticeVal &ValState = getValueState(I.getOperand(0));
1032 LatticeVal &EltState = getValueState(I.getOperand(1));
1033 LatticeVal &IdxState = getValueState(I.getOperand(2));
1035 if (ValState.isOverdefined() || EltState.isOverdefined() ||
1036 IdxState.isOverdefined())
1037 markOverdefined(&I);
1038 else if(ValState.isConstant() && EltState.isConstant() &&
1039 IdxState.isConstant())
1040 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
1041 EltState.getConstant(),
1042 IdxState.getConstant()));
1043 else if (ValState.isUndefined() && EltState.isConstant() &&
1044 IdxState.isConstant())
1045 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
1046 EltState.getConstant(),
1047 IdxState.getConstant()));
1051 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
1052 // FIXME : SCCP does not handle vectors properly.
1053 return markOverdefined(&I);
1055 LatticeVal &V1State = getValueState(I.getOperand(0));
1056 LatticeVal &V2State = getValueState(I.getOperand(1));
1057 LatticeVal &MaskState = getValueState(I.getOperand(2));
1059 if (MaskState.isUndefined() ||
1060 (V1State.isUndefined() && V2State.isUndefined()))
1061 return; // Undefined output if mask or both inputs undefined.
1063 if (V1State.isOverdefined() || V2State.isOverdefined() ||
1064 MaskState.isOverdefined()) {
1065 markOverdefined(&I);
1067 // A mix of constant/undef inputs.
1068 Constant *V1 = V1State.isConstant() ?
1069 V1State.getConstant() : UndefValue::get(I.getType());
1070 Constant *V2 = V2State.isConstant() ?
1071 V2State.getConstant() : UndefValue::get(I.getType());
1072 Constant *Mask = MaskState.isConstant() ?
1073 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
1074 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
1079 // Handle getelementptr instructions. If all operands are constants then we
1080 // can turn this into a getelementptr ConstantExpr.
1082 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1083 if (ValueState[&I].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(&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->isDeclaration()) {
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 (F && F->isDeclaration() && !isa<StructType>(I->getType()) &&
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 local function that doesn't have its address taken, mark its
1201 // entry block executable and merge in the actual arguments to the call into
1202 // the formal arguments of the function.
1203 if (!TrackingIncomingArguments.empty() && TrackingIncomingArguments.count(F)){
1204 MarkBlockExecutable(F->begin());
1206 // Propagate information from this call site into the callee.
1207 CallSite::arg_iterator CAI = CS.arg_begin();
1208 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1209 AI != E; ++AI, ++CAI) {
1210 // If this argument is byval, and if the function is not readonly, there
1211 // will be an implicit copy formed of the input aggregate.
1212 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1213 markOverdefined(AI);
1217 mergeInValue(AI, getValueState(*CAI));
1221 // If this is a single/zero retval case, see if we're tracking the function.
1222 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1223 if (TFRVI != TrackedRetVals.end()) {
1224 // If so, propagate the return value of the callee into this call result.
1225 mergeInValue(I, TFRVI->second);
1226 } else if (isa<StructType>(I->getType())) {
1227 // Check to see if we're tracking this callee, if not, handle it in the
1228 // common path above.
1229 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
1230 TMRVI = TrackedMultipleRetVals.find(std::make_pair(F, 0));
1231 if (TMRVI == TrackedMultipleRetVals.end())
1232 goto CallOverdefined;
1234 // Need to mark as overdefined, otherwise it stays undefined which
1235 // creates extractvalue undef, <idx>
1238 // If we are tracking this callee, propagate the return values of the call
1239 // into this call site. We do this by walking all the uses. Single-index
1240 // ExtractValueInst uses can be tracked; anything more complicated is
1241 // currently handled conservatively.
1242 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1244 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(*UI)) {
1245 if (EVI->getNumIndices() == 1) {
1247 TrackedMultipleRetVals[std::make_pair(F, *EVI->idx_begin())]);
1251 // The aggregate value is used in a way not handled here. Assume nothing.
1252 markOverdefined(*UI);
1255 // Otherwise we're not tracking this callee, so handle it in the
1256 // common path above.
1257 goto CallOverdefined;
1261 void SCCPSolver::Solve() {
1262 // Process the work lists until they are empty!
1263 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1264 !OverdefinedInstWorkList.empty()) {
1265 // Process the overdefined instruction's work list first, which drives other
1266 // things to overdefined more quickly.
1267 while (!OverdefinedInstWorkList.empty()) {
1268 Value *I = OverdefinedInstWorkList.pop_back_val();
1270 DEBUG(errs() << "\nPopped off OI-WL: " << *I << '\n');
1272 // "I" got into the work list because it either made the transition from
1273 // bottom to constant
1275 // Anything on this worklist that is overdefined need not be visited
1276 // since all of its users will have already been marked as overdefined
1277 // Update all of the users of this instruction's value.
1279 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1281 if (Instruction *I = dyn_cast<Instruction>(*UI))
1282 OperandChangedState(I);
1285 // Process the instruction work list.
1286 while (!InstWorkList.empty()) {
1287 Value *I = InstWorkList.pop_back_val();
1289 DEBUG(errs() << "\nPopped off I-WL: " << *I << '\n');
1291 // "I" got into the work list because it made the transition from undef to
1294 // Anything on this worklist that is overdefined need not be visited
1295 // since all of its users will have already been marked as overdefined.
1296 // Update all of the users of this instruction's value.
1298 if (!getValueState(I).isOverdefined())
1299 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1301 if (Instruction *I = dyn_cast<Instruction>(*UI))
1302 OperandChangedState(I);
1305 // Process the basic block work list.
1306 while (!BBWorkList.empty()) {
1307 BasicBlock *BB = BBWorkList.back();
1308 BBWorkList.pop_back();
1310 DEBUG(errs() << "\nPopped off BBWL: " << *BB << '\n');
1312 // Notify all instructions in this basic block that they are newly
1319 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1320 /// that branches on undef values cannot reach any of their successors.
1321 /// However, this is not a safe assumption. After we solve dataflow, this
1322 /// method should be use to handle this. If this returns true, the solver
1323 /// should be rerun.
1325 /// This method handles this by finding an unresolved branch and marking it one
1326 /// of the edges from the block as being feasible, even though the condition
1327 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1328 /// CFG and only slightly pessimizes the analysis results (by marking one,
1329 /// potentially infeasible, edge feasible). This cannot usefully modify the
1330 /// constraints on the condition of the branch, as that would impact other users
1333 /// This scan also checks for values that use undefs, whose results are actually
1334 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1335 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1336 /// even if X isn't defined.
1337 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1338 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1339 if (!BBExecutable.count(BB))
1342 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1343 // Look for instructions which produce undef values.
1344 if (I->getType()->isVoidTy()) continue;
1346 LatticeVal &LV = getValueState(I);
1347 if (!LV.isUndefined()) continue;
1349 // Get the lattice values of the first two operands for use below.
1350 LatticeVal Op0LV = getValueState(I->getOperand(0));
1352 if (I->getNumOperands() == 2) {
1353 // If this is a two-operand instruction, and if both operands are
1354 // undefs, the result stays undef.
1355 Op1LV = getValueState(I->getOperand(1));
1356 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1360 // If this is an instructions whose result is defined even if the input is
1361 // not fully defined, propagate the information.
1362 const Type *ITy = I->getType();
1363 switch (I->getOpcode()) {
1364 default: break; // Leave the instruction as an undef.
1365 case Instruction::ZExt:
1366 // After a zero extend, we know the top part is zero. SExt doesn't have
1367 // to be handled here, because we don't know whether the top part is 1's
1369 markForcedConstant(I, Constant::getNullValue(ITy));
1371 case Instruction::Mul:
1372 case Instruction::And:
1373 // undef * X -> 0. X could be zero.
1374 // undef & X -> 0. X could be zero.
1375 markForcedConstant(I, Constant::getNullValue(ITy));
1378 case Instruction::Or:
1379 // undef | X -> -1. X could be -1.
1380 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1383 case Instruction::SDiv:
1384 case Instruction::UDiv:
1385 case Instruction::SRem:
1386 case Instruction::URem:
1387 // X / undef -> undef. No change.
1388 // X % undef -> undef. No change.
1389 if (Op1LV.isUndefined()) break;
1391 // undef / X -> 0. X could be maxint.
1392 // undef % X -> 0. X could be 1.
1393 markForcedConstant(I, Constant::getNullValue(ITy));
1396 case Instruction::AShr:
1397 // undef >>s X -> undef. No change.
1398 if (Op0LV.isUndefined()) break;
1400 // X >>s undef -> X. X could be 0, X could have the high-bit known set.
1401 if (Op0LV.isConstant())
1402 markForcedConstant(I, Op0LV.getConstant());
1406 case Instruction::LShr:
1407 case Instruction::Shl:
1408 // undef >> X -> undef. No change.
1409 // undef << X -> undef. No change.
1410 if (Op0LV.isUndefined()) break;
1412 // X >> undef -> 0. X could be 0.
1413 // X << undef -> 0. X could be 0.
1414 markForcedConstant(I, Constant::getNullValue(ITy));
1416 case Instruction::Select:
1417 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1418 if (Op0LV.isUndefined()) {
1419 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1420 Op1LV = getValueState(I->getOperand(2));
1421 } else if (Op1LV.isUndefined()) {
1422 // c ? undef : undef -> undef. No change.
1423 Op1LV = getValueState(I->getOperand(2));
1424 if (Op1LV.isUndefined())
1426 // Otherwise, c ? undef : x -> x.
1428 // Leave Op1LV as Operand(1)'s LatticeValue.
1431 if (Op1LV.isConstant())
1432 markForcedConstant(I, Op1LV.getConstant());
1436 case Instruction::Call:
1437 // If a call has an undef result, it is because it is constant foldable
1438 // but one of the inputs was undef. Just force the result to
1445 TerminatorInst *TI = BB->getTerminator();
1446 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1447 if (!BI->isConditional()) continue;
1448 if (!getValueState(BI->getCondition()).isUndefined())
1450 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1451 if (SI->getNumSuccessors() < 2) // no cases
1453 if (!getValueState(SI->getCondition()).isUndefined())
1459 // If the edge to the second successor isn't thought to be feasible yet,
1460 // mark it so now. We pick the second one so that this goes to some
1461 // enumerated value in a switch instead of going to the default destination.
1462 if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(1))))
1465 // Otherwise, it isn't already thought to be feasible. Mark it as such now
1466 // and return. This will make other blocks reachable, which will allow new
1467 // values to be discovered and existing ones to be moved in the lattice.
1468 markEdgeExecutable(BB, TI->getSuccessor(1));
1470 // This must be a conditional branch of switch on undef. At this point,
1471 // force the old terminator to branch to the first successor. This is
1472 // required because we are now influencing the dataflow of the function with
1473 // the assumption that this edge is taken. If we leave the branch condition
1474 // as undef, then further analysis could think the undef went another way
1475 // leading to an inconsistent set of conclusions.
1476 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1477 BI->setCondition(ConstantInt::getFalse(BI->getContext()));
1479 SwitchInst *SI = cast<SwitchInst>(TI);
1480 SI->setCondition(SI->getCaseValue(1));
1491 //===--------------------------------------------------------------------===//
1493 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1494 /// Sparse Conditional Constant Propagator.
1496 struct SCCP : public FunctionPass {
1497 static char ID; // Pass identification, replacement for typeid
1498 SCCP() : FunctionPass(&ID) {}
1500 // runOnFunction - Run the Sparse Conditional Constant Propagation
1501 // algorithm, and return true if the function was modified.
1503 bool runOnFunction(Function &F);
1505 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1506 AU.setPreservesCFG();
1509 } // end anonymous namespace
1512 static RegisterPass<SCCP>
1513 X("sccp", "Sparse Conditional Constant Propagation");
1515 // createSCCPPass - This is the public interface to this file.
1516 FunctionPass *llvm::createSCCPPass() {
1520 static void DeleteInstructionInBlock(BasicBlock *BB) {
1521 DEBUG(errs() << " BasicBlock Dead:" << *BB);
1524 // Delete the instructions backwards, as it has a reduced likelihood of
1525 // having to update as many def-use and use-def chains.
1526 while (!isa<TerminatorInst>(BB->begin())) {
1527 Instruction *I = --BasicBlock::iterator(BB->getTerminator());
1529 if (!I->use_empty())
1530 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1531 BB->getInstList().erase(I);
1536 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1537 // and return true if the function was modified.
1539 bool SCCP::runOnFunction(Function &F) {
1540 DEBUG(errs() << "SCCP on function '" << F.getName() << "'\n");
1541 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1543 // Mark the first block of the function as being executable.
1544 Solver.MarkBlockExecutable(F.begin());
1546 // Mark all arguments to the function as being overdefined.
1547 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1548 Solver.markOverdefined(AI);
1550 // Solve for constants.
1551 bool ResolvedUndefs = true;
1552 while (ResolvedUndefs) {
1554 DEBUG(errs() << "RESOLVING UNDEFs\n");
1555 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1558 bool MadeChanges = false;
1560 // If we decided that there are basic blocks that are dead in this function,
1561 // delete their contents now. Note that we cannot actually delete the blocks,
1562 // as we cannot modify the CFG of the function.
1564 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1565 if (!Solver.isBlockExecutable(BB)) {
1566 DeleteInstructionInBlock(BB);
1571 // Iterate over all of the instructions in a function, replacing them with
1572 // constants if we have found them to be of constant values.
1574 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1575 Instruction *Inst = BI++;
1576 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst))
1579 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1580 if (IV.isOverdefined())
1583 Constant *Const = IV.isConstant()
1584 ? IV.getConstant() : UndefValue::get(Inst->getType());
1585 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1587 // Replaces all of the uses of a variable with uses of the constant.
1588 Inst->replaceAllUsesWith(Const);
1590 // Delete the instruction.
1591 Inst->eraseFromParent();
1593 // Hey, we just changed something!
1603 //===--------------------------------------------------------------------===//
1605 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1606 /// Constant Propagation.
1608 struct IPSCCP : public ModulePass {
1610 IPSCCP() : ModulePass(&ID) {}
1611 bool runOnModule(Module &M);
1613 } // end anonymous namespace
1615 char IPSCCP::ID = 0;
1616 static RegisterPass<IPSCCP>
1617 Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation");
1619 // createIPSCCPPass - This is the public interface to this file.
1620 ModulePass *llvm::createIPSCCPPass() {
1621 return new IPSCCP();
1625 static bool AddressIsTaken(GlobalValue *GV) {
1626 // Delete any dead constantexpr klingons.
1627 GV->removeDeadConstantUsers();
1629 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
1631 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1632 if (SI->getOperand(0) == GV || SI->isVolatile())
1633 return true; // Storing addr of GV.
1634 } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) {
1635 // Make sure we are calling the function, not passing the address.
1636 if (UI.getOperandNo() != 0)
1638 } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1639 if (LI->isVolatile())
1641 } else if (isa<BlockAddress>(*UI)) {
1642 // blockaddress doesn't take the address of the function, it takes addr
1650 bool IPSCCP::runOnModule(Module &M) {
1651 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1653 // Loop over all functions, marking arguments to those with their addresses
1654 // taken or that are external as overdefined.
1656 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1657 if (F->isDeclaration())
1660 // If this is a strong or ODR definition of this function, then we can
1661 // propagate information about its result into callsites of it.
1662 if (!F->mayBeOverridden() &&
1663 !isa<StructType>(F->getReturnType()))
1664 Solver.AddTrackedFunction(F);
1666 // If this function only has direct calls that we can see, we can track its
1667 // arguments and return value aggressively, and can assume it is not called
1668 // unless we see evidence to the contrary.
1669 if (F->hasLocalLinkage() && !AddressIsTaken(F)) {
1670 Solver.AddArgumentTrackedFunction(F);
1674 // Assume the function is called.
1675 Solver.MarkBlockExecutable(F->begin());
1677 // Assume nothing about the incoming arguments.
1678 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1680 Solver.markOverdefined(AI);
1683 // Loop over global variables. We inform the solver about any internal global
1684 // variables that do not have their 'addresses taken'. If they don't have
1685 // their addresses taken, we can propagate constants through them.
1686 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1688 if (!G->isConstant() && G->hasLocalLinkage() && !AddressIsTaken(G))
1689 Solver.TrackValueOfGlobalVariable(G);
1691 // Solve for constants.
1692 bool ResolvedUndefs = true;
1693 while (ResolvedUndefs) {
1696 DEBUG(errs() << "RESOLVING UNDEFS\n");
1697 ResolvedUndefs = false;
1698 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1699 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1702 bool MadeChanges = false;
1704 // Iterate over all of the instructions in the module, replacing them with
1705 // constants if we have found them to be of constant values.
1707 SmallVector<BasicBlock*, 512> BlocksToErase;
1709 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1710 if (Solver.isBlockExecutable(F->begin())) {
1711 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1713 if (AI->use_empty()) continue;
1715 LatticeVal IV = Solver.getLatticeValueFor(AI);
1716 if (IV.isOverdefined()) continue;
1718 Constant *CST = IV.isConstant() ?
1719 IV.getConstant() : UndefValue::get(AI->getType());
1720 DEBUG(errs() << "*** Arg " << *AI << " = " << *CST <<"\n");
1722 // Replaces all of the uses of a variable with uses of the
1724 AI->replaceAllUsesWith(CST);
1729 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1730 if (!Solver.isBlockExecutable(BB)) {
1731 DeleteInstructionInBlock(BB);
1734 TerminatorInst *TI = BB->getTerminator();
1735 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1736 BasicBlock *Succ = TI->getSuccessor(i);
1737 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1738 TI->getSuccessor(i)->removePredecessor(BB);
1740 if (!TI->use_empty())
1741 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1742 TI->eraseFromParent();
1744 if (&*BB != &F->front())
1745 BlocksToErase.push_back(BB);
1747 new UnreachableInst(M.getContext(), BB);
1751 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1752 Instruction *Inst = BI++;
1753 if (Inst->getType()->isVoidTy())
1756 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1757 if (IV.isOverdefined())
1760 Constant *Const = IV.isConstant()
1761 ? IV.getConstant() : UndefValue::get(Inst->getType());
1762 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1764 // Replaces all of the uses of a variable with uses of the
1766 Inst->replaceAllUsesWith(Const);
1768 // Delete the instruction.
1769 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst))
1770 Inst->eraseFromParent();
1772 // Hey, we just changed something!
1778 // Now that all instructions in the function are constant folded, erase dead
1779 // blocks, because we can now use ConstantFoldTerminator to get rid of
1781 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1782 // If there are any PHI nodes in this successor, drop entries for BB now.
1783 BasicBlock *DeadBB = BlocksToErase[i];
1784 while (!DeadBB->use_empty()) {
1785 Instruction *I = cast<Instruction>(DeadBB->use_back());
1786 bool Folded = ConstantFoldTerminator(I->getParent());
1788 // The constant folder may not have been able to fold the terminator
1789 // if this is a branch or switch on undef. Fold it manually as a
1790 // branch to the first successor.
1792 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1793 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1794 "Branch should be foldable!");
1795 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1796 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1798 llvm_unreachable("Didn't fold away reference to block!");
1802 // Make this an uncond branch to the first successor.
1803 TerminatorInst *TI = I->getParent()->getTerminator();
1804 BranchInst::Create(TI->getSuccessor(0), TI);
1806 // Remove entries in successor phi nodes to remove edges.
1807 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1808 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1810 // Remove the old terminator.
1811 TI->eraseFromParent();
1815 // Finally, delete the basic block.
1816 F->getBasicBlockList().erase(DeadBB);
1818 BlocksToErase.clear();
1821 // If we inferred constant or undef return values for a function, we replaced
1822 // all call uses with the inferred value. This means we don't need to bother
1823 // actually returning anything from the function. Replace all return
1824 // instructions with return undef.
1825 // TODO: Process multiple value ret instructions also.
1826 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1827 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1828 E = RV.end(); I != E; ++I) {
1829 Function *F = I->first;
1830 if (I->second.isOverdefined() || F->getReturnType()->isVoidTy())
1833 // We can only do this if we know that nothing else can call the function.
1834 if (!F->hasLocalLinkage() || AddressIsTaken(F))
1837 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1838 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1839 if (!isa<UndefValue>(RI->getOperand(0)))
1840 RI->setOperand(0, UndefValue::get(F->getReturnType()));
1843 // If we infered constant or undef values for globals variables, we can delete
1844 // the global and any stores that remain to it.
1845 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1846 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1847 E = TG.end(); I != E; ++I) {
1848 GlobalVariable *GV = I->first;
1849 assert(!I->second.isOverdefined() &&
1850 "Overdefined values should have been taken out of the map!");
1851 DEBUG(errs() << "Found that GV '" << GV->getName() << "' is constant!\n");
1852 while (!GV->use_empty()) {
1853 StoreInst *SI = cast<StoreInst>(GV->use_back());
1854 SI->eraseFromParent();
1856 M.getGlobalList().erase(GV);