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) {
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 /// The reason for two worklists is that overdefined is the lowest state
178 /// on the lattice, and moving things to overdefined as fast as possible
179 /// makes SCCP converge much faster.
181 /// By having a separate worklist, we accomplish this because everything
182 /// possibly overdefined will become overdefined at the soonest possible
184 SmallVector<Value*, 64> OverdefinedInstWorkList;
185 SmallVector<Value*, 64> InstWorkList;
188 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list
190 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
191 /// overdefined, despite the fact that the PHI node is overdefined.
192 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
194 /// KnownFeasibleEdges - Entries in this set are edges which have already had
195 /// PHI nodes retriggered.
196 typedef std::pair<BasicBlock*, BasicBlock*> Edge;
197 DenseSet<Edge> KnownFeasibleEdges;
199 SCCPSolver(const TargetData *td) : TD(td) {}
201 /// MarkBlockExecutable - This method can be used by clients to mark all of
202 /// the blocks that are known to be intrinsically live in the processed unit.
204 /// This returns true if the block was not considered live before.
205 bool MarkBlockExecutable(BasicBlock *BB) {
206 if (!BBExecutable.insert(BB)) return false;
207 DEBUG(errs() << "Marking Block Executable: " << BB->getName() << "\n");
208 BBWorkList.push_back(BB); // Add the block to the work list!
212 /// TrackValueOfGlobalVariable - Clients can use this method to
213 /// inform the SCCPSolver that it should track loads and stores to the
214 /// specified global variable if it can. This is only legal to call if
215 /// performing Interprocedural SCCP.
216 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
217 const Type *ElTy = GV->getType()->getElementType();
218 if (ElTy->isFirstClassType()) {
219 LatticeVal &IV = TrackedGlobals[GV];
220 if (!isa<UndefValue>(GV->getInitializer()))
221 IV.markConstant(GV->getInitializer());
225 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
226 /// and out of the specified function (which cannot have its address taken),
227 /// this method must be called.
228 void AddTrackedFunction(Function *F) {
229 // Add an entry, F -> undef.
230 if (const StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
231 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
232 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
235 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
238 /// Solve - Solve for constants and executable blocks.
242 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
243 /// that branches on undef values cannot reach any of their successors.
244 /// However, this is not a safe assumption. After we solve dataflow, this
245 /// method should be use to handle this. If this returns true, the solver
247 bool ResolvedUndefsIn(Function &F);
249 bool isBlockExecutable(BasicBlock *BB) const {
250 return BBExecutable.count(BB);
253 LatticeVal getLatticeValueFor(Value *V) const {
254 DenseMap<Value*, LatticeVal>::const_iterator I = ValueState.find(V);
255 assert(I != ValueState.end() && "V is not in valuemap!");
259 /// getTrackedRetVals - Get the inferred return value map.
261 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
262 return TrackedRetVals;
265 /// getTrackedGlobals - Get and return the set of inferred initializers for
266 /// global variables.
267 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
268 return TrackedGlobals;
271 void markOverdefined(Value *V) {
272 markOverdefined(ValueState[V], V);
276 // markConstant - Make a value be marked as "constant". If the value
277 // is not already a constant, add it to the instruction work list so that
278 // the users of the instruction are updated later.
280 void markConstant(LatticeVal &IV, Value *V, Constant *C) {
281 if (!IV.markConstant(C)) return;
282 DEBUG(errs() << "markConstant: " << *C << ": " << *V << '\n');
283 InstWorkList.push_back(V);
286 void markConstant(Value *V, Constant *C) {
287 markConstant(ValueState[V], V, C);
290 void markForcedConstant(Value *V, Constant *C) {
291 ValueState[V].markForcedConstant(C);
292 DEBUG(errs() << "markForcedConstant: " << *C << ": " << *V << '\n');
293 InstWorkList.push_back(V);
297 // markOverdefined - Make a value be marked as "overdefined". If the
298 // value is not already overdefined, add it to the overdefined instruction
299 // work list so that the users of the instruction are updated later.
300 void markOverdefined(LatticeVal &IV, Value *V) {
301 if (!IV.markOverdefined()) return;
303 DEBUG(errs() << "markOverdefined: ";
304 if (Function *F = dyn_cast<Function>(V))
305 errs() << "Function '" << F->getName() << "'\n";
307 errs() << *V << '\n');
308 // Only instructions go on the work list
309 OverdefinedInstWorkList.push_back(V);
312 void mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) {
313 if (IV.isOverdefined() || MergeWithV.isUndefined())
315 if (MergeWithV.isOverdefined())
316 markOverdefined(IV, V);
317 else if (IV.isUndefined())
318 markConstant(IV, V, MergeWithV.getConstant());
319 else if (IV.getConstant() != MergeWithV.getConstant())
320 markOverdefined(IV, V);
323 void mergeInValue(Value *V, LatticeVal MergeWithV) {
324 mergeInValue(ValueState[V], V, MergeWithV);
328 /// getValueState - Return the LatticeVal object that corresponds to the
329 /// value. This function handles the case when the value hasn't been seen yet
330 /// by properly seeding constants etc.
331 LatticeVal &getValueState(Value *V) {
332 DenseMap<Value*, LatticeVal>::iterator I = ValueState.find(V);
333 if (I != ValueState.end()) return I->second; // Common case, in the map
335 LatticeVal &LV = ValueState[V];
337 if (Constant *C = dyn_cast<Constant>(V)) {
338 // Undef values remain undefined.
339 if (!isa<UndefValue>(V))
340 LV.markConstant(C); // Constants are constant
343 // All others are underdefined by default.
347 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
348 /// work list if it is not already executable.
349 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
350 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
351 return; // This edge is already known to be executable!
353 if (!MarkBlockExecutable(Dest)) {
354 // If the destination is already executable, we just made an *edge*
355 // feasible that wasn't before. Revisit the PHI nodes in the block
356 // because they have potentially new operands.
357 DEBUG(errs() << "Marking Edge Executable: " << Source->getName()
358 << " -> " << Dest->getName() << "\n");
361 for (BasicBlock::iterator I = Dest->begin();
362 (PN = dyn_cast<PHINode>(I)); ++I)
367 // getFeasibleSuccessors - Return a vector of booleans to indicate which
368 // successors are reachable from a given terminator instruction.
370 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
372 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
373 // block to the 'To' basic block is currently feasible.
375 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
377 // OperandChangedState - This method is invoked on all of the users of an
378 // instruction that was just changed state somehow. Based on this
379 // information, we need to update the specified user of this instruction.
381 void OperandChangedState(Instruction *I) {
382 if (BBExecutable.count(I->getParent())) // Inst is executable?
386 /// RemoveFromOverdefinedPHIs - If I has any entries in the
387 /// UsersOfOverdefinedPHIs map for PN, remove them now.
388 void RemoveFromOverdefinedPHIs(Instruction *I, PHINode *PN) {
389 if (UsersOfOverdefinedPHIs.empty()) return;
390 std::multimap<PHINode*, Instruction*>::iterator It, E;
391 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN);
394 UsersOfOverdefinedPHIs.erase(It++);
401 friend class InstVisitor<SCCPSolver>;
403 // visit implementations - Something changed in this instruction. Either an
404 // operand made a transition, or the instruction is newly executable. Change
405 // the value type of I to reflect these changes if appropriate.
406 void visitPHINode(PHINode &I);
409 void visitReturnInst(ReturnInst &I);
410 void visitTerminatorInst(TerminatorInst &TI);
412 void visitCastInst(CastInst &I);
413 void visitSelectInst(SelectInst &I);
414 void visitBinaryOperator(Instruction &I);
415 void visitCmpInst(CmpInst &I);
416 void visitExtractElementInst(ExtractElementInst &I);
417 void visitInsertElementInst(InsertElementInst &I);
418 void visitShuffleVectorInst(ShuffleVectorInst &I);
419 void visitExtractValueInst(ExtractValueInst &EVI);
420 void visitInsertValueInst(InsertValueInst &IVI);
422 // Instructions that cannot be folded away.
423 void visitStoreInst (StoreInst &I);
424 void visitLoadInst (LoadInst &I);
425 void visitGetElementPtrInst(GetElementPtrInst &I);
426 void visitCallInst (CallInst &I) {
427 visitCallSite(CallSite::get(&I));
429 void visitInvokeInst (InvokeInst &II) {
430 visitCallSite(CallSite::get(&II));
431 visitTerminatorInst(II);
433 void visitCallSite (CallSite CS);
434 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ }
435 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
436 void visitAllocaInst (Instruction &I) { markOverdefined(&I); }
437 void visitVANextInst (Instruction &I) { markOverdefined(&I); }
438 void visitVAArgInst (Instruction &I) { markOverdefined(&I); }
440 void visitInstruction(Instruction &I) {
441 // If a new instruction is added to LLVM that we don't handle.
442 errs() << "SCCP: Don't know how to handle: " << I;
443 markOverdefined(&I); // Just in case
447 } // end anonymous namespace
450 // getFeasibleSuccessors - Return a vector of booleans to indicate which
451 // successors are reachable from a given terminator instruction.
453 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
454 SmallVector<bool, 16> &Succs) {
455 Succs.resize(TI.getNumSuccessors());
456 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
457 if (BI->isUnconditional()) {
462 LatticeVal BCValue = getValueState(BI->getCondition());
463 ConstantInt *CI = BCValue.getConstantInt();
465 // Overdefined condition variables, and branches on unfoldable constant
466 // conditions, mean the branch could go either way.
467 if (!BCValue.isUndefined())
468 Succs[0] = Succs[1] = true;
472 // Constant condition variables mean the branch can only go a single way.
473 Succs[CI->isZero()] = true;
477 if (isa<InvokeInst>(TI)) {
478 // Invoke instructions successors are always executable.
479 Succs[0] = Succs[1] = true;
483 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
484 LatticeVal SCValue = getValueState(SI->getCondition());
485 ConstantInt *CI = SCValue.getConstantInt();
487 if (CI == 0) { // Overdefined or undefined condition?
488 // All destinations are executable!
489 if (!SCValue.isUndefined())
490 Succs.assign(TI.getNumSuccessors(), true);
494 Succs[SI->findCaseValue(CI)] = true;
498 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
499 if (isa<IndirectBrInst>(&TI)) {
500 // Just mark all destinations executable!
501 Succs.assign(TI.getNumSuccessors(), true);
506 errs() << "Unknown terminator instruction: " << TI << '\n';
508 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
512 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
513 // block to the 'To' basic block is currently feasible.
515 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
516 assert(BBExecutable.count(To) && "Dest should always be alive!");
518 // Make sure the source basic block is executable!!
519 if (!BBExecutable.count(From)) return false;
521 // Check to make sure this edge itself is actually feasible now.
522 TerminatorInst *TI = From->getTerminator();
523 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
524 if (BI->isUnconditional())
527 LatticeVal BCValue = getValueState(BI->getCondition());
529 // Overdefined condition variables mean the branch could go either way,
530 // undef conditions mean that neither edge is feasible yet.
531 ConstantInt *CI = BCValue.getConstantInt();
533 return !BCValue.isUndefined();
535 // Constant condition variables mean the branch can only go a single way.
536 return BI->getSuccessor(CI->isZero()) == To;
539 // Invoke instructions successors are always executable.
540 if (isa<InvokeInst>(TI))
543 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
544 LatticeVal SCValue = getValueState(SI->getCondition());
545 ConstantInt *CI = SCValue.getConstantInt();
548 return !SCValue.isUndefined();
550 // Make sure to skip the "default value" which isn't a value
551 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i)
552 if (SI->getSuccessorValue(i) == CI) // Found the taken branch.
553 return SI->getSuccessor(i) == To;
555 // If the constant value is not equal to any of the branches, we must
556 // execute default branch.
557 return SI->getDefaultDest() == To;
560 // Just mark all destinations executable!
561 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
562 if (isa<IndirectBrInst>(&TI))
566 errs() << "Unknown terminator instruction: " << *TI << '\n';
571 // visit Implementations - Something changed in this instruction, either an
572 // operand made a transition, or the instruction is newly executable. Change
573 // the value type of I to reflect these changes if appropriate. This method
574 // makes sure to do the following actions:
576 // 1. If a phi node merges two constants in, and has conflicting value coming
577 // from different branches, or if the PHI node merges in an overdefined
578 // value, then the PHI node becomes overdefined.
579 // 2. If a phi node merges only constants in, and they all agree on value, the
580 // PHI node becomes a constant value equal to that.
581 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
582 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
583 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
584 // 6. If a conditional branch has a value that is constant, make the selected
585 // destination executable
586 // 7. If a conditional branch has a value that is overdefined, make all
587 // successors executable.
589 void SCCPSolver::visitPHINode(PHINode &PN) {
590 if (getValueState(&PN).isOverdefined()) {
591 // There may be instructions using this PHI node that are not overdefined
592 // themselves. If so, make sure that they know that the PHI node operand
594 std::multimap<PHINode*, Instruction*>::iterator I, E;
595 tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN);
599 SmallVector<Instruction*, 16> Users;
601 Users.push_back(I->second);
602 while (!Users.empty())
603 visit(Users.pop_back_val());
604 return; // Quick exit
607 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
608 // and slow us down a lot. Just mark them overdefined.
609 if (PN.getNumIncomingValues() > 64)
610 return markOverdefined(&PN);
612 // Look at all of the executable operands of the PHI node. If any of them
613 // are overdefined, the PHI becomes overdefined as well. If they are all
614 // constant, and they agree with each other, the PHI becomes the identical
615 // constant. If they are constant and don't agree, the PHI is overdefined.
616 // If there are no executable operands, the PHI remains undefined.
618 Constant *OperandVal = 0;
619 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
620 LatticeVal IV = getValueState(PN.getIncomingValue(i));
621 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
623 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
626 if (IV.isOverdefined()) // PHI node becomes overdefined!
627 return markOverdefined(&PN);
629 if (OperandVal == 0) { // Grab the first value.
630 OperandVal = IV.getConstant();
634 // There is already a reachable operand. If we conflict with it,
635 // then the PHI node becomes overdefined. If we agree with it, we
638 // Check to see if there are two different constants merging, if so, the PHI
639 // node is overdefined.
640 if (IV.getConstant() != OperandVal)
641 return markOverdefined(&PN);
644 // If we exited the loop, this means that the PHI node only has constant
645 // arguments that agree with each other(and OperandVal is the constant) or
646 // OperandVal is null because there are no defined incoming arguments. If
647 // this is the case, the PHI remains undefined.
650 markConstant(&PN, OperandVal); // Acquire operand value
656 void SCCPSolver::visitReturnInst(ReturnInst &I) {
657 if (I.getNumOperands() == 0) return; // ret void
659 Function *F = I.getParent()->getParent();
661 // If we are tracking the return value of this function, merge it in.
662 if (!TrackedRetVals.empty()) {
663 DenseMap<Function*, LatticeVal>::iterator TFRVI =
664 TrackedRetVals.find(F);
665 if (TFRVI != TrackedRetVals.end()) {
666 mergeInValue(TFRVI->second, F, getValueState(I.getOperand(0)));
671 // Handle functions that return multiple values.
672 if (!TrackedMultipleRetVals.empty() &&
673 isa<StructType>(I.getOperand(0)->getType())) {
674 for (unsigned i = 0, e = I.getOperand(0)->getType()->getNumContainedTypes();
676 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
677 It = TrackedMultipleRetVals.find(std::make_pair(F, i));
678 if (It == TrackedMultipleRetVals.end()) break;
679 if (Value *Val = FindInsertedValue(I.getOperand(0), i, I.getContext()))
680 mergeInValue(It->second, F, getValueState(Val));
685 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
686 SmallVector<bool, 16> SuccFeasible;
687 getFeasibleSuccessors(TI, SuccFeasible);
689 BasicBlock *BB = TI.getParent();
691 // Mark all feasible successors executable.
692 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
694 markEdgeExecutable(BB, TI.getSuccessor(i));
697 void SCCPSolver::visitCastInst(CastInst &I) {
698 LatticeVal OpSt = getValueState(I.getOperand(0));
699 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand
701 else if (OpSt.isConstant()) // Propagate constant value
702 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
703 OpSt.getConstant(), I.getType()));
706 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
707 Value *Aggr = EVI.getAggregateOperand();
709 // If the operand to the extractvalue is an undef, the result is undef.
710 if (isa<UndefValue>(Aggr))
713 // Currently only handle single-index extractvalues.
714 if (EVI.getNumIndices() != 1)
715 return markOverdefined(&EVI);
718 if (CallInst *CI = dyn_cast<CallInst>(Aggr))
719 F = CI->getCalledFunction();
720 else if (InvokeInst *II = dyn_cast<InvokeInst>(Aggr))
721 F = II->getCalledFunction();
723 // TODO: If IPSCCP resolves the callee of this function, we could propagate a
725 if (F == 0 || TrackedMultipleRetVals.empty())
726 return markOverdefined(&EVI);
728 // See if we are tracking the result of the callee. If not tracking this
729 // function (for example, it is a declaration) just move to overdefined.
730 if (!TrackedMultipleRetVals.count(std::make_pair(F, *EVI.idx_begin())))
731 return markOverdefined(&EVI);
733 // Otherwise, the value will be merged in here as a result of CallSite
737 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
738 Value *Aggr = IVI.getAggregateOperand();
739 Value *Val = IVI.getInsertedValueOperand();
741 // If the operands to the insertvalue are undef, the result is undef.
742 if (isa<UndefValue>(Aggr) && isa<UndefValue>(Val))
745 // Currently only handle single-index insertvalues.
746 if (IVI.getNumIndices() != 1)
747 return markOverdefined(&IVI);
749 // Currently only handle insertvalue instructions that are in a single-use
750 // chain that builds up a return value.
751 for (const InsertValueInst *TmpIVI = &IVI; ; ) {
752 if (!TmpIVI->hasOneUse())
753 return markOverdefined(&IVI);
755 const Value *V = *TmpIVI->use_begin();
756 if (isa<ReturnInst>(V))
758 TmpIVI = dyn_cast<InsertValueInst>(V);
760 return markOverdefined(&IVI);
763 // See if we are tracking the result of the callee.
764 Function *F = IVI.getParent()->getParent();
765 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
766 It = TrackedMultipleRetVals.find(std::make_pair(F, *IVI.idx_begin()));
768 // Merge in the inserted member value.
769 if (It != TrackedMultipleRetVals.end())
770 mergeInValue(It->second, F, getValueState(Val));
772 // Mark the aggregate result of the IVI overdefined; any tracking that we do
773 // will be done on the individual member values.
774 markOverdefined(&IVI);
777 void SCCPSolver::visitSelectInst(SelectInst &I) {
778 LatticeVal CondValue = getValueState(I.getCondition());
779 if (CondValue.isUndefined())
782 if (ConstantInt *CondCB = CondValue.getConstantInt()) {
783 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
784 mergeInValue(&I, getValueState(OpVal));
788 // Otherwise, the condition is overdefined or a constant we can't evaluate.
789 // See if we can produce something better than overdefined based on the T/F
791 LatticeVal TVal = getValueState(I.getTrueValue());
792 LatticeVal FVal = getValueState(I.getFalseValue());
794 // select ?, C, C -> C.
795 if (TVal.isConstant() && FVal.isConstant() &&
796 TVal.getConstant() == FVal.getConstant())
797 return markConstant(&I, FVal.getConstant());
799 if (TVal.isUndefined()) // select ?, undef, X -> X.
800 return mergeInValue(&I, FVal);
801 if (FVal.isUndefined()) // select ?, X, undef -> X.
802 return mergeInValue(&I, TVal);
806 // Handle Binary Operators.
807 void SCCPSolver::visitBinaryOperator(Instruction &I) {
808 LatticeVal V1State = getValueState(I.getOperand(0));
809 LatticeVal V2State = getValueState(I.getOperand(1));
811 LatticeVal &IV = ValueState[&I];
812 if (IV.isOverdefined()) return;
814 if (V1State.isConstant() && V2State.isConstant())
815 return markConstant(IV, &I,
816 ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
817 V2State.getConstant()));
819 // If something is undef, wait for it to resolve.
820 if (!V1State.isOverdefined() && !V2State.isOverdefined())
823 // Otherwise, one of our operands is overdefined. Try to produce something
824 // better than overdefined with some tricks.
826 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
827 // operand is overdefined.
828 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
829 LatticeVal *NonOverdefVal = 0;
830 if (!V1State.isOverdefined())
831 NonOverdefVal = &V1State;
832 else if (!V2State.isOverdefined())
833 NonOverdefVal = &V2State;
836 if (NonOverdefVal->isUndefined()) {
837 // Could annihilate value.
838 if (I.getOpcode() == Instruction::And)
839 markConstant(IV, &I, Constant::getNullValue(I.getType()));
840 else if (const VectorType *PT = dyn_cast<VectorType>(I.getType()))
841 markConstant(IV, &I, Constant::getAllOnesValue(PT));
844 Constant::getAllOnesValue(I.getType()));
848 if (I.getOpcode() == Instruction::And) {
850 if (NonOverdefVal->getConstant()->isNullValue())
851 return markConstant(IV, &I, NonOverdefVal->getConstant());
853 if (ConstantInt *CI = NonOverdefVal->getConstantInt())
854 if (CI->isAllOnesValue()) // X or -1 = -1
855 return markConstant(IV, &I, NonOverdefVal->getConstant());
861 // If both operands are PHI nodes, it is possible that this instruction has
862 // a constant value, despite the fact that the PHI node doesn't. Check for
863 // this condition now.
864 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
865 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
866 if (PN1->getParent() == PN2->getParent()) {
867 // Since the two PHI nodes are in the same basic block, they must have
868 // entries for the same predecessors. Walk the predecessor list, and
869 // if all of the incoming values are constants, and the result of
870 // evaluating this expression with all incoming value pairs is the
871 // same, then this expression is a constant even though the PHI node
872 // is not a constant!
874 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
875 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
876 BasicBlock *InBlock = PN1->getIncomingBlock(i);
877 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
879 if (In1.isOverdefined() || In2.isOverdefined()) {
880 Result.markOverdefined();
881 break; // Cannot fold this operation over the PHI nodes!
884 if (In1.isConstant() && In2.isConstant()) {
885 Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(),
887 if (Result.isUndefined())
888 Result.markConstant(V);
889 else if (Result.isConstant() && Result.getConstant() != V) {
890 Result.markOverdefined();
896 // If we found a constant value here, then we know the instruction is
897 // constant despite the fact that the PHI nodes are overdefined.
898 if (Result.isConstant()) {
899 markConstant(IV, &I, Result.getConstant());
900 // Remember that this instruction is virtually using the PHI node
902 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
903 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
907 if (Result.isUndefined())
910 // Okay, this really is overdefined now. Since we might have
911 // speculatively thought that this was not overdefined before, and
912 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
913 // make sure to clean out any entries that we put there, for
915 RemoveFromOverdefinedPHIs(&I, PN1);
916 RemoveFromOverdefinedPHIs(&I, PN2);
922 // Handle ICmpInst instruction.
923 void SCCPSolver::visitCmpInst(CmpInst &I) {
924 LatticeVal V1State = getValueState(I.getOperand(0));
925 LatticeVal V2State = getValueState(I.getOperand(1));
927 LatticeVal &IV = ValueState[&I];
928 if (IV.isOverdefined()) return;
930 if (V1State.isConstant() && V2State.isConstant())
931 return markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
932 V1State.getConstant(),
933 V2State.getConstant()));
935 // If operands are still undefined, wait for it to resolve.
936 if (!V1State.isOverdefined() && !V2State.isOverdefined())
939 // If something is overdefined, use some tricks to avoid ending up and over
940 // defined if we can.
942 // If both operands are PHI nodes, it is possible that this instruction has
943 // a constant value, despite the fact that the PHI node doesn't. Check for
944 // this condition now.
945 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
946 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
947 if (PN1->getParent() == PN2->getParent()) {
948 // Since the two PHI nodes are in the same basic block, they must have
949 // entries for the same predecessors. Walk the predecessor list, and
950 // if all of the incoming values are constants, and the result of
951 // evaluating this expression with all incoming value pairs is the
952 // same, then this expression is a constant even though the PHI node
953 // is not a constant!
955 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
956 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
957 BasicBlock *InBlock = PN1->getIncomingBlock(i);
958 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
960 if (In1.isOverdefined() || In2.isOverdefined()) {
961 Result.markOverdefined();
962 break; // Cannot fold this operation over the PHI nodes!
965 if (In1.isConstant() && In2.isConstant()) {
966 Constant *V = ConstantExpr::getCompare(I.getPredicate(),
969 if (Result.isUndefined())
970 Result.markConstant(V);
971 else if (Result.isConstant() && Result.getConstant() != V) {
972 Result.markOverdefined();
978 // If we found a constant value here, then we know the instruction is
979 // constant despite the fact that the PHI nodes are overdefined.
980 if (Result.isConstant()) {
981 markConstant(&I, Result.getConstant());
982 // Remember that this instruction is virtually using the PHI node
984 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
985 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
989 if (Result.isUndefined())
992 // Okay, this really is overdefined now. Since we might have
993 // speculatively thought that this was not overdefined before, and
994 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
995 // make sure to clean out any entries that we put there, for
997 RemoveFromOverdefinedPHIs(&I, PN1);
998 RemoveFromOverdefinedPHIs(&I, PN2);
1001 markOverdefined(&I);
1004 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
1005 // FIXME : SCCP does not handle vectors properly.
1006 return markOverdefined(&I);
1009 LatticeVal &ValState = getValueState(I.getOperand(0));
1010 LatticeVal &IdxState = getValueState(I.getOperand(1));
1012 if (ValState.isOverdefined() || IdxState.isOverdefined())
1013 markOverdefined(&I);
1014 else if(ValState.isConstant() && IdxState.isConstant())
1015 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
1016 IdxState.getConstant()));
1020 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
1021 // FIXME : SCCP does not handle vectors properly.
1022 return markOverdefined(&I);
1024 LatticeVal &ValState = getValueState(I.getOperand(0));
1025 LatticeVal &EltState = getValueState(I.getOperand(1));
1026 LatticeVal &IdxState = getValueState(I.getOperand(2));
1028 if (ValState.isOverdefined() || EltState.isOverdefined() ||
1029 IdxState.isOverdefined())
1030 markOverdefined(&I);
1031 else if(ValState.isConstant() && EltState.isConstant() &&
1032 IdxState.isConstant())
1033 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
1034 EltState.getConstant(),
1035 IdxState.getConstant()));
1036 else if (ValState.isUndefined() && EltState.isConstant() &&
1037 IdxState.isConstant())
1038 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
1039 EltState.getConstant(),
1040 IdxState.getConstant()));
1044 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
1045 // FIXME : SCCP does not handle vectors properly.
1046 return markOverdefined(&I);
1048 LatticeVal &V1State = getValueState(I.getOperand(0));
1049 LatticeVal &V2State = getValueState(I.getOperand(1));
1050 LatticeVal &MaskState = getValueState(I.getOperand(2));
1052 if (MaskState.isUndefined() ||
1053 (V1State.isUndefined() && V2State.isUndefined()))
1054 return; // Undefined output if mask or both inputs undefined.
1056 if (V1State.isOverdefined() || V2State.isOverdefined() ||
1057 MaskState.isOverdefined()) {
1058 markOverdefined(&I);
1060 // A mix of constant/undef inputs.
1061 Constant *V1 = V1State.isConstant() ?
1062 V1State.getConstant() : UndefValue::get(I.getType());
1063 Constant *V2 = V2State.isConstant() ?
1064 V2State.getConstant() : UndefValue::get(I.getType());
1065 Constant *Mask = MaskState.isConstant() ?
1066 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
1067 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
1072 // Handle getelementptr instructions. If all operands are constants then we
1073 // can turn this into a getelementptr ConstantExpr.
1075 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1076 if (ValueState[&I].isOverdefined()) return;
1078 SmallVector<Constant*, 8> Operands;
1079 Operands.reserve(I.getNumOperands());
1081 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1082 LatticeVal State = getValueState(I.getOperand(i));
1083 if (State.isUndefined())
1084 return; // Operands are not resolved yet.
1086 if (State.isOverdefined())
1087 return markOverdefined(&I);
1089 assert(State.isConstant() && "Unknown state!");
1090 Operands.push_back(State.getConstant());
1093 Constant *Ptr = Operands[0];
1094 markConstant(&I, ConstantExpr::getGetElementPtr(Ptr, &Operands[0]+1,
1095 Operands.size()-1));
1098 void SCCPSolver::visitStoreInst(StoreInst &SI) {
1099 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1102 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1103 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1104 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1106 // Get the value we are storing into the global, then merge it.
1107 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)));
1108 if (I->second.isOverdefined())
1109 TrackedGlobals.erase(I); // No need to keep tracking this!
1113 // Handle load instructions. If the operand is a constant pointer to a constant
1114 // global, we can replace the load with the loaded constant value!
1115 void SCCPSolver::visitLoadInst(LoadInst &I) {
1116 LatticeVal PtrVal = getValueState(I.getOperand(0));
1117 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1119 LatticeVal &IV = ValueState[&I];
1120 if (IV.isOverdefined()) return;
1122 if (!PtrVal.isConstant() || I.isVolatile())
1123 return markOverdefined(IV, &I);
1125 Constant *Ptr = PtrVal.getConstant();
1127 // load null -> null
1128 if (isa<ConstantPointerNull>(Ptr) && I.getPointerAddressSpace() == 0)
1129 return markConstant(IV, &I, Constant::getNullValue(I.getType()));
1131 // Transform load (constant global) into the value loaded.
1132 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1133 if (!TrackedGlobals.empty()) {
1134 // If we are tracking this global, merge in the known value for it.
1135 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1136 TrackedGlobals.find(GV);
1137 if (It != TrackedGlobals.end()) {
1138 mergeInValue(IV, &I, It->second);
1144 // Transform load from a constant into a constant if possible.
1145 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, TD))
1146 return markConstant(IV, &I, C);
1148 // Otherwise we cannot say for certain what value this load will produce.
1150 markOverdefined(IV, &I);
1153 void SCCPSolver::visitCallSite(CallSite CS) {
1154 Function *F = CS.getCalledFunction();
1155 Instruction *I = CS.getInstruction();
1157 // The common case is that we aren't tracking the callee, either because we
1158 // are not doing interprocedural analysis or the callee is indirect, or is
1159 // external. Handle these cases first.
1160 if (F == 0 || F->isDeclaration()) {
1162 // Void return and not tracking callee, just bail.
1163 if (I->getType()->isVoidTy()) return;
1165 // Otherwise, if we have a single return value case, and if the function is
1166 // a declaration, maybe we can constant fold it.
1167 if (F && F->isDeclaration() && !isa<StructType>(I->getType()) &&
1168 canConstantFoldCallTo(F)) {
1170 SmallVector<Constant*, 8> Operands;
1171 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1173 LatticeVal State = getValueState(*AI);
1175 if (State.isUndefined())
1176 return; // Operands are not resolved yet.
1177 if (State.isOverdefined())
1178 return markOverdefined(I);
1179 assert(State.isConstant() && "Unknown state!");
1180 Operands.push_back(State.getConstant());
1183 // If we can constant fold this, mark the result of the call as a
1185 if (Constant *C = ConstantFoldCall(F, Operands.data(), Operands.size()))
1186 return markConstant(I, C);
1189 // Otherwise, we don't know anything about this call, mark it overdefined.
1190 return markOverdefined(I);
1193 // If this is a single/zero retval case, see if we're tracking the function.
1194 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1195 if (TFRVI != TrackedRetVals.end()) {
1196 // If so, propagate the return value of the callee into this call result.
1197 mergeInValue(I, TFRVI->second);
1198 } else if (isa<StructType>(I->getType())) {
1199 // Check to see if we're tracking this callee, if not, handle it in the
1200 // common path above.
1201 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
1202 TMRVI = TrackedMultipleRetVals.find(std::make_pair(F, 0));
1203 if (TMRVI == TrackedMultipleRetVals.end())
1204 goto CallOverdefined;
1206 // Need to mark as overdefined, otherwise it stays undefined which
1207 // creates extractvalue undef, <idx>
1210 // If we are tracking this callee, propagate the return values of the call
1211 // into this call site. We do this by walking all the uses. Single-index
1212 // ExtractValueInst uses can be tracked; anything more complicated is
1213 // currently handled conservatively.
1214 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1216 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(*UI)) {
1217 if (EVI->getNumIndices() == 1) {
1219 TrackedMultipleRetVals[std::make_pair(F, *EVI->idx_begin())]);
1223 // The aggregate value is used in a way not handled here. Assume nothing.
1224 markOverdefined(*UI);
1227 // Otherwise we're not tracking this callee, so handle it in the
1228 // common path above.
1229 goto CallOverdefined;
1232 // Finally, if this is the first call to the function hit, mark its entry
1233 // block executable.
1234 MarkBlockExecutable(F->begin());
1236 // Propagate information from this call site into the callee.
1237 CallSite::arg_iterator CAI = CS.arg_begin();
1238 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1239 AI != E; ++AI, ++CAI) {
1240 // If this argument is byval, and if the function is not readonly, there
1241 // will be an implicit copy formed of the input aggregate.
1242 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1243 markOverdefined(AI);
1247 mergeInValue(AI, getValueState(*CAI));
1251 void SCCPSolver::Solve() {
1252 // Process the work lists until they are empty!
1253 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1254 !OverdefinedInstWorkList.empty()) {
1255 // Process the overdefined instruction's work list first, which drives other
1256 // things to overdefined more quickly.
1257 while (!OverdefinedInstWorkList.empty()) {
1258 Value *I = OverdefinedInstWorkList.pop_back_val();
1260 DEBUG(errs() << "\nPopped off OI-WL: " << *I << '\n');
1262 // "I" got into the work list because it either made the transition from
1263 // bottom to constant
1265 // Anything on this worklist that is overdefined need not be visited
1266 // since all of its users will have already been marked as overdefined
1267 // Update all of the users of this instruction's value.
1269 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1271 if (Instruction *I = dyn_cast<Instruction>(*UI))
1272 OperandChangedState(I);
1275 // Process the instruction work list.
1276 while (!InstWorkList.empty()) {
1277 Value *I = InstWorkList.pop_back_val();
1279 DEBUG(errs() << "\nPopped off I-WL: " << *I << '\n');
1281 // "I" got into the work list because it made the transition from undef to
1284 // Anything on this worklist that is overdefined need not be visited
1285 // since all of its users will have already been marked as overdefined.
1286 // Update all of the users of this instruction's value.
1288 if (!getValueState(I).isOverdefined())
1289 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1291 if (Instruction *I = dyn_cast<Instruction>(*UI))
1292 OperandChangedState(I);
1295 // Process the basic block work list.
1296 while (!BBWorkList.empty()) {
1297 BasicBlock *BB = BBWorkList.back();
1298 BBWorkList.pop_back();
1300 DEBUG(errs() << "\nPopped off BBWL: " << *BB << '\n');
1302 // Notify all instructions in this basic block that they are newly
1309 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1310 /// that branches on undef values cannot reach any of their successors.
1311 /// However, this is not a safe assumption. After we solve dataflow, this
1312 /// method should be use to handle this. If this returns true, the solver
1313 /// should be rerun.
1315 /// This method handles this by finding an unresolved branch and marking it one
1316 /// of the edges from the block as being feasible, even though the condition
1317 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1318 /// CFG and only slightly pessimizes the analysis results (by marking one,
1319 /// potentially infeasible, edge feasible). This cannot usefully modify the
1320 /// constraints on the condition of the branch, as that would impact other users
1323 /// This scan also checks for values that use undefs, whose results are actually
1324 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1325 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1326 /// even if X isn't defined.
1327 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1328 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1329 if (!BBExecutable.count(BB))
1332 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1333 // Look for instructions which produce undef values.
1334 if (I->getType()->isVoidTy()) continue;
1336 LatticeVal &LV = getValueState(I);
1337 if (!LV.isUndefined()) continue;
1339 // Get the lattice values of the first two operands for use below.
1340 LatticeVal Op0LV = getValueState(I->getOperand(0));
1342 if (I->getNumOperands() == 2) {
1343 // If this is a two-operand instruction, and if both operands are
1344 // undefs, the result stays undef.
1345 Op1LV = getValueState(I->getOperand(1));
1346 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1350 // If this is an instructions whose result is defined even if the input is
1351 // not fully defined, propagate the information.
1352 const Type *ITy = I->getType();
1353 switch (I->getOpcode()) {
1354 default: break; // Leave the instruction as an undef.
1355 case Instruction::ZExt:
1356 // After a zero extend, we know the top part is zero. SExt doesn't have
1357 // to be handled here, because we don't know whether the top part is 1's
1359 markForcedConstant(I, Constant::getNullValue(ITy));
1361 case Instruction::Mul:
1362 case Instruction::And:
1363 // undef * X -> 0. X could be zero.
1364 // undef & X -> 0. X could be zero.
1365 markForcedConstant(I, Constant::getNullValue(ITy));
1368 case Instruction::Or:
1369 // undef | X -> -1. X could be -1.
1370 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1373 case Instruction::SDiv:
1374 case Instruction::UDiv:
1375 case Instruction::SRem:
1376 case Instruction::URem:
1377 // X / undef -> undef. No change.
1378 // X % undef -> undef. No change.
1379 if (Op1LV.isUndefined()) break;
1381 // undef / X -> 0. X could be maxint.
1382 // undef % X -> 0. X could be 1.
1383 markForcedConstant(I, Constant::getNullValue(ITy));
1386 case Instruction::AShr:
1387 // undef >>s X -> undef. No change.
1388 if (Op0LV.isUndefined()) break;
1390 // X >>s undef -> X. X could be 0, X could have the high-bit known set.
1391 if (Op0LV.isConstant())
1392 markForcedConstant(I, Op0LV.getConstant());
1396 case Instruction::LShr:
1397 case Instruction::Shl:
1398 // undef >> X -> undef. No change.
1399 // undef << X -> undef. No change.
1400 if (Op0LV.isUndefined()) break;
1402 // X >> undef -> 0. X could be 0.
1403 // X << undef -> 0. X could be 0.
1404 markForcedConstant(I, Constant::getNullValue(ITy));
1406 case Instruction::Select:
1407 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1408 if (Op0LV.isUndefined()) {
1409 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1410 Op1LV = getValueState(I->getOperand(2));
1411 } else if (Op1LV.isUndefined()) {
1412 // c ? undef : undef -> undef. No change.
1413 Op1LV = getValueState(I->getOperand(2));
1414 if (Op1LV.isUndefined())
1416 // Otherwise, c ? undef : x -> x.
1418 // Leave Op1LV as Operand(1)'s LatticeValue.
1421 if (Op1LV.isConstant())
1422 markForcedConstant(I, Op1LV.getConstant());
1426 case Instruction::Call:
1427 // If a call has an undef result, it is because it is constant foldable
1428 // but one of the inputs was undef. Just force the result to
1435 TerminatorInst *TI = BB->getTerminator();
1436 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1437 if (!BI->isConditional()) continue;
1438 if (!getValueState(BI->getCondition()).isUndefined())
1440 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1441 if (SI->getNumSuccessors() < 2) // no cases
1443 if (!getValueState(SI->getCondition()).isUndefined())
1449 // If the edge to the second successor isn't thought to be feasible yet,
1450 // mark it so now. We pick the second one so that this goes to some
1451 // enumerated value in a switch instead of going to the default destination.
1452 if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(1))))
1455 // Otherwise, it isn't already thought to be feasible. Mark it as such now
1456 // and return. This will make other blocks reachable, which will allow new
1457 // values to be discovered and existing ones to be moved in the lattice.
1458 markEdgeExecutable(BB, TI->getSuccessor(1));
1460 // This must be a conditional branch of switch on undef. At this point,
1461 // force the old terminator to branch to the first successor. This is
1462 // required because we are now influencing the dataflow of the function with
1463 // the assumption that this edge is taken. If we leave the branch condition
1464 // as undef, then further analysis could think the undef went another way
1465 // leading to an inconsistent set of conclusions.
1466 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1467 BI->setCondition(ConstantInt::getFalse(BI->getContext()));
1469 SwitchInst *SI = cast<SwitchInst>(TI);
1470 SI->setCondition(SI->getCaseValue(1));
1481 //===--------------------------------------------------------------------===//
1483 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1484 /// Sparse Conditional Constant Propagator.
1486 struct SCCP : public FunctionPass {
1487 static char ID; // Pass identification, replacement for typeid
1488 SCCP() : FunctionPass(&ID) {}
1490 // runOnFunction - Run the Sparse Conditional Constant Propagation
1491 // algorithm, and return true if the function was modified.
1493 bool runOnFunction(Function &F);
1495 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1496 AU.setPreservesCFG();
1499 } // end anonymous namespace
1502 static RegisterPass<SCCP>
1503 X("sccp", "Sparse Conditional Constant Propagation");
1505 // createSCCPPass - This is the public interface to this file.
1506 FunctionPass *llvm::createSCCPPass() {
1510 static void DeleteInstructionInBlock(BasicBlock *BB) {
1511 DEBUG(errs() << " BasicBlock Dead:" << *BB);
1514 // Delete the instructions backwards, as it has a reduced likelihood of
1515 // having to update as many def-use and use-def chains.
1516 while (!isa<TerminatorInst>(BB->begin())) {
1517 Instruction *I = --BasicBlock::iterator(BB->getTerminator());
1519 if (!I->use_empty())
1520 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1521 BB->getInstList().erase(I);
1526 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1527 // and return true if the function was modified.
1529 bool SCCP::runOnFunction(Function &F) {
1530 DEBUG(errs() << "SCCP on function '" << F.getName() << "'\n");
1531 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1533 // Mark the first block of the function as being executable.
1534 Solver.MarkBlockExecutable(F.begin());
1536 // Mark all arguments to the function as being overdefined.
1537 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1538 Solver.markOverdefined(AI);
1540 // Solve for constants.
1541 bool ResolvedUndefs = true;
1542 while (ResolvedUndefs) {
1544 DEBUG(errs() << "RESOLVING UNDEFs\n");
1545 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1548 bool MadeChanges = false;
1550 // If we decided that there are basic blocks that are dead in this function,
1551 // delete their contents now. Note that we cannot actually delete the blocks,
1552 // as we cannot modify the CFG of the function.
1554 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1555 if (!Solver.isBlockExecutable(BB)) {
1556 DeleteInstructionInBlock(BB);
1561 // Iterate over all of the instructions in a function, replacing them with
1562 // constants if we have found them to be of constant values.
1564 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1565 Instruction *Inst = BI++;
1566 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst))
1569 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1570 if (IV.isOverdefined())
1573 Constant *Const = IV.isConstant()
1574 ? IV.getConstant() : UndefValue::get(Inst->getType());
1575 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1577 // Replaces all of the uses of a variable with uses of the constant.
1578 Inst->replaceAllUsesWith(Const);
1580 // Delete the instruction.
1581 Inst->eraseFromParent();
1583 // Hey, we just changed something!
1593 //===--------------------------------------------------------------------===//
1595 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1596 /// Constant Propagation.
1598 struct IPSCCP : public ModulePass {
1600 IPSCCP() : ModulePass(&ID) {}
1601 bool runOnModule(Module &M);
1603 } // end anonymous namespace
1605 char IPSCCP::ID = 0;
1606 static RegisterPass<IPSCCP>
1607 Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation");
1609 // createIPSCCPPass - This is the public interface to this file.
1610 ModulePass *llvm::createIPSCCPPass() {
1611 return new IPSCCP();
1615 static bool AddressIsTaken(GlobalValue *GV) {
1616 // Delete any dead constantexpr klingons.
1617 GV->removeDeadConstantUsers();
1619 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
1621 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1622 if (SI->getOperand(0) == GV || SI->isVolatile())
1623 return true; // Storing addr of GV.
1624 } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) {
1625 // Make sure we are calling the function, not passing the address.
1626 if (UI.getOperandNo() != 0)
1628 } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1629 if (LI->isVolatile())
1631 } else if (isa<BlockAddress>(*UI)) {
1632 // blockaddress doesn't take the address of the function, it takes addr
1640 bool IPSCCP::runOnModule(Module &M) {
1641 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1643 // Loop over all functions, marking arguments to those with their addresses
1644 // taken or that are external as overdefined.
1646 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1647 if (F->isDeclaration())
1650 // If this is a strong or ODR definition of this function, then we can
1651 // propagate information about its result into callsites of it.
1652 if (!F->mayBeOverridden() &&
1653 !isa<StructType>(F->getReturnType()))
1654 Solver.AddTrackedFunction(F);
1656 // If this function only has direct calls that we can see, we can track its
1657 // arguments and return value aggressively, and can assume it is not called
1658 // unless we see evidence to the contrary.
1659 if (F->hasLocalLinkage() && !AddressIsTaken(F))
1662 // Assume the function is called.
1663 Solver.MarkBlockExecutable(F->begin());
1665 // Assume nothing about the incoming arguments.
1666 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1668 Solver.markOverdefined(AI);
1671 // Loop over global variables. We inform the solver about any internal global
1672 // variables that do not have their 'addresses taken'. If they don't have
1673 // their addresses taken, we can propagate constants through them.
1674 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1676 if (!G->isConstant() && G->hasLocalLinkage() && !AddressIsTaken(G))
1677 Solver.TrackValueOfGlobalVariable(G);
1679 // Solve for constants.
1680 bool ResolvedUndefs = true;
1681 while (ResolvedUndefs) {
1684 DEBUG(errs() << "RESOLVING UNDEFS\n");
1685 ResolvedUndefs = false;
1686 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1687 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1690 bool MadeChanges = false;
1692 // Iterate over all of the instructions in the module, replacing them with
1693 // constants if we have found them to be of constant values.
1695 SmallVector<BasicBlock*, 512> BlocksToErase;
1697 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1698 if (Solver.isBlockExecutable(F->begin())) {
1699 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1701 if (AI->use_empty()) continue;
1703 LatticeVal IV = Solver.getLatticeValueFor(AI);
1704 if (IV.isOverdefined()) continue;
1706 Constant *CST = IV.isConstant() ?
1707 IV.getConstant() : UndefValue::get(AI->getType());
1708 DEBUG(errs() << "*** Arg " << *AI << " = " << *CST <<"\n");
1710 // Replaces all of the uses of a variable with uses of the
1712 AI->replaceAllUsesWith(CST);
1717 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1718 if (!Solver.isBlockExecutable(BB)) {
1719 DeleteInstructionInBlock(BB);
1722 TerminatorInst *TI = BB->getTerminator();
1723 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1724 BasicBlock *Succ = TI->getSuccessor(i);
1725 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1726 TI->getSuccessor(i)->removePredecessor(BB);
1728 if (!TI->use_empty())
1729 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1730 TI->eraseFromParent();
1732 if (&*BB != &F->front())
1733 BlocksToErase.push_back(BB);
1735 new UnreachableInst(M.getContext(), BB);
1739 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1740 Instruction *Inst = BI++;
1741 if (Inst->getType()->isVoidTy())
1744 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1745 if (IV.isOverdefined())
1748 Constant *Const = IV.isConstant()
1749 ? IV.getConstant() : UndefValue::get(Inst->getType());
1750 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1752 // Replaces all of the uses of a variable with uses of the
1754 Inst->replaceAllUsesWith(Const);
1756 // Delete the instruction.
1757 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst))
1758 Inst->eraseFromParent();
1760 // Hey, we just changed something!
1766 // Now that all instructions in the function are constant folded, erase dead
1767 // blocks, because we can now use ConstantFoldTerminator to get rid of
1769 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1770 // If there are any PHI nodes in this successor, drop entries for BB now.
1771 BasicBlock *DeadBB = BlocksToErase[i];
1772 while (!DeadBB->use_empty()) {
1773 Instruction *I = cast<Instruction>(DeadBB->use_back());
1774 bool Folded = ConstantFoldTerminator(I->getParent());
1776 // The constant folder may not have been able to fold the terminator
1777 // if this is a branch or switch on undef. Fold it manually as a
1778 // branch to the first successor.
1780 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1781 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1782 "Branch should be foldable!");
1783 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1784 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1786 llvm_unreachable("Didn't fold away reference to block!");
1790 // Make this an uncond branch to the first successor.
1791 TerminatorInst *TI = I->getParent()->getTerminator();
1792 BranchInst::Create(TI->getSuccessor(0), TI);
1794 // Remove entries in successor phi nodes to remove edges.
1795 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1796 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1798 // Remove the old terminator.
1799 TI->eraseFromParent();
1803 // Finally, delete the basic block.
1804 F->getBasicBlockList().erase(DeadBB);
1806 BlocksToErase.clear();
1809 // If we inferred constant or undef return values for a function, we replaced
1810 // all call uses with the inferred value. This means we don't need to bother
1811 // actually returning anything from the function. Replace all return
1812 // instructions with return undef.
1813 // TODO: Process multiple value ret instructions also.
1814 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1815 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1816 E = RV.end(); I != E; ++I) {
1817 Function *F = I->first;
1818 if (I->second.isOverdefined() || F->getReturnType()->isVoidTy())
1821 // We can only do this if we know that nothing else can call the function.
1822 if (!F->hasLocalLinkage() || AddressIsTaken(F))
1825 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1826 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1827 if (!isa<UndefValue>(RI->getOperand(0)))
1828 RI->setOperand(0, UndefValue::get(F->getReturnType()));
1831 // If we infered constant or undef values for globals variables, we can delete
1832 // the global and any stores that remain to it.
1833 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1834 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1835 E = TG.end(); I != E; ++I) {
1836 GlobalVariable *GV = I->first;
1837 assert(!I->second.isOverdefined() &&
1838 "Overdefined values should have been taken out of the map!");
1839 DEBUG(errs() << "Found that GV '" << GV->getName() << "' is constant!\n");
1840 while (!GV->use_empty()) {
1841 StoreInst *SI = cast<StoreInst>(GV->use_back());
1842 SI->eraseFromParent();
1844 M.getGlobalList().erase(GV);