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 for whose
178 /// arguments we make optimistic assumptions about and try to prove as
180 SmallPtrSet<Function*, 16> TrackingIncomingArguments;
182 /// The reason for two worklists is that overdefined is the lowest state
183 /// on the lattice, and moving things to overdefined as fast as possible
184 /// makes SCCP converge much faster.
186 /// By having a separate worklist, we accomplish this because everything
187 /// possibly overdefined will become overdefined at the soonest possible
189 SmallVector<Value*, 64> OverdefinedInstWorkList;
190 SmallVector<Value*, 64> InstWorkList;
193 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list
195 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
196 /// overdefined, despite the fact that the PHI node is overdefined.
197 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
199 /// KnownFeasibleEdges - Entries in this set are edges which have already had
200 /// PHI nodes retriggered.
201 typedef std::pair<BasicBlock*, BasicBlock*> Edge;
202 DenseSet<Edge> KnownFeasibleEdges;
204 SCCPSolver(const TargetData *td) : TD(td) {}
206 /// MarkBlockExecutable - This method can be used by clients to mark all of
207 /// the blocks that are known to be intrinsically live in the processed unit.
209 /// This returns true if the block was not considered live before.
210 bool MarkBlockExecutable(BasicBlock *BB) {
211 if (!BBExecutable.insert(BB)) return false;
212 DEBUG(errs() << "Marking Block Executable: " << BB->getName() << "\n");
213 BBWorkList.push_back(BB); // Add the block to the work list!
217 /// TrackValueOfGlobalVariable - Clients can use this method to
218 /// inform the SCCPSolver that it should track loads and stores to the
219 /// specified global variable if it can. This is only legal to call if
220 /// performing Interprocedural SCCP.
221 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
222 const Type *ElTy = GV->getType()->getElementType();
223 if (ElTy->isFirstClassType()) {
224 LatticeVal &IV = TrackedGlobals[GV];
225 if (!isa<UndefValue>(GV->getInitializer()))
226 IV.markConstant(GV->getInitializer());
230 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
231 /// and out of the specified function (which cannot have its address taken),
232 /// this method must be called.
233 void AddTrackedFunction(Function *F) {
234 // Add an entry, F -> undef.
235 if (const StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
236 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
237 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
240 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
243 void AddArgumentTrackedFunction(Function *F) {
244 TrackingIncomingArguments.insert(F);
247 /// Solve - Solve for constants and executable blocks.
251 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
252 /// that branches on undef values cannot reach any of their successors.
253 /// However, this is not a safe assumption. After we solve dataflow, this
254 /// method should be use to handle this. If this returns true, the solver
256 bool ResolvedUndefsIn(Function &F);
258 bool isBlockExecutable(BasicBlock *BB) const {
259 return BBExecutable.count(BB);
262 LatticeVal getLatticeValueFor(Value *V) const {
263 DenseMap<Value*, LatticeVal>::const_iterator I = ValueState.find(V);
264 assert(I != ValueState.end() && "V is not in valuemap!");
268 /// getTrackedRetVals - Get the inferred return value map.
270 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
271 return TrackedRetVals;
274 /// getTrackedGlobals - Get and return the set of inferred initializers for
275 /// global variables.
276 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
277 return TrackedGlobals;
280 void markOverdefined(Value *V) {
281 markOverdefined(ValueState[V], V);
285 // markConstant - Make a value be marked as "constant". If the value
286 // is not already a constant, add it to the instruction work list so that
287 // the users of the instruction are updated later.
289 void markConstant(LatticeVal &IV, Value *V, Constant *C) {
290 if (!IV.markConstant(C)) return;
291 DEBUG(errs() << "markConstant: " << *C << ": " << *V << '\n');
292 InstWorkList.push_back(V);
295 void markConstant(Value *V, Constant *C) {
296 markConstant(ValueState[V], V, C);
299 void markForcedConstant(Value *V, Constant *C) {
300 ValueState[V].markForcedConstant(C);
301 DEBUG(errs() << "markForcedConstant: " << *C << ": " << *V << '\n');
302 InstWorkList.push_back(V);
306 // markOverdefined - Make a value be marked as "overdefined". If the
307 // value is not already overdefined, add it to the overdefined instruction
308 // work list so that the users of the instruction are updated later.
309 void markOverdefined(LatticeVal &IV, Value *V) {
310 if (!IV.markOverdefined()) return;
312 DEBUG(errs() << "markOverdefined: ";
313 if (Function *F = dyn_cast<Function>(V))
314 errs() << "Function '" << F->getName() << "'\n";
316 errs() << *V << '\n');
317 // Only instructions go on the work list
318 OverdefinedInstWorkList.push_back(V);
321 void mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) {
322 if (IV.isOverdefined() || MergeWithV.isUndefined())
324 if (MergeWithV.isOverdefined())
325 markOverdefined(IV, V);
326 else if (IV.isUndefined())
327 markConstant(IV, V, MergeWithV.getConstant());
328 else if (IV.getConstant() != MergeWithV.getConstant())
329 markOverdefined(IV, V);
332 void mergeInValue(Value *V, LatticeVal MergeWithV) {
333 mergeInValue(ValueState[V], V, MergeWithV);
337 /// getValueState - Return the LatticeVal object that corresponds to the
338 /// value. This function handles the case when the value hasn't been seen yet
339 /// by properly seeding constants etc.
340 LatticeVal &getValueState(Value *V) {
341 DenseMap<Value*, LatticeVal>::iterator I = ValueState.find(V);
342 if (I != ValueState.end()) return I->second; // Common case, in the map
344 LatticeVal &LV = ValueState[V];
346 if (Constant *C = dyn_cast<Constant>(V)) {
347 // Undef values remain undefined.
348 if (!isa<UndefValue>(V))
349 LV.markConstant(C); // Constants are constant
352 // All others are underdefined by default.
356 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
357 /// work list if it is not already executable.
358 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
359 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
360 return; // This edge is already known to be executable!
362 if (!MarkBlockExecutable(Dest)) {
363 // If the destination is already executable, we just made an *edge*
364 // feasible that wasn't before. Revisit the PHI nodes in the block
365 // because they have potentially new operands.
366 DEBUG(errs() << "Marking Edge Executable: " << Source->getName()
367 << " -> " << Dest->getName() << "\n");
370 for (BasicBlock::iterator I = Dest->begin();
371 (PN = dyn_cast<PHINode>(I)); ++I)
376 // getFeasibleSuccessors - Return a vector of booleans to indicate which
377 // successors are reachable from a given terminator instruction.
379 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
381 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
382 // block to the 'To' basic block is currently feasible.
384 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
386 // OperandChangedState - This method is invoked on all of the users of an
387 // instruction that was just changed state somehow. Based on this
388 // information, we need to update the specified user of this instruction.
390 void OperandChangedState(Instruction *I) {
391 if (BBExecutable.count(I->getParent())) // Inst is executable?
395 /// RemoveFromOverdefinedPHIs - If I has any entries in the
396 /// UsersOfOverdefinedPHIs map for PN, remove them now.
397 void RemoveFromOverdefinedPHIs(Instruction *I, PHINode *PN) {
398 if (UsersOfOverdefinedPHIs.empty()) return;
399 std::multimap<PHINode*, Instruction*>::iterator It, E;
400 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN);
403 UsersOfOverdefinedPHIs.erase(It++);
410 friend class InstVisitor<SCCPSolver>;
412 // visit implementations - Something changed in this instruction. Either an
413 // operand made a transition, or the instruction is newly executable. Change
414 // the value type of I to reflect these changes if appropriate.
415 void visitPHINode(PHINode &I);
418 void visitReturnInst(ReturnInst &I);
419 void visitTerminatorInst(TerminatorInst &TI);
421 void visitCastInst(CastInst &I);
422 void visitSelectInst(SelectInst &I);
423 void visitBinaryOperator(Instruction &I);
424 void visitCmpInst(CmpInst &I);
425 void visitExtractElementInst(ExtractElementInst &I);
426 void visitInsertElementInst(InsertElementInst &I);
427 void visitShuffleVectorInst(ShuffleVectorInst &I);
428 void visitExtractValueInst(ExtractValueInst &EVI);
429 void visitInsertValueInst(InsertValueInst &IVI);
431 // Instructions that cannot be folded away.
432 void visitStoreInst (StoreInst &I);
433 void visitLoadInst (LoadInst &I);
434 void visitGetElementPtrInst(GetElementPtrInst &I);
435 void visitCallInst (CallInst &I) {
436 visitCallSite(CallSite::get(&I));
438 void visitInvokeInst (InvokeInst &II) {
439 visitCallSite(CallSite::get(&II));
440 visitTerminatorInst(II);
442 void visitCallSite (CallSite CS);
443 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ }
444 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
445 void visitAllocaInst (Instruction &I) { markOverdefined(&I); }
446 void visitVANextInst (Instruction &I) { markOverdefined(&I); }
447 void visitVAArgInst (Instruction &I) { markOverdefined(&I); }
449 void visitInstruction(Instruction &I) {
450 // If a new instruction is added to LLVM that we don't handle.
451 errs() << "SCCP: Don't know how to handle: " << I;
452 markOverdefined(&I); // Just in case
456 } // end anonymous namespace
459 // getFeasibleSuccessors - Return a vector of booleans to indicate which
460 // successors are reachable from a given terminator instruction.
462 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
463 SmallVector<bool, 16> &Succs) {
464 Succs.resize(TI.getNumSuccessors());
465 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
466 if (BI->isUnconditional()) {
471 LatticeVal BCValue = getValueState(BI->getCondition());
472 ConstantInt *CI = BCValue.getConstantInt();
474 // Overdefined condition variables, and branches on unfoldable constant
475 // conditions, mean the branch could go either way.
476 if (!BCValue.isUndefined())
477 Succs[0] = Succs[1] = true;
481 // Constant condition variables mean the branch can only go a single way.
482 Succs[CI->isZero()] = true;
486 if (isa<InvokeInst>(TI)) {
487 // Invoke instructions successors are always executable.
488 Succs[0] = Succs[1] = true;
492 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
493 LatticeVal SCValue = getValueState(SI->getCondition());
494 ConstantInt *CI = SCValue.getConstantInt();
496 if (CI == 0) { // Overdefined or undefined condition?
497 // All destinations are executable!
498 if (!SCValue.isUndefined())
499 Succs.assign(TI.getNumSuccessors(), true);
503 Succs[SI->findCaseValue(CI)] = true;
507 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
508 if (isa<IndirectBrInst>(&TI)) {
509 // Just mark all destinations executable!
510 Succs.assign(TI.getNumSuccessors(), true);
515 errs() << "Unknown terminator instruction: " << TI << '\n';
517 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
521 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
522 // block to the 'To' basic block is currently feasible.
524 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
525 assert(BBExecutable.count(To) && "Dest should always be alive!");
527 // Make sure the source basic block is executable!!
528 if (!BBExecutable.count(From)) return false;
530 // Check to make sure this edge itself is actually feasible now.
531 TerminatorInst *TI = From->getTerminator();
532 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
533 if (BI->isUnconditional())
536 LatticeVal BCValue = getValueState(BI->getCondition());
538 // Overdefined condition variables mean the branch could go either way,
539 // undef conditions mean that neither edge is feasible yet.
540 ConstantInt *CI = BCValue.getConstantInt();
542 return !BCValue.isUndefined();
544 // Constant condition variables mean the branch can only go a single way.
545 return BI->getSuccessor(CI->isZero()) == To;
548 // Invoke instructions successors are always executable.
549 if (isa<InvokeInst>(TI))
552 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
553 LatticeVal SCValue = getValueState(SI->getCondition());
554 ConstantInt *CI = SCValue.getConstantInt();
557 return !SCValue.isUndefined();
559 // Make sure to skip the "default value" which isn't a value
560 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i)
561 if (SI->getSuccessorValue(i) == CI) // Found the taken branch.
562 return SI->getSuccessor(i) == To;
564 // If the constant value is not equal to any of the branches, we must
565 // execute default branch.
566 return SI->getDefaultDest() == To;
569 // Just mark all destinations executable!
570 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
571 if (isa<IndirectBrInst>(&TI))
575 errs() << "Unknown terminator instruction: " << *TI << '\n';
580 // visit Implementations - Something changed in this instruction, either an
581 // operand made a transition, or the instruction is newly executable. Change
582 // the value type of I to reflect these changes if appropriate. This method
583 // makes sure to do the following actions:
585 // 1. If a phi node merges two constants in, and has conflicting value coming
586 // from different branches, or if the PHI node merges in an overdefined
587 // value, then the PHI node becomes overdefined.
588 // 2. If a phi node merges only constants in, and they all agree on value, the
589 // PHI node becomes a constant value equal to that.
590 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
591 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
592 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
593 // 6. If a conditional branch has a value that is constant, make the selected
594 // destination executable
595 // 7. If a conditional branch has a value that is overdefined, make all
596 // successors executable.
598 void SCCPSolver::visitPHINode(PHINode &PN) {
599 if (getValueState(&PN).isOverdefined()) {
600 // There may be instructions using this PHI node that are not overdefined
601 // themselves. If so, make sure that they know that the PHI node operand
603 std::multimap<PHINode*, Instruction*>::iterator I, E;
604 tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN);
608 SmallVector<Instruction*, 16> Users;
610 Users.push_back(I->second);
611 while (!Users.empty())
612 visit(Users.pop_back_val());
613 return; // Quick exit
616 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
617 // and slow us down a lot. Just mark them overdefined.
618 if (PN.getNumIncomingValues() > 64)
619 return markOverdefined(&PN);
621 // Look at all of the executable operands of the PHI node. If any of them
622 // are overdefined, the PHI becomes overdefined as well. If they are all
623 // constant, and they agree with each other, the PHI becomes the identical
624 // constant. If they are constant and don't agree, the PHI is overdefined.
625 // If there are no executable operands, the PHI remains undefined.
627 Constant *OperandVal = 0;
628 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
629 LatticeVal IV = getValueState(PN.getIncomingValue(i));
630 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
632 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
635 if (IV.isOverdefined()) // PHI node becomes overdefined!
636 return markOverdefined(&PN);
638 if (OperandVal == 0) { // Grab the first value.
639 OperandVal = IV.getConstant();
643 // There is already a reachable operand. If we conflict with it,
644 // then the PHI node becomes overdefined. If we agree with it, we
647 // Check to see if there are two different constants merging, if so, the PHI
648 // node is overdefined.
649 if (IV.getConstant() != OperandVal)
650 return markOverdefined(&PN);
653 // If we exited the loop, this means that the PHI node only has constant
654 // arguments that agree with each other(and OperandVal is the constant) or
655 // OperandVal is null because there are no defined incoming arguments. If
656 // this is the case, the PHI remains undefined.
659 markConstant(&PN, OperandVal); // Acquire operand value
665 void SCCPSolver::visitReturnInst(ReturnInst &I) {
666 if (I.getNumOperands() == 0) return; // ret void
668 Function *F = I.getParent()->getParent();
670 // If we are tracking the return value of this function, merge it in.
671 if (!TrackedRetVals.empty()) {
672 DenseMap<Function*, LatticeVal>::iterator TFRVI =
673 TrackedRetVals.find(F);
674 if (TFRVI != TrackedRetVals.end()) {
675 mergeInValue(TFRVI->second, F, getValueState(I.getOperand(0)));
680 // Handle functions that return multiple values.
681 if (!TrackedMultipleRetVals.empty() &&
682 isa<StructType>(I.getOperand(0)->getType())) {
683 for (unsigned i = 0, e = I.getOperand(0)->getType()->getNumContainedTypes();
685 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
686 It = TrackedMultipleRetVals.find(std::make_pair(F, i));
687 if (It == TrackedMultipleRetVals.end()) break;
688 if (Value *Val = FindInsertedValue(I.getOperand(0), i, I.getContext()))
689 mergeInValue(It->second, F, getValueState(Val));
694 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
695 SmallVector<bool, 16> SuccFeasible;
696 getFeasibleSuccessors(TI, SuccFeasible);
698 BasicBlock *BB = TI.getParent();
700 // Mark all feasible successors executable.
701 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
703 markEdgeExecutable(BB, TI.getSuccessor(i));
706 void SCCPSolver::visitCastInst(CastInst &I) {
707 LatticeVal OpSt = getValueState(I.getOperand(0));
708 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand
710 else if (OpSt.isConstant()) // Propagate constant value
711 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
712 OpSt.getConstant(), I.getType()));
715 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
716 Value *Aggr = EVI.getAggregateOperand();
718 // If the operand to the extractvalue is an undef, the result is undef.
719 if (isa<UndefValue>(Aggr))
722 // Currently only handle single-index extractvalues.
723 if (EVI.getNumIndices() != 1)
724 return markOverdefined(&EVI);
727 if (CallInst *CI = dyn_cast<CallInst>(Aggr))
728 F = CI->getCalledFunction();
729 else if (InvokeInst *II = dyn_cast<InvokeInst>(Aggr))
730 F = II->getCalledFunction();
732 // TODO: If IPSCCP resolves the callee of this function, we could propagate a
734 if (F == 0 || TrackedMultipleRetVals.empty())
735 return markOverdefined(&EVI);
737 // See if we are tracking the result of the callee. If not tracking this
738 // function (for example, it is a declaration) just move to overdefined.
739 if (!TrackedMultipleRetVals.count(std::make_pair(F, *EVI.idx_begin())))
740 return markOverdefined(&EVI);
742 // Otherwise, the value will be merged in here as a result of CallSite
746 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
747 Value *Aggr = IVI.getAggregateOperand();
748 Value *Val = IVI.getInsertedValueOperand();
750 // If the operands to the insertvalue are undef, the result is undef.
751 if (isa<UndefValue>(Aggr) && isa<UndefValue>(Val))
754 // Currently only handle single-index insertvalues.
755 if (IVI.getNumIndices() != 1)
756 return markOverdefined(&IVI);
758 // Currently only handle insertvalue instructions that are in a single-use
759 // chain that builds up a return value.
760 for (const InsertValueInst *TmpIVI = &IVI; ; ) {
761 if (!TmpIVI->hasOneUse())
762 return markOverdefined(&IVI);
764 const Value *V = *TmpIVI->use_begin();
765 if (isa<ReturnInst>(V))
767 TmpIVI = dyn_cast<InsertValueInst>(V);
769 return markOverdefined(&IVI);
772 // See if we are tracking the result of the callee.
773 Function *F = IVI.getParent()->getParent();
774 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
775 It = TrackedMultipleRetVals.find(std::make_pair(F, *IVI.idx_begin()));
777 // Merge in the inserted member value.
778 if (It != TrackedMultipleRetVals.end())
779 mergeInValue(It->second, F, getValueState(Val));
781 // Mark the aggregate result of the IVI overdefined; any tracking that we do
782 // will be done on the individual member values.
783 markOverdefined(&IVI);
786 void SCCPSolver::visitSelectInst(SelectInst &I) {
787 LatticeVal CondValue = getValueState(I.getCondition());
788 if (CondValue.isUndefined())
791 if (ConstantInt *CondCB = CondValue.getConstantInt()) {
792 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
793 mergeInValue(&I, getValueState(OpVal));
797 // Otherwise, the condition is overdefined or a constant we can't evaluate.
798 // See if we can produce something better than overdefined based on the T/F
800 LatticeVal TVal = getValueState(I.getTrueValue());
801 LatticeVal FVal = getValueState(I.getFalseValue());
803 // select ?, C, C -> C.
804 if (TVal.isConstant() && FVal.isConstant() &&
805 TVal.getConstant() == FVal.getConstant())
806 return markConstant(&I, FVal.getConstant());
808 if (TVal.isUndefined()) // select ?, undef, X -> X.
809 return mergeInValue(&I, FVal);
810 if (FVal.isUndefined()) // select ?, X, undef -> X.
811 return mergeInValue(&I, TVal);
815 // Handle Binary Operators.
816 void SCCPSolver::visitBinaryOperator(Instruction &I) {
817 LatticeVal V1State = getValueState(I.getOperand(0));
818 LatticeVal V2State = getValueState(I.getOperand(1));
820 LatticeVal &IV = ValueState[&I];
821 if (IV.isOverdefined()) return;
823 if (V1State.isConstant() && V2State.isConstant())
824 return markConstant(IV, &I,
825 ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
826 V2State.getConstant()));
828 // If something is undef, wait for it to resolve.
829 if (!V1State.isOverdefined() && !V2State.isOverdefined())
832 // Otherwise, one of our operands is overdefined. Try to produce something
833 // better than overdefined with some tricks.
835 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
836 // operand is overdefined.
837 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
838 LatticeVal *NonOverdefVal = 0;
839 if (!V1State.isOverdefined())
840 NonOverdefVal = &V1State;
841 else if (!V2State.isOverdefined())
842 NonOverdefVal = &V2State;
845 if (NonOverdefVal->isUndefined()) {
846 // Could annihilate value.
847 if (I.getOpcode() == Instruction::And)
848 markConstant(IV, &I, Constant::getNullValue(I.getType()));
849 else if (const VectorType *PT = dyn_cast<VectorType>(I.getType()))
850 markConstant(IV, &I, Constant::getAllOnesValue(PT));
853 Constant::getAllOnesValue(I.getType()));
857 if (I.getOpcode() == Instruction::And) {
859 if (NonOverdefVal->getConstant()->isNullValue())
860 return markConstant(IV, &I, NonOverdefVal->getConstant());
862 if (ConstantInt *CI = NonOverdefVal->getConstantInt())
863 if (CI->isAllOnesValue()) // X or -1 = -1
864 return markConstant(IV, &I, NonOverdefVal->getConstant());
870 // If both operands are PHI nodes, it is possible that this instruction has
871 // a constant value, despite the fact that the PHI node doesn't. Check for
872 // this condition now.
873 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
874 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
875 if (PN1->getParent() == PN2->getParent()) {
876 // Since the two PHI nodes are in the same basic block, they must have
877 // entries for the same predecessors. Walk the predecessor list, and
878 // if all of the incoming values are constants, and the result of
879 // evaluating this expression with all incoming value pairs is the
880 // same, then this expression is a constant even though the PHI node
881 // is not a constant!
883 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
884 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
885 BasicBlock *InBlock = PN1->getIncomingBlock(i);
886 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
888 if (In1.isOverdefined() || In2.isOverdefined()) {
889 Result.markOverdefined();
890 break; // Cannot fold this operation over the PHI nodes!
893 if (In1.isConstant() && In2.isConstant()) {
894 Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(),
896 if (Result.isUndefined())
897 Result.markConstant(V);
898 else if (Result.isConstant() && Result.getConstant() != V) {
899 Result.markOverdefined();
905 // If we found a constant value here, then we know the instruction is
906 // constant despite the fact that the PHI nodes are overdefined.
907 if (Result.isConstant()) {
908 markConstant(IV, &I, Result.getConstant());
909 // Remember that this instruction is virtually using the PHI node
911 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
912 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
916 if (Result.isUndefined())
919 // Okay, this really is overdefined now. Since we might have
920 // speculatively thought that this was not overdefined before, and
921 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
922 // make sure to clean out any entries that we put there, for
924 RemoveFromOverdefinedPHIs(&I, PN1);
925 RemoveFromOverdefinedPHIs(&I, PN2);
931 // Handle ICmpInst instruction.
932 void SCCPSolver::visitCmpInst(CmpInst &I) {
933 LatticeVal V1State = getValueState(I.getOperand(0));
934 LatticeVal V2State = getValueState(I.getOperand(1));
936 LatticeVal &IV = ValueState[&I];
937 if (IV.isOverdefined()) return;
939 if (V1State.isConstant() && V2State.isConstant())
940 return markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
941 V1State.getConstant(),
942 V2State.getConstant()));
944 // If operands are still undefined, wait for it to resolve.
945 if (!V1State.isOverdefined() && !V2State.isOverdefined())
948 // If something is overdefined, use some tricks to avoid ending up and over
949 // defined if we can.
951 // If both operands are PHI nodes, it is possible that this instruction has
952 // a constant value, despite the fact that the PHI node doesn't. Check for
953 // this condition now.
954 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
955 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
956 if (PN1->getParent() == PN2->getParent()) {
957 // Since the two PHI nodes are in the same basic block, they must have
958 // entries for the same predecessors. Walk the predecessor list, and
959 // if all of the incoming values are constants, and the result of
960 // evaluating this expression with all incoming value pairs is the
961 // same, then this expression is a constant even though the PHI node
962 // is not a constant!
964 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
965 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
966 BasicBlock *InBlock = PN1->getIncomingBlock(i);
967 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
969 if (In1.isOverdefined() || In2.isOverdefined()) {
970 Result.markOverdefined();
971 break; // Cannot fold this operation over the PHI nodes!
974 if (In1.isConstant() && In2.isConstant()) {
975 Constant *V = ConstantExpr::getCompare(I.getPredicate(),
978 if (Result.isUndefined())
979 Result.markConstant(V);
980 else if (Result.isConstant() && Result.getConstant() != V) {
981 Result.markOverdefined();
987 // If we found a constant value here, then we know the instruction is
988 // constant despite the fact that the PHI nodes are overdefined.
989 if (Result.isConstant()) {
990 markConstant(&I, Result.getConstant());
991 // Remember that this instruction is virtually using the PHI node
993 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
994 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
998 if (Result.isUndefined())
1001 // Okay, this really is overdefined now. Since we might have
1002 // speculatively thought that this was not overdefined before, and
1003 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
1004 // make sure to clean out any entries that we put there, for
1006 RemoveFromOverdefinedPHIs(&I, PN1);
1007 RemoveFromOverdefinedPHIs(&I, PN2);
1010 markOverdefined(&I);
1013 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
1014 // FIXME : SCCP does not handle vectors properly.
1015 return markOverdefined(&I);
1018 LatticeVal &ValState = getValueState(I.getOperand(0));
1019 LatticeVal &IdxState = getValueState(I.getOperand(1));
1021 if (ValState.isOverdefined() || IdxState.isOverdefined())
1022 markOverdefined(&I);
1023 else if(ValState.isConstant() && IdxState.isConstant())
1024 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
1025 IdxState.getConstant()));
1029 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
1030 // FIXME : SCCP does not handle vectors properly.
1031 return markOverdefined(&I);
1033 LatticeVal &ValState = getValueState(I.getOperand(0));
1034 LatticeVal &EltState = getValueState(I.getOperand(1));
1035 LatticeVal &IdxState = getValueState(I.getOperand(2));
1037 if (ValState.isOverdefined() || EltState.isOverdefined() ||
1038 IdxState.isOverdefined())
1039 markOverdefined(&I);
1040 else if(ValState.isConstant() && EltState.isConstant() &&
1041 IdxState.isConstant())
1042 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
1043 EltState.getConstant(),
1044 IdxState.getConstant()));
1045 else if (ValState.isUndefined() && EltState.isConstant() &&
1046 IdxState.isConstant())
1047 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
1048 EltState.getConstant(),
1049 IdxState.getConstant()));
1053 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
1054 // FIXME : SCCP does not handle vectors properly.
1055 return markOverdefined(&I);
1057 LatticeVal &V1State = getValueState(I.getOperand(0));
1058 LatticeVal &V2State = getValueState(I.getOperand(1));
1059 LatticeVal &MaskState = getValueState(I.getOperand(2));
1061 if (MaskState.isUndefined() ||
1062 (V1State.isUndefined() && V2State.isUndefined()))
1063 return; // Undefined output if mask or both inputs undefined.
1065 if (V1State.isOverdefined() || V2State.isOverdefined() ||
1066 MaskState.isOverdefined()) {
1067 markOverdefined(&I);
1069 // A mix of constant/undef inputs.
1070 Constant *V1 = V1State.isConstant() ?
1071 V1State.getConstant() : UndefValue::get(I.getType());
1072 Constant *V2 = V2State.isConstant() ?
1073 V2State.getConstant() : UndefValue::get(I.getType());
1074 Constant *Mask = MaskState.isConstant() ?
1075 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
1076 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
1081 // Handle getelementptr instructions. If all operands are constants then we
1082 // can turn this into a getelementptr ConstantExpr.
1084 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1085 if (ValueState[&I].isOverdefined()) return;
1087 SmallVector<Constant*, 8> Operands;
1088 Operands.reserve(I.getNumOperands());
1090 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1091 LatticeVal State = getValueState(I.getOperand(i));
1092 if (State.isUndefined())
1093 return; // Operands are not resolved yet.
1095 if (State.isOverdefined())
1096 return markOverdefined(&I);
1098 assert(State.isConstant() && "Unknown state!");
1099 Operands.push_back(State.getConstant());
1102 Constant *Ptr = Operands[0];
1103 markConstant(&I, ConstantExpr::getGetElementPtr(Ptr, &Operands[0]+1,
1104 Operands.size()-1));
1107 void SCCPSolver::visitStoreInst(StoreInst &SI) {
1108 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1111 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1112 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1113 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1115 // Get the value we are storing into the global, then merge it.
1116 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)));
1117 if (I->second.isOverdefined())
1118 TrackedGlobals.erase(I); // No need to keep tracking this!
1122 // Handle load instructions. If the operand is a constant pointer to a constant
1123 // global, we can replace the load with the loaded constant value!
1124 void SCCPSolver::visitLoadInst(LoadInst &I) {
1125 LatticeVal PtrVal = getValueState(I.getOperand(0));
1126 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1128 LatticeVal &IV = ValueState[&I];
1129 if (IV.isOverdefined()) return;
1131 if (!PtrVal.isConstant() || I.isVolatile())
1132 return markOverdefined(IV, &I);
1134 Constant *Ptr = PtrVal.getConstant();
1136 // load null -> null
1137 if (isa<ConstantPointerNull>(Ptr) && I.getPointerAddressSpace() == 0)
1138 return markConstant(IV, &I, Constant::getNullValue(I.getType()));
1140 // Transform load (constant global) into the value loaded.
1141 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1142 if (!TrackedGlobals.empty()) {
1143 // If we are tracking this global, merge in the known value for it.
1144 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1145 TrackedGlobals.find(GV);
1146 if (It != TrackedGlobals.end()) {
1147 mergeInValue(IV, &I, It->second);
1153 // Transform load from a constant into a constant if possible.
1154 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, TD))
1155 return markConstant(IV, &I, C);
1157 // Otherwise we cannot say for certain what value this load will produce.
1159 markOverdefined(IV, &I);
1162 void SCCPSolver::visitCallSite(CallSite CS) {
1163 Function *F = CS.getCalledFunction();
1164 Instruction *I = CS.getInstruction();
1166 // The common case is that we aren't tracking the callee, either because we
1167 // are not doing interprocedural analysis or the callee is indirect, or is
1168 // external. Handle these cases first.
1169 if (F == 0 || F->isDeclaration()) {
1171 // Void return and not tracking callee, just bail.
1172 if (I->getType()->isVoidTy()) return;
1174 // Otherwise, if we have a single return value case, and if the function is
1175 // a declaration, maybe we can constant fold it.
1176 if (F && F->isDeclaration() && !isa<StructType>(I->getType()) &&
1177 canConstantFoldCallTo(F)) {
1179 SmallVector<Constant*, 8> Operands;
1180 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1182 LatticeVal State = getValueState(*AI);
1184 if (State.isUndefined())
1185 return; // Operands are not resolved yet.
1186 if (State.isOverdefined())
1187 return markOverdefined(I);
1188 assert(State.isConstant() && "Unknown state!");
1189 Operands.push_back(State.getConstant());
1192 // If we can constant fold this, mark the result of the call as a
1194 if (Constant *C = ConstantFoldCall(F, Operands.data(), Operands.size()))
1195 return markConstant(I, C);
1198 // Otherwise, we don't know anything about this call, mark it overdefined.
1199 return markOverdefined(I);
1202 // If this is a local function that doesn't have its address taken, mark its
1203 // entry block executable and merge in the actual arguments to the call into
1204 // the formal arguments of the function.
1205 if (!TrackingIncomingArguments.empty() && TrackingIncomingArguments.count(F)){
1206 MarkBlockExecutable(F->begin());
1208 // Propagate information from this call site into the callee.
1209 CallSite::arg_iterator CAI = CS.arg_begin();
1210 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1211 AI != E; ++AI, ++CAI) {
1212 // If this argument is byval, and if the function is not readonly, there
1213 // will be an implicit copy formed of the input aggregate.
1214 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1215 markOverdefined(AI);
1219 mergeInValue(AI, getValueState(*CAI));
1223 // If this is a single/zero retval case, see if we're tracking the function.
1224 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1225 if (TFRVI != TrackedRetVals.end()) {
1226 // If so, propagate the return value of the callee into this call result.
1227 mergeInValue(I, TFRVI->second);
1228 } else if (isa<StructType>(I->getType())) {
1229 // Check to see if we're tracking this callee, if not, handle it in the
1230 // common path above.
1231 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
1232 TMRVI = TrackedMultipleRetVals.find(std::make_pair(F, 0));
1233 if (TMRVI == TrackedMultipleRetVals.end())
1234 goto CallOverdefined;
1236 // Need to mark as overdefined, otherwise it stays undefined which
1237 // creates extractvalue undef, <idx>
1240 // If we are tracking this callee, propagate the return values of the call
1241 // into this call site. We do this by walking all the uses. Single-index
1242 // ExtractValueInst uses can be tracked; anything more complicated is
1243 // currently handled conservatively.
1244 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1246 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(*UI)) {
1247 if (EVI->getNumIndices() == 1) {
1249 TrackedMultipleRetVals[std::make_pair(F, *EVI->idx_begin())]);
1253 // The aggregate value is used in a way not handled here. Assume nothing.
1254 markOverdefined(*UI);
1257 // Otherwise we're not tracking this callee, so handle it in the
1258 // common path above.
1259 goto CallOverdefined;
1263 void SCCPSolver::Solve() {
1264 // Process the work lists until they are empty!
1265 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1266 !OverdefinedInstWorkList.empty()) {
1267 // Process the overdefined instruction's work list first, which drives other
1268 // things to overdefined more quickly.
1269 while (!OverdefinedInstWorkList.empty()) {
1270 Value *I = OverdefinedInstWorkList.pop_back_val();
1272 DEBUG(errs() << "\nPopped off OI-WL: " << *I << '\n');
1274 // "I" got into the work list because it either made the transition from
1275 // bottom to constant
1277 // Anything on this worklist that is overdefined need not be visited
1278 // since all of its users will have already been marked as overdefined
1279 // Update all of the users of this instruction's value.
1281 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1283 if (Instruction *I = dyn_cast<Instruction>(*UI))
1284 OperandChangedState(I);
1287 // Process the instruction work list.
1288 while (!InstWorkList.empty()) {
1289 Value *I = InstWorkList.pop_back_val();
1291 DEBUG(errs() << "\nPopped off I-WL: " << *I << '\n');
1293 // "I" got into the work list because it made the transition from undef to
1296 // Anything on this worklist that is overdefined need not be visited
1297 // since all of its users will have already been marked as overdefined.
1298 // Update all of the users of this instruction's value.
1300 if (!getValueState(I).isOverdefined())
1301 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1303 if (Instruction *I = dyn_cast<Instruction>(*UI))
1304 OperandChangedState(I);
1307 // Process the basic block work list.
1308 while (!BBWorkList.empty()) {
1309 BasicBlock *BB = BBWorkList.back();
1310 BBWorkList.pop_back();
1312 DEBUG(errs() << "\nPopped off BBWL: " << *BB << '\n');
1314 // Notify all instructions in this basic block that they are newly
1321 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1322 /// that branches on undef values cannot reach any of their successors.
1323 /// However, this is not a safe assumption. After we solve dataflow, this
1324 /// method should be use to handle this. If this returns true, the solver
1325 /// should be rerun.
1327 /// This method handles this by finding an unresolved branch and marking it one
1328 /// of the edges from the block as being feasible, even though the condition
1329 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1330 /// CFG and only slightly pessimizes the analysis results (by marking one,
1331 /// potentially infeasible, edge feasible). This cannot usefully modify the
1332 /// constraints on the condition of the branch, as that would impact other users
1335 /// This scan also checks for values that use undefs, whose results are actually
1336 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1337 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1338 /// even if X isn't defined.
1339 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1340 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1341 if (!BBExecutable.count(BB))
1344 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1345 // Look for instructions which produce undef values.
1346 if (I->getType()->isVoidTy()) continue;
1348 LatticeVal &LV = getValueState(I);
1349 if (!LV.isUndefined()) continue;
1351 // Get the lattice values of the first two operands for use below.
1352 LatticeVal Op0LV = getValueState(I->getOperand(0));
1354 if (I->getNumOperands() == 2) {
1355 // If this is a two-operand instruction, and if both operands are
1356 // undefs, the result stays undef.
1357 Op1LV = getValueState(I->getOperand(1));
1358 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1362 // If this is an instructions whose result is defined even if the input is
1363 // not fully defined, propagate the information.
1364 const Type *ITy = I->getType();
1365 switch (I->getOpcode()) {
1366 default: break; // Leave the instruction as an undef.
1367 case Instruction::ZExt:
1368 // After a zero extend, we know the top part is zero. SExt doesn't have
1369 // to be handled here, because we don't know whether the top part is 1's
1371 markForcedConstant(I, Constant::getNullValue(ITy));
1373 case Instruction::Mul:
1374 case Instruction::And:
1375 // undef * X -> 0. X could be zero.
1376 // undef & X -> 0. X could be zero.
1377 markForcedConstant(I, Constant::getNullValue(ITy));
1380 case Instruction::Or:
1381 // undef | X -> -1. X could be -1.
1382 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1385 case Instruction::SDiv:
1386 case Instruction::UDiv:
1387 case Instruction::SRem:
1388 case Instruction::URem:
1389 // X / undef -> undef. No change.
1390 // X % undef -> undef. No change.
1391 if (Op1LV.isUndefined()) break;
1393 // undef / X -> 0. X could be maxint.
1394 // undef % X -> 0. X could be 1.
1395 markForcedConstant(I, Constant::getNullValue(ITy));
1398 case Instruction::AShr:
1399 // undef >>s X -> undef. No change.
1400 if (Op0LV.isUndefined()) break;
1402 // X >>s undef -> X. X could be 0, X could have the high-bit known set.
1403 if (Op0LV.isConstant())
1404 markForcedConstant(I, Op0LV.getConstant());
1408 case Instruction::LShr:
1409 case Instruction::Shl:
1410 // undef >> X -> undef. No change.
1411 // undef << X -> undef. No change.
1412 if (Op0LV.isUndefined()) break;
1414 // X >> undef -> 0. X could be 0.
1415 // X << undef -> 0. X could be 0.
1416 markForcedConstant(I, Constant::getNullValue(ITy));
1418 case Instruction::Select:
1419 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1420 if (Op0LV.isUndefined()) {
1421 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1422 Op1LV = getValueState(I->getOperand(2));
1423 } else if (Op1LV.isUndefined()) {
1424 // c ? undef : undef -> undef. No change.
1425 Op1LV = getValueState(I->getOperand(2));
1426 if (Op1LV.isUndefined())
1428 // Otherwise, c ? undef : x -> x.
1430 // Leave Op1LV as Operand(1)'s LatticeValue.
1433 if (Op1LV.isConstant())
1434 markForcedConstant(I, Op1LV.getConstant());
1438 case Instruction::Call:
1439 // If a call has an undef result, it is because it is constant foldable
1440 // but one of the inputs was undef. Just force the result to
1447 TerminatorInst *TI = BB->getTerminator();
1448 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1449 if (!BI->isConditional()) continue;
1450 if (!getValueState(BI->getCondition()).isUndefined())
1452 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1453 if (SI->getNumSuccessors() < 2) // no cases
1455 if (!getValueState(SI->getCondition()).isUndefined())
1461 // If the edge to the second successor isn't thought to be feasible yet,
1462 // mark it so now. We pick the second one so that this goes to some
1463 // enumerated value in a switch instead of going to the default destination.
1464 if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(1))))
1467 // Otherwise, it isn't already thought to be feasible. Mark it as such now
1468 // and return. This will make other blocks reachable, which will allow new
1469 // values to be discovered and existing ones to be moved in the lattice.
1470 markEdgeExecutable(BB, TI->getSuccessor(1));
1472 // This must be a conditional branch of switch on undef. At this point,
1473 // force the old terminator to branch to the first successor. This is
1474 // required because we are now influencing the dataflow of the function with
1475 // the assumption that this edge is taken. If we leave the branch condition
1476 // as undef, then further analysis could think the undef went another way
1477 // leading to an inconsistent set of conclusions.
1478 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1479 BI->setCondition(ConstantInt::getFalse(BI->getContext()));
1481 SwitchInst *SI = cast<SwitchInst>(TI);
1482 SI->setCondition(SI->getCaseValue(1));
1493 //===--------------------------------------------------------------------===//
1495 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1496 /// Sparse Conditional Constant Propagator.
1498 struct SCCP : public FunctionPass {
1499 static char ID; // Pass identification, replacement for typeid
1500 SCCP() : FunctionPass(&ID) {}
1502 // runOnFunction - Run the Sparse Conditional Constant Propagation
1503 // algorithm, and return true if the function was modified.
1505 bool runOnFunction(Function &F);
1507 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1508 AU.setPreservesCFG();
1511 } // end anonymous namespace
1514 static RegisterPass<SCCP>
1515 X("sccp", "Sparse Conditional Constant Propagation");
1517 // createSCCPPass - This is the public interface to this file.
1518 FunctionPass *llvm::createSCCPPass() {
1522 static void DeleteInstructionInBlock(BasicBlock *BB) {
1523 DEBUG(errs() << " BasicBlock Dead:" << *BB);
1526 // Delete the instructions backwards, as it has a reduced likelihood of
1527 // having to update as many def-use and use-def chains.
1528 while (!isa<TerminatorInst>(BB->begin())) {
1529 Instruction *I = --BasicBlock::iterator(BB->getTerminator());
1531 if (!I->use_empty())
1532 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1533 BB->getInstList().erase(I);
1538 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1539 // and return true if the function was modified.
1541 bool SCCP::runOnFunction(Function &F) {
1542 DEBUG(errs() << "SCCP on function '" << F.getName() << "'\n");
1543 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1545 // Mark the first block of the function as being executable.
1546 Solver.MarkBlockExecutable(F.begin());
1548 // Mark all arguments to the function as being overdefined.
1549 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1550 Solver.markOverdefined(AI);
1552 // Solve for constants.
1553 bool ResolvedUndefs = true;
1554 while (ResolvedUndefs) {
1556 DEBUG(errs() << "RESOLVING UNDEFs\n");
1557 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1560 bool MadeChanges = false;
1562 // If we decided that there are basic blocks that are dead in this function,
1563 // delete their contents now. Note that we cannot actually delete the blocks,
1564 // as we cannot modify the CFG of the function.
1566 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1567 if (!Solver.isBlockExecutable(BB)) {
1568 DeleteInstructionInBlock(BB);
1573 // Iterate over all of the instructions in a function, replacing them with
1574 // constants if we have found them to be of constant values.
1576 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1577 Instruction *Inst = BI++;
1578 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst))
1581 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1582 if (IV.isOverdefined())
1585 Constant *Const = IV.isConstant()
1586 ? IV.getConstant() : UndefValue::get(Inst->getType());
1587 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1589 // Replaces all of the uses of a variable with uses of the constant.
1590 Inst->replaceAllUsesWith(Const);
1592 // Delete the instruction.
1593 Inst->eraseFromParent();
1595 // Hey, we just changed something!
1605 //===--------------------------------------------------------------------===//
1607 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1608 /// Constant Propagation.
1610 struct IPSCCP : public ModulePass {
1612 IPSCCP() : ModulePass(&ID) {}
1613 bool runOnModule(Module &M);
1615 } // end anonymous namespace
1617 char IPSCCP::ID = 0;
1618 static RegisterPass<IPSCCP>
1619 Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation");
1621 // createIPSCCPPass - This is the public interface to this file.
1622 ModulePass *llvm::createIPSCCPPass() {
1623 return new IPSCCP();
1627 static bool AddressIsTaken(GlobalValue *GV) {
1628 // Delete any dead constantexpr klingons.
1629 GV->removeDeadConstantUsers();
1631 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
1633 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1634 if (SI->getOperand(0) == GV || SI->isVolatile())
1635 return true; // Storing addr of GV.
1636 } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) {
1637 // Make sure we are calling the function, not passing the address.
1638 if (UI.getOperandNo() != 0)
1640 } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1641 if (LI->isVolatile())
1643 } else if (isa<BlockAddress>(*UI)) {
1644 // blockaddress doesn't take the address of the function, it takes addr
1652 bool IPSCCP::runOnModule(Module &M) {
1653 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1655 // Loop over all functions, marking arguments to those with their addresses
1656 // taken or that are external as overdefined.
1658 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1659 if (F->isDeclaration())
1662 // If this is a strong or ODR definition of this function, then we can
1663 // propagate information about its result into callsites of it.
1664 if (!F->mayBeOverridden() &&
1665 !isa<StructType>(F->getReturnType()))
1666 Solver.AddTrackedFunction(F);
1668 // If this function only has direct calls that we can see, we can track its
1669 // arguments and return value aggressively, and can assume it is not called
1670 // unless we see evidence to the contrary.
1671 if (F->hasLocalLinkage() && !AddressIsTaken(F)) {
1672 Solver.AddArgumentTrackedFunction(F);
1676 // Assume the function is called.
1677 Solver.MarkBlockExecutable(F->begin());
1679 // Assume nothing about the incoming arguments.
1680 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1682 Solver.markOverdefined(AI);
1685 // Loop over global variables. We inform the solver about any internal global
1686 // variables that do not have their 'addresses taken'. If they don't have
1687 // their addresses taken, we can propagate constants through them.
1688 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1690 if (!G->isConstant() && G->hasLocalLinkage() && !AddressIsTaken(G))
1691 Solver.TrackValueOfGlobalVariable(G);
1693 // Solve for constants.
1694 bool ResolvedUndefs = true;
1695 while (ResolvedUndefs) {
1698 DEBUG(errs() << "RESOLVING UNDEFS\n");
1699 ResolvedUndefs = false;
1700 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1701 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1704 bool MadeChanges = false;
1706 // Iterate over all of the instructions in the module, replacing them with
1707 // constants if we have found them to be of constant values.
1709 SmallVector<BasicBlock*, 512> BlocksToErase;
1711 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1712 if (Solver.isBlockExecutable(F->begin())) {
1713 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1715 if (AI->use_empty()) continue;
1717 LatticeVal IV = Solver.getLatticeValueFor(AI);
1718 if (IV.isOverdefined()) continue;
1720 Constant *CST = IV.isConstant() ?
1721 IV.getConstant() : UndefValue::get(AI->getType());
1722 DEBUG(errs() << "*** Arg " << *AI << " = " << *CST <<"\n");
1724 // Replaces all of the uses of a variable with uses of the
1726 AI->replaceAllUsesWith(CST);
1731 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1732 if (!Solver.isBlockExecutable(BB)) {
1733 DeleteInstructionInBlock(BB);
1736 TerminatorInst *TI = BB->getTerminator();
1737 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1738 BasicBlock *Succ = TI->getSuccessor(i);
1739 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1740 TI->getSuccessor(i)->removePredecessor(BB);
1742 if (!TI->use_empty())
1743 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1744 TI->eraseFromParent();
1746 if (&*BB != &F->front())
1747 BlocksToErase.push_back(BB);
1749 new UnreachableInst(M.getContext(), BB);
1753 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1754 Instruction *Inst = BI++;
1755 if (Inst->getType()->isVoidTy())
1758 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1759 if (IV.isOverdefined())
1762 Constant *Const = IV.isConstant()
1763 ? IV.getConstant() : UndefValue::get(Inst->getType());
1764 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1766 // Replaces all of the uses of a variable with uses of the
1768 Inst->replaceAllUsesWith(Const);
1770 // Delete the instruction.
1771 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst))
1772 Inst->eraseFromParent();
1774 // Hey, we just changed something!
1780 // Now that all instructions in the function are constant folded, erase dead
1781 // blocks, because we can now use ConstantFoldTerminator to get rid of
1783 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1784 // If there are any PHI nodes in this successor, drop entries for BB now.
1785 BasicBlock *DeadBB = BlocksToErase[i];
1786 while (!DeadBB->use_empty()) {
1787 Instruction *I = cast<Instruction>(DeadBB->use_back());
1788 bool Folded = ConstantFoldTerminator(I->getParent());
1790 // The constant folder may not have been able to fold the terminator
1791 // if this is a branch or switch on undef. Fold it manually as a
1792 // branch to the first successor.
1794 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1795 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1796 "Branch should be foldable!");
1797 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1798 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1800 llvm_unreachable("Didn't fold away reference to block!");
1804 // Make this an uncond branch to the first successor.
1805 TerminatorInst *TI = I->getParent()->getTerminator();
1806 BranchInst::Create(TI->getSuccessor(0), TI);
1808 // Remove entries in successor phi nodes to remove edges.
1809 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1810 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1812 // Remove the old terminator.
1813 TI->eraseFromParent();
1817 // Finally, delete the basic block.
1818 F->getBasicBlockList().erase(DeadBB);
1820 BlocksToErase.clear();
1823 // If we inferred constant or undef return values for a function, we replaced
1824 // all call uses with the inferred value. This means we don't need to bother
1825 // actually returning anything from the function. Replace all return
1826 // instructions with return undef.
1827 // TODO: Process multiple value ret instructions also.
1828 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1829 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1830 E = RV.end(); I != E; ++I) {
1831 Function *F = I->first;
1832 if (I->second.isOverdefined() || F->getReturnType()->isVoidTy())
1835 // We can only do this if we know that nothing else can call the function.
1836 if (!F->hasLocalLinkage() || AddressIsTaken(F))
1839 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1840 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1841 if (!isa<UndefValue>(RI->getOperand(0)))
1842 RI->setOperand(0, UndefValue::get(F->getReturnType()));
1845 // If we infered constant or undef values for globals variables, we can delete
1846 // the global and any stores that remain to it.
1847 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1848 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1849 E = TG.end(); I != E; ++I) {
1850 GlobalVariable *GV = I->first;
1851 assert(!I->second.isOverdefined() &&
1852 "Overdefined values should have been taken out of the map!");
1853 DEBUG(errs() << "Found that GV '" << GV->getName() << "' is constant!\n");
1854 while (!GV->use_empty()) {
1855 StoreInst *SI = cast<StoreInst>(GV->use_back());
1856 SI->eraseFromParent();
1858 M.getGlobalList().erase(GV);