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 #include "llvm/Transforms/Scalar.h"
21 #include "llvm/ADT/DenseMap.h"
22 #include "llvm/ADT/DenseSet.h"
23 #include "llvm/ADT/PointerIntPair.h"
24 #include "llvm/ADT/SmallPtrSet.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/GlobalsModRef.h"
28 #include "llvm/Analysis/ConstantFolding.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/IR/CallSite.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/InstVisitor.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/Pass.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Transforms/IPO.h"
41 #include "llvm/Transforms/Utils/Local.h"
45 #define DEBUG_TYPE "sccp"
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(nullptr, 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 DataLayout &DL;
159 const TargetLibraryInfo *TLI;
160 SmallPtrSet<BasicBlock*, 8> BBExecutable; // The BBs that are executable.
161 DenseMap<Value*, LatticeVal> ValueState; // The state each value is in.
163 /// StructValueState - This maintains ValueState for values that have
164 /// StructType, for example for formal arguments, calls, insertelement, etc.
166 DenseMap<std::pair<Value*, unsigned>, LatticeVal> StructValueState;
168 /// GlobalValue - If we are tracking any values for the contents of a global
169 /// variable, we keep a mapping from the constant accessor to the element of
170 /// the global, to the currently known value. If the value becomes
171 /// overdefined, it's entry is simply removed from this map.
172 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
174 /// TrackedRetVals - If we are tracking arguments into and the return
175 /// value out of a function, it will have an entry in this map, indicating
176 /// what the known return value for the function is.
177 DenseMap<Function*, LatticeVal> TrackedRetVals;
179 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
180 /// that return multiple values.
181 DenseMap<std::pair<Function*, unsigned>, LatticeVal> TrackedMultipleRetVals;
183 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is
184 /// represented here for efficient lookup.
185 SmallPtrSet<Function*, 16> MRVFunctionsTracked;
187 /// TrackingIncomingArguments - This is the set of functions for whose
188 /// arguments we make optimistic assumptions about and try to prove as
190 SmallPtrSet<Function*, 16> TrackingIncomingArguments;
192 /// The reason for two worklists is that overdefined is the lowest state
193 /// on the lattice, and moving things to overdefined as fast as possible
194 /// makes SCCP converge much faster.
196 /// By having a separate worklist, we accomplish this because everything
197 /// possibly overdefined will become overdefined at the soonest possible
199 SmallVector<Value*, 64> OverdefinedInstWorkList;
200 SmallVector<Value*, 64> InstWorkList;
203 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list
205 /// KnownFeasibleEdges - Entries in this set are edges which have already had
206 /// PHI nodes retriggered.
207 typedef std::pair<BasicBlock*, BasicBlock*> Edge;
208 DenseSet<Edge> KnownFeasibleEdges;
210 SCCPSolver(const DataLayout &DL, const TargetLibraryInfo *tli)
211 : DL(DL), TLI(tli) {}
213 /// MarkBlockExecutable - This method can be used by clients to mark all of
214 /// the blocks that are known to be intrinsically live in the processed unit.
216 /// This returns true if the block was not considered live before.
217 bool MarkBlockExecutable(BasicBlock *BB) {
218 if (!BBExecutable.insert(BB).second)
220 DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n');
221 BBWorkList.push_back(BB); // Add the block to the work list!
225 /// TrackValueOfGlobalVariable - Clients can use this method to
226 /// inform the SCCPSolver that it should track loads and stores to the
227 /// specified global variable if it can. This is only legal to call if
228 /// performing Interprocedural SCCP.
229 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
230 // We only track the contents of scalar globals.
231 if (GV->getType()->getElementType()->isSingleValueType()) {
232 LatticeVal &IV = TrackedGlobals[GV];
233 if (!isa<UndefValue>(GV->getInitializer()))
234 IV.markConstant(GV->getInitializer());
238 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
239 /// and out of the specified function (which cannot have its address taken),
240 /// this method must be called.
241 void AddTrackedFunction(Function *F) {
242 // Add an entry, F -> undef.
243 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
244 MRVFunctionsTracked.insert(F);
245 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
246 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
249 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
252 void AddArgumentTrackedFunction(Function *F) {
253 TrackingIncomingArguments.insert(F);
256 /// Solve - Solve for constants and executable blocks.
260 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
261 /// that branches on undef values cannot reach any of their successors.
262 /// However, this is not a safe assumption. After we solve dataflow, this
263 /// method should be use to handle this. If this returns true, the solver
265 bool ResolvedUndefsIn(Function &F);
267 bool isBlockExecutable(BasicBlock *BB) const {
268 return BBExecutable.count(BB);
271 LatticeVal getLatticeValueFor(Value *V) const {
272 DenseMap<Value*, LatticeVal>::const_iterator I = ValueState.find(V);
273 assert(I != ValueState.end() && "V is not in valuemap!");
277 /// getTrackedRetVals - Get the inferred return value map.
279 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
280 return TrackedRetVals;
283 /// getTrackedGlobals - Get and return the set of inferred initializers for
284 /// global variables.
285 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
286 return TrackedGlobals;
289 void markOverdefined(Value *V) {
290 assert(!V->getType()->isStructTy() && "Should use other method");
291 markOverdefined(ValueState[V], V);
294 /// markAnythingOverdefined - Mark the specified value overdefined. This
295 /// works with both scalars and structs.
296 void markAnythingOverdefined(Value *V) {
297 if (StructType *STy = dyn_cast<StructType>(V->getType()))
298 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
299 markOverdefined(getStructValueState(V, i), V);
305 // markConstant - Make a value be marked as "constant". If the value
306 // is not already a constant, add it to the instruction work list so that
307 // the users of the instruction are updated later.
309 void markConstant(LatticeVal &IV, Value *V, Constant *C) {
310 if (!IV.markConstant(C)) return;
311 DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n');
312 if (IV.isOverdefined())
313 OverdefinedInstWorkList.push_back(V);
315 InstWorkList.push_back(V);
318 void markConstant(Value *V, Constant *C) {
319 assert(!V->getType()->isStructTy() && "Should use other method");
320 markConstant(ValueState[V], V, C);
323 void markForcedConstant(Value *V, Constant *C) {
324 assert(!V->getType()->isStructTy() && "Should use other method");
325 LatticeVal &IV = ValueState[V];
326 IV.markForcedConstant(C);
327 DEBUG(dbgs() << "markForcedConstant: " << *C << ": " << *V << '\n');
328 if (IV.isOverdefined())
329 OverdefinedInstWorkList.push_back(V);
331 InstWorkList.push_back(V);
335 // markOverdefined - Make a value be marked as "overdefined". If the
336 // value is not already overdefined, add it to the overdefined instruction
337 // work list so that the users of the instruction are updated later.
338 void markOverdefined(LatticeVal &IV, Value *V) {
339 if (!IV.markOverdefined()) return;
341 DEBUG(dbgs() << "markOverdefined: ";
342 if (Function *F = dyn_cast<Function>(V))
343 dbgs() << "Function '" << F->getName() << "'\n";
345 dbgs() << *V << '\n');
346 // Only instructions go on the work list
347 OverdefinedInstWorkList.push_back(V);
350 void mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) {
351 if (IV.isOverdefined() || MergeWithV.isUndefined())
353 if (MergeWithV.isOverdefined())
354 markOverdefined(IV, V);
355 else if (IV.isUndefined())
356 markConstant(IV, V, MergeWithV.getConstant());
357 else if (IV.getConstant() != MergeWithV.getConstant())
358 markOverdefined(IV, V);
361 void mergeInValue(Value *V, LatticeVal MergeWithV) {
362 assert(!V->getType()->isStructTy() && "Should use other method");
363 mergeInValue(ValueState[V], V, MergeWithV);
367 /// getValueState - Return the LatticeVal object that corresponds to the
368 /// value. This function handles the case when the value hasn't been seen yet
369 /// by properly seeding constants etc.
370 LatticeVal &getValueState(Value *V) {
371 assert(!V->getType()->isStructTy() && "Should use getStructValueState");
373 std::pair<DenseMap<Value*, LatticeVal>::iterator, bool> I =
374 ValueState.insert(std::make_pair(V, LatticeVal()));
375 LatticeVal &LV = I.first->second;
378 return LV; // Common case, already in the map.
380 if (Constant *C = dyn_cast<Constant>(V)) {
381 // Undef values remain undefined.
382 if (!isa<UndefValue>(V))
383 LV.markConstant(C); // Constants are constant
386 // All others are underdefined by default.
390 /// getStructValueState - Return the LatticeVal object that corresponds to the
391 /// value/field pair. This function handles the case when the value hasn't
392 /// been seen yet by properly seeding constants etc.
393 LatticeVal &getStructValueState(Value *V, unsigned i) {
394 assert(V->getType()->isStructTy() && "Should use getValueState");
395 assert(i < cast<StructType>(V->getType())->getNumElements() &&
396 "Invalid element #");
398 std::pair<DenseMap<std::pair<Value*, unsigned>, LatticeVal>::iterator,
399 bool> I = StructValueState.insert(
400 std::make_pair(std::make_pair(V, i), LatticeVal()));
401 LatticeVal &LV = I.first->second;
404 return LV; // Common case, already in the map.
406 if (Constant *C = dyn_cast<Constant>(V)) {
407 Constant *Elt = C->getAggregateElement(i);
410 LV.markOverdefined(); // Unknown sort of constant.
411 else if (isa<UndefValue>(Elt))
412 ; // Undef values remain undefined.
414 LV.markConstant(Elt); // Constants are constant.
417 // All others are underdefined by default.
422 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
423 /// work list if it is not already executable.
424 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
425 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
426 return; // This edge is already known to be executable!
428 if (!MarkBlockExecutable(Dest)) {
429 // If the destination is already executable, we just made an *edge*
430 // feasible that wasn't before. Revisit the PHI nodes in the block
431 // because they have potentially new operands.
432 DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName()
433 << " -> " << Dest->getName() << '\n');
436 for (BasicBlock::iterator I = Dest->begin();
437 (PN = dyn_cast<PHINode>(I)); ++I)
442 // getFeasibleSuccessors - Return a vector of booleans to indicate which
443 // successors are reachable from a given terminator instruction.
445 void getFeasibleSuccessors(TerminatorInst &TI, SmallVectorImpl<bool> &Succs);
447 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
448 // block to the 'To' basic block is currently feasible.
450 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
452 // OperandChangedState - This method is invoked on all of the users of an
453 // instruction that was just changed state somehow. Based on this
454 // information, we need to update the specified user of this instruction.
456 void OperandChangedState(Instruction *I) {
457 if (BBExecutable.count(I->getParent())) // Inst is executable?
462 friend class InstVisitor<SCCPSolver>;
464 // visit implementations - Something changed in this instruction. Either an
465 // operand made a transition, or the instruction is newly executable. Change
466 // the value type of I to reflect these changes if appropriate.
467 void visitPHINode(PHINode &I);
470 void visitReturnInst(ReturnInst &I);
471 void visitTerminatorInst(TerminatorInst &TI);
473 void visitCastInst(CastInst &I);
474 void visitSelectInst(SelectInst &I);
475 void visitBinaryOperator(Instruction &I);
476 void visitCmpInst(CmpInst &I);
477 void visitExtractElementInst(ExtractElementInst &I);
478 void visitInsertElementInst(InsertElementInst &I);
479 void visitShuffleVectorInst(ShuffleVectorInst &I);
480 void visitExtractValueInst(ExtractValueInst &EVI);
481 void visitInsertValueInst(InsertValueInst &IVI);
482 void visitLandingPadInst(LandingPadInst &I) { markAnythingOverdefined(&I); }
483 void visitCleanupPadInst(CleanupPadInst &CPI) { markAnythingOverdefined(&CPI); }
484 void visitCatchPadInst(CatchPadInst &CPI) {
485 markAnythingOverdefined(&CPI);
486 visitTerminatorInst(CPI);
489 // Instructions that cannot be folded away.
490 void visitStoreInst (StoreInst &I);
491 void visitLoadInst (LoadInst &I);
492 void visitGetElementPtrInst(GetElementPtrInst &I);
493 void visitCallInst (CallInst &I) {
496 void visitInvokeInst (InvokeInst &II) {
498 visitTerminatorInst(II);
500 void visitCallSite (CallSite CS);
501 void visitResumeInst (TerminatorInst &I) { /*returns void*/ }
502 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
503 void visitFenceInst (FenceInst &I) { /*returns void*/ }
504 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
505 markAnythingOverdefined(&I);
507 void visitAtomicRMWInst (AtomicRMWInst &I) { markOverdefined(&I); }
508 void visitAllocaInst (Instruction &I) { markOverdefined(&I); }
509 void visitVAArgInst (Instruction &I) { markAnythingOverdefined(&I); }
511 void visitInstruction(Instruction &I) {
512 // If a new instruction is added to LLVM that we don't handle.
513 dbgs() << "SCCP: Don't know how to handle: " << I << '\n';
514 markAnythingOverdefined(&I); // Just in case
518 } // end anonymous namespace
521 // getFeasibleSuccessors - Return a vector of booleans to indicate which
522 // successors are reachable from a given terminator instruction.
524 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
525 SmallVectorImpl<bool> &Succs) {
526 Succs.resize(TI.getNumSuccessors());
527 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
528 if (BI->isUnconditional()) {
533 LatticeVal BCValue = getValueState(BI->getCondition());
534 ConstantInt *CI = BCValue.getConstantInt();
536 // Overdefined condition variables, and branches on unfoldable constant
537 // conditions, mean the branch could go either way.
538 if (!BCValue.isUndefined())
539 Succs[0] = Succs[1] = true;
543 // Constant condition variables mean the branch can only go a single way.
544 Succs[CI->isZero()] = true;
548 // Unwinding instructions successors are always executable.
549 if (TI.isExceptional()) {
550 Succs.assign(TI.getNumSuccessors(), true);
554 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
555 if (!SI->getNumCases()) {
559 LatticeVal SCValue = getValueState(SI->getCondition());
560 ConstantInt *CI = SCValue.getConstantInt();
562 if (!CI) { // Overdefined or undefined condition?
563 // All destinations are executable!
564 if (!SCValue.isUndefined())
565 Succs.assign(TI.getNumSuccessors(), true);
569 Succs[SI->findCaseValue(CI).getSuccessorIndex()] = true;
573 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
574 if (isa<IndirectBrInst>(&TI)) {
575 // Just mark all destinations executable!
576 Succs.assign(TI.getNumSuccessors(), true);
581 dbgs() << "Unknown terminator instruction: " << TI << '\n';
583 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
587 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
588 // block to the 'To' basic block is currently feasible.
590 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
591 assert(BBExecutable.count(To) && "Dest should always be alive!");
593 // Make sure the source basic block is executable!!
594 if (!BBExecutable.count(From)) return false;
596 // Check to make sure this edge itself is actually feasible now.
597 TerminatorInst *TI = From->getTerminator();
598 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
599 if (BI->isUnconditional())
602 LatticeVal BCValue = getValueState(BI->getCondition());
604 // Overdefined condition variables mean the branch could go either way,
605 // undef conditions mean that neither edge is feasible yet.
606 ConstantInt *CI = BCValue.getConstantInt();
608 return !BCValue.isUndefined();
610 // Constant condition variables mean the branch can only go a single way.
611 return BI->getSuccessor(CI->isZero()) == To;
614 // Unwinding instructions successors are always executable.
615 if (TI->isExceptional())
618 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
619 if (SI->getNumCases() < 1)
622 LatticeVal SCValue = getValueState(SI->getCondition());
623 ConstantInt *CI = SCValue.getConstantInt();
626 return !SCValue.isUndefined();
628 return SI->findCaseValue(CI).getCaseSuccessor() == To;
631 // Just mark all destinations executable!
632 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
633 if (isa<IndirectBrInst>(TI))
637 dbgs() << "Unknown terminator instruction: " << *TI << '\n';
639 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
642 // visit Implementations - Something changed in this instruction, either an
643 // operand made a transition, or the instruction is newly executable. Change
644 // the value type of I to reflect these changes if appropriate. This method
645 // makes sure to do the following actions:
647 // 1. If a phi node merges two constants in, and has conflicting value coming
648 // from different branches, or if the PHI node merges in an overdefined
649 // value, then the PHI node becomes overdefined.
650 // 2. If a phi node merges only constants in, and they all agree on value, the
651 // PHI node becomes a constant value equal to that.
652 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
653 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
654 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
655 // 6. If a conditional branch has a value that is constant, make the selected
656 // destination executable
657 // 7. If a conditional branch has a value that is overdefined, make all
658 // successors executable.
660 void SCCPSolver::visitPHINode(PHINode &PN) {
661 // If this PN returns a struct, just mark the result overdefined.
662 // TODO: We could do a lot better than this if code actually uses this.
663 if (PN.getType()->isStructTy())
664 return markAnythingOverdefined(&PN);
666 if (getValueState(&PN).isOverdefined())
667 return; // Quick exit
669 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
670 // and slow us down a lot. Just mark them overdefined.
671 if (PN.getNumIncomingValues() > 64)
672 return markOverdefined(&PN);
674 // Look at all of the executable operands of the PHI node. If any of them
675 // are overdefined, the PHI becomes overdefined as well. If they are all
676 // constant, and they agree with each other, the PHI becomes the identical
677 // constant. If they are constant and don't agree, the PHI is overdefined.
678 // If there are no executable operands, the PHI remains undefined.
680 Constant *OperandVal = nullptr;
681 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
682 LatticeVal IV = getValueState(PN.getIncomingValue(i));
683 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
685 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
688 if (IV.isOverdefined()) // PHI node becomes overdefined!
689 return markOverdefined(&PN);
691 if (!OperandVal) { // Grab the first value.
692 OperandVal = IV.getConstant();
696 // There is already a reachable operand. If we conflict with it,
697 // then the PHI node becomes overdefined. If we agree with it, we
700 // Check to see if there are two different constants merging, if so, the PHI
701 // node is overdefined.
702 if (IV.getConstant() != OperandVal)
703 return markOverdefined(&PN);
706 // If we exited the loop, this means that the PHI node only has constant
707 // arguments that agree with each other(and OperandVal is the constant) or
708 // OperandVal is null because there are no defined incoming arguments. If
709 // this is the case, the PHI remains undefined.
712 markConstant(&PN, OperandVal); // Acquire operand value
715 void SCCPSolver::visitReturnInst(ReturnInst &I) {
716 if (I.getNumOperands() == 0) return; // ret void
718 Function *F = I.getParent()->getParent();
719 Value *ResultOp = I.getOperand(0);
721 // If we are tracking the return value of this function, merge it in.
722 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) {
723 DenseMap<Function*, LatticeVal>::iterator TFRVI =
724 TrackedRetVals.find(F);
725 if (TFRVI != TrackedRetVals.end()) {
726 mergeInValue(TFRVI->second, F, getValueState(ResultOp));
731 // Handle functions that return multiple values.
732 if (!TrackedMultipleRetVals.empty()) {
733 if (StructType *STy = dyn_cast<StructType>(ResultOp->getType()))
734 if (MRVFunctionsTracked.count(F))
735 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
736 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F,
737 getStructValueState(ResultOp, i));
742 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
743 SmallVector<bool, 16> SuccFeasible;
744 getFeasibleSuccessors(TI, SuccFeasible);
746 BasicBlock *BB = TI.getParent();
748 // Mark all feasible successors executable.
749 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
751 markEdgeExecutable(BB, TI.getSuccessor(i));
754 void SCCPSolver::visitCastInst(CastInst &I) {
755 LatticeVal OpSt = getValueState(I.getOperand(0));
756 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand
758 else if (OpSt.isConstant()) // Propagate constant value
759 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
760 OpSt.getConstant(), I.getType()));
764 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
765 // If this returns a struct, mark all elements over defined, we don't track
766 // structs in structs.
767 if (EVI.getType()->isStructTy())
768 return markAnythingOverdefined(&EVI);
770 // If this is extracting from more than one level of struct, we don't know.
771 if (EVI.getNumIndices() != 1)
772 return markOverdefined(&EVI);
774 Value *AggVal = EVI.getAggregateOperand();
775 if (AggVal->getType()->isStructTy()) {
776 unsigned i = *EVI.idx_begin();
777 LatticeVal EltVal = getStructValueState(AggVal, i);
778 mergeInValue(getValueState(&EVI), &EVI, EltVal);
780 // Otherwise, must be extracting from an array.
781 return markOverdefined(&EVI);
785 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
786 StructType *STy = dyn_cast<StructType>(IVI.getType());
788 return markOverdefined(&IVI);
790 // If this has more than one index, we can't handle it, drive all results to
792 if (IVI.getNumIndices() != 1)
793 return markAnythingOverdefined(&IVI);
795 Value *Aggr = IVI.getAggregateOperand();
796 unsigned Idx = *IVI.idx_begin();
798 // Compute the result based on what we're inserting.
799 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
800 // This passes through all values that aren't the inserted element.
802 LatticeVal EltVal = getStructValueState(Aggr, i);
803 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal);
807 Value *Val = IVI.getInsertedValueOperand();
808 if (Val->getType()->isStructTy())
809 // We don't track structs in structs.
810 markOverdefined(getStructValueState(&IVI, i), &IVI);
812 LatticeVal InVal = getValueState(Val);
813 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal);
818 void SCCPSolver::visitSelectInst(SelectInst &I) {
819 // If this select returns a struct, just mark the result overdefined.
820 // TODO: We could do a lot better than this if code actually uses this.
821 if (I.getType()->isStructTy())
822 return markAnythingOverdefined(&I);
824 LatticeVal CondValue = getValueState(I.getCondition());
825 if (CondValue.isUndefined())
828 if (ConstantInt *CondCB = CondValue.getConstantInt()) {
829 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
830 mergeInValue(&I, getValueState(OpVal));
834 // Otherwise, the condition is overdefined or a constant we can't evaluate.
835 // See if we can produce something better than overdefined based on the T/F
837 LatticeVal TVal = getValueState(I.getTrueValue());
838 LatticeVal FVal = getValueState(I.getFalseValue());
840 // select ?, C, C -> C.
841 if (TVal.isConstant() && FVal.isConstant() &&
842 TVal.getConstant() == FVal.getConstant())
843 return markConstant(&I, FVal.getConstant());
845 if (TVal.isUndefined()) // select ?, undef, X -> X.
846 return mergeInValue(&I, FVal);
847 if (FVal.isUndefined()) // select ?, X, undef -> X.
848 return mergeInValue(&I, TVal);
852 // Handle Binary Operators.
853 void SCCPSolver::visitBinaryOperator(Instruction &I) {
854 LatticeVal V1State = getValueState(I.getOperand(0));
855 LatticeVal V2State = getValueState(I.getOperand(1));
857 LatticeVal &IV = ValueState[&I];
858 if (IV.isOverdefined()) return;
860 if (V1State.isConstant() && V2State.isConstant())
861 return markConstant(IV, &I,
862 ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
863 V2State.getConstant()));
865 // If something is undef, wait for it to resolve.
866 if (!V1State.isOverdefined() && !V2State.isOverdefined())
869 // Otherwise, one of our operands is overdefined. Try to produce something
870 // better than overdefined with some tricks.
872 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
873 // operand is overdefined.
874 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
875 LatticeVal *NonOverdefVal = nullptr;
876 if (!V1State.isOverdefined())
877 NonOverdefVal = &V1State;
878 else if (!V2State.isOverdefined())
879 NonOverdefVal = &V2State;
882 if (NonOverdefVal->isUndefined()) {
883 // Could annihilate value.
884 if (I.getOpcode() == Instruction::And)
885 markConstant(IV, &I, Constant::getNullValue(I.getType()));
886 else if (VectorType *PT = dyn_cast<VectorType>(I.getType()))
887 markConstant(IV, &I, Constant::getAllOnesValue(PT));
890 Constant::getAllOnesValue(I.getType()));
894 if (I.getOpcode() == Instruction::And) {
896 if (NonOverdefVal->getConstant()->isNullValue())
897 return markConstant(IV, &I, NonOverdefVal->getConstant());
899 if (ConstantInt *CI = NonOverdefVal->getConstantInt())
900 if (CI->isAllOnesValue()) // X or -1 = -1
901 return markConstant(IV, &I, NonOverdefVal->getConstant());
910 // Handle ICmpInst instruction.
911 void SCCPSolver::visitCmpInst(CmpInst &I) {
912 LatticeVal V1State = getValueState(I.getOperand(0));
913 LatticeVal V2State = getValueState(I.getOperand(1));
915 LatticeVal &IV = ValueState[&I];
916 if (IV.isOverdefined()) return;
918 if (V1State.isConstant() && V2State.isConstant())
919 return markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
920 V1State.getConstant(),
921 V2State.getConstant()));
923 // If operands are still undefined, wait for it to resolve.
924 if (!V1State.isOverdefined() && !V2State.isOverdefined())
930 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
931 // TODO : SCCP does not handle vectors properly.
932 return markOverdefined(&I);
935 LatticeVal &ValState = getValueState(I.getOperand(0));
936 LatticeVal &IdxState = getValueState(I.getOperand(1));
938 if (ValState.isOverdefined() || IdxState.isOverdefined())
940 else if(ValState.isConstant() && IdxState.isConstant())
941 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
942 IdxState.getConstant()));
946 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
947 // TODO : SCCP does not handle vectors properly.
948 return markOverdefined(&I);
950 LatticeVal &ValState = getValueState(I.getOperand(0));
951 LatticeVal &EltState = getValueState(I.getOperand(1));
952 LatticeVal &IdxState = getValueState(I.getOperand(2));
954 if (ValState.isOverdefined() || EltState.isOverdefined() ||
955 IdxState.isOverdefined())
957 else if(ValState.isConstant() && EltState.isConstant() &&
958 IdxState.isConstant())
959 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
960 EltState.getConstant(),
961 IdxState.getConstant()));
962 else if (ValState.isUndefined() && EltState.isConstant() &&
963 IdxState.isConstant())
964 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
965 EltState.getConstant(),
966 IdxState.getConstant()));
970 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
971 // TODO : SCCP does not handle vectors properly.
972 return markOverdefined(&I);
974 LatticeVal &V1State = getValueState(I.getOperand(0));
975 LatticeVal &V2State = getValueState(I.getOperand(1));
976 LatticeVal &MaskState = getValueState(I.getOperand(2));
978 if (MaskState.isUndefined() ||
979 (V1State.isUndefined() && V2State.isUndefined()))
980 return; // Undefined output if mask or both inputs undefined.
982 if (V1State.isOverdefined() || V2State.isOverdefined() ||
983 MaskState.isOverdefined()) {
986 // A mix of constant/undef inputs.
987 Constant *V1 = V1State.isConstant() ?
988 V1State.getConstant() : UndefValue::get(I.getType());
989 Constant *V2 = V2State.isConstant() ?
990 V2State.getConstant() : UndefValue::get(I.getType());
991 Constant *Mask = MaskState.isConstant() ?
992 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
993 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
998 // Handle getelementptr instructions. If all operands are constants then we
999 // can turn this into a getelementptr ConstantExpr.
1001 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1002 if (ValueState[&I].isOverdefined()) return;
1004 SmallVector<Constant*, 8> Operands;
1005 Operands.reserve(I.getNumOperands());
1007 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1008 LatticeVal State = getValueState(I.getOperand(i));
1009 if (State.isUndefined())
1010 return; // Operands are not resolved yet.
1012 if (State.isOverdefined())
1013 return markOverdefined(&I);
1015 assert(State.isConstant() && "Unknown state!");
1016 Operands.push_back(State.getConstant());
1019 Constant *Ptr = Operands[0];
1020 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end());
1021 markConstant(&I, ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr,
1025 void SCCPSolver::visitStoreInst(StoreInst &SI) {
1026 // If this store is of a struct, ignore it.
1027 if (SI.getOperand(0)->getType()->isStructTy())
1030 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1033 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1034 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1035 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1037 // Get the value we are storing into the global, then merge it.
1038 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)));
1039 if (I->second.isOverdefined())
1040 TrackedGlobals.erase(I); // No need to keep tracking this!
1044 // Handle load instructions. If the operand is a constant pointer to a constant
1045 // global, we can replace the load with the loaded constant value!
1046 void SCCPSolver::visitLoadInst(LoadInst &I) {
1047 // If this load is of a struct, just mark the result overdefined.
1048 if (I.getType()->isStructTy())
1049 return markAnythingOverdefined(&I);
1051 LatticeVal PtrVal = getValueState(I.getOperand(0));
1052 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1054 LatticeVal &IV = ValueState[&I];
1055 if (IV.isOverdefined()) return;
1057 if (!PtrVal.isConstant() || I.isVolatile())
1058 return markOverdefined(IV, &I);
1060 Constant *Ptr = PtrVal.getConstant();
1062 // load null -> null
1063 if (isa<ConstantPointerNull>(Ptr) && I.getPointerAddressSpace() == 0)
1064 return markConstant(IV, &I, UndefValue::get(I.getType()));
1066 // Transform load (constant global) into the value loaded.
1067 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1068 if (!TrackedGlobals.empty()) {
1069 // If we are tracking this global, merge in the known value for it.
1070 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1071 TrackedGlobals.find(GV);
1072 if (It != TrackedGlobals.end()) {
1073 mergeInValue(IV, &I, It->second);
1079 // Transform load from a constant into a constant if possible.
1080 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, DL))
1081 return markConstant(IV, &I, C);
1083 // Otherwise we cannot say for certain what value this load will produce.
1085 markOverdefined(IV, &I);
1088 void SCCPSolver::visitCallSite(CallSite CS) {
1089 Function *F = CS.getCalledFunction();
1090 Instruction *I = CS.getInstruction();
1092 // The common case is that we aren't tracking the callee, either because we
1093 // are not doing interprocedural analysis or the callee is indirect, or is
1094 // external. Handle these cases first.
1095 if (!F || F->isDeclaration()) {
1097 // Void return and not tracking callee, just bail.
1098 if (I->getType()->isVoidTy()) return;
1100 // Otherwise, if we have a single return value case, and if the function is
1101 // a declaration, maybe we can constant fold it.
1102 if (F && F->isDeclaration() && !I->getType()->isStructTy() &&
1103 canConstantFoldCallTo(F)) {
1105 SmallVector<Constant*, 8> Operands;
1106 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1108 LatticeVal State = getValueState(*AI);
1110 if (State.isUndefined())
1111 return; // Operands are not resolved yet.
1112 if (State.isOverdefined())
1113 return markOverdefined(I);
1114 assert(State.isConstant() && "Unknown state!");
1115 Operands.push_back(State.getConstant());
1118 if (getValueState(I).isOverdefined())
1121 // If we can constant fold this, mark the result of the call as a
1123 if (Constant *C = ConstantFoldCall(F, Operands, TLI))
1124 return markConstant(I, C);
1127 // Otherwise, we don't know anything about this call, mark it overdefined.
1128 return markAnythingOverdefined(I);
1131 // If this is a local function that doesn't have its address taken, mark its
1132 // entry block executable and merge in the actual arguments to the call into
1133 // the formal arguments of the function.
1134 if (!TrackingIncomingArguments.empty() && TrackingIncomingArguments.count(F)){
1135 MarkBlockExecutable(&F->front());
1137 // Propagate information from this call site into the callee.
1138 CallSite::arg_iterator CAI = CS.arg_begin();
1139 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1140 AI != E; ++AI, ++CAI) {
1141 // If this argument is byval, and if the function is not readonly, there
1142 // will be an implicit copy formed of the input aggregate.
1143 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1144 markOverdefined(&*AI);
1148 if (StructType *STy = dyn_cast<StructType>(AI->getType())) {
1149 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1150 LatticeVal CallArg = getStructValueState(*CAI, i);
1151 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg);
1154 mergeInValue(&*AI, getValueState(*CAI));
1159 // If this is a single/zero retval case, see if we're tracking the function.
1160 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
1161 if (!MRVFunctionsTracked.count(F))
1162 goto CallOverdefined; // Not tracking this callee.
1164 // If we are tracking this callee, propagate the result of the function
1165 // into this call site.
1166 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1167 mergeInValue(getStructValueState(I, i), I,
1168 TrackedMultipleRetVals[std::make_pair(F, i)]);
1170 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1171 if (TFRVI == TrackedRetVals.end())
1172 goto CallOverdefined; // Not tracking this callee.
1174 // If so, propagate the return value of the callee into this call result.
1175 mergeInValue(I, TFRVI->second);
1179 void SCCPSolver::Solve() {
1180 // Process the work lists until they are empty!
1181 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1182 !OverdefinedInstWorkList.empty()) {
1183 // Process the overdefined instruction's work list first, which drives other
1184 // things to overdefined more quickly.
1185 while (!OverdefinedInstWorkList.empty()) {
1186 Value *I = OverdefinedInstWorkList.pop_back_val();
1188 DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n');
1190 // "I" got into the work list because it either made the transition from
1191 // bottom to constant, or to overdefined.
1193 // Anything on this worklist that is overdefined need not be visited
1194 // since all of its users will have already been marked as overdefined
1195 // Update all of the users of this instruction's value.
1197 for (User *U : I->users())
1198 if (Instruction *UI = dyn_cast<Instruction>(U))
1199 OperandChangedState(UI);
1202 // Process the instruction work list.
1203 while (!InstWorkList.empty()) {
1204 Value *I = InstWorkList.pop_back_val();
1206 DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n');
1208 // "I" got into the work list because it made the transition from undef to
1211 // Anything on this worklist that is overdefined need not be visited
1212 // since all of its users will have already been marked as overdefined.
1213 // Update all of the users of this instruction's value.
1215 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined())
1216 for (User *U : I->users())
1217 if (Instruction *UI = dyn_cast<Instruction>(U))
1218 OperandChangedState(UI);
1221 // Process the basic block work list.
1222 while (!BBWorkList.empty()) {
1223 BasicBlock *BB = BBWorkList.back();
1224 BBWorkList.pop_back();
1226 DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n');
1228 // Notify all instructions in this basic block that they are newly
1235 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1236 /// that branches on undef values cannot reach any of their successors.
1237 /// However, this is not a safe assumption. After we solve dataflow, this
1238 /// method should be use to handle this. If this returns true, the solver
1239 /// should be rerun.
1241 /// This method handles this by finding an unresolved branch and marking it one
1242 /// of the edges from the block as being feasible, even though the condition
1243 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1244 /// CFG and only slightly pessimizes the analysis results (by marking one,
1245 /// potentially infeasible, edge feasible). This cannot usefully modify the
1246 /// constraints on the condition of the branch, as that would impact other users
1249 /// This scan also checks for values that use undefs, whose results are actually
1250 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1251 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1252 /// even if X isn't defined.
1253 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1254 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1255 if (!BBExecutable.count(&*BB))
1258 for (Instruction &I : *BB) {
1259 // Look for instructions which produce undef values.
1260 if (I.getType()->isVoidTy()) continue;
1262 if (StructType *STy = dyn_cast<StructType>(I.getType())) {
1263 // Only a few things that can be structs matter for undef.
1265 // Tracked calls must never be marked overdefined in ResolvedUndefsIn.
1266 if (CallSite CS = CallSite(&I))
1267 if (Function *F = CS.getCalledFunction())
1268 if (MRVFunctionsTracked.count(F))
1271 // extractvalue and insertvalue don't need to be marked; they are
1272 // tracked as precisely as their operands.
1273 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I))
1276 // Send the results of everything else to overdefined. We could be
1277 // more precise than this but it isn't worth bothering.
1278 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1279 LatticeVal &LV = getStructValueState(&I, i);
1280 if (LV.isUndefined())
1281 markOverdefined(LV, &I);
1286 LatticeVal &LV = getValueState(&I);
1287 if (!LV.isUndefined()) continue;
1289 // extractvalue is safe; check here because the argument is a struct.
1290 if (isa<ExtractValueInst>(I))
1293 // Compute the operand LatticeVals, for convenience below.
1294 // Anything taking a struct is conservatively assumed to require
1295 // overdefined markings.
1296 if (I.getOperand(0)->getType()->isStructTy()) {
1297 markOverdefined(&I);
1300 LatticeVal Op0LV = getValueState(I.getOperand(0));
1302 if (I.getNumOperands() == 2) {
1303 if (I.getOperand(1)->getType()->isStructTy()) {
1304 markOverdefined(&I);
1308 Op1LV = getValueState(I.getOperand(1));
1310 // If this is an instructions whose result is defined even if the input is
1311 // not fully defined, propagate the information.
1312 Type *ITy = I.getType();
1313 switch (I.getOpcode()) {
1314 case Instruction::Add:
1315 case Instruction::Sub:
1316 case Instruction::Trunc:
1317 case Instruction::FPTrunc:
1318 case Instruction::BitCast:
1319 break; // Any undef -> undef
1320 case Instruction::FSub:
1321 case Instruction::FAdd:
1322 case Instruction::FMul:
1323 case Instruction::FDiv:
1324 case Instruction::FRem:
1325 // Floating-point binary operation: be conservative.
1326 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1327 markForcedConstant(&I, Constant::getNullValue(ITy));
1329 markOverdefined(&I);
1331 case Instruction::ZExt:
1332 case Instruction::SExt:
1333 case Instruction::FPToUI:
1334 case Instruction::FPToSI:
1335 case Instruction::FPExt:
1336 case Instruction::PtrToInt:
1337 case Instruction::IntToPtr:
1338 case Instruction::SIToFP:
1339 case Instruction::UIToFP:
1340 // undef -> 0; some outputs are impossible
1341 markForcedConstant(&I, Constant::getNullValue(ITy));
1343 case Instruction::Mul:
1344 case Instruction::And:
1345 // Both operands undef -> undef
1346 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1348 // undef * X -> 0. X could be zero.
1349 // undef & X -> 0. X could be zero.
1350 markForcedConstant(&I, Constant::getNullValue(ITy));
1353 case Instruction::Or:
1354 // Both operands undef -> undef
1355 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1357 // undef | X -> -1. X could be -1.
1358 markForcedConstant(&I, Constant::getAllOnesValue(ITy));
1361 case Instruction::Xor:
1362 // undef ^ undef -> 0; strictly speaking, this is not strictly
1363 // necessary, but we try to be nice to people who expect this
1364 // behavior in simple cases
1365 if (Op0LV.isUndefined() && Op1LV.isUndefined()) {
1366 markForcedConstant(&I, Constant::getNullValue(ITy));
1369 // undef ^ X -> undef
1372 case Instruction::SDiv:
1373 case Instruction::UDiv:
1374 case Instruction::SRem:
1375 case Instruction::URem:
1376 // X / undef -> undef. No change.
1377 // X % undef -> undef. No change.
1378 if (Op1LV.isUndefined()) break;
1380 // undef / X -> 0. X could be maxint.
1381 // undef % X -> 0. X could be 1.
1382 markForcedConstant(&I, Constant::getNullValue(ITy));
1385 case Instruction::AShr:
1386 // X >>a undef -> undef.
1387 if (Op1LV.isUndefined()) break;
1389 // undef >>a X -> all ones
1390 markForcedConstant(&I, Constant::getAllOnesValue(ITy));
1392 case Instruction::LShr:
1393 case Instruction::Shl:
1394 // X << undef -> undef.
1395 // X >> undef -> undef.
1396 if (Op1LV.isUndefined()) break;
1400 markForcedConstant(&I, Constant::getNullValue(ITy));
1402 case Instruction::Select:
1403 Op1LV = getValueState(I.getOperand(1));
1404 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1405 if (Op0LV.isUndefined()) {
1406 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1407 Op1LV = getValueState(I.getOperand(2));
1408 } else if (Op1LV.isUndefined()) {
1409 // c ? undef : undef -> undef. No change.
1410 Op1LV = getValueState(I.getOperand(2));
1411 if (Op1LV.isUndefined())
1413 // Otherwise, c ? undef : x -> x.
1415 // Leave Op1LV as Operand(1)'s LatticeValue.
1418 if (Op1LV.isConstant())
1419 markForcedConstant(&I, Op1LV.getConstant());
1421 markOverdefined(&I);
1423 case Instruction::Load:
1424 // A load here means one of two things: a load of undef from a global,
1425 // a load from an unknown pointer. Either way, having it return undef
1428 case Instruction::ICmp:
1429 // X == undef -> undef. Other comparisons get more complicated.
1430 if (cast<ICmpInst>(&I)->isEquality())
1432 markOverdefined(&I);
1434 case Instruction::Call:
1435 case Instruction::Invoke: {
1436 // There are two reasons a call can have an undef result
1437 // 1. It could be tracked.
1438 // 2. It could be constant-foldable.
1439 // Because of the way we solve return values, tracked calls must
1440 // never be marked overdefined in ResolvedUndefsIn.
1441 if (Function *F = CallSite(&I).getCalledFunction())
1442 if (TrackedRetVals.count(F))
1445 // If the call is constant-foldable, we mark it overdefined because
1446 // we do not know what return values are valid.
1447 markOverdefined(&I);
1451 // If we don't know what should happen here, conservatively mark it
1453 markOverdefined(&I);
1458 // Check to see if we have a branch or switch on an undefined value. If so
1459 // we force the branch to go one way or the other to make the successor
1460 // values live. It doesn't really matter which way we force it.
1461 TerminatorInst *TI = BB->getTerminator();
1462 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1463 if (!BI->isConditional()) continue;
1464 if (!getValueState(BI->getCondition()).isUndefined())
1467 // If the input to SCCP is actually branch on undef, fix the undef to
1469 if (isa<UndefValue>(BI->getCondition())) {
1470 BI->setCondition(ConstantInt::getFalse(BI->getContext()));
1471 markEdgeExecutable(&*BB, TI->getSuccessor(1));
1475 // Otherwise, it is a branch on a symbolic value which is currently
1476 // considered to be undef. Handle this by forcing the input value to the
1478 markForcedConstant(BI->getCondition(),
1479 ConstantInt::getFalse(TI->getContext()));
1483 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1484 if (!SI->getNumCases())
1486 if (!getValueState(SI->getCondition()).isUndefined())
1489 // If the input to SCCP is actually switch on undef, fix the undef to
1490 // the first constant.
1491 if (isa<UndefValue>(SI->getCondition())) {
1492 SI->setCondition(SI->case_begin().getCaseValue());
1493 markEdgeExecutable(&*BB, SI->case_begin().getCaseSuccessor());
1497 markForcedConstant(SI->getCondition(), SI->case_begin().getCaseValue());
1507 //===--------------------------------------------------------------------===//
1509 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1510 /// Sparse Conditional Constant Propagator.
1512 struct SCCP : public FunctionPass {
1513 void getAnalysisUsage(AnalysisUsage &AU) const override {
1514 AU.addRequired<TargetLibraryInfoWrapperPass>();
1515 AU.addPreserved<GlobalsAAWrapperPass>();
1517 static char ID; // Pass identification, replacement for typeid
1518 SCCP() : FunctionPass(ID) {
1519 initializeSCCPPass(*PassRegistry::getPassRegistry());
1522 // runOnFunction - Run the Sparse Conditional Constant Propagation
1523 // algorithm, and return true if the function was modified.
1525 bool runOnFunction(Function &F) override;
1527 } // end anonymous namespace
1530 INITIALIZE_PASS(SCCP, "sccp",
1531 "Sparse Conditional Constant Propagation", false, false)
1533 // createSCCPPass - This is the public interface to this file.
1534 FunctionPass *llvm::createSCCPPass() {
1538 static void DeleteInstructionInBlock(BasicBlock *BB) {
1539 DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
1542 // Check to see if there are non-terminating instructions to delete.
1543 if (isa<TerminatorInst>(BB->begin()))
1546 // Delete the instructions backwards, as it has a reduced likelihood of having
1547 // to update as many def-use and use-def chains.
1548 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
1549 while (EndInst != BB->begin()) {
1550 // Delete the next to last instruction.
1551 Instruction *Inst = &*--EndInst->getIterator();
1552 if (!Inst->use_empty())
1553 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
1554 if (Inst->isEHPad()) {
1558 BB->getInstList().erase(Inst);
1563 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1564 // and return true if the function was modified.
1566 bool SCCP::runOnFunction(Function &F) {
1567 if (skipOptnoneFunction(F))
1570 DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n");
1571 const DataLayout &DL = F.getParent()->getDataLayout();
1572 const TargetLibraryInfo *TLI =
1573 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1574 SCCPSolver Solver(DL, TLI);
1576 // Mark the first block of the function as being executable.
1577 Solver.MarkBlockExecutable(&F.front());
1579 // Mark all arguments to the function as being overdefined.
1580 for (Argument &AI : F.args())
1581 Solver.markAnythingOverdefined(&AI);
1583 // Solve for constants.
1584 bool ResolvedUndefs = true;
1585 while (ResolvedUndefs) {
1587 DEBUG(dbgs() << "RESOLVING UNDEFs\n");
1588 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1591 bool MadeChanges = false;
1593 // If we decided that there are basic blocks that are dead in this function,
1594 // delete their contents now. Note that we cannot actually delete the blocks,
1595 // as we cannot modify the CFG of the function.
1597 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1598 if (!Solver.isBlockExecutable(&*BB)) {
1599 DeleteInstructionInBlock(&*BB);
1604 // Iterate over all of the instructions in a function, replacing them with
1605 // constants if we have found them to be of constant values.
1607 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1608 Instruction *Inst = &*BI++;
1609 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst))
1612 // TODO: Reconstruct structs from their elements.
1613 if (Inst->getType()->isStructTy())
1616 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1617 if (IV.isOverdefined())
1620 Constant *Const = IV.isConstant()
1621 ? IV.getConstant() : UndefValue::get(Inst->getType());
1622 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst << '\n');
1624 // Replaces all of the uses of a variable with uses of the constant.
1625 Inst->replaceAllUsesWith(Const);
1627 // Delete the instruction.
1628 Inst->eraseFromParent();
1630 // Hey, we just changed something!
1640 //===--------------------------------------------------------------------===//
1642 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1643 /// Constant Propagation.
1645 struct IPSCCP : public ModulePass {
1646 void getAnalysisUsage(AnalysisUsage &AU) const override {
1647 AU.addRequired<TargetLibraryInfoWrapperPass>();
1650 IPSCCP() : ModulePass(ID) {
1651 initializeIPSCCPPass(*PassRegistry::getPassRegistry());
1653 bool runOnModule(Module &M) override;
1655 } // end anonymous namespace
1657 char IPSCCP::ID = 0;
1658 INITIALIZE_PASS_BEGIN(IPSCCP, "ipsccp",
1659 "Interprocedural Sparse Conditional Constant Propagation",
1661 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1662 INITIALIZE_PASS_END(IPSCCP, "ipsccp",
1663 "Interprocedural Sparse Conditional Constant Propagation",
1666 // createIPSCCPPass - This is the public interface to this file.
1667 ModulePass *llvm::createIPSCCPPass() {
1668 return new IPSCCP();
1672 static bool AddressIsTaken(const GlobalValue *GV) {
1673 // Delete any dead constantexpr klingons.
1674 GV->removeDeadConstantUsers();
1676 for (const Use &U : GV->uses()) {
1677 const User *UR = U.getUser();
1678 if (const StoreInst *SI = dyn_cast<StoreInst>(UR)) {
1679 if (SI->getOperand(0) == GV || SI->isVolatile())
1680 return true; // Storing addr of GV.
1681 } else if (isa<InvokeInst>(UR) || isa<CallInst>(UR)) {
1682 // Make sure we are calling the function, not passing the address.
1683 ImmutableCallSite CS(cast<Instruction>(UR));
1684 if (!CS.isCallee(&U))
1686 } else if (const LoadInst *LI = dyn_cast<LoadInst>(UR)) {
1687 if (LI->isVolatile())
1689 } else if (isa<BlockAddress>(UR)) {
1690 // blockaddress doesn't take the address of the function, it takes addr
1699 bool IPSCCP::runOnModule(Module &M) {
1700 const DataLayout &DL = M.getDataLayout();
1701 const TargetLibraryInfo *TLI =
1702 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1703 SCCPSolver Solver(DL, TLI);
1705 // AddressTakenFunctions - This set keeps track of the address-taken functions
1706 // that are in the input. As IPSCCP runs through and simplifies code,
1707 // functions that were address taken can end up losing their
1708 // address-taken-ness. Because of this, we keep track of their addresses from
1709 // the first pass so we can use them for the later simplification pass.
1710 SmallPtrSet<Function*, 32> AddressTakenFunctions;
1712 // Loop over all functions, marking arguments to those with their addresses
1713 // taken or that are external as overdefined.
1715 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1716 if (F->isDeclaration())
1719 // If this is a strong or ODR definition of this function, then we can
1720 // propagate information about its result into callsites of it.
1721 if (!F->mayBeOverridden())
1722 Solver.AddTrackedFunction(&*F);
1724 // If this function only has direct calls that we can see, we can track its
1725 // arguments and return value aggressively, and can assume it is not called
1726 // unless we see evidence to the contrary.
1727 if (F->hasLocalLinkage()) {
1728 if (AddressIsTaken(&*F))
1729 AddressTakenFunctions.insert(&*F);
1731 Solver.AddArgumentTrackedFunction(&*F);
1736 // Assume the function is called.
1737 Solver.MarkBlockExecutable(&F->front());
1739 // Assume nothing about the incoming arguments.
1740 for (Argument &AI : F->args())
1741 Solver.markAnythingOverdefined(&AI);
1744 // Loop over global variables. We inform the solver about any internal global
1745 // variables that do not have their 'addresses taken'. If they don't have
1746 // their addresses taken, we can propagate constants through them.
1747 for (GlobalVariable &G : M.globals())
1748 if (!G.isConstant() && G.hasLocalLinkage() && !AddressIsTaken(&G))
1749 Solver.TrackValueOfGlobalVariable(&G);
1751 // Solve for constants.
1752 bool ResolvedUndefs = true;
1753 while (ResolvedUndefs) {
1756 DEBUG(dbgs() << "RESOLVING UNDEFS\n");
1757 ResolvedUndefs = false;
1758 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1759 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1762 bool MadeChanges = false;
1764 // Iterate over all of the instructions in the module, replacing them with
1765 // constants if we have found them to be of constant values.
1767 SmallVector<BasicBlock*, 512> BlocksToErase;
1769 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1770 if (F->isDeclaration())
1773 if (Solver.isBlockExecutable(&F->front())) {
1774 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1776 if (AI->use_empty() || AI->getType()->isStructTy()) continue;
1778 // TODO: Could use getStructLatticeValueFor to find out if the entire
1779 // result is a constant and replace it entirely if so.
1781 LatticeVal IV = Solver.getLatticeValueFor(&*AI);
1782 if (IV.isOverdefined()) continue;
1784 Constant *CST = IV.isConstant() ?
1785 IV.getConstant() : UndefValue::get(AI->getType());
1786 DEBUG(dbgs() << "*** Arg " << *AI << " = " << *CST <<"\n");
1788 // Replaces all of the uses of a variable with uses of the
1790 AI->replaceAllUsesWith(CST);
1795 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1796 if (!Solver.isBlockExecutable(&*BB)) {
1797 DeleteInstructionInBlock(&*BB);
1800 TerminatorInst *TI = BB->getTerminator();
1801 for (BasicBlock *Succ : TI->successors()) {
1802 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1803 Succ->removePredecessor(&*BB);
1805 if (!TI->use_empty())
1806 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1807 TI->eraseFromParent();
1808 new UnreachableInst(M.getContext(), &*BB);
1810 if (&*BB != &F->front())
1811 BlocksToErase.push_back(&*BB);
1815 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1816 Instruction *Inst = &*BI++;
1817 if (Inst->getType()->isVoidTy() || Inst->getType()->isStructTy())
1820 // TODO: Could use getStructLatticeValueFor to find out if the entire
1821 // result is a constant and replace it entirely if so.
1823 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1824 if (IV.isOverdefined())
1827 Constant *Const = IV.isConstant()
1828 ? IV.getConstant() : UndefValue::get(Inst->getType());
1829 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst << '\n');
1831 // Replaces all of the uses of a variable with uses of the
1833 Inst->replaceAllUsesWith(Const);
1835 // Delete the instruction.
1836 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst))
1837 Inst->eraseFromParent();
1839 // Hey, we just changed something!
1845 // Now that all instructions in the function are constant folded, erase dead
1846 // blocks, because we can now use ConstantFoldTerminator to get rid of
1848 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1849 // If there are any PHI nodes in this successor, drop entries for BB now.
1850 BasicBlock *DeadBB = BlocksToErase[i];
1851 for (Value::user_iterator UI = DeadBB->user_begin(),
1852 UE = DeadBB->user_end();
1854 // Grab the user and then increment the iterator early, as the user
1855 // will be deleted. Step past all adjacent uses from the same user.
1856 Instruction *I = dyn_cast<Instruction>(*UI);
1857 do { ++UI; } while (UI != UE && *UI == I);
1859 // Ignore blockaddress users; BasicBlock's dtor will handle them.
1862 bool Folded = ConstantFoldTerminator(I->getParent());
1864 // The constant folder may not have been able to fold the terminator
1865 // if this is a branch or switch on undef. Fold it manually as a
1866 // branch to the first successor.
1868 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1869 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1870 "Branch should be foldable!");
1871 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1872 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1874 llvm_unreachable("Didn't fold away reference to block!");
1878 // Make this an uncond branch to the first successor.
1879 TerminatorInst *TI = I->getParent()->getTerminator();
1880 BranchInst::Create(TI->getSuccessor(0), TI);
1882 // Remove entries in successor phi nodes to remove edges.
1883 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1884 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1886 // Remove the old terminator.
1887 TI->eraseFromParent();
1891 // Finally, delete the basic block.
1892 F->getBasicBlockList().erase(DeadBB);
1894 BlocksToErase.clear();
1897 // If we inferred constant or undef return values for a function, we replaced
1898 // all call uses with the inferred value. This means we don't need to bother
1899 // actually returning anything from the function. Replace all return
1900 // instructions with return undef.
1902 // Do this in two stages: first identify the functions we should process, then
1903 // actually zap their returns. This is important because we can only do this
1904 // if the address of the function isn't taken. In cases where a return is the
1905 // last use of a function, the order of processing functions would affect
1906 // whether other functions are optimizable.
1907 SmallVector<ReturnInst*, 8> ReturnsToZap;
1909 // TODO: Process multiple value ret instructions also.
1910 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1911 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1912 E = RV.end(); I != E; ++I) {
1913 Function *F = I->first;
1914 if (I->second.isOverdefined() || F->getReturnType()->isVoidTy())
1917 // We can only do this if we know that nothing else can call the function.
1918 if (!F->hasLocalLinkage() || AddressTakenFunctions.count(F))
1921 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1922 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1923 if (!isa<UndefValue>(RI->getOperand(0)))
1924 ReturnsToZap.push_back(RI);
1927 // Zap all returns which we've identified as zap to change.
1928 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) {
1929 Function *F = ReturnsToZap[i]->getParent()->getParent();
1930 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType()));
1933 // If we inferred constant or undef values for globals variables, we can
1934 // delete the global and any stores that remain to it.
1935 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1936 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1937 E = TG.end(); I != E; ++I) {
1938 GlobalVariable *GV = I->first;
1939 assert(!I->second.isOverdefined() &&
1940 "Overdefined values should have been taken out of the map!");
1941 DEBUG(dbgs() << "Found that GV '" << GV->getName() << "' is constant!\n");
1942 while (!GV->use_empty()) {
1943 StoreInst *SI = cast<StoreInst>(GV->user_back());
1944 SI->eraseFromParent();
1946 M.getGlobalList().erase(GV);