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/ConstantFolding.h"
28 #include "llvm/IR/CallSite.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/InstVisitor.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/Target/TargetLibraryInfo.h"
39 #include "llvm/Transforms/IPO.h"
40 #include "llvm/Transforms/Utils/Local.h"
44 #define DEBUG_TYPE "sccp"
46 STATISTIC(NumInstRemoved, "Number of instructions removed");
47 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
49 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP");
50 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
51 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
54 /// LatticeVal class - This class represents the different lattice values that
55 /// an LLVM value may occupy. It is a simple class with value semantics.
59 /// undefined - This LLVM Value has no known value yet.
62 /// constant - This LLVM Value has a specific constant value.
65 /// forcedconstant - This LLVM Value was thought to be undef until
66 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged
67 /// with another (different) constant, it goes to overdefined, instead of
71 /// overdefined - This instruction is not known to be constant, and we know
76 /// Val: This stores the current lattice value along with the Constant* for
77 /// the constant if this is a 'constant' or 'forcedconstant' value.
78 PointerIntPair<Constant *, 2, LatticeValueTy> Val;
80 LatticeValueTy getLatticeValue() const {
85 LatticeVal() : Val(nullptr, undefined) {}
87 bool isUndefined() const { return getLatticeValue() == undefined; }
88 bool isConstant() const {
89 return getLatticeValue() == constant || getLatticeValue() == forcedconstant;
91 bool isOverdefined() const { return getLatticeValue() == overdefined; }
93 Constant *getConstant() const {
94 assert(isConstant() && "Cannot get the constant of a non-constant!");
95 return Val.getPointer();
98 /// markOverdefined - Return true if this is a change in status.
99 bool markOverdefined() {
103 Val.setInt(overdefined);
107 /// markConstant - Return true if this is a change in status.
108 bool markConstant(Constant *V) {
109 if (getLatticeValue() == constant) { // Constant but not forcedconstant.
110 assert(getConstant() == V && "Marking constant with different value");
115 Val.setInt(constant);
116 assert(V && "Marking constant with NULL");
119 assert(getLatticeValue() == forcedconstant &&
120 "Cannot move from overdefined to constant!");
121 // Stay at forcedconstant if the constant is the same.
122 if (V == getConstant()) return false;
124 // Otherwise, we go to overdefined. Assumptions made based on the
125 // forced value are possibly wrong. Assuming this is another constant
126 // could expose a contradiction.
127 Val.setInt(overdefined);
132 /// getConstantInt - If this is a constant with a ConstantInt value, return it
133 /// otherwise return null.
134 ConstantInt *getConstantInt() const {
136 return dyn_cast<ConstantInt>(getConstant());
140 void markForcedConstant(Constant *V) {
141 assert(isUndefined() && "Can't force a defined value!");
142 Val.setInt(forcedconstant);
146 } // end anonymous namespace.
151 //===----------------------------------------------------------------------===//
153 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional
154 /// Constant Propagation.
156 class SCCPSolver : public InstVisitor<SCCPSolver> {
157 const DataLayout *DL;
158 const TargetLibraryInfo *TLI;
159 SmallPtrSet<BasicBlock*, 8> BBExecutable; // The BBs that are executable.
160 DenseMap<Value*, LatticeVal> ValueState; // The state each value is in.
162 /// StructValueState - This maintains ValueState for values that have
163 /// StructType, for example for formal arguments, calls, insertelement, etc.
165 DenseMap<std::pair<Value*, unsigned>, LatticeVal> StructValueState;
167 /// GlobalValue - If we are tracking any values for the contents of a global
168 /// variable, we keep a mapping from the constant accessor to the element of
169 /// the global, to the currently known value. If the value becomes
170 /// overdefined, it's entry is simply removed from this map.
171 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
173 /// TrackedRetVals - If we are tracking arguments into and the return
174 /// value out of a function, it will have an entry in this map, indicating
175 /// what the known return value for the function is.
176 DenseMap<Function*, LatticeVal> TrackedRetVals;
178 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
179 /// that return multiple values.
180 DenseMap<std::pair<Function*, unsigned>, LatticeVal> TrackedMultipleRetVals;
182 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is
183 /// represented here for efficient lookup.
184 SmallPtrSet<Function*, 16> MRVFunctionsTracked;
186 /// TrackingIncomingArguments - This is the set of functions for whose
187 /// arguments we make optimistic assumptions about and try to prove as
189 SmallPtrSet<Function*, 16> TrackingIncomingArguments;
191 /// The reason for two worklists is that overdefined is the lowest state
192 /// on the lattice, and moving things to overdefined as fast as possible
193 /// makes SCCP converge much faster.
195 /// By having a separate worklist, we accomplish this because everything
196 /// possibly overdefined will become overdefined at the soonest possible
198 SmallVector<Value*, 64> OverdefinedInstWorkList;
199 SmallVector<Value*, 64> InstWorkList;
202 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list
204 /// KnownFeasibleEdges - Entries in this set are edges which have already had
205 /// PHI nodes retriggered.
206 typedef std::pair<BasicBlock*, BasicBlock*> Edge;
207 DenseSet<Edge> KnownFeasibleEdges;
209 SCCPSolver(const DataLayout *DL, const TargetLibraryInfo *tli)
210 : DL(DL), TLI(tli) {}
212 /// MarkBlockExecutable - This method can be used by clients to mark all of
213 /// the blocks that are known to be intrinsically live in the processed unit.
215 /// This returns true if the block was not considered live before.
216 bool MarkBlockExecutable(BasicBlock *BB) {
217 if (!BBExecutable.insert(BB)) return false;
218 DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n');
219 BBWorkList.push_back(BB); // Add the block to the work list!
223 /// TrackValueOfGlobalVariable - Clients can use this method to
224 /// inform the SCCPSolver that it should track loads and stores to the
225 /// specified global variable if it can. This is only legal to call if
226 /// performing Interprocedural SCCP.
227 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
228 // We only track the contents of scalar globals.
229 if (GV->getType()->getElementType()->isSingleValueType()) {
230 LatticeVal &IV = TrackedGlobals[GV];
231 if (!isa<UndefValue>(GV->getInitializer()))
232 IV.markConstant(GV->getInitializer());
236 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
237 /// and out of the specified function (which cannot have its address taken),
238 /// this method must be called.
239 void AddTrackedFunction(Function *F) {
240 // Add an entry, F -> undef.
241 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
242 MRVFunctionsTracked.insert(F);
243 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
244 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
247 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
250 void AddArgumentTrackedFunction(Function *F) {
251 TrackingIncomingArguments.insert(F);
254 /// Solve - Solve for constants and executable blocks.
258 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
259 /// that branches on undef values cannot reach any of their successors.
260 /// However, this is not a safe assumption. After we solve dataflow, this
261 /// method should be use to handle this. If this returns true, the solver
263 bool ResolvedUndefsIn(Function &F);
265 bool isBlockExecutable(BasicBlock *BB) const {
266 return BBExecutable.count(BB);
269 LatticeVal getLatticeValueFor(Value *V) const {
270 DenseMap<Value*, LatticeVal>::const_iterator I = ValueState.find(V);
271 assert(I != ValueState.end() && "V is not in valuemap!");
275 /// getTrackedRetVals - Get the inferred return value map.
277 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
278 return TrackedRetVals;
281 /// getTrackedGlobals - Get and return the set of inferred initializers for
282 /// global variables.
283 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
284 return TrackedGlobals;
287 void markOverdefined(Value *V) {
288 assert(!V->getType()->isStructTy() && "Should use other method");
289 markOverdefined(ValueState[V], V);
292 /// markAnythingOverdefined - Mark the specified value overdefined. This
293 /// works with both scalars and structs.
294 void markAnythingOverdefined(Value *V) {
295 if (StructType *STy = dyn_cast<StructType>(V->getType()))
296 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
297 markOverdefined(getStructValueState(V, i), V);
303 // markConstant - Make a value be marked as "constant". If the value
304 // is not already a constant, add it to the instruction work list so that
305 // the users of the instruction are updated later.
307 void markConstant(LatticeVal &IV, Value *V, Constant *C) {
308 if (!IV.markConstant(C)) return;
309 DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n');
310 if (IV.isOverdefined())
311 OverdefinedInstWorkList.push_back(V);
313 InstWorkList.push_back(V);
316 void markConstant(Value *V, Constant *C) {
317 assert(!V->getType()->isStructTy() && "Should use other method");
318 markConstant(ValueState[V], V, C);
321 void markForcedConstant(Value *V, Constant *C) {
322 assert(!V->getType()->isStructTy() && "Should use other method");
323 LatticeVal &IV = ValueState[V];
324 IV.markForcedConstant(C);
325 DEBUG(dbgs() << "markForcedConstant: " << *C << ": " << *V << '\n');
326 if (IV.isOverdefined())
327 OverdefinedInstWorkList.push_back(V);
329 InstWorkList.push_back(V);
333 // markOverdefined - Make a value be marked as "overdefined". If the
334 // value is not already overdefined, add it to the overdefined instruction
335 // work list so that the users of the instruction are updated later.
336 void markOverdefined(LatticeVal &IV, Value *V) {
337 if (!IV.markOverdefined()) return;
339 DEBUG(dbgs() << "markOverdefined: ";
340 if (Function *F = dyn_cast<Function>(V))
341 dbgs() << "Function '" << F->getName() << "'\n";
343 dbgs() << *V << '\n');
344 // Only instructions go on the work list
345 OverdefinedInstWorkList.push_back(V);
348 void mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) {
349 if (IV.isOverdefined() || MergeWithV.isUndefined())
351 if (MergeWithV.isOverdefined())
352 markOverdefined(IV, V);
353 else if (IV.isUndefined())
354 markConstant(IV, V, MergeWithV.getConstant());
355 else if (IV.getConstant() != MergeWithV.getConstant())
356 markOverdefined(IV, V);
359 void mergeInValue(Value *V, LatticeVal MergeWithV) {
360 assert(!V->getType()->isStructTy() && "Should use other method");
361 mergeInValue(ValueState[V], V, MergeWithV);
365 /// getValueState - Return the LatticeVal object that corresponds to the
366 /// value. This function handles the case when the value hasn't been seen yet
367 /// by properly seeding constants etc.
368 LatticeVal &getValueState(Value *V) {
369 assert(!V->getType()->isStructTy() && "Should use getStructValueState");
371 std::pair<DenseMap<Value*, LatticeVal>::iterator, bool> I =
372 ValueState.insert(std::make_pair(V, LatticeVal()));
373 LatticeVal &LV = I.first->second;
376 return LV; // Common case, already in the map.
378 if (Constant *C = dyn_cast<Constant>(V)) {
379 // Undef values remain undefined.
380 if (!isa<UndefValue>(V))
381 LV.markConstant(C); // Constants are constant
384 // All others are underdefined by default.
388 /// getStructValueState - Return the LatticeVal object that corresponds to the
389 /// value/field pair. This function handles the case when the value hasn't
390 /// been seen yet by properly seeding constants etc.
391 LatticeVal &getStructValueState(Value *V, unsigned i) {
392 assert(V->getType()->isStructTy() && "Should use getValueState");
393 assert(i < cast<StructType>(V->getType())->getNumElements() &&
394 "Invalid element #");
396 std::pair<DenseMap<std::pair<Value*, unsigned>, LatticeVal>::iterator,
397 bool> I = StructValueState.insert(
398 std::make_pair(std::make_pair(V, i), LatticeVal()));
399 LatticeVal &LV = I.first->second;
402 return LV; // Common case, already in the map.
404 if (Constant *C = dyn_cast<Constant>(V)) {
405 Constant *Elt = C->getAggregateElement(i);
408 LV.markOverdefined(); // Unknown sort of constant.
409 else if (isa<UndefValue>(Elt))
410 ; // Undef values remain undefined.
412 LV.markConstant(Elt); // Constants are constant.
415 // All others are underdefined by default.
420 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
421 /// work list if it is not already executable.
422 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
423 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
424 return; // This edge is already known to be executable!
426 if (!MarkBlockExecutable(Dest)) {
427 // If the destination is already executable, we just made an *edge*
428 // feasible that wasn't before. Revisit the PHI nodes in the block
429 // because they have potentially new operands.
430 DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName()
431 << " -> " << Dest->getName() << '\n');
434 for (BasicBlock::iterator I = Dest->begin();
435 (PN = dyn_cast<PHINode>(I)); ++I)
440 // getFeasibleSuccessors - Return a vector of booleans to indicate which
441 // successors are reachable from a given terminator instruction.
443 void getFeasibleSuccessors(TerminatorInst &TI, SmallVectorImpl<bool> &Succs);
445 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
446 // block to the 'To' basic block is currently feasible.
448 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
450 // OperandChangedState - This method is invoked on all of the users of an
451 // instruction that was just changed state somehow. Based on this
452 // information, we need to update the specified user of this instruction.
454 void OperandChangedState(Instruction *I) {
455 if (BBExecutable.count(I->getParent())) // Inst is executable?
460 friend class InstVisitor<SCCPSolver>;
462 // visit implementations - Something changed in this instruction. Either an
463 // operand made a transition, or the instruction is newly executable. Change
464 // the value type of I to reflect these changes if appropriate.
465 void visitPHINode(PHINode &I);
468 void visitReturnInst(ReturnInst &I);
469 void visitTerminatorInst(TerminatorInst &TI);
471 void visitCastInst(CastInst &I);
472 void visitSelectInst(SelectInst &I);
473 void visitBinaryOperator(Instruction &I);
474 void visitCmpInst(CmpInst &I);
475 void visitExtractElementInst(ExtractElementInst &I);
476 void visitInsertElementInst(InsertElementInst &I);
477 void visitShuffleVectorInst(ShuffleVectorInst &I);
478 void visitExtractValueInst(ExtractValueInst &EVI);
479 void visitInsertValueInst(InsertValueInst &IVI);
480 void visitLandingPadInst(LandingPadInst &I) { markAnythingOverdefined(&I); }
482 // Instructions that cannot be folded away.
483 void visitStoreInst (StoreInst &I);
484 void visitLoadInst (LoadInst &I);
485 void visitGetElementPtrInst(GetElementPtrInst &I);
486 void visitCallInst (CallInst &I) {
489 void visitInvokeInst (InvokeInst &II) {
491 visitTerminatorInst(II);
493 void visitCallSite (CallSite CS);
494 void visitResumeInst (TerminatorInst &I) { /*returns void*/ }
495 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
496 void visitFenceInst (FenceInst &I) { /*returns void*/ }
497 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
498 markAnythingOverdefined(&I);
500 void visitAtomicRMWInst (AtomicRMWInst &I) { markOverdefined(&I); }
501 void visitAllocaInst (Instruction &I) { markOverdefined(&I); }
502 void visitVAArgInst (Instruction &I) { markAnythingOverdefined(&I); }
504 void visitInstruction(Instruction &I) {
505 // If a new instruction is added to LLVM that we don't handle.
506 dbgs() << "SCCP: Don't know how to handle: " << I << '\n';
507 markAnythingOverdefined(&I); // Just in case
511 } // end anonymous namespace
514 // getFeasibleSuccessors - Return a vector of booleans to indicate which
515 // successors are reachable from a given terminator instruction.
517 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
518 SmallVectorImpl<bool> &Succs) {
519 Succs.resize(TI.getNumSuccessors());
520 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
521 if (BI->isUnconditional()) {
526 LatticeVal BCValue = getValueState(BI->getCondition());
527 ConstantInt *CI = BCValue.getConstantInt();
529 // Overdefined condition variables, and branches on unfoldable constant
530 // conditions, mean the branch could go either way.
531 if (!BCValue.isUndefined())
532 Succs[0] = Succs[1] = true;
536 // Constant condition variables mean the branch can only go a single way.
537 Succs[CI->isZero()] = true;
541 if (isa<InvokeInst>(TI)) {
542 // Invoke instructions successors are always executable.
543 Succs[0] = Succs[1] = true;
547 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
548 if (!SI->getNumCases()) {
552 LatticeVal SCValue = getValueState(SI->getCondition());
553 ConstantInt *CI = SCValue.getConstantInt();
555 if (!CI) { // Overdefined or undefined condition?
556 // All destinations are executable!
557 if (!SCValue.isUndefined())
558 Succs.assign(TI.getNumSuccessors(), true);
562 Succs[SI->findCaseValue(CI).getSuccessorIndex()] = true;
566 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
567 if (isa<IndirectBrInst>(&TI)) {
568 // Just mark all destinations executable!
569 Succs.assign(TI.getNumSuccessors(), true);
574 dbgs() << "Unknown terminator instruction: " << TI << '\n';
576 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
580 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
581 // block to the 'To' basic block is currently feasible.
583 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
584 assert(BBExecutable.count(To) && "Dest should always be alive!");
586 // Make sure the source basic block is executable!!
587 if (!BBExecutable.count(From)) return false;
589 // Check to make sure this edge itself is actually feasible now.
590 TerminatorInst *TI = From->getTerminator();
591 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
592 if (BI->isUnconditional())
595 LatticeVal BCValue = getValueState(BI->getCondition());
597 // Overdefined condition variables mean the branch could go either way,
598 // undef conditions mean that neither edge is feasible yet.
599 ConstantInt *CI = BCValue.getConstantInt();
601 return !BCValue.isUndefined();
603 // Constant condition variables mean the branch can only go a single way.
604 return BI->getSuccessor(CI->isZero()) == To;
607 // Invoke instructions successors are always executable.
608 if (isa<InvokeInst>(TI))
611 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
612 if (SI->getNumCases() < 1)
615 LatticeVal SCValue = getValueState(SI->getCondition());
616 ConstantInt *CI = SCValue.getConstantInt();
619 return !SCValue.isUndefined();
621 return SI->findCaseValue(CI).getCaseSuccessor() == To;
624 // Just mark all destinations executable!
625 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
626 if (isa<IndirectBrInst>(TI))
630 dbgs() << "Unknown terminator instruction: " << *TI << '\n';
632 llvm_unreachable(nullptr);
635 // visit Implementations - Something changed in this instruction, either an
636 // operand made a transition, or the instruction is newly executable. Change
637 // the value type of I to reflect these changes if appropriate. This method
638 // makes sure to do the following actions:
640 // 1. If a phi node merges two constants in, and has conflicting value coming
641 // from different branches, or if the PHI node merges in an overdefined
642 // value, then the PHI node becomes overdefined.
643 // 2. If a phi node merges only constants in, and they all agree on value, the
644 // PHI node becomes a constant value equal to that.
645 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
646 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
647 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
648 // 6. If a conditional branch has a value that is constant, make the selected
649 // destination executable
650 // 7. If a conditional branch has a value that is overdefined, make all
651 // successors executable.
653 void SCCPSolver::visitPHINode(PHINode &PN) {
654 // If this PN returns a struct, just mark the result overdefined.
655 // TODO: We could do a lot better than this if code actually uses this.
656 if (PN.getType()->isStructTy())
657 return markAnythingOverdefined(&PN);
659 if (getValueState(&PN).isOverdefined())
660 return; // Quick exit
662 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
663 // and slow us down a lot. Just mark them overdefined.
664 if (PN.getNumIncomingValues() > 64)
665 return markOverdefined(&PN);
667 // Look at all of the executable operands of the PHI node. If any of them
668 // are overdefined, the PHI becomes overdefined as well. If they are all
669 // constant, and they agree with each other, the PHI becomes the identical
670 // constant. If they are constant and don't agree, the PHI is overdefined.
671 // If there are no executable operands, the PHI remains undefined.
673 Constant *OperandVal = nullptr;
674 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
675 LatticeVal IV = getValueState(PN.getIncomingValue(i));
676 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
678 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
681 if (IV.isOverdefined()) // PHI node becomes overdefined!
682 return markOverdefined(&PN);
684 if (!OperandVal) { // Grab the first value.
685 OperandVal = IV.getConstant();
689 // There is already a reachable operand. If we conflict with it,
690 // then the PHI node becomes overdefined. If we agree with it, we
693 // Check to see if there are two different constants merging, if so, the PHI
694 // node is overdefined.
695 if (IV.getConstant() != OperandVal)
696 return markOverdefined(&PN);
699 // If we exited the loop, this means that the PHI node only has constant
700 // arguments that agree with each other(and OperandVal is the constant) or
701 // OperandVal is null because there are no defined incoming arguments. If
702 // this is the case, the PHI remains undefined.
705 markConstant(&PN, OperandVal); // Acquire operand value
708 void SCCPSolver::visitReturnInst(ReturnInst &I) {
709 if (I.getNumOperands() == 0) return; // ret void
711 Function *F = I.getParent()->getParent();
712 Value *ResultOp = I.getOperand(0);
714 // If we are tracking the return value of this function, merge it in.
715 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) {
716 DenseMap<Function*, LatticeVal>::iterator TFRVI =
717 TrackedRetVals.find(F);
718 if (TFRVI != TrackedRetVals.end()) {
719 mergeInValue(TFRVI->second, F, getValueState(ResultOp));
724 // Handle functions that return multiple values.
725 if (!TrackedMultipleRetVals.empty()) {
726 if (StructType *STy = dyn_cast<StructType>(ResultOp->getType()))
727 if (MRVFunctionsTracked.count(F))
728 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
729 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F,
730 getStructValueState(ResultOp, i));
735 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
736 SmallVector<bool, 16> SuccFeasible;
737 getFeasibleSuccessors(TI, SuccFeasible);
739 BasicBlock *BB = TI.getParent();
741 // Mark all feasible successors executable.
742 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
744 markEdgeExecutable(BB, TI.getSuccessor(i));
747 void SCCPSolver::visitCastInst(CastInst &I) {
748 LatticeVal OpSt = getValueState(I.getOperand(0));
749 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand
751 else if (OpSt.isConstant()) // Propagate constant value
752 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
753 OpSt.getConstant(), I.getType()));
757 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
758 // If this returns a struct, mark all elements over defined, we don't track
759 // structs in structs.
760 if (EVI.getType()->isStructTy())
761 return markAnythingOverdefined(&EVI);
763 // If this is extracting from more than one level of struct, we don't know.
764 if (EVI.getNumIndices() != 1)
765 return markOverdefined(&EVI);
767 Value *AggVal = EVI.getAggregateOperand();
768 if (AggVal->getType()->isStructTy()) {
769 unsigned i = *EVI.idx_begin();
770 LatticeVal EltVal = getStructValueState(AggVal, i);
771 mergeInValue(getValueState(&EVI), &EVI, EltVal);
773 // Otherwise, must be extracting from an array.
774 return markOverdefined(&EVI);
778 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
779 StructType *STy = dyn_cast<StructType>(IVI.getType());
781 return markOverdefined(&IVI);
783 // If this has more than one index, we can't handle it, drive all results to
785 if (IVI.getNumIndices() != 1)
786 return markAnythingOverdefined(&IVI);
788 Value *Aggr = IVI.getAggregateOperand();
789 unsigned Idx = *IVI.idx_begin();
791 // Compute the result based on what we're inserting.
792 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
793 // This passes through all values that aren't the inserted element.
795 LatticeVal EltVal = getStructValueState(Aggr, i);
796 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal);
800 Value *Val = IVI.getInsertedValueOperand();
801 if (Val->getType()->isStructTy())
802 // We don't track structs in structs.
803 markOverdefined(getStructValueState(&IVI, i), &IVI);
805 LatticeVal InVal = getValueState(Val);
806 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal);
811 void SCCPSolver::visitSelectInst(SelectInst &I) {
812 // If this select returns a struct, just mark the result overdefined.
813 // TODO: We could do a lot better than this if code actually uses this.
814 if (I.getType()->isStructTy())
815 return markAnythingOverdefined(&I);
817 LatticeVal CondValue = getValueState(I.getCondition());
818 if (CondValue.isUndefined())
821 if (ConstantInt *CondCB = CondValue.getConstantInt()) {
822 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
823 mergeInValue(&I, getValueState(OpVal));
827 // Otherwise, the condition is overdefined or a constant we can't evaluate.
828 // See if we can produce something better than overdefined based on the T/F
830 LatticeVal TVal = getValueState(I.getTrueValue());
831 LatticeVal FVal = getValueState(I.getFalseValue());
833 // select ?, C, C -> C.
834 if (TVal.isConstant() && FVal.isConstant() &&
835 TVal.getConstant() == FVal.getConstant())
836 return markConstant(&I, FVal.getConstant());
838 if (TVal.isUndefined()) // select ?, undef, X -> X.
839 return mergeInValue(&I, FVal);
840 if (FVal.isUndefined()) // select ?, X, undef -> X.
841 return mergeInValue(&I, TVal);
845 // Handle Binary Operators.
846 void SCCPSolver::visitBinaryOperator(Instruction &I) {
847 LatticeVal V1State = getValueState(I.getOperand(0));
848 LatticeVal V2State = getValueState(I.getOperand(1));
850 LatticeVal &IV = ValueState[&I];
851 if (IV.isOverdefined()) return;
853 if (V1State.isConstant() && V2State.isConstant())
854 return markConstant(IV, &I,
855 ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
856 V2State.getConstant()));
858 // If something is undef, wait for it to resolve.
859 if (!V1State.isOverdefined() && !V2State.isOverdefined())
862 // Otherwise, one of our operands is overdefined. Try to produce something
863 // better than overdefined with some tricks.
865 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
866 // operand is overdefined.
867 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
868 LatticeVal *NonOverdefVal = nullptr;
869 if (!V1State.isOverdefined())
870 NonOverdefVal = &V1State;
871 else if (!V2State.isOverdefined())
872 NonOverdefVal = &V2State;
875 if (NonOverdefVal->isUndefined()) {
876 // Could annihilate value.
877 if (I.getOpcode() == Instruction::And)
878 markConstant(IV, &I, Constant::getNullValue(I.getType()));
879 else if (VectorType *PT = dyn_cast<VectorType>(I.getType()))
880 markConstant(IV, &I, Constant::getAllOnesValue(PT));
883 Constant::getAllOnesValue(I.getType()));
887 if (I.getOpcode() == Instruction::And) {
889 if (NonOverdefVal->getConstant()->isNullValue())
890 return markConstant(IV, &I, NonOverdefVal->getConstant());
892 if (ConstantInt *CI = NonOverdefVal->getConstantInt())
893 if (CI->isAllOnesValue()) // X or -1 = -1
894 return markConstant(IV, &I, NonOverdefVal->getConstant());
903 // Handle ICmpInst instruction.
904 void SCCPSolver::visitCmpInst(CmpInst &I) {
905 LatticeVal V1State = getValueState(I.getOperand(0));
906 LatticeVal V2State = getValueState(I.getOperand(1));
908 LatticeVal &IV = ValueState[&I];
909 if (IV.isOverdefined()) return;
911 if (V1State.isConstant() && V2State.isConstant())
912 return markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
913 V1State.getConstant(),
914 V2State.getConstant()));
916 // If operands are still undefined, wait for it to resolve.
917 if (!V1State.isOverdefined() && !V2State.isOverdefined())
923 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
924 // TODO : SCCP does not handle vectors properly.
925 return markOverdefined(&I);
928 LatticeVal &ValState = getValueState(I.getOperand(0));
929 LatticeVal &IdxState = getValueState(I.getOperand(1));
931 if (ValState.isOverdefined() || IdxState.isOverdefined())
933 else if(ValState.isConstant() && IdxState.isConstant())
934 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
935 IdxState.getConstant()));
939 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
940 // TODO : SCCP does not handle vectors properly.
941 return markOverdefined(&I);
943 LatticeVal &ValState = getValueState(I.getOperand(0));
944 LatticeVal &EltState = getValueState(I.getOperand(1));
945 LatticeVal &IdxState = getValueState(I.getOperand(2));
947 if (ValState.isOverdefined() || EltState.isOverdefined() ||
948 IdxState.isOverdefined())
950 else if(ValState.isConstant() && EltState.isConstant() &&
951 IdxState.isConstant())
952 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
953 EltState.getConstant(),
954 IdxState.getConstant()));
955 else if (ValState.isUndefined() && EltState.isConstant() &&
956 IdxState.isConstant())
957 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
958 EltState.getConstant(),
959 IdxState.getConstant()));
963 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
964 // TODO : SCCP does not handle vectors properly.
965 return markOverdefined(&I);
967 LatticeVal &V1State = getValueState(I.getOperand(0));
968 LatticeVal &V2State = getValueState(I.getOperand(1));
969 LatticeVal &MaskState = getValueState(I.getOperand(2));
971 if (MaskState.isUndefined() ||
972 (V1State.isUndefined() && V2State.isUndefined()))
973 return; // Undefined output if mask or both inputs undefined.
975 if (V1State.isOverdefined() || V2State.isOverdefined() ||
976 MaskState.isOverdefined()) {
979 // A mix of constant/undef inputs.
980 Constant *V1 = V1State.isConstant() ?
981 V1State.getConstant() : UndefValue::get(I.getType());
982 Constant *V2 = V2State.isConstant() ?
983 V2State.getConstant() : UndefValue::get(I.getType());
984 Constant *Mask = MaskState.isConstant() ?
985 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
986 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
991 // Handle getelementptr instructions. If all operands are constants then we
992 // can turn this into a getelementptr ConstantExpr.
994 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
995 if (ValueState[&I].isOverdefined()) return;
997 SmallVector<Constant*, 8> Operands;
998 Operands.reserve(I.getNumOperands());
1000 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1001 LatticeVal State = getValueState(I.getOperand(i));
1002 if (State.isUndefined())
1003 return; // Operands are not resolved yet.
1005 if (State.isOverdefined())
1006 return markOverdefined(&I);
1008 assert(State.isConstant() && "Unknown state!");
1009 Operands.push_back(State.getConstant());
1012 Constant *Ptr = Operands[0];
1013 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end());
1014 markConstant(&I, ConstantExpr::getGetElementPtr(Ptr, Indices));
1017 void SCCPSolver::visitStoreInst(StoreInst &SI) {
1018 // If this store is of a struct, ignore it.
1019 if (SI.getOperand(0)->getType()->isStructTy())
1022 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1025 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1026 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1027 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1029 // Get the value we are storing into the global, then merge it.
1030 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)));
1031 if (I->second.isOverdefined())
1032 TrackedGlobals.erase(I); // No need to keep tracking this!
1036 // Handle load instructions. If the operand is a constant pointer to a constant
1037 // global, we can replace the load with the loaded constant value!
1038 void SCCPSolver::visitLoadInst(LoadInst &I) {
1039 // If this load is of a struct, just mark the result overdefined.
1040 if (I.getType()->isStructTy())
1041 return markAnythingOverdefined(&I);
1043 LatticeVal PtrVal = getValueState(I.getOperand(0));
1044 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1046 LatticeVal &IV = ValueState[&I];
1047 if (IV.isOverdefined()) return;
1049 if (!PtrVal.isConstant() || I.isVolatile())
1050 return markOverdefined(IV, &I);
1052 Constant *Ptr = PtrVal.getConstant();
1054 // load null -> null
1055 if (isa<ConstantPointerNull>(Ptr) && I.getPointerAddressSpace() == 0)
1056 return markConstant(IV, &I, Constant::getNullValue(I.getType()));
1058 // Transform load (constant global) into the value loaded.
1059 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1060 if (!TrackedGlobals.empty()) {
1061 // If we are tracking this global, merge in the known value for it.
1062 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1063 TrackedGlobals.find(GV);
1064 if (It != TrackedGlobals.end()) {
1065 mergeInValue(IV, &I, It->second);
1071 // Transform load from a constant into a constant if possible.
1072 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, DL))
1073 return markConstant(IV, &I, C);
1075 // Otherwise we cannot say for certain what value this load will produce.
1077 markOverdefined(IV, &I);
1080 void SCCPSolver::visitCallSite(CallSite CS) {
1081 Function *F = CS.getCalledFunction();
1082 Instruction *I = CS.getInstruction();
1084 // The common case is that we aren't tracking the callee, either because we
1085 // are not doing interprocedural analysis or the callee is indirect, or is
1086 // external. Handle these cases first.
1087 if (!F || F->isDeclaration()) {
1089 // Void return and not tracking callee, just bail.
1090 if (I->getType()->isVoidTy()) return;
1092 // Otherwise, if we have a single return value case, and if the function is
1093 // a declaration, maybe we can constant fold it.
1094 if (F && F->isDeclaration() && !I->getType()->isStructTy() &&
1095 canConstantFoldCallTo(F)) {
1097 SmallVector<Constant*, 8> Operands;
1098 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1100 LatticeVal State = getValueState(*AI);
1102 if (State.isUndefined())
1103 return; // Operands are not resolved yet.
1104 if (State.isOverdefined())
1105 return markOverdefined(I);
1106 assert(State.isConstant() && "Unknown state!");
1107 Operands.push_back(State.getConstant());
1110 if (getValueState(I).isOverdefined())
1113 // If we can constant fold this, mark the result of the call as a
1115 if (Constant *C = ConstantFoldCall(F, Operands, TLI))
1116 return markConstant(I, C);
1119 // Otherwise, we don't know anything about this call, mark it overdefined.
1120 return markAnythingOverdefined(I);
1123 // If this is a local function that doesn't have its address taken, mark its
1124 // entry block executable and merge in the actual arguments to the call into
1125 // the formal arguments of the function.
1126 if (!TrackingIncomingArguments.empty() && TrackingIncomingArguments.count(F)){
1127 MarkBlockExecutable(F->begin());
1129 // Propagate information from this call site into the callee.
1130 CallSite::arg_iterator CAI = CS.arg_begin();
1131 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1132 AI != E; ++AI, ++CAI) {
1133 // If this argument is byval, and if the function is not readonly, there
1134 // will be an implicit copy formed of the input aggregate.
1135 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1136 markOverdefined(AI);
1140 if (StructType *STy = dyn_cast<StructType>(AI->getType())) {
1141 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1142 LatticeVal CallArg = getStructValueState(*CAI, i);
1143 mergeInValue(getStructValueState(AI, i), AI, CallArg);
1146 mergeInValue(AI, getValueState(*CAI));
1151 // If this is a single/zero retval case, see if we're tracking the function.
1152 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
1153 if (!MRVFunctionsTracked.count(F))
1154 goto CallOverdefined; // Not tracking this callee.
1156 // If we are tracking this callee, propagate the result of the function
1157 // into this call site.
1158 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1159 mergeInValue(getStructValueState(I, i), I,
1160 TrackedMultipleRetVals[std::make_pair(F, i)]);
1162 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1163 if (TFRVI == TrackedRetVals.end())
1164 goto CallOverdefined; // Not tracking this callee.
1166 // If so, propagate the return value of the callee into this call result.
1167 mergeInValue(I, TFRVI->second);
1171 void SCCPSolver::Solve() {
1172 // Process the work lists until they are empty!
1173 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1174 !OverdefinedInstWorkList.empty()) {
1175 // Process the overdefined instruction's work list first, which drives other
1176 // things to overdefined more quickly.
1177 while (!OverdefinedInstWorkList.empty()) {
1178 Value *I = OverdefinedInstWorkList.pop_back_val();
1180 DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n');
1182 // "I" got into the work list because it either made the transition from
1183 // bottom to constant, or to overdefined.
1185 // Anything on this worklist that is overdefined need not be visited
1186 // since all of its users will have already been marked as overdefined
1187 // Update all of the users of this instruction's value.
1189 for (User *U : I->users())
1190 if (Instruction *UI = dyn_cast<Instruction>(U))
1191 OperandChangedState(UI);
1194 // Process the instruction work list.
1195 while (!InstWorkList.empty()) {
1196 Value *I = InstWorkList.pop_back_val();
1198 DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n');
1200 // "I" got into the work list because it made the transition from undef to
1203 // Anything on this worklist that is overdefined need not be visited
1204 // since all of its users will have already been marked as overdefined.
1205 // Update all of the users of this instruction's value.
1207 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined())
1208 for (User *U : I->users())
1209 if (Instruction *UI = dyn_cast<Instruction>(U))
1210 OperandChangedState(UI);
1213 // Process the basic block work list.
1214 while (!BBWorkList.empty()) {
1215 BasicBlock *BB = BBWorkList.back();
1216 BBWorkList.pop_back();
1218 DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n');
1220 // Notify all instructions in this basic block that they are newly
1227 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1228 /// that branches on undef values cannot reach any of their successors.
1229 /// However, this is not a safe assumption. After we solve dataflow, this
1230 /// method should be use to handle this. If this returns true, the solver
1231 /// should be rerun.
1233 /// This method handles this by finding an unresolved branch and marking it one
1234 /// of the edges from the block as being feasible, even though the condition
1235 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1236 /// CFG and only slightly pessimizes the analysis results (by marking one,
1237 /// potentially infeasible, edge feasible). This cannot usefully modify the
1238 /// constraints on the condition of the branch, as that would impact other users
1241 /// This scan also checks for values that use undefs, whose results are actually
1242 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1243 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1244 /// even if X isn't defined.
1245 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1246 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1247 if (!BBExecutable.count(BB))
1250 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1251 // Look for instructions which produce undef values.
1252 if (I->getType()->isVoidTy()) continue;
1254 if (StructType *STy = dyn_cast<StructType>(I->getType())) {
1255 // Only a few things that can be structs matter for undef.
1257 // Tracked calls must never be marked overdefined in ResolvedUndefsIn.
1258 if (CallSite CS = CallSite(I))
1259 if (Function *F = CS.getCalledFunction())
1260 if (MRVFunctionsTracked.count(F))
1263 // extractvalue and insertvalue don't need to be marked; they are
1264 // tracked as precisely as their operands.
1265 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I))
1268 // Send the results of everything else to overdefined. We could be
1269 // more precise than this but it isn't worth bothering.
1270 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1271 LatticeVal &LV = getStructValueState(I, i);
1272 if (LV.isUndefined())
1273 markOverdefined(LV, I);
1278 LatticeVal &LV = getValueState(I);
1279 if (!LV.isUndefined()) continue;
1281 // extractvalue is safe; check here because the argument is a struct.
1282 if (isa<ExtractValueInst>(I))
1285 // Compute the operand LatticeVals, for convenience below.
1286 // Anything taking a struct is conservatively assumed to require
1287 // overdefined markings.
1288 if (I->getOperand(0)->getType()->isStructTy()) {
1292 LatticeVal Op0LV = getValueState(I->getOperand(0));
1294 if (I->getNumOperands() == 2) {
1295 if (I->getOperand(1)->getType()->isStructTy()) {
1300 Op1LV = getValueState(I->getOperand(1));
1302 // If this is an instructions whose result is defined even if the input is
1303 // not fully defined, propagate the information.
1304 Type *ITy = I->getType();
1305 switch (I->getOpcode()) {
1306 case Instruction::Add:
1307 case Instruction::Sub:
1308 case Instruction::Trunc:
1309 case Instruction::FPTrunc:
1310 case Instruction::BitCast:
1311 break; // Any undef -> undef
1312 case Instruction::FSub:
1313 case Instruction::FAdd:
1314 case Instruction::FMul:
1315 case Instruction::FDiv:
1316 case Instruction::FRem:
1317 // Floating-point binary operation: be conservative.
1318 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1319 markForcedConstant(I, Constant::getNullValue(ITy));
1323 case Instruction::ZExt:
1324 case Instruction::SExt:
1325 case Instruction::FPToUI:
1326 case Instruction::FPToSI:
1327 case Instruction::FPExt:
1328 case Instruction::PtrToInt:
1329 case Instruction::IntToPtr:
1330 case Instruction::SIToFP:
1331 case Instruction::UIToFP:
1332 // undef -> 0; some outputs are impossible
1333 markForcedConstant(I, Constant::getNullValue(ITy));
1335 case Instruction::Mul:
1336 case Instruction::And:
1337 // Both operands undef -> undef
1338 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1340 // undef * X -> 0. X could be zero.
1341 // undef & X -> 0. X could be zero.
1342 markForcedConstant(I, Constant::getNullValue(ITy));
1345 case Instruction::Or:
1346 // Both operands undef -> undef
1347 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1349 // undef | X -> -1. X could be -1.
1350 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1353 case Instruction::Xor:
1354 // undef ^ undef -> 0; strictly speaking, this is not strictly
1355 // necessary, but we try to be nice to people who expect this
1356 // behavior in simple cases
1357 if (Op0LV.isUndefined() && Op1LV.isUndefined()) {
1358 markForcedConstant(I, Constant::getNullValue(ITy));
1361 // undef ^ X -> undef
1364 case Instruction::SDiv:
1365 case Instruction::UDiv:
1366 case Instruction::SRem:
1367 case Instruction::URem:
1368 // X / undef -> undef. No change.
1369 // X % undef -> undef. No change.
1370 if (Op1LV.isUndefined()) break;
1372 // undef / X -> 0. X could be maxint.
1373 // undef % X -> 0. X could be 1.
1374 markForcedConstant(I, Constant::getNullValue(ITy));
1377 case Instruction::AShr:
1378 // X >>a undef -> undef.
1379 if (Op1LV.isUndefined()) break;
1381 // undef >>a X -> all ones
1382 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1384 case Instruction::LShr:
1385 case Instruction::Shl:
1386 // X << undef -> undef.
1387 // X >> undef -> undef.
1388 if (Op1LV.isUndefined()) break;
1392 markForcedConstant(I, Constant::getNullValue(ITy));
1394 case Instruction::Select:
1395 Op1LV = getValueState(I->getOperand(1));
1396 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1397 if (Op0LV.isUndefined()) {
1398 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1399 Op1LV = getValueState(I->getOperand(2));
1400 } else if (Op1LV.isUndefined()) {
1401 // c ? undef : undef -> undef. No change.
1402 Op1LV = getValueState(I->getOperand(2));
1403 if (Op1LV.isUndefined())
1405 // Otherwise, c ? undef : x -> x.
1407 // Leave Op1LV as Operand(1)'s LatticeValue.
1410 if (Op1LV.isConstant())
1411 markForcedConstant(I, Op1LV.getConstant());
1415 case Instruction::Load:
1416 // A load here means one of two things: a load of undef from a global,
1417 // a load from an unknown pointer. Either way, having it return undef
1420 case Instruction::ICmp:
1421 // X == undef -> undef. Other comparisons get more complicated.
1422 if (cast<ICmpInst>(I)->isEquality())
1426 case Instruction::Call:
1427 case Instruction::Invoke: {
1428 // There are two reasons a call can have an undef result
1429 // 1. It could be tracked.
1430 // 2. It could be constant-foldable.
1431 // Because of the way we solve return values, tracked calls must
1432 // never be marked overdefined in ResolvedUndefsIn.
1433 if (Function *F = CallSite(I).getCalledFunction())
1434 if (TrackedRetVals.count(F))
1437 // If the call is constant-foldable, we mark it overdefined because
1438 // we do not know what return values are valid.
1443 // If we don't know what should happen here, conservatively mark it
1450 // Check to see if we have a branch or switch on an undefined value. If so
1451 // we force the branch to go one way or the other to make the successor
1452 // values live. It doesn't really matter which way we force it.
1453 TerminatorInst *TI = BB->getTerminator();
1454 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1455 if (!BI->isConditional()) continue;
1456 if (!getValueState(BI->getCondition()).isUndefined())
1459 // If the input to SCCP is actually branch on undef, fix the undef to
1461 if (isa<UndefValue>(BI->getCondition())) {
1462 BI->setCondition(ConstantInt::getFalse(BI->getContext()));
1463 markEdgeExecutable(BB, TI->getSuccessor(1));
1467 // Otherwise, it is a branch on a symbolic value which is currently
1468 // considered to be undef. Handle this by forcing the input value to the
1470 markForcedConstant(BI->getCondition(),
1471 ConstantInt::getFalse(TI->getContext()));
1475 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1476 if (!SI->getNumCases())
1478 if (!getValueState(SI->getCondition()).isUndefined())
1481 // If the input to SCCP is actually switch on undef, fix the undef to
1482 // the first constant.
1483 if (isa<UndefValue>(SI->getCondition())) {
1484 SI->setCondition(SI->case_begin().getCaseValue());
1485 markEdgeExecutable(BB, SI->case_begin().getCaseSuccessor());
1489 markForcedConstant(SI->getCondition(), SI->case_begin().getCaseValue());
1499 //===--------------------------------------------------------------------===//
1501 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1502 /// Sparse Conditional Constant Propagator.
1504 struct SCCP : public FunctionPass {
1505 void getAnalysisUsage(AnalysisUsage &AU) const override {
1506 AU.addRequired<TargetLibraryInfo>();
1508 static char ID; // Pass identification, replacement for typeid
1509 SCCP() : FunctionPass(ID) {
1510 initializeSCCPPass(*PassRegistry::getPassRegistry());
1513 // runOnFunction - Run the Sparse Conditional Constant Propagation
1514 // algorithm, and return true if the function was modified.
1516 bool runOnFunction(Function &F) override;
1518 } // end anonymous namespace
1521 INITIALIZE_PASS(SCCP, "sccp",
1522 "Sparse Conditional Constant Propagation", false, false)
1524 // createSCCPPass - This is the public interface to this file.
1525 FunctionPass *llvm::createSCCPPass() {
1529 static void DeleteInstructionInBlock(BasicBlock *BB) {
1530 DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
1533 // Check to see if there are non-terminating instructions to delete.
1534 if (isa<TerminatorInst>(BB->begin()))
1537 // Delete the instructions backwards, as it has a reduced likelihood of having
1538 // to update as many def-use and use-def chains.
1539 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
1540 while (EndInst != BB->begin()) {
1541 // Delete the next to last instruction.
1542 BasicBlock::iterator I = EndInst;
1543 Instruction *Inst = --I;
1544 if (!Inst->use_empty())
1545 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
1546 if (isa<LandingPadInst>(Inst)) {
1550 BB->getInstList().erase(Inst);
1555 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1556 // and return true if the function was modified.
1558 bool SCCP::runOnFunction(Function &F) {
1559 if (skipOptnoneFunction(F))
1562 DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n");
1563 const DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1564 const DataLayout *DL = DLP ? &DLP->getDataLayout() : nullptr;
1565 const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>();
1566 SCCPSolver Solver(DL, TLI);
1568 // Mark the first block of the function as being executable.
1569 Solver.MarkBlockExecutable(F.begin());
1571 // Mark all arguments to the function as being overdefined.
1572 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1573 Solver.markAnythingOverdefined(AI);
1575 // Solve for constants.
1576 bool ResolvedUndefs = true;
1577 while (ResolvedUndefs) {
1579 DEBUG(dbgs() << "RESOLVING UNDEFs\n");
1580 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1583 bool MadeChanges = false;
1585 // If we decided that there are basic blocks that are dead in this function,
1586 // delete their contents now. Note that we cannot actually delete the blocks,
1587 // as we cannot modify the CFG of the function.
1589 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1590 if (!Solver.isBlockExecutable(BB)) {
1591 DeleteInstructionInBlock(BB);
1596 // Iterate over all of the instructions in a function, replacing them with
1597 // constants if we have found them to be of constant values.
1599 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1600 Instruction *Inst = BI++;
1601 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst))
1604 // TODO: Reconstruct structs from their elements.
1605 if (Inst->getType()->isStructTy())
1608 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1609 if (IV.isOverdefined())
1612 Constant *Const = IV.isConstant()
1613 ? IV.getConstant() : UndefValue::get(Inst->getType());
1614 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst << '\n');
1616 // Replaces all of the uses of a variable with uses of the constant.
1617 Inst->replaceAllUsesWith(Const);
1619 // Delete the instruction.
1620 Inst->eraseFromParent();
1622 // Hey, we just changed something!
1632 //===--------------------------------------------------------------------===//
1634 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1635 /// Constant Propagation.
1637 struct IPSCCP : public ModulePass {
1638 void getAnalysisUsage(AnalysisUsage &AU) const override {
1639 AU.addRequired<TargetLibraryInfo>();
1642 IPSCCP() : ModulePass(ID) {
1643 initializeIPSCCPPass(*PassRegistry::getPassRegistry());
1645 bool runOnModule(Module &M) override;
1647 } // end anonymous namespace
1649 char IPSCCP::ID = 0;
1650 INITIALIZE_PASS_BEGIN(IPSCCP, "ipsccp",
1651 "Interprocedural Sparse Conditional Constant Propagation",
1653 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
1654 INITIALIZE_PASS_END(IPSCCP, "ipsccp",
1655 "Interprocedural Sparse Conditional Constant Propagation",
1658 // createIPSCCPPass - This is the public interface to this file.
1659 ModulePass *llvm::createIPSCCPPass() {
1660 return new IPSCCP();
1664 static bool AddressIsTaken(const GlobalValue *GV) {
1665 // Delete any dead constantexpr klingons.
1666 GV->removeDeadConstantUsers();
1668 for (const Use &U : GV->uses()) {
1669 const User *UR = U.getUser();
1670 if (const StoreInst *SI = dyn_cast<StoreInst>(UR)) {
1671 if (SI->getOperand(0) == GV || SI->isVolatile())
1672 return true; // Storing addr of GV.
1673 } else if (isa<InvokeInst>(UR) || isa<CallInst>(UR)) {
1674 // Make sure we are calling the function, not passing the address.
1675 ImmutableCallSite CS(cast<Instruction>(UR));
1676 if (!CS.isCallee(&U))
1678 } else if (const LoadInst *LI = dyn_cast<LoadInst>(UR)) {
1679 if (LI->isVolatile())
1681 } else if (isa<BlockAddress>(UR)) {
1682 // blockaddress doesn't take the address of the function, it takes addr
1691 bool IPSCCP::runOnModule(Module &M) {
1692 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1693 const DataLayout *DL = DLP ? &DLP->getDataLayout() : nullptr;
1694 const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>();
1695 SCCPSolver Solver(DL, TLI);
1697 // AddressTakenFunctions - This set keeps track of the address-taken functions
1698 // that are in the input. As IPSCCP runs through and simplifies code,
1699 // functions that were address taken can end up losing their
1700 // address-taken-ness. Because of this, we keep track of their addresses from
1701 // the first pass so we can use them for the later simplification pass.
1702 SmallPtrSet<Function*, 32> AddressTakenFunctions;
1704 // Loop over all functions, marking arguments to those with their addresses
1705 // taken or that are external as overdefined.
1707 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1708 if (F->isDeclaration())
1711 // If this is a strong or ODR definition of this function, then we can
1712 // propagate information about its result into callsites of it.
1713 if (!F->mayBeOverridden())
1714 Solver.AddTrackedFunction(F);
1716 // If this function only has direct calls that we can see, we can track its
1717 // arguments and return value aggressively, and can assume it is not called
1718 // unless we see evidence to the contrary.
1719 if (F->hasLocalLinkage()) {
1720 if (AddressIsTaken(F))
1721 AddressTakenFunctions.insert(F);
1723 Solver.AddArgumentTrackedFunction(F);
1728 // Assume the function is called.
1729 Solver.MarkBlockExecutable(F->begin());
1731 // Assume nothing about the incoming arguments.
1732 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1734 Solver.markAnythingOverdefined(AI);
1737 // Loop over global variables. We inform the solver about any internal global
1738 // variables that do not have their 'addresses taken'. If they don't have
1739 // their addresses taken, we can propagate constants through them.
1740 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1742 if (!G->isConstant() && G->hasLocalLinkage() && !AddressIsTaken(G))
1743 Solver.TrackValueOfGlobalVariable(G);
1745 // Solve for constants.
1746 bool ResolvedUndefs = true;
1747 while (ResolvedUndefs) {
1750 DEBUG(dbgs() << "RESOLVING UNDEFS\n");
1751 ResolvedUndefs = false;
1752 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1753 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1756 bool MadeChanges = false;
1758 // Iterate over all of the instructions in the module, replacing them with
1759 // constants if we have found them to be of constant values.
1761 SmallVector<BasicBlock*, 512> BlocksToErase;
1763 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1764 if (Solver.isBlockExecutable(F->begin())) {
1765 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1767 if (AI->use_empty() || AI->getType()->isStructTy()) continue;
1769 // TODO: Could use getStructLatticeValueFor to find out if the entire
1770 // result is a constant and replace it entirely if so.
1772 LatticeVal IV = Solver.getLatticeValueFor(AI);
1773 if (IV.isOverdefined()) continue;
1775 Constant *CST = IV.isConstant() ?
1776 IV.getConstant() : UndefValue::get(AI->getType());
1777 DEBUG(dbgs() << "*** Arg " << *AI << " = " << *CST <<"\n");
1779 // Replaces all of the uses of a variable with uses of the
1781 AI->replaceAllUsesWith(CST);
1786 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1787 if (!Solver.isBlockExecutable(BB)) {
1788 DeleteInstructionInBlock(BB);
1791 TerminatorInst *TI = BB->getTerminator();
1792 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1793 BasicBlock *Succ = TI->getSuccessor(i);
1794 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1795 TI->getSuccessor(i)->removePredecessor(BB);
1797 if (!TI->use_empty())
1798 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1799 TI->eraseFromParent();
1801 if (&*BB != &F->front())
1802 BlocksToErase.push_back(BB);
1804 new UnreachableInst(M.getContext(), BB);
1808 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1809 Instruction *Inst = BI++;
1810 if (Inst->getType()->isVoidTy() || Inst->getType()->isStructTy())
1813 // TODO: Could use getStructLatticeValueFor to find out if the entire
1814 // result is a constant and replace it entirely if so.
1816 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1817 if (IV.isOverdefined())
1820 Constant *Const = IV.isConstant()
1821 ? IV.getConstant() : UndefValue::get(Inst->getType());
1822 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst << '\n');
1824 // Replaces all of the uses of a variable with uses of the
1826 Inst->replaceAllUsesWith(Const);
1828 // Delete the instruction.
1829 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst))
1830 Inst->eraseFromParent();
1832 // Hey, we just changed something!
1838 // Now that all instructions in the function are constant folded, erase dead
1839 // blocks, because we can now use ConstantFoldTerminator to get rid of
1841 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1842 // If there are any PHI nodes in this successor, drop entries for BB now.
1843 BasicBlock *DeadBB = BlocksToErase[i];
1844 for (Value::user_iterator UI = DeadBB->user_begin(),
1845 UE = DeadBB->user_end();
1847 // Grab the user and then increment the iterator early, as the user
1848 // will be deleted. Step past all adjacent uses from the same user.
1849 Instruction *I = dyn_cast<Instruction>(*UI);
1850 do { ++UI; } while (UI != UE && *UI == I);
1852 // Ignore blockaddress users; BasicBlock's dtor will handle them.
1855 bool Folded = ConstantFoldTerminator(I->getParent());
1857 // The constant folder may not have been able to fold the terminator
1858 // if this is a branch or switch on undef. Fold it manually as a
1859 // branch to the first successor.
1861 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1862 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1863 "Branch should be foldable!");
1864 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1865 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1867 llvm_unreachable("Didn't fold away reference to block!");
1871 // Make this an uncond branch to the first successor.
1872 TerminatorInst *TI = I->getParent()->getTerminator();
1873 BranchInst::Create(TI->getSuccessor(0), TI);
1875 // Remove entries in successor phi nodes to remove edges.
1876 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1877 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1879 // Remove the old terminator.
1880 TI->eraseFromParent();
1884 // Finally, delete the basic block.
1885 F->getBasicBlockList().erase(DeadBB);
1887 BlocksToErase.clear();
1890 // If we inferred constant or undef return values for a function, we replaced
1891 // all call uses with the inferred value. This means we don't need to bother
1892 // actually returning anything from the function. Replace all return
1893 // instructions with return undef.
1895 // Do this in two stages: first identify the functions we should process, then
1896 // actually zap their returns. This is important because we can only do this
1897 // if the address of the function isn't taken. In cases where a return is the
1898 // last use of a function, the order of processing functions would affect
1899 // whether other functions are optimizable.
1900 SmallVector<ReturnInst*, 8> ReturnsToZap;
1902 // TODO: Process multiple value ret instructions also.
1903 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1904 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1905 E = RV.end(); I != E; ++I) {
1906 Function *F = I->first;
1907 if (I->second.isOverdefined() || F->getReturnType()->isVoidTy())
1910 // We can only do this if we know that nothing else can call the function.
1911 if (!F->hasLocalLinkage() || AddressTakenFunctions.count(F))
1914 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1915 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1916 if (!isa<UndefValue>(RI->getOperand(0)))
1917 ReturnsToZap.push_back(RI);
1920 // Zap all returns which we've identified as zap to change.
1921 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) {
1922 Function *F = ReturnsToZap[i]->getParent()->getParent();
1923 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType()));
1926 // If we inferred constant or undef values for globals variables, we can
1927 // delete the global and any stores that remain to it.
1928 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1929 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1930 E = TG.end(); I != E; ++I) {
1931 GlobalVariable *GV = I->first;
1932 assert(!I->second.isOverdefined() &&
1933 "Overdefined values should have been taken out of the map!");
1934 DEBUG(dbgs() << "Found that GV '" << GV->getName() << "' is constant!\n");
1935 while (!GV->use_empty()) {
1936 StoreInst *SI = cast<StoreInst>(GV->user_back());
1937 SI->eraseFromParent();
1939 M.getGlobalList().erase(GV);