1 //===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements sparse conditional constant propagation and merging:
12 // Specifically, this:
13 // * Assumes values are constant unless proven otherwise
14 // * Assumes BasicBlocks are dead unless proven otherwise
15 // * Proves values to be constant, and replaces them with constants
16 // * Proves conditional branches to be unconditional
18 //===----------------------------------------------------------------------===//
20 #define DEBUG_TYPE "sccp"
21 #include "llvm/Transforms/Scalar.h"
22 #include "llvm/ADT/DenseMap.h"
23 #include "llvm/ADT/DenseSet.h"
24 #include "llvm/ADT/PointerIntPair.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Analysis/ConstantFolding.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/InstVisitor.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/CallSite.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetLibraryInfo.h"
40 #include "llvm/Transforms/IPO.h"
41 #include "llvm/Transforms/Utils/Local.h"
45 STATISTIC(NumInstRemoved, "Number of instructions removed");
46 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
48 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP");
49 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
50 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
53 /// LatticeVal class - This class represents the different lattice values that
54 /// an LLVM value may occupy. It is a simple class with value semantics.
58 /// undefined - This LLVM Value has no known value yet.
61 /// constant - This LLVM Value has a specific constant value.
64 /// forcedconstant - This LLVM Value was thought to be undef until
65 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged
66 /// with another (different) constant, it goes to overdefined, instead of
70 /// overdefined - This instruction is not known to be constant, and we know
75 /// Val: This stores the current lattice value along with the Constant* for
76 /// the constant if this is a 'constant' or 'forcedconstant' value.
77 PointerIntPair<Constant *, 2, LatticeValueTy> Val;
79 LatticeValueTy getLatticeValue() const {
84 LatticeVal() : Val(0, undefined) {}
86 bool isUndefined() const { return getLatticeValue() == undefined; }
87 bool isConstant() const {
88 return getLatticeValue() == constant || getLatticeValue() == forcedconstant;
90 bool isOverdefined() const { return getLatticeValue() == overdefined; }
92 Constant *getConstant() const {
93 assert(isConstant() && "Cannot get the constant of a non-constant!");
94 return Val.getPointer();
97 /// markOverdefined - Return true if this is a change in status.
98 bool markOverdefined() {
102 Val.setInt(overdefined);
106 /// markConstant - Return true if this is a change in status.
107 bool markConstant(Constant *V) {
108 if (getLatticeValue() == constant) { // Constant but not forcedconstant.
109 assert(getConstant() == V && "Marking constant with different value");
114 Val.setInt(constant);
115 assert(V && "Marking constant with NULL");
118 assert(getLatticeValue() == forcedconstant &&
119 "Cannot move from overdefined to constant!");
120 // Stay at forcedconstant if the constant is the same.
121 if (V == getConstant()) return false;
123 // Otherwise, we go to overdefined. Assumptions made based on the
124 // forced value are possibly wrong. Assuming this is another constant
125 // could expose a contradiction.
126 Val.setInt(overdefined);
131 /// getConstantInt - If this is a constant with a ConstantInt value, return it
132 /// otherwise return null.
133 ConstantInt *getConstantInt() const {
135 return dyn_cast<ConstantInt>(getConstant());
139 void markForcedConstant(Constant *V) {
140 assert(isUndefined() && "Can't force a defined value!");
141 Val.setInt(forcedconstant);
145 } // end anonymous namespace.
150 //===----------------------------------------------------------------------===//
152 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional
153 /// Constant Propagation.
155 class SCCPSolver : public InstVisitor<SCCPSolver> {
156 const DataLayout *TD;
157 const TargetLibraryInfo *TLI;
158 SmallPtrSet<BasicBlock*, 8> BBExecutable; // The BBs that are executable.
159 DenseMap<Value*, LatticeVal> ValueState; // The state each value is in.
161 /// StructValueState - This maintains ValueState for values that have
162 /// StructType, for example for formal arguments, calls, insertelement, etc.
164 DenseMap<std::pair<Value*, unsigned>, LatticeVal> StructValueState;
166 /// GlobalValue - If we are tracking any values for the contents of a global
167 /// variable, we keep a mapping from the constant accessor to the element of
168 /// the global, to the currently known value. If the value becomes
169 /// overdefined, it's entry is simply removed from this map.
170 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
172 /// TrackedRetVals - If we are tracking arguments into and the return
173 /// value out of a function, it will have an entry in this map, indicating
174 /// what the known return value for the function is.
175 DenseMap<Function*, LatticeVal> TrackedRetVals;
177 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
178 /// that return multiple values.
179 DenseMap<std::pair<Function*, unsigned>, LatticeVal> TrackedMultipleRetVals;
181 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is
182 /// represented here for efficient lookup.
183 SmallPtrSet<Function*, 16> MRVFunctionsTracked;
185 /// TrackingIncomingArguments - This is the set of functions for whose
186 /// arguments we make optimistic assumptions about and try to prove as
188 SmallPtrSet<Function*, 16> TrackingIncomingArguments;
190 /// The reason for two worklists is that overdefined is the lowest state
191 /// on the lattice, and moving things to overdefined as fast as possible
192 /// makes SCCP converge much faster.
194 /// By having a separate worklist, we accomplish this because everything
195 /// possibly overdefined will become overdefined at the soonest possible
197 SmallVector<Value*, 64> OverdefinedInstWorkList;
198 SmallVector<Value*, 64> InstWorkList;
201 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list
203 /// KnownFeasibleEdges - Entries in this set are edges which have already had
204 /// PHI nodes retriggered.
205 typedef std::pair<BasicBlock*, BasicBlock*> Edge;
206 DenseSet<Edge> KnownFeasibleEdges;
208 SCCPSolver(const DataLayout *td, const TargetLibraryInfo *tli)
209 : TD(td), TLI(tli) {}
211 /// MarkBlockExecutable - This method can be used by clients to mark all of
212 /// the blocks that are known to be intrinsically live in the processed unit.
214 /// This returns true if the block was not considered live before.
215 bool MarkBlockExecutable(BasicBlock *BB) {
216 if (!BBExecutable.insert(BB)) return false;
217 DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << "\n");
218 BBWorkList.push_back(BB); // Add the block to the work list!
222 /// TrackValueOfGlobalVariable - Clients can use this method to
223 /// inform the SCCPSolver that it should track loads and stores to the
224 /// specified global variable if it can. This is only legal to call if
225 /// performing Interprocedural SCCP.
226 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
227 // We only track the contents of scalar globals.
228 if (GV->getType()->getElementType()->isSingleValueType()) {
229 LatticeVal &IV = TrackedGlobals[GV];
230 if (!isa<UndefValue>(GV->getInitializer()))
231 IV.markConstant(GV->getInitializer());
235 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
236 /// and out of the specified function (which cannot have its address taken),
237 /// this method must be called.
238 void AddTrackedFunction(Function *F) {
239 // Add an entry, F -> undef.
240 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
241 MRVFunctionsTracked.insert(F);
242 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
243 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
246 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
249 void AddArgumentTrackedFunction(Function *F) {
250 TrackingIncomingArguments.insert(F);
253 /// Solve - Solve for constants and executable blocks.
257 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
258 /// that branches on undef values cannot reach any of their successors.
259 /// However, this is not a safe assumption. After we solve dataflow, this
260 /// method should be use to handle this. If this returns true, the solver
262 bool ResolvedUndefsIn(Function &F);
264 bool isBlockExecutable(BasicBlock *BB) const {
265 return BBExecutable.count(BB);
268 LatticeVal getLatticeValueFor(Value *V) const {
269 DenseMap<Value*, LatticeVal>::const_iterator I = ValueState.find(V);
270 assert(I != ValueState.end() && "V is not in valuemap!");
274 /// getTrackedRetVals - Get the inferred return value map.
276 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
277 return TrackedRetVals;
280 /// getTrackedGlobals - Get and return the set of inferred initializers for
281 /// global variables.
282 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
283 return TrackedGlobals;
286 void markOverdefined(Value *V) {
287 assert(!V->getType()->isStructTy() && "Should use other method");
288 markOverdefined(ValueState[V], V);
291 /// markAnythingOverdefined - Mark the specified value overdefined. This
292 /// works with both scalars and structs.
293 void markAnythingOverdefined(Value *V) {
294 if (StructType *STy = dyn_cast<StructType>(V->getType()))
295 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
296 markOverdefined(getStructValueState(V, i), V);
302 // markConstant - Make a value be marked as "constant". If the value
303 // is not already a constant, add it to the instruction work list so that
304 // the users of the instruction are updated later.
306 void markConstant(LatticeVal &IV, Value *V, Constant *C) {
307 if (!IV.markConstant(C)) return;
308 DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n');
309 if (IV.isOverdefined())
310 OverdefinedInstWorkList.push_back(V);
312 InstWorkList.push_back(V);
315 void markConstant(Value *V, Constant *C) {
316 assert(!V->getType()->isStructTy() && "Should use other method");
317 markConstant(ValueState[V], V, C);
320 void markForcedConstant(Value *V, Constant *C) {
321 assert(!V->getType()->isStructTy() && "Should use other method");
322 LatticeVal &IV = ValueState[V];
323 IV.markForcedConstant(C);
324 DEBUG(dbgs() << "markForcedConstant: " << *C << ": " << *V << '\n');
325 if (IV.isOverdefined())
326 OverdefinedInstWorkList.push_back(V);
328 InstWorkList.push_back(V);
332 // markOverdefined - Make a value be marked as "overdefined". If the
333 // value is not already overdefined, add it to the overdefined instruction
334 // work list so that the users of the instruction are updated later.
335 void markOverdefined(LatticeVal &IV, Value *V) {
336 if (!IV.markOverdefined()) return;
338 DEBUG(dbgs() << "markOverdefined: ";
339 if (Function *F = dyn_cast<Function>(V))
340 dbgs() << "Function '" << F->getName() << "'\n";
342 dbgs() << *V << '\n');
343 // Only instructions go on the work list
344 OverdefinedInstWorkList.push_back(V);
347 void mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) {
348 if (IV.isOverdefined() || MergeWithV.isUndefined())
350 if (MergeWithV.isOverdefined())
351 markOverdefined(IV, V);
352 else if (IV.isUndefined())
353 markConstant(IV, V, MergeWithV.getConstant());
354 else if (IV.getConstant() != MergeWithV.getConstant())
355 markOverdefined(IV, V);
358 void mergeInValue(Value *V, LatticeVal MergeWithV) {
359 assert(!V->getType()->isStructTy() && "Should use other method");
360 mergeInValue(ValueState[V], V, MergeWithV);
364 /// getValueState - Return the LatticeVal object that corresponds to the
365 /// value. This function handles the case when the value hasn't been seen yet
366 /// by properly seeding constants etc.
367 LatticeVal &getValueState(Value *V) {
368 assert(!V->getType()->isStructTy() && "Should use getStructValueState");
370 std::pair<DenseMap<Value*, LatticeVal>::iterator, bool> I =
371 ValueState.insert(std::make_pair(V, LatticeVal()));
372 LatticeVal &LV = I.first->second;
375 return LV; // Common case, already in the map.
377 if (Constant *C = dyn_cast<Constant>(V)) {
378 // Undef values remain undefined.
379 if (!isa<UndefValue>(V))
380 LV.markConstant(C); // Constants are constant
383 // All others are underdefined by default.
387 /// getStructValueState - Return the LatticeVal object that corresponds to the
388 /// value/field pair. This function handles the case when the value hasn't
389 /// been seen yet by properly seeding constants etc.
390 LatticeVal &getStructValueState(Value *V, unsigned i) {
391 assert(V->getType()->isStructTy() && "Should use getValueState");
392 assert(i < cast<StructType>(V->getType())->getNumElements() &&
393 "Invalid element #");
395 std::pair<DenseMap<std::pair<Value*, unsigned>, LatticeVal>::iterator,
396 bool> I = StructValueState.insert(
397 std::make_pair(std::make_pair(V, i), LatticeVal()));
398 LatticeVal &LV = I.first->second;
401 return LV; // Common case, already in the map.
403 if (Constant *C = dyn_cast<Constant>(V)) {
404 Constant *Elt = C->getAggregateElement(i);
407 LV.markOverdefined(); // Unknown sort of constant.
408 else if (isa<UndefValue>(Elt))
409 ; // Undef values remain undefined.
411 LV.markConstant(Elt); // Constants are constant.
414 // All others are underdefined by default.
419 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
420 /// work list if it is not already executable.
421 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
422 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
423 return; // This edge is already known to be executable!
425 if (!MarkBlockExecutable(Dest)) {
426 // If the destination is already executable, we just made an *edge*
427 // feasible that wasn't before. Revisit the PHI nodes in the block
428 // because they have potentially new operands.
429 DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName()
430 << " -> " << Dest->getName() << "\n");
433 for (BasicBlock::iterator I = Dest->begin();
434 (PN = dyn_cast<PHINode>(I)); ++I)
439 // getFeasibleSuccessors - Return a vector of booleans to indicate which
440 // successors are reachable from a given terminator instruction.
442 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
444 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
445 // block to the 'To' basic block is currently feasible.
447 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
449 // OperandChangedState - This method is invoked on all of the users of an
450 // instruction that was just changed state somehow. Based on this
451 // information, we need to update the specified user of this instruction.
453 void OperandChangedState(Instruction *I) {
454 if (BBExecutable.count(I->getParent())) // Inst is executable?
459 friend class InstVisitor<SCCPSolver>;
461 // visit implementations - Something changed in this instruction. Either an
462 // operand made a transition, or the instruction is newly executable. Change
463 // the value type of I to reflect these changes if appropriate.
464 void visitPHINode(PHINode &I);
467 void visitReturnInst(ReturnInst &I);
468 void visitTerminatorInst(TerminatorInst &TI);
470 void visitCastInst(CastInst &I);
471 void visitSelectInst(SelectInst &I);
472 void visitBinaryOperator(Instruction &I);
473 void visitCmpInst(CmpInst &I);
474 void visitExtractElementInst(ExtractElementInst &I);
475 void visitInsertElementInst(InsertElementInst &I);
476 void visitShuffleVectorInst(ShuffleVectorInst &I);
477 void visitExtractValueInst(ExtractValueInst &EVI);
478 void visitInsertValueInst(InsertValueInst &IVI);
479 void visitLandingPadInst(LandingPadInst &I) { markAnythingOverdefined(&I); }
481 // Instructions that cannot be folded away.
482 void visitStoreInst (StoreInst &I);
483 void visitLoadInst (LoadInst &I);
484 void visitGetElementPtrInst(GetElementPtrInst &I);
485 void visitCallInst (CallInst &I) {
488 void visitInvokeInst (InvokeInst &II) {
490 visitTerminatorInst(II);
492 void visitCallSite (CallSite CS);
493 void visitResumeInst (TerminatorInst &I) { /*returns void*/ }
494 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ }
495 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
496 void visitFenceInst (FenceInst &I) { /*returns void*/ }
497 void visitAtomicCmpXchgInst (AtomicCmpXchgInst &I) { markOverdefined(&I); }
498 void visitAtomicRMWInst (AtomicRMWInst &I) { markOverdefined(&I); }
499 void visitAllocaInst (Instruction &I) { markOverdefined(&I); }
500 void visitVAArgInst (Instruction &I) { markAnythingOverdefined(&I); }
502 void visitInstruction(Instruction &I) {
503 // If a new instruction is added to LLVM that we don't handle.
504 dbgs() << "SCCP: Don't know how to handle: " << I;
505 markAnythingOverdefined(&I); // Just in case
509 } // end anonymous namespace
512 // getFeasibleSuccessors - Return a vector of booleans to indicate which
513 // successors are reachable from a given terminator instruction.
515 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
516 SmallVector<bool, 16> &Succs) {
517 Succs.resize(TI.getNumSuccessors());
518 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
519 if (BI->isUnconditional()) {
524 LatticeVal BCValue = getValueState(BI->getCondition());
525 ConstantInt *CI = BCValue.getConstantInt();
527 // Overdefined condition variables, and branches on unfoldable constant
528 // conditions, mean the branch could go either way.
529 if (!BCValue.isUndefined())
530 Succs[0] = Succs[1] = true;
534 // Constant condition variables mean the branch can only go a single way.
535 Succs[CI->isZero()] = true;
539 if (isa<InvokeInst>(TI)) {
540 // Invoke instructions successors are always executable.
541 Succs[0] = Succs[1] = true;
545 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
546 if (!SI->getNumCases()) {
550 LatticeVal SCValue = getValueState(SI->getCondition());
551 ConstantInt *CI = SCValue.getConstantInt();
553 if (CI == 0) { // Overdefined or undefined condition?
554 // All destinations are executable!
555 if (!SCValue.isUndefined())
556 Succs.assign(TI.getNumSuccessors(), true);
560 Succs[SI->findCaseValue(CI).getSuccessorIndex()] = true;
564 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
565 if (isa<IndirectBrInst>(&TI)) {
566 // Just mark all destinations executable!
567 Succs.assign(TI.getNumSuccessors(), true);
572 dbgs() << "Unknown terminator instruction: " << TI << '\n';
574 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
578 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
579 // block to the 'To' basic block is currently feasible.
581 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
582 assert(BBExecutable.count(To) && "Dest should always be alive!");
584 // Make sure the source basic block is executable!!
585 if (!BBExecutable.count(From)) return false;
587 // Check to make sure this edge itself is actually feasible now.
588 TerminatorInst *TI = From->getTerminator();
589 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
590 if (BI->isUnconditional())
593 LatticeVal BCValue = getValueState(BI->getCondition());
595 // Overdefined condition variables mean the branch could go either way,
596 // undef conditions mean that neither edge is feasible yet.
597 ConstantInt *CI = BCValue.getConstantInt();
599 return !BCValue.isUndefined();
601 // Constant condition variables mean the branch can only go a single way.
602 return BI->getSuccessor(CI->isZero()) == To;
605 // Invoke instructions successors are always executable.
606 if (isa<InvokeInst>(TI))
609 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
610 if (SI->getNumCases() < 1)
613 LatticeVal SCValue = getValueState(SI->getCondition());
614 ConstantInt *CI = SCValue.getConstantInt();
617 return !SCValue.isUndefined();
619 return SI->findCaseValue(CI).getCaseSuccessor() == To;
622 // Just mark all destinations executable!
623 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
624 if (isa<IndirectBrInst>(TI))
628 dbgs() << "Unknown terminator instruction: " << *TI << '\n';
633 // visit Implementations - Something changed in this instruction, either an
634 // operand made a transition, or the instruction is newly executable. Change
635 // the value type of I to reflect these changes if appropriate. This method
636 // makes sure to do the following actions:
638 // 1. If a phi node merges two constants in, and has conflicting value coming
639 // from different branches, or if the PHI node merges in an overdefined
640 // value, then the PHI node becomes overdefined.
641 // 2. If a phi node merges only constants in, and they all agree on value, the
642 // PHI node becomes a constant value equal to that.
643 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
644 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
645 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
646 // 6. If a conditional branch has a value that is constant, make the selected
647 // destination executable
648 // 7. If a conditional branch has a value that is overdefined, make all
649 // successors executable.
651 void SCCPSolver::visitPHINode(PHINode &PN) {
652 // If this PN returns a struct, just mark the result overdefined.
653 // TODO: We could do a lot better than this if code actually uses this.
654 if (PN.getType()->isStructTy())
655 return markAnythingOverdefined(&PN);
657 if (getValueState(&PN).isOverdefined())
658 return; // Quick exit
660 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
661 // and slow us down a lot. Just mark them overdefined.
662 if (PN.getNumIncomingValues() > 64)
663 return markOverdefined(&PN);
665 // Look at all of the executable operands of the PHI node. If any of them
666 // are overdefined, the PHI becomes overdefined as well. If they are all
667 // constant, and they agree with each other, the PHI becomes the identical
668 // constant. If they are constant and don't agree, the PHI is overdefined.
669 // If there are no executable operands, the PHI remains undefined.
671 Constant *OperandVal = 0;
672 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
673 LatticeVal IV = getValueState(PN.getIncomingValue(i));
674 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
676 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
679 if (IV.isOverdefined()) // PHI node becomes overdefined!
680 return markOverdefined(&PN);
682 if (OperandVal == 0) { // Grab the first value.
683 OperandVal = IV.getConstant();
687 // There is already a reachable operand. If we conflict with it,
688 // then the PHI node becomes overdefined. If we agree with it, we
691 // Check to see if there are two different constants merging, if so, the PHI
692 // node is overdefined.
693 if (IV.getConstant() != OperandVal)
694 return markOverdefined(&PN);
697 // If we exited the loop, this means that the PHI node only has constant
698 // arguments that agree with each other(and OperandVal is the constant) or
699 // OperandVal is null because there are no defined incoming arguments. If
700 // this is the case, the PHI remains undefined.
703 markConstant(&PN, OperandVal); // Acquire operand value
706 void SCCPSolver::visitReturnInst(ReturnInst &I) {
707 if (I.getNumOperands() == 0) return; // ret void
709 Function *F = I.getParent()->getParent();
710 Value *ResultOp = I.getOperand(0);
712 // If we are tracking the return value of this function, merge it in.
713 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) {
714 DenseMap<Function*, LatticeVal>::iterator TFRVI =
715 TrackedRetVals.find(F);
716 if (TFRVI != TrackedRetVals.end()) {
717 mergeInValue(TFRVI->second, F, getValueState(ResultOp));
722 // Handle functions that return multiple values.
723 if (!TrackedMultipleRetVals.empty()) {
724 if (StructType *STy = dyn_cast<StructType>(ResultOp->getType()))
725 if (MRVFunctionsTracked.count(F))
726 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
727 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F,
728 getStructValueState(ResultOp, i));
733 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
734 SmallVector<bool, 16> SuccFeasible;
735 getFeasibleSuccessors(TI, SuccFeasible);
737 BasicBlock *BB = TI.getParent();
739 // Mark all feasible successors executable.
740 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
742 markEdgeExecutable(BB, TI.getSuccessor(i));
745 void SCCPSolver::visitCastInst(CastInst &I) {
746 LatticeVal OpSt = getValueState(I.getOperand(0));
747 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand
749 else if (OpSt.isConstant()) // Propagate constant value
750 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
751 OpSt.getConstant(), I.getType()));
755 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
756 // If this returns a struct, mark all elements over defined, we don't track
757 // structs in structs.
758 if (EVI.getType()->isStructTy())
759 return markAnythingOverdefined(&EVI);
761 // If this is extracting from more than one level of struct, we don't know.
762 if (EVI.getNumIndices() != 1)
763 return markOverdefined(&EVI);
765 Value *AggVal = EVI.getAggregateOperand();
766 if (AggVal->getType()->isStructTy()) {
767 unsigned i = *EVI.idx_begin();
768 LatticeVal EltVal = getStructValueState(AggVal, i);
769 mergeInValue(getValueState(&EVI), &EVI, EltVal);
771 // Otherwise, must be extracting from an array.
772 return markOverdefined(&EVI);
776 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
777 StructType *STy = dyn_cast<StructType>(IVI.getType());
779 return markOverdefined(&IVI);
781 // If this has more than one index, we can't handle it, drive all results to
783 if (IVI.getNumIndices() != 1)
784 return markAnythingOverdefined(&IVI);
786 Value *Aggr = IVI.getAggregateOperand();
787 unsigned Idx = *IVI.idx_begin();
789 // Compute the result based on what we're inserting.
790 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
791 // This passes through all values that aren't the inserted element.
793 LatticeVal EltVal = getStructValueState(Aggr, i);
794 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal);
798 Value *Val = IVI.getInsertedValueOperand();
799 if (Val->getType()->isStructTy())
800 // We don't track structs in structs.
801 markOverdefined(getStructValueState(&IVI, i), &IVI);
803 LatticeVal InVal = getValueState(Val);
804 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal);
809 void SCCPSolver::visitSelectInst(SelectInst &I) {
810 // If this select returns a struct, just mark the result overdefined.
811 // TODO: We could do a lot better than this if code actually uses this.
812 if (I.getType()->isStructTy())
813 return markAnythingOverdefined(&I);
815 LatticeVal CondValue = getValueState(I.getCondition());
816 if (CondValue.isUndefined())
819 if (ConstantInt *CondCB = CondValue.getConstantInt()) {
820 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
821 mergeInValue(&I, getValueState(OpVal));
825 // Otherwise, the condition is overdefined or a constant we can't evaluate.
826 // See if we can produce something better than overdefined based on the T/F
828 LatticeVal TVal = getValueState(I.getTrueValue());
829 LatticeVal FVal = getValueState(I.getFalseValue());
831 // select ?, C, C -> C.
832 if (TVal.isConstant() && FVal.isConstant() &&
833 TVal.getConstant() == FVal.getConstant())
834 return markConstant(&I, FVal.getConstant());
836 if (TVal.isUndefined()) // select ?, undef, X -> X.
837 return mergeInValue(&I, FVal);
838 if (FVal.isUndefined()) // select ?, X, undef -> X.
839 return mergeInValue(&I, TVal);
843 // Handle Binary Operators.
844 void SCCPSolver::visitBinaryOperator(Instruction &I) {
845 LatticeVal V1State = getValueState(I.getOperand(0));
846 LatticeVal V2State = getValueState(I.getOperand(1));
848 LatticeVal &IV = ValueState[&I];
849 if (IV.isOverdefined()) return;
851 if (V1State.isConstant() && V2State.isConstant())
852 return markConstant(IV, &I,
853 ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
854 V2State.getConstant()));
856 // If something is undef, wait for it to resolve.
857 if (!V1State.isOverdefined() && !V2State.isOverdefined())
860 // Otherwise, one of our operands is overdefined. Try to produce something
861 // better than overdefined with some tricks.
863 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
864 // operand is overdefined.
865 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
866 LatticeVal *NonOverdefVal = 0;
867 if (!V1State.isOverdefined())
868 NonOverdefVal = &V1State;
869 else if (!V2State.isOverdefined())
870 NonOverdefVal = &V2State;
873 if (NonOverdefVal->isUndefined()) {
874 // Could annihilate value.
875 if (I.getOpcode() == Instruction::And)
876 markConstant(IV, &I, Constant::getNullValue(I.getType()));
877 else if (VectorType *PT = dyn_cast<VectorType>(I.getType()))
878 markConstant(IV, &I, Constant::getAllOnesValue(PT));
881 Constant::getAllOnesValue(I.getType()));
885 if (I.getOpcode() == Instruction::And) {
887 if (NonOverdefVal->getConstant()->isNullValue())
888 return markConstant(IV, &I, NonOverdefVal->getConstant());
890 if (ConstantInt *CI = NonOverdefVal->getConstantInt())
891 if (CI->isAllOnesValue()) // X or -1 = -1
892 return markConstant(IV, &I, NonOverdefVal->getConstant());
901 // Handle ICmpInst instruction.
902 void SCCPSolver::visitCmpInst(CmpInst &I) {
903 LatticeVal V1State = getValueState(I.getOperand(0));
904 LatticeVal V2State = getValueState(I.getOperand(1));
906 LatticeVal &IV = ValueState[&I];
907 if (IV.isOverdefined()) return;
909 if (V1State.isConstant() && V2State.isConstant())
910 return markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
911 V1State.getConstant(),
912 V2State.getConstant()));
914 // If operands are still undefined, wait for it to resolve.
915 if (!V1State.isOverdefined() && !V2State.isOverdefined())
921 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
922 // TODO : SCCP does not handle vectors properly.
923 return markOverdefined(&I);
926 LatticeVal &ValState = getValueState(I.getOperand(0));
927 LatticeVal &IdxState = getValueState(I.getOperand(1));
929 if (ValState.isOverdefined() || IdxState.isOverdefined())
931 else if(ValState.isConstant() && IdxState.isConstant())
932 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
933 IdxState.getConstant()));
937 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
938 // TODO : SCCP does not handle vectors properly.
939 return markOverdefined(&I);
941 LatticeVal &ValState = getValueState(I.getOperand(0));
942 LatticeVal &EltState = getValueState(I.getOperand(1));
943 LatticeVal &IdxState = getValueState(I.getOperand(2));
945 if (ValState.isOverdefined() || EltState.isOverdefined() ||
946 IdxState.isOverdefined())
948 else if(ValState.isConstant() && EltState.isConstant() &&
949 IdxState.isConstant())
950 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
951 EltState.getConstant(),
952 IdxState.getConstant()));
953 else if (ValState.isUndefined() && EltState.isConstant() &&
954 IdxState.isConstant())
955 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
956 EltState.getConstant(),
957 IdxState.getConstant()));
961 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
962 // TODO : SCCP does not handle vectors properly.
963 return markOverdefined(&I);
965 LatticeVal &V1State = getValueState(I.getOperand(0));
966 LatticeVal &V2State = getValueState(I.getOperand(1));
967 LatticeVal &MaskState = getValueState(I.getOperand(2));
969 if (MaskState.isUndefined() ||
970 (V1State.isUndefined() && V2State.isUndefined()))
971 return; // Undefined output if mask or both inputs undefined.
973 if (V1State.isOverdefined() || V2State.isOverdefined() ||
974 MaskState.isOverdefined()) {
977 // A mix of constant/undef inputs.
978 Constant *V1 = V1State.isConstant() ?
979 V1State.getConstant() : UndefValue::get(I.getType());
980 Constant *V2 = V2State.isConstant() ?
981 V2State.getConstant() : UndefValue::get(I.getType());
982 Constant *Mask = MaskState.isConstant() ?
983 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
984 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
989 // Handle getelementptr instructions. If all operands are constants then we
990 // can turn this into a getelementptr ConstantExpr.
992 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
993 if (ValueState[&I].isOverdefined()) return;
995 SmallVector<Constant*, 8> Operands;
996 Operands.reserve(I.getNumOperands());
998 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
999 LatticeVal State = getValueState(I.getOperand(i));
1000 if (State.isUndefined())
1001 return; // Operands are not resolved yet.
1003 if (State.isOverdefined())
1004 return markOverdefined(&I);
1006 assert(State.isConstant() && "Unknown state!");
1007 Operands.push_back(State.getConstant());
1010 Constant *Ptr = Operands[0];
1011 ArrayRef<Constant *> Indices(Operands.begin() + 1, Operands.end());
1012 markConstant(&I, ConstantExpr::getGetElementPtr(Ptr, Indices));
1015 void SCCPSolver::visitStoreInst(StoreInst &SI) {
1016 // If this store is of a struct, ignore it.
1017 if (SI.getOperand(0)->getType()->isStructTy())
1020 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1023 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1024 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1025 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1027 // Get the value we are storing into the global, then merge it.
1028 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)));
1029 if (I->second.isOverdefined())
1030 TrackedGlobals.erase(I); // No need to keep tracking this!
1034 // Handle load instructions. If the operand is a constant pointer to a constant
1035 // global, we can replace the load with the loaded constant value!
1036 void SCCPSolver::visitLoadInst(LoadInst &I) {
1037 // If this load is of a struct, just mark the result overdefined.
1038 if (I.getType()->isStructTy())
1039 return markAnythingOverdefined(&I);
1041 LatticeVal PtrVal = getValueState(I.getOperand(0));
1042 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1044 LatticeVal &IV = ValueState[&I];
1045 if (IV.isOverdefined()) return;
1047 if (!PtrVal.isConstant() || I.isVolatile())
1048 return markOverdefined(IV, &I);
1050 Constant *Ptr = PtrVal.getConstant();
1052 // load null -> null
1053 if (isa<ConstantPointerNull>(Ptr) && I.getPointerAddressSpace() == 0)
1054 return markConstant(IV, &I, Constant::getNullValue(I.getType()));
1056 // Transform load (constant global) into the value loaded.
1057 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1058 if (!TrackedGlobals.empty()) {
1059 // If we are tracking this global, merge in the known value for it.
1060 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1061 TrackedGlobals.find(GV);
1062 if (It != TrackedGlobals.end()) {
1063 mergeInValue(IV, &I, It->second);
1069 // Transform load from a constant into a constant if possible.
1070 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, TD))
1071 return markConstant(IV, &I, C);
1073 // Otherwise we cannot say for certain what value this load will produce.
1075 markOverdefined(IV, &I);
1078 void SCCPSolver::visitCallSite(CallSite CS) {
1079 Function *F = CS.getCalledFunction();
1080 Instruction *I = CS.getInstruction();
1082 // The common case is that we aren't tracking the callee, either because we
1083 // are not doing interprocedural analysis or the callee is indirect, or is
1084 // external. Handle these cases first.
1085 if (F == 0 || F->isDeclaration()) {
1087 // Void return and not tracking callee, just bail.
1088 if (I->getType()->isVoidTy()) return;
1090 // Otherwise, if we have a single return value case, and if the function is
1091 // a declaration, maybe we can constant fold it.
1092 if (F && F->isDeclaration() && !I->getType()->isStructTy() &&
1093 canConstantFoldCallTo(F)) {
1095 SmallVector<Constant*, 8> Operands;
1096 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1098 LatticeVal State = getValueState(*AI);
1100 if (State.isUndefined())
1101 return; // Operands are not resolved yet.
1102 if (State.isOverdefined())
1103 return markOverdefined(I);
1104 assert(State.isConstant() && "Unknown state!");
1105 Operands.push_back(State.getConstant());
1108 // If we can constant fold this, mark the result of the call as a
1110 if (Constant *C = ConstantFoldCall(F, Operands, TLI))
1111 return markConstant(I, C);
1114 // Otherwise, we don't know anything about this call, mark it overdefined.
1115 return markAnythingOverdefined(I);
1118 // If this is a local function that doesn't have its address taken, mark its
1119 // entry block executable and merge in the actual arguments to the call into
1120 // the formal arguments of the function.
1121 if (!TrackingIncomingArguments.empty() && TrackingIncomingArguments.count(F)){
1122 MarkBlockExecutable(F->begin());
1124 // Propagate information from this call site into the callee.
1125 CallSite::arg_iterator CAI = CS.arg_begin();
1126 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1127 AI != E; ++AI, ++CAI) {
1128 // If this argument is byval, and if the function is not readonly, there
1129 // will be an implicit copy formed of the input aggregate.
1130 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1131 markOverdefined(AI);
1135 if (StructType *STy = dyn_cast<StructType>(AI->getType())) {
1136 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1137 LatticeVal CallArg = getStructValueState(*CAI, i);
1138 mergeInValue(getStructValueState(AI, i), AI, CallArg);
1141 mergeInValue(AI, getValueState(*CAI));
1146 // If this is a single/zero retval case, see if we're tracking the function.
1147 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
1148 if (!MRVFunctionsTracked.count(F))
1149 goto CallOverdefined; // Not tracking this callee.
1151 // If we are tracking this callee, propagate the result of the function
1152 // into this call site.
1153 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1154 mergeInValue(getStructValueState(I, i), I,
1155 TrackedMultipleRetVals[std::make_pair(F, i)]);
1157 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1158 if (TFRVI == TrackedRetVals.end())
1159 goto CallOverdefined; // Not tracking this callee.
1161 // If so, propagate the return value of the callee into this call result.
1162 mergeInValue(I, TFRVI->second);
1166 void SCCPSolver::Solve() {
1167 // Process the work lists until they are empty!
1168 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1169 !OverdefinedInstWorkList.empty()) {
1170 // Process the overdefined instruction's work list first, which drives other
1171 // things to overdefined more quickly.
1172 while (!OverdefinedInstWorkList.empty()) {
1173 Value *I = OverdefinedInstWorkList.pop_back_val();
1175 DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n');
1177 // "I" got into the work list because it either made the transition from
1178 // bottom to constant, or to overdefined.
1180 // Anything on this worklist that is overdefined need not be visited
1181 // since all of its users will have already been marked as overdefined
1182 // Update all of the users of this instruction's value.
1184 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1186 if (Instruction *I = dyn_cast<Instruction>(*UI))
1187 OperandChangedState(I);
1190 // Process the instruction work list.
1191 while (!InstWorkList.empty()) {
1192 Value *I = InstWorkList.pop_back_val();
1194 DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n');
1196 // "I" got into the work list because it made the transition from undef to
1199 // Anything on this worklist that is overdefined need not be visited
1200 // since all of its users will have already been marked as overdefined.
1201 // Update all of the users of this instruction's value.
1203 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined())
1204 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1206 if (Instruction *I = dyn_cast<Instruction>(*UI))
1207 OperandChangedState(I);
1210 // Process the basic block work list.
1211 while (!BBWorkList.empty()) {
1212 BasicBlock *BB = BBWorkList.back();
1213 BBWorkList.pop_back();
1215 DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n');
1217 // Notify all instructions in this basic block that they are newly
1224 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1225 /// that branches on undef values cannot reach any of their successors.
1226 /// However, this is not a safe assumption. After we solve dataflow, this
1227 /// method should be use to handle this. If this returns true, the solver
1228 /// should be rerun.
1230 /// This method handles this by finding an unresolved branch and marking it one
1231 /// of the edges from the block as being feasible, even though the condition
1232 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1233 /// CFG and only slightly pessimizes the analysis results (by marking one,
1234 /// potentially infeasible, edge feasible). This cannot usefully modify the
1235 /// constraints on the condition of the branch, as that would impact other users
1238 /// This scan also checks for values that use undefs, whose results are actually
1239 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1240 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1241 /// even if X isn't defined.
1242 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1243 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1244 if (!BBExecutable.count(BB))
1247 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1248 // Look for instructions which produce undef values.
1249 if (I->getType()->isVoidTy()) continue;
1251 if (StructType *STy = dyn_cast<StructType>(I->getType())) {
1252 // Only a few things that can be structs matter for undef.
1254 // Tracked calls must never be marked overdefined in ResolvedUndefsIn.
1255 if (CallSite CS = CallSite(I))
1256 if (Function *F = CS.getCalledFunction())
1257 if (MRVFunctionsTracked.count(F))
1260 // extractvalue and insertvalue don't need to be marked; they are
1261 // tracked as precisely as their operands.
1262 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I))
1265 // Send the results of everything else to overdefined. We could be
1266 // more precise than this but it isn't worth bothering.
1267 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1268 LatticeVal &LV = getStructValueState(I, i);
1269 if (LV.isUndefined())
1270 markOverdefined(LV, I);
1275 LatticeVal &LV = getValueState(I);
1276 if (!LV.isUndefined()) continue;
1278 // extractvalue is safe; check here because the argument is a struct.
1279 if (isa<ExtractValueInst>(I))
1282 // Compute the operand LatticeVals, for convenience below.
1283 // Anything taking a struct is conservatively assumed to require
1284 // overdefined markings.
1285 if (I->getOperand(0)->getType()->isStructTy()) {
1289 LatticeVal Op0LV = getValueState(I->getOperand(0));
1291 if (I->getNumOperands() == 2) {
1292 if (I->getOperand(1)->getType()->isStructTy()) {
1297 Op1LV = getValueState(I->getOperand(1));
1299 // If this is an instructions whose result is defined even if the input is
1300 // not fully defined, propagate the information.
1301 Type *ITy = I->getType();
1302 switch (I->getOpcode()) {
1303 case Instruction::Add:
1304 case Instruction::Sub:
1305 case Instruction::Trunc:
1306 case Instruction::FPTrunc:
1307 case Instruction::BitCast:
1308 break; // Any undef -> undef
1309 case Instruction::FSub:
1310 case Instruction::FAdd:
1311 case Instruction::FMul:
1312 case Instruction::FDiv:
1313 case Instruction::FRem:
1314 // Floating-point binary operation: be conservative.
1315 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1316 markForcedConstant(I, Constant::getNullValue(ITy));
1320 case Instruction::ZExt:
1321 case Instruction::SExt:
1322 case Instruction::FPToUI:
1323 case Instruction::FPToSI:
1324 case Instruction::FPExt:
1325 case Instruction::PtrToInt:
1326 case Instruction::IntToPtr:
1327 case Instruction::SIToFP:
1328 case Instruction::UIToFP:
1329 // undef -> 0; some outputs are impossible
1330 markForcedConstant(I, Constant::getNullValue(ITy));
1332 case Instruction::Mul:
1333 case Instruction::And:
1334 // Both operands undef -> undef
1335 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1337 // undef * X -> 0. X could be zero.
1338 // undef & X -> 0. X could be zero.
1339 markForcedConstant(I, Constant::getNullValue(ITy));
1342 case Instruction::Or:
1343 // Both operands undef -> undef
1344 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1346 // undef | X -> -1. X could be -1.
1347 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1350 case Instruction::Xor:
1351 // undef ^ undef -> 0; strictly speaking, this is not strictly
1352 // necessary, but we try to be nice to people who expect this
1353 // behavior in simple cases
1354 if (Op0LV.isUndefined() && Op1LV.isUndefined()) {
1355 markForcedConstant(I, Constant::getNullValue(ITy));
1358 // undef ^ X -> undef
1361 case Instruction::SDiv:
1362 case Instruction::UDiv:
1363 case Instruction::SRem:
1364 case Instruction::URem:
1365 // X / undef -> undef. No change.
1366 // X % undef -> undef. No change.
1367 if (Op1LV.isUndefined()) break;
1369 // undef / X -> 0. X could be maxint.
1370 // undef % X -> 0. X could be 1.
1371 markForcedConstant(I, Constant::getNullValue(ITy));
1374 case Instruction::AShr:
1375 // X >>a undef -> undef.
1376 if (Op1LV.isUndefined()) break;
1378 // undef >>a X -> all ones
1379 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1381 case Instruction::LShr:
1382 case Instruction::Shl:
1383 // X << undef -> undef.
1384 // X >> undef -> undef.
1385 if (Op1LV.isUndefined()) break;
1389 markForcedConstant(I, Constant::getNullValue(ITy));
1391 case Instruction::Select:
1392 Op1LV = getValueState(I->getOperand(1));
1393 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1394 if (Op0LV.isUndefined()) {
1395 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1396 Op1LV = getValueState(I->getOperand(2));
1397 } else if (Op1LV.isUndefined()) {
1398 // c ? undef : undef -> undef. No change.
1399 Op1LV = getValueState(I->getOperand(2));
1400 if (Op1LV.isUndefined())
1402 // Otherwise, c ? undef : x -> x.
1404 // Leave Op1LV as Operand(1)'s LatticeValue.
1407 if (Op1LV.isConstant())
1408 markForcedConstant(I, Op1LV.getConstant());
1412 case Instruction::Load:
1413 // A load here means one of two things: a load of undef from a global,
1414 // a load from an unknown pointer. Either way, having it return undef
1417 case Instruction::ICmp:
1418 // X == undef -> undef. Other comparisons get more complicated.
1419 if (cast<ICmpInst>(I)->isEquality())
1423 case Instruction::Call:
1424 case Instruction::Invoke: {
1425 // There are two reasons a call can have an undef result
1426 // 1. It could be tracked.
1427 // 2. It could be constant-foldable.
1428 // Because of the way we solve return values, tracked calls must
1429 // never be marked overdefined in ResolvedUndefsIn.
1430 if (Function *F = CallSite(I).getCalledFunction())
1431 if (TrackedRetVals.count(F))
1434 // If the call is constant-foldable, we mark it overdefined because
1435 // we do not know what return values are valid.
1440 // If we don't know what should happen here, conservatively mark it
1447 // Check to see if we have a branch or switch on an undefined value. If so
1448 // we force the branch to go one way or the other to make the successor
1449 // values live. It doesn't really matter which way we force it.
1450 TerminatorInst *TI = BB->getTerminator();
1451 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1452 if (!BI->isConditional()) continue;
1453 if (!getValueState(BI->getCondition()).isUndefined())
1456 // If the input to SCCP is actually branch on undef, fix the undef to
1458 if (isa<UndefValue>(BI->getCondition())) {
1459 BI->setCondition(ConstantInt::getFalse(BI->getContext()));
1460 markEdgeExecutable(BB, TI->getSuccessor(1));
1464 // Otherwise, it is a branch on a symbolic value which is currently
1465 // considered to be undef. Handle this by forcing the input value to the
1467 markForcedConstant(BI->getCondition(),
1468 ConstantInt::getFalse(TI->getContext()));
1472 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1473 if (!SI->getNumCases())
1475 if (!getValueState(SI->getCondition()).isUndefined())
1478 // If the input to SCCP is actually switch on undef, fix the undef to
1479 // the first constant.
1480 if (isa<UndefValue>(SI->getCondition())) {
1481 SI->setCondition(SI->case_begin().getCaseValue());
1482 markEdgeExecutable(BB, SI->case_begin().getCaseSuccessor());
1486 markForcedConstant(SI->getCondition(), SI->case_begin().getCaseValue());
1496 //===--------------------------------------------------------------------===//
1498 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1499 /// Sparse Conditional Constant Propagator.
1501 struct SCCP : public FunctionPass {
1502 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1503 AU.addRequired<TargetLibraryInfo>();
1505 static char ID; // Pass identification, replacement for typeid
1506 SCCP() : FunctionPass(ID) {
1507 initializeSCCPPass(*PassRegistry::getPassRegistry());
1510 // runOnFunction - Run the Sparse Conditional Constant Propagation
1511 // algorithm, and return true if the function was modified.
1513 bool runOnFunction(Function &F);
1515 } // end anonymous namespace
1518 INITIALIZE_PASS(SCCP, "sccp",
1519 "Sparse Conditional Constant Propagation", false, false)
1521 // createSCCPPass - This is the public interface to this file.
1522 FunctionPass *llvm::createSCCPPass() {
1526 static void DeleteInstructionInBlock(BasicBlock *BB) {
1527 DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
1530 // Check to see if there are non-terminating instructions to delete.
1531 if (isa<TerminatorInst>(BB->begin()))
1534 // Delete the instructions backwards, as it has a reduced likelihood of having
1535 // to update as many def-use and use-def chains.
1536 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
1537 while (EndInst != BB->begin()) {
1538 // Delete the next to last instruction.
1539 BasicBlock::iterator I = EndInst;
1540 Instruction *Inst = --I;
1541 if (!Inst->use_empty())
1542 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
1543 if (isa<LandingPadInst>(Inst)) {
1547 BB->getInstList().erase(Inst);
1552 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1553 // and return true if the function was modified.
1555 bool SCCP::runOnFunction(Function &F) {
1556 DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n");
1557 const DataLayout *TD = getAnalysisIfAvailable<DataLayout>();
1558 const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>();
1559 SCCPSolver Solver(TD, TLI);
1561 // Mark the first block of the function as being executable.
1562 Solver.MarkBlockExecutable(F.begin());
1564 // Mark all arguments to the function as being overdefined.
1565 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1566 Solver.markAnythingOverdefined(AI);
1568 // Solve for constants.
1569 bool ResolvedUndefs = true;
1570 while (ResolvedUndefs) {
1572 DEBUG(dbgs() << "RESOLVING UNDEFs\n");
1573 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1576 bool MadeChanges = false;
1578 // If we decided that there are basic blocks that are dead in this function,
1579 // delete their contents now. Note that we cannot actually delete the blocks,
1580 // as we cannot modify the CFG of the function.
1582 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1583 if (!Solver.isBlockExecutable(BB)) {
1584 DeleteInstructionInBlock(BB);
1589 // Iterate over all of the instructions in a function, replacing them with
1590 // constants if we have found them to be of constant values.
1592 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1593 Instruction *Inst = BI++;
1594 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst))
1597 // TODO: Reconstruct structs from their elements.
1598 if (Inst->getType()->isStructTy())
1601 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1602 if (IV.isOverdefined())
1605 Constant *Const = IV.isConstant()
1606 ? IV.getConstant() : UndefValue::get(Inst->getType());
1607 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst);
1609 // Replaces all of the uses of a variable with uses of the constant.
1610 Inst->replaceAllUsesWith(Const);
1612 // Delete the instruction.
1613 Inst->eraseFromParent();
1615 // Hey, we just changed something!
1625 //===--------------------------------------------------------------------===//
1627 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1628 /// Constant Propagation.
1630 struct IPSCCP : public ModulePass {
1631 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1632 AU.addRequired<TargetLibraryInfo>();
1635 IPSCCP() : ModulePass(ID) {
1636 initializeIPSCCPPass(*PassRegistry::getPassRegistry());
1638 bool runOnModule(Module &M);
1640 } // end anonymous namespace
1642 char IPSCCP::ID = 0;
1643 INITIALIZE_PASS_BEGIN(IPSCCP, "ipsccp",
1644 "Interprocedural Sparse Conditional Constant Propagation",
1646 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
1647 INITIALIZE_PASS_END(IPSCCP, "ipsccp",
1648 "Interprocedural Sparse Conditional Constant Propagation",
1651 // createIPSCCPPass - This is the public interface to this file.
1652 ModulePass *llvm::createIPSCCPPass() {
1653 return new IPSCCP();
1657 static bool AddressIsTaken(const GlobalValue *GV) {
1658 // Delete any dead constantexpr klingons.
1659 GV->removeDeadConstantUsers();
1661 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
1663 const User *U = *UI;
1664 if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
1665 if (SI->getOperand(0) == GV || SI->isVolatile())
1666 return true; // Storing addr of GV.
1667 } else if (isa<InvokeInst>(U) || isa<CallInst>(U)) {
1668 // Make sure we are calling the function, not passing the address.
1669 ImmutableCallSite CS(cast<Instruction>(U));
1670 if (!CS.isCallee(UI))
1672 } else if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
1673 if (LI->isVolatile())
1675 } else if (isa<BlockAddress>(U)) {
1676 // blockaddress doesn't take the address of the function, it takes addr
1685 bool IPSCCP::runOnModule(Module &M) {
1686 const DataLayout *TD = getAnalysisIfAvailable<DataLayout>();
1687 const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>();
1688 SCCPSolver Solver(TD, TLI);
1690 // AddressTakenFunctions - This set keeps track of the address-taken functions
1691 // that are in the input. As IPSCCP runs through and simplifies code,
1692 // functions that were address taken can end up losing their
1693 // address-taken-ness. Because of this, we keep track of their addresses from
1694 // the first pass so we can use them for the later simplification pass.
1695 SmallPtrSet<Function*, 32> AddressTakenFunctions;
1697 // Loop over all functions, marking arguments to those with their addresses
1698 // taken or that are external as overdefined.
1700 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1701 if (F->isDeclaration())
1704 // If this is a strong or ODR definition of this function, then we can
1705 // propagate information about its result into callsites of it.
1706 if (!F->mayBeOverridden())
1707 Solver.AddTrackedFunction(F);
1709 // If this function only has direct calls that we can see, we can track its
1710 // arguments and return value aggressively, and can assume it is not called
1711 // unless we see evidence to the contrary.
1712 if (F->hasLocalLinkage()) {
1713 if (AddressIsTaken(F))
1714 AddressTakenFunctions.insert(F);
1716 Solver.AddArgumentTrackedFunction(F);
1721 // Assume the function is called.
1722 Solver.MarkBlockExecutable(F->begin());
1724 // Assume nothing about the incoming arguments.
1725 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1727 Solver.markAnythingOverdefined(AI);
1730 // Loop over global variables. We inform the solver about any internal global
1731 // variables that do not have their 'addresses taken'. If they don't have
1732 // their addresses taken, we can propagate constants through them.
1733 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1735 if (!G->isConstant() && G->hasLocalLinkage() && !AddressIsTaken(G))
1736 Solver.TrackValueOfGlobalVariable(G);
1738 // Solve for constants.
1739 bool ResolvedUndefs = true;
1740 while (ResolvedUndefs) {
1743 DEBUG(dbgs() << "RESOLVING UNDEFS\n");
1744 ResolvedUndefs = false;
1745 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1746 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1749 bool MadeChanges = false;
1751 // Iterate over all of the instructions in the module, replacing them with
1752 // constants if we have found them to be of constant values.
1754 SmallVector<BasicBlock*, 512> BlocksToErase;
1756 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1757 if (Solver.isBlockExecutable(F->begin())) {
1758 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1760 if (AI->use_empty() || AI->getType()->isStructTy()) continue;
1762 // TODO: Could use getStructLatticeValueFor to find out if the entire
1763 // result is a constant and replace it entirely if so.
1765 LatticeVal IV = Solver.getLatticeValueFor(AI);
1766 if (IV.isOverdefined()) continue;
1768 Constant *CST = IV.isConstant() ?
1769 IV.getConstant() : UndefValue::get(AI->getType());
1770 DEBUG(dbgs() << "*** Arg " << *AI << " = " << *CST <<"\n");
1772 // Replaces all of the uses of a variable with uses of the
1774 AI->replaceAllUsesWith(CST);
1779 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1780 if (!Solver.isBlockExecutable(BB)) {
1781 DeleteInstructionInBlock(BB);
1784 TerminatorInst *TI = BB->getTerminator();
1785 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1786 BasicBlock *Succ = TI->getSuccessor(i);
1787 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1788 TI->getSuccessor(i)->removePredecessor(BB);
1790 if (!TI->use_empty())
1791 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1792 TI->eraseFromParent();
1794 if (&*BB != &F->front())
1795 BlocksToErase.push_back(BB);
1797 new UnreachableInst(M.getContext(), BB);
1801 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1802 Instruction *Inst = BI++;
1803 if (Inst->getType()->isVoidTy() || Inst->getType()->isStructTy())
1806 // TODO: Could use getStructLatticeValueFor to find out if the entire
1807 // result is a constant and replace it entirely if so.
1809 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1810 if (IV.isOverdefined())
1813 Constant *Const = IV.isConstant()
1814 ? IV.getConstant() : UndefValue::get(Inst->getType());
1815 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst);
1817 // Replaces all of the uses of a variable with uses of the
1819 Inst->replaceAllUsesWith(Const);
1821 // Delete the instruction.
1822 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst))
1823 Inst->eraseFromParent();
1825 // Hey, we just changed something!
1831 // Now that all instructions in the function are constant folded, erase dead
1832 // blocks, because we can now use ConstantFoldTerminator to get rid of
1834 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1835 // If there are any PHI nodes in this successor, drop entries for BB now.
1836 BasicBlock *DeadBB = BlocksToErase[i];
1837 for (Value::use_iterator UI = DeadBB->use_begin(), UE = DeadBB->use_end();
1839 // Grab the user and then increment the iterator early, as the user
1840 // will be deleted. Step past all adjacent uses from the same user.
1841 Instruction *I = dyn_cast<Instruction>(*UI);
1842 do { ++UI; } while (UI != UE && *UI == I);
1844 // Ignore blockaddress users; BasicBlock's dtor will handle them.
1847 bool Folded = ConstantFoldTerminator(I->getParent());
1849 // The constant folder may not have been able to fold the terminator
1850 // if this is a branch or switch on undef. Fold it manually as a
1851 // branch to the first successor.
1853 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1854 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1855 "Branch should be foldable!");
1856 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1857 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1859 llvm_unreachable("Didn't fold away reference to block!");
1863 // Make this an uncond branch to the first successor.
1864 TerminatorInst *TI = I->getParent()->getTerminator();
1865 BranchInst::Create(TI->getSuccessor(0), TI);
1867 // Remove entries in successor phi nodes to remove edges.
1868 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1869 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1871 // Remove the old terminator.
1872 TI->eraseFromParent();
1876 // Finally, delete the basic block.
1877 F->getBasicBlockList().erase(DeadBB);
1879 BlocksToErase.clear();
1882 // If we inferred constant or undef return values for a function, we replaced
1883 // all call uses with the inferred value. This means we don't need to bother
1884 // actually returning anything from the function. Replace all return
1885 // instructions with return undef.
1887 // Do this in two stages: first identify the functions we should process, then
1888 // actually zap their returns. This is important because we can only do this
1889 // if the address of the function isn't taken. In cases where a return is the
1890 // last use of a function, the order of processing functions would affect
1891 // whether other functions are optimizable.
1892 SmallVector<ReturnInst*, 8> ReturnsToZap;
1894 // TODO: Process multiple value ret instructions also.
1895 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1896 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1897 E = RV.end(); I != E; ++I) {
1898 Function *F = I->first;
1899 if (I->second.isOverdefined() || F->getReturnType()->isVoidTy())
1902 // We can only do this if we know that nothing else can call the function.
1903 if (!F->hasLocalLinkage() || AddressTakenFunctions.count(F))
1906 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1907 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1908 if (!isa<UndefValue>(RI->getOperand(0)))
1909 ReturnsToZap.push_back(RI);
1912 // Zap all returns which we've identified as zap to change.
1913 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) {
1914 Function *F = ReturnsToZap[i]->getParent()->getParent();
1915 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType()));
1918 // If we inferred constant or undef values for globals variables, we can
1919 // delete the global and any stores that remain to it.
1920 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1921 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1922 E = TG.end(); I != E; ++I) {
1923 GlobalVariable *GV = I->first;
1924 assert(!I->second.isOverdefined() &&
1925 "Overdefined values should have been taken out of the map!");
1926 DEBUG(dbgs() << "Found that GV '" << GV->getName() << "' is constant!\n");
1927 while (!GV->use_empty()) {
1928 StoreInst *SI = cast<StoreInst>(GV->use_back());
1929 SI->eraseFromParent();
1931 M.getGlobalList().erase(GV);