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
19 // * This pass has a habit of making definitions be dead. It is a good idea
20 // to to run a DCE pass sometime after running this pass.
22 //===----------------------------------------------------------------------===//
24 #define DEBUG_TYPE "sccp"
25 #include "llvm/Transforms/Scalar.h"
26 #include "llvm/Transforms/IPO.h"
27 #include "llvm/Constants.h"
28 #include "llvm/DerivedTypes.h"
29 #include "llvm/Instructions.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Analysis/ConstantFolding.h"
33 #include "llvm/Analysis/MemoryBuiltins.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/Transforms/Utils/Local.h"
36 #include "llvm/Support/CallSite.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/InstVisitor.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/ADT/DenseMap.h"
42 #include "llvm/ADT/DenseSet.h"
43 #include "llvm/ADT/SmallSet.h"
44 #include "llvm/ADT/SmallVector.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/STLExtras.h"
51 STATISTIC(NumInstRemoved, "Number of instructions removed");
52 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
54 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP");
55 STATISTIC(IPNumDeadBlocks , "Number of basic blocks unreachable by IPSCCP");
56 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
57 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
60 /// LatticeVal class - This class represents the different lattice values that
61 /// an LLVM value may occupy. It is a simple class with value semantics.
65 /// undefined - This LLVM Value has no known value yet.
68 /// constant - This LLVM Value has a specific constant value.
71 /// forcedconstant - This LLVM Value was thought to be undef until
72 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged
73 /// with another (different) constant, it goes to overdefined, instead of
77 /// overdefined - This instruction is not known to be constant, and we know
80 } LatticeValue; // The current lattice position
82 Constant *ConstantVal; // If Constant value, the current value
84 inline LatticeVal() : LatticeValue(undefined), ConstantVal(0) {}
86 // markOverdefined - Return true if this is a new status to be in...
87 inline bool markOverdefined() {
88 if (LatticeValue != overdefined) {
89 LatticeValue = overdefined;
95 // markConstant - Return true if this is a new status for us.
96 inline bool markConstant(Constant *V) {
97 if (LatticeValue != constant) {
98 if (LatticeValue == undefined) {
99 LatticeValue = constant;
100 assert(V && "Marking constant with NULL");
103 assert(LatticeValue == forcedconstant &&
104 "Cannot move from overdefined to constant!");
105 // Stay at forcedconstant if the constant is the same.
106 if (V == ConstantVal) return false;
108 // Otherwise, we go to overdefined. Assumptions made based on the
109 // forced value are possibly wrong. Assuming this is another constant
110 // could expose a contradiction.
111 LatticeValue = overdefined;
115 assert(ConstantVal == V && "Marking constant with different value");
120 inline void markForcedConstant(Constant *V) {
121 assert(LatticeValue == undefined && "Can't force a defined value!");
122 LatticeValue = forcedconstant;
126 inline bool isUndefined() const { return LatticeValue == undefined; }
127 inline bool isConstant() const {
128 return LatticeValue == constant || LatticeValue == forcedconstant;
130 inline bool isOverdefined() const { return LatticeValue == overdefined; }
132 inline Constant *getConstant() const {
133 assert(isConstant() && "Cannot get the constant of a non-constant!");
138 //===----------------------------------------------------------------------===//
140 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional
141 /// Constant Propagation.
143 class SCCPSolver : public InstVisitor<SCCPSolver> {
144 LLVMContext *Context;
145 DenseSet<BasicBlock*> BBExecutable;// The basic blocks that are executable
146 std::map<Value*, LatticeVal> ValueState; // The state each value is in.
148 /// GlobalValue - If we are tracking any values for the contents of a global
149 /// variable, we keep a mapping from the constant accessor to the element of
150 /// the global, to the currently known value. If the value becomes
151 /// overdefined, it's entry is simply removed from this map.
152 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
154 /// TrackedRetVals - If we are tracking arguments into and the return
155 /// value out of a function, it will have an entry in this map, indicating
156 /// what the known return value for the function is.
157 DenseMap<Function*, LatticeVal> TrackedRetVals;
159 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
160 /// that return multiple values.
161 DenseMap<std::pair<Function*, unsigned>, LatticeVal> TrackedMultipleRetVals;
163 // The reason for two worklists is that overdefined is the lowest state
164 // on the lattice, and moving things to overdefined as fast as possible
165 // makes SCCP converge much faster.
166 // By having a separate worklist, we accomplish this because everything
167 // possibly overdefined will become overdefined at the soonest possible
169 SmallVector<Value*, 64> OverdefinedInstWorkList;
170 SmallVector<Value*, 64> InstWorkList;
173 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list
175 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
176 /// overdefined, despite the fact that the PHI node is overdefined.
177 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
179 /// KnownFeasibleEdges - Entries in this set are edges which have already had
180 /// PHI nodes retriggered.
181 typedef std::pair<BasicBlock*, BasicBlock*> Edge;
182 DenseSet<Edge> KnownFeasibleEdges;
184 void setContext(LLVMContext *C) { Context = C; }
186 /// MarkBlockExecutable - This method can be used by clients to mark all of
187 /// the blocks that are known to be intrinsically live in the processed unit.
188 void MarkBlockExecutable(BasicBlock *BB) {
189 DEBUG(errs() << "Marking Block Executable: " << BB->getName() << "\n");
190 BBExecutable.insert(BB); // Basic block is executable!
191 BBWorkList.push_back(BB); // Add the block to the work list!
194 /// TrackValueOfGlobalVariable - Clients can use this method to
195 /// inform the SCCPSolver that it should track loads and stores to the
196 /// specified global variable if it can. This is only legal to call if
197 /// performing Interprocedural SCCP.
198 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
199 const Type *ElTy = GV->getType()->getElementType();
200 if (ElTy->isFirstClassType()) {
201 LatticeVal &IV = TrackedGlobals[GV];
202 if (!isa<UndefValue>(GV->getInitializer()))
203 IV.markConstant(GV->getInitializer());
207 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
208 /// and out of the specified function (which cannot have its address taken),
209 /// this method must be called.
210 void AddTrackedFunction(Function *F) {
211 assert(F->hasLocalLinkage() && "Can only track internal functions!");
212 // Add an entry, F -> undef.
213 if (const StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
214 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
215 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
218 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
221 /// Solve - Solve for constants and executable blocks.
225 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
226 /// that branches on undef values cannot reach any of their successors.
227 /// However, this is not a safe assumption. After we solve dataflow, this
228 /// method should be use to handle this. If this returns true, the solver
230 bool ResolvedUndefsIn(Function &F);
232 bool isBlockExecutable(BasicBlock *BB) const {
233 return BBExecutable.count(BB);
236 /// getValueMapping - Once we have solved for constants, return the mapping of
237 /// LLVM values to LatticeVals.
238 std::map<Value*, LatticeVal> &getValueMapping() {
242 /// getTrackedRetVals - Get the inferred return value map.
244 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
245 return TrackedRetVals;
248 /// getTrackedGlobals - Get and return the set of inferred initializers for
249 /// global variables.
250 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
251 return TrackedGlobals;
254 inline void markOverdefined(Value *V) {
255 markOverdefined(ValueState[V], V);
259 // markConstant - Make a value be marked as "constant". If the value
260 // is not already a constant, add it to the instruction work list so that
261 // the users of the instruction are updated later.
263 inline void markConstant(LatticeVal &IV, Value *V, Constant *C) {
264 if (IV.markConstant(C)) {
265 DEBUG(errs() << "markConstant: " << *C << ": " << *V << '\n');
266 InstWorkList.push_back(V);
270 inline void markForcedConstant(LatticeVal &IV, Value *V, Constant *C) {
271 IV.markForcedConstant(C);
272 DEBUG(errs() << "markForcedConstant: " << *C << ": " << *V << '\n');
273 InstWorkList.push_back(V);
276 inline void markConstant(Value *V, Constant *C) {
277 markConstant(ValueState[V], V, C);
280 // markOverdefined - Make a value be marked as "overdefined". If the
281 // value is not already overdefined, add it to the overdefined instruction
282 // work list so that the users of the instruction are updated later.
283 inline void markOverdefined(LatticeVal &IV, Value *V) {
284 if (IV.markOverdefined()) {
285 DEBUG(errs() << "markOverdefined: ";
286 if (Function *F = dyn_cast<Function>(V))
287 errs() << "Function '" << F->getName() << "'\n";
289 errs() << *V << '\n');
290 // Only instructions go on the work list
291 OverdefinedInstWorkList.push_back(V);
295 inline void mergeInValue(LatticeVal &IV, Value *V, LatticeVal &MergeWithV) {
296 if (IV.isOverdefined() || MergeWithV.isUndefined())
298 if (MergeWithV.isOverdefined())
299 markOverdefined(IV, V);
300 else if (IV.isUndefined())
301 markConstant(IV, V, MergeWithV.getConstant());
302 else if (IV.getConstant() != MergeWithV.getConstant())
303 markOverdefined(IV, V);
306 inline void mergeInValue(Value *V, LatticeVal &MergeWithV) {
307 return mergeInValue(ValueState[V], V, MergeWithV);
311 // getValueState - Return the LatticeVal object that corresponds to the value.
312 // This function is necessary because not all values should start out in the
313 // underdefined state... Argument's should be overdefined, and
314 // constants should be marked as constants. If a value is not known to be an
315 // Instruction object, then use this accessor to get its value from the map.
317 inline LatticeVal &getValueState(Value *V) {
318 std::map<Value*, LatticeVal>::iterator I = ValueState.find(V);
319 if (I != ValueState.end()) return I->second; // Common case, in the map
321 if (Constant *C = dyn_cast<Constant>(V)) {
322 if (isa<UndefValue>(V)) {
323 // Nothing to do, remain undefined.
325 LatticeVal &LV = ValueState[C];
326 LV.markConstant(C); // Constants are constant
330 // All others are underdefined by default...
331 return ValueState[V];
334 // markEdgeExecutable - Mark a basic block as executable, adding it to the BB
335 // work list if it is not already executable...
337 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
338 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
339 return; // This edge is already known to be executable!
341 if (BBExecutable.count(Dest)) {
342 DEBUG(errs() << "Marking Edge Executable: " << Source->getName()
343 << " -> " << Dest->getName() << "\n");
345 // The destination is already executable, but we just made an edge
346 // feasible that wasn't before. Revisit the PHI nodes in the block
347 // because they have potentially new operands.
348 for (BasicBlock::iterator I = Dest->begin(); isa<PHINode>(I); ++I)
349 visitPHINode(*cast<PHINode>(I));
352 MarkBlockExecutable(Dest);
356 // getFeasibleSuccessors - Return a vector of booleans to indicate which
357 // successors are reachable from a given terminator instruction.
359 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
361 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
362 // block to the 'To' basic block is currently feasible...
364 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
366 // OperandChangedState - This method is invoked on all of the users of an
367 // instruction that was just changed state somehow.... Based on this
368 // information, we need to update the specified user of this instruction.
370 void OperandChangedState(User *U) {
371 // Only instructions use other variable values!
372 Instruction &I = cast<Instruction>(*U);
373 if (BBExecutable.count(I.getParent())) // Inst is executable?
378 friend class InstVisitor<SCCPSolver>;
380 // visit implementations - Something changed in this instruction... Either an
381 // operand made a transition, or the instruction is newly executable. Change
382 // the value type of I to reflect these changes if appropriate.
384 void visitPHINode(PHINode &I);
387 void visitReturnInst(ReturnInst &I);
388 void visitTerminatorInst(TerminatorInst &TI);
390 void visitCastInst(CastInst &I);
391 void visitSelectInst(SelectInst &I);
392 void visitBinaryOperator(Instruction &I);
393 void visitCmpInst(CmpInst &I);
394 void visitExtractElementInst(ExtractElementInst &I);
395 void visitInsertElementInst(InsertElementInst &I);
396 void visitShuffleVectorInst(ShuffleVectorInst &I);
397 void visitExtractValueInst(ExtractValueInst &EVI);
398 void visitInsertValueInst(InsertValueInst &IVI);
400 // Instructions that cannot be folded away...
401 void visitStoreInst (Instruction &I);
402 void visitLoadInst (LoadInst &I);
403 void visitGetElementPtrInst(GetElementPtrInst &I);
404 void visitCallInst (CallInst &I) {
407 visitCallSite(CallSite::get(&I));
409 void visitInvokeInst (InvokeInst &II) {
410 visitCallSite(CallSite::get(&II));
411 visitTerminatorInst(II);
413 void visitCallSite (CallSite CS);
414 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ }
415 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
416 void visitAllocaInst (Instruction &I) { markOverdefined(&I); }
417 void visitVANextInst (Instruction &I) { markOverdefined(&I); }
418 void visitVAArgInst (Instruction &I) { markOverdefined(&I); }
420 void visitInstruction(Instruction &I) {
421 // If a new instruction is added to LLVM that we don't handle...
422 errs() << "SCCP: Don't know how to handle: " << I;
423 markOverdefined(&I); // Just in case
427 } // end anonymous namespace
430 // getFeasibleSuccessors - Return a vector of booleans to indicate which
431 // successors are reachable from a given terminator instruction.
433 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
434 SmallVector<bool, 16> &Succs) {
435 Succs.resize(TI.getNumSuccessors());
436 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
437 if (BI->isUnconditional()) {
440 LatticeVal &BCValue = getValueState(BI->getCondition());
441 if (BCValue.isOverdefined() ||
442 (BCValue.isConstant() && !isa<ConstantInt>(BCValue.getConstant()))) {
443 // Overdefined condition variables, and branches on unfoldable constant
444 // conditions, mean the branch could go either way.
445 Succs[0] = Succs[1] = true;
446 } else if (BCValue.isConstant()) {
447 // Constant condition variables mean the branch can only go a single way
448 Succs[BCValue.getConstant() == ConstantInt::getFalse(*Context)] = true;
454 if (isa<InvokeInst>(&TI)) {
455 // Invoke instructions successors are always executable.
456 Succs[0] = Succs[1] = true;
460 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
461 LatticeVal &SCValue = getValueState(SI->getCondition());
462 if (SCValue.isOverdefined() || // Overdefined condition?
463 (SCValue.isConstant() && !isa<ConstantInt>(SCValue.getConstant()))) {
464 // All destinations are executable!
465 Succs.assign(TI.getNumSuccessors(), true);
466 } else if (SCValue.isConstant())
467 Succs[SI->findCaseValue(cast<ConstantInt>(SCValue.getConstant()))] = true;
471 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
472 if (isa<IndirectBrInst>(&TI)) {
473 // Just mark all destinations executable!
474 Succs.assign(TI.getNumSuccessors(), true);
479 errs() << "Unknown terminator instruction: " << TI << '\n';
481 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
485 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
486 // block to the 'To' basic block is currently feasible...
488 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
489 assert(BBExecutable.count(To) && "Dest should always be alive!");
491 // Make sure the source basic block is executable!!
492 if (!BBExecutable.count(From)) return false;
494 // Check to make sure this edge itself is actually feasible now...
495 TerminatorInst *TI = From->getTerminator();
496 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
497 if (BI->isUnconditional())
500 LatticeVal &BCValue = getValueState(BI->getCondition());
501 if (BCValue.isOverdefined()) {
502 // Overdefined condition variables mean the branch could go either way.
504 } else if (BCValue.isConstant()) {
505 // Not branching on an evaluatable constant?
506 if (!isa<ConstantInt>(BCValue.getConstant())) return true;
508 // Constant condition variables mean the branch can only go a single way
509 return BI->getSuccessor(BCValue.getConstant() ==
510 ConstantInt::getFalse(*Context)) == To;
515 // Invoke instructions successors are always executable.
516 if (isa<InvokeInst>(TI))
519 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
520 LatticeVal &SCValue = getValueState(SI->getCondition());
521 if (SCValue.isOverdefined()) { // Overdefined condition?
522 // All destinations are executable!
524 } else if (SCValue.isConstant()) {
525 Constant *CPV = SCValue.getConstant();
526 if (!isa<ConstantInt>(CPV))
527 return true; // not a foldable constant?
529 // Make sure to skip the "default value" which isn't a value
530 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i)
531 if (SI->getSuccessorValue(i) == CPV) // Found the taken branch...
532 return SI->getSuccessor(i) == To;
534 // Constant value not equal to any of the branches... must execute
535 // default branch then...
536 return SI->getDefaultDest() == To;
541 // Just mark all destinations executable!
542 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
543 if (isa<IndirectBrInst>(&TI))
547 errs() << "Unknown terminator instruction: " << *TI << '\n';
552 // visit Implementations - Something changed in this instruction... Either an
553 // operand made a transition, or the instruction is newly executable. Change
554 // the value type of I to reflect these changes if appropriate. This method
555 // makes sure to do the following actions:
557 // 1. If a phi node merges two constants in, and has conflicting value coming
558 // from different branches, or if the PHI node merges in an overdefined
559 // value, then the PHI node becomes overdefined.
560 // 2. If a phi node merges only constants in, and they all agree on value, the
561 // PHI node becomes a constant value equal to that.
562 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
563 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
564 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
565 // 6. If a conditional branch has a value that is constant, make the selected
566 // destination executable
567 // 7. If a conditional branch has a value that is overdefined, make all
568 // successors executable.
570 void SCCPSolver::visitPHINode(PHINode &PN) {
571 LatticeVal &PNIV = getValueState(&PN);
572 if (PNIV.isOverdefined()) {
573 // There may be instructions using this PHI node that are not overdefined
574 // themselves. If so, make sure that they know that the PHI node operand
576 std::multimap<PHINode*, Instruction*>::iterator I, E;
577 tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN);
579 SmallVector<Instruction*, 16> Users;
580 for (; I != E; ++I) Users.push_back(I->second);
581 while (!Users.empty()) {
586 return; // Quick exit
589 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
590 // and slow us down a lot. Just mark them overdefined.
591 if (PN.getNumIncomingValues() > 64) {
592 markOverdefined(PNIV, &PN);
596 // Look at all of the executable operands of the PHI node. If any of them
597 // are overdefined, the PHI becomes overdefined as well. If they are all
598 // constant, and they agree with each other, the PHI becomes the identical
599 // constant. If they are constant and don't agree, the PHI is overdefined.
600 // If there are no executable operands, the PHI remains undefined.
602 Constant *OperandVal = 0;
603 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
604 LatticeVal &IV = getValueState(PN.getIncomingValue(i));
605 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
607 if (isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) {
608 if (IV.isOverdefined()) { // PHI node becomes overdefined!
609 markOverdefined(&PN);
613 if (OperandVal == 0) { // Grab the first value...
614 OperandVal = IV.getConstant();
615 } else { // Another value is being merged in!
616 // There is already a reachable operand. If we conflict with it,
617 // then the PHI node becomes overdefined. If we agree with it, we
620 // Check to see if there are two different constants merging...
621 if (IV.getConstant() != OperandVal) {
622 // Yes there is. This means the PHI node is not constant.
623 // You must be overdefined poor PHI.
625 markOverdefined(&PN); // The PHI node now becomes overdefined
626 return; // I'm done analyzing you
632 // If we exited the loop, this means that the PHI node only has constant
633 // arguments that agree with each other(and OperandVal is the constant) or
634 // OperandVal is null because there are no defined incoming arguments. If
635 // this is the case, the PHI remains undefined.
638 markConstant(&PN, OperandVal); // Acquire operand value
641 void SCCPSolver::visitReturnInst(ReturnInst &I) {
642 if (I.getNumOperands() == 0) return; // Ret void
644 Function *F = I.getParent()->getParent();
645 // If we are tracking the return value of this function, merge it in.
646 if (!F->hasLocalLinkage())
649 if (!TrackedRetVals.empty() && I.getNumOperands() == 1) {
650 DenseMap<Function*, LatticeVal>::iterator TFRVI =
651 TrackedRetVals.find(F);
652 if (TFRVI != TrackedRetVals.end() &&
653 !TFRVI->second.isOverdefined()) {
654 LatticeVal &IV = getValueState(I.getOperand(0));
655 mergeInValue(TFRVI->second, F, IV);
660 // Handle functions that return multiple values.
661 if (!TrackedMultipleRetVals.empty() && I.getNumOperands() > 1) {
662 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
663 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
664 It = TrackedMultipleRetVals.find(std::make_pair(F, i));
665 if (It == TrackedMultipleRetVals.end()) break;
666 mergeInValue(It->second, F, getValueState(I.getOperand(i)));
668 } else if (!TrackedMultipleRetVals.empty() &&
669 I.getNumOperands() == 1 &&
670 isa<StructType>(I.getOperand(0)->getType())) {
671 for (unsigned i = 0, e = I.getOperand(0)->getType()->getNumContainedTypes();
673 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
674 It = TrackedMultipleRetVals.find(std::make_pair(F, i));
675 if (It == TrackedMultipleRetVals.end()) break;
676 if (Value *Val = FindInsertedValue(I.getOperand(0), i, I.getContext()))
677 mergeInValue(It->second, F, getValueState(Val));
682 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
683 SmallVector<bool, 16> SuccFeasible;
684 getFeasibleSuccessors(TI, SuccFeasible);
686 BasicBlock *BB = TI.getParent();
688 // Mark all feasible successors executable...
689 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
691 markEdgeExecutable(BB, TI.getSuccessor(i));
694 void SCCPSolver::visitCastInst(CastInst &I) {
695 Value *V = I.getOperand(0);
696 LatticeVal &VState = getValueState(V);
697 if (VState.isOverdefined()) // Inherit overdefinedness of operand
699 else if (VState.isConstant()) // Propagate constant value
700 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
701 VState.getConstant(), I.getType()));
704 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
705 Value *Aggr = EVI.getAggregateOperand();
707 // If the operand to the extractvalue is an undef, the result is undef.
708 if (isa<UndefValue>(Aggr))
711 // Currently only handle single-index extractvalues.
712 if (EVI.getNumIndices() != 1) {
713 markOverdefined(&EVI);
718 if (CallInst *CI = dyn_cast<CallInst>(Aggr))
719 F = CI->getCalledFunction();
720 else if (InvokeInst *II = dyn_cast<InvokeInst>(Aggr))
721 F = II->getCalledFunction();
723 // TODO: If IPSCCP resolves the callee of this function, we could propagate a
725 if (F == 0 || TrackedMultipleRetVals.empty()) {
726 markOverdefined(&EVI);
730 // See if we are tracking the result of the callee. If not tracking this
731 // function (for example, it is a declaration) just move to overdefined.
732 if (!TrackedMultipleRetVals.count(std::make_pair(F, *EVI.idx_begin()))) {
733 markOverdefined(&EVI);
737 // Otherwise, the value will be merged in here as a result of CallSite
741 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
742 Value *Aggr = IVI.getAggregateOperand();
743 Value *Val = IVI.getInsertedValueOperand();
745 // If the operands to the insertvalue are undef, the result is undef.
746 if (isa<UndefValue>(Aggr) && isa<UndefValue>(Val))
749 // Currently only handle single-index insertvalues.
750 if (IVI.getNumIndices() != 1) {
751 markOverdefined(&IVI);
755 // Currently only handle insertvalue instructions that are in a single-use
756 // chain that builds up a return value.
757 for (const InsertValueInst *TmpIVI = &IVI; ; ) {
758 if (!TmpIVI->hasOneUse()) {
759 markOverdefined(&IVI);
762 const Value *V = *TmpIVI->use_begin();
763 if (isa<ReturnInst>(V))
765 TmpIVI = dyn_cast<InsertValueInst>(V);
767 markOverdefined(&IVI);
772 // See if we are tracking the result of the callee.
773 Function *F = IVI.getParent()->getParent();
774 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
775 It = TrackedMultipleRetVals.find(std::make_pair(F, *IVI.idx_begin()));
777 // Merge in the inserted member value.
778 if (It != TrackedMultipleRetVals.end())
779 mergeInValue(It->second, F, getValueState(Val));
781 // Mark the aggregate result of the IVI overdefined; any tracking that we do
782 // will be done on the individual member values.
783 markOverdefined(&IVI);
786 void SCCPSolver::visitSelectInst(SelectInst &I) {
787 LatticeVal &CondValue = getValueState(I.getCondition());
788 if (CondValue.isUndefined())
790 if (CondValue.isConstant()) {
791 if (ConstantInt *CondCB = dyn_cast<ConstantInt>(CondValue.getConstant())){
792 mergeInValue(&I, getValueState(CondCB->getZExtValue() ? I.getTrueValue()
793 : I.getFalseValue()));
798 // Otherwise, the condition is overdefined or a constant we can't evaluate.
799 // See if we can produce something better than overdefined based on the T/F
801 LatticeVal &TVal = getValueState(I.getTrueValue());
802 LatticeVal &FVal = getValueState(I.getFalseValue());
804 // select ?, C, C -> C.
805 if (TVal.isConstant() && FVal.isConstant() &&
806 TVal.getConstant() == FVal.getConstant()) {
807 markConstant(&I, FVal.getConstant());
811 if (TVal.isUndefined()) { // select ?, undef, X -> X.
812 mergeInValue(&I, FVal);
813 } else if (FVal.isUndefined()) { // select ?, X, undef -> X.
814 mergeInValue(&I, TVal);
820 // Handle BinaryOperators and Shift Instructions...
821 void SCCPSolver::visitBinaryOperator(Instruction &I) {
822 LatticeVal &IV = ValueState[&I];
823 if (IV.isOverdefined()) return;
825 LatticeVal &V1State = getValueState(I.getOperand(0));
826 LatticeVal &V2State = getValueState(I.getOperand(1));
828 if (V1State.isOverdefined() || V2State.isOverdefined()) {
829 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
830 // operand is overdefined.
831 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
832 LatticeVal *NonOverdefVal = 0;
833 if (!V1State.isOverdefined()) {
834 NonOverdefVal = &V1State;
835 } else if (!V2State.isOverdefined()) {
836 NonOverdefVal = &V2State;
840 if (NonOverdefVal->isUndefined()) {
841 // Could annihilate value.
842 if (I.getOpcode() == Instruction::And)
843 markConstant(IV, &I, Constant::getNullValue(I.getType()));
844 else if (const VectorType *PT = dyn_cast<VectorType>(I.getType()))
845 markConstant(IV, &I, Constant::getAllOnesValue(PT));
848 Constant::getAllOnesValue(I.getType()));
851 if (I.getOpcode() == Instruction::And) {
852 if (NonOverdefVal->getConstant()->isNullValue()) {
853 markConstant(IV, &I, NonOverdefVal->getConstant());
854 return; // X and 0 = 0
857 if (ConstantInt *CI =
858 dyn_cast<ConstantInt>(NonOverdefVal->getConstant()))
859 if (CI->isAllOnesValue()) {
860 markConstant(IV, &I, NonOverdefVal->getConstant());
861 return; // X or -1 = -1
869 // If both operands are PHI nodes, it is possible that this instruction has
870 // a constant value, despite the fact that the PHI node doesn't. Check for
871 // this condition now.
872 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
873 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
874 if (PN1->getParent() == PN2->getParent()) {
875 // Since the two PHI nodes are in the same basic block, they must have
876 // entries for the same predecessors. Walk the predecessor list, and
877 // if all of the incoming values are constants, and the result of
878 // evaluating this expression with all incoming value pairs is the
879 // same, then this expression is a constant even though the PHI node
880 // is not a constant!
882 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
883 LatticeVal &In1 = getValueState(PN1->getIncomingValue(i));
884 BasicBlock *InBlock = PN1->getIncomingBlock(i);
886 getValueState(PN2->getIncomingValueForBlock(InBlock));
888 if (In1.isOverdefined() || In2.isOverdefined()) {
889 Result.markOverdefined();
890 break; // Cannot fold this operation over the PHI nodes!
891 } else if (In1.isConstant() && In2.isConstant()) {
893 ConstantExpr::get(I.getOpcode(), In1.getConstant(),
895 if (Result.isUndefined())
896 Result.markConstant(V);
897 else if (Result.isConstant() && Result.getConstant() != V) {
898 Result.markOverdefined();
904 // If we found a constant value here, then we know the instruction is
905 // constant despite the fact that the PHI nodes are overdefined.
906 if (Result.isConstant()) {
907 markConstant(IV, &I, Result.getConstant());
908 // Remember that this instruction is virtually using the PHI node
910 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
911 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
913 } else if (Result.isUndefined()) {
917 // Okay, this really is overdefined now. Since we might have
918 // speculatively thought that this was not overdefined before, and
919 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
920 // make sure to clean out any entries that we put there, for
922 std::multimap<PHINode*, Instruction*>::iterator It, E;
923 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1);
925 if (It->second == &I) {
926 UsersOfOverdefinedPHIs.erase(It++);
930 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2);
932 if (It->second == &I) {
933 UsersOfOverdefinedPHIs.erase(It++);
939 markOverdefined(IV, &I);
940 } else if (V1State.isConstant() && V2State.isConstant()) {
942 ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
943 V2State.getConstant()));
947 // Handle ICmpInst instruction...
948 void SCCPSolver::visitCmpInst(CmpInst &I) {
949 LatticeVal &IV = ValueState[&I];
950 if (IV.isOverdefined()) return;
952 LatticeVal &V1State = getValueState(I.getOperand(0));
953 LatticeVal &V2State = getValueState(I.getOperand(1));
955 if (V1State.isOverdefined() || V2State.isOverdefined()) {
956 // If both operands are PHI nodes, it is possible that this instruction has
957 // a constant value, despite the fact that the PHI node doesn't. Check for
958 // this condition now.
959 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
960 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
961 if (PN1->getParent() == PN2->getParent()) {
962 // Since the two PHI nodes are in the same basic block, they must have
963 // entries for the same predecessors. Walk the predecessor list, and
964 // if all of the incoming values are constants, and the result of
965 // evaluating this expression with all incoming value pairs is the
966 // same, then this expression is a constant even though the PHI node
967 // is not a constant!
969 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
970 LatticeVal &In1 = getValueState(PN1->getIncomingValue(i));
971 BasicBlock *InBlock = PN1->getIncomingBlock(i);
973 getValueState(PN2->getIncomingValueForBlock(InBlock));
975 if (In1.isOverdefined() || In2.isOverdefined()) {
976 Result.markOverdefined();
977 break; // Cannot fold this operation over the PHI nodes!
978 } else if (In1.isConstant() && In2.isConstant()) {
979 Constant *V = ConstantExpr::getCompare(I.getPredicate(),
982 if (Result.isUndefined())
983 Result.markConstant(V);
984 else if (Result.isConstant() && Result.getConstant() != V) {
985 Result.markOverdefined();
991 // If we found a constant value here, then we know the instruction is
992 // constant despite the fact that the PHI nodes are overdefined.
993 if (Result.isConstant()) {
994 markConstant(IV, &I, Result.getConstant());
995 // Remember that this instruction is virtually using the PHI node
997 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
998 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
1000 } else if (Result.isUndefined()) {
1004 // Okay, this really is overdefined now. Since we might have
1005 // speculatively thought that this was not overdefined before, and
1006 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
1007 // make sure to clean out any entries that we put there, for
1009 std::multimap<PHINode*, Instruction*>::iterator It, E;
1010 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1);
1012 if (It->second == &I) {
1013 UsersOfOverdefinedPHIs.erase(It++);
1017 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2);
1019 if (It->second == &I) {
1020 UsersOfOverdefinedPHIs.erase(It++);
1026 markOverdefined(IV, &I);
1027 } else if (V1State.isConstant() && V2State.isConstant()) {
1028 markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
1029 V1State.getConstant(),
1030 V2State.getConstant()));
1034 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
1035 // FIXME : SCCP does not handle vectors properly.
1036 markOverdefined(&I);
1040 LatticeVal &ValState = getValueState(I.getOperand(0));
1041 LatticeVal &IdxState = getValueState(I.getOperand(1));
1043 if (ValState.isOverdefined() || IdxState.isOverdefined())
1044 markOverdefined(&I);
1045 else if(ValState.isConstant() && IdxState.isConstant())
1046 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
1047 IdxState.getConstant()));
1051 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
1052 // FIXME : SCCP does not handle vectors properly.
1053 markOverdefined(&I);
1056 LatticeVal &ValState = getValueState(I.getOperand(0));
1057 LatticeVal &EltState = getValueState(I.getOperand(1));
1058 LatticeVal &IdxState = getValueState(I.getOperand(2));
1060 if (ValState.isOverdefined() || EltState.isOverdefined() ||
1061 IdxState.isOverdefined())
1062 markOverdefined(&I);
1063 else if(ValState.isConstant() && EltState.isConstant() &&
1064 IdxState.isConstant())
1065 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
1066 EltState.getConstant(),
1067 IdxState.getConstant()));
1068 else if (ValState.isUndefined() && EltState.isConstant() &&
1069 IdxState.isConstant())
1070 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
1071 EltState.getConstant(),
1072 IdxState.getConstant()));
1076 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
1077 // FIXME : SCCP does not handle vectors properly.
1078 markOverdefined(&I);
1081 LatticeVal &V1State = getValueState(I.getOperand(0));
1082 LatticeVal &V2State = getValueState(I.getOperand(1));
1083 LatticeVal &MaskState = getValueState(I.getOperand(2));
1085 if (MaskState.isUndefined() ||
1086 (V1State.isUndefined() && V2State.isUndefined()))
1087 return; // Undefined output if mask or both inputs undefined.
1089 if (V1State.isOverdefined() || V2State.isOverdefined() ||
1090 MaskState.isOverdefined()) {
1091 markOverdefined(&I);
1093 // A mix of constant/undef inputs.
1094 Constant *V1 = V1State.isConstant() ?
1095 V1State.getConstant() : UndefValue::get(I.getType());
1096 Constant *V2 = V2State.isConstant() ?
1097 V2State.getConstant() : UndefValue::get(I.getType());
1098 Constant *Mask = MaskState.isConstant() ?
1099 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
1100 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
1105 // Handle getelementptr instructions... if all operands are constants then we
1106 // can turn this into a getelementptr ConstantExpr.
1108 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1109 LatticeVal &IV = ValueState[&I];
1110 if (IV.isOverdefined()) return;
1112 SmallVector<Constant*, 8> Operands;
1113 Operands.reserve(I.getNumOperands());
1115 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1116 LatticeVal &State = getValueState(I.getOperand(i));
1117 if (State.isUndefined())
1118 return; // Operands are not resolved yet...
1119 else if (State.isOverdefined()) {
1120 markOverdefined(IV, &I);
1123 assert(State.isConstant() && "Unknown state!");
1124 Operands.push_back(State.getConstant());
1127 Constant *Ptr = Operands[0];
1128 Operands.erase(Operands.begin()); // Erase the pointer from idx list...
1130 markConstant(IV, &I, ConstantExpr::getGetElementPtr(Ptr, &Operands[0],
1134 void SCCPSolver::visitStoreInst(Instruction &SI) {
1135 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1137 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1138 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1139 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1141 // Get the value we are storing into the global.
1142 LatticeVal &PtrVal = getValueState(SI.getOperand(0));
1144 mergeInValue(I->second, GV, PtrVal);
1145 if (I->second.isOverdefined())
1146 TrackedGlobals.erase(I); // No need to keep tracking this!
1150 // Handle load instructions. If the operand is a constant pointer to a constant
1151 // global, we can replace the load with the loaded constant value!
1152 void SCCPSolver::visitLoadInst(LoadInst &I) {
1153 LatticeVal &IV = ValueState[&I];
1154 if (IV.isOverdefined()) return;
1156 LatticeVal &PtrVal = getValueState(I.getOperand(0));
1157 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1158 if (PtrVal.isConstant() && !I.isVolatile()) {
1159 Value *Ptr = PtrVal.getConstant();
1160 // TODO: Consider a target hook for valid address spaces for this xform.
1161 if (isa<ConstantPointerNull>(Ptr) && I.getPointerAddressSpace() == 0) {
1162 // load null -> null
1163 markConstant(IV, &I, Constant::getNullValue(I.getType()));
1167 // Transform load (constant global) into the value loaded.
1168 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1169 if (GV->isConstant()) {
1170 if (GV->hasDefinitiveInitializer()) {
1171 markConstant(IV, &I, GV->getInitializer());
1174 } else if (!TrackedGlobals.empty()) {
1175 // If we are tracking this global, merge in the known value for it.
1176 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1177 TrackedGlobals.find(GV);
1178 if (It != TrackedGlobals.end()) {
1179 mergeInValue(IV, &I, It->second);
1185 // Transform load (constantexpr_GEP global, 0, ...) into the value loaded.
1186 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
1187 if (CE->getOpcode() == Instruction::GetElementPtr)
1188 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
1189 if (GV->isConstant() && GV->hasDefinitiveInitializer())
1191 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) {
1192 markConstant(IV, &I, V);
1197 // Otherwise we cannot say for certain what value this load will produce.
1199 markOverdefined(IV, &I);
1202 void SCCPSolver::visitCallSite(CallSite CS) {
1203 Function *F = CS.getCalledFunction();
1204 Instruction *I = CS.getInstruction();
1206 // The common case is that we aren't tracking the callee, either because we
1207 // are not doing interprocedural analysis or the callee is indirect, or is
1208 // external. Handle these cases first.
1209 if (F == 0 || !F->hasLocalLinkage()) {
1211 // Void return and not tracking callee, just bail.
1212 if (I->getType()->isVoidTy()) return;
1214 // Otherwise, if we have a single return value case, and if the function is
1215 // a declaration, maybe we can constant fold it.
1216 if (!isa<StructType>(I->getType()) && F && F->isDeclaration() &&
1217 canConstantFoldCallTo(F)) {
1219 SmallVector<Constant*, 8> Operands;
1220 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1222 LatticeVal &State = getValueState(*AI);
1223 if (State.isUndefined())
1224 return; // Operands are not resolved yet.
1225 else if (State.isOverdefined()) {
1229 assert(State.isConstant() && "Unknown state!");
1230 Operands.push_back(State.getConstant());
1233 // If we can constant fold this, mark the result of the call as a
1235 if (Constant *C = ConstantFoldCall(F, Operands.data(), Operands.size())) {
1241 // Otherwise, we don't know anything about this call, mark it overdefined.
1246 // If this is a single/zero retval case, see if we're tracking the function.
1247 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1248 if (TFRVI != TrackedRetVals.end()) {
1249 // If so, propagate the return value of the callee into this call result.
1250 mergeInValue(I, TFRVI->second);
1251 } else if (isa<StructType>(I->getType())) {
1252 // Check to see if we're tracking this callee, if not, handle it in the
1253 // common path above.
1254 DenseMap<std::pair<Function*, unsigned>, LatticeVal>::iterator
1255 TMRVI = TrackedMultipleRetVals.find(std::make_pair(F, 0));
1256 if (TMRVI == TrackedMultipleRetVals.end())
1257 goto CallOverdefined;
1259 // Need to mark as overdefined, otherwise it stays undefined which
1260 // creates extractvalue undef, <idx>
1262 // If we are tracking this callee, propagate the return values of the call
1263 // into this call site. We do this by walking all the uses. Single-index
1264 // ExtractValueInst uses can be tracked; anything more complicated is
1265 // currently handled conservatively.
1266 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1268 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(*UI)) {
1269 if (EVI->getNumIndices() == 1) {
1271 TrackedMultipleRetVals[std::make_pair(F, *EVI->idx_begin())]);
1275 // The aggregate value is used in a way not handled here. Assume nothing.
1276 markOverdefined(*UI);
1279 // Otherwise we're not tracking this callee, so handle it in the
1280 // common path above.
1281 goto CallOverdefined;
1284 // Finally, if this is the first call to the function hit, mark its entry
1285 // block executable.
1286 if (!BBExecutable.count(F->begin()))
1287 MarkBlockExecutable(F->begin());
1289 // Propagate information from this call site into the callee.
1290 CallSite::arg_iterator CAI = CS.arg_begin();
1291 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1292 AI != E; ++AI, ++CAI) {
1293 LatticeVal &IV = ValueState[AI];
1294 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1295 IV.markOverdefined();
1298 if (!IV.isOverdefined())
1299 mergeInValue(IV, AI, getValueState(*CAI));
1303 void SCCPSolver::Solve() {
1304 // Process the work lists until they are empty!
1305 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1306 !OverdefinedInstWorkList.empty()) {
1307 // Process the instruction work list...
1308 while (!OverdefinedInstWorkList.empty()) {
1309 Value *I = OverdefinedInstWorkList.back();
1310 OverdefinedInstWorkList.pop_back();
1312 DEBUG(errs() << "\nPopped off OI-WL: " << *I << '\n');
1314 // "I" got into the work list because it either made the transition from
1315 // bottom to constant
1317 // Anything on this worklist that is overdefined need not be visited
1318 // since all of its users will have already been marked as overdefined
1319 // Update all of the users of this instruction's value...
1321 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1323 OperandChangedState(*UI);
1325 // Process the instruction work list...
1326 while (!InstWorkList.empty()) {
1327 Value *I = InstWorkList.back();
1328 InstWorkList.pop_back();
1330 DEBUG(errs() << "\nPopped off I-WL: " << *I << '\n');
1332 // "I" got into the work list because it either made the transition from
1333 // bottom to constant
1335 // Anything on this worklist that is overdefined need not be visited
1336 // since all of its users will have already been marked as overdefined.
1337 // Update all of the users of this instruction's value...
1339 if (!getValueState(I).isOverdefined())
1340 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1342 OperandChangedState(*UI);
1345 // Process the basic block work list...
1346 while (!BBWorkList.empty()) {
1347 BasicBlock *BB = BBWorkList.back();
1348 BBWorkList.pop_back();
1350 DEBUG(errs() << "\nPopped off BBWL: " << *BB << '\n');
1352 // Notify all instructions in this basic block that they are newly
1359 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1360 /// that branches on undef values cannot reach any of their successors.
1361 /// However, this is not a safe assumption. After we solve dataflow, this
1362 /// method should be use to handle this. If this returns true, the solver
1363 /// should be rerun.
1365 /// This method handles this by finding an unresolved branch and marking it one
1366 /// of the edges from the block as being feasible, even though the condition
1367 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1368 /// CFG and only slightly pessimizes the analysis results (by marking one,
1369 /// potentially infeasible, edge feasible). This cannot usefully modify the
1370 /// constraints on the condition of the branch, as that would impact other users
1373 /// This scan also checks for values that use undefs, whose results are actually
1374 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1375 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1376 /// even if X isn't defined.
1377 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1378 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1379 if (!BBExecutable.count(BB))
1382 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1383 // Look for instructions which produce undef values.
1384 if (I->getType()->isVoidTy()) continue;
1386 LatticeVal &LV = getValueState(I);
1387 if (!LV.isUndefined()) continue;
1389 // Get the lattice values of the first two operands for use below.
1390 LatticeVal &Op0LV = getValueState(I->getOperand(0));
1392 if (I->getNumOperands() == 2) {
1393 // If this is a two-operand instruction, and if both operands are
1394 // undefs, the result stays undef.
1395 Op1LV = getValueState(I->getOperand(1));
1396 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1400 // If this is an instructions whose result is defined even if the input is
1401 // not fully defined, propagate the information.
1402 const Type *ITy = I->getType();
1403 switch (I->getOpcode()) {
1404 default: break; // Leave the instruction as an undef.
1405 case Instruction::ZExt:
1406 // After a zero extend, we know the top part is zero. SExt doesn't have
1407 // to be handled here, because we don't know whether the top part is 1's
1409 assert(Op0LV.isUndefined());
1410 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1412 case Instruction::Mul:
1413 case Instruction::And:
1414 // undef * X -> 0. X could be zero.
1415 // undef & X -> 0. X could be zero.
1416 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1419 case Instruction::Or:
1420 // undef | X -> -1. X could be -1.
1421 if (const VectorType *PTy = dyn_cast<VectorType>(ITy))
1422 markForcedConstant(LV, I,
1423 Constant::getAllOnesValue(PTy));
1425 markForcedConstant(LV, I, Constant::getAllOnesValue(ITy));
1428 case Instruction::SDiv:
1429 case Instruction::UDiv:
1430 case Instruction::SRem:
1431 case Instruction::URem:
1432 // X / undef -> undef. No change.
1433 // X % undef -> undef. No change.
1434 if (Op1LV.isUndefined()) break;
1436 // undef / X -> 0. X could be maxint.
1437 // undef % X -> 0. X could be 1.
1438 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1441 case Instruction::AShr:
1442 // undef >>s X -> undef. No change.
1443 if (Op0LV.isUndefined()) break;
1445 // X >>s undef -> X. X could be 0, X could have the high-bit known set.
1446 if (Op0LV.isConstant())
1447 markForcedConstant(LV, I, Op0LV.getConstant());
1449 markOverdefined(LV, I);
1451 case Instruction::LShr:
1452 case Instruction::Shl:
1453 // undef >> X -> undef. No change.
1454 // undef << X -> undef. No change.
1455 if (Op0LV.isUndefined()) break;
1457 // X >> undef -> 0. X could be 0.
1458 // X << undef -> 0. X could be 0.
1459 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1461 case Instruction::Select:
1462 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1463 if (Op0LV.isUndefined()) {
1464 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1465 Op1LV = getValueState(I->getOperand(2));
1466 } else if (Op1LV.isUndefined()) {
1467 // c ? undef : undef -> undef. No change.
1468 Op1LV = getValueState(I->getOperand(2));
1469 if (Op1LV.isUndefined())
1471 // Otherwise, c ? undef : x -> x.
1473 // Leave Op1LV as Operand(1)'s LatticeValue.
1476 if (Op1LV.isConstant())
1477 markForcedConstant(LV, I, Op1LV.getConstant());
1479 markOverdefined(LV, I);
1481 case Instruction::Call:
1482 // If a call has an undef result, it is because it is constant foldable
1483 // but one of the inputs was undef. Just force the result to
1485 markOverdefined(LV, I);
1490 TerminatorInst *TI = BB->getTerminator();
1491 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1492 if (!BI->isConditional()) continue;
1493 if (!getValueState(BI->getCondition()).isUndefined())
1495 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1496 if (SI->getNumSuccessors()<2) // no cases
1498 if (!getValueState(SI->getCondition()).isUndefined())
1504 // If the edge to the second successor isn't thought to be feasible yet,
1505 // mark it so now. We pick the second one so that this goes to some
1506 // enumerated value in a switch instead of going to the default destination.
1507 if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(1))))
1510 // Otherwise, it isn't already thought to be feasible. Mark it as such now
1511 // and return. This will make other blocks reachable, which will allow new
1512 // values to be discovered and existing ones to be moved in the lattice.
1513 markEdgeExecutable(BB, TI->getSuccessor(1));
1515 // This must be a conditional branch of switch on undef. At this point,
1516 // force the old terminator to branch to the first successor. This is
1517 // required because we are now influencing the dataflow of the function with
1518 // the assumption that this edge is taken. If we leave the branch condition
1519 // as undef, then further analysis could think the undef went another way
1520 // leading to an inconsistent set of conclusions.
1521 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1522 BI->setCondition(ConstantInt::getFalse(*Context));
1524 SwitchInst *SI = cast<SwitchInst>(TI);
1525 SI->setCondition(SI->getCaseValue(1));
1536 //===--------------------------------------------------------------------===//
1538 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1539 /// Sparse Conditional Constant Propagator.
1541 struct SCCP : public FunctionPass {
1542 static char ID; // Pass identification, replacement for typeid
1543 SCCP() : FunctionPass(&ID) {}
1545 // runOnFunction - Run the Sparse Conditional Constant Propagation
1546 // algorithm, and return true if the function was modified.
1548 bool runOnFunction(Function &F);
1550 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1551 AU.setPreservesCFG();
1554 } // end anonymous namespace
1557 static RegisterPass<SCCP>
1558 X("sccp", "Sparse Conditional Constant Propagation");
1560 // createSCCPPass - This is the public interface to this file...
1561 FunctionPass *llvm::createSCCPPass() {
1566 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1567 // and return true if the function was modified.
1569 bool SCCP::runOnFunction(Function &F) {
1570 DEBUG(errs() << "SCCP on function '" << F.getName() << "'\n");
1572 Solver.setContext(&F.getContext());
1574 // Mark the first block of the function as being executable.
1575 Solver.MarkBlockExecutable(F.begin());
1577 // Mark all arguments to the function as being overdefined.
1578 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1579 Solver.markOverdefined(AI);
1581 // Solve for constants.
1582 bool ResolvedUndefs = true;
1583 while (ResolvedUndefs) {
1585 DEBUG(errs() << "RESOLVING UNDEFs\n");
1586 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1589 bool MadeChanges = false;
1591 // If we decided that there are basic blocks that are dead in this function,
1592 // delete their contents now. Note that we cannot actually delete the blocks,
1593 // as we cannot modify the CFG of the function.
1595 SmallVector<Instruction*, 512> Insts;
1596 std::map<Value*, LatticeVal> &Values = Solver.getValueMapping();
1598 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1599 if (!Solver.isBlockExecutable(BB)) {
1600 DEBUG(errs() << " BasicBlock Dead:" << *BB);
1603 // Delete the instructions backwards, as it has a reduced likelihood of
1604 // having to update as many def-use and use-def chains.
1605 for (BasicBlock::iterator I = BB->begin(), E = BB->getTerminator();
1608 while (!Insts.empty()) {
1609 Instruction *I = Insts.back();
1611 if (!I->use_empty())
1612 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1613 BB->getInstList().erase(I);
1618 // Iterate over all of the instructions in a function, replacing them with
1619 // constants if we have found them to be of constant values.
1621 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1622 Instruction *Inst = BI++;
1623 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst))
1626 LatticeVal &IV = Values[Inst];
1627 if (!IV.isConstant() && !IV.isUndefined())
1630 Constant *Const = IV.isConstant()
1631 ? IV.getConstant() : UndefValue::get(Inst->getType());
1632 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1634 // Replaces all of the uses of a variable with uses of the constant.
1635 Inst->replaceAllUsesWith(Const);
1637 // Delete the instruction.
1638 Inst->eraseFromParent();
1640 // Hey, we just changed something!
1650 //===--------------------------------------------------------------------===//
1652 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1653 /// Constant Propagation.
1655 struct IPSCCP : public ModulePass {
1657 IPSCCP() : ModulePass(&ID) {}
1658 bool runOnModule(Module &M);
1660 } // end anonymous namespace
1662 char IPSCCP::ID = 0;
1663 static RegisterPass<IPSCCP>
1664 Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation");
1666 // createIPSCCPPass - This is the public interface to this file...
1667 ModulePass *llvm::createIPSCCPPass() {
1668 return new IPSCCP();
1672 static bool AddressIsTaken(GlobalValue *GV) {
1673 // Delete any dead constantexpr klingons.
1674 GV->removeDeadConstantUsers();
1676 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
1678 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1679 if (SI->getOperand(0) == GV || SI->isVolatile())
1680 return true; // Storing addr of GV.
1681 } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) {
1682 // Make sure we are calling the function, not passing the address.
1683 CallSite CS = CallSite::get(cast<Instruction>(*UI));
1684 if (CS.hasArgument(GV))
1686 } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1687 if (LI->isVolatile())
1695 bool IPSCCP::runOnModule(Module &M) {
1696 LLVMContext *Context = &M.getContext();
1699 Solver.setContext(Context);
1701 // Loop over all functions, marking arguments to those with their addresses
1702 // taken or that are external as overdefined.
1704 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1705 if (!F->hasLocalLinkage() || AddressIsTaken(F)) {
1706 if (!F->isDeclaration())
1707 Solver.MarkBlockExecutable(F->begin());
1708 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1710 Solver.markOverdefined(AI);
1712 Solver.AddTrackedFunction(F);
1715 // Loop over global variables. We inform the solver about any internal global
1716 // variables that do not have their 'addresses taken'. If they don't have
1717 // their addresses taken, we can propagate constants through them.
1718 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1720 if (!G->isConstant() && G->hasLocalLinkage() && !AddressIsTaken(G))
1721 Solver.TrackValueOfGlobalVariable(G);
1723 // Solve for constants.
1724 bool ResolvedUndefs = true;
1725 while (ResolvedUndefs) {
1728 DEBUG(errs() << "RESOLVING UNDEFS\n");
1729 ResolvedUndefs = false;
1730 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1731 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1734 bool MadeChanges = false;
1736 // Iterate over all of the instructions in the module, replacing them with
1737 // constants if we have found them to be of constant values.
1739 SmallVector<Instruction*, 512> Insts;
1740 SmallVector<BasicBlock*, 512> BlocksToErase;
1741 std::map<Value*, LatticeVal> &Values = Solver.getValueMapping();
1743 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1744 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1746 if (!AI->use_empty()) {
1747 LatticeVal &IV = Values[AI];
1748 if (IV.isConstant() || IV.isUndefined()) {
1749 Constant *CST = IV.isConstant() ?
1750 IV.getConstant() : UndefValue::get(AI->getType());
1751 DEBUG(errs() << "*** Arg " << *AI << " = " << *CST <<"\n");
1753 // Replaces all of the uses of a variable with uses of the
1755 AI->replaceAllUsesWith(CST);
1760 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1761 if (!Solver.isBlockExecutable(BB)) {
1762 DEBUG(errs() << " BasicBlock Dead:" << *BB);
1765 // Delete the instructions backwards, as it has a reduced likelihood of
1766 // having to update as many def-use and use-def chains.
1767 TerminatorInst *TI = BB->getTerminator();
1768 for (BasicBlock::iterator I = BB->begin(), E = TI; I != E; ++I)
1771 while (!Insts.empty()) {
1772 Instruction *I = Insts.back();
1774 if (!I->use_empty())
1775 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1776 BB->getInstList().erase(I);
1781 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1782 BasicBlock *Succ = TI->getSuccessor(i);
1783 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1784 TI->getSuccessor(i)->removePredecessor(BB);
1786 if (!TI->use_empty())
1787 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1788 BB->getInstList().erase(TI);
1790 if (&*BB != &F->front())
1791 BlocksToErase.push_back(BB);
1793 new UnreachableInst(M.getContext(), BB);
1796 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1797 Instruction *Inst = BI++;
1798 if (Inst->getType()->isVoidTy())
1801 LatticeVal &IV = Values[Inst];
1802 if (!IV.isConstant() && !IV.isUndefined())
1805 Constant *Const = IV.isConstant()
1806 ? IV.getConstant() : UndefValue::get(Inst->getType());
1807 DEBUG(errs() << " Constant: " << *Const << " = " << *Inst);
1809 // Replaces all of the uses of a variable with uses of the
1811 Inst->replaceAllUsesWith(Const);
1813 // Delete the instruction.
1814 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst))
1815 Inst->eraseFromParent();
1817 // Hey, we just changed something!
1823 // Now that all instructions in the function are constant folded, erase dead
1824 // blocks, because we can now use ConstantFoldTerminator to get rid of
1826 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1827 // If there are any PHI nodes in this successor, drop entries for BB now.
1828 BasicBlock *DeadBB = BlocksToErase[i];
1829 while (!DeadBB->use_empty()) {
1830 Instruction *I = cast<Instruction>(DeadBB->use_back());
1831 bool Folded = ConstantFoldTerminator(I->getParent());
1833 // The constant folder may not have been able to fold the terminator
1834 // if this is a branch or switch on undef. Fold it manually as a
1835 // branch to the first successor.
1837 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1838 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1839 "Branch should be foldable!");
1840 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1841 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1843 llvm_unreachable("Didn't fold away reference to block!");
1847 // Make this an uncond branch to the first successor.
1848 TerminatorInst *TI = I->getParent()->getTerminator();
1849 BranchInst::Create(TI->getSuccessor(0), TI);
1851 // Remove entries in successor phi nodes to remove edges.
1852 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1853 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1855 // Remove the old terminator.
1856 TI->eraseFromParent();
1860 // Finally, delete the basic block.
1861 F->getBasicBlockList().erase(DeadBB);
1863 BlocksToErase.clear();
1866 // If we inferred constant or undef return values for a function, we replaced
1867 // all call uses with the inferred value. This means we don't need to bother
1868 // actually returning anything from the function. Replace all return
1869 // instructions with return undef.
1870 // TODO: Process multiple value ret instructions also.
1871 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1872 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1873 E = RV.end(); I != E; ++I)
1874 if (!I->second.isOverdefined() &&
1875 !I->first->getReturnType()->isVoidTy()) {
1876 Function *F = I->first;
1877 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1878 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1879 if (!isa<UndefValue>(RI->getOperand(0)))
1880 RI->setOperand(0, UndefValue::get(F->getReturnType()));
1883 // If we infered constant or undef values for globals variables, we can delete
1884 // the global and any stores that remain to it.
1885 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1886 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1887 E = TG.end(); I != E; ++I) {
1888 GlobalVariable *GV = I->first;
1889 assert(!I->second.isOverdefined() &&
1890 "Overdefined values should have been taken out of the map!");
1891 DEBUG(errs() << "Found that GV '" << GV->getName() << "' is constant!\n");
1892 while (!GV->use_empty()) {
1893 StoreInst *SI = cast<StoreInst>(GV->use_back());
1894 SI->eraseFromParent();
1896 M.getGlobalList().erase(GV);