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/Pass.h"
31 #include "llvm/Analysis/ConstantFolding.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/Support/CallSite.h"
34 #include "llvm/Support/Compiler.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/InstVisitor.h"
37 #include "llvm/ADT/DenseMap.h"
38 #include "llvm/ADT/SmallSet.h"
39 #include "llvm/ADT/SmallVector.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/ADT/STLExtras.h"
46 STATISTIC(NumInstRemoved, "Number of instructions removed");
47 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
49 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP");
50 STATISTIC(IPNumDeadBlocks , "Number of basic blocks unreachable by IPSCCP");
51 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
52 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
55 /// LatticeVal class - This class represents the different lattice values that
56 /// an LLVM value may occupy. It is a simple class with value semantics.
58 class VISIBILITY_HIDDEN LatticeVal {
60 /// undefined - This LLVM Value has no known value yet.
63 /// constant - This LLVM Value has a specific constant value.
66 /// forcedconstant - This LLVM Value was thought to be undef until
67 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged
68 /// with another (different) constant, it goes to overdefined, instead of
72 /// overdefined - This instruction is not known to be constant, and we know
75 } LatticeValue; // The current lattice position
77 Constant *ConstantVal; // If Constant value, the current value
79 inline LatticeVal() : LatticeValue(undefined), ConstantVal(0) {}
81 // markOverdefined - Return true if this is a new status to be in...
82 inline bool markOverdefined() {
83 if (LatticeValue != overdefined) {
84 LatticeValue = overdefined;
90 // markConstant - Return true if this is a new status for us.
91 inline bool markConstant(Constant *V) {
92 if (LatticeValue != constant) {
93 if (LatticeValue == undefined) {
94 LatticeValue = constant;
95 assert(V && "Marking constant with NULL");
98 assert(LatticeValue == forcedconstant &&
99 "Cannot move from overdefined to constant!");
100 // Stay at forcedconstant if the constant is the same.
101 if (V == ConstantVal) return false;
103 // Otherwise, we go to overdefined. Assumptions made based on the
104 // forced value are possibly wrong. Assuming this is another constant
105 // could expose a contradiction.
106 LatticeValue = overdefined;
110 assert(ConstantVal == V && "Marking constant with different value");
115 inline void markForcedConstant(Constant *V) {
116 assert(LatticeValue == undefined && "Can't force a defined value!");
117 LatticeValue = forcedconstant;
121 inline bool isUndefined() const { return LatticeValue == undefined; }
122 inline bool isConstant() const {
123 return LatticeValue == constant || LatticeValue == forcedconstant;
125 inline bool isOverdefined() const { return LatticeValue == overdefined; }
127 inline Constant *getConstant() const {
128 assert(isConstant() && "Cannot get the constant of a non-constant!");
133 /// LatticeValIndex - LatticeVal and associated Index. This is used
134 /// to track individual operand Lattice values for multi value ret instructions.
135 class VISIBILITY_HIDDEN LatticeValIndexed {
137 LatticeValIndexed(unsigned I = 0) { Index = I; }
138 LatticeVal &getLatticeVal() { return LV; }
139 unsigned getIndex() const { return Index; }
141 void setLatticeVal(LatticeVal &L) { LV = L; }
142 void setIndex(unsigned I) { Index = I; }
148 //===----------------------------------------------------------------------===//
150 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional
151 /// Constant Propagation.
153 class SCCPSolver : public InstVisitor<SCCPSolver> {
154 SmallSet<BasicBlock*, 16> BBExecutable;// The basic blocks that are executable
155 std::map<Value*, LatticeVal> ValueState; // The state each value is in.
157 /// GlobalValue - If we are tracking any values for the contents of a global
158 /// variable, we keep a mapping from the constant accessor to the element of
159 /// the global, to the currently known value. If the value becomes
160 /// overdefined, it's entry is simply removed from this map.
161 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
163 /// TrackedRetVals - If we are tracking arguments into and the return
164 /// value out of a function, it will have an entry in this map, indicating
165 /// what the known return value for the function is.
166 DenseMap<Function*, LatticeVal> TrackedRetVals;
168 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
169 /// that return multiple values.
170 std::multimap<Function*, LatticeValIndexed> TrackedMultipleRetVals;
172 // The reason for two worklists is that overdefined is the lowest state
173 // on the lattice, and moving things to overdefined as fast as possible
174 // makes SCCP converge much faster.
175 // By having a separate worklist, we accomplish this because everything
176 // possibly overdefined will become overdefined at the soonest possible
178 std::vector<Value*> OverdefinedInstWorkList;
179 std::vector<Value*> InstWorkList;
182 std::vector<BasicBlock*> BBWorkList; // The BasicBlock work list
184 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
185 /// overdefined, despite the fact that the PHI node is overdefined.
186 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
188 /// KnownFeasibleEdges - Entries in this set are edges which have already had
189 /// PHI nodes retriggered.
190 typedef std::pair<BasicBlock*,BasicBlock*> Edge;
191 std::set<Edge> KnownFeasibleEdges;
194 /// MarkBlockExecutable - This method can be used by clients to mark all of
195 /// the blocks that are known to be intrinsically live in the processed unit.
196 void MarkBlockExecutable(BasicBlock *BB) {
197 DOUT << "Marking Block Executable: " << BB->getName() << "\n";
198 BBExecutable.insert(BB); // Basic block is executable!
199 BBWorkList.push_back(BB); // Add the block to the work list!
202 /// TrackValueOfGlobalVariable - Clients can use this method to
203 /// inform the SCCPSolver that it should track loads and stores to the
204 /// specified global variable if it can. This is only legal to call if
205 /// performing Interprocedural SCCP.
206 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
207 const Type *ElTy = GV->getType()->getElementType();
208 if (ElTy->isFirstClassType()) {
209 LatticeVal &IV = TrackedGlobals[GV];
210 if (!isa<UndefValue>(GV->getInitializer()))
211 IV.markConstant(GV->getInitializer());
215 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
216 /// and out of the specified function (which cannot have its address taken),
217 /// this method must be called.
218 void AddTrackedFunction(Function *F) {
219 assert(F->hasInternalLinkage() && "Can only track internal functions!");
220 // Add an entry, F -> undef.
221 if (const StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
222 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
223 TrackedMultipleRetVals.insert(std::pair<Function *, LatticeValIndexed>
224 (F, LatticeValIndexed(i)));
230 /// Solve - Solve for constants and executable blocks.
234 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
235 /// that branches on undef values cannot reach any of their successors.
236 /// However, this is not a safe assumption. After we solve dataflow, this
237 /// method should be use to handle this. If this returns true, the solver
239 bool ResolvedUndefsIn(Function &F);
241 /// getExecutableBlocks - Once we have solved for constants, return the set of
242 /// blocks that is known to be executable.
243 SmallSet<BasicBlock*, 16> &getExecutableBlocks() {
247 /// getValueMapping - Once we have solved for constants, return the mapping of
248 /// LLVM values to LatticeVals.
249 std::map<Value*, LatticeVal> &getValueMapping() {
253 /// getTrackedRetVals - Get the inferred return value map.
255 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
256 return TrackedRetVals;
259 /// getTrackedGlobals - Get and return the set of inferred initializers for
260 /// global variables.
261 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
262 return TrackedGlobals;
265 inline void markOverdefined(Value *V) {
266 markOverdefined(ValueState[V], V);
270 // markConstant - Make a value be marked as "constant". If the value
271 // is not already a constant, add it to the instruction work list so that
272 // the users of the instruction are updated later.
274 inline void markConstant(LatticeVal &IV, Value *V, Constant *C) {
275 if (IV.markConstant(C)) {
276 DOUT << "markConstant: " << *C << ": " << *V;
277 InstWorkList.push_back(V);
281 inline void markForcedConstant(LatticeVal &IV, Value *V, Constant *C) {
282 IV.markForcedConstant(C);
283 DOUT << "markForcedConstant: " << *C << ": " << *V;
284 InstWorkList.push_back(V);
287 inline void markConstant(Value *V, Constant *C) {
288 markConstant(ValueState[V], V, C);
291 // markOverdefined - Make a value be marked as "overdefined". If the
292 // value is not already overdefined, add it to the overdefined instruction
293 // work list so that the users of the instruction are updated later.
295 inline void markOverdefined(LatticeVal &IV, Value *V) {
296 if (IV.markOverdefined()) {
297 DEBUG(DOUT << "markOverdefined: ";
298 if (Function *F = dyn_cast<Function>(V))
299 DOUT << "Function '" << F->getName() << "'\n";
302 // Only instructions go on the work list
303 OverdefinedInstWorkList.push_back(V);
307 inline void mergeInValue(LatticeVal &IV, Value *V, LatticeVal &MergeWithV) {
308 if (IV.isOverdefined() || MergeWithV.isUndefined())
310 if (MergeWithV.isOverdefined())
311 markOverdefined(IV, V);
312 else if (IV.isUndefined())
313 markConstant(IV, V, MergeWithV.getConstant());
314 else if (IV.getConstant() != MergeWithV.getConstant())
315 markOverdefined(IV, V);
318 inline void mergeInValue(Value *V, LatticeVal &MergeWithV) {
319 return mergeInValue(ValueState[V], V, MergeWithV);
323 // getValueState - Return the LatticeVal object that corresponds to the value.
324 // This function is necessary because not all values should start out in the
325 // underdefined state... Argument's should be overdefined, and
326 // constants should be marked as constants. If a value is not known to be an
327 // Instruction object, then use this accessor to get its value from the map.
329 inline LatticeVal &getValueState(Value *V) {
330 std::map<Value*, LatticeVal>::iterator I = ValueState.find(V);
331 if (I != ValueState.end()) return I->second; // Common case, in the map
333 if (Constant *C = dyn_cast<Constant>(V)) {
334 if (isa<UndefValue>(V)) {
335 // Nothing to do, remain undefined.
337 LatticeVal &LV = ValueState[C];
338 LV.markConstant(C); // Constants are constant
342 // All others are underdefined by default...
343 return ValueState[V];
346 // markEdgeExecutable - Mark a basic block as executable, adding it to the BB
347 // work list if it is not already executable...
349 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
350 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
351 return; // This edge is already known to be executable!
353 if (BBExecutable.count(Dest)) {
354 DOUT << "Marking Edge Executable: " << Source->getName()
355 << " -> " << Dest->getName() << "\n";
357 // The destination is already executable, but we just made an edge
358 // feasible that wasn't before. Revisit the PHI nodes in the block
359 // because they have potentially new operands.
360 for (BasicBlock::iterator I = Dest->begin(); isa<PHINode>(I); ++I)
361 visitPHINode(*cast<PHINode>(I));
364 MarkBlockExecutable(Dest);
368 // getFeasibleSuccessors - Return a vector of booleans to indicate which
369 // successors are reachable from a given terminator instruction.
371 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
373 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
374 // block to the 'To' basic block is currently feasible...
376 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
378 // OperandChangedState - This method is invoked on all of the users of an
379 // instruction that was just changed state somehow.... Based on this
380 // information, we need to update the specified user of this instruction.
382 void OperandChangedState(User *U) {
383 // Only instructions use other variable values!
384 Instruction &I = cast<Instruction>(*U);
385 if (BBExecutable.count(I.getParent())) // Inst is executable?
390 friend class InstVisitor<SCCPSolver>;
392 // visit implementations - Something changed in this instruction... Either an
393 // operand made a transition, or the instruction is newly executable. Change
394 // the value type of I to reflect these changes if appropriate.
396 void visitPHINode(PHINode &I);
399 void visitReturnInst(ReturnInst &I);
400 void visitTerminatorInst(TerminatorInst &TI);
402 void visitCastInst(CastInst &I);
403 void visitGetResultInst(GetResultInst &GRI);
404 void visitSelectInst(SelectInst &I);
405 void visitBinaryOperator(Instruction &I);
406 void visitCmpInst(CmpInst &I);
407 void visitExtractElementInst(ExtractElementInst &I);
408 void visitInsertElementInst(InsertElementInst &I);
409 void visitShuffleVectorInst(ShuffleVectorInst &I);
411 // Instructions that cannot be folded away...
412 void visitStoreInst (Instruction &I);
413 void visitLoadInst (LoadInst &I);
414 void visitGetElementPtrInst(GetElementPtrInst &I);
415 void visitCallInst (CallInst &I) { visitCallSite(CallSite::get(&I)); }
416 void visitInvokeInst (InvokeInst &II) {
417 visitCallSite(CallSite::get(&II));
418 visitTerminatorInst(II);
420 void visitCallSite (CallSite CS);
421 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ }
422 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
423 void visitAllocationInst(Instruction &I) { markOverdefined(&I); }
424 void visitVANextInst (Instruction &I) { markOverdefined(&I); }
425 void visitVAArgInst (Instruction &I) { markOverdefined(&I); }
426 void visitFreeInst (Instruction &I) { /*returns void*/ }
428 void visitInstruction(Instruction &I) {
429 // If a new instruction is added to LLVM that we don't handle...
430 cerr << "SCCP: Don't know how to handle: " << I;
431 markOverdefined(&I); // Just in case
435 } // end anonymous namespace
438 // getFeasibleSuccessors - Return a vector of booleans to indicate which
439 // successors are reachable from a given terminator instruction.
441 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
442 SmallVector<bool, 16> &Succs) {
443 Succs.resize(TI.getNumSuccessors());
444 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
445 if (BI->isUnconditional()) {
448 LatticeVal &BCValue = getValueState(BI->getCondition());
449 if (BCValue.isOverdefined() ||
450 (BCValue.isConstant() && !isa<ConstantInt>(BCValue.getConstant()))) {
451 // Overdefined condition variables, and branches on unfoldable constant
452 // conditions, mean the branch could go either way.
453 Succs[0] = Succs[1] = true;
454 } else if (BCValue.isConstant()) {
455 // Constant condition variables mean the branch can only go a single way
456 Succs[BCValue.getConstant() == ConstantInt::getFalse()] = true;
459 } else if (isa<InvokeInst>(&TI)) {
460 // Invoke instructions successors are always executable.
461 Succs[0] = Succs[1] = true;
462 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
463 LatticeVal &SCValue = getValueState(SI->getCondition());
464 if (SCValue.isOverdefined() || // Overdefined condition?
465 (SCValue.isConstant() && !isa<ConstantInt>(SCValue.getConstant()))) {
466 // All destinations are executable!
467 Succs.assign(TI.getNumSuccessors(), true);
468 } else if (SCValue.isConstant()) {
469 Constant *CPV = SCValue.getConstant();
470 // Make sure to skip the "default value" which isn't a value
471 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i) {
472 if (SI->getSuccessorValue(i) == CPV) {// Found the right branch...
478 // Constant value not equal to any of the branches... must execute
479 // default branch then...
483 assert(0 && "SCCP: Don't know how to handle this terminator!");
488 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
489 // block to the 'To' basic block is currently feasible...
491 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
492 assert(BBExecutable.count(To) && "Dest should always be alive!");
494 // Make sure the source basic block is executable!!
495 if (!BBExecutable.count(From)) return false;
497 // Check to make sure this edge itself is actually feasible now...
498 TerminatorInst *TI = From->getTerminator();
499 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
500 if (BI->isUnconditional())
503 LatticeVal &BCValue = getValueState(BI->getCondition());
504 if (BCValue.isOverdefined()) {
505 // Overdefined condition variables mean the branch could go either way.
507 } else if (BCValue.isConstant()) {
508 // Not branching on an evaluatable constant?
509 if (!isa<ConstantInt>(BCValue.getConstant())) return true;
511 // Constant condition variables mean the branch can only go a single way
512 return BI->getSuccessor(BCValue.getConstant() ==
513 ConstantInt::getFalse()) == To;
517 } else if (isa<InvokeInst>(TI)) {
518 // Invoke instructions successors are always executable.
520 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
521 LatticeVal &SCValue = getValueState(SI->getCondition());
522 if (SCValue.isOverdefined()) { // Overdefined condition?
523 // All destinations are executable!
525 } else if (SCValue.isConstant()) {
526 Constant *CPV = SCValue.getConstant();
527 if (!isa<ConstantInt>(CPV))
528 return true; // not a foldable constant?
530 // Make sure to skip the "default value" which isn't a value
531 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i)
532 if (SI->getSuccessorValue(i) == CPV) // Found the taken branch...
533 return SI->getSuccessor(i) == To;
535 // Constant value not equal to any of the branches... must execute
536 // default branch then...
537 return SI->getDefaultDest() == To;
541 cerr << "Unknown terminator instruction: " << *TI;
546 // visit Implementations - Something changed in this instruction... Either an
547 // operand made a transition, or the instruction is newly executable. Change
548 // the value type of I to reflect these changes if appropriate. This method
549 // makes sure to do the following actions:
551 // 1. If a phi node merges two constants in, and has conflicting value coming
552 // from different branches, or if the PHI node merges in an overdefined
553 // value, then the PHI node becomes overdefined.
554 // 2. If a phi node merges only constants in, and they all agree on value, the
555 // PHI node becomes a constant value equal to that.
556 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
557 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
558 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
559 // 6. If a conditional branch has a value that is constant, make the selected
560 // destination executable
561 // 7. If a conditional branch has a value that is overdefined, make all
562 // successors executable.
564 void SCCPSolver::visitPHINode(PHINode &PN) {
565 LatticeVal &PNIV = getValueState(&PN);
566 if (PNIV.isOverdefined()) {
567 // There may be instructions using this PHI node that are not overdefined
568 // themselves. If so, make sure that they know that the PHI node operand
570 std::multimap<PHINode*, Instruction*>::iterator I, E;
571 tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN);
573 SmallVector<Instruction*, 16> Users;
574 for (; I != E; ++I) Users.push_back(I->second);
575 while (!Users.empty()) {
580 return; // Quick exit
583 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
584 // and slow us down a lot. Just mark them overdefined.
585 if (PN.getNumIncomingValues() > 64) {
586 markOverdefined(PNIV, &PN);
590 // Look at all of the executable operands of the PHI node. If any of them
591 // are overdefined, the PHI becomes overdefined as well. If they are all
592 // constant, and they agree with each other, the PHI becomes the identical
593 // constant. If they are constant and don't agree, the PHI is overdefined.
594 // If there are no executable operands, the PHI remains undefined.
596 Constant *OperandVal = 0;
597 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
598 LatticeVal &IV = getValueState(PN.getIncomingValue(i));
599 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
601 if (isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) {
602 if (IV.isOverdefined()) { // PHI node becomes overdefined!
603 markOverdefined(PNIV, &PN);
607 if (OperandVal == 0) { // Grab the first value...
608 OperandVal = IV.getConstant();
609 } else { // Another value is being merged in!
610 // There is already a reachable operand. If we conflict with it,
611 // then the PHI node becomes overdefined. If we agree with it, we
614 // Check to see if there are two different constants merging...
615 if (IV.getConstant() != OperandVal) {
616 // Yes there is. This means the PHI node is not constant.
617 // You must be overdefined poor PHI.
619 markOverdefined(PNIV, &PN); // The PHI node now becomes overdefined
620 return; // I'm done analyzing you
626 // If we exited the loop, this means that the PHI node only has constant
627 // arguments that agree with each other(and OperandVal is the constant) or
628 // OperandVal is null because there are no defined incoming arguments. If
629 // this is the case, the PHI remains undefined.
632 markConstant(PNIV, &PN, OperandVal); // Acquire operand value
635 void SCCPSolver::visitReturnInst(ReturnInst &I) {
636 if (I.getNumOperands() == 0) return; // Ret void
638 Function *F = I.getParent()->getParent();
639 // If we are tracking the return value of this function, merge it in.
640 if (!F->hasInternalLinkage())
643 if (!TrackedRetVals.empty()) {
644 DenseMap<Function*, LatticeVal>::iterator TFRVI =
645 TrackedRetVals.find(F);
646 if (TFRVI != TrackedRetVals.end() &&
647 !TFRVI->second.isOverdefined()) {
648 LatticeVal &IV = getValueState(I.getOperand(0));
649 mergeInValue(TFRVI->second, F, IV);
654 // Handle function that returns multiple values.
655 std::multimap<Function*, LatticeValIndexed>::iterator It, E;
656 tie(It, E) = TrackedMultipleRetVals.equal_range(F);
658 for (; It != E; ++It) {
659 LatticeValIndexed &LV = It->second;
660 unsigned Idx = LV.getIndex();
661 Value *V = I.getOperand(Idx);
662 mergeInValue(LV.getLatticeVal(), V, getValueState(V));
667 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
668 SmallVector<bool, 16> SuccFeasible;
669 getFeasibleSuccessors(TI, SuccFeasible);
671 BasicBlock *BB = TI.getParent();
673 // Mark all feasible successors executable...
674 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
676 markEdgeExecutable(BB, TI.getSuccessor(i));
679 void SCCPSolver::visitCastInst(CastInst &I) {
680 Value *V = I.getOperand(0);
681 LatticeVal &VState = getValueState(V);
682 if (VState.isOverdefined()) // Inherit overdefinedness of operand
684 else if (VState.isConstant()) // Propagate constant value
685 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
686 VState.getConstant(), I.getType()));
689 void SCCPSolver::visitGetResultInst(GetResultInst &GRI) {
690 unsigned Idx = GRI.getIndex();
691 Value *Aggr = GRI.getOperand(0);
693 if (CallInst *CI = dyn_cast<CallInst>(Aggr))
694 F = CI->getCalledFunction();
695 else if (InvokeInst *II = dyn_cast<InvokeInst>(Aggr))
696 F = II->getCalledFunction();
698 assert (F && "Invalid GetResultInst operands!");
700 std::multimap<Function*, LatticeValIndexed>::iterator It, E;
701 tie(It, E) = TrackedMultipleRetVals.equal_range(F);
705 for (; It != E; ++It) {
706 LatticeValIndexed &LIV = It->second;
707 if (LIV.getIndex() == Idx) {
708 mergeInValue(&GRI, LIV.getLatticeVal());
713 void SCCPSolver::visitSelectInst(SelectInst &I) {
714 LatticeVal &CondValue = getValueState(I.getCondition());
715 if (CondValue.isUndefined())
717 if (CondValue.isConstant()) {
718 if (ConstantInt *CondCB = dyn_cast<ConstantInt>(CondValue.getConstant())){
719 mergeInValue(&I, getValueState(CondCB->getZExtValue() ? I.getTrueValue()
720 : I.getFalseValue()));
725 // Otherwise, the condition is overdefined or a constant we can't evaluate.
726 // See if we can produce something better than overdefined based on the T/F
728 LatticeVal &TVal = getValueState(I.getTrueValue());
729 LatticeVal &FVal = getValueState(I.getFalseValue());
731 // select ?, C, C -> C.
732 if (TVal.isConstant() && FVal.isConstant() &&
733 TVal.getConstant() == FVal.getConstant()) {
734 markConstant(&I, FVal.getConstant());
738 if (TVal.isUndefined()) { // select ?, undef, X -> X.
739 mergeInValue(&I, FVal);
740 } else if (FVal.isUndefined()) { // select ?, X, undef -> X.
741 mergeInValue(&I, TVal);
747 // Handle BinaryOperators and Shift Instructions...
748 void SCCPSolver::visitBinaryOperator(Instruction &I) {
749 LatticeVal &IV = ValueState[&I];
750 if (IV.isOverdefined()) return;
752 LatticeVal &V1State = getValueState(I.getOperand(0));
753 LatticeVal &V2State = getValueState(I.getOperand(1));
755 if (V1State.isOverdefined() || V2State.isOverdefined()) {
756 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
757 // operand is overdefined.
758 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
759 LatticeVal *NonOverdefVal = 0;
760 if (!V1State.isOverdefined()) {
761 NonOverdefVal = &V1State;
762 } else if (!V2State.isOverdefined()) {
763 NonOverdefVal = &V2State;
767 if (NonOverdefVal->isUndefined()) {
768 // Could annihilate value.
769 if (I.getOpcode() == Instruction::And)
770 markConstant(IV, &I, Constant::getNullValue(I.getType()));
771 else if (const VectorType *PT = dyn_cast<VectorType>(I.getType()))
772 markConstant(IV, &I, ConstantVector::getAllOnesValue(PT));
774 markConstant(IV, &I, ConstantInt::getAllOnesValue(I.getType()));
777 if (I.getOpcode() == Instruction::And) {
778 if (NonOverdefVal->getConstant()->isNullValue()) {
779 markConstant(IV, &I, NonOverdefVal->getConstant());
780 return; // X and 0 = 0
783 if (ConstantInt *CI =
784 dyn_cast<ConstantInt>(NonOverdefVal->getConstant()))
785 if (CI->isAllOnesValue()) {
786 markConstant(IV, &I, NonOverdefVal->getConstant());
787 return; // X or -1 = -1
795 // If both operands are PHI nodes, it is possible that this instruction has
796 // a constant value, despite the fact that the PHI node doesn't. Check for
797 // this condition now.
798 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
799 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
800 if (PN1->getParent() == PN2->getParent()) {
801 // Since the two PHI nodes are in the same basic block, they must have
802 // entries for the same predecessors. Walk the predecessor list, and
803 // if all of the incoming values are constants, and the result of
804 // evaluating this expression with all incoming value pairs is the
805 // same, then this expression is a constant even though the PHI node
806 // is not a constant!
808 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
809 LatticeVal &In1 = getValueState(PN1->getIncomingValue(i));
810 BasicBlock *InBlock = PN1->getIncomingBlock(i);
812 getValueState(PN2->getIncomingValueForBlock(InBlock));
814 if (In1.isOverdefined() || In2.isOverdefined()) {
815 Result.markOverdefined();
816 break; // Cannot fold this operation over the PHI nodes!
817 } else if (In1.isConstant() && In2.isConstant()) {
818 Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(),
820 if (Result.isUndefined())
821 Result.markConstant(V);
822 else if (Result.isConstant() && Result.getConstant() != V) {
823 Result.markOverdefined();
829 // If we found a constant value here, then we know the instruction is
830 // constant despite the fact that the PHI nodes are overdefined.
831 if (Result.isConstant()) {
832 markConstant(IV, &I, Result.getConstant());
833 // Remember that this instruction is virtually using the PHI node
835 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
836 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
838 } else if (Result.isUndefined()) {
842 // Okay, this really is overdefined now. Since we might have
843 // speculatively thought that this was not overdefined before, and
844 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
845 // make sure to clean out any entries that we put there, for
847 std::multimap<PHINode*, Instruction*>::iterator It, E;
848 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1);
850 if (It->second == &I) {
851 UsersOfOverdefinedPHIs.erase(It++);
855 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2);
857 if (It->second == &I) {
858 UsersOfOverdefinedPHIs.erase(It++);
864 markOverdefined(IV, &I);
865 } else if (V1State.isConstant() && V2State.isConstant()) {
866 markConstant(IV, &I, ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
867 V2State.getConstant()));
871 // Handle ICmpInst instruction...
872 void SCCPSolver::visitCmpInst(CmpInst &I) {
873 LatticeVal &IV = ValueState[&I];
874 if (IV.isOverdefined()) return;
876 LatticeVal &V1State = getValueState(I.getOperand(0));
877 LatticeVal &V2State = getValueState(I.getOperand(1));
879 if (V1State.isOverdefined() || V2State.isOverdefined()) {
880 // If both operands are PHI nodes, it is possible that this instruction has
881 // a constant value, despite the fact that the PHI node doesn't. Check for
882 // this condition now.
883 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
884 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
885 if (PN1->getParent() == PN2->getParent()) {
886 // Since the two PHI nodes are in the same basic block, they must have
887 // entries for the same predecessors. Walk the predecessor list, and
888 // if all of the incoming values are constants, and the result of
889 // evaluating this expression with all incoming value pairs is the
890 // same, then this expression is a constant even though the PHI node
891 // is not a constant!
893 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
894 LatticeVal &In1 = getValueState(PN1->getIncomingValue(i));
895 BasicBlock *InBlock = PN1->getIncomingBlock(i);
897 getValueState(PN2->getIncomingValueForBlock(InBlock));
899 if (In1.isOverdefined() || In2.isOverdefined()) {
900 Result.markOverdefined();
901 break; // Cannot fold this operation over the PHI nodes!
902 } else if (In1.isConstant() && In2.isConstant()) {
903 Constant *V = ConstantExpr::getCompare(I.getPredicate(),
906 if (Result.isUndefined())
907 Result.markConstant(V);
908 else if (Result.isConstant() && Result.getConstant() != V) {
909 Result.markOverdefined();
915 // If we found a constant value here, then we know the instruction is
916 // constant despite the fact that the PHI nodes are overdefined.
917 if (Result.isConstant()) {
918 markConstant(IV, &I, Result.getConstant());
919 // Remember that this instruction is virtually using the PHI node
921 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
922 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
924 } else if (Result.isUndefined()) {
928 // Okay, this really is overdefined now. Since we might have
929 // speculatively thought that this was not overdefined before, and
930 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
931 // make sure to clean out any entries that we put there, for
933 std::multimap<PHINode*, Instruction*>::iterator It, E;
934 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1);
936 if (It->second == &I) {
937 UsersOfOverdefinedPHIs.erase(It++);
941 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2);
943 if (It->second == &I) {
944 UsersOfOverdefinedPHIs.erase(It++);
950 markOverdefined(IV, &I);
951 } else if (V1State.isConstant() && V2State.isConstant()) {
952 markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
953 V1State.getConstant(),
954 V2State.getConstant()));
958 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
959 // FIXME : SCCP does not handle vectors properly.
964 LatticeVal &ValState = getValueState(I.getOperand(0));
965 LatticeVal &IdxState = getValueState(I.getOperand(1));
967 if (ValState.isOverdefined() || IdxState.isOverdefined())
969 else if(ValState.isConstant() && IdxState.isConstant())
970 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
971 IdxState.getConstant()));
975 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
976 // FIXME : SCCP does not handle vectors properly.
980 LatticeVal &ValState = getValueState(I.getOperand(0));
981 LatticeVal &EltState = getValueState(I.getOperand(1));
982 LatticeVal &IdxState = getValueState(I.getOperand(2));
984 if (ValState.isOverdefined() || EltState.isOverdefined() ||
985 IdxState.isOverdefined())
987 else if(ValState.isConstant() && EltState.isConstant() &&
988 IdxState.isConstant())
989 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
990 EltState.getConstant(),
991 IdxState.getConstant()));
992 else if (ValState.isUndefined() && EltState.isConstant() &&
993 IdxState.isConstant())
994 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
995 EltState.getConstant(),
996 IdxState.getConstant()));
1000 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
1001 // FIXME : SCCP does not handle vectors properly.
1002 markOverdefined(&I);
1005 LatticeVal &V1State = getValueState(I.getOperand(0));
1006 LatticeVal &V2State = getValueState(I.getOperand(1));
1007 LatticeVal &MaskState = getValueState(I.getOperand(2));
1009 if (MaskState.isUndefined() ||
1010 (V1State.isUndefined() && V2State.isUndefined()))
1011 return; // Undefined output if mask or both inputs undefined.
1013 if (V1State.isOverdefined() || V2State.isOverdefined() ||
1014 MaskState.isOverdefined()) {
1015 markOverdefined(&I);
1017 // A mix of constant/undef inputs.
1018 Constant *V1 = V1State.isConstant() ?
1019 V1State.getConstant() : UndefValue::get(I.getType());
1020 Constant *V2 = V2State.isConstant() ?
1021 V2State.getConstant() : UndefValue::get(I.getType());
1022 Constant *Mask = MaskState.isConstant() ?
1023 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
1024 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
1029 // Handle getelementptr instructions... if all operands are constants then we
1030 // can turn this into a getelementptr ConstantExpr.
1032 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1033 LatticeVal &IV = ValueState[&I];
1034 if (IV.isOverdefined()) return;
1036 SmallVector<Constant*, 8> Operands;
1037 Operands.reserve(I.getNumOperands());
1039 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1040 LatticeVal &State = getValueState(I.getOperand(i));
1041 if (State.isUndefined())
1042 return; // Operands are not resolved yet...
1043 else if (State.isOverdefined()) {
1044 markOverdefined(IV, &I);
1047 assert(State.isConstant() && "Unknown state!");
1048 Operands.push_back(State.getConstant());
1051 Constant *Ptr = Operands[0];
1052 Operands.erase(Operands.begin()); // Erase the pointer from idx list...
1054 markConstant(IV, &I, ConstantExpr::getGetElementPtr(Ptr, &Operands[0],
1058 void SCCPSolver::visitStoreInst(Instruction &SI) {
1059 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1061 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1062 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1063 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1065 // Get the value we are storing into the global.
1066 LatticeVal &PtrVal = getValueState(SI.getOperand(0));
1068 mergeInValue(I->second, GV, PtrVal);
1069 if (I->second.isOverdefined())
1070 TrackedGlobals.erase(I); // No need to keep tracking this!
1074 // Handle load instructions. If the operand is a constant pointer to a constant
1075 // global, we can replace the load with the loaded constant value!
1076 void SCCPSolver::visitLoadInst(LoadInst &I) {
1077 LatticeVal &IV = ValueState[&I];
1078 if (IV.isOverdefined()) return;
1080 LatticeVal &PtrVal = getValueState(I.getOperand(0));
1081 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1082 if (PtrVal.isConstant() && !I.isVolatile()) {
1083 Value *Ptr = PtrVal.getConstant();
1084 // TODO: Consider a target hook for valid address spaces for this xform.
1085 if (isa<ConstantPointerNull>(Ptr) &&
1086 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
1087 // load null -> null
1088 markConstant(IV, &I, Constant::getNullValue(I.getType()));
1092 // Transform load (constant global) into the value loaded.
1093 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1094 if (GV->isConstant()) {
1095 if (!GV->isDeclaration()) {
1096 markConstant(IV, &I, GV->getInitializer());
1099 } else if (!TrackedGlobals.empty()) {
1100 // If we are tracking this global, merge in the known value for it.
1101 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1102 TrackedGlobals.find(GV);
1103 if (It != TrackedGlobals.end()) {
1104 mergeInValue(IV, &I, It->second);
1110 // Transform load (constantexpr_GEP global, 0, ...) into the value loaded.
1111 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
1112 if (CE->getOpcode() == Instruction::GetElementPtr)
1113 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
1114 if (GV->isConstant() && !GV->isDeclaration())
1116 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) {
1117 markConstant(IV, &I, V);
1122 // Otherwise we cannot say for certain what value this load will produce.
1124 markOverdefined(IV, &I);
1127 void SCCPSolver::visitCallSite(CallSite CS) {
1128 Function *F = CS.getCalledFunction();
1130 DenseMap<Function*, LatticeVal>::iterator TFRVI =TrackedRetVals.end();
1131 // If we are tracking this function, we must make sure to bind arguments as
1133 bool FirstCall = false;
1134 if (F && F->hasInternalLinkage()) {
1135 TFRVI = TrackedRetVals.find(F);
1136 if (TFRVI != TrackedRetVals.end())
1139 std::multimap<Function*, LatticeValIndexed>::iterator It, E;
1140 tie(It, E) = TrackedMultipleRetVals.equal_range(F);
1147 // If this is the first call to the function hit, mark its entry block
1149 if (!BBExecutable.count(F->begin()))
1150 MarkBlockExecutable(F->begin());
1152 CallSite::arg_iterator CAI = CS.arg_begin();
1153 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1154 AI != E; ++AI, ++CAI) {
1155 LatticeVal &IV = ValueState[AI];
1156 if (!IV.isOverdefined())
1157 mergeInValue(IV, AI, getValueState(*CAI));
1160 Instruction *I = CS.getInstruction();
1162 if (!CS.doesNotThrow() && I->getParent()->getUnwindDest())
1163 markEdgeExecutable(I->getParent(), I->getParent()->getUnwindDest());
1165 if (I->getType() == Type::VoidTy) return;
1167 LatticeVal &IV = ValueState[I];
1168 if (IV.isOverdefined()) return;
1170 // Propagate the single return value of the function to the value of the
1172 if (TFRVI != TrackedRetVals.end()) {
1173 mergeInValue(IV, I, TFRVI->second);
1177 if (F == 0 || !F->isDeclaration() || !canConstantFoldCallTo(F)) {
1178 markOverdefined(IV, I);
1182 SmallVector<Constant*, 8> Operands;
1183 Operands.reserve(I->getNumOperands()-1);
1185 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1187 LatticeVal &State = getValueState(*AI);
1188 if (State.isUndefined())
1189 return; // Operands are not resolved yet...
1190 else if (State.isOverdefined()) {
1191 markOverdefined(IV, I);
1194 assert(State.isConstant() && "Unknown state!");
1195 Operands.push_back(State.getConstant());
1198 if (Constant *C = ConstantFoldCall(F, &Operands[0], Operands.size()))
1199 markConstant(IV, I, C);
1201 markOverdefined(IV, I);
1205 void SCCPSolver::Solve() {
1206 // Process the work lists until they are empty!
1207 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1208 !OverdefinedInstWorkList.empty()) {
1209 // Process the instruction work list...
1210 while (!OverdefinedInstWorkList.empty()) {
1211 Value *I = OverdefinedInstWorkList.back();
1212 OverdefinedInstWorkList.pop_back();
1214 DOUT << "\nPopped off OI-WL: " << *I;
1216 // "I" got into the work list because it either made the transition from
1217 // bottom to constant
1219 // Anything on this worklist that is overdefined need not be visited
1220 // since all of its users will have already been marked as overdefined
1221 // Update all of the users of this instruction's value...
1223 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1225 OperandChangedState(*UI);
1227 // Process the instruction work list...
1228 while (!InstWorkList.empty()) {
1229 Value *I = InstWorkList.back();
1230 InstWorkList.pop_back();
1232 DOUT << "\nPopped off I-WL: " << *I;
1234 // "I" got into the work list because it either made the transition from
1235 // bottom to constant
1237 // Anything on this worklist that is overdefined need not be visited
1238 // since all of its users will have already been marked as overdefined.
1239 // Update all of the users of this instruction's value...
1241 if (!getValueState(I).isOverdefined())
1242 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1244 OperandChangedState(*UI);
1247 // Process the basic block work list...
1248 while (!BBWorkList.empty()) {
1249 BasicBlock *BB = BBWorkList.back();
1250 BBWorkList.pop_back();
1252 DOUT << "\nPopped off BBWL: " << *BB;
1254 // Notify all instructions in this basic block that they are newly
1261 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1262 /// that branches on undef values cannot reach any of their successors.
1263 /// However, this is not a safe assumption. After we solve dataflow, this
1264 /// method should be use to handle this. If this returns true, the solver
1265 /// should be rerun.
1267 /// This method handles this by finding an unresolved branch and marking it one
1268 /// of the edges from the block as being feasible, even though the condition
1269 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1270 /// CFG and only slightly pessimizes the analysis results (by marking one,
1271 /// potentially infeasible, edge feasible). This cannot usefully modify the
1272 /// constraints on the condition of the branch, as that would impact other users
1275 /// This scan also checks for values that use undefs, whose results are actually
1276 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1277 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1278 /// even if X isn't defined.
1279 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1280 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1281 if (!BBExecutable.count(BB))
1284 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1285 // Look for instructions which produce undef values.
1286 if (I->getType() == Type::VoidTy) continue;
1288 LatticeVal &LV = getValueState(I);
1289 if (!LV.isUndefined()) continue;
1291 // Get the lattice values of the first two operands for use below.
1292 LatticeVal &Op0LV = getValueState(I->getOperand(0));
1294 if (I->getNumOperands() == 2) {
1295 // If this is a two-operand instruction, and if both operands are
1296 // undefs, the result stays undef.
1297 Op1LV = getValueState(I->getOperand(1));
1298 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1302 // If this is an instructions whose result is defined even if the input is
1303 // not fully defined, propagate the information.
1304 const Type *ITy = I->getType();
1305 switch (I->getOpcode()) {
1306 default: break; // Leave the instruction as an undef.
1307 case Instruction::ZExt:
1308 // After a zero extend, we know the top part is zero. SExt doesn't have
1309 // to be handled here, because we don't know whether the top part is 1's
1311 assert(Op0LV.isUndefined());
1312 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1314 case Instruction::Mul:
1315 case Instruction::And:
1316 // undef * X -> 0. X could be zero.
1317 // undef & X -> 0. X could be zero.
1318 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1321 case Instruction::Or:
1322 // undef | X -> -1. X could be -1.
1323 if (const VectorType *PTy = dyn_cast<VectorType>(ITy))
1324 markForcedConstant(LV, I, ConstantVector::getAllOnesValue(PTy));
1326 markForcedConstant(LV, I, ConstantInt::getAllOnesValue(ITy));
1329 case Instruction::SDiv:
1330 case Instruction::UDiv:
1331 case Instruction::SRem:
1332 case Instruction::URem:
1333 // X / undef -> undef. No change.
1334 // X % undef -> undef. No change.
1335 if (Op1LV.isUndefined()) break;
1337 // undef / X -> 0. X could be maxint.
1338 // undef % X -> 0. X could be 1.
1339 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1342 case Instruction::AShr:
1343 // undef >>s X -> undef. No change.
1344 if (Op0LV.isUndefined()) break;
1346 // X >>s undef -> X. X could be 0, X could have the high-bit known set.
1347 if (Op0LV.isConstant())
1348 markForcedConstant(LV, I, Op0LV.getConstant());
1350 markOverdefined(LV, I);
1352 case Instruction::LShr:
1353 case Instruction::Shl:
1354 // undef >> X -> undef. No change.
1355 // undef << X -> undef. No change.
1356 if (Op0LV.isUndefined()) break;
1358 // X >> undef -> 0. X could be 0.
1359 // X << undef -> 0. X could be 0.
1360 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1362 case Instruction::Select:
1363 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1364 if (Op0LV.isUndefined()) {
1365 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1366 Op1LV = getValueState(I->getOperand(2));
1367 } else if (Op1LV.isUndefined()) {
1368 // c ? undef : undef -> undef. No change.
1369 Op1LV = getValueState(I->getOperand(2));
1370 if (Op1LV.isUndefined())
1372 // Otherwise, c ? undef : x -> x.
1374 // Leave Op1LV as Operand(1)'s LatticeValue.
1377 if (Op1LV.isConstant())
1378 markForcedConstant(LV, I, Op1LV.getConstant());
1380 markOverdefined(LV, I);
1385 TerminatorInst *TI = BB->getTerminator();
1386 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1387 if (!BI->isConditional()) continue;
1388 if (!getValueState(BI->getCondition()).isUndefined())
1390 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1391 if (!getValueState(SI->getCondition()).isUndefined())
1397 // If the edge to the second successor isn't thought to be feasible yet,
1398 // mark it so now. We pick the second one so that this goes to some
1399 // enumerated value in a switch instead of going to the default destination.
1400 if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(1))))
1403 // Otherwise, it isn't already thought to be feasible. Mark it as such now
1404 // and return. This will make other blocks reachable, which will allow new
1405 // values to be discovered and existing ones to be moved in the lattice.
1406 markEdgeExecutable(BB, TI->getSuccessor(1));
1408 // This must be a conditional branch of switch on undef. At this point,
1409 // force the old terminator to branch to the first successor. This is
1410 // required because we are now influencing the dataflow of the function with
1411 // the assumption that this edge is taken. If we leave the branch condition
1412 // as undef, then further analysis could think the undef went another way
1413 // leading to an inconsistent set of conclusions.
1414 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1415 BI->setCondition(ConstantInt::getFalse());
1417 SwitchInst *SI = cast<SwitchInst>(TI);
1418 SI->setCondition(SI->getCaseValue(1));
1429 //===--------------------------------------------------------------------===//
1431 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1432 /// Sparse Conditional Constant Propagator.
1434 struct VISIBILITY_HIDDEN SCCP : public FunctionPass {
1435 static char ID; // Pass identification, replacement for typeid
1436 SCCP() : FunctionPass((intptr_t)&ID) {}
1438 // runOnFunction - Run the Sparse Conditional Constant Propagation
1439 // algorithm, and return true if the function was modified.
1441 bool runOnFunction(Function &F);
1443 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1444 AU.setPreservesCFG();
1449 RegisterPass<SCCP> X("sccp", "Sparse Conditional Constant Propagation");
1450 } // end anonymous namespace
1453 // createSCCPPass - This is the public interface to this file...
1454 FunctionPass *llvm::createSCCPPass() {
1459 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1460 // and return true if the function was modified.
1462 bool SCCP::runOnFunction(Function &F) {
1463 DOUT << "SCCP on function '" << F.getName() << "'\n";
1466 // Mark the first block of the function as being executable.
1467 Solver.MarkBlockExecutable(F.begin());
1469 // Mark all arguments to the function as being overdefined.
1470 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1471 Solver.markOverdefined(AI);
1473 // Solve for constants.
1474 bool ResolvedUndefs = true;
1475 while (ResolvedUndefs) {
1477 DOUT << "RESOLVING UNDEFs\n";
1478 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1481 bool MadeChanges = false;
1483 // If we decided that there are basic blocks that are dead in this function,
1484 // delete their contents now. Note that we cannot actually delete the blocks,
1485 // as we cannot modify the CFG of the function.
1487 SmallSet<BasicBlock*, 16> &ExecutableBBs = Solver.getExecutableBlocks();
1488 SmallVector<Instruction*, 32> Insts;
1489 std::map<Value*, LatticeVal> &Values = Solver.getValueMapping();
1491 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1492 if (!ExecutableBBs.count(BB)) {
1493 DOUT << " BasicBlock Dead:" << *BB;
1496 // Delete the instructions backwards, as it has a reduced likelihood of
1497 // having to update as many def-use and use-def chains.
1498 for (BasicBlock::iterator I = BB->begin(), E = BB->getTerminator();
1501 while (!Insts.empty()) {
1502 Instruction *I = Insts.back();
1504 if (!I->use_empty())
1505 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1506 BB->getInstList().erase(I);
1511 // Iterate over all of the instructions in a function, replacing them with
1512 // constants if we have found them to be of constant values.
1514 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1515 Instruction *Inst = BI++;
1516 if (Inst->getType() != Type::VoidTy) {
1517 LatticeVal &IV = Values[Inst];
1518 if ((IV.isConstant() || IV.isUndefined()) &&
1519 !isa<TerminatorInst>(Inst)) {
1520 Constant *Const = IV.isConstant()
1521 ? IV.getConstant() : UndefValue::get(Inst->getType());
1522 DOUT << " Constant: " << *Const << " = " << *Inst;
1524 // Replaces all of the uses of a variable with uses of the constant.
1525 Inst->replaceAllUsesWith(Const);
1527 // Delete the instruction.
1528 BB->getInstList().erase(Inst);
1530 // Hey, we just changed something!
1542 //===--------------------------------------------------------------------===//
1544 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1545 /// Constant Propagation.
1547 struct VISIBILITY_HIDDEN IPSCCP : public ModulePass {
1549 IPSCCP() : ModulePass((intptr_t)&ID) {}
1550 bool runOnModule(Module &M);
1553 char IPSCCP::ID = 0;
1554 RegisterPass<IPSCCP>
1555 Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation");
1556 } // end anonymous namespace
1558 // createIPSCCPPass - This is the public interface to this file...
1559 ModulePass *llvm::createIPSCCPPass() {
1560 return new IPSCCP();
1564 static bool AddressIsTaken(GlobalValue *GV) {
1565 // Delete any dead constantexpr klingons.
1566 GV->removeDeadConstantUsers();
1568 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
1570 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1571 if (SI->getOperand(0) == GV || SI->isVolatile())
1572 return true; // Storing addr of GV.
1573 } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) {
1574 // Make sure we are calling the function, not passing the address.
1575 CallSite CS = CallSite::get(cast<Instruction>(*UI));
1576 for (CallSite::arg_iterator AI = CS.arg_begin(),
1577 E = CS.arg_end(); AI != E; ++AI)
1580 } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1581 if (LI->isVolatile())
1589 bool IPSCCP::runOnModule(Module &M) {
1592 // Loop over all functions, marking arguments to those with their addresses
1593 // taken or that are external as overdefined.
1595 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1596 if (!F->hasInternalLinkage() || AddressIsTaken(F)) {
1597 if (!F->isDeclaration())
1598 Solver.MarkBlockExecutable(F->begin());
1599 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1601 Solver.markOverdefined(AI);
1603 Solver.AddTrackedFunction(F);
1606 // Loop over global variables. We inform the solver about any internal global
1607 // variables that do not have their 'addresses taken'. If they don't have
1608 // their addresses taken, we can propagate constants through them.
1609 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1611 if (!G->isConstant() && G->hasInternalLinkage() && !AddressIsTaken(G))
1612 Solver.TrackValueOfGlobalVariable(G);
1614 // Solve for constants.
1615 bool ResolvedUndefs = true;
1616 while (ResolvedUndefs) {
1619 DOUT << "RESOLVING UNDEFS\n";
1620 ResolvedUndefs = false;
1621 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1622 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1625 bool MadeChanges = false;
1627 // Iterate over all of the instructions in the module, replacing them with
1628 // constants if we have found them to be of constant values.
1630 SmallSet<BasicBlock*, 16> &ExecutableBBs = Solver.getExecutableBlocks();
1631 SmallVector<Instruction*, 32> Insts;
1632 SmallVector<BasicBlock*, 32> BlocksToErase;
1633 std::map<Value*, LatticeVal> &Values = Solver.getValueMapping();
1635 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1636 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1638 if (!AI->use_empty()) {
1639 LatticeVal &IV = Values[AI];
1640 if (IV.isConstant() || IV.isUndefined()) {
1641 Constant *CST = IV.isConstant() ?
1642 IV.getConstant() : UndefValue::get(AI->getType());
1643 DOUT << "*** Arg " << *AI << " = " << *CST <<"\n";
1645 // Replaces all of the uses of a variable with uses of the
1647 AI->replaceAllUsesWith(CST);
1652 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1653 if (!ExecutableBBs.count(BB)) {
1654 DOUT << " BasicBlock Dead:" << *BB;
1657 // Delete the instructions backwards, as it has a reduced likelihood of
1658 // having to update as many def-use and use-def chains.
1659 TerminatorInst *TI = BB->getTerminator();
1660 for (BasicBlock::iterator I = BB->begin(), E = TI; I != E; ++I)
1663 while (!Insts.empty()) {
1664 Instruction *I = Insts.back();
1666 if (!I->use_empty())
1667 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1668 BB->getInstList().erase(I);
1673 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1674 BasicBlock *Succ = TI->getSuccessor(i);
1675 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1676 TI->getSuccessor(i)->removePredecessor(BB);
1678 if (!TI->use_empty())
1679 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1680 BB->getInstList().erase(TI);
1682 if (&*BB != &F->front())
1683 BlocksToErase.push_back(BB);
1685 new UnreachableInst(BB);
1688 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1689 Instruction *Inst = BI++;
1690 if (Inst->getType() != Type::VoidTy) {
1691 LatticeVal &IV = Values[Inst];
1692 if (IV.isConstant() ||
1693 (IV.isUndefined() && !isa<TerminatorInst>(Inst))) {
1694 Constant *Const = IV.isConstant()
1695 ? IV.getConstant() : UndefValue::get(Inst->getType());
1696 DOUT << " Constant: " << *Const << " = " << *Inst;
1698 // Replaces all of the uses of a variable with uses of the
1700 Inst->replaceAllUsesWith(Const);
1702 // Delete the instruction.
1703 if (!isa<TerminatorInst>(Inst) && !isa<CallInst>(Inst))
1704 BB->getInstList().erase(Inst);
1706 // Hey, we just changed something!
1714 // Now that all instructions in the function are constant folded, erase dead
1715 // blocks, because we can now use ConstantFoldTerminator to get rid of
1717 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1718 // If there are any PHI nodes in this successor, drop entries for BB now.
1719 BasicBlock *DeadBB = BlocksToErase[i];
1720 while (!DeadBB->use_empty()) {
1721 if (BasicBlock *PredBB = dyn_cast<BasicBlock>(DeadBB->use_back())) {
1722 PredBB->setUnwindDest(NULL);
1726 Instruction *I = cast<Instruction>(DeadBB->use_back());
1727 bool Folded = ConstantFoldTerminator(I->getParent());
1729 // The constant folder may not have been able to fold the terminator
1730 // if this is a branch or switch on undef. Fold it manually as a
1731 // branch to the first successor.
1732 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1733 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1734 "Branch should be foldable!");
1735 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1736 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1738 assert(0 && "Didn't fold away reference to block!");
1741 // Make this an uncond branch to the first successor.
1742 TerminatorInst *TI = I->getParent()->getTerminator();
1743 BranchInst::Create(TI->getSuccessor(0), TI);
1745 // Remove entries in successor phi nodes to remove edges.
1746 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1747 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1749 // Remove the old terminator.
1750 TI->eraseFromParent();
1754 // Finally, delete the basic block.
1755 F->getBasicBlockList().erase(DeadBB);
1757 BlocksToErase.clear();
1760 // If we inferred constant or undef return values for a function, we replaced
1761 // all call uses with the inferred value. This means we don't need to bother
1762 // actually returning anything from the function. Replace all return
1763 // instructions with return undef.
1764 // TODO: Process multiple value ret instructions also.
1765 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1766 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1767 E = RV.end(); I != E; ++I)
1768 if (!I->second.isOverdefined() &&
1769 I->first->getReturnType() != Type::VoidTy) {
1770 Function *F = I->first;
1771 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1772 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1773 if (!isa<UndefValue>(RI->getOperand(0)))
1774 RI->setOperand(0, UndefValue::get(F->getReturnType()));
1777 // If we infered constant or undef values for globals variables, we can delete
1778 // the global and any stores that remain to it.
1779 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1780 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1781 E = TG.end(); I != E; ++I) {
1782 GlobalVariable *GV = I->first;
1783 assert(!I->second.isOverdefined() &&
1784 "Overdefined values should have been taken out of the map!");
1785 DOUT << "Found that GV '" << GV->getName()<< "' is constant!\n";
1786 while (!GV->use_empty()) {
1787 StoreInst *SI = cast<StoreInst>(GV->use_back());
1788 SI->eraseFromParent();
1790 M.getGlobalList().erase(GV);