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/Analysis/ValueTracking.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 #include "llvm/Support/CallSite.h"
35 #include "llvm/Support/Compiler.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/InstVisitor.h"
38 #include "llvm/ADT/DenseMap.h"
39 #include "llvm/ADT/SmallSet.h"
40 #include "llvm/ADT/SmallVector.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/ADT/STLExtras.h"
47 STATISTIC(NumInstRemoved, "Number of instructions removed");
48 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
50 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP");
51 STATISTIC(IPNumDeadBlocks , "Number of basic blocks unreachable by IPSCCP");
52 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
53 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
56 /// LatticeVal class - This class represents the different lattice values that
57 /// an LLVM value may occupy. It is a simple class with value semantics.
59 class VISIBILITY_HIDDEN LatticeVal {
61 /// undefined - This LLVM Value has no known value yet.
64 /// constant - This LLVM Value has a specific constant value.
67 /// forcedconstant - This LLVM Value was thought to be undef until
68 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged
69 /// with another (different) constant, it goes to overdefined, instead of
73 /// overdefined - This instruction is not known to be constant, and we know
76 } LatticeValue; // The current lattice position
78 Constant *ConstantVal; // If Constant value, the current value
80 inline LatticeVal() : LatticeValue(undefined), ConstantVal(0) {}
82 // markOverdefined - Return true if this is a new status to be in...
83 inline bool markOverdefined() {
84 if (LatticeValue != overdefined) {
85 LatticeValue = overdefined;
91 // markConstant - Return true if this is a new status for us.
92 inline bool markConstant(Constant *V) {
93 if (LatticeValue != constant) {
94 if (LatticeValue == undefined) {
95 LatticeValue = constant;
96 assert(V && "Marking constant with NULL");
99 assert(LatticeValue == forcedconstant &&
100 "Cannot move from overdefined to constant!");
101 // Stay at forcedconstant if the constant is the same.
102 if (V == ConstantVal) return false;
104 // Otherwise, we go to overdefined. Assumptions made based on the
105 // forced value are possibly wrong. Assuming this is another constant
106 // could expose a contradiction.
107 LatticeValue = overdefined;
111 assert(ConstantVal == V && "Marking constant with different value");
116 inline void markForcedConstant(Constant *V) {
117 assert(LatticeValue == undefined && "Can't force a defined value!");
118 LatticeValue = forcedconstant;
122 inline bool isUndefined() const { return LatticeValue == undefined; }
123 inline bool isConstant() const {
124 return LatticeValue == constant || LatticeValue == forcedconstant;
126 inline bool isOverdefined() const { return LatticeValue == overdefined; }
128 inline Constant *getConstant() const {
129 assert(isConstant() && "Cannot get the constant of a non-constant!");
134 //===----------------------------------------------------------------------===//
136 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional
137 /// Constant Propagation.
139 class SCCPSolver : public InstVisitor<SCCPSolver> {
140 SmallSet<BasicBlock*, 16> BBExecutable;// The basic blocks that are executable
141 DenseMap<Value*, LatticeVal> ValueState; // The state each value is in.
143 /// GlobalValue - If we are tracking any values for the contents of a global
144 /// variable, we keep a mapping from the constant accessor to the element of
145 /// the global, to the currently known value. If the value becomes
146 /// overdefined, it's entry is simply removed from this map.
147 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
149 /// TrackedRetVals - If we are tracking arguments into and the return
150 /// value out of a function, it will have an entry in this map, indicating
151 /// what the known return value for the function is.
152 DenseMap<Function*, LatticeVal> TrackedRetVals;
154 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
155 /// that return multiple values.
156 std::map<std::pair<Function*, unsigned>, LatticeVal> TrackedMultipleRetVals;
158 // The reason for two worklists is that overdefined is the lowest state
159 // on the lattice, and moving things to overdefined as fast as possible
160 // makes SCCP converge much faster.
161 // By having a separate worklist, we accomplish this because everything
162 // possibly overdefined will become overdefined at the soonest possible
164 std::vector<Value*> OverdefinedInstWorkList;
165 std::vector<Value*> InstWorkList;
168 std::vector<BasicBlock*> BBWorkList; // The BasicBlock work list
170 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
171 /// overdefined, despite the fact that the PHI node is overdefined.
172 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
174 /// KnownFeasibleEdges - Entries in this set are edges which have already had
175 /// PHI nodes retriggered.
176 typedef std::pair<BasicBlock*,BasicBlock*> Edge;
177 std::set<Edge> KnownFeasibleEdges;
180 /// MarkBlockExecutable - This method can be used by clients to mark all of
181 /// the blocks that are known to be intrinsically live in the processed unit.
182 void MarkBlockExecutable(BasicBlock *BB) {
183 DOUT << "Marking Block Executable: " << BB->getNameStart() << "\n";
184 BBExecutable.insert(BB); // Basic block is executable!
185 BBWorkList.push_back(BB); // Add the block to the work list!
188 /// TrackValueOfGlobalVariable - Clients can use this method to
189 /// inform the SCCPSolver that it should track loads and stores to the
190 /// specified global variable if it can. This is only legal to call if
191 /// performing Interprocedural SCCP.
192 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
193 const Type *ElTy = GV->getType()->getElementType();
194 if (ElTy->isFirstClassType()) {
195 LatticeVal &IV = TrackedGlobals[GV];
196 if (!isa<UndefValue>(GV->getInitializer()))
197 IV.markConstant(GV->getInitializer());
201 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
202 /// and out of the specified function (which cannot have its address taken),
203 /// this method must be called.
204 void AddTrackedFunction(Function *F) {
205 assert(F->hasInternalLinkage() && "Can only track internal functions!");
206 // Add an entry, F -> undef.
207 if (const StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
208 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
209 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
212 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
215 /// Solve - Solve for constants and executable blocks.
219 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
220 /// that branches on undef values cannot reach any of their successors.
221 /// However, this is not a safe assumption. After we solve dataflow, this
222 /// method should be use to handle this. If this returns true, the solver
224 bool ResolvedUndefsIn(Function &F);
226 /// getExecutableBlocks - Once we have solved for constants, return the set of
227 /// blocks that is known to be executable.
228 SmallSet<BasicBlock*, 16> &getExecutableBlocks() {
232 /// getValueMapping - Once we have solved for constants, return the mapping of
233 /// LLVM values to LatticeVals.
234 DenseMap<Value*, LatticeVal> &getValueMapping() {
238 /// getTrackedRetVals - Get the inferred return value map.
240 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
241 return TrackedRetVals;
244 /// getTrackedGlobals - Get and return the set of inferred initializers for
245 /// global variables.
246 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
247 return TrackedGlobals;
250 inline void markOverdefined(Value *V) {
251 markOverdefined(ValueState[V], V);
255 // markConstant - Make a value be marked as "constant". If the value
256 // is not already a constant, add it to the instruction work list so that
257 // the users of the instruction are updated later.
259 inline void markConstant(LatticeVal &IV, Value *V, Constant *C) {
260 if (IV.markConstant(C)) {
261 DOUT << "markConstant: " << *C << ": " << *V;
262 InstWorkList.push_back(V);
266 inline void markForcedConstant(LatticeVal &IV, Value *V, Constant *C) {
267 IV.markForcedConstant(C);
268 DOUT << "markForcedConstant: " << *C << ": " << *V;
269 InstWorkList.push_back(V);
272 inline void markConstant(Value *V, Constant *C) {
273 markConstant(ValueState[V], V, C);
276 // markOverdefined - Make a value be marked as "overdefined". If the
277 // value is not already overdefined, add it to the overdefined instruction
278 // work list so that the users of the instruction are updated later.
279 inline void markOverdefined(LatticeVal &IV, Value *V) {
280 if (IV.markOverdefined()) {
281 DEBUG(DOUT << "markOverdefined: ";
282 if (Function *F = dyn_cast<Function>(V))
283 DOUT << "Function '" << F->getName() << "'\n";
286 // Only instructions go on the work list
287 OverdefinedInstWorkList.push_back(V);
291 inline void mergeInValue(LatticeVal &IV, Value *V, LatticeVal &MergeWithV) {
292 if (IV.isOverdefined() || MergeWithV.isUndefined())
294 if (MergeWithV.isOverdefined())
295 markOverdefined(IV, V);
296 else if (IV.isUndefined())
297 markConstant(IV, V, MergeWithV.getConstant());
298 else if (IV.getConstant() != MergeWithV.getConstant())
299 markOverdefined(IV, V);
302 inline void mergeInValue(Value *V, LatticeVal &MergeWithV) {
303 return mergeInValue(ValueState[V], V, MergeWithV);
307 // getValueState - Return the LatticeVal object that corresponds to the value.
308 // This function is necessary because not all values should start out in the
309 // underdefined state... Argument's should be overdefined, and
310 // constants should be marked as constants. If a value is not known to be an
311 // Instruction object, then use this accessor to get its value from the map.
313 inline LatticeVal &getValueState(Value *V) {
314 DenseMap<Value*, LatticeVal>::iterator I = ValueState.find(V);
315 if (I != ValueState.end()) return I->second; // Common case, in the map
317 if (Constant *C = dyn_cast<Constant>(V)) {
318 if (isa<UndefValue>(V)) {
319 // Nothing to do, remain undefined.
321 LatticeVal &LV = ValueState[C];
322 LV.markConstant(C); // Constants are constant
326 // All others are underdefined by default...
327 return ValueState[V];
330 // markEdgeExecutable - Mark a basic block as executable, adding it to the BB
331 // work list if it is not already executable...
333 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
334 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
335 return; // This edge is already known to be executable!
337 if (BBExecutable.count(Dest)) {
338 DOUT << "Marking Edge Executable: " << Source->getNameStart()
339 << " -> " << Dest->getNameStart() << "\n";
341 // The destination is already executable, but we just made an edge
342 // feasible that wasn't before. Revisit the PHI nodes in the block
343 // because they have potentially new operands.
344 for (BasicBlock::iterator I = Dest->begin(); isa<PHINode>(I); ++I)
345 visitPHINode(*cast<PHINode>(I));
348 MarkBlockExecutable(Dest);
352 // getFeasibleSuccessors - Return a vector of booleans to indicate which
353 // successors are reachable from a given terminator instruction.
355 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
357 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
358 // block to the 'To' basic block is currently feasible...
360 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
362 // OperandChangedState - This method is invoked on all of the users of an
363 // instruction that was just changed state somehow.... Based on this
364 // information, we need to update the specified user of this instruction.
366 void OperandChangedState(User *U) {
367 // Only instructions use other variable values!
368 Instruction &I = cast<Instruction>(*U);
369 if (BBExecutable.count(I.getParent())) // Inst is executable?
374 friend class InstVisitor<SCCPSolver>;
376 // visit implementations - Something changed in this instruction... Either an
377 // operand made a transition, or the instruction is newly executable. Change
378 // the value type of I to reflect these changes if appropriate.
380 void visitPHINode(PHINode &I);
383 void visitReturnInst(ReturnInst &I);
384 void visitTerminatorInst(TerminatorInst &TI);
386 void visitCastInst(CastInst &I);
387 void visitSelectInst(SelectInst &I);
388 void visitBinaryOperator(Instruction &I);
389 void visitCmpInst(CmpInst &I);
390 void visitExtractElementInst(ExtractElementInst &I);
391 void visitInsertElementInst(InsertElementInst &I);
392 void visitShuffleVectorInst(ShuffleVectorInst &I);
393 void visitExtractValueInst(ExtractValueInst &EVI);
394 void visitInsertValueInst(InsertValueInst &IVI);
396 // Instructions that cannot be folded away...
397 void visitStoreInst (Instruction &I);
398 void visitLoadInst (LoadInst &I);
399 void visitGetElementPtrInst(GetElementPtrInst &I);
400 void visitCallInst (CallInst &I) { visitCallSite(CallSite::get(&I)); }
401 void visitInvokeInst (InvokeInst &II) {
402 visitCallSite(CallSite::get(&II));
403 visitTerminatorInst(II);
405 void visitCallSite (CallSite CS);
406 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ }
407 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
408 void visitAllocationInst(Instruction &I) { markOverdefined(&I); }
409 void visitVANextInst (Instruction &I) { markOverdefined(&I); }
410 void visitVAArgInst (Instruction &I) { markOverdefined(&I); }
411 void visitFreeInst (Instruction &I) { /*returns void*/ }
413 void visitInstruction(Instruction &I) {
414 // If a new instruction is added to LLVM that we don't handle...
415 cerr << "SCCP: Don't know how to handle: " << I;
416 markOverdefined(&I); // Just in case
420 } // end anonymous namespace
423 // getFeasibleSuccessors - Return a vector of booleans to indicate which
424 // successors are reachable from a given terminator instruction.
426 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
427 SmallVector<bool, 16> &Succs) {
428 Succs.resize(TI.getNumSuccessors());
429 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
430 if (BI->isUnconditional()) {
433 LatticeVal &BCValue = getValueState(BI->getCondition());
434 if (BCValue.isOverdefined() ||
435 (BCValue.isConstant() && !isa<ConstantInt>(BCValue.getConstant()))) {
436 // Overdefined condition variables, and branches on unfoldable constant
437 // conditions, mean the branch could go either way.
438 Succs[0] = Succs[1] = true;
439 } else if (BCValue.isConstant()) {
440 // Constant condition variables mean the branch can only go a single way
441 Succs[BCValue.getConstant() == ConstantInt::getFalse()] = true;
444 } else if (isa<InvokeInst>(&TI)) {
445 // Invoke instructions successors are always executable.
446 Succs[0] = Succs[1] = true;
447 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
448 LatticeVal &SCValue = getValueState(SI->getCondition());
449 if (SCValue.isOverdefined() || // Overdefined condition?
450 (SCValue.isConstant() && !isa<ConstantInt>(SCValue.getConstant()))) {
451 // All destinations are executable!
452 Succs.assign(TI.getNumSuccessors(), true);
453 } else if (SCValue.isConstant())
454 Succs[SI->findCaseValue(cast<ConstantInt>(SCValue.getConstant()))] = true;
456 assert(0 && "SCCP: Don't know how to handle this terminator!");
461 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
462 // block to the 'To' basic block is currently feasible...
464 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
465 assert(BBExecutable.count(To) && "Dest should always be alive!");
467 // Make sure the source basic block is executable!!
468 if (!BBExecutable.count(From)) return false;
470 // Check to make sure this edge itself is actually feasible now...
471 TerminatorInst *TI = From->getTerminator();
472 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
473 if (BI->isUnconditional())
476 LatticeVal &BCValue = getValueState(BI->getCondition());
477 if (BCValue.isOverdefined()) {
478 // Overdefined condition variables mean the branch could go either way.
480 } else if (BCValue.isConstant()) {
481 // Not branching on an evaluatable constant?
482 if (!isa<ConstantInt>(BCValue.getConstant())) return true;
484 // Constant condition variables mean the branch can only go a single way
485 return BI->getSuccessor(BCValue.getConstant() ==
486 ConstantInt::getFalse()) == To;
490 } else if (isa<InvokeInst>(TI)) {
491 // Invoke instructions successors are always executable.
493 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
494 LatticeVal &SCValue = getValueState(SI->getCondition());
495 if (SCValue.isOverdefined()) { // Overdefined condition?
496 // All destinations are executable!
498 } else if (SCValue.isConstant()) {
499 Constant *CPV = SCValue.getConstant();
500 if (!isa<ConstantInt>(CPV))
501 return true; // not a foldable constant?
503 // Make sure to skip the "default value" which isn't a value
504 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i)
505 if (SI->getSuccessorValue(i) == CPV) // Found the taken branch...
506 return SI->getSuccessor(i) == To;
508 // Constant value not equal to any of the branches... must execute
509 // default branch then...
510 return SI->getDefaultDest() == To;
514 cerr << "Unknown terminator instruction: " << *TI;
519 // visit Implementations - Something changed in this instruction... Either an
520 // operand made a transition, or the instruction is newly executable. Change
521 // the value type of I to reflect these changes if appropriate. This method
522 // makes sure to do the following actions:
524 // 1. If a phi node merges two constants in, and has conflicting value coming
525 // from different branches, or if the PHI node merges in an overdefined
526 // value, then the PHI node becomes overdefined.
527 // 2. If a phi node merges only constants in, and they all agree on value, the
528 // PHI node becomes a constant value equal to that.
529 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
530 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
531 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
532 // 6. If a conditional branch has a value that is constant, make the selected
533 // destination executable
534 // 7. If a conditional branch has a value that is overdefined, make all
535 // successors executable.
537 void SCCPSolver::visitPHINode(PHINode &PN) {
538 LatticeVal &PNIV = getValueState(&PN);
539 if (PNIV.isOverdefined()) {
540 // There may be instructions using this PHI node that are not overdefined
541 // themselves. If so, make sure that they know that the PHI node operand
543 std::multimap<PHINode*, Instruction*>::iterator I, E;
544 tie(I, E) = UsersOfOverdefinedPHIs.equal_range(&PN);
546 SmallVector<Instruction*, 16> Users;
547 for (; I != E; ++I) Users.push_back(I->second);
548 while (!Users.empty()) {
553 return; // Quick exit
556 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
557 // and slow us down a lot. Just mark them overdefined.
558 if (PN.getNumIncomingValues() > 64) {
559 markOverdefined(PNIV, &PN);
563 // Look at all of the executable operands of the PHI node. If any of them
564 // are overdefined, the PHI becomes overdefined as well. If they are all
565 // constant, and they agree with each other, the PHI becomes the identical
566 // constant. If they are constant and don't agree, the PHI is overdefined.
567 // If there are no executable operands, the PHI remains undefined.
569 Constant *OperandVal = 0;
570 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
571 LatticeVal &IV = getValueState(PN.getIncomingValue(i));
572 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
574 if (isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent())) {
575 if (IV.isOverdefined()) { // PHI node becomes overdefined!
576 markOverdefined(PNIV, &PN);
580 if (OperandVal == 0) { // Grab the first value...
581 OperandVal = IV.getConstant();
582 } else { // Another value is being merged in!
583 // There is already a reachable operand. If we conflict with it,
584 // then the PHI node becomes overdefined. If we agree with it, we
587 // Check to see if there are two different constants merging...
588 if (IV.getConstant() != OperandVal) {
589 // Yes there is. This means the PHI node is not constant.
590 // You must be overdefined poor PHI.
592 markOverdefined(PNIV, &PN); // The PHI node now becomes overdefined
593 return; // I'm done analyzing you
599 // If we exited the loop, this means that the PHI node only has constant
600 // arguments that agree with each other(and OperandVal is the constant) or
601 // OperandVal is null because there are no defined incoming arguments. If
602 // this is the case, the PHI remains undefined.
605 markConstant(PNIV, &PN, OperandVal); // Acquire operand value
608 void SCCPSolver::visitReturnInst(ReturnInst &I) {
609 if (I.getNumOperands() == 0) return; // Ret void
611 Function *F = I.getParent()->getParent();
612 // If we are tracking the return value of this function, merge it in.
613 if (!F->hasInternalLinkage())
616 if (!TrackedRetVals.empty() && I.getNumOperands() == 1) {
617 DenseMap<Function*, LatticeVal>::iterator TFRVI =
618 TrackedRetVals.find(F);
619 if (TFRVI != TrackedRetVals.end() &&
620 !TFRVI->second.isOverdefined()) {
621 LatticeVal &IV = getValueState(I.getOperand(0));
622 mergeInValue(TFRVI->second, F, IV);
627 // Handle functions that return multiple values.
628 if (!TrackedMultipleRetVals.empty() && I.getNumOperands() > 1) {
629 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
630 std::map<std::pair<Function*, unsigned>, LatticeVal>::iterator
631 It = TrackedMultipleRetVals.find(std::make_pair(F, i));
632 if (It == TrackedMultipleRetVals.end()) break;
633 mergeInValue(It->second, F, getValueState(I.getOperand(i)));
635 } else if (!TrackedMultipleRetVals.empty() &&
636 I.getNumOperands() == 1 &&
637 isa<StructType>(I.getOperand(0)->getType())) {
638 for (unsigned i = 0, e = I.getOperand(0)->getType()->getNumContainedTypes();
640 std::map<std::pair<Function*, unsigned>, LatticeVal>::iterator
641 It = TrackedMultipleRetVals.find(std::make_pair(F, i));
642 if (It == TrackedMultipleRetVals.end()) break;
643 Value *Val = FindInsertedValue(I.getOperand(0), i);
644 mergeInValue(It->second, F, getValueState(Val));
649 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
650 SmallVector<bool, 16> SuccFeasible;
651 getFeasibleSuccessors(TI, SuccFeasible);
653 BasicBlock *BB = TI.getParent();
655 // Mark all feasible successors executable...
656 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
658 markEdgeExecutable(BB, TI.getSuccessor(i));
661 void SCCPSolver::visitCastInst(CastInst &I) {
662 Value *V = I.getOperand(0);
663 LatticeVal &VState = getValueState(V);
664 if (VState.isOverdefined()) // Inherit overdefinedness of operand
666 else if (VState.isConstant()) // Propagate constant value
667 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
668 VState.getConstant(), I.getType()));
671 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
672 Value *Aggr = EVI.getAggregateOperand();
674 // If the operand to the extractvalue is an undef, the result is undef.
675 if (isa<UndefValue>(Aggr))
678 // Currently only handle single-index extractvalues.
679 if (EVI.getNumIndices() != 1) {
680 markOverdefined(&EVI);
685 if (CallInst *CI = dyn_cast<CallInst>(Aggr))
686 F = CI->getCalledFunction();
687 else if (InvokeInst *II = dyn_cast<InvokeInst>(Aggr))
688 F = II->getCalledFunction();
690 // TODO: If IPSCCP resolves the callee of this function, we could propagate a
692 if (F == 0 || TrackedMultipleRetVals.empty()) {
693 markOverdefined(&EVI);
697 // See if we are tracking the result of the callee.
698 std::map<std::pair<Function*, unsigned>, LatticeVal>::iterator
699 It = TrackedMultipleRetVals.find(std::make_pair(F, *EVI.idx_begin()));
701 // If not tracking this function (for example, it is a declaration) just move
703 if (It == TrackedMultipleRetVals.end()) {
704 markOverdefined(&EVI);
708 // Otherwise, the value will be merged in here as a result of CallSite
712 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
713 Value *Aggr = IVI.getAggregateOperand();
714 Value *Val = IVI.getInsertedValueOperand();
716 // If the operands to the insertvalue are undef, the result is undef.
717 if (isa<UndefValue>(Aggr) && isa<UndefValue>(Val))
720 // Currently only handle single-index insertvalues.
721 if (IVI.getNumIndices() != 1) {
722 markOverdefined(&IVI);
726 // Currently only handle insertvalue instructions that are in a single-use
727 // chain that builds up a return value.
728 for (const InsertValueInst *TmpIVI = &IVI; ; ) {
729 if (!TmpIVI->hasOneUse()) {
730 markOverdefined(&IVI);
733 const Value *V = *TmpIVI->use_begin();
734 if (isa<ReturnInst>(V))
736 TmpIVI = dyn_cast<InsertValueInst>(V);
738 markOverdefined(&IVI);
743 // See if we are tracking the result of the callee.
744 Function *F = IVI.getParent()->getParent();
745 std::map<std::pair<Function*, unsigned>, LatticeVal>::iterator
746 It = TrackedMultipleRetVals.find(std::make_pair(F, *IVI.idx_begin()));
748 // Merge in the inserted member value.
749 if (It != TrackedMultipleRetVals.end())
750 mergeInValue(It->second, F, getValueState(Val));
752 // Mark the aggregate result of the IVI overdefined; any tracking that we do
753 // will be done on the individual member values.
754 markOverdefined(&IVI);
757 void SCCPSolver::visitSelectInst(SelectInst &I) {
758 LatticeVal &CondValue = getValueState(I.getCondition());
759 if (CondValue.isUndefined())
761 if (CondValue.isConstant()) {
762 if (ConstantInt *CondCB = dyn_cast<ConstantInt>(CondValue.getConstant())){
763 mergeInValue(&I, getValueState(CondCB->getZExtValue() ? I.getTrueValue()
764 : I.getFalseValue()));
769 // Otherwise, the condition is overdefined or a constant we can't evaluate.
770 // See if we can produce something better than overdefined based on the T/F
772 LatticeVal &TVal = getValueState(I.getTrueValue());
773 LatticeVal &FVal = getValueState(I.getFalseValue());
775 // select ?, C, C -> C.
776 if (TVal.isConstant() && FVal.isConstant() &&
777 TVal.getConstant() == FVal.getConstant()) {
778 markConstant(&I, FVal.getConstant());
782 if (TVal.isUndefined()) { // select ?, undef, X -> X.
783 mergeInValue(&I, FVal);
784 } else if (FVal.isUndefined()) { // select ?, X, undef -> X.
785 mergeInValue(&I, TVal);
791 // Handle BinaryOperators and Shift Instructions...
792 void SCCPSolver::visitBinaryOperator(Instruction &I) {
793 LatticeVal &IV = ValueState[&I];
794 if (IV.isOverdefined()) return;
796 LatticeVal &V1State = getValueState(I.getOperand(0));
797 LatticeVal &V2State = getValueState(I.getOperand(1));
799 if (V1State.isOverdefined() || V2State.isOverdefined()) {
800 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
801 // operand is overdefined.
802 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
803 LatticeVal *NonOverdefVal = 0;
804 if (!V1State.isOverdefined()) {
805 NonOverdefVal = &V1State;
806 } else if (!V2State.isOverdefined()) {
807 NonOverdefVal = &V2State;
811 if (NonOverdefVal->isUndefined()) {
812 // Could annihilate value.
813 if (I.getOpcode() == Instruction::And)
814 markConstant(IV, &I, Constant::getNullValue(I.getType()));
815 else if (const VectorType *PT = dyn_cast<VectorType>(I.getType()))
816 markConstant(IV, &I, ConstantVector::getAllOnesValue(PT));
818 markConstant(IV, &I, ConstantInt::getAllOnesValue(I.getType()));
821 if (I.getOpcode() == Instruction::And) {
822 if (NonOverdefVal->getConstant()->isNullValue()) {
823 markConstant(IV, &I, NonOverdefVal->getConstant());
824 return; // X and 0 = 0
827 if (ConstantInt *CI =
828 dyn_cast<ConstantInt>(NonOverdefVal->getConstant()))
829 if (CI->isAllOnesValue()) {
830 markConstant(IV, &I, NonOverdefVal->getConstant());
831 return; // X or -1 = -1
839 // If both operands are PHI nodes, it is possible that this instruction has
840 // a constant value, despite the fact that the PHI node doesn't. Check for
841 // this condition now.
842 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
843 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
844 if (PN1->getParent() == PN2->getParent()) {
845 // Since the two PHI nodes are in the same basic block, they must have
846 // entries for the same predecessors. Walk the predecessor list, and
847 // if all of the incoming values are constants, and the result of
848 // evaluating this expression with all incoming value pairs is the
849 // same, then this expression is a constant even though the PHI node
850 // is not a constant!
852 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
853 LatticeVal &In1 = getValueState(PN1->getIncomingValue(i));
854 BasicBlock *InBlock = PN1->getIncomingBlock(i);
856 getValueState(PN2->getIncomingValueForBlock(InBlock));
858 if (In1.isOverdefined() || In2.isOverdefined()) {
859 Result.markOverdefined();
860 break; // Cannot fold this operation over the PHI nodes!
861 } else if (In1.isConstant() && In2.isConstant()) {
862 Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(),
864 if (Result.isUndefined())
865 Result.markConstant(V);
866 else if (Result.isConstant() && Result.getConstant() != V) {
867 Result.markOverdefined();
873 // If we found a constant value here, then we know the instruction is
874 // constant despite the fact that the PHI nodes are overdefined.
875 if (Result.isConstant()) {
876 markConstant(IV, &I, Result.getConstant());
877 // Remember that this instruction is virtually using the PHI node
879 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
880 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
882 } else if (Result.isUndefined()) {
886 // Okay, this really is overdefined now. Since we might have
887 // speculatively thought that this was not overdefined before, and
888 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
889 // make sure to clean out any entries that we put there, for
891 std::multimap<PHINode*, Instruction*>::iterator It, E;
892 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1);
894 if (It->second == &I) {
895 UsersOfOverdefinedPHIs.erase(It++);
899 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2);
901 if (It->second == &I) {
902 UsersOfOverdefinedPHIs.erase(It++);
908 markOverdefined(IV, &I);
909 } else if (V1State.isConstant() && V2State.isConstant()) {
910 markConstant(IV, &I, ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
911 V2State.getConstant()));
915 // Handle ICmpInst instruction...
916 void SCCPSolver::visitCmpInst(CmpInst &I) {
917 LatticeVal &IV = ValueState[&I];
918 if (IV.isOverdefined()) return;
920 LatticeVal &V1State = getValueState(I.getOperand(0));
921 LatticeVal &V2State = getValueState(I.getOperand(1));
923 if (V1State.isOverdefined() || V2State.isOverdefined()) {
924 // If both operands are PHI nodes, it is possible that this instruction has
925 // a constant value, despite the fact that the PHI node doesn't. Check for
926 // this condition now.
927 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
928 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
929 if (PN1->getParent() == PN2->getParent()) {
930 // Since the two PHI nodes are in the same basic block, they must have
931 // entries for the same predecessors. Walk the predecessor list, and
932 // if all of the incoming values are constants, and the result of
933 // evaluating this expression with all incoming value pairs is the
934 // same, then this expression is a constant even though the PHI node
935 // is not a constant!
937 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
938 LatticeVal &In1 = getValueState(PN1->getIncomingValue(i));
939 BasicBlock *InBlock = PN1->getIncomingBlock(i);
941 getValueState(PN2->getIncomingValueForBlock(InBlock));
943 if (In1.isOverdefined() || In2.isOverdefined()) {
944 Result.markOverdefined();
945 break; // Cannot fold this operation over the PHI nodes!
946 } else if (In1.isConstant() && In2.isConstant()) {
947 Constant *V = ConstantExpr::getCompare(I.getPredicate(),
950 if (Result.isUndefined())
951 Result.markConstant(V);
952 else if (Result.isConstant() && Result.getConstant() != V) {
953 Result.markOverdefined();
959 // If we found a constant value here, then we know the instruction is
960 // constant despite the fact that the PHI nodes are overdefined.
961 if (Result.isConstant()) {
962 markConstant(IV, &I, Result.getConstant());
963 // Remember that this instruction is virtually using the PHI node
965 UsersOfOverdefinedPHIs.insert(std::make_pair(PN1, &I));
966 UsersOfOverdefinedPHIs.insert(std::make_pair(PN2, &I));
968 } else if (Result.isUndefined()) {
972 // Okay, this really is overdefined now. Since we might have
973 // speculatively thought that this was not overdefined before, and
974 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
975 // make sure to clean out any entries that we put there, for
977 std::multimap<PHINode*, Instruction*>::iterator It, E;
978 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN1);
980 if (It->second == &I) {
981 UsersOfOverdefinedPHIs.erase(It++);
985 tie(It, E) = UsersOfOverdefinedPHIs.equal_range(PN2);
987 if (It->second == &I) {
988 UsersOfOverdefinedPHIs.erase(It++);
994 markOverdefined(IV, &I);
995 } else if (V1State.isConstant() && V2State.isConstant()) {
996 markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
997 V1State.getConstant(),
998 V2State.getConstant()));
1002 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
1003 // FIXME : SCCP does not handle vectors properly.
1004 markOverdefined(&I);
1008 LatticeVal &ValState = getValueState(I.getOperand(0));
1009 LatticeVal &IdxState = getValueState(I.getOperand(1));
1011 if (ValState.isOverdefined() || IdxState.isOverdefined())
1012 markOverdefined(&I);
1013 else if(ValState.isConstant() && IdxState.isConstant())
1014 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
1015 IdxState.getConstant()));
1019 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
1020 // FIXME : SCCP does not handle vectors properly.
1021 markOverdefined(&I);
1024 LatticeVal &ValState = getValueState(I.getOperand(0));
1025 LatticeVal &EltState = getValueState(I.getOperand(1));
1026 LatticeVal &IdxState = getValueState(I.getOperand(2));
1028 if (ValState.isOverdefined() || EltState.isOverdefined() ||
1029 IdxState.isOverdefined())
1030 markOverdefined(&I);
1031 else if(ValState.isConstant() && EltState.isConstant() &&
1032 IdxState.isConstant())
1033 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
1034 EltState.getConstant(),
1035 IdxState.getConstant()));
1036 else if (ValState.isUndefined() && EltState.isConstant() &&
1037 IdxState.isConstant())
1038 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
1039 EltState.getConstant(),
1040 IdxState.getConstant()));
1044 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
1045 // FIXME : SCCP does not handle vectors properly.
1046 markOverdefined(&I);
1049 LatticeVal &V1State = getValueState(I.getOperand(0));
1050 LatticeVal &V2State = getValueState(I.getOperand(1));
1051 LatticeVal &MaskState = getValueState(I.getOperand(2));
1053 if (MaskState.isUndefined() ||
1054 (V1State.isUndefined() && V2State.isUndefined()))
1055 return; // Undefined output if mask or both inputs undefined.
1057 if (V1State.isOverdefined() || V2State.isOverdefined() ||
1058 MaskState.isOverdefined()) {
1059 markOverdefined(&I);
1061 // A mix of constant/undef inputs.
1062 Constant *V1 = V1State.isConstant() ?
1063 V1State.getConstant() : UndefValue::get(I.getType());
1064 Constant *V2 = V2State.isConstant() ?
1065 V2State.getConstant() : UndefValue::get(I.getType());
1066 Constant *Mask = MaskState.isConstant() ?
1067 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
1068 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
1073 // Handle getelementptr instructions... if all operands are constants then we
1074 // can turn this into a getelementptr ConstantExpr.
1076 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1077 LatticeVal &IV = ValueState[&I];
1078 if (IV.isOverdefined()) return;
1080 SmallVector<Constant*, 8> Operands;
1081 Operands.reserve(I.getNumOperands());
1083 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1084 LatticeVal &State = getValueState(I.getOperand(i));
1085 if (State.isUndefined())
1086 return; // Operands are not resolved yet...
1087 else if (State.isOverdefined()) {
1088 markOverdefined(IV, &I);
1091 assert(State.isConstant() && "Unknown state!");
1092 Operands.push_back(State.getConstant());
1095 Constant *Ptr = Operands[0];
1096 Operands.erase(Operands.begin()); // Erase the pointer from idx list...
1098 markConstant(IV, &I, ConstantExpr::getGetElementPtr(Ptr, &Operands[0],
1102 void SCCPSolver::visitStoreInst(Instruction &SI) {
1103 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1105 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1106 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1107 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1109 // Get the value we are storing into the global.
1110 LatticeVal &PtrVal = getValueState(SI.getOperand(0));
1112 mergeInValue(I->second, GV, PtrVal);
1113 if (I->second.isOverdefined())
1114 TrackedGlobals.erase(I); // No need to keep tracking this!
1118 // Handle load instructions. If the operand is a constant pointer to a constant
1119 // global, we can replace the load with the loaded constant value!
1120 void SCCPSolver::visitLoadInst(LoadInst &I) {
1121 LatticeVal &IV = ValueState[&I];
1122 if (IV.isOverdefined()) return;
1124 LatticeVal &PtrVal = getValueState(I.getOperand(0));
1125 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1126 if (PtrVal.isConstant() && !I.isVolatile()) {
1127 Value *Ptr = PtrVal.getConstant();
1128 // TODO: Consider a target hook for valid address spaces for this xform.
1129 if (isa<ConstantPointerNull>(Ptr) &&
1130 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
1131 // load null -> null
1132 markConstant(IV, &I, Constant::getNullValue(I.getType()));
1136 // Transform load (constant global) into the value loaded.
1137 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1138 if (GV->isConstant()) {
1139 if (!GV->isDeclaration()) {
1140 markConstant(IV, &I, GV->getInitializer());
1143 } else if (!TrackedGlobals.empty()) {
1144 // If we are tracking this global, merge in the known value for it.
1145 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1146 TrackedGlobals.find(GV);
1147 if (It != TrackedGlobals.end()) {
1148 mergeInValue(IV, &I, It->second);
1154 // Transform load (constantexpr_GEP global, 0, ...) into the value loaded.
1155 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
1156 if (CE->getOpcode() == Instruction::GetElementPtr)
1157 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
1158 if (GV->isConstant() && !GV->isDeclaration())
1160 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE)) {
1161 markConstant(IV, &I, V);
1166 // Otherwise we cannot say for certain what value this load will produce.
1168 markOverdefined(IV, &I);
1171 void SCCPSolver::visitCallSite(CallSite CS) {
1172 Function *F = CS.getCalledFunction();
1173 Instruction *I = CS.getInstruction();
1175 // The common case is that we aren't tracking the callee, either because we
1176 // are not doing interprocedural analysis or the callee is indirect, or is
1177 // external. Handle these cases first.
1178 if (F == 0 || !F->hasInternalLinkage()) {
1180 // Void return and not tracking callee, just bail.
1181 if (I->getType() == Type::VoidTy) return;
1183 // Otherwise, if we have a single return value case, and if the function is
1184 // a declaration, maybe we can constant fold it.
1185 if (!isa<StructType>(I->getType()) && F && F->isDeclaration() &&
1186 canConstantFoldCallTo(F)) {
1188 SmallVector<Constant*, 8> Operands;
1189 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1191 LatticeVal &State = getValueState(*AI);
1192 if (State.isUndefined())
1193 return; // Operands are not resolved yet.
1194 else if (State.isOverdefined()) {
1198 assert(State.isConstant() && "Unknown state!");
1199 Operands.push_back(State.getConstant());
1202 // If we can constant fold this, mark the result of the call as a
1204 if (Constant *C = ConstantFoldCall(F, &Operands[0], Operands.size())) {
1210 // Otherwise, we don't know anything about this call, mark it overdefined.
1215 // If this is a single/zero retval case, see if we're tracking the function.
1216 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1217 if (TFRVI != TrackedRetVals.end()) {
1218 // If so, propagate the return value of the callee into this call result.
1219 mergeInValue(I, TFRVI->second);
1220 } else if (isa<StructType>(I->getType())) {
1221 // Check to see if we're tracking this callee, if not, handle it in the
1222 // common path above.
1223 std::map<std::pair<Function*, unsigned>, LatticeVal>::iterator
1224 TMRVI = TrackedMultipleRetVals.find(std::make_pair(F, 0));
1225 if (TMRVI == TrackedMultipleRetVals.end())
1226 goto CallOverdefined;
1228 // If we are tracking this callee, propagate the return values of the call
1229 // into this call site. We do this by walking all the uses. Single-index
1230 // ExtractValueInst uses can be tracked; anything more complicated is
1231 // currently handled conservatively.
1232 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1234 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(*UI)) {
1235 if (EVI->getNumIndices() == 1) {
1237 TrackedMultipleRetVals[std::make_pair(F, *EVI->idx_begin())]);
1241 // The aggregate value is used in a way not handled here. Assume nothing.
1242 markOverdefined(*UI);
1245 // Otherwise we're not tracking this callee, so handle it in the
1246 // common path above.
1247 goto CallOverdefined;
1250 // Finally, if this is the first call to the function hit, mark its entry
1251 // block executable.
1252 if (!BBExecutable.count(F->begin()))
1253 MarkBlockExecutable(F->begin());
1255 // Propagate information from this call site into the callee.
1256 CallSite::arg_iterator CAI = CS.arg_begin();
1257 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1258 AI != E; ++AI, ++CAI) {
1259 LatticeVal &IV = ValueState[AI];
1260 if (!IV.isOverdefined())
1261 mergeInValue(IV, AI, getValueState(*CAI));
1266 void SCCPSolver::Solve() {
1267 // Process the work lists until they are empty!
1268 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1269 !OverdefinedInstWorkList.empty()) {
1270 // Process the instruction work list...
1271 while (!OverdefinedInstWorkList.empty()) {
1272 Value *I = OverdefinedInstWorkList.back();
1273 OverdefinedInstWorkList.pop_back();
1275 DOUT << "\nPopped off OI-WL: " << *I;
1277 // "I" got into the work list because it either made the transition from
1278 // bottom to constant
1280 // Anything on this worklist that is overdefined need not be visited
1281 // since all of its users will have already been marked as overdefined
1282 // Update all of the users of this instruction's value...
1284 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1286 OperandChangedState(*UI);
1288 // Process the instruction work list...
1289 while (!InstWorkList.empty()) {
1290 Value *I = InstWorkList.back();
1291 InstWorkList.pop_back();
1293 DOUT << "\nPopped off I-WL: " << *I;
1295 // "I" got into the work list because it either made the transition from
1296 // bottom to constant
1298 // Anything on this worklist that is overdefined need not be visited
1299 // since all of its users will have already been marked as overdefined.
1300 // Update all of the users of this instruction's value...
1302 if (!getValueState(I).isOverdefined())
1303 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1305 OperandChangedState(*UI);
1308 // Process the basic block work list...
1309 while (!BBWorkList.empty()) {
1310 BasicBlock *BB = BBWorkList.back();
1311 BBWorkList.pop_back();
1313 DOUT << "\nPopped off BBWL: " << *BB;
1315 // Notify all instructions in this basic block that they are newly
1322 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1323 /// that branches on undef values cannot reach any of their successors.
1324 /// However, this is not a safe assumption. After we solve dataflow, this
1325 /// method should be use to handle this. If this returns true, the solver
1326 /// should be rerun.
1328 /// This method handles this by finding an unresolved branch and marking it one
1329 /// of the edges from the block as being feasible, even though the condition
1330 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1331 /// CFG and only slightly pessimizes the analysis results (by marking one,
1332 /// potentially infeasible, edge feasible). This cannot usefully modify the
1333 /// constraints on the condition of the branch, as that would impact other users
1336 /// This scan also checks for values that use undefs, whose results are actually
1337 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1338 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1339 /// even if X isn't defined.
1340 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1341 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1342 if (!BBExecutable.count(BB))
1345 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1346 // Look for instructions which produce undef values.
1347 if (I->getType() == Type::VoidTy) continue;
1349 LatticeVal &LV = getValueState(I);
1350 if (!LV.isUndefined()) continue;
1352 // Get the lattice values of the first two operands for use below.
1353 LatticeVal &Op0LV = getValueState(I->getOperand(0));
1355 if (I->getNumOperands() == 2) {
1356 // If this is a two-operand instruction, and if both operands are
1357 // undefs, the result stays undef.
1358 Op1LV = getValueState(I->getOperand(1));
1359 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1363 // If this is an instructions whose result is defined even if the input is
1364 // not fully defined, propagate the information.
1365 const Type *ITy = I->getType();
1366 switch (I->getOpcode()) {
1367 default: break; // Leave the instruction as an undef.
1368 case Instruction::ZExt:
1369 // After a zero extend, we know the top part is zero. SExt doesn't have
1370 // to be handled here, because we don't know whether the top part is 1's
1372 assert(Op0LV.isUndefined());
1373 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1375 case Instruction::Mul:
1376 case Instruction::And:
1377 // undef * X -> 0. X could be zero.
1378 // undef & X -> 0. X could be zero.
1379 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1382 case Instruction::Or:
1383 // undef | X -> -1. X could be -1.
1384 if (const VectorType *PTy = dyn_cast<VectorType>(ITy))
1385 markForcedConstant(LV, I, ConstantVector::getAllOnesValue(PTy));
1387 markForcedConstant(LV, I, ConstantInt::getAllOnesValue(ITy));
1390 case Instruction::SDiv:
1391 case Instruction::UDiv:
1392 case Instruction::SRem:
1393 case Instruction::URem:
1394 // X / undef -> undef. No change.
1395 // X % undef -> undef. No change.
1396 if (Op1LV.isUndefined()) break;
1398 // undef / X -> 0. X could be maxint.
1399 // undef % X -> 0. X could be 1.
1400 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1403 case Instruction::AShr:
1404 // undef >>s X -> undef. No change.
1405 if (Op0LV.isUndefined()) break;
1407 // X >>s undef -> X. X could be 0, X could have the high-bit known set.
1408 if (Op0LV.isConstant())
1409 markForcedConstant(LV, I, Op0LV.getConstant());
1411 markOverdefined(LV, I);
1413 case Instruction::LShr:
1414 case Instruction::Shl:
1415 // undef >> X -> undef. No change.
1416 // undef << X -> undef. No change.
1417 if (Op0LV.isUndefined()) break;
1419 // X >> undef -> 0. X could be 0.
1420 // X << undef -> 0. X could be 0.
1421 markForcedConstant(LV, I, Constant::getNullValue(ITy));
1423 case Instruction::Select:
1424 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1425 if (Op0LV.isUndefined()) {
1426 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1427 Op1LV = getValueState(I->getOperand(2));
1428 } else if (Op1LV.isUndefined()) {
1429 // c ? undef : undef -> undef. No change.
1430 Op1LV = getValueState(I->getOperand(2));
1431 if (Op1LV.isUndefined())
1433 // Otherwise, c ? undef : x -> x.
1435 // Leave Op1LV as Operand(1)'s LatticeValue.
1438 if (Op1LV.isConstant())
1439 markForcedConstant(LV, I, Op1LV.getConstant());
1441 markOverdefined(LV, I);
1443 case Instruction::Call:
1444 // If a call has an undef result, it is because it is constant foldable
1445 // but one of the inputs was undef. Just force the result to
1447 markOverdefined(LV, I);
1452 TerminatorInst *TI = BB->getTerminator();
1453 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1454 if (!BI->isConditional()) continue;
1455 if (!getValueState(BI->getCondition()).isUndefined())
1457 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1458 if (SI->getNumSuccessors()<2) // no cases
1460 if (!getValueState(SI->getCondition()).isUndefined())
1466 // If the edge to the second successor isn't thought to be feasible yet,
1467 // mark it so now. We pick the second one so that this goes to some
1468 // enumerated value in a switch instead of going to the default destination.
1469 if (KnownFeasibleEdges.count(Edge(BB, TI->getSuccessor(1))))
1472 // Otherwise, it isn't already thought to be feasible. Mark it as such now
1473 // and return. This will make other blocks reachable, which will allow new
1474 // values to be discovered and existing ones to be moved in the lattice.
1475 markEdgeExecutable(BB, TI->getSuccessor(1));
1477 // This must be a conditional branch of switch on undef. At this point,
1478 // force the old terminator to branch to the first successor. This is
1479 // required because we are now influencing the dataflow of the function with
1480 // the assumption that this edge is taken. If we leave the branch condition
1481 // as undef, then further analysis could think the undef went another way
1482 // leading to an inconsistent set of conclusions.
1483 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1484 BI->setCondition(ConstantInt::getFalse());
1486 SwitchInst *SI = cast<SwitchInst>(TI);
1487 SI->setCondition(SI->getCaseValue(1));
1498 //===--------------------------------------------------------------------===//
1500 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1501 /// Sparse Conditional Constant Propagator.
1503 struct VISIBILITY_HIDDEN SCCP : public FunctionPass {
1504 static char ID; // Pass identification, replacement for typeid
1505 SCCP() : FunctionPass((intptr_t)&ID) {}
1507 // runOnFunction - Run the Sparse Conditional Constant Propagation
1508 // algorithm, and return true if the function was modified.
1510 bool runOnFunction(Function &F);
1512 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1513 AU.setPreservesCFG();
1516 } // end anonymous namespace
1519 static RegisterPass<SCCP>
1520 X("sccp", "Sparse Conditional Constant Propagation");
1522 // createSCCPPass - This is the public interface to this file...
1523 FunctionPass *llvm::createSCCPPass() {
1528 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1529 // and return true if the function was modified.
1531 bool SCCP::runOnFunction(Function &F) {
1532 DOUT << "SCCP on function '" << F.getNameStart() << "'\n";
1535 // Mark the first block of the function as being executable.
1536 Solver.MarkBlockExecutable(F.begin());
1538 // Mark all arguments to the function as being overdefined.
1539 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1540 Solver.markOverdefined(AI);
1542 // Solve for constants.
1543 bool ResolvedUndefs = true;
1544 while (ResolvedUndefs) {
1546 DOUT << "RESOLVING UNDEFs\n";
1547 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1550 bool MadeChanges = false;
1552 // If we decided that there are basic blocks that are dead in this function,
1553 // delete their contents now. Note that we cannot actually delete the blocks,
1554 // as we cannot modify the CFG of the function.
1556 SmallSet<BasicBlock*, 16> &ExecutableBBs = Solver.getExecutableBlocks();
1557 SmallVector<Instruction*, 32> Insts;
1558 DenseMap<Value*, LatticeVal> &Values = Solver.getValueMapping();
1560 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1561 if (!ExecutableBBs.count(BB)) {
1562 DOUT << " BasicBlock Dead:" << *BB;
1565 // Delete the instructions backwards, as it has a reduced likelihood of
1566 // having to update as many def-use and use-def chains.
1567 for (BasicBlock::iterator I = BB->begin(), E = BB->getTerminator();
1570 while (!Insts.empty()) {
1571 Instruction *I = Insts.back();
1573 if (!I->use_empty())
1574 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1575 BB->getInstList().erase(I);
1580 // Iterate over all of the instructions in a function, replacing them with
1581 // constants if we have found them to be of constant values.
1583 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1584 Instruction *Inst = BI++;
1585 if (Inst->getType() == Type::VoidTy ||
1586 isa<TerminatorInst>(Inst))
1589 LatticeVal &IV = Values[Inst];
1590 if (!IV.isConstant() && !IV.isUndefined())
1593 Constant *Const = IV.isConstant()
1594 ? IV.getConstant() : UndefValue::get(Inst->getType());
1595 DOUT << " Constant: " << *Const << " = " << *Inst;
1597 // Replaces all of the uses of a variable with uses of the constant.
1598 Inst->replaceAllUsesWith(Const);
1600 // Delete the instruction.
1601 Inst->eraseFromParent();
1603 // Hey, we just changed something!
1613 //===--------------------------------------------------------------------===//
1615 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1616 /// Constant Propagation.
1618 struct VISIBILITY_HIDDEN IPSCCP : public ModulePass {
1620 IPSCCP() : ModulePass((intptr_t)&ID) {}
1621 bool runOnModule(Module &M);
1623 } // end anonymous namespace
1625 char IPSCCP::ID = 0;
1626 static RegisterPass<IPSCCP>
1627 Y("ipsccp", "Interprocedural Sparse Conditional Constant Propagation");
1629 // createIPSCCPPass - This is the public interface to this file...
1630 ModulePass *llvm::createIPSCCPPass() {
1631 return new IPSCCP();
1635 static bool AddressIsTaken(GlobalValue *GV) {
1636 // Delete any dead constantexpr klingons.
1637 GV->removeDeadConstantUsers();
1639 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
1641 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
1642 if (SI->getOperand(0) == GV || SI->isVolatile())
1643 return true; // Storing addr of GV.
1644 } else if (isa<InvokeInst>(*UI) || isa<CallInst>(*UI)) {
1645 // Make sure we are calling the function, not passing the address.
1646 CallSite CS = CallSite::get(cast<Instruction>(*UI));
1647 for (CallSite::arg_iterator AI = CS.arg_begin(),
1648 E = CS.arg_end(); AI != E; ++AI)
1651 } else if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1652 if (LI->isVolatile())
1660 bool IPSCCP::runOnModule(Module &M) {
1663 // Loop over all functions, marking arguments to those with their addresses
1664 // taken or that are external as overdefined.
1666 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1667 if (!F->hasInternalLinkage() || AddressIsTaken(F)) {
1668 if (!F->isDeclaration())
1669 Solver.MarkBlockExecutable(F->begin());
1670 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1672 Solver.markOverdefined(AI);
1674 Solver.AddTrackedFunction(F);
1677 // Loop over global variables. We inform the solver about any internal global
1678 // variables that do not have their 'addresses taken'. If they don't have
1679 // their addresses taken, we can propagate constants through them.
1680 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1682 if (!G->isConstant() && G->hasInternalLinkage() && !AddressIsTaken(G))
1683 Solver.TrackValueOfGlobalVariable(G);
1685 // Solve for constants.
1686 bool ResolvedUndefs = true;
1687 while (ResolvedUndefs) {
1690 DOUT << "RESOLVING UNDEFS\n";
1691 ResolvedUndefs = false;
1692 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1693 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1696 bool MadeChanges = false;
1698 // Iterate over all of the instructions in the module, replacing them with
1699 // constants if we have found them to be of constant values.
1701 SmallSet<BasicBlock*, 16> &ExecutableBBs = Solver.getExecutableBlocks();
1702 SmallVector<Instruction*, 32> Insts;
1703 SmallVector<BasicBlock*, 32> BlocksToErase;
1704 DenseMap<Value*, LatticeVal> &Values = Solver.getValueMapping();
1706 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1707 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1709 if (!AI->use_empty()) {
1710 LatticeVal &IV = Values[AI];
1711 if (IV.isConstant() || IV.isUndefined()) {
1712 Constant *CST = IV.isConstant() ?
1713 IV.getConstant() : UndefValue::get(AI->getType());
1714 DOUT << "*** Arg " << *AI << " = " << *CST <<"\n";
1716 // Replaces all of the uses of a variable with uses of the
1718 AI->replaceAllUsesWith(CST);
1723 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1724 if (!ExecutableBBs.count(BB)) {
1725 DOUT << " BasicBlock Dead:" << *BB;
1728 // Delete the instructions backwards, as it has a reduced likelihood of
1729 // having to update as many def-use and use-def chains.
1730 TerminatorInst *TI = BB->getTerminator();
1731 for (BasicBlock::iterator I = BB->begin(), E = TI; I != E; ++I)
1734 while (!Insts.empty()) {
1735 Instruction *I = Insts.back();
1737 if (!I->use_empty())
1738 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1739 BB->getInstList().erase(I);
1744 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1745 BasicBlock *Succ = TI->getSuccessor(i);
1746 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1747 TI->getSuccessor(i)->removePredecessor(BB);
1749 if (!TI->use_empty())
1750 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1751 BB->getInstList().erase(TI);
1753 if (&*BB != &F->front())
1754 BlocksToErase.push_back(BB);
1756 new UnreachableInst(BB);
1759 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1760 Instruction *Inst = BI++;
1761 if (Inst->getType() == Type::VoidTy ||
1762 isa<TerminatorInst>(Inst))
1765 LatticeVal &IV = Values[Inst];
1766 if (!IV.isConstant() && !IV.isUndefined())
1769 Constant *Const = IV.isConstant()
1770 ? IV.getConstant() : UndefValue::get(Inst->getType());
1771 DOUT << " Constant: " << *Const << " = " << *Inst;
1773 // Replaces all of the uses of a variable with uses of the
1775 Inst->replaceAllUsesWith(Const);
1777 // Delete the instruction.
1778 if (!isa<CallInst>(Inst))
1779 Inst->eraseFromParent();
1781 // Hey, we just changed something!
1787 // Now that all instructions in the function are constant folded, erase dead
1788 // blocks, because we can now use ConstantFoldTerminator to get rid of
1790 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
1791 // If there are any PHI nodes in this successor, drop entries for BB now.
1792 BasicBlock *DeadBB = BlocksToErase[i];
1793 while (!DeadBB->use_empty()) {
1794 Instruction *I = cast<Instruction>(DeadBB->use_back());
1795 bool Folded = ConstantFoldTerminator(I->getParent());
1797 // The constant folder may not have been able to fold the terminator
1798 // if this is a branch or switch on undef. Fold it manually as a
1799 // branch to the first successor.
1800 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1801 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
1802 "Branch should be foldable!");
1803 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1804 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
1806 assert(0 && "Didn't fold away reference to block!");
1809 // Make this an uncond branch to the first successor.
1810 TerminatorInst *TI = I->getParent()->getTerminator();
1811 BranchInst::Create(TI->getSuccessor(0), TI);
1813 // Remove entries in successor phi nodes to remove edges.
1814 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
1815 TI->getSuccessor(i)->removePredecessor(TI->getParent());
1817 // Remove the old terminator.
1818 TI->eraseFromParent();
1822 // Finally, delete the basic block.
1823 F->getBasicBlockList().erase(DeadBB);
1825 BlocksToErase.clear();
1828 // If we inferred constant or undef return values for a function, we replaced
1829 // all call uses with the inferred value. This means we don't need to bother
1830 // actually returning anything from the function. Replace all return
1831 // instructions with return undef.
1832 // TODO: Process multiple value ret instructions also.
1833 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
1834 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
1835 E = RV.end(); I != E; ++I)
1836 if (!I->second.isOverdefined() &&
1837 I->first->getReturnType() != Type::VoidTy) {
1838 Function *F = I->first;
1839 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1840 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
1841 if (!isa<UndefValue>(RI->getOperand(0)))
1842 RI->setOperand(0, UndefValue::get(F->getReturnType()));
1845 // If we infered constant or undef values for globals variables, we can delete
1846 // the global and any stores that remain to it.
1847 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
1848 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
1849 E = TG.end(); I != E; ++I) {
1850 GlobalVariable *GV = I->first;
1851 assert(!I->second.isOverdefined() &&
1852 "Overdefined values should have been taken out of the map!");
1853 DOUT << "Found that GV '" << GV->getNameStart() << "' is constant!\n";
1854 while (!GV->use_empty()) {
1855 StoreInst *SI = cast<StoreInst>(GV->use_back());
1856 SI->eraseFromParent();
1858 M.getGlobalList().erase(GV);