1 //===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements sparse conditional constant propagation and merging:
11 //
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
17 //
18 //===----------------------------------------------------------------------===//
19
20 #define DEBUG_TYPE "sccp"
21 #include "llvm/Transforms/Scalar.h"
22 #include "llvm/Transforms/IPO.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/Transforms/Utils/Local.h"
30 #include "llvm/Target/TargetData.h"
31 #include "llvm/Support/CallSite.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/InstVisitor.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/ADT/DenseMap.h"
37 #include "llvm/ADT/DenseSet.h"
38 #include "llvm/ADT/PointerIntPair.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/SmallVector.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/ADT/STLExtras.h"
43 #include <algorithm>
44 #include <map>
45 using namespace llvm;
46
47 STATISTIC(NumInstRemoved, "Number of instructions removed");
48 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
49
50 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP");
51 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
52 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
53
54 namespace {
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.
57 ///
58 class LatticeVal {
59 enum LatticeValueTy {
60 /// undefined - This LLVM Value has no known value yet.
61 undefined,
62
63 /// constant - This LLVM Value has a specific constant value.
64 constant,
65
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
69 /// asserting.
70 forcedconstant,
71
72 /// overdefined - This instruction is not known to be constant, and we know
73 /// it has a value.
74 overdefined
75 };
76
77 /// Val: This stores the current lattice value along with the Constant* for
78 /// the constant if this is a 'constant' or 'forcedconstant' value.
79 PointerIntPair<Constant *, 2, LatticeValueTy> Val;
80
getLatticeValue() const81 LatticeValueTy getLatticeValue() const {
82 return Val.getInt();
83 }
84
85 public:
LatticeVal()86 LatticeVal() : Val(0, undefined) {}
87
isUndefined() const88 bool isUndefined() const { return getLatticeValue() == undefined; }
isConstant() const89 bool isConstant() const {
90 return getLatticeValue() == constant || getLatticeValue() == forcedconstant;
91 }
isOverdefined() const92 bool isOverdefined() const { return getLatticeValue() == overdefined; }
93
getConstant() const94 Constant *getConstant() const {
95 assert(isConstant() && "Cannot get the constant of a non-constant!");
96 return Val.getPointer();
97 }
98
99 /// markOverdefined - Return true if this is a change in status.
markOverdefined()100 bool markOverdefined() {
101 if (isOverdefined())
102 return false;
103
104 Val.setInt(overdefined);
105 return true;
106 }
107
108 /// markConstant - Return true if this is a change in status.
markConstant(Constant * V)109 bool markConstant(Constant *V) {
110 if (getLatticeValue() == constant) { // Constant but not forcedconstant.
111 assert(getConstant() == V && "Marking constant with different value");
112 return false;
113 }
114
115 if (isUndefined()) {
116 Val.setInt(constant);
117 assert(V && "Marking constant with NULL");
118 Val.setPointer(V);
119 } else {
120 assert(getLatticeValue() == forcedconstant &&
121 "Cannot move from overdefined to constant!");
122 // Stay at forcedconstant if the constant is the same.
123 if (V == getConstant()) return false;
124
125 // Otherwise, we go to overdefined. Assumptions made based on the
126 // forced value are possibly wrong. Assuming this is another constant
127 // could expose a contradiction.
128 Val.setInt(overdefined);
129 }
130 return true;
131 }
132
133 /// getConstantInt - If this is a constant with a ConstantInt value, return it
134 /// otherwise return null.
getConstantInt() const135 ConstantInt *getConstantInt() const {
136 if (isConstant())
137 return dyn_cast<ConstantInt>(getConstant());
138 return 0;
139 }
140
markForcedConstant(Constant * V)141 void markForcedConstant(Constant *V) {
142 assert(isUndefined() && "Can't force a defined value!");
143 Val.setInt(forcedconstant);
144 Val.setPointer(V);
145 }
146 };
147 } // end anonymous namespace.
148
149
150 namespace {
151
152 //===----------------------------------------------------------------------===//
153 //
154 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional
155 /// Constant Propagation.
156 ///
157 class SCCPSolver : public InstVisitor<SCCPSolver> {
158 const TargetData *TD;
159 SmallPtrSet<BasicBlock*, 8> BBExecutable; // The BBs that are executable.
160 DenseMap<Value*, LatticeVal> ValueState; // The state each value is in.
161
162 /// StructValueState - This maintains ValueState for values that have
163 /// StructType, for example for formal arguments, calls, insertelement, etc.
164 ///
165 DenseMap<std::pair<Value*, unsigned>, LatticeVal> StructValueState;
166
167 /// GlobalValue - If we are tracking any values for the contents of a global
168 /// variable, we keep a mapping from the constant accessor to the element of
169 /// the global, to the currently known value. If the value becomes
170 /// overdefined, it's entry is simply removed from this map.
171 DenseMap<GlobalVariable*, LatticeVal> TrackedGlobals;
172
173 /// TrackedRetVals - If we are tracking arguments into and the return
174 /// value out of a function, it will have an entry in this map, indicating
175 /// what the known return value for the function is.
176 DenseMap<Function*, LatticeVal> TrackedRetVals;
177
178 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
179 /// that return multiple values.
180 DenseMap<std::pair<Function*, unsigned>, LatticeVal> TrackedMultipleRetVals;
181
182 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is
183 /// represented here for efficient lookup.
184 SmallPtrSet<Function*, 16> MRVFunctionsTracked;
185
186 /// TrackingIncomingArguments - This is the set of functions for whose
187 /// arguments we make optimistic assumptions about and try to prove as
188 /// constants.
189 SmallPtrSet<Function*, 16> TrackingIncomingArguments;
190
191 /// The reason for two worklists is that overdefined is the lowest state
192 /// on the lattice, and moving things to overdefined as fast as possible
193 /// makes SCCP converge much faster.
194 ///
195 /// By having a separate worklist, we accomplish this because everything
196 /// possibly overdefined will become overdefined at the soonest possible
197 /// point.
198 SmallVector<Value*, 64> OverdefinedInstWorkList;
199 SmallVector<Value*, 64> InstWorkList;
200
201
202 SmallVector<BasicBlock*, 64> BBWorkList; // The BasicBlock work list
203
204 /// UsersOfOverdefinedPHIs - Keep track of any users of PHI nodes that are not
205 /// overdefined, despite the fact that the PHI node is overdefined.
206 std::multimap<PHINode*, Instruction*> UsersOfOverdefinedPHIs;
207
208 /// KnownFeasibleEdges - Entries in this set are edges which have already had
209 /// PHI nodes retriggered.
210 typedef std::pair<BasicBlock*, BasicBlock*> Edge;
211 DenseSet<Edge> KnownFeasibleEdges;
212 public:
SCCPSolver(const TargetData * td)213 SCCPSolver(const TargetData *td) : TD(td) {}
214
215 /// MarkBlockExecutable - This method can be used by clients to mark all of
216 /// the blocks that are known to be intrinsically live in the processed unit.
217 ///
218 /// This returns true if the block was not considered live before.
MarkBlockExecutable(BasicBlock * BB)219 bool MarkBlockExecutable(BasicBlock *BB) {
220 if (!BBExecutable.insert(BB)) return false;
221 DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << "\n");
222 BBWorkList.push_back(BB); // Add the block to the work list!
223 return true;
224 }
225
226 /// TrackValueOfGlobalVariable - Clients can use this method to
227 /// inform the SCCPSolver that it should track loads and stores to the
228 /// specified global variable if it can. This is only legal to call if
229 /// performing Interprocedural SCCP.
TrackValueOfGlobalVariable(GlobalVariable * GV)230 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
231 // We only track the contents of scalar globals.
232 if (GV->getType()->getElementType()->isSingleValueType()) {
233 LatticeVal &IV = TrackedGlobals[GV];
234 if (!isa<UndefValue>(GV->getInitializer()))
235 IV.markConstant(GV->getInitializer());
236 }
237 }
238
239 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
240 /// and out of the specified function (which cannot have its address taken),
241 /// this method must be called.
AddTrackedFunction(Function * F)242 void AddTrackedFunction(Function *F) {
243 // Add an entry, F -> undef.
244 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
245 MRVFunctionsTracked.insert(F);
246 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
247 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
248 LatticeVal()));
249 } else
250 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
251 }
252
AddArgumentTrackedFunction(Function * F)253 void AddArgumentTrackedFunction(Function *F) {
254 TrackingIncomingArguments.insert(F);
255 }
256
257 /// Solve - Solve for constants and executable blocks.
258 ///
259 void Solve();
260
261 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
262 /// that branches on undef values cannot reach any of their successors.
263 /// However, this is not a safe assumption. After we solve dataflow, this
264 /// method should be use to handle this. If this returns true, the solver
265 /// should be rerun.
266 bool ResolvedUndefsIn(Function &F);
267
isBlockExecutable(BasicBlock * BB) const268 bool isBlockExecutable(BasicBlock *BB) const {
269 return BBExecutable.count(BB);
270 }
271
getLatticeValueFor(Value * V) const272 LatticeVal getLatticeValueFor(Value *V) const {
273 DenseMap<Value*, LatticeVal>::const_iterator I = ValueState.find(V);
274 assert(I != ValueState.end() && "V is not in valuemap!");
275 return I->second;
276 }
277
278 /*LatticeVal getStructLatticeValueFor(Value *V, unsigned i) const {
279 DenseMap<std::pair<Value*, unsigned>, LatticeVal>::const_iterator I =
280 StructValueState.find(std::make_pair(V, i));
281 assert(I != StructValueState.end() && "V is not in valuemap!");
282 return I->second;
283 }*/
284
285 /// getTrackedRetVals - Get the inferred return value map.
286 ///
getTrackedRetVals()287 const DenseMap<Function*, LatticeVal> &getTrackedRetVals() {
288 return TrackedRetVals;
289 }
290
291 /// getTrackedGlobals - Get and return the set of inferred initializers for
292 /// global variables.
getTrackedGlobals()293 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
294 return TrackedGlobals;
295 }
296
markOverdefined(Value * V)297 void markOverdefined(Value *V) {
298 assert(!V->getType()->isStructTy() && "Should use other method");
299 markOverdefined(ValueState[V], V);
300 }
301
302 /// markAnythingOverdefined - Mark the specified value overdefined. This
303 /// works with both scalars and structs.
markAnythingOverdefined(Value * V)304 void markAnythingOverdefined(Value *V) {
305 if (StructType *STy = dyn_cast<StructType>(V->getType()))
306 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
307 markOverdefined(getStructValueState(V, i), V);
308 else
309 markOverdefined(V);
310 }
311
312 private:
313 // markConstant - Make a value be marked as "constant". If the value
314 // is not already a constant, add it to the instruction work list so that
315 // the users of the instruction are updated later.
316 //
markConstant(LatticeVal & IV,Value * V,Constant * C)317 void markConstant(LatticeVal &IV, Value *V, Constant *C) {
318 if (!IV.markConstant(C)) return;
319 DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n');
320 if (IV.isOverdefined())
321 OverdefinedInstWorkList.push_back(V);
322 else
323 InstWorkList.push_back(V);
324 }
325
markConstant(Value * V,Constant * C)326 void markConstant(Value *V, Constant *C) {
327 assert(!V->getType()->isStructTy() && "Should use other method");
328 markConstant(ValueState[V], V, C);
329 }
330
markForcedConstant(Value * V,Constant * C)331 void markForcedConstant(Value *V, Constant *C) {
332 assert(!V->getType()->isStructTy() && "Should use other method");
333 LatticeVal &IV = ValueState[V];
334 IV.markForcedConstant(C);
335 DEBUG(dbgs() << "markForcedConstant: " << *C << ": " << *V << '\n');
336 if (IV.isOverdefined())
337 OverdefinedInstWorkList.push_back(V);
338 else
339 InstWorkList.push_back(V);
340 }
341
342
343 // markOverdefined - Make a value be marked as "overdefined". If the
344 // value is not already overdefined, add it to the overdefined instruction
345 // work list so that the users of the instruction are updated later.
markOverdefined(LatticeVal & IV,Value * V)346 void markOverdefined(LatticeVal &IV, Value *V) {
347 if (!IV.markOverdefined()) return;
348
349 DEBUG(dbgs() << "markOverdefined: ";
350 if (Function *F = dyn_cast<Function>(V))
351 dbgs() << "Function '" << F->getName() << "'\n";
352 else
353 dbgs() << *V << '\n');
354 // Only instructions go on the work list
355 OverdefinedInstWorkList.push_back(V);
356 }
357
mergeInValue(LatticeVal & IV,Value * V,LatticeVal MergeWithV)358 void mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) {
359 if (IV.isOverdefined() || MergeWithV.isUndefined())
360 return; // Noop.
361 if (MergeWithV.isOverdefined())
362 markOverdefined(IV, V);
363 else if (IV.isUndefined())
364 markConstant(IV, V, MergeWithV.getConstant());
365 else if (IV.getConstant() != MergeWithV.getConstant())
366 markOverdefined(IV, V);
367 }
368
mergeInValue(Value * V,LatticeVal MergeWithV)369 void mergeInValue(Value *V, LatticeVal MergeWithV) {
370 assert(!V->getType()->isStructTy() && "Should use other method");
371 mergeInValue(ValueState[V], V, MergeWithV);
372 }
373
374
375 /// getValueState - Return the LatticeVal object that corresponds to the
376 /// value. This function handles the case when the value hasn't been seen yet
377 /// by properly seeding constants etc.
getValueState(Value * V)378 LatticeVal &getValueState(Value *V) {
379 assert(!V->getType()->isStructTy() && "Should use getStructValueState");
380
381 std::pair<DenseMap<Value*, LatticeVal>::iterator, bool> I =
382 ValueState.insert(std::make_pair(V, LatticeVal()));
383 LatticeVal &LV = I.first->second;
384
385 if (!I.second)
386 return LV; // Common case, already in the map.
387
388 if (Constant *C = dyn_cast<Constant>(V)) {
389 // Undef values remain undefined.
390 if (!isa<UndefValue>(V))
391 LV.markConstant(C); // Constants are constant
392 }
393
394 // All others are underdefined by default.
395 return LV;
396 }
397
398 /// getStructValueState - Return the LatticeVal object that corresponds to the
399 /// value/field pair. This function handles the case when the value hasn't
400 /// been seen yet by properly seeding constants etc.
getStructValueState(Value * V,unsigned i)401 LatticeVal &getStructValueState(Value *V, unsigned i) {
402 assert(V->getType()->isStructTy() && "Should use getValueState");
403 assert(i < cast<StructType>(V->getType())->getNumElements() &&
404 "Invalid element #");
405
406 std::pair<DenseMap<std::pair<Value*, unsigned>, LatticeVal>::iterator,
407 bool> I = StructValueState.insert(
408 std::make_pair(std::make_pair(V, i), LatticeVal()));
409 LatticeVal &LV = I.first->second;
410
411 if (!I.second)
412 return LV; // Common case, already in the map.
413
414 if (Constant *C = dyn_cast<Constant>(V)) {
415 if (isa<UndefValue>(C))
416 ; // Undef values remain undefined.
417 else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C))
418 LV.markConstant(CS->getOperand(i)); // Constants are constant.
419 else if (isa<ConstantAggregateZero>(C)) {
420 Type *FieldTy = cast<StructType>(V->getType())->getElementType(i);
421 LV.markConstant(Constant::getNullValue(FieldTy));
422 } else
423 LV.markOverdefined(); // Unknown sort of constant.
424 }
425
426 // All others are underdefined by default.
427 return LV;
428 }
429
430
431 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
432 /// work list if it is not already executable.
markEdgeExecutable(BasicBlock * Source,BasicBlock * Dest)433 void markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
434 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
435 return; // This edge is already known to be executable!
436
437 if (!MarkBlockExecutable(Dest)) {
438 // If the destination is already executable, we just made an *edge*
439 // feasible that wasn't before. Revisit the PHI nodes in the block
440 // because they have potentially new operands.
441 DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName()
442 << " -> " << Dest->getName() << "\n");
443
444 PHINode *PN;
445 for (BasicBlock::iterator I = Dest->begin();
446 (PN = dyn_cast<PHINode>(I)); ++I)
447 visitPHINode(*PN);
448 }
449 }
450
451 // getFeasibleSuccessors - Return a vector of booleans to indicate which
452 // successors are reachable from a given terminator instruction.
453 //
454 void getFeasibleSuccessors(TerminatorInst &TI, SmallVector<bool, 16> &Succs);
455
456 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
457 // block to the 'To' basic block is currently feasible.
458 //
459 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
460
461 // OperandChangedState - This method is invoked on all of the users of an
462 // instruction that was just changed state somehow. Based on this
463 // information, we need to update the specified user of this instruction.
464 //
OperandChangedState(Instruction * I)465 void OperandChangedState(Instruction *I) {
466 if (BBExecutable.count(I->getParent())) // Inst is executable?
467 visit(*I);
468 }
469
470 /// RemoveFromOverdefinedPHIs - If I has any entries in the
471 /// UsersOfOverdefinedPHIs map for PN, remove them now.
RemoveFromOverdefinedPHIs(Instruction * I,PHINode * PN)472 void RemoveFromOverdefinedPHIs(Instruction *I, PHINode *PN) {
473 if (UsersOfOverdefinedPHIs.empty()) return;
474 typedef std::multimap<PHINode*, Instruction*>::iterator ItTy;
475 std::pair<ItTy, ItTy> Range = UsersOfOverdefinedPHIs.equal_range(PN);
476 for (ItTy It = Range.first, E = Range.second; It != E;) {
477 if (It->second == I)
478 UsersOfOverdefinedPHIs.erase(It++);
479 else
480 ++It;
481 }
482 }
483
484 /// InsertInOverdefinedPHIs - Insert an entry in the UsersOfOverdefinedPHIS
485 /// map for I and PN, but if one is there already, do not create another.
486 /// (Duplicate entries do not break anything directly, but can lead to
487 /// exponential growth of the table in rare cases.)
InsertInOverdefinedPHIs(Instruction * I,PHINode * PN)488 void InsertInOverdefinedPHIs(Instruction *I, PHINode *PN) {
489 typedef std::multimap<PHINode*, Instruction*>::iterator ItTy;
490 std::pair<ItTy, ItTy> Range = UsersOfOverdefinedPHIs.equal_range(PN);
491 for (ItTy J = Range.first, E = Range.second; J != E; ++J)
492 if (J->second == I)
493 return;
494 UsersOfOverdefinedPHIs.insert(std::make_pair(PN, I));
495 }
496
497 private:
498 friend class InstVisitor<SCCPSolver>;
499
500 // visit implementations - Something changed in this instruction. Either an
501 // operand made a transition, or the instruction is newly executable. Change
502 // the value type of I to reflect these changes if appropriate.
503 void visitPHINode(PHINode &I);
504
505 // Terminators
506 void visitReturnInst(ReturnInst &I);
507 void visitTerminatorInst(TerminatorInst &TI);
508
509 void visitCastInst(CastInst &I);
510 void visitSelectInst(SelectInst &I);
511 void visitBinaryOperator(Instruction &I);
512 void visitCmpInst(CmpInst &I);
513 void visitExtractElementInst(ExtractElementInst &I);
514 void visitInsertElementInst(InsertElementInst &I);
515 void visitShuffleVectorInst(ShuffleVectorInst &I);
516 void visitExtractValueInst(ExtractValueInst &EVI);
517 void visitInsertValueInst(InsertValueInst &IVI);
visitLandingPadInst(LandingPadInst & I)518 void visitLandingPadInst(LandingPadInst &I) { markAnythingOverdefined(&I); }
519
520 // Instructions that cannot be folded away.
521 void visitStoreInst (StoreInst &I);
522 void visitLoadInst (LoadInst &I);
523 void visitGetElementPtrInst(GetElementPtrInst &I);
visitCallInst(CallInst & I)524 void visitCallInst (CallInst &I) {
525 visitCallSite(&I);
526 }
visitInvokeInst(InvokeInst & II)527 void visitInvokeInst (InvokeInst &II) {
528 visitCallSite(&II);
529 visitTerminatorInst(II);
530 }
531 void visitCallSite (CallSite CS);
visitResumeInst(TerminatorInst & I)532 void visitResumeInst (TerminatorInst &I) { /*returns void*/ }
visitUnwindInst(TerminatorInst & I)533 void visitUnwindInst (TerminatorInst &I) { /*returns void*/ }
visitUnreachableInst(TerminatorInst & I)534 void visitUnreachableInst(TerminatorInst &I) { /*returns void*/ }
visitFenceInst(FenceInst & I)535 void visitFenceInst (FenceInst &I) { /*returns void*/ }
visitAtomicCmpXchgInst(AtomicCmpXchgInst & I)536 void visitAtomicCmpXchgInst (AtomicCmpXchgInst &I) { markOverdefined(&I); }
visitAtomicRMWInst(AtomicRMWInst & I)537 void visitAtomicRMWInst (AtomicRMWInst &I) { markOverdefined(&I); }
visitAllocaInst(Instruction & I)538 void visitAllocaInst (Instruction &I) { markOverdefined(&I); }
visitVAArgInst(Instruction & I)539 void visitVAArgInst (Instruction &I) { markAnythingOverdefined(&I); }
540
visitInstruction(Instruction & I)541 void visitInstruction(Instruction &I) {
542 // If a new instruction is added to LLVM that we don't handle.
543 dbgs() << "SCCP: Don't know how to handle: " << I;
544 markAnythingOverdefined(&I); // Just in case
545 }
546 };
547
548 } // end anonymous namespace
549
550
551 // getFeasibleSuccessors - Return a vector of booleans to indicate which
552 // successors are reachable from a given terminator instruction.
553 //
getFeasibleSuccessors(TerminatorInst & TI,SmallVector<bool,16> & Succs)554 void SCCPSolver::getFeasibleSuccessors(TerminatorInst &TI,
555 SmallVector<bool, 16> &Succs) {
556 Succs.resize(TI.getNumSuccessors());
557 if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
558 if (BI->isUnconditional()) {
559 Succs[0] = true;
560 return;
561 }
562
563 LatticeVal BCValue = getValueState(BI->getCondition());
564 ConstantInt *CI = BCValue.getConstantInt();
565 if (CI == 0) {
566 // Overdefined condition variables, and branches on unfoldable constant
567 // conditions, mean the branch could go either way.
568 if (!BCValue.isUndefined())
569 Succs[0] = Succs[1] = true;
570 return;
571 }
572
573 // Constant condition variables mean the branch can only go a single way.
574 Succs[CI->isZero()] = true;
575 return;
576 }
577
578 if (isa<InvokeInst>(TI)) {
579 // Invoke instructions successors are always executable.
580 Succs[0] = Succs[1] = true;
581 return;
582 }
583
584 if (SwitchInst *SI = dyn_cast<SwitchInst>(&TI)) {
585 if (TI.getNumSuccessors() < 2) {
586 Succs[0] = true;
587 return;
588 }
589 LatticeVal SCValue = getValueState(SI->getCondition());
590 ConstantInt *CI = SCValue.getConstantInt();
591
592 if (CI == 0) { // Overdefined or undefined condition?
593 // All destinations are executable!
594 if (!SCValue.isUndefined())
595 Succs.assign(TI.getNumSuccessors(), true);
596 return;
597 }
598
599 Succs[SI->findCaseValue(CI)] = true;
600 return;
601 }
602
603 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
604 if (isa<IndirectBrInst>(&TI)) {
605 // Just mark all destinations executable!
606 Succs.assign(TI.getNumSuccessors(), true);
607 return;
608 }
609
610 #ifndef NDEBUG
611 dbgs() << "Unknown terminator instruction: " << TI << '\n';
612 #endif
613 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
614 }
615
616
617 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
618 // block to the 'To' basic block is currently feasible.
619 //
isEdgeFeasible(BasicBlock * From,BasicBlock * To)620 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
621 assert(BBExecutable.count(To) && "Dest should always be alive!");
622
623 // Make sure the source basic block is executable!!
624 if (!BBExecutable.count(From)) return false;
625
626 // Check to make sure this edge itself is actually feasible now.
627 TerminatorInst *TI = From->getTerminator();
628 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
629 if (BI->isUnconditional())
630 return true;
631
632 LatticeVal BCValue = getValueState(BI->getCondition());
633
634 // Overdefined condition variables mean the branch could go either way,
635 // undef conditions mean that neither edge is feasible yet.
636 ConstantInt *CI = BCValue.getConstantInt();
637 if (CI == 0)
638 return !BCValue.isUndefined();
639
640 // Constant condition variables mean the branch can only go a single way.
641 return BI->getSuccessor(CI->isZero()) == To;
642 }
643
644 // Invoke instructions successors are always executable.
645 if (isa<InvokeInst>(TI))
646 return true;
647
648 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
649 if (SI->getNumSuccessors() < 2)
650 return true;
651
652 LatticeVal SCValue = getValueState(SI->getCondition());
653 ConstantInt *CI = SCValue.getConstantInt();
654
655 if (CI == 0)
656 return !SCValue.isUndefined();
657
658 // Make sure to skip the "default value" which isn't a value
659 for (unsigned i = 1, E = SI->getNumSuccessors(); i != E; ++i)
660 if (SI->getSuccessorValue(i) == CI) // Found the taken branch.
661 return SI->getSuccessor(i) == To;
662
663 // If the constant value is not equal to any of the branches, we must
664 // execute default branch.
665 return SI->getDefaultDest() == To;
666 }
667
668 // Just mark all destinations executable!
669 // TODO: This could be improved if the operand is a [cast of a] BlockAddress.
670 if (isa<IndirectBrInst>(TI))
671 return true;
672
673 #ifndef NDEBUG
674 dbgs() << "Unknown terminator instruction: " << *TI << '\n';
675 #endif
676 llvm_unreachable(0);
677 }
678
679 // visit Implementations - Something changed in this instruction, either an
680 // operand made a transition, or the instruction is newly executable. Change
681 // the value type of I to reflect these changes if appropriate. This method
682 // makes sure to do the following actions:
683 //
684 // 1. If a phi node merges two constants in, and has conflicting value coming
685 // from different branches, or if the PHI node merges in an overdefined
686 // value, then the PHI node becomes overdefined.
687 // 2. If a phi node merges only constants in, and they all agree on value, the
688 // PHI node becomes a constant value equal to that.
689 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
690 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
691 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
692 // 6. If a conditional branch has a value that is constant, make the selected
693 // destination executable
694 // 7. If a conditional branch has a value that is overdefined, make all
695 // successors executable.
696 //
visitPHINode(PHINode & PN)697 void SCCPSolver::visitPHINode(PHINode &PN) {
698 // If this PN returns a struct, just mark the result overdefined.
699 // TODO: We could do a lot better than this if code actually uses this.
700 if (PN.getType()->isStructTy())
701 return markAnythingOverdefined(&PN);
702
703 if (getValueState(&PN).isOverdefined()) {
704 // There may be instructions using this PHI node that are not overdefined
705 // themselves. If so, make sure that they know that the PHI node operand
706 // changed.
707 typedef std::multimap<PHINode*, Instruction*>::iterator ItTy;
708 std::pair<ItTy, ItTy> Range = UsersOfOverdefinedPHIs.equal_range(&PN);
709
710 if (Range.first == Range.second)
711 return;
712
713 SmallVector<Instruction*, 16> Users;
714 for (ItTy I = Range.first, E = Range.second; I != E; ++I)
715 Users.push_back(I->second);
716 while (!Users.empty())
717 visit(Users.pop_back_val());
718 return; // Quick exit
719 }
720
721 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
722 // and slow us down a lot. Just mark them overdefined.
723 if (PN.getNumIncomingValues() > 64)
724 return markOverdefined(&PN);
725
726 // Look at all of the executable operands of the PHI node. If any of them
727 // are overdefined, the PHI becomes overdefined as well. If they are all
728 // constant, and they agree with each other, the PHI becomes the identical
729 // constant. If they are constant and don't agree, the PHI is overdefined.
730 // If there are no executable operands, the PHI remains undefined.
731 //
732 Constant *OperandVal = 0;
733 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
734 LatticeVal IV = getValueState(PN.getIncomingValue(i));
735 if (IV.isUndefined()) continue; // Doesn't influence PHI node.
736
737 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
738 continue;
739
740 if (IV.isOverdefined()) // PHI node becomes overdefined!
741 return markOverdefined(&PN);
742
743 if (OperandVal == 0) { // Grab the first value.
744 OperandVal = IV.getConstant();
745 continue;
746 }
747
748 // There is already a reachable operand. If we conflict with it,
749 // then the PHI node becomes overdefined. If we agree with it, we
750 // can continue on.
751
752 // Check to see if there are two different constants merging, if so, the PHI
753 // node is overdefined.
754 if (IV.getConstant() != OperandVal)
755 return markOverdefined(&PN);
756 }
757
758 // If we exited the loop, this means that the PHI node only has constant
759 // arguments that agree with each other(and OperandVal is the constant) or
760 // OperandVal is null because there are no defined incoming arguments. If
761 // this is the case, the PHI remains undefined.
762 //
763 if (OperandVal)
764 markConstant(&PN, OperandVal); // Acquire operand value
765 }
766
767
768
769
visitReturnInst(ReturnInst & I)770 void SCCPSolver::visitReturnInst(ReturnInst &I) {
771 if (I.getNumOperands() == 0) return; // ret void
772
773 Function *F = I.getParent()->getParent();
774 Value *ResultOp = I.getOperand(0);
775
776 // If we are tracking the return value of this function, merge it in.
777 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) {
778 DenseMap<Function*, LatticeVal>::iterator TFRVI =
779 TrackedRetVals.find(F);
780 if (TFRVI != TrackedRetVals.end()) {
781 mergeInValue(TFRVI->second, F, getValueState(ResultOp));
782 return;
783 }
784 }
785
786 // Handle functions that return multiple values.
787 if (!TrackedMultipleRetVals.empty()) {
788 if (StructType *STy = dyn_cast<StructType>(ResultOp->getType()))
789 if (MRVFunctionsTracked.count(F))
790 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
791 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F,
792 getStructValueState(ResultOp, i));
793
794 }
795 }
796
visitTerminatorInst(TerminatorInst & TI)797 void SCCPSolver::visitTerminatorInst(TerminatorInst &TI) {
798 SmallVector<bool, 16> SuccFeasible;
799 getFeasibleSuccessors(TI, SuccFeasible);
800
801 BasicBlock *BB = TI.getParent();
802
803 // Mark all feasible successors executable.
804 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
805 if (SuccFeasible[i])
806 markEdgeExecutable(BB, TI.getSuccessor(i));
807 }
808
visitCastInst(CastInst & I)809 void SCCPSolver::visitCastInst(CastInst &I) {
810 LatticeVal OpSt = getValueState(I.getOperand(0));
811 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand
812 markOverdefined(&I);
813 else if (OpSt.isConstant()) // Propagate constant value
814 markConstant(&I, ConstantExpr::getCast(I.getOpcode(),
815 OpSt.getConstant(), I.getType()));
816 }
817
818
visitExtractValueInst(ExtractValueInst & EVI)819 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
820 // If this returns a struct, mark all elements over defined, we don't track
821 // structs in structs.
822 if (EVI.getType()->isStructTy())
823 return markAnythingOverdefined(&EVI);
824
825 // If this is extracting from more than one level of struct, we don't know.
826 if (EVI.getNumIndices() != 1)
827 return markOverdefined(&EVI);
828
829 Value *AggVal = EVI.getAggregateOperand();
830 if (AggVal->getType()->isStructTy()) {
831 unsigned i = *EVI.idx_begin();
832 LatticeVal EltVal = getStructValueState(AggVal, i);
833 mergeInValue(getValueState(&EVI), &EVI, EltVal);
834 } else {
835 // Otherwise, must be extracting from an array.
836 return markOverdefined(&EVI);
837 }
838 }
839
visitInsertValueInst(InsertValueInst & IVI)840 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
841 StructType *STy = dyn_cast<StructType>(IVI.getType());
842 if (STy == 0)
843 return markOverdefined(&IVI);
844
845 // If this has more than one index, we can't handle it, drive all results to
846 // undef.
847 if (IVI.getNumIndices() != 1)
848 return markAnythingOverdefined(&IVI);
849
850 Value *Aggr = IVI.getAggregateOperand();
851 unsigned Idx = *IVI.idx_begin();
852
853 // Compute the result based on what we're inserting.
854 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
855 // This passes through all values that aren't the inserted element.
856 if (i != Idx) {
857 LatticeVal EltVal = getStructValueState(Aggr, i);
858 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal);
859 continue;
860 }
861
862 Value *Val = IVI.getInsertedValueOperand();
863 if (Val->getType()->isStructTy())
864 // We don't track structs in structs.
865 markOverdefined(getStructValueState(&IVI, i), &IVI);
866 else {
867 LatticeVal InVal = getValueState(Val);
868 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal);
869 }
870 }
871 }
872
visitSelectInst(SelectInst & I)873 void SCCPSolver::visitSelectInst(SelectInst &I) {
874 // If this select returns a struct, just mark the result overdefined.
875 // TODO: We could do a lot better than this if code actually uses this.
876 if (I.getType()->isStructTy())
877 return markAnythingOverdefined(&I);
878
879 LatticeVal CondValue = getValueState(I.getCondition());
880 if (CondValue.isUndefined())
881 return;
882
883 if (ConstantInt *CondCB = CondValue.getConstantInt()) {
884 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
885 mergeInValue(&I, getValueState(OpVal));
886 return;
887 }
888
889 // Otherwise, the condition is overdefined or a constant we can't evaluate.
890 // See if we can produce something better than overdefined based on the T/F
891 // value.
892 LatticeVal TVal = getValueState(I.getTrueValue());
893 LatticeVal FVal = getValueState(I.getFalseValue());
894
895 // select ?, C, C -> C.
896 if (TVal.isConstant() && FVal.isConstant() &&
897 TVal.getConstant() == FVal.getConstant())
898 return markConstant(&I, FVal.getConstant());
899
900 if (TVal.isUndefined()) // select ?, undef, X -> X.
901 return mergeInValue(&I, FVal);
902 if (FVal.isUndefined()) // select ?, X, undef -> X.
903 return mergeInValue(&I, TVal);
904 markOverdefined(&I);
905 }
906
907 // Handle Binary Operators.
visitBinaryOperator(Instruction & I)908 void SCCPSolver::visitBinaryOperator(Instruction &I) {
909 LatticeVal V1State = getValueState(I.getOperand(0));
910 LatticeVal V2State = getValueState(I.getOperand(1));
911
912 LatticeVal &IV = ValueState[&I];
913 if (IV.isOverdefined()) return;
914
915 if (V1State.isConstant() && V2State.isConstant())
916 return markConstant(IV, &I,
917 ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
918 V2State.getConstant()));
919
920 // If something is undef, wait for it to resolve.
921 if (!V1State.isOverdefined() && !V2State.isOverdefined())
922 return;
923
924 // Otherwise, one of our operands is overdefined. Try to produce something
925 // better than overdefined with some tricks.
926
927 // If this is an AND or OR with 0 or -1, it doesn't matter that the other
928 // operand is overdefined.
929 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Or) {
930 LatticeVal *NonOverdefVal = 0;
931 if (!V1State.isOverdefined())
932 NonOverdefVal = &V1State;
933 else if (!V2State.isOverdefined())
934 NonOverdefVal = &V2State;
935
936 if (NonOverdefVal) {
937 if (NonOverdefVal->isUndefined()) {
938 // Could annihilate value.
939 if (I.getOpcode() == Instruction::And)
940 markConstant(IV, &I, Constant::getNullValue(I.getType()));
941 else if (VectorType *PT = dyn_cast<VectorType>(I.getType()))
942 markConstant(IV, &I, Constant::getAllOnesValue(PT));
943 else
944 markConstant(IV, &I,
945 Constant::getAllOnesValue(I.getType()));
946 return;
947 }
948
949 if (I.getOpcode() == Instruction::And) {
950 // X and 0 = 0
951 if (NonOverdefVal->getConstant()->isNullValue())
952 return markConstant(IV, &I, NonOverdefVal->getConstant());
953 } else {
954 if (ConstantInt *CI = NonOverdefVal->getConstantInt())
955 if (CI->isAllOnesValue()) // X or -1 = -1
956 return markConstant(IV, &I, NonOverdefVal->getConstant());
957 }
958 }
959 }
960
961
962 // If both operands are PHI nodes, it is possible that this instruction has
963 // a constant value, despite the fact that the PHI node doesn't. Check for
964 // this condition now.
965 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
966 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
967 if (PN1->getParent() == PN2->getParent()) {
968 // Since the two PHI nodes are in the same basic block, they must have
969 // entries for the same predecessors. Walk the predecessor list, and
970 // if all of the incoming values are constants, and the result of
971 // evaluating this expression with all incoming value pairs is the
972 // same, then this expression is a constant even though the PHI node
973 // is not a constant!
974 LatticeVal Result;
975 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
976 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
977 BasicBlock *InBlock = PN1->getIncomingBlock(i);
978 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
979
980 if (In1.isOverdefined() || In2.isOverdefined()) {
981 Result.markOverdefined();
982 break; // Cannot fold this operation over the PHI nodes!
983 }
984
985 if (In1.isConstant() && In2.isConstant()) {
986 Constant *V = ConstantExpr::get(I.getOpcode(), In1.getConstant(),
987 In2.getConstant());
988 if (Result.isUndefined())
989 Result.markConstant(V);
990 else if (Result.isConstant() && Result.getConstant() != V) {
991 Result.markOverdefined();
992 break;
993 }
994 }
995 }
996
997 // If we found a constant value here, then we know the instruction is
998 // constant despite the fact that the PHI nodes are overdefined.
999 if (Result.isConstant()) {
1000 markConstant(IV, &I, Result.getConstant());
1001 // Remember that this instruction is virtually using the PHI node
1002 // operands.
1003 InsertInOverdefinedPHIs(&I, PN1);
1004 InsertInOverdefinedPHIs(&I, PN2);
1005 return;
1006 }
1007
1008 if (Result.isUndefined())
1009 return;
1010
1011 // Okay, this really is overdefined now. Since we might have
1012 // speculatively thought that this was not overdefined before, and
1013 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
1014 // make sure to clean out any entries that we put there, for
1015 // efficiency.
1016 RemoveFromOverdefinedPHIs(&I, PN1);
1017 RemoveFromOverdefinedPHIs(&I, PN2);
1018 }
1019
1020 markOverdefined(&I);
1021 }
1022
1023 // Handle ICmpInst instruction.
visitCmpInst(CmpInst & I)1024 void SCCPSolver::visitCmpInst(CmpInst &I) {
1025 LatticeVal V1State = getValueState(I.getOperand(0));
1026 LatticeVal V2State = getValueState(I.getOperand(1));
1027
1028 LatticeVal &IV = ValueState[&I];
1029 if (IV.isOverdefined()) return;
1030
1031 if (V1State.isConstant() && V2State.isConstant())
1032 return markConstant(IV, &I, ConstantExpr::getCompare(I.getPredicate(),
1033 V1State.getConstant(),
1034 V2State.getConstant()));
1035
1036 // If operands are still undefined, wait for it to resolve.
1037 if (!V1State.isOverdefined() && !V2State.isOverdefined())
1038 return;
1039
1040 // If something is overdefined, use some tricks to avoid ending up and over
1041 // defined if we can.
1042
1043 // If both operands are PHI nodes, it is possible that this instruction has
1044 // a constant value, despite the fact that the PHI node doesn't. Check for
1045 // this condition now.
1046 if (PHINode *PN1 = dyn_cast<PHINode>(I.getOperand(0)))
1047 if (PHINode *PN2 = dyn_cast<PHINode>(I.getOperand(1)))
1048 if (PN1->getParent() == PN2->getParent()) {
1049 // Since the two PHI nodes are in the same basic block, they must have
1050 // entries for the same predecessors. Walk the predecessor list, and
1051 // if all of the incoming values are constants, and the result of
1052 // evaluating this expression with all incoming value pairs is the
1053 // same, then this expression is a constant even though the PHI node
1054 // is not a constant!
1055 LatticeVal Result;
1056 for (unsigned i = 0, e = PN1->getNumIncomingValues(); i != e; ++i) {
1057 LatticeVal In1 = getValueState(PN1->getIncomingValue(i));
1058 BasicBlock *InBlock = PN1->getIncomingBlock(i);
1059 LatticeVal In2 =getValueState(PN2->getIncomingValueForBlock(InBlock));
1060
1061 if (In1.isOverdefined() || In2.isOverdefined()) {
1062 Result.markOverdefined();
1063 break; // Cannot fold this operation over the PHI nodes!
1064 }
1065
1066 if (In1.isConstant() && In2.isConstant()) {
1067 Constant *V = ConstantExpr::getCompare(I.getPredicate(),
1068 In1.getConstant(),
1069 In2.getConstant());
1070 if (Result.isUndefined())
1071 Result.markConstant(V);
1072 else if (Result.isConstant() && Result.getConstant() != V) {
1073 Result.markOverdefined();
1074 break;
1075 }
1076 }
1077 }
1078
1079 // If we found a constant value here, then we know the instruction is
1080 // constant despite the fact that the PHI nodes are overdefined.
1081 if (Result.isConstant()) {
1082 markConstant(&I, Result.getConstant());
1083 // Remember that this instruction is virtually using the PHI node
1084 // operands.
1085 InsertInOverdefinedPHIs(&I, PN1);
1086 InsertInOverdefinedPHIs(&I, PN2);
1087 return;
1088 }
1089
1090 if (Result.isUndefined())
1091 return;
1092
1093 // Okay, this really is overdefined now. Since we might have
1094 // speculatively thought that this was not overdefined before, and
1095 // added ourselves to the UsersOfOverdefinedPHIs list for the PHIs,
1096 // make sure to clean out any entries that we put there, for
1097 // efficiency.
1098 RemoveFromOverdefinedPHIs(&I, PN1);
1099 RemoveFromOverdefinedPHIs(&I, PN2);
1100 }
1101
1102 markOverdefined(&I);
1103 }
1104
visitExtractElementInst(ExtractElementInst & I)1105 void SCCPSolver::visitExtractElementInst(ExtractElementInst &I) {
1106 // TODO : SCCP does not handle vectors properly.
1107 return markOverdefined(&I);
1108
1109 #if 0
1110 LatticeVal &ValState = getValueState(I.getOperand(0));
1111 LatticeVal &IdxState = getValueState(I.getOperand(1));
1112
1113 if (ValState.isOverdefined() || IdxState.isOverdefined())
1114 markOverdefined(&I);
1115 else if(ValState.isConstant() && IdxState.isConstant())
1116 markConstant(&I, ConstantExpr::getExtractElement(ValState.getConstant(),
1117 IdxState.getConstant()));
1118 #endif
1119 }
1120
visitInsertElementInst(InsertElementInst & I)1121 void SCCPSolver::visitInsertElementInst(InsertElementInst &I) {
1122 // TODO : SCCP does not handle vectors properly.
1123 return markOverdefined(&I);
1124 #if 0
1125 LatticeVal &ValState = getValueState(I.getOperand(0));
1126 LatticeVal &EltState = getValueState(I.getOperand(1));
1127 LatticeVal &IdxState = getValueState(I.getOperand(2));
1128
1129 if (ValState.isOverdefined() || EltState.isOverdefined() ||
1130 IdxState.isOverdefined())
1131 markOverdefined(&I);
1132 else if(ValState.isConstant() && EltState.isConstant() &&
1133 IdxState.isConstant())
1134 markConstant(&I, ConstantExpr::getInsertElement(ValState.getConstant(),
1135 EltState.getConstant(),
1136 IdxState.getConstant()));
1137 else if (ValState.isUndefined() && EltState.isConstant() &&
1138 IdxState.isConstant())
1139 markConstant(&I,ConstantExpr::getInsertElement(UndefValue::get(I.getType()),
1140 EltState.getConstant(),
1141 IdxState.getConstant()));
1142 #endif
1143 }
1144
visitShuffleVectorInst(ShuffleVectorInst & I)1145 void SCCPSolver::visitShuffleVectorInst(ShuffleVectorInst &I) {
1146 // TODO : SCCP does not handle vectors properly.
1147 return markOverdefined(&I);
1148 #if 0
1149 LatticeVal &V1State = getValueState(I.getOperand(0));
1150 LatticeVal &V2State = getValueState(I.getOperand(1));
1151 LatticeVal &MaskState = getValueState(I.getOperand(2));
1152
1153 if (MaskState.isUndefined() ||
1154 (V1State.isUndefined() && V2State.isUndefined()))
1155 return; // Undefined output if mask or both inputs undefined.
1156
1157 if (V1State.isOverdefined() || V2State.isOverdefined() ||
1158 MaskState.isOverdefined()) {
1159 markOverdefined(&I);
1160 } else {
1161 // A mix of constant/undef inputs.
1162 Constant *V1 = V1State.isConstant() ?
1163 V1State.getConstant() : UndefValue::get(I.getType());
1164 Constant *V2 = V2State.isConstant() ?
1165 V2State.getConstant() : UndefValue::get(I.getType());
1166 Constant *Mask = MaskState.isConstant() ?
1167 MaskState.getConstant() : UndefValue::get(I.getOperand(2)->getType());
1168 markConstant(&I, ConstantExpr::getShuffleVector(V1, V2, Mask));
1169 }
1170 #endif
1171 }
1172
1173 // Handle getelementptr instructions. If all operands are constants then we
1174 // can turn this into a getelementptr ConstantExpr.
1175 //
visitGetElementPtrInst(GetElementPtrInst & I)1176 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1177 if (ValueState[&I].isOverdefined()) return;
1178
1179 SmallVector<Constant*, 8> Operands;
1180 Operands.reserve(I.getNumOperands());
1181
1182 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1183 LatticeVal State = getValueState(I.getOperand(i));
1184 if (State.isUndefined())
1185 return; // Operands are not resolved yet.
1186
1187 if (State.isOverdefined())
1188 return markOverdefined(&I);
1189
1190 assert(State.isConstant() && "Unknown state!");
1191 Operands.push_back(State.getConstant());
1192 }
1193
1194 Constant *Ptr = Operands[0];
1195 ArrayRef<Constant *> Indices(Operands.begin() + 1, Operands.end());
1196 markConstant(&I, ConstantExpr::getGetElementPtr(Ptr, Indices));
1197 }
1198
visitStoreInst(StoreInst & SI)1199 void SCCPSolver::visitStoreInst(StoreInst &SI) {
1200 // If this store is of a struct, ignore it.
1201 if (SI.getOperand(0)->getType()->isStructTy())
1202 return;
1203
1204 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1205 return;
1206
1207 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1208 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1209 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1210
1211 // Get the value we are storing into the global, then merge it.
1212 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)));
1213 if (I->second.isOverdefined())
1214 TrackedGlobals.erase(I); // No need to keep tracking this!
1215 }
1216
1217
1218 // Handle load instructions. If the operand is a constant pointer to a constant
1219 // global, we can replace the load with the loaded constant value!
visitLoadInst(LoadInst & I)1220 void SCCPSolver::visitLoadInst(LoadInst &I) {
1221 // If this load is of a struct, just mark the result overdefined.
1222 if (I.getType()->isStructTy())
1223 return markAnythingOverdefined(&I);
1224
1225 LatticeVal PtrVal = getValueState(I.getOperand(0));
1226 if (PtrVal.isUndefined()) return; // The pointer is not resolved yet!
1227
1228 LatticeVal &IV = ValueState[&I];
1229 if (IV.isOverdefined()) return;
1230
1231 if (!PtrVal.isConstant() || I.isVolatile())
1232 return markOverdefined(IV, &I);
1233
1234 Constant *Ptr = PtrVal.getConstant();
1235
1236 // load null -> null
1237 if (isa<ConstantPointerNull>(Ptr) && I.getPointerAddressSpace() == 0)
1238 return markConstant(IV, &I, Constant::getNullValue(I.getType()));
1239
1240 // Transform load (constant global) into the value loaded.
1241 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
1242 if (!TrackedGlobals.empty()) {
1243 // If we are tracking this global, merge in the known value for it.
1244 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1245 TrackedGlobals.find(GV);
1246 if (It != TrackedGlobals.end()) {
1247 mergeInValue(IV, &I, It->second);
1248 return;
1249 }
1250 }
1251 }
1252
1253 // Transform load from a constant into a constant if possible.
1254 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, TD))
1255 return markConstant(IV, &I, C);
1256
1257 // Otherwise we cannot say for certain what value this load will produce.
1258 // Bail out.
1259 markOverdefined(IV, &I);
1260 }
1261
visitCallSite(CallSite CS)1262 void SCCPSolver::visitCallSite(CallSite CS) {
1263 Function *F = CS.getCalledFunction();
1264 Instruction *I = CS.getInstruction();
1265
1266 // The common case is that we aren't tracking the callee, either because we
1267 // are not doing interprocedural analysis or the callee is indirect, or is
1268 // external. Handle these cases first.
1269 if (F == 0 || F->isDeclaration()) {
1270 CallOverdefined:
1271 // Void return and not tracking callee, just bail.
1272 if (I->getType()->isVoidTy()) return;
1273
1274 // Otherwise, if we have a single return value case, and if the function is
1275 // a declaration, maybe we can constant fold it.
1276 if (F && F->isDeclaration() && !I->getType()->isStructTy() &&
1277 canConstantFoldCallTo(F)) {
1278
1279 SmallVector<Constant*, 8> Operands;
1280 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1281 AI != E; ++AI) {
1282 LatticeVal State = getValueState(*AI);
1283
1284 if (State.isUndefined())
1285 return; // Operands are not resolved yet.
1286 if (State.isOverdefined())
1287 return markOverdefined(I);
1288 assert(State.isConstant() && "Unknown state!");
1289 Operands.push_back(State.getConstant());
1290 }
1291
1292 // If we can constant fold this, mark the result of the call as a
1293 // constant.
1294 if (Constant *C = ConstantFoldCall(F, Operands))
1295 return markConstant(I, C);
1296 }
1297
1298 // Otherwise, we don't know anything about this call, mark it overdefined.
1299 return markAnythingOverdefined(I);
1300 }
1301
1302 // If this is a local function that doesn't have its address taken, mark its
1303 // entry block executable and merge in the actual arguments to the call into
1304 // the formal arguments of the function.
1305 if (!TrackingIncomingArguments.empty() && TrackingIncomingArguments.count(F)){
1306 MarkBlockExecutable(F->begin());
1307
1308 // Propagate information from this call site into the callee.
1309 CallSite::arg_iterator CAI = CS.arg_begin();
1310 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1311 AI != E; ++AI, ++CAI) {
1312 // If this argument is byval, and if the function is not readonly, there
1313 // will be an implicit copy formed of the input aggregate.
1314 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1315 markOverdefined(AI);
1316 continue;
1317 }
1318
1319 if (StructType *STy = dyn_cast<StructType>(AI->getType())) {
1320 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1321 LatticeVal CallArg = getStructValueState(*CAI, i);
1322 mergeInValue(getStructValueState(AI, i), AI, CallArg);
1323 }
1324 } else {
1325 mergeInValue(AI, getValueState(*CAI));
1326 }
1327 }
1328 }
1329
1330 // If this is a single/zero retval case, see if we're tracking the function.
1331 if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) {
1332 if (!MRVFunctionsTracked.count(F))
1333 goto CallOverdefined; // Not tracking this callee.
1334
1335 // If we are tracking this callee, propagate the result of the function
1336 // into this call site.
1337 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1338 mergeInValue(getStructValueState(I, i), I,
1339 TrackedMultipleRetVals[std::make_pair(F, i)]);
1340 } else {
1341 DenseMap<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1342 if (TFRVI == TrackedRetVals.end())
1343 goto CallOverdefined; // Not tracking this callee.
1344
1345 // If so, propagate the return value of the callee into this call result.
1346 mergeInValue(I, TFRVI->second);
1347 }
1348 }
1349
Solve()1350 void SCCPSolver::Solve() {
1351 // Process the work lists until they are empty!
1352 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1353 !OverdefinedInstWorkList.empty()) {
1354 // Process the overdefined instruction's work list first, which drives other
1355 // things to overdefined more quickly.
1356 while (!OverdefinedInstWorkList.empty()) {
1357 Value *I = OverdefinedInstWorkList.pop_back_val();
1358
1359 DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n');
1360
1361 // "I" got into the work list because it either made the transition from
1362 // bottom to constant
1363 //
1364 // Anything on this worklist that is overdefined need not be visited
1365 // since all of its users will have already been marked as overdefined
1366 // Update all of the users of this instruction's value.
1367 //
1368 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1369 UI != E; ++UI)
1370 if (Instruction *I = dyn_cast<Instruction>(*UI))
1371 OperandChangedState(I);
1372 }
1373
1374 // Process the instruction work list.
1375 while (!InstWorkList.empty()) {
1376 Value *I = InstWorkList.pop_back_val();
1377
1378 DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n');
1379
1380 // "I" got into the work list because it made the transition from undef to
1381 // constant.
1382 //
1383 // Anything on this worklist that is overdefined need not be visited
1384 // since all of its users will have already been marked as overdefined.
1385 // Update all of the users of this instruction's value.
1386 //
1387 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined())
1388 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1389 UI != E; ++UI)
1390 if (Instruction *I = dyn_cast<Instruction>(*UI))
1391 OperandChangedState(I);
1392 }
1393
1394 // Process the basic block work list.
1395 while (!BBWorkList.empty()) {
1396 BasicBlock *BB = BBWorkList.back();
1397 BBWorkList.pop_back();
1398
1399 DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n');
1400
1401 // Notify all instructions in this basic block that they are newly
1402 // executable.
1403 visit(BB);
1404 }
1405 }
1406 }
1407
1408 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1409 /// that branches on undef values cannot reach any of their successors.
1410 /// However, this is not a safe assumption. After we solve dataflow, this
1411 /// method should be use to handle this. If this returns true, the solver
1412 /// should be rerun.
1413 ///
1414 /// This method handles this by finding an unresolved branch and marking it one
1415 /// of the edges from the block as being feasible, even though the condition
1416 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1417 /// CFG and only slightly pessimizes the analysis results (by marking one,
1418 /// potentially infeasible, edge feasible). This cannot usefully modify the
1419 /// constraints on the condition of the branch, as that would impact other users
1420 /// of the value.
1421 ///
1422 /// This scan also checks for values that use undefs, whose results are actually
1423 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1424 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1425 /// even if X isn't defined.
ResolvedUndefsIn(Function & F)1426 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1427 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1428 if (!BBExecutable.count(BB))
1429 continue;
1430
1431 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1432 // Look for instructions which produce undef values.
1433 if (I->getType()->isVoidTy()) continue;
1434
1435 if (StructType *STy = dyn_cast<StructType>(I->getType())) {
1436 // Only a few things that can be structs matter for undef.
1437
1438 // Tracked calls must never be marked overdefined in ResolvedUndefsIn.
1439 if (CallSite CS = CallSite(I))
1440 if (Function *F = CS.getCalledFunction())
1441 if (MRVFunctionsTracked.count(F))
1442 continue;
1443
1444 // extractvalue and insertvalue don't need to be marked; they are
1445 // tracked as precisely as their operands.
1446 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I))
1447 continue;
1448
1449 // Send the results of everything else to overdefined. We could be
1450 // more precise than this but it isn't worth bothering.
1451 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1452 LatticeVal &LV = getStructValueState(I, i);
1453 if (LV.isUndefined())
1454 markOverdefined(LV, I);
1455 }
1456 continue;
1457 }
1458
1459 LatticeVal &LV = getValueState(I);
1460 if (!LV.isUndefined()) continue;
1461
1462 // extractvalue is safe; check here because the argument is a struct.
1463 if (isa<ExtractValueInst>(I))
1464 continue;
1465
1466 // Compute the operand LatticeVals, for convenience below.
1467 // Anything taking a struct is conservatively assumed to require
1468 // overdefined markings.
1469 if (I->getOperand(0)->getType()->isStructTy()) {
1470 markOverdefined(I);
1471 return true;
1472 }
1473 LatticeVal Op0LV = getValueState(I->getOperand(0));
1474 LatticeVal Op1LV;
1475 if (I->getNumOperands() == 2) {
1476 if (I->getOperand(1)->getType()->isStructTy()) {
1477 markOverdefined(I);
1478 return true;
1479 }
1480
1481 Op1LV = getValueState(I->getOperand(1));
1482 }
1483 // If this is an instructions whose result is defined even if the input is
1484 // not fully defined, propagate the information.
1485 Type *ITy = I->getType();
1486 switch (I->getOpcode()) {
1487 case Instruction::Add:
1488 case Instruction::Sub:
1489 case Instruction::Trunc:
1490 case Instruction::FPTrunc:
1491 case Instruction::BitCast:
1492 break; // Any undef -> undef
1493 case Instruction::FSub:
1494 case Instruction::FAdd:
1495 case Instruction::FMul:
1496 case Instruction::FDiv:
1497 case Instruction::FRem:
1498 // Floating-point binary operation: be conservative.
1499 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1500 markForcedConstant(I, Constant::getNullValue(ITy));
1501 else
1502 markOverdefined(I);
1503 return true;
1504 case Instruction::ZExt:
1505 case Instruction::SExt:
1506 case Instruction::FPToUI:
1507 case Instruction::FPToSI:
1508 case Instruction::FPExt:
1509 case Instruction::PtrToInt:
1510 case Instruction::IntToPtr:
1511 case Instruction::SIToFP:
1512 case Instruction::UIToFP:
1513 // undef -> 0; some outputs are impossible
1514 markForcedConstant(I, Constant::getNullValue(ITy));
1515 return true;
1516 case Instruction::Mul:
1517 case Instruction::And:
1518 // Both operands undef -> undef
1519 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1520 break;
1521 // undef * X -> 0. X could be zero.
1522 // undef & X -> 0. X could be zero.
1523 markForcedConstant(I, Constant::getNullValue(ITy));
1524 return true;
1525
1526 case Instruction::Or:
1527 // Both operands undef -> undef
1528 if (Op0LV.isUndefined() && Op1LV.isUndefined())
1529 break;
1530 // undef | X -> -1. X could be -1.
1531 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1532 return true;
1533
1534 case Instruction::Xor:
1535 // undef ^ undef -> 0; strictly speaking, this is not strictly
1536 // necessary, but we try to be nice to people who expect this
1537 // behavior in simple cases
1538 if (Op0LV.isUndefined() && Op1LV.isUndefined()) {
1539 markForcedConstant(I, Constant::getNullValue(ITy));
1540 return true;
1541 }
1542 // undef ^ X -> undef
1543 break;
1544
1545 case Instruction::SDiv:
1546 case Instruction::UDiv:
1547 case Instruction::SRem:
1548 case Instruction::URem:
1549 // X / undef -> undef. No change.
1550 // X % undef -> undef. No change.
1551 if (Op1LV.isUndefined()) break;
1552
1553 // undef / X -> 0. X could be maxint.
1554 // undef % X -> 0. X could be 1.
1555 markForcedConstant(I, Constant::getNullValue(ITy));
1556 return true;
1557
1558 case Instruction::AShr:
1559 // X >>a undef -> undef.
1560 if (Op1LV.isUndefined()) break;
1561
1562 // undef >>a X -> all ones
1563 markForcedConstant(I, Constant::getAllOnesValue(ITy));
1564 return true;
1565 case Instruction::LShr:
1566 case Instruction::Shl:
1567 // X << undef -> undef.
1568 // X >> undef -> undef.
1569 if (Op1LV.isUndefined()) break;
1570
1571 // undef << X -> 0
1572 // undef >> X -> 0
1573 markForcedConstant(I, Constant::getNullValue(ITy));
1574 return true;
1575 case Instruction::Select:
1576 Op1LV = getValueState(I->getOperand(1));
1577 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1578 if (Op0LV.isUndefined()) {
1579 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1580 Op1LV = getValueState(I->getOperand(2));
1581 } else if (Op1LV.isUndefined()) {
1582 // c ? undef : undef -> undef. No change.
1583 Op1LV = getValueState(I->getOperand(2));
1584 if (Op1LV.isUndefined())
1585 break;
1586 // Otherwise, c ? undef : x -> x.
1587 } else {
1588 // Leave Op1LV as Operand(1)'s LatticeValue.
1589 }
1590
1591 if (Op1LV.isConstant())
1592 markForcedConstant(I, Op1LV.getConstant());
1593 else
1594 markOverdefined(I);
1595 return true;
1596 case Instruction::Load:
1597 // A load here means one of two things: a load of undef from a global,
1598 // a load from an unknown pointer. Either way, having it return undef
1599 // is okay.
1600 break;
1601 case Instruction::ICmp:
1602 // X == undef -> undef. Other comparisons get more complicated.
1603 if (cast<ICmpInst>(I)->isEquality())
1604 break;
1605 markOverdefined(I);
1606 return true;
1607 case Instruction::Call:
1608 case Instruction::Invoke: {
1609 // There are two reasons a call can have an undef result
1610 // 1. It could be tracked.
1611 // 2. It could be constant-foldable.
1612 // Because of the way we solve return values, tracked calls must
1613 // never be marked overdefined in ResolvedUndefsIn.
1614 if (Function *F = CallSite(I).getCalledFunction())
1615 if (TrackedRetVals.count(F))
1616 break;
1617
1618 // If the call is constant-foldable, we mark it overdefined because
1619 // we do not know what return values are valid.
1620 markOverdefined(I);
1621 return true;
1622 }
1623 default:
1624 // If we don't know what should happen here, conservatively mark it
1625 // overdefined.
1626 markOverdefined(I);
1627 return true;
1628 }
1629 }
1630
1631 // Check to see if we have a branch or switch on an undefined value. If so
1632 // we force the branch to go one way or the other to make the successor
1633 // values live. It doesn't really matter which way we force it.
1634 TerminatorInst *TI = BB->getTerminator();
1635 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1636 if (!BI->isConditional()) continue;
1637 if (!getValueState(BI->getCondition()).isUndefined())
1638 continue;
1639
1640 // If the input to SCCP is actually branch on undef, fix the undef to
1641 // false.
1642 if (isa<UndefValue>(BI->getCondition())) {
1643 BI->setCondition(ConstantInt::getFalse(BI->getContext()));
1644 markEdgeExecutable(BB, TI->getSuccessor(1));
1645 return true;
1646 }
1647
1648 // Otherwise, it is a branch on a symbolic value which is currently
1649 // considered to be undef. Handle this by forcing the input value to the
1650 // branch to false.
1651 markForcedConstant(BI->getCondition(),
1652 ConstantInt::getFalse(TI->getContext()));
1653 return true;
1654 }
1655
1656 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1657 if (SI->getNumSuccessors() < 2) // no cases
1658 continue;
1659 if (!getValueState(SI->getCondition()).isUndefined())
1660 continue;
1661
1662 // If the input to SCCP is actually switch on undef, fix the undef to
1663 // the first constant.
1664 if (isa<UndefValue>(SI->getCondition())) {
1665 SI->setCondition(SI->getCaseValue(1));
1666 markEdgeExecutable(BB, TI->getSuccessor(1));
1667 return true;
1668 }
1669
1670 markForcedConstant(SI->getCondition(), SI->getCaseValue(1));
1671 return true;
1672 }
1673 }
1674
1675 return false;
1676 }
1677
1678
1679 namespace {
1680 //===--------------------------------------------------------------------===//
1681 //
1682 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1683 /// Sparse Conditional Constant Propagator.
1684 ///
1685 struct SCCP : public FunctionPass {
1686 static char ID; // Pass identification, replacement for typeid
SCCP__anon967d13930311::SCCP1687 SCCP() : FunctionPass(ID) {
1688 initializeSCCPPass(*PassRegistry::getPassRegistry());
1689 }
1690
1691 // runOnFunction - Run the Sparse Conditional Constant Propagation
1692 // algorithm, and return true if the function was modified.
1693 //
1694 bool runOnFunction(Function &F);
1695 };
1696 } // end anonymous namespace
1697
1698 char SCCP::ID = 0;
1699 INITIALIZE_PASS(SCCP, "sccp",
1700 "Sparse Conditional Constant Propagation", false, false)
1701
1702 // createSCCPPass - This is the public interface to this file.
createSCCPPass()1703 FunctionPass *llvm::createSCCPPass() {
1704 return new SCCP();
1705 }
1706
DeleteInstructionInBlock(BasicBlock * BB)1707 static void DeleteInstructionInBlock(BasicBlock *BB) {
1708 DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
1709 ++NumDeadBlocks;
1710
1711 // Check to see if there are non-terminating instructions to delete.
1712 if (isa<TerminatorInst>(BB->begin()))
1713 return;
1714
1715 // Delete the instructions backwards, as it has a reduced likelihood of having
1716 // to update as many def-use and use-def chains.
1717 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
1718 while (EndInst != BB->begin()) {
1719 // Delete the next to last instruction.
1720 BasicBlock::iterator I = EndInst;
1721 Instruction *Inst = --I;
1722 if (!Inst->use_empty())
1723 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
1724 if (isa<LandingPadInst>(Inst)) {
1725 EndInst = Inst;
1726 continue;
1727 }
1728 BB->getInstList().erase(Inst);
1729 ++NumInstRemoved;
1730 }
1731 }
1732
1733 // runOnFunction() - Run the Sparse Conditional Constant Propagation algorithm,
1734 // and return true if the function was modified.
1735 //
runOnFunction(Function & F)1736 bool SCCP::runOnFunction(Function &F) {
1737 DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n");
1738 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1739
1740 // Mark the first block of the function as being executable.
1741 Solver.MarkBlockExecutable(F.begin());
1742
1743 // Mark all arguments to the function as being overdefined.
1744 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;++AI)
1745 Solver.markAnythingOverdefined(AI);
1746
1747 // Solve for constants.
1748 bool ResolvedUndefs = true;
1749 while (ResolvedUndefs) {
1750 Solver.Solve();
1751 DEBUG(dbgs() << "RESOLVING UNDEFs\n");
1752 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1753 }
1754
1755 bool MadeChanges = false;
1756
1757 // If we decided that there are basic blocks that are dead in this function,
1758 // delete their contents now. Note that we cannot actually delete the blocks,
1759 // as we cannot modify the CFG of the function.
1760
1761 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
1762 if (!Solver.isBlockExecutable(BB)) {
1763 DeleteInstructionInBlock(BB);
1764 MadeChanges = true;
1765 continue;
1766 }
1767
1768 // Iterate over all of the instructions in a function, replacing them with
1769 // constants if we have found them to be of constant values.
1770 //
1771 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1772 Instruction *Inst = BI++;
1773 if (Inst->getType()->isVoidTy() || isa<TerminatorInst>(Inst))
1774 continue;
1775
1776 // TODO: Reconstruct structs from their elements.
1777 if (Inst->getType()->isStructTy())
1778 continue;
1779
1780 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1781 if (IV.isOverdefined())
1782 continue;
1783
1784 Constant *Const = IV.isConstant()
1785 ? IV.getConstant() : UndefValue::get(Inst->getType());
1786 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst);
1787
1788 // Replaces all of the uses of a variable with uses of the constant.
1789 Inst->replaceAllUsesWith(Const);
1790
1791 // Delete the instruction.
1792 Inst->eraseFromParent();
1793
1794 // Hey, we just changed something!
1795 MadeChanges = true;
1796 ++NumInstRemoved;
1797 }
1798 }
1799
1800 return MadeChanges;
1801 }
1802
1803 namespace {
1804 //===--------------------------------------------------------------------===//
1805 //
1806 /// IPSCCP Class - This class implements interprocedural Sparse Conditional
1807 /// Constant Propagation.
1808 ///
1809 struct IPSCCP : public ModulePass {
1810 static char ID;
IPSCCP__anon967d13930411::IPSCCP1811 IPSCCP() : ModulePass(ID) {
1812 initializeIPSCCPPass(*PassRegistry::getPassRegistry());
1813 }
1814 bool runOnModule(Module &M);
1815 };
1816 } // end anonymous namespace
1817
1818 char IPSCCP::ID = 0;
1819 INITIALIZE_PASS(IPSCCP, "ipsccp",
1820 "Interprocedural Sparse Conditional Constant Propagation",
1821 false, false)
1822
1823 // createIPSCCPPass - This is the public interface to this file.
createIPSCCPPass()1824 ModulePass *llvm::createIPSCCPPass() {
1825 return new IPSCCP();
1826 }
1827
1828
AddressIsTaken(const GlobalValue * GV)1829 static bool AddressIsTaken(const GlobalValue *GV) {
1830 // Delete any dead constantexpr klingons.
1831 GV->removeDeadConstantUsers();
1832
1833 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
1834 UI != E; ++UI) {
1835 const User *U = *UI;
1836 if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
1837 if (SI->getOperand(0) == GV || SI->isVolatile())
1838 return true; // Storing addr of GV.
1839 } else if (isa<InvokeInst>(U) || isa<CallInst>(U)) {
1840 // Make sure we are calling the function, not passing the address.
1841 ImmutableCallSite CS(cast<Instruction>(U));
1842 if (!CS.isCallee(UI))
1843 return true;
1844 } else if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
1845 if (LI->isVolatile())
1846 return true;
1847 } else if (isa<BlockAddress>(U)) {
1848 // blockaddress doesn't take the address of the function, it takes addr
1849 // of label.
1850 } else {
1851 return true;
1852 }
1853 }
1854 return false;
1855 }
1856
runOnModule(Module & M)1857 bool IPSCCP::runOnModule(Module &M) {
1858 SCCPSolver Solver(getAnalysisIfAvailable<TargetData>());
1859
1860 // AddressTakenFunctions - This set keeps track of the address-taken functions
1861 // that are in the input. As IPSCCP runs through and simplifies code,
1862 // functions that were address taken can end up losing their
1863 // address-taken-ness. Because of this, we keep track of their addresses from
1864 // the first pass so we can use them for the later simplification pass.
1865 SmallPtrSet<Function*, 32> AddressTakenFunctions;
1866
1867 // Loop over all functions, marking arguments to those with their addresses
1868 // taken or that are external as overdefined.
1869 //
1870 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1871 if (F->isDeclaration())
1872 continue;
1873
1874 // If this is a strong or ODR definition of this function, then we can
1875 // propagate information about its result into callsites of it.
1876 if (!F->mayBeOverridden())
1877 Solver.AddTrackedFunction(F);
1878
1879 // If this function only has direct calls that we can see, we can track its
1880 // arguments and return value aggressively, and can assume it is not called
1881 // unless we see evidence to the contrary.
1882 if (F->hasLocalLinkage()) {
1883 if (AddressIsTaken(F))
1884 AddressTakenFunctions.insert(F);
1885 else {
1886 Solver.AddArgumentTrackedFunction(F);
1887 continue;
1888 }
1889 }
1890
1891 // Assume the function is called.
1892 Solver.MarkBlockExecutable(F->begin());
1893
1894 // Assume nothing about the incoming arguments.
1895 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1896 AI != E; ++AI)
1897 Solver.markAnythingOverdefined(AI);
1898 }
1899
1900 // Loop over global variables. We inform the solver about any internal global
1901 // variables that do not have their 'addresses taken'. If they don't have
1902 // their addresses taken, we can propagate constants through them.
1903 for (Module::global_iterator G = M.global_begin(), E = M.global_end();
1904 G != E; ++G)
1905 if (!G->isConstant() && G->hasLocalLinkage() && !AddressIsTaken(G))
1906 Solver.TrackValueOfGlobalVariable(G);
1907
1908 // Solve for constants.
1909 bool ResolvedUndefs = true;
1910 while (ResolvedUndefs) {
1911 Solver.Solve();
1912
1913 DEBUG(dbgs() << "RESOLVING UNDEFS\n");
1914 ResolvedUndefs = false;
1915 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F)
1916 ResolvedUndefs |= Solver.ResolvedUndefsIn(*F);
1917 }
1918
1919 bool MadeChanges = false;
1920
1921 // Iterate over all of the instructions in the module, replacing them with
1922 // constants if we have found them to be of constant values.
1923 //
1924 SmallVector<BasicBlock*, 512> BlocksToErase;
1925
1926 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
1927 if (Solver.isBlockExecutable(F->begin())) {
1928 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1929 AI != E; ++AI) {
1930 if (AI->use_empty() || AI->getType()->isStructTy()) continue;
1931
1932 // TODO: Could use getStructLatticeValueFor to find out if the entire
1933 // result is a constant and replace it entirely if so.
1934
1935 LatticeVal IV = Solver.getLatticeValueFor(AI);
1936 if (IV.isOverdefined()) continue;
1937
1938 Constant *CST = IV.isConstant() ?
1939 IV.getConstant() : UndefValue::get(AI->getType());
1940 DEBUG(dbgs() << "*** Arg " << *AI << " = " << *CST <<"\n");
1941
1942 // Replaces all of the uses of a variable with uses of the
1943 // constant.
1944 AI->replaceAllUsesWith(CST);
1945 ++IPNumArgsElimed;
1946 }
1947 }
1948
1949 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
1950 if (!Solver.isBlockExecutable(BB)) {
1951 DeleteInstructionInBlock(BB);
1952 MadeChanges = true;
1953
1954 TerminatorInst *TI = BB->getTerminator();
1955 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1956 BasicBlock *Succ = TI->getSuccessor(i);
1957 if (!Succ->empty() && isa<PHINode>(Succ->begin()))
1958 TI->getSuccessor(i)->removePredecessor(BB);
1959 }
1960 if (!TI->use_empty())
1961 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1962 TI->eraseFromParent();
1963
1964 if (&*BB != &F->front())
1965 BlocksToErase.push_back(BB);
1966 else
1967 new UnreachableInst(M.getContext(), BB);
1968 continue;
1969 }
1970
1971 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
1972 Instruction *Inst = BI++;
1973 if (Inst->getType()->isVoidTy() || Inst->getType()->isStructTy())
1974 continue;
1975
1976 // TODO: Could use getStructLatticeValueFor to find out if the entire
1977 // result is a constant and replace it entirely if so.
1978
1979 LatticeVal IV = Solver.getLatticeValueFor(Inst);
1980 if (IV.isOverdefined())
1981 continue;
1982
1983 Constant *Const = IV.isConstant()
1984 ? IV.getConstant() : UndefValue::get(Inst->getType());
1985 DEBUG(dbgs() << " Constant: " << *Const << " = " << *Inst);
1986
1987 // Replaces all of the uses of a variable with uses of the
1988 // constant.
1989 Inst->replaceAllUsesWith(Const);
1990
1991 // Delete the instruction.
1992 if (!isa<CallInst>(Inst) && !isa<TerminatorInst>(Inst))
1993 Inst->eraseFromParent();
1994
1995 // Hey, we just changed something!
1996 MadeChanges = true;
1997 ++IPNumInstRemoved;
1998 }
1999 }
2000
2001 // Now that all instructions in the function are constant folded, erase dead
2002 // blocks, because we can now use ConstantFoldTerminator to get rid of
2003 // in-edges.
2004 for (unsigned i = 0, e = BlocksToErase.size(); i != e; ++i) {
2005 // If there are any PHI nodes in this successor, drop entries for BB now.
2006 BasicBlock *DeadBB = BlocksToErase[i];
2007 for (Value::use_iterator UI = DeadBB->use_begin(), UE = DeadBB->use_end();
2008 UI != UE; ) {
2009 // Grab the user and then increment the iterator early, as the user
2010 // will be deleted. Step past all adjacent uses from the same user.
2011 Instruction *I = dyn_cast<Instruction>(*UI);
2012 do { ++UI; } while (UI != UE && *UI == I);
2013
2014 // Ignore blockaddress users; BasicBlock's dtor will handle them.
2015 if (!I) continue;
2016
2017 bool Folded = ConstantFoldTerminator(I->getParent());
2018 if (!Folded) {
2019 // The constant folder may not have been able to fold the terminator
2020 // if this is a branch or switch on undef. Fold it manually as a
2021 // branch to the first successor.
2022 #ifndef NDEBUG
2023 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2024 assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
2025 "Branch should be foldable!");
2026 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2027 assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
2028 } else {
2029 llvm_unreachable("Didn't fold away reference to block!");
2030 }
2031 #endif
2032
2033 // Make this an uncond branch to the first successor.
2034 TerminatorInst *TI = I->getParent()->getTerminator();
2035 BranchInst::Create(TI->getSuccessor(0), TI);
2036
2037 // Remove entries in successor phi nodes to remove edges.
2038 for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
2039 TI->getSuccessor(i)->removePredecessor(TI->getParent());
2040
2041 // Remove the old terminator.
2042 TI->eraseFromParent();
2043 }
2044 }
2045
2046 // Finally, delete the basic block.
2047 F->getBasicBlockList().erase(DeadBB);
2048 }
2049 BlocksToErase.clear();
2050 }
2051
2052 // If we inferred constant or undef return values for a function, we replaced
2053 // all call uses with the inferred value. This means we don't need to bother
2054 // actually returning anything from the function. Replace all return
2055 // instructions with return undef.
2056 //
2057 // Do this in two stages: first identify the functions we should process, then
2058 // actually zap their returns. This is important because we can only do this
2059 // if the address of the function isn't taken. In cases where a return is the
2060 // last use of a function, the order of processing functions would affect
2061 // whether other functions are optimizable.
2062 SmallVector<ReturnInst*, 8> ReturnsToZap;
2063
2064 // TODO: Process multiple value ret instructions also.
2065 const DenseMap<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
2066 for (DenseMap<Function*, LatticeVal>::const_iterator I = RV.begin(),
2067 E = RV.end(); I != E; ++I) {
2068 Function *F = I->first;
2069 if (I->second.isOverdefined() || F->getReturnType()->isVoidTy())
2070 continue;
2071
2072 // We can only do this if we know that nothing else can call the function.
2073 if (!F->hasLocalLinkage() || AddressTakenFunctions.count(F))
2074 continue;
2075
2076 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
2077 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()))
2078 if (!isa<UndefValue>(RI->getOperand(0)))
2079 ReturnsToZap.push_back(RI);
2080 }
2081
2082 // Zap all returns which we've identified as zap to change.
2083 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) {
2084 Function *F = ReturnsToZap[i]->getParent()->getParent();
2085 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType()));
2086 }
2087
2088 // If we inferred constant or undef values for globals variables, we can delete
2089 // the global and any stores that remain to it.
2090 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
2091 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
2092 E = TG.end(); I != E; ++I) {
2093 GlobalVariable *GV = I->first;
2094 assert(!I->second.isOverdefined() &&
2095 "Overdefined values should have been taken out of the map!");
2096 DEBUG(dbgs() << "Found that GV '" << GV->getName() << "' is constant!\n");
2097 while (!GV->use_empty()) {
2098 StoreInst *SI = cast<StoreInst>(GV->use_back());
2099 SI->eraseFromParent();
2100 }
2101 M.getGlobalList().erase(GV);
2102 ++IPNumGlobalConst;
2103 }
2104
2105 return MadeChanges;
2106 }
2107