1 //===- SCCP.cpp - Sparse Conditional Constant Propagation -----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements sparse conditional constant propagation and merging:
10 //
11 // Specifically, this:
12 // * Assumes values are constant unless proven otherwise
13 // * Assumes BasicBlocks are dead unless proven otherwise
14 // * Proves values to be constant, and replaces them with constants
15 // * Proves conditional branches to be unconditional
16 //
17 //===----------------------------------------------------------------------===//
18
19 #include "llvm/Transforms/Scalar/SCCP.h"
20 #include "llvm/ADT/ArrayRef.h"
21 #include "llvm/ADT/DenseMap.h"
22 #include "llvm/ADT/DenseSet.h"
23 #include "llvm/ADT/MapVector.h"
24 #include "llvm/ADT/PointerIntPair.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/Analysis/ConstantFolding.h"
30 #include "llvm/Analysis/GlobalsModRef.h"
31 #include "llvm/Analysis/TargetLibraryInfo.h"
32 #include "llvm/Analysis/ValueLattice.h"
33 #include "llvm/Analysis/ValueLatticeUtils.h"
34 #include "llvm/IR/BasicBlock.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/InstVisitor.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/Module.h"
47 #include "llvm/IR/PassManager.h"
48 #include "llvm/IR/Type.h"
49 #include "llvm/IR/User.h"
50 #include "llvm/IR/Value.h"
51 #include "llvm/InitializePasses.h"
52 #include "llvm/Pass.h"
53 #include "llvm/Support/Casting.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/ErrorHandling.h"
56 #include "llvm/Support/raw_ostream.h"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 #include "llvm/Transforms/Utils/PredicateInfo.h"
60 #include <cassert>
61 #include <utility>
62 #include <vector>
63
64 using namespace llvm;
65
66 #define DEBUG_TYPE "sccp"
67
68 STATISTIC(NumInstRemoved, "Number of instructions removed");
69 STATISTIC(NumDeadBlocks , "Number of basic blocks unreachable");
70
71 STATISTIC(IPNumInstRemoved, "Number of instructions removed by IPSCCP");
72 STATISTIC(IPNumArgsElimed ,"Number of arguments constant propagated by IPSCCP");
73 STATISTIC(IPNumGlobalConst, "Number of globals found to be constant by IPSCCP");
74
75 namespace {
76
77 /// LatticeVal class - This class represents the different lattice values that
78 /// an LLVM value may occupy. It is a simple class with value semantics.
79 ///
80 class LatticeVal {
81 enum LatticeValueTy {
82 /// unknown - This LLVM Value has no known value yet.
83 unknown,
84
85 /// constant - This LLVM Value has a specific constant value.
86 constant,
87
88 /// forcedconstant - This LLVM Value was thought to be undef until
89 /// ResolvedUndefsIn. This is treated just like 'constant', but if merged
90 /// with another (different) constant, it goes to overdefined, instead of
91 /// asserting.
92 forcedconstant,
93
94 /// overdefined - This instruction is not known to be constant, and we know
95 /// it has a value.
96 overdefined
97 };
98
99 /// Val: This stores the current lattice value along with the Constant* for
100 /// the constant if this is a 'constant' or 'forcedconstant' value.
101 PointerIntPair<Constant *, 2, LatticeValueTy> Val;
102
getLatticeValue() const103 LatticeValueTy getLatticeValue() const {
104 return Val.getInt();
105 }
106
107 public:
LatticeVal()108 LatticeVal() : Val(nullptr, unknown) {}
109
isUnknown() const110 bool isUnknown() const { return getLatticeValue() == unknown; }
111
isConstant() const112 bool isConstant() const {
113 return getLatticeValue() == constant || getLatticeValue() == forcedconstant;
114 }
115
isOverdefined() const116 bool isOverdefined() const { return getLatticeValue() == overdefined; }
117
getConstant() const118 Constant *getConstant() const {
119 assert(isConstant() && "Cannot get the constant of a non-constant!");
120 return Val.getPointer();
121 }
122
123 /// markOverdefined - Return true if this is a change in status.
markOverdefined()124 bool markOverdefined() {
125 if (isOverdefined())
126 return false;
127
128 Val.setInt(overdefined);
129 return true;
130 }
131
132 /// markConstant - Return true if this is a change in status.
markConstant(Constant * V)133 bool markConstant(Constant *V) {
134 if (getLatticeValue() == constant) { // Constant but not forcedconstant.
135 assert(getConstant() == V && "Marking constant with different value");
136 return false;
137 }
138
139 if (isUnknown()) {
140 Val.setInt(constant);
141 assert(V && "Marking constant with NULL");
142 Val.setPointer(V);
143 } else {
144 assert(getLatticeValue() == forcedconstant &&
145 "Cannot move from overdefined to constant!");
146 // Stay at forcedconstant if the constant is the same.
147 if (V == getConstant()) return false;
148
149 // Otherwise, we go to overdefined. Assumptions made based on the
150 // forced value are possibly wrong. Assuming this is another constant
151 // could expose a contradiction.
152 Val.setInt(overdefined);
153 }
154 return true;
155 }
156
157 /// getConstantInt - If this is a constant with a ConstantInt value, return it
158 /// otherwise return null.
getConstantInt() const159 ConstantInt *getConstantInt() const {
160 if (isConstant())
161 return dyn_cast<ConstantInt>(getConstant());
162 return nullptr;
163 }
164
165 /// getBlockAddress - If this is a constant with a BlockAddress value, return
166 /// it, otherwise return null.
getBlockAddress() const167 BlockAddress *getBlockAddress() const {
168 if (isConstant())
169 return dyn_cast<BlockAddress>(getConstant());
170 return nullptr;
171 }
172
markForcedConstant(Constant * V)173 void markForcedConstant(Constant *V) {
174 assert(isUnknown() && "Can't force a defined value!");
175 Val.setInt(forcedconstant);
176 Val.setPointer(V);
177 }
178
toValueLattice() const179 ValueLatticeElement toValueLattice() const {
180 if (isOverdefined())
181 return ValueLatticeElement::getOverdefined();
182 if (isConstant())
183 return ValueLatticeElement::get(getConstant());
184 return ValueLatticeElement();
185 }
186 };
187
188 //===----------------------------------------------------------------------===//
189 //
190 /// SCCPSolver - This class is a general purpose solver for Sparse Conditional
191 /// Constant Propagation.
192 ///
193 class SCCPSolver : public InstVisitor<SCCPSolver> {
194 const DataLayout &DL;
195 std::function<const TargetLibraryInfo &(Function &)> GetTLI;
196 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable.
197 DenseMap<Value *, LatticeVal> ValueState; // The state each value is in.
198 // The state each parameter is in.
199 DenseMap<Value *, ValueLatticeElement> ParamState;
200
201 /// StructValueState - This maintains ValueState for values that have
202 /// StructType, for example for formal arguments, calls, insertelement, etc.
203 DenseMap<std::pair<Value *, unsigned>, LatticeVal> StructValueState;
204
205 /// GlobalValue - If we are tracking any values for the contents of a global
206 /// variable, we keep a mapping from the constant accessor to the element of
207 /// the global, to the currently known value. If the value becomes
208 /// overdefined, it's entry is simply removed from this map.
209 DenseMap<GlobalVariable *, LatticeVal> TrackedGlobals;
210
211 /// TrackedRetVals - If we are tracking arguments into and the return
212 /// value out of a function, it will have an entry in this map, indicating
213 /// what the known return value for the function is.
214 MapVector<Function *, LatticeVal> TrackedRetVals;
215
216 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
217 /// that return multiple values.
218 MapVector<std::pair<Function *, unsigned>, LatticeVal> TrackedMultipleRetVals;
219
220 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is
221 /// represented here for efficient lookup.
222 SmallPtrSet<Function *, 16> MRVFunctionsTracked;
223
224 /// MustTailFunctions - Each function here is a callee of non-removable
225 /// musttail call site.
226 SmallPtrSet<Function *, 16> MustTailCallees;
227
228 /// TrackingIncomingArguments - This is the set of functions for whose
229 /// arguments we make optimistic assumptions about and try to prove as
230 /// constants.
231 SmallPtrSet<Function *, 16> TrackingIncomingArguments;
232
233 /// The reason for two worklists is that overdefined is the lowest state
234 /// on the lattice, and moving things to overdefined as fast as possible
235 /// makes SCCP converge much faster.
236 ///
237 /// By having a separate worklist, we accomplish this because everything
238 /// possibly overdefined will become overdefined at the soonest possible
239 /// point.
240 SmallVector<Value *, 64> OverdefinedInstWorkList;
241 SmallVector<Value *, 64> InstWorkList;
242
243 // The BasicBlock work list
244 SmallVector<BasicBlock *, 64> BBWorkList;
245
246 /// KnownFeasibleEdges - Entries in this set are edges which have already had
247 /// PHI nodes retriggered.
248 using Edge = std::pair<BasicBlock *, BasicBlock *>;
249 DenseSet<Edge> KnownFeasibleEdges;
250
251 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults;
252 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers;
253
254 public:
addAnalysis(Function & F,AnalysisResultsForFn A)255 void addAnalysis(Function &F, AnalysisResultsForFn A) {
256 AnalysisResults.insert({&F, std::move(A)});
257 }
258
getPredicateInfoFor(Instruction * I)259 const PredicateBase *getPredicateInfoFor(Instruction *I) {
260 auto A = AnalysisResults.find(I->getParent()->getParent());
261 if (A == AnalysisResults.end())
262 return nullptr;
263 return A->second.PredInfo->getPredicateInfoFor(I);
264 }
265
getDTU(Function & F)266 DomTreeUpdater getDTU(Function &F) {
267 auto A = AnalysisResults.find(&F);
268 assert(A != AnalysisResults.end() && "Need analysis results for function.");
269 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy};
270 }
271
SCCPSolver(const DataLayout & DL,std::function<const TargetLibraryInfo & (Function &)> GetTLI)272 SCCPSolver(const DataLayout &DL,
273 std::function<const TargetLibraryInfo &(Function &)> GetTLI)
274 : DL(DL), GetTLI(std::move(GetTLI)) {}
275
276 /// MarkBlockExecutable - This method can be used by clients to mark all of
277 /// the blocks that are known to be intrinsically live in the processed unit.
278 ///
279 /// This returns true if the block was not considered live before.
MarkBlockExecutable(BasicBlock * BB)280 bool MarkBlockExecutable(BasicBlock *BB) {
281 if (!BBExecutable.insert(BB).second)
282 return false;
283 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n');
284 BBWorkList.push_back(BB); // Add the block to the work list!
285 return true;
286 }
287
288 /// TrackValueOfGlobalVariable - Clients can use this method to
289 /// inform the SCCPSolver that it should track loads and stores to the
290 /// specified global variable if it can. This is only legal to call if
291 /// performing Interprocedural SCCP.
TrackValueOfGlobalVariable(GlobalVariable * GV)292 void TrackValueOfGlobalVariable(GlobalVariable *GV) {
293 // We only track the contents of scalar globals.
294 if (GV->getValueType()->isSingleValueType()) {
295 LatticeVal &IV = TrackedGlobals[GV];
296 if (!isa<UndefValue>(GV->getInitializer()))
297 IV.markConstant(GV->getInitializer());
298 }
299 }
300
301 /// AddTrackedFunction - If the SCCP solver is supposed to track calls into
302 /// and out of the specified function (which cannot have its address taken),
303 /// this method must be called.
AddTrackedFunction(Function * F)304 void AddTrackedFunction(Function *F) {
305 // Add an entry, F -> undef.
306 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) {
307 MRVFunctionsTracked.insert(F);
308 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
309 TrackedMultipleRetVals.insert(std::make_pair(std::make_pair(F, i),
310 LatticeVal()));
311 } else
312 TrackedRetVals.insert(std::make_pair(F, LatticeVal()));
313 }
314
315 /// AddMustTailCallee - If the SCCP solver finds that this function is called
316 /// from non-removable musttail call site.
AddMustTailCallee(Function * F)317 void AddMustTailCallee(Function *F) {
318 MustTailCallees.insert(F);
319 }
320
321 /// Returns true if the given function is called from non-removable musttail
322 /// call site.
isMustTailCallee(Function * F)323 bool isMustTailCallee(Function *F) {
324 return MustTailCallees.count(F);
325 }
326
AddArgumentTrackedFunction(Function * F)327 void AddArgumentTrackedFunction(Function *F) {
328 TrackingIncomingArguments.insert(F);
329 }
330
331 /// Returns true if the given function is in the solver's set of
332 /// argument-tracked functions.
isArgumentTrackedFunction(Function * F)333 bool isArgumentTrackedFunction(Function *F) {
334 return TrackingIncomingArguments.count(F);
335 }
336
337 /// Solve - Solve for constants and executable blocks.
338 void Solve();
339
340 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
341 /// that branches on undef values cannot reach any of their successors.
342 /// However, this is not a safe assumption. After we solve dataflow, this
343 /// method should be use to handle this. If this returns true, the solver
344 /// should be rerun.
345 bool ResolvedUndefsIn(Function &F);
346
isBlockExecutable(BasicBlock * BB) const347 bool isBlockExecutable(BasicBlock *BB) const {
348 return BBExecutable.count(BB);
349 }
350
351 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
352 // block to the 'To' basic block is currently feasible.
353 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To);
354
getStructLatticeValueFor(Value * V) const355 std::vector<LatticeVal> getStructLatticeValueFor(Value *V) const {
356 std::vector<LatticeVal> StructValues;
357 auto *STy = dyn_cast<StructType>(V->getType());
358 assert(STy && "getStructLatticeValueFor() can be called only on structs");
359 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
360 auto I = StructValueState.find(std::make_pair(V, i));
361 assert(I != StructValueState.end() && "Value not in valuemap!");
362 StructValues.push_back(I->second);
363 }
364 return StructValues;
365 }
366
getLatticeValueFor(Value * V) const367 const LatticeVal &getLatticeValueFor(Value *V) const {
368 assert(!V->getType()->isStructTy() &&
369 "Should use getStructLatticeValueFor");
370 DenseMap<Value *, LatticeVal>::const_iterator I = ValueState.find(V);
371 assert(I != ValueState.end() &&
372 "V not found in ValueState nor Paramstate map!");
373 return I->second;
374 }
375
376 /// getTrackedRetVals - Get the inferred return value map.
getTrackedRetVals()377 const MapVector<Function*, LatticeVal> &getTrackedRetVals() {
378 return TrackedRetVals;
379 }
380
381 /// getTrackedGlobals - Get and return the set of inferred initializers for
382 /// global variables.
getTrackedGlobals()383 const DenseMap<GlobalVariable*, LatticeVal> &getTrackedGlobals() {
384 return TrackedGlobals;
385 }
386
387 /// getMRVFunctionsTracked - Get the set of functions which return multiple
388 /// values tracked by the pass.
getMRVFunctionsTracked()389 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() {
390 return MRVFunctionsTracked;
391 }
392
393 /// getMustTailCallees - Get the set of functions which are called
394 /// from non-removable musttail call sites.
getMustTailCallees()395 const SmallPtrSet<Function *, 16> getMustTailCallees() {
396 return MustTailCallees;
397 }
398
399 /// markOverdefined - Mark the specified value overdefined. This
400 /// works with both scalars and structs.
markOverdefined(Value * V)401 void markOverdefined(Value *V) {
402 if (auto *STy = dyn_cast<StructType>(V->getType()))
403 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
404 markOverdefined(getStructValueState(V, i), V);
405 else
406 markOverdefined(ValueState[V], V);
407 }
408
409 // isStructLatticeConstant - Return true if all the lattice values
410 // corresponding to elements of the structure are not overdefined,
411 // false otherwise.
isStructLatticeConstant(Function * F,StructType * STy)412 bool isStructLatticeConstant(Function *F, StructType *STy) {
413 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
414 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i));
415 assert(It != TrackedMultipleRetVals.end());
416 LatticeVal LV = It->second;
417 if (LV.isOverdefined())
418 return false;
419 }
420 return true;
421 }
422
423 private:
424 // pushToWorkList - Helper for markConstant/markForcedConstant/markOverdefined
pushToWorkList(LatticeVal & IV,Value * V)425 void pushToWorkList(LatticeVal &IV, Value *V) {
426 if (IV.isOverdefined())
427 return OverdefinedInstWorkList.push_back(V);
428 InstWorkList.push_back(V);
429 }
430
431 // markConstant - Make a value be marked as "constant". If the value
432 // is not already a constant, add it to the instruction work list so that
433 // the users of the instruction are updated later.
markConstant(LatticeVal & IV,Value * V,Constant * C)434 bool markConstant(LatticeVal &IV, Value *V, Constant *C) {
435 if (!IV.markConstant(C)) return false;
436 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n');
437 pushToWorkList(IV, V);
438 return true;
439 }
440
markConstant(Value * V,Constant * C)441 bool markConstant(Value *V, Constant *C) {
442 assert(!V->getType()->isStructTy() && "structs should use mergeInValue");
443 return markConstant(ValueState[V], V, C);
444 }
445
markForcedConstant(Value * V,Constant * C)446 void markForcedConstant(Value *V, Constant *C) {
447 assert(!V->getType()->isStructTy() && "structs should use mergeInValue");
448 LatticeVal &IV = ValueState[V];
449 IV.markForcedConstant(C);
450 LLVM_DEBUG(dbgs() << "markForcedConstant: " << *C << ": " << *V << '\n');
451 pushToWorkList(IV, V);
452 }
453
454 // markOverdefined - Make a value be marked as "overdefined". If the
455 // value is not already overdefined, add it to the overdefined instruction
456 // work list so that the users of the instruction are updated later.
markOverdefined(LatticeVal & IV,Value * V)457 bool markOverdefined(LatticeVal &IV, Value *V) {
458 if (!IV.markOverdefined()) return false;
459
460 LLVM_DEBUG(dbgs() << "markOverdefined: ";
461 if (auto *F = dyn_cast<Function>(V)) dbgs()
462 << "Function '" << F->getName() << "'\n";
463 else dbgs() << *V << '\n');
464 // Only instructions go on the work list
465 pushToWorkList(IV, V);
466 return true;
467 }
468
mergeInValue(LatticeVal & IV,Value * V,LatticeVal MergeWithV)469 bool mergeInValue(LatticeVal &IV, Value *V, LatticeVal MergeWithV) {
470 if (IV.isOverdefined() || MergeWithV.isUnknown())
471 return false; // Noop.
472 if (MergeWithV.isOverdefined())
473 return markOverdefined(IV, V);
474 if (IV.isUnknown())
475 return markConstant(IV, V, MergeWithV.getConstant());
476 if (IV.getConstant() != MergeWithV.getConstant())
477 return markOverdefined(IV, V);
478 return false;
479 }
480
mergeInValue(Value * V,LatticeVal MergeWithV)481 bool mergeInValue(Value *V, LatticeVal MergeWithV) {
482 assert(!V->getType()->isStructTy() &&
483 "non-structs should use markConstant");
484 return mergeInValue(ValueState[V], V, MergeWithV);
485 }
486
487 /// getValueState - Return the LatticeVal object that corresponds to the
488 /// value. This function handles the case when the value hasn't been seen yet
489 /// by properly seeding constants etc.
getValueState(Value * V)490 LatticeVal &getValueState(Value *V) {
491 assert(!V->getType()->isStructTy() && "Should use getStructValueState");
492
493 std::pair<DenseMap<Value*, LatticeVal>::iterator, bool> I =
494 ValueState.insert(std::make_pair(V, LatticeVal()));
495 LatticeVal &LV = I.first->second;
496
497 if (!I.second)
498 return LV; // Common case, already in the map.
499
500 if (auto *C = dyn_cast<Constant>(V)) {
501 // Undef values remain unknown.
502 if (!isa<UndefValue>(V))
503 LV.markConstant(C); // Constants are constant
504 }
505
506 // All others are underdefined by default.
507 return LV;
508 }
509
getParamState(Value * V)510 ValueLatticeElement &getParamState(Value *V) {
511 assert(!V->getType()->isStructTy() && "Should use getStructValueState");
512
513 std::pair<DenseMap<Value*, ValueLatticeElement>::iterator, bool>
514 PI = ParamState.insert(std::make_pair(V, ValueLatticeElement()));
515 ValueLatticeElement &LV = PI.first->second;
516 if (PI.second)
517 LV = getValueState(V).toValueLattice();
518
519 return LV;
520 }
521
522 /// getStructValueState - Return the LatticeVal object that corresponds to the
523 /// value/field pair. This function handles the case when the value hasn't
524 /// been seen yet by properly seeding constants etc.
getStructValueState(Value * V,unsigned i)525 LatticeVal &getStructValueState(Value *V, unsigned i) {
526 assert(V->getType()->isStructTy() && "Should use getValueState");
527 assert(i < cast<StructType>(V->getType())->getNumElements() &&
528 "Invalid element #");
529
530 std::pair<DenseMap<std::pair<Value*, unsigned>, LatticeVal>::iterator,
531 bool> I = StructValueState.insert(
532 std::make_pair(std::make_pair(V, i), LatticeVal()));
533 LatticeVal &LV = I.first->second;
534
535 if (!I.second)
536 return LV; // Common case, already in the map.
537
538 if (auto *C = dyn_cast<Constant>(V)) {
539 Constant *Elt = C->getAggregateElement(i);
540
541 if (!Elt)
542 LV.markOverdefined(); // Unknown sort of constant.
543 else if (isa<UndefValue>(Elt))
544 ; // Undef values remain unknown.
545 else
546 LV.markConstant(Elt); // Constants are constant.
547 }
548
549 // All others are underdefined by default.
550 return LV;
551 }
552
553 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
554 /// work list if it is not already executable.
markEdgeExecutable(BasicBlock * Source,BasicBlock * Dest)555 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
556 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
557 return false; // This edge is already known to be executable!
558
559 if (!MarkBlockExecutable(Dest)) {
560 // If the destination is already executable, we just made an *edge*
561 // feasible that wasn't before. Revisit the PHI nodes in the block
562 // because they have potentially new operands.
563 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName()
564 << " -> " << Dest->getName() << '\n');
565
566 for (PHINode &PN : Dest->phis())
567 visitPHINode(PN);
568 }
569 return true;
570 }
571
572 // getFeasibleSuccessors - Return a vector of booleans to indicate which
573 // successors are reachable from a given terminator instruction.
574 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs);
575
576 // OperandChangedState - This method is invoked on all of the users of an
577 // instruction that was just changed state somehow. Based on this
578 // information, we need to update the specified user of this instruction.
OperandChangedState(Instruction * I)579 void OperandChangedState(Instruction *I) {
580 if (BBExecutable.count(I->getParent())) // Inst is executable?
581 visit(*I);
582 }
583
584 // Add U as additional user of V.
addAdditionalUser(Value * V,User * U)585 void addAdditionalUser(Value *V, User *U) {
586 auto Iter = AdditionalUsers.insert({V, {}});
587 Iter.first->second.insert(U);
588 }
589
590 // Mark I's users as changed, including AdditionalUsers.
markUsersAsChanged(Value * I)591 void markUsersAsChanged(Value *I) {
592 for (User *U : I->users())
593 if (auto *UI = dyn_cast<Instruction>(U))
594 OperandChangedState(UI);
595
596 auto Iter = AdditionalUsers.find(I);
597 if (Iter != AdditionalUsers.end()) {
598 for (User *U : Iter->second)
599 if (auto *UI = dyn_cast<Instruction>(U))
600 OperandChangedState(UI);
601 }
602 }
603
604 private:
605 friend class InstVisitor<SCCPSolver>;
606
607 // visit implementations - Something changed in this instruction. Either an
608 // operand made a transition, or the instruction is newly executable. Change
609 // the value type of I to reflect these changes if appropriate.
610 void visitPHINode(PHINode &I);
611
612 // Terminators
613
614 void visitReturnInst(ReturnInst &I);
615 void visitTerminator(Instruction &TI);
616
617 void visitCastInst(CastInst &I);
618 void visitSelectInst(SelectInst &I);
619 void visitUnaryOperator(Instruction &I);
620 void visitBinaryOperator(Instruction &I);
621 void visitCmpInst(CmpInst &I);
622 void visitExtractValueInst(ExtractValueInst &EVI);
623 void visitInsertValueInst(InsertValueInst &IVI);
624
visitCatchSwitchInst(CatchSwitchInst & CPI)625 void visitCatchSwitchInst(CatchSwitchInst &CPI) {
626 markOverdefined(&CPI);
627 visitTerminator(CPI);
628 }
629
630 // Instructions that cannot be folded away.
631
632 void visitStoreInst (StoreInst &I);
633 void visitLoadInst (LoadInst &I);
634 void visitGetElementPtrInst(GetElementPtrInst &I);
635
visitCallInst(CallInst & I)636 void visitCallInst (CallInst &I) {
637 visitCallSite(&I);
638 }
639
visitInvokeInst(InvokeInst & II)640 void visitInvokeInst (InvokeInst &II) {
641 visitCallSite(&II);
642 visitTerminator(II);
643 }
644
visitCallBrInst(CallBrInst & CBI)645 void visitCallBrInst (CallBrInst &CBI) {
646 visitCallSite(&CBI);
647 visitTerminator(CBI);
648 }
649
650 void visitCallSite (CallSite CS);
visitResumeInst(ResumeInst & I)651 void visitResumeInst (ResumeInst &I) { /*returns void*/ }
visitUnreachableInst(UnreachableInst & I)652 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/ }
visitFenceInst(FenceInst & I)653 void visitFenceInst (FenceInst &I) { /*returns void*/ }
654
visitInstruction(Instruction & I)655 void visitInstruction(Instruction &I) {
656 // All the instructions we don't do any special handling for just
657 // go to overdefined.
658 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n');
659 markOverdefined(&I);
660 }
661 };
662
663 } // end anonymous namespace
664
665 // getFeasibleSuccessors - Return a vector of booleans to indicate which
666 // successors are reachable from a given terminator instruction.
getFeasibleSuccessors(Instruction & TI,SmallVectorImpl<bool> & Succs)667 void SCCPSolver::getFeasibleSuccessors(Instruction &TI,
668 SmallVectorImpl<bool> &Succs) {
669 Succs.resize(TI.getNumSuccessors());
670 if (auto *BI = dyn_cast<BranchInst>(&TI)) {
671 if (BI->isUnconditional()) {
672 Succs[0] = true;
673 return;
674 }
675
676 LatticeVal BCValue = getValueState(BI->getCondition());
677 ConstantInt *CI = BCValue.getConstantInt();
678 if (!CI) {
679 // Overdefined condition variables, and branches on unfoldable constant
680 // conditions, mean the branch could go either way.
681 if (!BCValue.isUnknown())
682 Succs[0] = Succs[1] = true;
683 return;
684 }
685
686 // Constant condition variables mean the branch can only go a single way.
687 Succs[CI->isZero()] = true;
688 return;
689 }
690
691 // Unwinding instructions successors are always executable.
692 if (TI.isExceptionalTerminator()) {
693 Succs.assign(TI.getNumSuccessors(), true);
694 return;
695 }
696
697 if (auto *SI = dyn_cast<SwitchInst>(&TI)) {
698 if (!SI->getNumCases()) {
699 Succs[0] = true;
700 return;
701 }
702 LatticeVal SCValue = getValueState(SI->getCondition());
703 ConstantInt *CI = SCValue.getConstantInt();
704
705 if (!CI) { // Overdefined or unknown condition?
706 // All destinations are executable!
707 if (!SCValue.isUnknown())
708 Succs.assign(TI.getNumSuccessors(), true);
709 return;
710 }
711
712 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true;
713 return;
714 }
715
716 // In case of indirect branch and its address is a blockaddress, we mark
717 // the target as executable.
718 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) {
719 // Casts are folded by visitCastInst.
720 LatticeVal IBRValue = getValueState(IBR->getAddress());
721 BlockAddress *Addr = IBRValue.getBlockAddress();
722 if (!Addr) { // Overdefined or unknown condition?
723 // All destinations are executable!
724 if (!IBRValue.isUnknown())
725 Succs.assign(TI.getNumSuccessors(), true);
726 return;
727 }
728
729 BasicBlock* T = Addr->getBasicBlock();
730 assert(Addr->getFunction() == T->getParent() &&
731 "Block address of a different function ?");
732 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) {
733 // This is the target.
734 if (IBR->getDestination(i) == T) {
735 Succs[i] = true;
736 return;
737 }
738 }
739
740 // If we didn't find our destination in the IBR successor list, then we
741 // have undefined behavior. Its ok to assume no successor is executable.
742 return;
743 }
744
745 // In case of callbr, we pessimistically assume that all successors are
746 // feasible.
747 if (isa<CallBrInst>(&TI)) {
748 Succs.assign(TI.getNumSuccessors(), true);
749 return;
750 }
751
752 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n');
753 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
754 }
755
756 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
757 // block to the 'To' basic block is currently feasible.
isEdgeFeasible(BasicBlock * From,BasicBlock * To)758 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) {
759 // Check if we've called markEdgeExecutable on the edge yet. (We could
760 // be more aggressive and try to consider edges which haven't been marked
761 // yet, but there isn't any need.)
762 return KnownFeasibleEdges.count(Edge(From, To));
763 }
764
765 // visit Implementations - Something changed in this instruction, either an
766 // operand made a transition, or the instruction is newly executable. Change
767 // the value type of I to reflect these changes if appropriate. This method
768 // makes sure to do the following actions:
769 //
770 // 1. If a phi node merges two constants in, and has conflicting value coming
771 // from different branches, or if the PHI node merges in an overdefined
772 // value, then the PHI node becomes overdefined.
773 // 2. If a phi node merges only constants in, and they all agree on value, the
774 // PHI node becomes a constant value equal to that.
775 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
776 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
777 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
778 // 6. If a conditional branch has a value that is constant, make the selected
779 // destination executable
780 // 7. If a conditional branch has a value that is overdefined, make all
781 // successors executable.
visitPHINode(PHINode & PN)782 void SCCPSolver::visitPHINode(PHINode &PN) {
783 // If this PN returns a struct, just mark the result overdefined.
784 // TODO: We could do a lot better than this if code actually uses this.
785 if (PN.getType()->isStructTy())
786 return (void)markOverdefined(&PN);
787
788 if (getValueState(&PN).isOverdefined())
789 return; // Quick exit
790
791 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
792 // and slow us down a lot. Just mark them overdefined.
793 if (PN.getNumIncomingValues() > 64)
794 return (void)markOverdefined(&PN);
795
796 // Look at all of the executable operands of the PHI node. If any of them
797 // are overdefined, the PHI becomes overdefined as well. If they are all
798 // constant, and they agree with each other, the PHI becomes the identical
799 // constant. If they are constant and don't agree, the PHI is overdefined.
800 // If there are no executable operands, the PHI remains unknown.
801 Constant *OperandVal = nullptr;
802 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
803 LatticeVal IV = getValueState(PN.getIncomingValue(i));
804 if (IV.isUnknown()) continue; // Doesn't influence PHI node.
805
806 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
807 continue;
808
809 if (IV.isOverdefined()) // PHI node becomes overdefined!
810 return (void)markOverdefined(&PN);
811
812 if (!OperandVal) { // Grab the first value.
813 OperandVal = IV.getConstant();
814 continue;
815 }
816
817 // There is already a reachable operand. If we conflict with it,
818 // then the PHI node becomes overdefined. If we agree with it, we
819 // can continue on.
820
821 // Check to see if there are two different constants merging, if so, the PHI
822 // node is overdefined.
823 if (IV.getConstant() != OperandVal)
824 return (void)markOverdefined(&PN);
825 }
826
827 // If we exited the loop, this means that the PHI node only has constant
828 // arguments that agree with each other(and OperandVal is the constant) or
829 // OperandVal is null because there are no defined incoming arguments. If
830 // this is the case, the PHI remains unknown.
831 if (OperandVal)
832 markConstant(&PN, OperandVal); // Acquire operand value
833 }
834
visitReturnInst(ReturnInst & I)835 void SCCPSolver::visitReturnInst(ReturnInst &I) {
836 if (I.getNumOperands() == 0) return; // ret void
837
838 Function *F = I.getParent()->getParent();
839 Value *ResultOp = I.getOperand(0);
840
841 // If we are tracking the return value of this function, merge it in.
842 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) {
843 MapVector<Function*, LatticeVal>::iterator TFRVI =
844 TrackedRetVals.find(F);
845 if (TFRVI != TrackedRetVals.end()) {
846 mergeInValue(TFRVI->second, F, getValueState(ResultOp));
847 return;
848 }
849 }
850
851 // Handle functions that return multiple values.
852 if (!TrackedMultipleRetVals.empty()) {
853 if (auto *STy = dyn_cast<StructType>(ResultOp->getType()))
854 if (MRVFunctionsTracked.count(F))
855 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
856 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F,
857 getStructValueState(ResultOp, i));
858 }
859 }
860
visitTerminator(Instruction & TI)861 void SCCPSolver::visitTerminator(Instruction &TI) {
862 SmallVector<bool, 16> SuccFeasible;
863 getFeasibleSuccessors(TI, SuccFeasible);
864
865 BasicBlock *BB = TI.getParent();
866
867 // Mark all feasible successors executable.
868 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
869 if (SuccFeasible[i])
870 markEdgeExecutable(BB, TI.getSuccessor(i));
871 }
872
visitCastInst(CastInst & I)873 void SCCPSolver::visitCastInst(CastInst &I) {
874 LatticeVal OpSt = getValueState(I.getOperand(0));
875 if (OpSt.isOverdefined()) // Inherit overdefinedness of operand
876 markOverdefined(&I);
877 else if (OpSt.isConstant()) {
878 // Fold the constant as we build.
879 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpSt.getConstant(),
880 I.getType(), DL);
881 if (isa<UndefValue>(C))
882 return;
883 // Propagate constant value
884 markConstant(&I, C);
885 }
886 }
887
visitExtractValueInst(ExtractValueInst & EVI)888 void SCCPSolver::visitExtractValueInst(ExtractValueInst &EVI) {
889 // If this returns a struct, mark all elements over defined, we don't track
890 // structs in structs.
891 if (EVI.getType()->isStructTy())
892 return (void)markOverdefined(&EVI);
893
894 // If this is extracting from more than one level of struct, we don't know.
895 if (EVI.getNumIndices() != 1)
896 return (void)markOverdefined(&EVI);
897
898 Value *AggVal = EVI.getAggregateOperand();
899 if (AggVal->getType()->isStructTy()) {
900 unsigned i = *EVI.idx_begin();
901 LatticeVal EltVal = getStructValueState(AggVal, i);
902 mergeInValue(getValueState(&EVI), &EVI, EltVal);
903 } else {
904 // Otherwise, must be extracting from an array.
905 return (void)markOverdefined(&EVI);
906 }
907 }
908
visitInsertValueInst(InsertValueInst & IVI)909 void SCCPSolver::visitInsertValueInst(InsertValueInst &IVI) {
910 auto *STy = dyn_cast<StructType>(IVI.getType());
911 if (!STy)
912 return (void)markOverdefined(&IVI);
913
914 // If this has more than one index, we can't handle it, drive all results to
915 // undef.
916 if (IVI.getNumIndices() != 1)
917 return (void)markOverdefined(&IVI);
918
919 Value *Aggr = IVI.getAggregateOperand();
920 unsigned Idx = *IVI.idx_begin();
921
922 // Compute the result based on what we're inserting.
923 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
924 // This passes through all values that aren't the inserted element.
925 if (i != Idx) {
926 LatticeVal EltVal = getStructValueState(Aggr, i);
927 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal);
928 continue;
929 }
930
931 Value *Val = IVI.getInsertedValueOperand();
932 if (Val->getType()->isStructTy())
933 // We don't track structs in structs.
934 markOverdefined(getStructValueState(&IVI, i), &IVI);
935 else {
936 LatticeVal InVal = getValueState(Val);
937 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal);
938 }
939 }
940 }
941
visitSelectInst(SelectInst & I)942 void SCCPSolver::visitSelectInst(SelectInst &I) {
943 // If this select returns a struct, just mark the result overdefined.
944 // TODO: We could do a lot better than this if code actually uses this.
945 if (I.getType()->isStructTy())
946 return (void)markOverdefined(&I);
947
948 LatticeVal CondValue = getValueState(I.getCondition());
949 if (CondValue.isUnknown())
950 return;
951
952 if (ConstantInt *CondCB = CondValue.getConstantInt()) {
953 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
954 mergeInValue(&I, getValueState(OpVal));
955 return;
956 }
957
958 // Otherwise, the condition is overdefined or a constant we can't evaluate.
959 // See if we can produce something better than overdefined based on the T/F
960 // value.
961 LatticeVal TVal = getValueState(I.getTrueValue());
962 LatticeVal FVal = getValueState(I.getFalseValue());
963
964 // select ?, C, C -> C.
965 if (TVal.isConstant() && FVal.isConstant() &&
966 TVal.getConstant() == FVal.getConstant())
967 return (void)markConstant(&I, FVal.getConstant());
968
969 if (TVal.isUnknown()) // select ?, undef, X -> X.
970 return (void)mergeInValue(&I, FVal);
971 if (FVal.isUnknown()) // select ?, X, undef -> X.
972 return (void)mergeInValue(&I, TVal);
973 markOverdefined(&I);
974 }
975
976 // Handle Unary Operators.
visitUnaryOperator(Instruction & I)977 void SCCPSolver::visitUnaryOperator(Instruction &I) {
978 LatticeVal V0State = getValueState(I.getOperand(0));
979
980 LatticeVal &IV = ValueState[&I];
981 if (IV.isOverdefined()) return;
982
983 if (V0State.isConstant()) {
984 Constant *C = ConstantExpr::get(I.getOpcode(), V0State.getConstant());
985
986 // op Y -> undef.
987 if (isa<UndefValue>(C))
988 return;
989 return (void)markConstant(IV, &I, C);
990 }
991
992 // If something is undef, wait for it to resolve.
993 if (!V0State.isOverdefined())
994 return;
995
996 markOverdefined(&I);
997 }
998
999 // Handle Binary Operators.
visitBinaryOperator(Instruction & I)1000 void SCCPSolver::visitBinaryOperator(Instruction &I) {
1001 LatticeVal V1State = getValueState(I.getOperand(0));
1002 LatticeVal V2State = getValueState(I.getOperand(1));
1003
1004 LatticeVal &IV = ValueState[&I];
1005 if (IV.isOverdefined()) return;
1006
1007 if (V1State.isConstant() && V2State.isConstant()) {
1008 Constant *C = ConstantExpr::get(I.getOpcode(), V1State.getConstant(),
1009 V2State.getConstant());
1010 // X op Y -> undef.
1011 if (isa<UndefValue>(C))
1012 return;
1013 return (void)markConstant(IV, &I, C);
1014 }
1015
1016 // If something is undef, wait for it to resolve.
1017 if (!V1State.isOverdefined() && !V2State.isOverdefined())
1018 return;
1019
1020 // Otherwise, one of our operands is overdefined. Try to produce something
1021 // better than overdefined with some tricks.
1022 // If this is 0 / Y, it doesn't matter that the second operand is
1023 // overdefined, and we can replace it with zero.
1024 if (I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv)
1025 if (V1State.isConstant() && V1State.getConstant()->isNullValue())
1026 return (void)markConstant(IV, &I, V1State.getConstant());
1027
1028 // If this is:
1029 // -> AND/MUL with 0
1030 // -> OR with -1
1031 // it doesn't matter that the other operand is overdefined.
1032 if (I.getOpcode() == Instruction::And || I.getOpcode() == Instruction::Mul ||
1033 I.getOpcode() == Instruction::Or) {
1034 LatticeVal *NonOverdefVal = nullptr;
1035 if (!V1State.isOverdefined())
1036 NonOverdefVal = &V1State;
1037 else if (!V2State.isOverdefined())
1038 NonOverdefVal = &V2State;
1039
1040 if (NonOverdefVal) {
1041 if (NonOverdefVal->isUnknown())
1042 return;
1043
1044 if (I.getOpcode() == Instruction::And ||
1045 I.getOpcode() == Instruction::Mul) {
1046 // X and 0 = 0
1047 // X * 0 = 0
1048 if (NonOverdefVal->getConstant()->isNullValue())
1049 return (void)markConstant(IV, &I, NonOverdefVal->getConstant());
1050 } else {
1051 // X or -1 = -1
1052 if (ConstantInt *CI = NonOverdefVal->getConstantInt())
1053 if (CI->isMinusOne())
1054 return (void)markConstant(IV, &I, NonOverdefVal->getConstant());
1055 }
1056 }
1057 }
1058
1059 markOverdefined(&I);
1060 }
1061
1062 // Handle ICmpInst instruction.
visitCmpInst(CmpInst & I)1063 void SCCPSolver::visitCmpInst(CmpInst &I) {
1064 // Do not cache this lookup, getValueState calls later in the function might
1065 // invalidate the reference.
1066 if (ValueState[&I].isOverdefined()) return;
1067
1068 Value *Op1 = I.getOperand(0);
1069 Value *Op2 = I.getOperand(1);
1070
1071 // For parameters, use ParamState which includes constant range info if
1072 // available.
1073 auto V1Param = ParamState.find(Op1);
1074 ValueLatticeElement V1State = (V1Param != ParamState.end())
1075 ? V1Param->second
1076 : getValueState(Op1).toValueLattice();
1077
1078 auto V2Param = ParamState.find(Op2);
1079 ValueLatticeElement V2State = V2Param != ParamState.end()
1080 ? V2Param->second
1081 : getValueState(Op2).toValueLattice();
1082
1083 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State);
1084 if (C) {
1085 if (isa<UndefValue>(C))
1086 return;
1087 LatticeVal CV;
1088 CV.markConstant(C);
1089 mergeInValue(&I, CV);
1090 return;
1091 }
1092
1093 // If operands are still unknown, wait for it to resolve.
1094 if (!V1State.isOverdefined() && !V2State.isOverdefined() &&
1095 !ValueState[&I].isConstant())
1096 return;
1097
1098 markOverdefined(&I);
1099 }
1100
1101 // Handle getelementptr instructions. If all operands are constants then we
1102 // can turn this into a getelementptr ConstantExpr.
visitGetElementPtrInst(GetElementPtrInst & I)1103 void SCCPSolver::visitGetElementPtrInst(GetElementPtrInst &I) {
1104 if (ValueState[&I].isOverdefined()) return;
1105
1106 SmallVector<Constant*, 8> Operands;
1107 Operands.reserve(I.getNumOperands());
1108
1109 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1110 LatticeVal State = getValueState(I.getOperand(i));
1111 if (State.isUnknown())
1112 return; // Operands are not resolved yet.
1113
1114 if (State.isOverdefined())
1115 return (void)markOverdefined(&I);
1116
1117 assert(State.isConstant() && "Unknown state!");
1118 Operands.push_back(State.getConstant());
1119 }
1120
1121 Constant *Ptr = Operands[0];
1122 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end());
1123 Constant *C =
1124 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices);
1125 if (isa<UndefValue>(C))
1126 return;
1127 markConstant(&I, C);
1128 }
1129
visitStoreInst(StoreInst & SI)1130 void SCCPSolver::visitStoreInst(StoreInst &SI) {
1131 // If this store is of a struct, ignore it.
1132 if (SI.getOperand(0)->getType()->isStructTy())
1133 return;
1134
1135 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1136 return;
1137
1138 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1139 DenseMap<GlobalVariable*, LatticeVal>::iterator I = TrackedGlobals.find(GV);
1140 if (I == TrackedGlobals.end() || I->second.isOverdefined()) return;
1141
1142 // Get the value we are storing into the global, then merge it.
1143 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)));
1144 if (I->second.isOverdefined())
1145 TrackedGlobals.erase(I); // No need to keep tracking this!
1146 }
1147
1148 // Handle load instructions. If the operand is a constant pointer to a constant
1149 // global, we can replace the load with the loaded constant value!
visitLoadInst(LoadInst & I)1150 void SCCPSolver::visitLoadInst(LoadInst &I) {
1151 // If this load is of a struct, just mark the result overdefined.
1152 if (I.getType()->isStructTy())
1153 return (void)markOverdefined(&I);
1154
1155 LatticeVal PtrVal = getValueState(I.getOperand(0));
1156 if (PtrVal.isUnknown()) return; // The pointer is not resolved yet!
1157
1158 LatticeVal &IV = ValueState[&I];
1159 if (IV.isOverdefined()) return;
1160
1161 if (!PtrVal.isConstant() || I.isVolatile())
1162 return (void)markOverdefined(IV, &I);
1163
1164 Constant *Ptr = PtrVal.getConstant();
1165
1166 // load null is undefined.
1167 if (isa<ConstantPointerNull>(Ptr)) {
1168 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace()))
1169 return (void)markOverdefined(IV, &I);
1170 else
1171 return;
1172 }
1173
1174 // Transform load (constant global) into the value loaded.
1175 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) {
1176 if (!TrackedGlobals.empty()) {
1177 // If we are tracking this global, merge in the known value for it.
1178 DenseMap<GlobalVariable*, LatticeVal>::iterator It =
1179 TrackedGlobals.find(GV);
1180 if (It != TrackedGlobals.end()) {
1181 mergeInValue(IV, &I, It->second);
1182 return;
1183 }
1184 }
1185 }
1186
1187 // Transform load from a constant into a constant if possible.
1188 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL)) {
1189 if (isa<UndefValue>(C))
1190 return;
1191 return (void)markConstant(IV, &I, C);
1192 }
1193
1194 // Otherwise we cannot say for certain what value this load will produce.
1195 // Bail out.
1196 markOverdefined(IV, &I);
1197 }
1198
visitCallSite(CallSite CS)1199 void SCCPSolver::visitCallSite(CallSite CS) {
1200 Function *F = CS.getCalledFunction();
1201 Instruction *I = CS.getInstruction();
1202
1203 if (auto *II = dyn_cast<IntrinsicInst>(I)) {
1204 if (II->getIntrinsicID() == Intrinsic::ssa_copy) {
1205 if (ValueState[I].isOverdefined())
1206 return;
1207
1208 auto *PI = getPredicateInfoFor(I);
1209 if (!PI)
1210 return;
1211
1212 Value *CopyOf = I->getOperand(0);
1213 auto *PBranch = dyn_cast<PredicateBranch>(PI);
1214 if (!PBranch) {
1215 mergeInValue(ValueState[I], I, getValueState(CopyOf));
1216 return;
1217 }
1218
1219 Value *Cond = PBranch->Condition;
1220
1221 // Everything below relies on the condition being a comparison.
1222 auto *Cmp = dyn_cast<CmpInst>(Cond);
1223 if (!Cmp) {
1224 mergeInValue(ValueState[I], I, getValueState(CopyOf));
1225 return;
1226 }
1227
1228 Value *CmpOp0 = Cmp->getOperand(0);
1229 Value *CmpOp1 = Cmp->getOperand(1);
1230 if (CopyOf != CmpOp0 && CopyOf != CmpOp1) {
1231 mergeInValue(ValueState[I], I, getValueState(CopyOf));
1232 return;
1233 }
1234
1235 if (CmpOp0 != CopyOf)
1236 std::swap(CmpOp0, CmpOp1);
1237
1238 LatticeVal OriginalVal = getValueState(CopyOf);
1239 LatticeVal EqVal = getValueState(CmpOp1);
1240 LatticeVal &IV = ValueState[I];
1241 if (PBranch->TrueEdge && Cmp->getPredicate() == CmpInst::ICMP_EQ) {
1242 addAdditionalUser(CmpOp1, I);
1243 if (OriginalVal.isConstant())
1244 mergeInValue(IV, I, OriginalVal);
1245 else
1246 mergeInValue(IV, I, EqVal);
1247 return;
1248 }
1249 if (!PBranch->TrueEdge && Cmp->getPredicate() == CmpInst::ICMP_NE) {
1250 addAdditionalUser(CmpOp1, I);
1251 if (OriginalVal.isConstant())
1252 mergeInValue(IV, I, OriginalVal);
1253 else
1254 mergeInValue(IV, I, EqVal);
1255 return;
1256 }
1257
1258 return (void)mergeInValue(IV, I, getValueState(CopyOf));
1259 }
1260 }
1261
1262 // The common case is that we aren't tracking the callee, either because we
1263 // are not doing interprocedural analysis or the callee is indirect, or is
1264 // external. Handle these cases first.
1265 if (!F || F->isDeclaration()) {
1266 CallOverdefined:
1267 // Void return and not tracking callee, just bail.
1268 if (I->getType()->isVoidTy()) return;
1269
1270 // Otherwise, if we have a single return value case, and if the function is
1271 // a declaration, maybe we can constant fold it.
1272 if (F && F->isDeclaration() && !I->getType()->isStructTy() &&
1273 canConstantFoldCallTo(cast<CallBase>(CS.getInstruction()), F)) {
1274 SmallVector<Constant*, 8> Operands;
1275 for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
1276 AI != E; ++AI) {
1277 if (AI->get()->getType()->isStructTy())
1278 return markOverdefined(I); // Can't handle struct args.
1279 LatticeVal State = getValueState(*AI);
1280
1281 if (State.isUnknown())
1282 return; // Operands are not resolved yet.
1283 if (State.isOverdefined())
1284 return (void)markOverdefined(I);
1285 assert(State.isConstant() && "Unknown state!");
1286 Operands.push_back(State.getConstant());
1287 }
1288
1289 if (getValueState(I).isOverdefined())
1290 return;
1291
1292 // If we can constant fold this, mark the result of the call as a
1293 // constant.
1294 if (Constant *C = ConstantFoldCall(cast<CallBase>(CS.getInstruction()), F,
1295 Operands, &GetTLI(*F))) {
1296 // call -> undef.
1297 if (isa<UndefValue>(C))
1298 return;
1299 return (void)markConstant(I, C);
1300 }
1301 }
1302
1303 // Otherwise, we don't know anything about this call, mark it overdefined.
1304 return (void)markOverdefined(I);
1305 }
1306
1307 // If this is a local function that doesn't have its address taken, mark its
1308 // entry block executable and merge in the actual arguments to the call into
1309 // the formal arguments of the function.
1310 if (!TrackingIncomingArguments.empty() && TrackingIncomingArguments.count(F)){
1311 MarkBlockExecutable(&F->front());
1312
1313 // Propagate information from this call site into the callee.
1314 CallSite::arg_iterator CAI = CS.arg_begin();
1315 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end();
1316 AI != E; ++AI, ++CAI) {
1317 // If this argument is byval, and if the function is not readonly, there
1318 // will be an implicit copy formed of the input aggregate.
1319 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1320 markOverdefined(&*AI);
1321 continue;
1322 }
1323
1324 if (auto *STy = dyn_cast<StructType>(AI->getType())) {
1325 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1326 LatticeVal CallArg = getStructValueState(*CAI, i);
1327 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg);
1328 }
1329 } else {
1330 // Most other parts of the Solver still only use the simpler value
1331 // lattice, so we propagate changes for parameters to both lattices.
1332 LatticeVal ConcreteArgument = getValueState(*CAI);
1333 bool ParamChanged =
1334 getParamState(&*AI).mergeIn(ConcreteArgument.toValueLattice(), DL);
1335 bool ValueChanged = mergeInValue(&*AI, ConcreteArgument);
1336 // Add argument to work list, if the state of a parameter changes but
1337 // ValueState does not change (because it is already overdefined there),
1338 // We have to take changes in ParamState into account, as it is used
1339 // when evaluating Cmp instructions.
1340 if (!ValueChanged && ParamChanged)
1341 pushToWorkList(ValueState[&*AI], &*AI);
1342 }
1343 }
1344 }
1345
1346 // If this is a single/zero retval case, see if we're tracking the function.
1347 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) {
1348 if (!MRVFunctionsTracked.count(F))
1349 goto CallOverdefined; // Not tracking this callee.
1350
1351 // If we are tracking this callee, propagate the result of the function
1352 // into this call site.
1353 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1354 mergeInValue(getStructValueState(I, i), I,
1355 TrackedMultipleRetVals[std::make_pair(F, i)]);
1356 } else {
1357 MapVector<Function*, LatticeVal>::iterator TFRVI = TrackedRetVals.find(F);
1358 if (TFRVI == TrackedRetVals.end())
1359 goto CallOverdefined; // Not tracking this callee.
1360
1361 // If so, propagate the return value of the callee into this call result.
1362 mergeInValue(I, TFRVI->second);
1363 }
1364 }
1365
Solve()1366 void SCCPSolver::Solve() {
1367 // Process the work lists until they are empty!
1368 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1369 !OverdefinedInstWorkList.empty()) {
1370 // Process the overdefined instruction's work list first, which drives other
1371 // things to overdefined more quickly.
1372 while (!OverdefinedInstWorkList.empty()) {
1373 Value *I = OverdefinedInstWorkList.pop_back_val();
1374
1375 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n');
1376
1377 // "I" got into the work list because it either made the transition from
1378 // bottom to constant, or to overdefined.
1379 //
1380 // Anything on this worklist that is overdefined need not be visited
1381 // since all of its users will have already been marked as overdefined
1382 // Update all of the users of this instruction's value.
1383 //
1384 markUsersAsChanged(I);
1385 }
1386
1387 // Process the instruction work list.
1388 while (!InstWorkList.empty()) {
1389 Value *I = InstWorkList.pop_back_val();
1390
1391 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n');
1392
1393 // "I" got into the work list because it made the transition from undef to
1394 // constant.
1395 //
1396 // Anything on this worklist that is overdefined need not be visited
1397 // since all of its users will have already been marked as overdefined.
1398 // Update all of the users of this instruction's value.
1399 //
1400 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined())
1401 markUsersAsChanged(I);
1402 }
1403
1404 // Process the basic block work list.
1405 while (!BBWorkList.empty()) {
1406 BasicBlock *BB = BBWorkList.back();
1407 BBWorkList.pop_back();
1408
1409 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n');
1410
1411 // Notify all instructions in this basic block that they are newly
1412 // executable.
1413 visit(BB);
1414 }
1415 }
1416 }
1417
1418 /// ResolvedUndefsIn - While solving the dataflow for a function, we assume
1419 /// that branches on undef values cannot reach any of their successors.
1420 /// However, this is not a safe assumption. After we solve dataflow, this
1421 /// method should be use to handle this. If this returns true, the solver
1422 /// should be rerun.
1423 ///
1424 /// This method handles this by finding an unresolved branch and marking it one
1425 /// of the edges from the block as being feasible, even though the condition
1426 /// doesn't say it would otherwise be. This allows SCCP to find the rest of the
1427 /// CFG and only slightly pessimizes the analysis results (by marking one,
1428 /// potentially infeasible, edge feasible). This cannot usefully modify the
1429 /// constraints on the condition of the branch, as that would impact other users
1430 /// of the value.
1431 ///
1432 /// This scan also checks for values that use undefs, whose results are actually
1433 /// defined. For example, 'zext i8 undef to i32' should produce all zeros
1434 /// conservatively, as "(zext i8 X -> i32) & 0xFF00" must always return zero,
1435 /// even if X isn't defined.
ResolvedUndefsIn(Function & F)1436 bool SCCPSolver::ResolvedUndefsIn(Function &F) {
1437 for (BasicBlock &BB : F) {
1438 if (!BBExecutable.count(&BB))
1439 continue;
1440
1441 for (Instruction &I : BB) {
1442 // Look for instructions which produce undef values.
1443 if (I.getType()->isVoidTy()) continue;
1444
1445 if (auto *STy = dyn_cast<StructType>(I.getType())) {
1446 // Only a few things that can be structs matter for undef.
1447
1448 // Tracked calls must never be marked overdefined in ResolvedUndefsIn.
1449 if (CallSite CS = CallSite(&I))
1450 if (Function *F = CS.getCalledFunction())
1451 if (MRVFunctionsTracked.count(F))
1452 continue;
1453
1454 // extractvalue and insertvalue don't need to be marked; they are
1455 // tracked as precisely as their operands.
1456 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I))
1457 continue;
1458
1459 // Send the results of everything else to overdefined. We could be
1460 // more precise than this but it isn't worth bothering.
1461 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1462 LatticeVal &LV = getStructValueState(&I, i);
1463 if (LV.isUnknown())
1464 markOverdefined(LV, &I);
1465 }
1466 continue;
1467 }
1468
1469 LatticeVal &LV = getValueState(&I);
1470 if (!LV.isUnknown())
1471 continue;
1472
1473 // There are two reasons a call can have an undef result
1474 // 1. It could be tracked.
1475 // 2. It could be constant-foldable.
1476 // Because of the way we solve return values, tracked calls must
1477 // never be marked overdefined in ResolvedUndefsIn.
1478 if (CallSite CS = CallSite(&I)) {
1479 if (Function *F = CS.getCalledFunction())
1480 if (TrackedRetVals.count(F))
1481 continue;
1482
1483 // If the call is constant-foldable, we mark it overdefined because
1484 // we do not know what return values are valid.
1485 markOverdefined(&I);
1486 return true;
1487 }
1488
1489 // extractvalue is safe; check here because the argument is a struct.
1490 if (isa<ExtractValueInst>(I))
1491 continue;
1492
1493 // Compute the operand LatticeVals, for convenience below.
1494 // Anything taking a struct is conservatively assumed to require
1495 // overdefined markings.
1496 if (I.getOperand(0)->getType()->isStructTy()) {
1497 markOverdefined(&I);
1498 return true;
1499 }
1500 LatticeVal Op0LV = getValueState(I.getOperand(0));
1501 LatticeVal Op1LV;
1502 if (I.getNumOperands() == 2) {
1503 if (I.getOperand(1)->getType()->isStructTy()) {
1504 markOverdefined(&I);
1505 return true;
1506 }
1507
1508 Op1LV = getValueState(I.getOperand(1));
1509 }
1510 // If this is an instructions whose result is defined even if the input is
1511 // not fully defined, propagate the information.
1512 Type *ITy = I.getType();
1513 switch (I.getOpcode()) {
1514 case Instruction::Add:
1515 case Instruction::Sub:
1516 case Instruction::Trunc:
1517 case Instruction::FPTrunc:
1518 case Instruction::BitCast:
1519 break; // Any undef -> undef
1520 case Instruction::FSub:
1521 case Instruction::FAdd:
1522 case Instruction::FMul:
1523 case Instruction::FDiv:
1524 case Instruction::FRem:
1525 // Floating-point binary operation: be conservative.
1526 if (Op0LV.isUnknown() && Op1LV.isUnknown())
1527 markForcedConstant(&I, Constant::getNullValue(ITy));
1528 else
1529 markOverdefined(&I);
1530 return true;
1531 case Instruction::FNeg:
1532 break; // fneg undef -> undef
1533 case Instruction::ZExt:
1534 case Instruction::SExt:
1535 case Instruction::FPToUI:
1536 case Instruction::FPToSI:
1537 case Instruction::FPExt:
1538 case Instruction::PtrToInt:
1539 case Instruction::IntToPtr:
1540 case Instruction::SIToFP:
1541 case Instruction::UIToFP:
1542 // undef -> 0; some outputs are impossible
1543 markForcedConstant(&I, Constant::getNullValue(ITy));
1544 return true;
1545 case Instruction::Mul:
1546 case Instruction::And:
1547 // Both operands undef -> undef
1548 if (Op0LV.isUnknown() && Op1LV.isUnknown())
1549 break;
1550 // undef * X -> 0. X could be zero.
1551 // undef & X -> 0. X could be zero.
1552 markForcedConstant(&I, Constant::getNullValue(ITy));
1553 return true;
1554 case Instruction::Or:
1555 // Both operands undef -> undef
1556 if (Op0LV.isUnknown() && Op1LV.isUnknown())
1557 break;
1558 // undef | X -> -1. X could be -1.
1559 markForcedConstant(&I, Constant::getAllOnesValue(ITy));
1560 return true;
1561 case Instruction::Xor:
1562 // undef ^ undef -> 0; strictly speaking, this is not strictly
1563 // necessary, but we try to be nice to people who expect this
1564 // behavior in simple cases
1565 if (Op0LV.isUnknown() && Op1LV.isUnknown()) {
1566 markForcedConstant(&I, Constant::getNullValue(ITy));
1567 return true;
1568 }
1569 // undef ^ X -> undef
1570 break;
1571 case Instruction::SDiv:
1572 case Instruction::UDiv:
1573 case Instruction::SRem:
1574 case Instruction::URem:
1575 // X / undef -> undef. No change.
1576 // X % undef -> undef. No change.
1577 if (Op1LV.isUnknown()) break;
1578
1579 // X / 0 -> undef. No change.
1580 // X % 0 -> undef. No change.
1581 if (Op1LV.isConstant() && Op1LV.getConstant()->isZeroValue())
1582 break;
1583
1584 // undef / X -> 0. X could be maxint.
1585 // undef % X -> 0. X could be 1.
1586 markForcedConstant(&I, Constant::getNullValue(ITy));
1587 return true;
1588 case Instruction::AShr:
1589 // X >>a undef -> undef.
1590 if (Op1LV.isUnknown()) break;
1591
1592 // Shifting by the bitwidth or more is undefined.
1593 if (Op1LV.isConstant()) {
1594 if (auto *ShiftAmt = Op1LV.getConstantInt())
1595 if (ShiftAmt->getLimitedValue() >=
1596 ShiftAmt->getType()->getScalarSizeInBits())
1597 break;
1598 }
1599
1600 // undef >>a X -> 0
1601 markForcedConstant(&I, Constant::getNullValue(ITy));
1602 return true;
1603 case Instruction::LShr:
1604 case Instruction::Shl:
1605 // X << undef -> undef.
1606 // X >> undef -> undef.
1607 if (Op1LV.isUnknown()) break;
1608
1609 // Shifting by the bitwidth or more is undefined.
1610 if (Op1LV.isConstant()) {
1611 if (auto *ShiftAmt = Op1LV.getConstantInt())
1612 if (ShiftAmt->getLimitedValue() >=
1613 ShiftAmt->getType()->getScalarSizeInBits())
1614 break;
1615 }
1616
1617 // undef << X -> 0
1618 // undef >> X -> 0
1619 markForcedConstant(&I, Constant::getNullValue(ITy));
1620 return true;
1621 case Instruction::Select:
1622 Op1LV = getValueState(I.getOperand(1));
1623 // undef ? X : Y -> X or Y. There could be commonality between X/Y.
1624 if (Op0LV.isUnknown()) {
1625 if (!Op1LV.isConstant()) // Pick the constant one if there is any.
1626 Op1LV = getValueState(I.getOperand(2));
1627 } else if (Op1LV.isUnknown()) {
1628 // c ? undef : undef -> undef. No change.
1629 Op1LV = getValueState(I.getOperand(2));
1630 if (Op1LV.isUnknown())
1631 break;
1632 // Otherwise, c ? undef : x -> x.
1633 } else {
1634 // Leave Op1LV as Operand(1)'s LatticeValue.
1635 }
1636
1637 if (Op1LV.isConstant())
1638 markForcedConstant(&I, Op1LV.getConstant());
1639 else
1640 markOverdefined(&I);
1641 return true;
1642 case Instruction::Load:
1643 // A load here means one of two things: a load of undef from a global,
1644 // a load from an unknown pointer. Either way, having it return undef
1645 // is okay.
1646 break;
1647 case Instruction::ICmp:
1648 // X == undef -> undef. Other comparisons get more complicated.
1649 Op0LV = getValueState(I.getOperand(0));
1650 Op1LV = getValueState(I.getOperand(1));
1651
1652 if ((Op0LV.isUnknown() || Op1LV.isUnknown()) &&
1653 cast<ICmpInst>(&I)->isEquality())
1654 break;
1655 markOverdefined(&I);
1656 return true;
1657 case Instruction::Call:
1658 case Instruction::Invoke:
1659 case Instruction::CallBr:
1660 llvm_unreachable("Call-like instructions should have be handled early");
1661 default:
1662 // If we don't know what should happen here, conservatively mark it
1663 // overdefined.
1664 markOverdefined(&I);
1665 return true;
1666 }
1667 }
1668
1669 // Check to see if we have a branch or switch on an undefined value. If so
1670 // we force the branch to go one way or the other to make the successor
1671 // values live. It doesn't really matter which way we force it.
1672 Instruction *TI = BB.getTerminator();
1673 if (auto *BI = dyn_cast<BranchInst>(TI)) {
1674 if (!BI->isConditional()) continue;
1675 if (!getValueState(BI->getCondition()).isUnknown())
1676 continue;
1677
1678 // If the input to SCCP is actually branch on undef, fix the undef to
1679 // false.
1680 if (isa<UndefValue>(BI->getCondition())) {
1681 BI->setCondition(ConstantInt::getFalse(BI->getContext()));
1682 markEdgeExecutable(&BB, TI->getSuccessor(1));
1683 return true;
1684 }
1685
1686 // Otherwise, it is a branch on a symbolic value which is currently
1687 // considered to be undef. Make sure some edge is executable, so a
1688 // branch on "undef" always flows somewhere.
1689 // FIXME: Distinguish between dead code and an LLVM "undef" value.
1690 BasicBlock *DefaultSuccessor = TI->getSuccessor(1);
1691 if (markEdgeExecutable(&BB, DefaultSuccessor))
1692 return true;
1693
1694 continue;
1695 }
1696
1697 if (auto *IBR = dyn_cast<IndirectBrInst>(TI)) {
1698 // Indirect branch with no successor ?. Its ok to assume it branches
1699 // to no target.
1700 if (IBR->getNumSuccessors() < 1)
1701 continue;
1702
1703 if (!getValueState(IBR->getAddress()).isUnknown())
1704 continue;
1705
1706 // If the input to SCCP is actually branch on undef, fix the undef to
1707 // the first successor of the indirect branch.
1708 if (isa<UndefValue>(IBR->getAddress())) {
1709 IBR->setAddress(BlockAddress::get(IBR->getSuccessor(0)));
1710 markEdgeExecutable(&BB, IBR->getSuccessor(0));
1711 return true;
1712 }
1713
1714 // Otherwise, it is a branch on a symbolic value which is currently
1715 // considered to be undef. Make sure some edge is executable, so a
1716 // branch on "undef" always flows somewhere.
1717 // FIXME: IndirectBr on "undef" doesn't actually need to go anywhere:
1718 // we can assume the branch has undefined behavior instead.
1719 BasicBlock *DefaultSuccessor = IBR->getSuccessor(0);
1720 if (markEdgeExecutable(&BB, DefaultSuccessor))
1721 return true;
1722
1723 continue;
1724 }
1725
1726 if (auto *SI = dyn_cast<SwitchInst>(TI)) {
1727 if (!SI->getNumCases() || !getValueState(SI->getCondition()).isUnknown())
1728 continue;
1729
1730 // If the input to SCCP is actually switch on undef, fix the undef to
1731 // the first constant.
1732 if (isa<UndefValue>(SI->getCondition())) {
1733 SI->setCondition(SI->case_begin()->getCaseValue());
1734 markEdgeExecutable(&BB, SI->case_begin()->getCaseSuccessor());
1735 return true;
1736 }
1737
1738 // Otherwise, it is a branch on a symbolic value which is currently
1739 // considered to be undef. Make sure some edge is executable, so a
1740 // branch on "undef" always flows somewhere.
1741 // FIXME: Distinguish between dead code and an LLVM "undef" value.
1742 BasicBlock *DefaultSuccessor = SI->case_begin()->getCaseSuccessor();
1743 if (markEdgeExecutable(&BB, DefaultSuccessor))
1744 return true;
1745
1746 continue;
1747 }
1748 }
1749
1750 return false;
1751 }
1752
tryToReplaceWithConstant(SCCPSolver & Solver,Value * V)1753 static bool tryToReplaceWithConstant(SCCPSolver &Solver, Value *V) {
1754 Constant *Const = nullptr;
1755 if (V->getType()->isStructTy()) {
1756 std::vector<LatticeVal> IVs = Solver.getStructLatticeValueFor(V);
1757 if (llvm::any_of(IVs,
1758 [](const LatticeVal &LV) { return LV.isOverdefined(); }))
1759 return false;
1760 std::vector<Constant *> ConstVals;
1761 auto *ST = cast<StructType>(V->getType());
1762 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1763 LatticeVal V = IVs[i];
1764 ConstVals.push_back(V.isConstant()
1765 ? V.getConstant()
1766 : UndefValue::get(ST->getElementType(i)));
1767 }
1768 Const = ConstantStruct::get(ST, ConstVals);
1769 } else {
1770 const LatticeVal &IV = Solver.getLatticeValueFor(V);
1771 if (IV.isOverdefined())
1772 return false;
1773
1774 Const = IV.isConstant() ? IV.getConstant() : UndefValue::get(V->getType());
1775 }
1776 assert(Const && "Constant is nullptr here!");
1777
1778 // Replacing `musttail` instructions with constant breaks `musttail` invariant
1779 // unless the call itself can be removed
1780 CallInst *CI = dyn_cast<CallInst>(V);
1781 if (CI && CI->isMustTailCall() && !CI->isSafeToRemove()) {
1782 CallSite CS(CI);
1783 Function *F = CS.getCalledFunction();
1784
1785 // Don't zap returns of the callee
1786 if (F)
1787 Solver.AddMustTailCallee(F);
1788
1789 LLVM_DEBUG(dbgs() << " Can\'t treat the result of musttail call : " << *CI
1790 << " as a constant\n");
1791 return false;
1792 }
1793
1794 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n');
1795
1796 // Replaces all of the uses of a variable with uses of the constant.
1797 V->replaceAllUsesWith(Const);
1798 return true;
1799 }
1800
1801 // runSCCP() - Run the Sparse Conditional Constant Propagation algorithm,
1802 // and return true if the function was modified.
runSCCP(Function & F,const DataLayout & DL,const TargetLibraryInfo * TLI)1803 static bool runSCCP(Function &F, const DataLayout &DL,
1804 const TargetLibraryInfo *TLI) {
1805 LLVM_DEBUG(dbgs() << "SCCP on function '" << F.getName() << "'\n");
1806 SCCPSolver Solver(
1807 DL, [TLI](Function &F) -> const TargetLibraryInfo & { return *TLI; });
1808
1809 // Mark the first block of the function as being executable.
1810 Solver.MarkBlockExecutable(&F.front());
1811
1812 // Mark all arguments to the function as being overdefined.
1813 for (Argument &AI : F.args())
1814 Solver.markOverdefined(&AI);
1815
1816 // Solve for constants.
1817 bool ResolvedUndefs = true;
1818 while (ResolvedUndefs) {
1819 Solver.Solve();
1820 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFs\n");
1821 ResolvedUndefs = Solver.ResolvedUndefsIn(F);
1822 }
1823
1824 bool MadeChanges = false;
1825
1826 // If we decided that there are basic blocks that are dead in this function,
1827 // delete their contents now. Note that we cannot actually delete the blocks,
1828 // as we cannot modify the CFG of the function.
1829
1830 for (BasicBlock &BB : F) {
1831 if (!Solver.isBlockExecutable(&BB)) {
1832 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << BB);
1833
1834 ++NumDeadBlocks;
1835 NumInstRemoved += removeAllNonTerminatorAndEHPadInstructions(&BB);
1836
1837 MadeChanges = true;
1838 continue;
1839 }
1840
1841 // Iterate over all of the instructions in a function, replacing them with
1842 // constants if we have found them to be of constant values.
1843 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) {
1844 Instruction *Inst = &*BI++;
1845 if (Inst->getType()->isVoidTy() || Inst->isTerminator())
1846 continue;
1847
1848 if (tryToReplaceWithConstant(Solver, Inst)) {
1849 if (isInstructionTriviallyDead(Inst))
1850 Inst->eraseFromParent();
1851 // Hey, we just changed something!
1852 MadeChanges = true;
1853 ++NumInstRemoved;
1854 }
1855 }
1856 }
1857
1858 return MadeChanges;
1859 }
1860
run(Function & F,FunctionAnalysisManager & AM)1861 PreservedAnalyses SCCPPass::run(Function &F, FunctionAnalysisManager &AM) {
1862 const DataLayout &DL = F.getParent()->getDataLayout();
1863 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1864 if (!runSCCP(F, DL, &TLI))
1865 return PreservedAnalyses::all();
1866
1867 auto PA = PreservedAnalyses();
1868 PA.preserve<GlobalsAA>();
1869 PA.preserveSet<CFGAnalyses>();
1870 return PA;
1871 }
1872
1873 namespace {
1874
1875 //===--------------------------------------------------------------------===//
1876 //
1877 /// SCCP Class - This class uses the SCCPSolver to implement a per-function
1878 /// Sparse Conditional Constant Propagator.
1879 ///
1880 class SCCPLegacyPass : public FunctionPass {
1881 public:
1882 // Pass identification, replacement for typeid
1883 static char ID;
1884
SCCPLegacyPass()1885 SCCPLegacyPass() : FunctionPass(ID) {
1886 initializeSCCPLegacyPassPass(*PassRegistry::getPassRegistry());
1887 }
1888
getAnalysisUsage(AnalysisUsage & AU) const1889 void getAnalysisUsage(AnalysisUsage &AU) const override {
1890 AU.addRequired<TargetLibraryInfoWrapperPass>();
1891 AU.addPreserved<GlobalsAAWrapperPass>();
1892 AU.setPreservesCFG();
1893 }
1894
1895 // runOnFunction - Run the Sparse Conditional Constant Propagation
1896 // algorithm, and return true if the function was modified.
runOnFunction(Function & F)1897 bool runOnFunction(Function &F) override {
1898 if (skipFunction(F))
1899 return false;
1900 const DataLayout &DL = F.getParent()->getDataLayout();
1901 const TargetLibraryInfo *TLI =
1902 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1903 return runSCCP(F, DL, TLI);
1904 }
1905 };
1906
1907 } // end anonymous namespace
1908
1909 char SCCPLegacyPass::ID = 0;
1910
1911 INITIALIZE_PASS_BEGIN(SCCPLegacyPass, "sccp",
1912 "Sparse Conditional Constant Propagation", false, false)
INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)1913 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1914 INITIALIZE_PASS_END(SCCPLegacyPass, "sccp",
1915 "Sparse Conditional Constant Propagation", false, false)
1916
1917 // createSCCPPass - This is the public interface to this file.
1918 FunctionPass *llvm::createSCCPPass() { return new SCCPLegacyPass(); }
1919
findReturnsToZap(Function & F,SmallVector<ReturnInst *,8> & ReturnsToZap,SCCPSolver & Solver)1920 static void findReturnsToZap(Function &F,
1921 SmallVector<ReturnInst *, 8> &ReturnsToZap,
1922 SCCPSolver &Solver) {
1923 // We can only do this if we know that nothing else can call the function.
1924 if (!Solver.isArgumentTrackedFunction(&F))
1925 return;
1926
1927 // There is a non-removable musttail call site of this function. Zapping
1928 // returns is not allowed.
1929 if (Solver.isMustTailCallee(&F)) {
1930 LLVM_DEBUG(dbgs() << "Can't zap returns of the function : " << F.getName()
1931 << " due to present musttail call of it\n");
1932 return;
1933 }
1934
1935 assert(
1936 all_of(F.users(),
1937 [&Solver](User *U) {
1938 if (isa<Instruction>(U) &&
1939 !Solver.isBlockExecutable(cast<Instruction>(U)->getParent()))
1940 return true;
1941 // Non-callsite uses are not impacted by zapping. Also, constant
1942 // uses (like blockaddresses) could stuck around, without being
1943 // used in the underlying IR, meaning we do not have lattice
1944 // values for them.
1945 if (!CallSite(U))
1946 return true;
1947 if (U->getType()->isStructTy()) {
1948 return all_of(
1949 Solver.getStructLatticeValueFor(U),
1950 [](const LatticeVal &LV) { return !LV.isOverdefined(); });
1951 }
1952 return !Solver.getLatticeValueFor(U).isOverdefined();
1953 }) &&
1954 "We can only zap functions where all live users have a concrete value");
1955
1956 for (BasicBlock &BB : F) {
1957 if (CallInst *CI = BB.getTerminatingMustTailCall()) {
1958 LLVM_DEBUG(dbgs() << "Can't zap return of the block due to present "
1959 << "musttail call : " << *CI << "\n");
1960 (void)CI;
1961 return;
1962 }
1963
1964 if (auto *RI = dyn_cast<ReturnInst>(BB.getTerminator()))
1965 if (!isa<UndefValue>(RI->getOperand(0)))
1966 ReturnsToZap.push_back(RI);
1967 }
1968 }
1969
1970 // Update the condition for terminators that are branching on indeterminate
1971 // values, forcing them to use a specific edge.
forceIndeterminateEdge(Instruction * I,SCCPSolver & Solver)1972 static void forceIndeterminateEdge(Instruction* I, SCCPSolver &Solver) {
1973 BasicBlock *Dest = nullptr;
1974 Constant *C = nullptr;
1975 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1976 if (!isa<ConstantInt>(SI->getCondition())) {
1977 // Indeterminate switch; use first case value.
1978 Dest = SI->case_begin()->getCaseSuccessor();
1979 C = SI->case_begin()->getCaseValue();
1980 }
1981 } else if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1982 if (!isa<ConstantInt>(BI->getCondition())) {
1983 // Indeterminate branch; use false.
1984 Dest = BI->getSuccessor(1);
1985 C = ConstantInt::getFalse(BI->getContext());
1986 }
1987 } else if (IndirectBrInst *IBR = dyn_cast<IndirectBrInst>(I)) {
1988 if (!isa<BlockAddress>(IBR->getAddress()->stripPointerCasts())) {
1989 // Indeterminate indirectbr; use successor 0.
1990 Dest = IBR->getSuccessor(0);
1991 C = BlockAddress::get(IBR->getSuccessor(0));
1992 }
1993 } else {
1994 llvm_unreachable("Unexpected terminator instruction");
1995 }
1996 if (C) {
1997 assert(Solver.isEdgeFeasible(I->getParent(), Dest) &&
1998 "Didn't find feasible edge?");
1999 (void)Dest;
2000
2001 I->setOperand(0, C);
2002 }
2003 }
2004
runIPSCCP(Module & M,const DataLayout & DL,std::function<const TargetLibraryInfo & (Function &)> GetTLI,function_ref<AnalysisResultsForFn (Function &)> getAnalysis)2005 bool llvm::runIPSCCP(
2006 Module &M, const DataLayout &DL,
2007 std::function<const TargetLibraryInfo &(Function &)> GetTLI,
2008 function_ref<AnalysisResultsForFn(Function &)> getAnalysis) {
2009 SCCPSolver Solver(DL, GetTLI);
2010
2011 // Loop over all functions, marking arguments to those with their addresses
2012 // taken or that are external as overdefined.
2013 for (Function &F : M) {
2014 if (F.isDeclaration())
2015 continue;
2016
2017 Solver.addAnalysis(F, getAnalysis(F));
2018
2019 // Determine if we can track the function's return values. If so, add the
2020 // function to the solver's set of return-tracked functions.
2021 if (canTrackReturnsInterprocedurally(&F))
2022 Solver.AddTrackedFunction(&F);
2023
2024 // Determine if we can track the function's arguments. If so, add the
2025 // function to the solver's set of argument-tracked functions.
2026 if (canTrackArgumentsInterprocedurally(&F)) {
2027 Solver.AddArgumentTrackedFunction(&F);
2028 continue;
2029 }
2030
2031 // Assume the function is called.
2032 Solver.MarkBlockExecutable(&F.front());
2033
2034 // Assume nothing about the incoming arguments.
2035 for (Argument &AI : F.args())
2036 Solver.markOverdefined(&AI);
2037 }
2038
2039 // Determine if we can track any of the module's global variables. If so, add
2040 // the global variables we can track to the solver's set of tracked global
2041 // variables.
2042 for (GlobalVariable &G : M.globals()) {
2043 G.removeDeadConstantUsers();
2044 if (canTrackGlobalVariableInterprocedurally(&G))
2045 Solver.TrackValueOfGlobalVariable(&G);
2046 }
2047
2048 // Solve for constants.
2049 bool ResolvedUndefs = true;
2050 Solver.Solve();
2051 while (ResolvedUndefs) {
2052 LLVM_DEBUG(dbgs() << "RESOLVING UNDEFS\n");
2053 ResolvedUndefs = false;
2054 for (Function &F : M)
2055 if (Solver.ResolvedUndefsIn(F)) {
2056 // We run Solve() after we resolved an undef in a function, because
2057 // we might deduce a fact that eliminates an undef in another function.
2058 Solver.Solve();
2059 ResolvedUndefs = true;
2060 }
2061 }
2062
2063 bool MadeChanges = false;
2064
2065 // Iterate over all of the instructions in the module, replacing them with
2066 // constants if we have found them to be of constant values.
2067
2068 for (Function &F : M) {
2069 if (F.isDeclaration())
2070 continue;
2071
2072 SmallVector<BasicBlock *, 512> BlocksToErase;
2073
2074 if (Solver.isBlockExecutable(&F.front()))
2075 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); AI != E;
2076 ++AI) {
2077 if (!AI->use_empty() && tryToReplaceWithConstant(Solver, &*AI)) {
2078 ++IPNumArgsElimed;
2079 continue;
2080 }
2081 }
2082
2083 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2084 if (!Solver.isBlockExecutable(&*BB)) {
2085 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
2086 ++NumDeadBlocks;
2087
2088 MadeChanges = true;
2089
2090 if (&*BB != &F.front())
2091 BlocksToErase.push_back(&*BB);
2092 continue;
2093 }
2094
2095 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) {
2096 Instruction *Inst = &*BI++;
2097 if (Inst->getType()->isVoidTy())
2098 continue;
2099 if (tryToReplaceWithConstant(Solver, Inst)) {
2100 if (Inst->isSafeToRemove())
2101 Inst->eraseFromParent();
2102 // Hey, we just changed something!
2103 MadeChanges = true;
2104 ++IPNumInstRemoved;
2105 }
2106 }
2107 }
2108
2109 DomTreeUpdater DTU = Solver.getDTU(F);
2110 // Change dead blocks to unreachable. We do it after replacing constants
2111 // in all executable blocks, because changeToUnreachable may remove PHI
2112 // nodes in executable blocks we found values for. The function's entry
2113 // block is not part of BlocksToErase, so we have to handle it separately.
2114 for (BasicBlock *BB : BlocksToErase) {
2115 NumInstRemoved +=
2116 changeToUnreachable(BB->getFirstNonPHI(), /*UseLLVMTrap=*/false,
2117 /*PreserveLCSSA=*/false, &DTU);
2118 }
2119 if (!Solver.isBlockExecutable(&F.front()))
2120 NumInstRemoved += changeToUnreachable(F.front().getFirstNonPHI(),
2121 /*UseLLVMTrap=*/false,
2122 /*PreserveLCSSA=*/false, &DTU);
2123
2124 // Now that all instructions in the function are constant folded,
2125 // use ConstantFoldTerminator to get rid of in-edges, record DT updates and
2126 // delete dead BBs.
2127 for (BasicBlock *DeadBB : BlocksToErase) {
2128 // If there are any PHI nodes in this successor, drop entries for BB now.
2129 for (Value::user_iterator UI = DeadBB->user_begin(),
2130 UE = DeadBB->user_end();
2131 UI != UE;) {
2132 // Grab the user and then increment the iterator early, as the user
2133 // will be deleted. Step past all adjacent uses from the same user.
2134 auto *I = dyn_cast<Instruction>(*UI);
2135 do { ++UI; } while (UI != UE && *UI == I);
2136
2137 // Ignore blockaddress users; BasicBlock's dtor will handle them.
2138 if (!I) continue;
2139
2140 // If we have forced an edge for an indeterminate value, then force the
2141 // terminator to fold to that edge.
2142 forceIndeterminateEdge(I, Solver);
2143 BasicBlock *InstBB = I->getParent();
2144 bool Folded = ConstantFoldTerminator(InstBB,
2145 /*DeleteDeadConditions=*/false,
2146 /*TLI=*/nullptr, &DTU);
2147 assert(Folded &&
2148 "Expect TermInst on constantint or blockaddress to be folded");
2149 (void) Folded;
2150 // If we folded the terminator to an unconditional branch to another
2151 // dead block, replace it with Unreachable, to avoid trying to fold that
2152 // branch again.
2153 BranchInst *BI = cast<BranchInst>(InstBB->getTerminator());
2154 if (BI && BI->isUnconditional() &&
2155 !Solver.isBlockExecutable(BI->getSuccessor(0))) {
2156 InstBB->getTerminator()->eraseFromParent();
2157 new UnreachableInst(InstBB->getContext(), InstBB);
2158 }
2159 }
2160 // Mark dead BB for deletion.
2161 DTU.deleteBB(DeadBB);
2162 }
2163
2164 for (BasicBlock &BB : F) {
2165 for (BasicBlock::iterator BI = BB.begin(), E = BB.end(); BI != E;) {
2166 Instruction *Inst = &*BI++;
2167 if (Solver.getPredicateInfoFor(Inst)) {
2168 if (auto *II = dyn_cast<IntrinsicInst>(Inst)) {
2169 if (II->getIntrinsicID() == Intrinsic::ssa_copy) {
2170 Value *Op = II->getOperand(0);
2171 Inst->replaceAllUsesWith(Op);
2172 Inst->eraseFromParent();
2173 }
2174 }
2175 }
2176 }
2177 }
2178 }
2179
2180 // If we inferred constant or undef return values for a function, we replaced
2181 // all call uses with the inferred value. This means we don't need to bother
2182 // actually returning anything from the function. Replace all return
2183 // instructions with return undef.
2184 //
2185 // Do this in two stages: first identify the functions we should process, then
2186 // actually zap their returns. This is important because we can only do this
2187 // if the address of the function isn't taken. In cases where a return is the
2188 // last use of a function, the order of processing functions would affect
2189 // whether other functions are optimizable.
2190 SmallVector<ReturnInst*, 8> ReturnsToZap;
2191
2192 const MapVector<Function*, LatticeVal> &RV = Solver.getTrackedRetVals();
2193 for (const auto &I : RV) {
2194 Function *F = I.first;
2195 if (I.second.isOverdefined() || F->getReturnType()->isVoidTy())
2196 continue;
2197 findReturnsToZap(*F, ReturnsToZap, Solver);
2198 }
2199
2200 for (auto F : Solver.getMRVFunctionsTracked()) {
2201 assert(F->getReturnType()->isStructTy() &&
2202 "The return type should be a struct");
2203 StructType *STy = cast<StructType>(F->getReturnType());
2204 if (Solver.isStructLatticeConstant(F, STy))
2205 findReturnsToZap(*F, ReturnsToZap, Solver);
2206 }
2207
2208 // Zap all returns which we've identified as zap to change.
2209 for (unsigned i = 0, e = ReturnsToZap.size(); i != e; ++i) {
2210 Function *F = ReturnsToZap[i]->getParent()->getParent();
2211 ReturnsToZap[i]->setOperand(0, UndefValue::get(F->getReturnType()));
2212 }
2213
2214 // If we inferred constant or undef values for globals variables, we can
2215 // delete the global and any stores that remain to it.
2216 const DenseMap<GlobalVariable*, LatticeVal> &TG = Solver.getTrackedGlobals();
2217 for (DenseMap<GlobalVariable*, LatticeVal>::const_iterator I = TG.begin(),
2218 E = TG.end(); I != E; ++I) {
2219 GlobalVariable *GV = I->first;
2220 assert(!I->second.isOverdefined() &&
2221 "Overdefined values should have been taken out of the map!");
2222 LLVM_DEBUG(dbgs() << "Found that GV '" << GV->getName()
2223 << "' is constant!\n");
2224 while (!GV->use_empty()) {
2225 StoreInst *SI = cast<StoreInst>(GV->user_back());
2226 SI->eraseFromParent();
2227 }
2228 M.getGlobalList().erase(GV);
2229 ++IPNumGlobalConst;
2230 }
2231
2232 return MadeChanges;
2233 }
2234