1 //===---- NewGVN.cpp - Global Value Numbering Pass --------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file implements the new LLVM's Global Value Numbering pass.
11 /// GVN partitions values computed by a function into congruence classes.
12 /// Values ending up in the same congruence class are guaranteed to be the same
13 /// for every execution of the program. In that respect, congruency is a
14 /// compile-time approximation of equivalence of values at runtime.
15 /// The algorithm implemented here uses a sparse formulation and it's based
16 /// on the ideas described in the paper:
17 /// "A Sparse Algorithm for Predicated Global Value Numbering" from
20 /// A brief overview of the algorithm: The algorithm is essentially the same as
21 /// the standard RPO value numbering algorithm (a good reference is the paper
22 /// "SCC based value numbering" by L. Taylor Simpson) with one major difference:
23 /// The RPO algorithm proceeds, on every iteration, to process every reachable
24 /// block and every instruction in that block. This is because the standard RPO
25 /// algorithm does not track what things have the same value number, it only
26 /// tracks what the value number of a given operation is (the mapping is
27 /// operation -> value number). Thus, when a value number of an operation
28 /// changes, it must reprocess everything to ensure all uses of a value number
29 /// get updated properly. In constrast, the sparse algorithm we use *also*
30 /// tracks what operations have a given value number (IE it also tracks the
31 /// reverse mapping from value number -> operations with that value number), so
32 /// that it only needs to reprocess the instructions that are affected when
33 /// something's value number changes. The vast majority of complexity and code
34 /// in this file is devoted to tracking what value numbers could change for what
35 /// instructions when various things happen. The rest of the algorithm is
36 /// devoted to performing symbolic evaluation, forward propagation, and
37 /// simplification of operations based on the value numbers deduced so far
39 /// In order to make the GVN mostly-complete, we use a technique derived from
40 /// "Detection of Redundant Expressions: A Complete and Polynomial-time
41 /// Algorithm in SSA" by R.R. Pai. The source of incompleteness in most SSA
42 /// based GVN algorithms is related to their inability to detect equivalence
43 /// between phi of ops (IE phi(a+b, c+d)) and op of phis (phi(a,c) + phi(b, d)).
44 /// We resolve this issue by generating the equivalent "phi of ops" form for
45 /// each op of phis we see, in a way that only takes polynomial time to resolve.
47 /// We also do not perform elimination by using any published algorithm. All
48 /// published algorithms are O(Instructions). Instead, we use a technique that
49 /// is O(number of operations with the same value number), enabling us to skip
50 /// trying to eliminate things that have unique value numbers.
51 //===----------------------------------------------------------------------===//
53 #include "llvm/Transforms/Scalar/NewGVN.h"
54 #include "llvm/ADT/BitVector.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/DenseSet.h"
57 #include "llvm/ADT/DepthFirstIterator.h"
58 #include "llvm/ADT/Hashing.h"
59 #include "llvm/ADT/MapVector.h"
60 #include "llvm/ADT/PostOrderIterator.h"
61 #include "llvm/ADT/STLExtras.h"
62 #include "llvm/ADT/SmallPtrSet.h"
63 #include "llvm/ADT/SmallSet.h"
64 #include "llvm/ADT/Statistic.h"
65 #include "llvm/ADT/TinyPtrVector.h"
66 #include "llvm/Analysis/AliasAnalysis.h"
67 #include "llvm/Analysis/AssumptionCache.h"
68 #include "llvm/Analysis/CFG.h"
69 #include "llvm/Analysis/CFGPrinter.h"
70 #include "llvm/Analysis/ConstantFolding.h"
71 #include "llvm/Analysis/GlobalsModRef.h"
72 #include "llvm/Analysis/InstructionSimplify.h"
73 #include "llvm/Analysis/MemoryBuiltins.h"
74 #include "llvm/Analysis/MemoryLocation.h"
75 #include "llvm/Analysis/MemorySSA.h"
76 #include "llvm/Analysis/TargetLibraryInfo.h"
77 #include "llvm/IR/DataLayout.h"
78 #include "llvm/IR/Dominators.h"
79 #include "llvm/IR/GlobalVariable.h"
80 #include "llvm/IR/IRBuilder.h"
81 #include "llvm/IR/IntrinsicInst.h"
82 #include "llvm/IR/LLVMContext.h"
83 #include "llvm/IR/Metadata.h"
84 #include "llvm/IR/PatternMatch.h"
85 #include "llvm/IR/Type.h"
86 #include "llvm/Support/Allocator.h"
87 #include "llvm/Support/CommandLine.h"
88 #include "llvm/Support/Debug.h"
89 #include "llvm/Support/DebugCounter.h"
90 #include "llvm/Transforms/Scalar.h"
91 #include "llvm/Transforms/Scalar/GVNExpression.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include "llvm/Transforms/Utils/Local.h"
94 #include "llvm/Transforms/Utils/PredicateInfo.h"
95 #include "llvm/Transforms/Utils/VNCoercion.h"
97 #include <unordered_map>
100 using namespace llvm;
101 using namespace PatternMatch;
102 using namespace llvm::GVNExpression;
103 using namespace llvm::VNCoercion;
104 #define DEBUG_TYPE "newgvn"
106 STATISTIC(NumGVNInstrDeleted, "Number of instructions deleted");
107 STATISTIC(NumGVNBlocksDeleted, "Number of blocks deleted");
108 STATISTIC(NumGVNOpsSimplified, "Number of Expressions simplified");
109 STATISTIC(NumGVNPhisAllSame, "Number of PHIs whos arguments are all the same");
110 STATISTIC(NumGVNMaxIterations,
111 "Maximum Number of iterations it took to converge GVN");
112 STATISTIC(NumGVNLeaderChanges, "Number of leader changes");
113 STATISTIC(NumGVNSortedLeaderChanges, "Number of sorted leader changes");
114 STATISTIC(NumGVNAvoidedSortedLeaderChanges,
115 "Number of avoided sorted leader changes");
116 STATISTIC(NumGVNDeadStores, "Number of redundant/dead stores eliminated");
117 STATISTIC(NumGVNPHIOfOpsCreated, "Number of PHI of ops created");
118 STATISTIC(NumGVNPHIOfOpsEliminations,
119 "Number of things eliminated using PHI of ops");
120 DEBUG_COUNTER(VNCounter, "newgvn-vn",
121 "Controls which instructions are value numbered")
122 DEBUG_COUNTER(PHIOfOpsCounter, "newgvn-phi",
123 "Controls which instructions we create phi of ops for")
124 // Currently store defining access refinement is too slow due to basicaa being
125 // egregiously slow. This flag lets us keep it working while we work on this
127 static cl::opt<bool> EnableStoreRefinement("enable-store-refinement",
128 cl::init(false), cl::Hidden);
130 //===----------------------------------------------------------------------===//
132 //===----------------------------------------------------------------------===//
136 namespace GVNExpression {
137 Expression::~Expression() = default;
138 BasicExpression::~BasicExpression() = default;
139 CallExpression::~CallExpression() = default;
140 LoadExpression::~LoadExpression() = default;
141 StoreExpression::~StoreExpression() = default;
142 AggregateValueExpression::~AggregateValueExpression() = default;
143 PHIExpression::~PHIExpression() = default;
147 // Tarjan's SCC finding algorithm with Nuutila's improvements
148 // SCCIterator is actually fairly complex for the simple thing we want.
149 // It also wants to hand us SCC's that are unrelated to the phi node we ask
150 // about, and have us process them there or risk redoing work.
151 // Graph traits over a filter iterator also doesn't work that well here.
152 // This SCC finder is specialized to walk use-def chains, and only follows
154 // not generic values (arguments, etc).
157 TarjanSCC() : Components(1) {}
159 void Start(const Instruction *Start) {
160 if (Root.lookup(Start) == 0)
164 const SmallPtrSetImpl<const Value *> &getComponentFor(const Value *V) const {
165 unsigned ComponentID = ValueToComponent.lookup(V);
167 assert(ComponentID > 0 &&
168 "Asking for a component for a value we never processed");
169 return Components[ComponentID];
173 void FindSCC(const Instruction *I) {
175 // Store the DFS Number we had before it possibly gets incremented.
176 unsigned int OurDFS = DFSNum;
177 for (auto &Op : I->operands()) {
178 if (auto *InstOp = dyn_cast<Instruction>(Op)) {
179 if (Root.lookup(Op) == 0)
181 if (!InComponent.count(Op))
182 Root[I] = std::min(Root.lookup(I), Root.lookup(Op));
185 // See if we really were the root of a component, by seeing if we still have
186 // our DFSNumber. If we do, we are the root of the component, and we have
187 // completed a component. If we do not, we are not the root of a component,
188 // and belong on the component stack.
189 if (Root.lookup(I) == OurDFS) {
190 unsigned ComponentID = Components.size();
191 Components.resize(Components.size() + 1);
192 auto &Component = Components.back();
194 DEBUG(dbgs() << "Component root is " << *I << "\n");
195 InComponent.insert(I);
196 ValueToComponent[I] = ComponentID;
197 // Pop a component off the stack and label it.
198 while (!Stack.empty() && Root.lookup(Stack.back()) >= OurDFS) {
199 auto *Member = Stack.back();
200 DEBUG(dbgs() << "Component member is " << *Member << "\n");
201 Component.insert(Member);
202 InComponent.insert(Member);
203 ValueToComponent[Member] = ComponentID;
207 // Part of a component, push to stack
211 unsigned int DFSNum = 1;
212 SmallPtrSet<const Value *, 8> InComponent;
213 DenseMap<const Value *, unsigned int> Root;
214 SmallVector<const Value *, 8> Stack;
215 // Store the components as vector of ptr sets, because we need the topo order
216 // of SCC's, but not individual member order
217 SmallVector<SmallPtrSet<const Value *, 8>, 8> Components;
218 DenseMap<const Value *, unsigned> ValueToComponent;
220 // Congruence classes represent the set of expressions/instructions
221 // that are all the same *during some scope in the function*.
222 // That is, because of the way we perform equality propagation, and
223 // because of memory value numbering, it is not correct to assume
224 // you can willy-nilly replace any member with any other at any
225 // point in the function.
227 // For any Value in the Member set, it is valid to replace any dominated member
230 // Every congruence class has a leader, and the leader is used to symbolize
231 // instructions in a canonical way (IE every operand of an instruction that is a
232 // member of the same congruence class will always be replaced with leader
233 // during symbolization). To simplify symbolization, we keep the leader as a
234 // constant if class can be proved to be a constant value. Otherwise, the
235 // leader is the member of the value set with the smallest DFS number. Each
236 // congruence class also has a defining expression, though the expression may be
237 // null. If it exists, it can be used for forward propagation and reassociation
240 // For memory, we also track a representative MemoryAccess, and a set of memory
241 // members for MemoryPhis (which have no real instructions). Note that for
242 // memory, it seems tempting to try to split the memory members into a
243 // MemoryCongruenceClass or something. Unfortunately, this does not work
244 // easily. The value numbering of a given memory expression depends on the
245 // leader of the memory congruence class, and the leader of memory congruence
246 // class depends on the value numbering of a given memory expression. This
247 // leads to wasted propagation, and in some cases, missed optimization. For
248 // example: If we had value numbered two stores together before, but now do not,
249 // we move them to a new value congruence class. This in turn will move at one
250 // of the memorydefs to a new memory congruence class. Which in turn, affects
251 // the value numbering of the stores we just value numbered (because the memory
252 // congruence class is part of the value number). So while theoretically
253 // possible to split them up, it turns out to be *incredibly* complicated to get
254 // it to work right, because of the interdependency. While structurally
255 // slightly messier, it is algorithmically much simpler and faster to do what we
256 // do here, and track them both at once in the same class.
257 // Note: The default iterators for this class iterate over values
258 class CongruenceClass {
260 using MemberType = Value;
261 using MemberSet = SmallPtrSet<MemberType *, 4>;
262 using MemoryMemberType = MemoryPhi;
263 using MemoryMemberSet = SmallPtrSet<const MemoryMemberType *, 2>;
265 explicit CongruenceClass(unsigned ID) : ID(ID) {}
266 CongruenceClass(unsigned ID, Value *Leader, const Expression *E)
267 : ID(ID), RepLeader(Leader), DefiningExpr(E) {}
268 unsigned getID() const { return ID; }
269 // True if this class has no members left. This is mainly used for assertion
270 // purposes, and for skipping empty classes.
271 bool isDead() const {
272 // If it's both dead from a value perspective, and dead from a memory
273 // perspective, it's really dead.
274 return empty() && memory_empty();
277 Value *getLeader() const { return RepLeader; }
278 void setLeader(Value *Leader) { RepLeader = Leader; }
279 const std::pair<Value *, unsigned int> &getNextLeader() const {
282 void resetNextLeader() { NextLeader = {nullptr, ~0}; }
284 void addPossibleNextLeader(std::pair<Value *, unsigned int> LeaderPair) {
285 if (LeaderPair.second < NextLeader.second)
286 NextLeader = LeaderPair;
289 Value *getStoredValue() const { return RepStoredValue; }
290 void setStoredValue(Value *Leader) { RepStoredValue = Leader; }
291 const MemoryAccess *getMemoryLeader() const { return RepMemoryAccess; }
292 void setMemoryLeader(const MemoryAccess *Leader) { RepMemoryAccess = Leader; }
294 // Forward propagation info
295 const Expression *getDefiningExpr() const { return DefiningExpr; }
298 bool empty() const { return Members.empty(); }
299 unsigned size() const { return Members.size(); }
300 MemberSet::const_iterator begin() const { return Members.begin(); }
301 MemberSet::const_iterator end() const { return Members.end(); }
302 void insert(MemberType *M) { Members.insert(M); }
303 void erase(MemberType *M) { Members.erase(M); }
304 void swap(MemberSet &Other) { Members.swap(Other); }
307 bool memory_empty() const { return MemoryMembers.empty(); }
308 unsigned memory_size() const { return MemoryMembers.size(); }
309 MemoryMemberSet::const_iterator memory_begin() const {
310 return MemoryMembers.begin();
312 MemoryMemberSet::const_iterator memory_end() const {
313 return MemoryMembers.end();
315 iterator_range<MemoryMemberSet::const_iterator> memory() const {
316 return make_range(memory_begin(), memory_end());
318 void memory_insert(const MemoryMemberType *M) { MemoryMembers.insert(M); }
319 void memory_erase(const MemoryMemberType *M) { MemoryMembers.erase(M); }
322 unsigned getStoreCount() const { return StoreCount; }
323 void incStoreCount() { ++StoreCount; }
324 void decStoreCount() {
325 assert(StoreCount != 0 && "Store count went negative");
329 // True if this class has no memory members.
330 bool definesNoMemory() const { return StoreCount == 0 && memory_empty(); }
332 // Return true if two congruence classes are equivalent to each other. This
334 // that every field but the ID number and the dead field are equivalent.
335 bool isEquivalentTo(const CongruenceClass *Other) const {
341 if (std::tie(StoreCount, RepLeader, RepStoredValue, RepMemoryAccess) !=
342 std::tie(Other->StoreCount, Other->RepLeader, Other->RepStoredValue,
343 Other->RepMemoryAccess))
345 if (DefiningExpr != Other->DefiningExpr)
346 if (!DefiningExpr || !Other->DefiningExpr ||
347 *DefiningExpr != *Other->DefiningExpr)
349 // We need some ordered set
350 std::set<Value *> AMembers(Members.begin(), Members.end());
351 std::set<Value *> BMembers(Members.begin(), Members.end());
352 return AMembers == BMembers;
357 // Representative leader.
358 Value *RepLeader = nullptr;
359 // The most dominating leader after our current leader, because the member set
360 // is not sorted and is expensive to keep sorted all the time.
361 std::pair<Value *, unsigned int> NextLeader = {nullptr, ~0U};
362 // If this is represented by a store, the value of the store.
363 Value *RepStoredValue = nullptr;
364 // If this class contains MemoryDefs or MemoryPhis, this is the leading memory
366 const MemoryAccess *RepMemoryAccess = nullptr;
367 // Defining Expression.
368 const Expression *DefiningExpr = nullptr;
369 // Actual members of this class.
371 // This is the set of MemoryPhis that exist in the class. MemoryDefs and
372 // MemoryUses have real instructions representing them, so we only need to
373 // track MemoryPhis here.
374 MemoryMemberSet MemoryMembers;
375 // Number of stores in this congruence class.
376 // This is used so we can detect store equivalence changes properly.
381 template <> struct DenseMapInfo<const Expression *> {
382 static const Expression *getEmptyKey() {
383 auto Val = static_cast<uintptr_t>(-1);
384 Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
385 return reinterpret_cast<const Expression *>(Val);
387 static const Expression *getTombstoneKey() {
388 auto Val = static_cast<uintptr_t>(~1U);
389 Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
390 return reinterpret_cast<const Expression *>(Val);
392 static unsigned getHashValue(const Expression *E) {
393 return static_cast<unsigned>(E->getComputedHash());
395 static bool isEqual(const Expression *LHS, const Expression *RHS) {
398 if (LHS == getTombstoneKey() || RHS == getTombstoneKey() ||
399 LHS == getEmptyKey() || RHS == getEmptyKey())
401 // Compare hashes before equality. This is *not* what the hashtable does,
402 // since it is computing it modulo the number of buckets, whereas we are
403 // using the full hash keyspace. Since the hashes are precomputed, this
404 // check is *much* faster than equality.
405 if (LHS->getComputedHash() != RHS->getComputedHash())
410 } // end namespace llvm
416 const TargetLibraryInfo *TLI;
419 MemorySSAWalker *MSSAWalker;
420 const DataLayout &DL;
421 std::unique_ptr<PredicateInfo> PredInfo;
423 // These are the only two things the create* functions should have
424 // side-effects on due to allocating memory.
425 mutable BumpPtrAllocator ExpressionAllocator;
426 mutable ArrayRecycler<Value *> ArgRecycler;
427 mutable TarjanSCC SCCFinder;
428 const SimplifyQuery SQ;
430 // Number of function arguments, used by ranking
431 unsigned int NumFuncArgs;
433 // RPOOrdering of basic blocks
434 DenseMap<const DomTreeNode *, unsigned> RPOOrdering;
436 // Congruence class info.
438 // This class is called INITIAL in the paper. It is the class everything
439 // startsout in, and represents any value. Being an optimistic analysis,
440 // anything in the TOP class has the value TOP, which is indeterminate and
441 // equivalent to everything.
442 CongruenceClass *TOPClass;
443 std::vector<CongruenceClass *> CongruenceClasses;
444 unsigned NextCongruenceNum;
447 DenseMap<Value *, CongruenceClass *> ValueToClass;
448 DenseMap<Value *, const Expression *> ValueToExpression;
449 // Value PHI handling, used to make equivalence between phi(op, op) and
451 // These mappings just store various data that would normally be part of the
453 DenseSet<const Instruction *> PHINodeUses;
454 // Map a temporary instruction we created to a parent block.
455 DenseMap<const Value *, BasicBlock *> TempToBlock;
456 // Map between the temporary phis we created and the real instructions they
457 // are known equivalent to.
458 DenseMap<const Value *, PHINode *> RealToTemp;
459 // In order to know when we should re-process instructions that have
460 // phi-of-ops, we track the set of expressions that they needed as
461 // leaders. When we discover new leaders for those expressions, we process the
462 // associated phi-of-op instructions again in case they have changed. The
463 // other way they may change is if they had leaders, and those leaders
464 // disappear. However, at the point they have leaders, there are uses of the
465 // relevant operands in the created phi node, and so they will get reprocessed
466 // through the normal user marking we perform.
467 mutable DenseMap<const Value *, SmallPtrSet<Value *, 2>> AdditionalUsers;
468 DenseMap<const Expression *, SmallPtrSet<Instruction *, 2>>
469 ExpressionToPhiOfOps;
470 // Map from basic block to the temporary operations we created
471 DenseMap<const BasicBlock *, SmallVector<PHINode *, 8>> PHIOfOpsPHIs;
472 // Map from temporary operation to MemoryAccess.
473 DenseMap<const Instruction *, MemoryUseOrDef *> TempToMemory;
474 // Set of all temporary instructions we created.
475 DenseSet<Instruction *> AllTempInstructions;
477 // Mapping from predicate info we used to the instructions we used it with.
478 // In order to correctly ensure propagation, we must keep track of what
479 // comparisons we used, so that when the values of the comparisons change, we
480 // propagate the information to the places we used the comparison.
481 mutable DenseMap<const Value *, SmallPtrSet<Instruction *, 2>>
483 // the same reasoning as PredicateToUsers. When we skip MemoryAccesses for
484 // stores, we no longer can rely solely on the def-use chains of MemorySSA.
485 mutable DenseMap<const MemoryAccess *, SmallPtrSet<MemoryAccess *, 2>>
488 // A table storing which memorydefs/phis represent a memory state provably
489 // equivalent to another memory state.
490 // We could use the congruence class machinery, but the MemoryAccess's are
491 // abstract memory states, so they can only ever be equivalent to each other,
492 // and not to constants, etc.
493 DenseMap<const MemoryAccess *, CongruenceClass *> MemoryAccessToClass;
495 // We could, if we wanted, build MemoryPhiExpressions and
496 // MemoryVariableExpressions, etc, and value number them the same way we value
497 // number phi expressions. For the moment, this seems like overkill. They
498 // can only exist in one of three states: they can be TOP (equal to
499 // everything), Equivalent to something else, or unique. Because we do not
500 // create expressions for them, we need to simulate leader change not just
501 // when they change class, but when they change state. Note: We can do the
502 // same thing for phis, and avoid having phi expressions if we wanted, We
503 // should eventually unify in one direction or the other, so this is a little
504 // bit of an experiment in which turns out easier to maintain.
505 enum MemoryPhiState { MPS_Invalid, MPS_TOP, MPS_Equivalent, MPS_Unique };
506 DenseMap<const MemoryPhi *, MemoryPhiState> MemoryPhiState;
508 enum InstCycleState { ICS_Unknown, ICS_CycleFree, ICS_Cycle };
509 mutable DenseMap<const Instruction *, InstCycleState> InstCycleState;
510 // Expression to class mapping.
511 using ExpressionClassMap = DenseMap<const Expression *, CongruenceClass *>;
512 ExpressionClassMap ExpressionToClass;
514 // We have a single expression that represents currently DeadExpressions.
515 // For dead expressions we can prove will stay dead, we mark them with
516 // DFS number zero. However, it's possible in the case of phi nodes
517 // for us to assume/prove all arguments are dead during fixpointing.
518 // We use DeadExpression for that case.
519 DeadExpression *SingletonDeadExpression = nullptr;
521 // Which values have changed as a result of leader changes.
522 SmallPtrSet<Value *, 8> LeaderChanges;
524 // Reachability info.
525 using BlockEdge = BasicBlockEdge;
526 DenseSet<BlockEdge> ReachableEdges;
527 SmallPtrSet<const BasicBlock *, 8> ReachableBlocks;
529 // This is a bitvector because, on larger functions, we may have
530 // thousands of touched instructions at once (entire blocks,
531 // instructions with hundreds of uses, etc). Even with optimization
532 // for when we mark whole blocks as touched, when this was a
533 // SmallPtrSet or DenseSet, for some functions, we spent >20% of all
534 // the time in GVN just managing this list. The bitvector, on the
535 // other hand, efficiently supports test/set/clear of both
536 // individual and ranges, as well as "find next element" This
537 // enables us to use it as a worklist with essentially 0 cost.
538 BitVector TouchedInstructions;
540 DenseMap<const BasicBlock *, std::pair<unsigned, unsigned>> BlockInstRange;
543 // Debugging for how many times each block and instruction got processed.
544 DenseMap<const Value *, unsigned> ProcessedCount;
548 // This contains a mapping from Instructions to DFS numbers.
549 // The numbering starts at 1. An instruction with DFS number zero
550 // means that the instruction is dead.
551 DenseMap<const Value *, unsigned> InstrDFS;
553 // This contains the mapping DFS numbers to instructions.
554 SmallVector<Value *, 32> DFSToInstr;
557 SmallPtrSet<Instruction *, 8> InstructionsToErase;
560 NewGVN(Function &F, DominatorTree *DT, AssumptionCache *AC,
561 TargetLibraryInfo *TLI, AliasAnalysis *AA, MemorySSA *MSSA,
562 const DataLayout &DL)
563 : F(F), DT(DT), TLI(TLI), AA(AA), MSSA(MSSA), DL(DL),
564 PredInfo(make_unique<PredicateInfo>(F, *DT, *AC)), SQ(DL, TLI, DT, AC) {
569 // Expression handling.
570 const Expression *createExpression(Instruction *) const;
571 const Expression *createBinaryExpression(unsigned, Type *, Value *,
573 PHIExpression *createPHIExpression(Instruction *, bool &HasBackEdge,
574 bool &OriginalOpsConstant) const;
575 const DeadExpression *createDeadExpression() const;
576 const VariableExpression *createVariableExpression(Value *) const;
577 const ConstantExpression *createConstantExpression(Constant *) const;
578 const Expression *createVariableOrConstant(Value *V) const;
579 const UnknownExpression *createUnknownExpression(Instruction *) const;
580 const StoreExpression *createStoreExpression(StoreInst *,
581 const MemoryAccess *) const;
582 LoadExpression *createLoadExpression(Type *, Value *, LoadInst *,
583 const MemoryAccess *) const;
584 const CallExpression *createCallExpression(CallInst *,
585 const MemoryAccess *) const;
586 const AggregateValueExpression *
587 createAggregateValueExpression(Instruction *) const;
588 bool setBasicExpressionInfo(Instruction *, BasicExpression *) const;
590 // Congruence class handling.
591 CongruenceClass *createCongruenceClass(Value *Leader, const Expression *E) {
592 auto *result = new CongruenceClass(NextCongruenceNum++, Leader, E);
593 CongruenceClasses.emplace_back(result);
597 CongruenceClass *createMemoryClass(MemoryAccess *MA) {
598 auto *CC = createCongruenceClass(nullptr, nullptr);
599 CC->setMemoryLeader(MA);
602 CongruenceClass *ensureLeaderOfMemoryClass(MemoryAccess *MA) {
603 auto *CC = getMemoryClass(MA);
604 if (CC->getMemoryLeader() != MA)
605 CC = createMemoryClass(MA);
609 CongruenceClass *createSingletonCongruenceClass(Value *Member) {
610 CongruenceClass *CClass = createCongruenceClass(Member, nullptr);
611 CClass->insert(Member);
612 ValueToClass[Member] = CClass;
615 void initializeCongruenceClasses(Function &F);
616 const Expression *makePossiblePhiOfOps(Instruction *, bool,
617 SmallPtrSetImpl<Value *> &);
618 void addPhiOfOps(PHINode *Op, BasicBlock *BB, Instruction *ExistingValue);
620 // Value number an Instruction or MemoryPhi.
621 void valueNumberMemoryPhi(MemoryPhi *);
622 void valueNumberInstruction(Instruction *);
624 // Symbolic evaluation.
625 const Expression *checkSimplificationResults(Expression *, Instruction *,
627 const Expression *performSymbolicEvaluation(Value *,
628 SmallPtrSetImpl<Value *> &) const;
629 const Expression *performSymbolicLoadCoercion(Type *, Value *, LoadInst *,
631 MemoryAccess *) const;
632 const Expression *performSymbolicLoadEvaluation(Instruction *) const;
633 const Expression *performSymbolicStoreEvaluation(Instruction *) const;
634 const Expression *performSymbolicCallEvaluation(Instruction *) const;
635 const Expression *performSymbolicPHIEvaluation(Instruction *) const;
636 const Expression *performSymbolicAggrValueEvaluation(Instruction *) const;
637 const Expression *performSymbolicCmpEvaluation(Instruction *) const;
638 const Expression *performSymbolicPredicateInfoEvaluation(Instruction *) const;
640 // Congruence finding.
641 bool someEquivalentDominates(const Instruction *, const Instruction *) const;
642 Value *lookupOperandLeader(Value *) const;
643 void performCongruenceFinding(Instruction *, const Expression *);
644 void moveValueToNewCongruenceClass(Instruction *, const Expression *,
645 CongruenceClass *, CongruenceClass *);
646 void moveMemoryToNewCongruenceClass(Instruction *, MemoryAccess *,
647 CongruenceClass *, CongruenceClass *);
648 Value *getNextValueLeader(CongruenceClass *) const;
649 const MemoryAccess *getNextMemoryLeader(CongruenceClass *) const;
650 bool setMemoryClass(const MemoryAccess *From, CongruenceClass *To);
651 CongruenceClass *getMemoryClass(const MemoryAccess *MA) const;
652 const MemoryAccess *lookupMemoryLeader(const MemoryAccess *) const;
653 bool isMemoryAccessTOP(const MemoryAccess *) const;
656 unsigned int getRank(const Value *) const;
657 bool shouldSwapOperands(const Value *, const Value *) const;
659 // Reachability handling.
660 void updateReachableEdge(BasicBlock *, BasicBlock *);
661 void processOutgoingEdges(TerminatorInst *, BasicBlock *);
662 Value *findConditionEquivalence(Value *) const;
666 void convertClassToDFSOrdered(const CongruenceClass &,
667 SmallVectorImpl<ValueDFS> &,
668 DenseMap<const Value *, unsigned int> &,
669 SmallPtrSetImpl<Instruction *> &) const;
670 void convertClassToLoadsAndStores(const CongruenceClass &,
671 SmallVectorImpl<ValueDFS> &) const;
673 bool eliminateInstructions(Function &);
674 void replaceInstruction(Instruction *, Value *);
675 void markInstructionForDeletion(Instruction *);
676 void deleteInstructionsInBlock(BasicBlock *);
677 Value *findPhiOfOpsLeader(const Expression *E, const BasicBlock *BB) const;
679 // New instruction creation.
680 void handleNewInstruction(Instruction *){};
682 // Various instruction touch utilities
683 template <typename Map, typename KeyType, typename Func>
684 void for_each_found(Map &, const KeyType &, Func);
685 template <typename Map, typename KeyType>
686 void touchAndErase(Map &, const KeyType &);
687 void markUsersTouched(Value *);
688 void markMemoryUsersTouched(const MemoryAccess *);
689 void markMemoryDefTouched(const MemoryAccess *);
690 void markPredicateUsersTouched(Instruction *);
691 void markValueLeaderChangeTouched(CongruenceClass *CC);
692 void markMemoryLeaderChangeTouched(CongruenceClass *CC);
693 void markPhiOfOpsChanged(const Expression *E);
694 void addPredicateUsers(const PredicateBase *, Instruction *) const;
695 void addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const;
696 void addAdditionalUsers(Value *To, Value *User) const;
698 // Main loop of value numbering
699 void iterateTouchedInstructions();
702 void cleanupTables();
703 std::pair<unsigned, unsigned> assignDFSNumbers(BasicBlock *, unsigned);
704 void updateProcessedCount(const Value *V);
705 void verifyMemoryCongruency() const;
706 void verifyIterationSettled(Function &F);
707 void verifyStoreExpressions() const;
708 bool singleReachablePHIPath(SmallPtrSet<const MemoryAccess *, 8> &,
709 const MemoryAccess *, const MemoryAccess *) const;
710 BasicBlock *getBlockForValue(Value *V) const;
711 void deleteExpression(const Expression *E) const;
712 MemoryUseOrDef *getMemoryAccess(const Instruction *) const;
713 MemoryAccess *getDefiningAccess(const MemoryAccess *) const;
714 MemoryPhi *getMemoryAccess(const BasicBlock *) const;
715 template <class T, class Range> T *getMinDFSOfRange(const Range &) const;
716 unsigned InstrToDFSNum(const Value *V) const {
717 assert(isa<Instruction>(V) && "This should not be used for MemoryAccesses");
718 return InstrDFS.lookup(V);
721 unsigned InstrToDFSNum(const MemoryAccess *MA) const {
722 return MemoryToDFSNum(MA);
724 Value *InstrFromDFSNum(unsigned DFSNum) { return DFSToInstr[DFSNum]; }
725 // Given a MemoryAccess, return the relevant instruction DFS number. Note:
726 // This deliberately takes a value so it can be used with Use's, which will
727 // auto-convert to Value's but not to MemoryAccess's.
728 unsigned MemoryToDFSNum(const Value *MA) const {
729 assert(isa<MemoryAccess>(MA) &&
730 "This should not be used with instructions");
731 return isa<MemoryUseOrDef>(MA)
732 ? InstrToDFSNum(cast<MemoryUseOrDef>(MA)->getMemoryInst())
733 : InstrDFS.lookup(MA);
735 bool isCycleFree(const Instruction *) const;
736 bool isBackedge(BasicBlock *From, BasicBlock *To) const;
737 // Debug counter info. When verifying, we have to reset the value numbering
738 // debug counter to the same state it started in to get the same results.
739 std::pair<int, int> StartingVNCounter;
741 } // end anonymous namespace
743 template <typename T>
744 static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS) {
745 if (!isa<LoadExpression>(RHS) && !isa<StoreExpression>(RHS))
747 return LHS.MemoryExpression::equals(RHS);
750 bool LoadExpression::equals(const Expression &Other) const {
751 return equalsLoadStoreHelper(*this, Other);
754 bool StoreExpression::equals(const Expression &Other) const {
755 if (!equalsLoadStoreHelper(*this, Other))
757 // Make sure that store vs store includes the value operand.
758 if (const auto *S = dyn_cast<StoreExpression>(&Other))
759 if (getStoredValue() != S->getStoredValue())
764 // Determine if the edge From->To is a backedge
765 bool NewGVN::isBackedge(BasicBlock *From, BasicBlock *To) const {
768 auto *FromDTN = DT->getNode(From);
769 auto *ToDTN = DT->getNode(To);
770 return RPOOrdering.lookup(FromDTN) >= RPOOrdering.lookup(ToDTN);
774 static std::string getBlockName(const BasicBlock *B) {
775 return DOTGraphTraits<const Function *>::getSimpleNodeLabel(B, nullptr);
779 // Get a MemoryAccess for an instruction, fake or real.
780 MemoryUseOrDef *NewGVN::getMemoryAccess(const Instruction *I) const {
781 auto *Result = MSSA->getMemoryAccess(I);
782 return Result ? Result : TempToMemory.lookup(I);
785 // Get a MemoryPhi for a basic block. These are all real.
786 MemoryPhi *NewGVN::getMemoryAccess(const BasicBlock *BB) const {
787 return MSSA->getMemoryAccess(BB);
790 // Get the basic block from an instruction/memory value.
791 BasicBlock *NewGVN::getBlockForValue(Value *V) const {
792 if (auto *I = dyn_cast<Instruction>(V)) {
793 auto *Parent = I->getParent();
796 Parent = TempToBlock.lookup(V);
797 assert(Parent && "Every fake instruction should have a block");
801 auto *MP = dyn_cast<MemoryPhi>(V);
802 assert(MP && "Should have been an instruction or a MemoryPhi");
803 return MP->getBlock();
806 // Delete a definitely dead expression, so it can be reused by the expression
807 // allocator. Some of these are not in creation functions, so we have to accept
809 void NewGVN::deleteExpression(const Expression *E) const {
810 assert(isa<BasicExpression>(E));
811 auto *BE = cast<BasicExpression>(E);
812 const_cast<BasicExpression *>(BE)->deallocateOperands(ArgRecycler);
813 ExpressionAllocator.Deallocate(E);
815 PHIExpression *NewGVN::createPHIExpression(Instruction *I, bool &HasBackedge,
816 bool &OriginalOpsConstant) const {
817 BasicBlock *PHIBlock = getBlockForValue(I);
818 auto *PN = cast<PHINode>(I);
820 new (ExpressionAllocator) PHIExpression(PN->getNumOperands(), PHIBlock);
822 E->allocateOperands(ArgRecycler, ExpressionAllocator);
823 E->setType(I->getType());
824 E->setOpcode(I->getOpcode());
826 // NewGVN assumes the operands of a PHI node are in a consistent order across
827 // PHIs. LLVM doesn't seem to always guarantee this. While we need to fix
828 // this in LLVM at some point we don't want GVN to find wrong congruences.
829 // Therefore, here we sort uses in predecessor order.
830 // We're sorting the values by pointer. In theory this might be cause of
831 // non-determinism, but here we don't rely on the ordering for anything
832 // significant, e.g. we don't create new instructions based on it so we're
834 SmallVector<const Use *, 4> PHIOperands;
835 for (const Use &U : PN->operands())
836 PHIOperands.push_back(&U);
837 std::sort(PHIOperands.begin(), PHIOperands.end(),
838 [&](const Use *U1, const Use *U2) {
839 return PN->getIncomingBlock(*U1) < PN->getIncomingBlock(*U2);
842 // Filter out unreachable phi operands.
843 auto Filtered = make_filter_range(PHIOperands, [&](const Use *U) {
846 if (!ReachableEdges.count({PN->getIncomingBlock(*U), PHIBlock}))
848 // Things in TOPClass are equivalent to everything.
849 if (ValueToClass.lookup(*U) == TOPClass)
853 std::transform(Filtered.begin(), Filtered.end(), op_inserter(E),
854 [&](const Use *U) -> Value * {
855 auto *BB = PN->getIncomingBlock(*U);
856 HasBackedge = HasBackedge || isBackedge(BB, PHIBlock);
857 OriginalOpsConstant =
858 OriginalOpsConstant && isa<Constant>(*U);
859 return lookupOperandLeader(*U);
864 // Set basic expression info (Arguments, type, opcode) for Expression
865 // E from Instruction I in block B.
866 bool NewGVN::setBasicExpressionInfo(Instruction *I, BasicExpression *E) const {
867 bool AllConstant = true;
868 if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
869 E->setType(GEP->getSourceElementType());
871 E->setType(I->getType());
872 E->setOpcode(I->getOpcode());
873 E->allocateOperands(ArgRecycler, ExpressionAllocator);
875 // Transform the operand array into an operand leader array, and keep track of
876 // whether all members are constant.
877 std::transform(I->op_begin(), I->op_end(), op_inserter(E), [&](Value *O) {
878 auto Operand = lookupOperandLeader(O);
879 AllConstant = AllConstant && isa<Constant>(Operand);
886 const Expression *NewGVN::createBinaryExpression(unsigned Opcode, Type *T,
889 auto *E = new (ExpressionAllocator) BasicExpression(2);
892 E->setOpcode(Opcode);
893 E->allocateOperands(ArgRecycler, ExpressionAllocator);
894 if (Instruction::isCommutative(Opcode)) {
895 // Ensure that commutative instructions that only differ by a permutation
896 // of their operands get the same value number by sorting the operand value
897 // numbers. Since all commutative instructions have two operands it is more
898 // efficient to sort by hand rather than using, say, std::sort.
899 if (shouldSwapOperands(Arg1, Arg2))
900 std::swap(Arg1, Arg2);
902 E->op_push_back(lookupOperandLeader(Arg1));
903 E->op_push_back(lookupOperandLeader(Arg2));
905 Value *V = SimplifyBinOp(Opcode, E->getOperand(0), E->getOperand(1), SQ);
906 if (const Expression *SimplifiedE = checkSimplificationResults(E, nullptr, V))
911 // Take a Value returned by simplification of Expression E/Instruction
912 // I, and see if it resulted in a simpler expression. If so, return
914 // TODO: Once finished, this should not take an Instruction, we only
915 // use it for printing.
916 const Expression *NewGVN::checkSimplificationResults(Expression *E,
921 if (auto *C = dyn_cast<Constant>(V)) {
923 DEBUG(dbgs() << "Simplified " << *I << " to "
924 << " constant " << *C << "\n");
925 NumGVNOpsSimplified++;
926 assert(isa<BasicExpression>(E) &&
927 "We should always have had a basic expression here");
929 return createConstantExpression(C);
930 } else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
932 DEBUG(dbgs() << "Simplified " << *I << " to "
933 << " variable " << *V << "\n");
935 return createVariableExpression(V);
938 CongruenceClass *CC = ValueToClass.lookup(V);
939 if (CC && CC->getDefiningExpr()) {
940 // If we simplified to something else, we need to communicate
941 // that we're users of the value we simplified to.
943 // Don't add temporary instructions to the user lists.
944 if (!AllTempInstructions.count(I))
945 addAdditionalUsers(V, I);
949 DEBUG(dbgs() << "Simplified " << *I << " to "
950 << " expression " << *CC->getDefiningExpr() << "\n");
951 NumGVNOpsSimplified++;
953 return CC->getDefiningExpr();
958 const Expression *NewGVN::createExpression(Instruction *I) const {
959 auto *E = new (ExpressionAllocator) BasicExpression(I->getNumOperands());
961 bool AllConstant = setBasicExpressionInfo(I, E);
963 if (I->isCommutative()) {
964 // Ensure that commutative instructions that only differ by a permutation
965 // of their operands get the same value number by sorting the operand value
966 // numbers. Since all commutative instructions have two operands it is more
967 // efficient to sort by hand rather than using, say, std::sort.
968 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
969 if (shouldSwapOperands(E->getOperand(0), E->getOperand(1)))
970 E->swapOperands(0, 1);
973 // Perform simplificaiton
974 // TODO: Right now we only check to see if we get a constant result.
975 // We may get a less than constant, but still better, result for
980 // We should handle this by simply rewriting the expression.
981 if (auto *CI = dyn_cast<CmpInst>(I)) {
982 // Sort the operand value numbers so x<y and y>x get the same value
984 CmpInst::Predicate Predicate = CI->getPredicate();
985 if (shouldSwapOperands(E->getOperand(0), E->getOperand(1))) {
986 E->swapOperands(0, 1);
987 Predicate = CmpInst::getSwappedPredicate(Predicate);
989 E->setOpcode((CI->getOpcode() << 8) | Predicate);
990 // TODO: 25% of our time is spent in SimplifyCmpInst with pointer operands
991 assert(I->getOperand(0)->getType() == I->getOperand(1)->getType() &&
992 "Wrong types on cmp instruction");
993 assert((E->getOperand(0)->getType() == I->getOperand(0)->getType() &&
994 E->getOperand(1)->getType() == I->getOperand(1)->getType()));
996 SimplifyCmpInst(Predicate, E->getOperand(0), E->getOperand(1), SQ);
997 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
999 } else if (isa<SelectInst>(I)) {
1000 if (isa<Constant>(E->getOperand(0)) ||
1001 E->getOperand(0) == E->getOperand(1)) {
1002 assert(E->getOperand(1)->getType() == I->getOperand(1)->getType() &&
1003 E->getOperand(2)->getType() == I->getOperand(2)->getType());
1004 Value *V = SimplifySelectInst(E->getOperand(0), E->getOperand(1),
1005 E->getOperand(2), SQ);
1006 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1009 } else if (I->isBinaryOp()) {
1011 SimplifyBinOp(E->getOpcode(), E->getOperand(0), E->getOperand(1), SQ);
1012 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1014 } else if (auto *BI = dyn_cast<BitCastInst>(I)) {
1016 SimplifyCastInst(BI->getOpcode(), BI->getOperand(0), BI->getType(), SQ);
1017 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1019 } else if (isa<GetElementPtrInst>(I)) {
1020 Value *V = SimplifyGEPInst(
1021 E->getType(), ArrayRef<Value *>(E->op_begin(), E->op_end()), SQ);
1022 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1024 } else if (AllConstant) {
1025 // We don't bother trying to simplify unless all of the operands
1027 // TODO: There are a lot of Simplify*'s we could call here, if we
1028 // wanted to. The original motivating case for this code was a
1029 // zext i1 false to i8, which we don't have an interface to
1030 // simplify (IE there is no SimplifyZExt).
1032 SmallVector<Constant *, 8> C;
1033 for (Value *Arg : E->operands())
1034 C.emplace_back(cast<Constant>(Arg));
1036 if (Value *V = ConstantFoldInstOperands(I, C, DL, TLI))
1037 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1043 const AggregateValueExpression *
1044 NewGVN::createAggregateValueExpression(Instruction *I) const {
1045 if (auto *II = dyn_cast<InsertValueInst>(I)) {
1046 auto *E = new (ExpressionAllocator)
1047 AggregateValueExpression(I->getNumOperands(), II->getNumIndices());
1048 setBasicExpressionInfo(I, E);
1049 E->allocateIntOperands(ExpressionAllocator);
1050 std::copy(II->idx_begin(), II->idx_end(), int_op_inserter(E));
1052 } else if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1053 auto *E = new (ExpressionAllocator)
1054 AggregateValueExpression(I->getNumOperands(), EI->getNumIndices());
1055 setBasicExpressionInfo(EI, E);
1056 E->allocateIntOperands(ExpressionAllocator);
1057 std::copy(EI->idx_begin(), EI->idx_end(), int_op_inserter(E));
1060 llvm_unreachable("Unhandled type of aggregate value operation");
1063 const DeadExpression *NewGVN::createDeadExpression() const {
1064 // DeadExpression has no arguments and all DeadExpression's are the same,
1065 // so we only need one of them.
1066 return SingletonDeadExpression;
1069 const VariableExpression *NewGVN::createVariableExpression(Value *V) const {
1070 auto *E = new (ExpressionAllocator) VariableExpression(V);
1071 E->setOpcode(V->getValueID());
1075 const Expression *NewGVN::createVariableOrConstant(Value *V) const {
1076 if (auto *C = dyn_cast<Constant>(V))
1077 return createConstantExpression(C);
1078 return createVariableExpression(V);
1081 const ConstantExpression *NewGVN::createConstantExpression(Constant *C) const {
1082 auto *E = new (ExpressionAllocator) ConstantExpression(C);
1083 E->setOpcode(C->getValueID());
1087 const UnknownExpression *NewGVN::createUnknownExpression(Instruction *I) const {
1088 auto *E = new (ExpressionAllocator) UnknownExpression(I);
1089 E->setOpcode(I->getOpcode());
1093 const CallExpression *
1094 NewGVN::createCallExpression(CallInst *CI, const MemoryAccess *MA) const {
1095 // FIXME: Add operand bundles for calls.
1097 new (ExpressionAllocator) CallExpression(CI->getNumOperands(), CI, MA);
1098 setBasicExpressionInfo(CI, E);
1102 // Return true if some equivalent of instruction Inst dominates instruction U.
1103 bool NewGVN::someEquivalentDominates(const Instruction *Inst,
1104 const Instruction *U) const {
1105 auto *CC = ValueToClass.lookup(Inst);
1106 // This must be an instruction because we are only called from phi nodes
1107 // in the case that the value it needs to check against is an instruction.
1109 // The most likely candiates for dominance are the leader and the next leader.
1110 // The leader or nextleader will dominate in all cases where there is an
1111 // equivalent that is higher up in the dom tree.
1112 // We can't *only* check them, however, because the
1113 // dominator tree could have an infinite number of non-dominating siblings
1114 // with instructions that are in the right congruence class.
1119 // Instruction U could be in H, with equivalents in every other sibling.
1120 // Depending on the rpo order picked, the leader could be the equivalent in
1121 // any of these siblings.
1124 if (DT->dominates(cast<Instruction>(CC->getLeader()), U))
1126 if (CC->getNextLeader().first &&
1127 DT->dominates(cast<Instruction>(CC->getNextLeader().first), U))
1129 return llvm::any_of(*CC, [&](const Value *Member) {
1130 return Member != CC->getLeader() &&
1131 DT->dominates(cast<Instruction>(Member), U);
1135 // See if we have a congruence class and leader for this operand, and if so,
1136 // return it. Otherwise, return the operand itself.
1137 Value *NewGVN::lookupOperandLeader(Value *V) const {
1138 CongruenceClass *CC = ValueToClass.lookup(V);
1140 // Everything in TOP is represented by undef, as it can be any value.
1141 // We do have to make sure we get the type right though, so we can't set the
1142 // RepLeader to undef.
1144 return UndefValue::get(V->getType());
1145 return CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
1151 const MemoryAccess *NewGVN::lookupMemoryLeader(const MemoryAccess *MA) const {
1152 auto *CC = getMemoryClass(MA);
1153 assert(CC->getMemoryLeader() &&
1154 "Every MemoryAccess should be mapped to a congruence class with a "
1155 "representative memory access");
1156 return CC->getMemoryLeader();
1159 // Return true if the MemoryAccess is really equivalent to everything. This is
1160 // equivalent to the lattice value "TOP" in most lattices. This is the initial
1161 // state of all MemoryAccesses.
1162 bool NewGVN::isMemoryAccessTOP(const MemoryAccess *MA) const {
1163 return getMemoryClass(MA) == TOPClass;
1166 LoadExpression *NewGVN::createLoadExpression(Type *LoadType, Value *PointerOp,
1168 const MemoryAccess *MA) const {
1170 new (ExpressionAllocator) LoadExpression(1, LI, lookupMemoryLeader(MA));
1171 E->allocateOperands(ArgRecycler, ExpressionAllocator);
1172 E->setType(LoadType);
1174 // Give store and loads same opcode so they value number together.
1176 E->op_push_back(PointerOp);
1178 E->setAlignment(LI->getAlignment());
1180 // TODO: Value number heap versions. We may be able to discover
1181 // things alias analysis can't on it's own (IE that a store and a
1182 // load have the same value, and thus, it isn't clobbering the load).
1186 const StoreExpression *
1187 NewGVN::createStoreExpression(StoreInst *SI, const MemoryAccess *MA) const {
1188 auto *StoredValueLeader = lookupOperandLeader(SI->getValueOperand());
1189 auto *E = new (ExpressionAllocator)
1190 StoreExpression(SI->getNumOperands(), SI, StoredValueLeader, MA);
1191 E->allocateOperands(ArgRecycler, ExpressionAllocator);
1192 E->setType(SI->getValueOperand()->getType());
1194 // Give store and loads same opcode so they value number together.
1196 E->op_push_back(lookupOperandLeader(SI->getPointerOperand()));
1198 // TODO: Value number heap versions. We may be able to discover
1199 // things alias analysis can't on it's own (IE that a store and a
1200 // load have the same value, and thus, it isn't clobbering the load).
1204 const Expression *NewGVN::performSymbolicStoreEvaluation(Instruction *I) const {
1205 // Unlike loads, we never try to eliminate stores, so we do not check if they
1206 // are simple and avoid value numbering them.
1207 auto *SI = cast<StoreInst>(I);
1208 auto *StoreAccess = getMemoryAccess(SI);
1209 // Get the expression, if any, for the RHS of the MemoryDef.
1210 const MemoryAccess *StoreRHS = StoreAccess->getDefiningAccess();
1211 if (EnableStoreRefinement)
1212 StoreRHS = MSSAWalker->getClobberingMemoryAccess(StoreAccess);
1213 // If we bypassed the use-def chains, make sure we add a use.
1214 if (StoreRHS != StoreAccess->getDefiningAccess())
1215 addMemoryUsers(StoreRHS, StoreAccess);
1216 StoreRHS = lookupMemoryLeader(StoreRHS);
1217 // If we are defined by ourselves, use the live on entry def.
1218 if (StoreRHS == StoreAccess)
1219 StoreRHS = MSSA->getLiveOnEntryDef();
1221 if (SI->isSimple()) {
1222 // See if we are defined by a previous store expression, it already has a
1223 // value, and it's the same value as our current store. FIXME: Right now, we
1224 // only do this for simple stores, we should expand to cover memcpys, etc.
1225 const auto *LastStore = createStoreExpression(SI, StoreRHS);
1226 const auto *LastCC = ExpressionToClass.lookup(LastStore);
1227 // Basically, check if the congruence class the store is in is defined by a
1228 // store that isn't us, and has the same value. MemorySSA takes care of
1229 // ensuring the store has the same memory state as us already.
1230 // The RepStoredValue gets nulled if all the stores disappear in a class, so
1231 // we don't need to check if the class contains a store besides us.
1233 LastCC->getStoredValue() == lookupOperandLeader(SI->getValueOperand()))
1235 deleteExpression(LastStore);
1236 // Also check if our value operand is defined by a load of the same memory
1237 // location, and the memory state is the same as it was then (otherwise, it
1238 // could have been overwritten later. See test32 in
1239 // transforms/DeadStoreElimination/simple.ll).
1241 dyn_cast<LoadInst>(lookupOperandLeader(SI->getValueOperand()))) {
1242 if ((lookupOperandLeader(LI->getPointerOperand()) ==
1243 lookupOperandLeader(SI->getPointerOperand())) &&
1244 (lookupMemoryLeader(getMemoryAccess(LI)->getDefiningAccess()) ==
1246 return createStoreExpression(SI, StoreRHS);
1250 // If the store is not equivalent to anything, value number it as a store that
1251 // produces a unique memory state (instead of using it's MemoryUse, we use
1253 return createStoreExpression(SI, StoreAccess);
1256 // See if we can extract the value of a loaded pointer from a load, a store, or
1257 // a memory instruction.
1259 NewGVN::performSymbolicLoadCoercion(Type *LoadType, Value *LoadPtr,
1260 LoadInst *LI, Instruction *DepInst,
1261 MemoryAccess *DefiningAccess) const {
1262 assert((!LI || LI->isSimple()) && "Not a simple load");
1263 if (auto *DepSI = dyn_cast<StoreInst>(DepInst)) {
1264 // Can't forward from non-atomic to atomic without violating memory model.
1265 // Also don't need to coerce if they are the same type, we will just
1267 if (LI->isAtomic() > DepSI->isAtomic() ||
1268 LoadType == DepSI->getValueOperand()->getType())
1270 int Offset = analyzeLoadFromClobberingStore(LoadType, LoadPtr, DepSI, DL);
1272 if (auto *C = dyn_cast<Constant>(
1273 lookupOperandLeader(DepSI->getValueOperand()))) {
1274 DEBUG(dbgs() << "Coercing load from store " << *DepSI << " to constant "
1276 return createConstantExpression(
1277 getConstantStoreValueForLoad(C, Offset, LoadType, DL));
1281 } else if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1282 // Can't forward from non-atomic to atomic without violating memory model.
1283 if (LI->isAtomic() > DepLI->isAtomic())
1285 int Offset = analyzeLoadFromClobberingLoad(LoadType, LoadPtr, DepLI, DL);
1287 // We can coerce a constant load into a load
1288 if (auto *C = dyn_cast<Constant>(lookupOperandLeader(DepLI)))
1289 if (auto *PossibleConstant =
1290 getConstantLoadValueForLoad(C, Offset, LoadType, DL)) {
1291 DEBUG(dbgs() << "Coercing load from load " << *LI << " to constant "
1292 << *PossibleConstant << "\n");
1293 return createConstantExpression(PossibleConstant);
1297 } else if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
1298 int Offset = analyzeLoadFromClobberingMemInst(LoadType, LoadPtr, DepMI, DL);
1300 if (auto *PossibleConstant =
1301 getConstantMemInstValueForLoad(DepMI, Offset, LoadType, DL)) {
1302 DEBUG(dbgs() << "Coercing load from meminst " << *DepMI
1303 << " to constant " << *PossibleConstant << "\n");
1304 return createConstantExpression(PossibleConstant);
1309 // All of the below are only true if the loaded pointer is produced
1310 // by the dependent instruction.
1311 if (LoadPtr != lookupOperandLeader(DepInst) &&
1312 !AA->isMustAlias(LoadPtr, DepInst))
1314 // If this load really doesn't depend on anything, then we must be loading an
1315 // undef value. This can happen when loading for a fresh allocation with no
1316 // intervening stores, for example. Note that this is only true in the case
1317 // that the result of the allocation is pointer equal to the load ptr.
1318 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1319 return createConstantExpression(UndefValue::get(LoadType));
1321 // If this load occurs either right after a lifetime begin,
1322 // then the loaded value is undefined.
1323 else if (auto *II = dyn_cast<IntrinsicInst>(DepInst)) {
1324 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1325 return createConstantExpression(UndefValue::get(LoadType));
1327 // If this load follows a calloc (which zero initializes memory),
1328 // then the loaded value is zero
1329 else if (isCallocLikeFn(DepInst, TLI)) {
1330 return createConstantExpression(Constant::getNullValue(LoadType));
1336 const Expression *NewGVN::performSymbolicLoadEvaluation(Instruction *I) const {
1337 auto *LI = cast<LoadInst>(I);
1339 // We can eliminate in favor of non-simple loads, but we won't be able to
1340 // eliminate the loads themselves.
1341 if (!LI->isSimple())
1344 Value *LoadAddressLeader = lookupOperandLeader(LI->getPointerOperand());
1345 // Load of undef is undef.
1346 if (isa<UndefValue>(LoadAddressLeader))
1347 return createConstantExpression(UndefValue::get(LI->getType()));
1348 MemoryAccess *OriginalAccess = getMemoryAccess(I);
1349 MemoryAccess *DefiningAccess =
1350 MSSAWalker->getClobberingMemoryAccess(OriginalAccess);
1352 if (!MSSA->isLiveOnEntryDef(DefiningAccess)) {
1353 if (auto *MD = dyn_cast<MemoryDef>(DefiningAccess)) {
1354 Instruction *DefiningInst = MD->getMemoryInst();
1355 // If the defining instruction is not reachable, replace with undef.
1356 if (!ReachableBlocks.count(DefiningInst->getParent()))
1357 return createConstantExpression(UndefValue::get(LI->getType()));
1358 // This will handle stores and memory insts. We only do if it the
1359 // defining access has a different type, or it is a pointer produced by
1360 // certain memory operations that cause the memory to have a fixed value
1361 // (IE things like calloc).
1362 if (const auto *CoercionResult =
1363 performSymbolicLoadCoercion(LI->getType(), LoadAddressLeader, LI,
1364 DefiningInst, DefiningAccess))
1365 return CoercionResult;
1369 const Expression *E = createLoadExpression(LI->getType(), LoadAddressLeader,
1370 LI, DefiningAccess);
1375 NewGVN::performSymbolicPredicateInfoEvaluation(Instruction *I) const {
1376 auto *PI = PredInfo->getPredicateInfoFor(I);
1380 DEBUG(dbgs() << "Found predicate info from instruction !\n");
1382 auto *PWC = dyn_cast<PredicateWithCondition>(PI);
1386 auto *CopyOf = I->getOperand(0);
1387 auto *Cond = PWC->Condition;
1389 // If this a copy of the condition, it must be either true or false depending
1390 // on the predicate info type and edge
1391 if (CopyOf == Cond) {
1392 // We should not need to add predicate users because the predicate info is
1393 // already a use of this operand.
1394 if (isa<PredicateAssume>(PI))
1395 return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1396 if (auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1397 if (PBranch->TrueEdge)
1398 return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1399 return createConstantExpression(ConstantInt::getFalse(Cond->getType()));
1401 if (auto *PSwitch = dyn_cast<PredicateSwitch>(PI))
1402 return createConstantExpression(cast<Constant>(PSwitch->CaseValue));
1405 // Not a copy of the condition, so see what the predicates tell us about this
1406 // value. First, though, we check to make sure the value is actually a copy
1407 // of one of the condition operands. It's possible, in certain cases, for it
1408 // to be a copy of a predicateinfo copy. In particular, if two branch
1409 // operations use the same condition, and one branch dominates the other, we
1410 // will end up with a copy of a copy. This is currently a small deficiency in
1411 // predicateinfo. What will end up happening here is that we will value
1412 // number both copies the same anyway.
1414 // Everything below relies on the condition being a comparison.
1415 auto *Cmp = dyn_cast<CmpInst>(Cond);
1419 if (CopyOf != Cmp->getOperand(0) && CopyOf != Cmp->getOperand(1)) {
1420 DEBUG(dbgs() << "Copy is not of any condition operands!\n");
1423 Value *FirstOp = lookupOperandLeader(Cmp->getOperand(0));
1424 Value *SecondOp = lookupOperandLeader(Cmp->getOperand(1));
1425 bool SwappedOps = false;
1427 if (shouldSwapOperands(FirstOp, SecondOp)) {
1428 std::swap(FirstOp, SecondOp);
1431 CmpInst::Predicate Predicate =
1432 SwappedOps ? Cmp->getSwappedPredicate() : Cmp->getPredicate();
1434 if (isa<PredicateAssume>(PI)) {
1435 // If the comparison is true when the operands are equal, then we know the
1436 // operands are equal, because assumes must always be true.
1437 if (CmpInst::isTrueWhenEqual(Predicate)) {
1438 addPredicateUsers(PI, I);
1439 addAdditionalUsers(Cmp->getOperand(0), I);
1440 return createVariableOrConstant(FirstOp);
1443 if (const auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1444 // If we are *not* a copy of the comparison, we may equal to the other
1445 // operand when the predicate implies something about equality of
1446 // operations. In particular, if the comparison is true/false when the
1447 // operands are equal, and we are on the right edge, we know this operation
1448 // is equal to something.
1449 if ((PBranch->TrueEdge && Predicate == CmpInst::ICMP_EQ) ||
1450 (!PBranch->TrueEdge && Predicate == CmpInst::ICMP_NE)) {
1451 addPredicateUsers(PI, I);
1452 addAdditionalUsers(Cmp->getOperand(0), I);
1453 return createVariableOrConstant(FirstOp);
1455 // Handle the special case of floating point.
1456 if (((PBranch->TrueEdge && Predicate == CmpInst::FCMP_OEQ) ||
1457 (!PBranch->TrueEdge && Predicate == CmpInst::FCMP_UNE)) &&
1458 isa<ConstantFP>(FirstOp) && !cast<ConstantFP>(FirstOp)->isZero()) {
1459 addPredicateUsers(PI, I);
1460 addAdditionalUsers(Cmp->getOperand(0), I);
1461 return createConstantExpression(cast<Constant>(FirstOp));
1467 // Evaluate read only and pure calls, and create an expression result.
1468 const Expression *NewGVN::performSymbolicCallEvaluation(Instruction *I) const {
1469 auto *CI = cast<CallInst>(I);
1470 if (auto *II = dyn_cast<IntrinsicInst>(I)) {
1471 // Instrinsics with the returned attribute are copies of arguments.
1472 if (auto *ReturnedValue = II->getReturnedArgOperand()) {
1473 if (II->getIntrinsicID() == Intrinsic::ssa_copy)
1474 if (const auto *Result = performSymbolicPredicateInfoEvaluation(I))
1476 return createVariableOrConstant(ReturnedValue);
1479 if (AA->doesNotAccessMemory(CI)) {
1480 return createCallExpression(CI, TOPClass->getMemoryLeader());
1481 } else if (AA->onlyReadsMemory(CI)) {
1482 MemoryAccess *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(CI);
1483 return createCallExpression(CI, DefiningAccess);
1488 // Retrieve the memory class for a given MemoryAccess.
1489 CongruenceClass *NewGVN::getMemoryClass(const MemoryAccess *MA) const {
1491 auto *Result = MemoryAccessToClass.lookup(MA);
1492 assert(Result && "Should have found memory class");
1496 // Update the MemoryAccess equivalence table to say that From is equal to To,
1497 // and return true if this is different from what already existed in the table.
1498 bool NewGVN::setMemoryClass(const MemoryAccess *From,
1499 CongruenceClass *NewClass) {
1501 "Every MemoryAccess should be getting mapped to a non-null class");
1502 DEBUG(dbgs() << "Setting " << *From);
1503 DEBUG(dbgs() << " equivalent to congruence class ");
1504 DEBUG(dbgs() << NewClass->getID() << " with current MemoryAccess leader ");
1505 DEBUG(dbgs() << *NewClass->getMemoryLeader() << "\n");
1507 auto LookupResult = MemoryAccessToClass.find(From);
1508 bool Changed = false;
1509 // If it's already in the table, see if the value changed.
1510 if (LookupResult != MemoryAccessToClass.end()) {
1511 auto *OldClass = LookupResult->second;
1512 if (OldClass != NewClass) {
1513 // If this is a phi, we have to handle memory member updates.
1514 if (auto *MP = dyn_cast<MemoryPhi>(From)) {
1515 OldClass->memory_erase(MP);
1516 NewClass->memory_insert(MP);
1517 // This may have killed the class if it had no non-memory members
1518 if (OldClass->getMemoryLeader() == From) {
1519 if (OldClass->definesNoMemory()) {
1520 OldClass->setMemoryLeader(nullptr);
1522 OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
1523 DEBUG(dbgs() << "Memory class leader change for class "
1524 << OldClass->getID() << " to "
1525 << *OldClass->getMemoryLeader()
1526 << " due to removal of a memory member " << *From
1528 markMemoryLeaderChangeTouched(OldClass);
1532 // It wasn't equivalent before, and now it is.
1533 LookupResult->second = NewClass;
1541 // Determine if a instruction is cycle-free. That means the values in the
1542 // instruction don't depend on any expressions that can change value as a result
1543 // of the instruction. For example, a non-cycle free instruction would be v =
1545 bool NewGVN::isCycleFree(const Instruction *I) const {
1546 // In order to compute cycle-freeness, we do SCC finding on the instruction,
1547 // and see what kind of SCC it ends up in. If it is a singleton, it is
1548 // cycle-free. If it is not in a singleton, it is only cycle free if the
1549 // other members are all phi nodes (as they do not compute anything, they are
1551 auto ICS = InstCycleState.lookup(I);
1552 if (ICS == ICS_Unknown) {
1554 auto &SCC = SCCFinder.getComponentFor(I);
1555 // It's cycle free if it's size 1 or or the SCC is *only* phi nodes.
1556 if (SCC.size() == 1)
1557 InstCycleState.insert({I, ICS_CycleFree});
1560 llvm::all_of(SCC, [](const Value *V) { return isa<PHINode>(V); });
1561 ICS = AllPhis ? ICS_CycleFree : ICS_Cycle;
1562 for (auto *Member : SCC)
1563 if (auto *MemberPhi = dyn_cast<PHINode>(Member))
1564 InstCycleState.insert({MemberPhi, ICS});
1567 if (ICS == ICS_Cycle)
1572 // Evaluate PHI nodes symbolically, and create an expression result.
1573 const Expression *NewGVN::performSymbolicPHIEvaluation(Instruction *I) const {
1574 // Resolve irreducible and reducible phi cycles.
1575 // FIXME: This is hopefully a temporary solution while we resolve the issues
1576 // with fixpointing self-cycles. It currently should be "guaranteed" to be
1577 // correct, but non-optimal. The SCCFinder does not, for example, take
1578 // reachability of arguments into account, etc.
1580 bool CanOptimize = true;
1581 SmallPtrSet<Value *, 8> OuterOps;
1583 auto &Component = SCCFinder.getComponentFor(I);
1584 for (auto *Member : Component) {
1585 if (!isa<PHINode>(Member)) {
1586 CanOptimize = false;
1589 for (auto &PHIOp : cast<PHINode>(Member)->operands())
1590 if (!isa<PHINode>(PHIOp) || !Component.count(cast<PHINode>(PHIOp)))
1591 OuterOps.insert(PHIOp);
1593 if (CanOptimize && OuterOps.size() == 1) {
1594 DEBUG(dbgs() << "Resolving cyclic phi to value " << *(*OuterOps.begin())
1596 return createVariableOrConstant(*OuterOps.begin());
1598 // True if one of the incoming phi edges is a backedge.
1599 bool HasBackedge = false;
1600 // All constant tracks the state of whether all the *original* phi operands
1601 // This is really shorthand for "this phi cannot cycle due to forward
1602 // change in value of the phi is guaranteed not to later change the value of
1603 // the phi. IE it can't be v = phi(undef, v+1)
1604 bool AllConstant = true;
1606 cast<PHIExpression>(createPHIExpression(I, HasBackedge, AllConstant));
1607 // We match the semantics of SimplifyPhiNode from InstructionSimplify here.
1608 // See if all arguments are the same.
1609 // We track if any were undef because they need special handling.
1610 bool HasUndef = false;
1611 auto Filtered = make_filter_range(E->operands(), [&](Value *Arg) {
1612 if (isa<UndefValue>(Arg)) {
1618 // If we are left with no operands, it's dead.
1619 if (Filtered.begin() == Filtered.end()) {
1620 // If it has undef at this point, it means there are no-non-undef arguments,
1621 // and thus, the value of the phi node must be undef.
1623 DEBUG(dbgs() << "PHI Node " << *I
1624 << " has no non-undef arguments, valuing it as undef\n");
1625 return createConstantExpression(UndefValue::get(I->getType()));
1628 DEBUG(dbgs() << "No arguments of PHI node " << *I << " are live\n");
1629 deleteExpression(E);
1630 return createDeadExpression();
1632 unsigned NumOps = 0;
1633 Value *AllSameValue = *(Filtered.begin());
1635 // Can't use std::equal here, sadly, because filter.begin moves.
1636 if (llvm::all_of(Filtered, [&](Value *Arg) {
1638 return Arg == AllSameValue;
1640 // In LLVM's non-standard representation of phi nodes, it's possible to have
1641 // phi nodes with cycles (IE dependent on other phis that are .... dependent
1642 // on the original phi node), especially in weird CFG's where some arguments
1643 // are unreachable, or uninitialized along certain paths. This can cause
1644 // infinite loops during evaluation. We work around this by not trying to
1645 // really evaluate them independently, but instead using a variable
1646 // expression to say if one is equivalent to the other.
1647 // We also special case undef, so that if we have an undef, we can't use the
1648 // common value unless it dominates the phi block.
1650 // If we have undef and at least one other value, this is really a
1651 // multivalued phi, and we need to know if it's cycle free in order to
1652 // evaluate whether we can ignore the undef. The other parts of this are
1653 // just shortcuts. If there is no backedge, or all operands are
1654 // constants, or all operands are ignored but the undef, it also must be
1656 if (!AllConstant && HasBackedge && NumOps > 0 &&
1657 !isa<UndefValue>(AllSameValue) && !isCycleFree(I))
1660 // Only have to check for instructions
1661 if (auto *AllSameInst = dyn_cast<Instruction>(AllSameValue))
1662 if (!someEquivalentDominates(AllSameInst, I))
1666 NumGVNPhisAllSame++;
1667 DEBUG(dbgs() << "Simplified PHI node " << *I << " to " << *AllSameValue
1669 deleteExpression(E);
1670 return createVariableOrConstant(AllSameValue);
1676 NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) const {
1677 if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1678 auto *II = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
1679 if (II && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
1680 unsigned Opcode = 0;
1681 // EI might be an extract from one of our recognised intrinsics. If it
1682 // is we'll synthesize a semantically equivalent expression instead on
1683 // an extract value expression.
1684 switch (II->getIntrinsicID()) {
1685 case Intrinsic::sadd_with_overflow:
1686 case Intrinsic::uadd_with_overflow:
1687 Opcode = Instruction::Add;
1689 case Intrinsic::ssub_with_overflow:
1690 case Intrinsic::usub_with_overflow:
1691 Opcode = Instruction::Sub;
1693 case Intrinsic::smul_with_overflow:
1694 case Intrinsic::umul_with_overflow:
1695 Opcode = Instruction::Mul;
1702 // Intrinsic recognized. Grab its args to finish building the
1704 assert(II->getNumArgOperands() == 2 &&
1705 "Expect two args for recognised intrinsics.");
1706 return createBinaryExpression(
1707 Opcode, EI->getType(), II->getArgOperand(0), II->getArgOperand(1));
1712 return createAggregateValueExpression(I);
1714 const Expression *NewGVN::performSymbolicCmpEvaluation(Instruction *I) const {
1715 auto *CI = dyn_cast<CmpInst>(I);
1716 // See if our operands are equal to those of a previous predicate, and if so,
1717 // if it implies true or false.
1718 auto Op0 = lookupOperandLeader(CI->getOperand(0));
1719 auto Op1 = lookupOperandLeader(CI->getOperand(1));
1720 auto OurPredicate = CI->getPredicate();
1721 if (shouldSwapOperands(Op0, Op1)) {
1722 std::swap(Op0, Op1);
1723 OurPredicate = CI->getSwappedPredicate();
1726 // Avoid processing the same info twice
1727 const PredicateBase *LastPredInfo = nullptr;
1728 // See if we know something about the comparison itself, like it is the target
1730 auto *CmpPI = PredInfo->getPredicateInfoFor(I);
1731 if (dyn_cast_or_null<PredicateAssume>(CmpPI))
1732 return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1735 // This condition does not depend on predicates, no need to add users
1736 if (CI->isTrueWhenEqual())
1737 return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1738 else if (CI->isFalseWhenEqual())
1739 return createConstantExpression(ConstantInt::getFalse(CI->getType()));
1742 // NOTE: Because we are comparing both operands here and below, and using
1743 // previous comparisons, we rely on fact that predicateinfo knows to mark
1744 // comparisons that use renamed operands as users of the earlier comparisons.
1745 // It is *not* enough to just mark predicateinfo renamed operands as users of
1746 // the earlier comparisons, because the *other* operand may have changed in a
1747 // previous iteration.
1750 // %b.0 = ssa.copy(%b)
1752 // icmp slt %c, %b.0
1754 // %c and %a may start out equal, and thus, the code below will say the second
1755 // %icmp is false. c may become equal to something else, and in that case the
1756 // %second icmp *must* be reexamined, but would not if only the renamed
1757 // %operands are considered users of the icmp.
1759 // *Currently* we only check one level of comparisons back, and only mark one
1760 // level back as touched when changes appen . If you modify this code to look
1761 // back farther through comparisons, you *must* mark the appropriate
1762 // comparisons as users in PredicateInfo.cpp, or you will cause bugs. See if
1763 // we know something just from the operands themselves
1765 // See if our operands have predicate info, so that we may be able to derive
1766 // something from a previous comparison.
1767 for (const auto &Op : CI->operands()) {
1768 auto *PI = PredInfo->getPredicateInfoFor(Op);
1769 if (const auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI)) {
1770 if (PI == LastPredInfo)
1774 // TODO: Along the false edge, we may know more things too, like icmp of
1775 // same operands is false.
1776 // TODO: We only handle actual comparison conditions below, not and/or.
1777 auto *BranchCond = dyn_cast<CmpInst>(PBranch->Condition);
1780 auto *BranchOp0 = lookupOperandLeader(BranchCond->getOperand(0));
1781 auto *BranchOp1 = lookupOperandLeader(BranchCond->getOperand(1));
1782 auto BranchPredicate = BranchCond->getPredicate();
1783 if (shouldSwapOperands(BranchOp0, BranchOp1)) {
1784 std::swap(BranchOp0, BranchOp1);
1785 BranchPredicate = BranchCond->getSwappedPredicate();
1787 if (BranchOp0 == Op0 && BranchOp1 == Op1) {
1788 if (PBranch->TrueEdge) {
1789 // If we know the previous predicate is true and we are in the true
1790 // edge then we may be implied true or false.
1791 if (CmpInst::isImpliedTrueByMatchingCmp(BranchPredicate,
1793 addPredicateUsers(PI, I);
1794 return createConstantExpression(
1795 ConstantInt::getTrue(CI->getType()));
1798 if (CmpInst::isImpliedFalseByMatchingCmp(BranchPredicate,
1800 addPredicateUsers(PI, I);
1801 return createConstantExpression(
1802 ConstantInt::getFalse(CI->getType()));
1806 // Just handle the ne and eq cases, where if we have the same
1807 // operands, we may know something.
1808 if (BranchPredicate == OurPredicate) {
1809 addPredicateUsers(PI, I);
1810 // Same predicate, same ops,we know it was false, so this is false.
1811 return createConstantExpression(
1812 ConstantInt::getFalse(CI->getType()));
1813 } else if (BranchPredicate ==
1814 CmpInst::getInversePredicate(OurPredicate)) {
1815 addPredicateUsers(PI, I);
1816 // Inverse predicate, we know the other was false, so this is true.
1817 return createConstantExpression(
1818 ConstantInt::getTrue(CI->getType()));
1824 // Create expression will take care of simplifyCmpInst
1825 return createExpression(I);
1828 // Return true if V is a value that will always be available (IE can
1829 // be placed anywhere) in the function. We don't do globals here
1830 // because they are often worse to put in place.
1831 // TODO: Separate cost from availability
1832 static bool alwaysAvailable(Value *V) {
1833 return isa<Constant>(V) || isa<Argument>(V);
1836 // Substitute and symbolize the value before value numbering.
1838 NewGVN::performSymbolicEvaluation(Value *V,
1839 SmallPtrSetImpl<Value *> &Visited) const {
1840 const Expression *E = nullptr;
1841 if (auto *C = dyn_cast<Constant>(V))
1842 E = createConstantExpression(C);
1843 else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1844 E = createVariableExpression(V);
1846 // TODO: memory intrinsics.
1847 // TODO: Some day, we should do the forward propagation and reassociation
1848 // parts of the algorithm.
1849 auto *I = cast<Instruction>(V);
1850 switch (I->getOpcode()) {
1851 case Instruction::ExtractValue:
1852 case Instruction::InsertValue:
1853 E = performSymbolicAggrValueEvaluation(I);
1855 case Instruction::PHI:
1856 E = performSymbolicPHIEvaluation(I);
1858 case Instruction::Call:
1859 E = performSymbolicCallEvaluation(I);
1861 case Instruction::Store:
1862 E = performSymbolicStoreEvaluation(I);
1864 case Instruction::Load:
1865 E = performSymbolicLoadEvaluation(I);
1867 case Instruction::BitCast: {
1868 E = createExpression(I);
1870 case Instruction::ICmp:
1871 case Instruction::FCmp: {
1872 E = performSymbolicCmpEvaluation(I);
1874 case Instruction::Add:
1875 case Instruction::FAdd:
1876 case Instruction::Sub:
1877 case Instruction::FSub:
1878 case Instruction::Mul:
1879 case Instruction::FMul:
1880 case Instruction::UDiv:
1881 case Instruction::SDiv:
1882 case Instruction::FDiv:
1883 case Instruction::URem:
1884 case Instruction::SRem:
1885 case Instruction::FRem:
1886 case Instruction::Shl:
1887 case Instruction::LShr:
1888 case Instruction::AShr:
1889 case Instruction::And:
1890 case Instruction::Or:
1891 case Instruction::Xor:
1892 case Instruction::Trunc:
1893 case Instruction::ZExt:
1894 case Instruction::SExt:
1895 case Instruction::FPToUI:
1896 case Instruction::FPToSI:
1897 case Instruction::UIToFP:
1898 case Instruction::SIToFP:
1899 case Instruction::FPTrunc:
1900 case Instruction::FPExt:
1901 case Instruction::PtrToInt:
1902 case Instruction::IntToPtr:
1903 case Instruction::Select:
1904 case Instruction::ExtractElement:
1905 case Instruction::InsertElement:
1906 case Instruction::ShuffleVector:
1907 case Instruction::GetElementPtr:
1908 E = createExpression(I);
1917 // Look up a container in a map, and then call a function for each thing in the
1919 template <typename Map, typename KeyType, typename Func>
1920 void NewGVN::for_each_found(Map &M, const KeyType &Key, Func F) {
1921 const auto Result = M.find_as(Key);
1922 if (Result != M.end())
1923 for (typename Map::mapped_type::value_type Mapped : Result->second)
1927 // Look up a container of values/instructions in a map, and touch all the
1928 // instructions in the container. Then erase value from the map.
1929 template <typename Map, typename KeyType>
1930 void NewGVN::touchAndErase(Map &M, const KeyType &Key) {
1931 const auto Result = M.find_as(Key);
1932 if (Result != M.end()) {
1933 for (const typename Map::mapped_type::value_type Mapped : Result->second)
1934 TouchedInstructions.set(InstrToDFSNum(Mapped));
1939 void NewGVN::addAdditionalUsers(Value *To, Value *User) const {
1940 AdditionalUsers[To].insert(User);
1943 void NewGVN::markUsersTouched(Value *V) {
1944 // Now mark the users as touched.
1945 for (auto *User : V->users()) {
1946 assert(isa<Instruction>(User) && "Use of value not within an instruction?");
1947 TouchedInstructions.set(InstrToDFSNum(User));
1949 touchAndErase(AdditionalUsers, V);
1952 void NewGVN::addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const {
1953 DEBUG(dbgs() << "Adding memory user " << *U << " to " << *To << "\n");
1954 MemoryToUsers[To].insert(U);
1957 void NewGVN::markMemoryDefTouched(const MemoryAccess *MA) {
1958 TouchedInstructions.set(MemoryToDFSNum(MA));
1961 void NewGVN::markMemoryUsersTouched(const MemoryAccess *MA) {
1962 if (isa<MemoryUse>(MA))
1964 for (auto U : MA->users())
1965 TouchedInstructions.set(MemoryToDFSNum(U));
1966 touchAndErase(MemoryToUsers, MA);
1969 // Add I to the set of users of a given predicate.
1970 void NewGVN::addPredicateUsers(const PredicateBase *PB, Instruction *I) const {
1971 // Don't add temporary instructions to the user lists.
1972 if (AllTempInstructions.count(I))
1975 if (auto *PBranch = dyn_cast<PredicateBranch>(PB))
1976 PredicateToUsers[PBranch->Condition].insert(I);
1977 else if (auto *PAssume = dyn_cast<PredicateBranch>(PB))
1978 PredicateToUsers[PAssume->Condition].insert(I);
1981 // Touch all the predicates that depend on this instruction.
1982 void NewGVN::markPredicateUsersTouched(Instruction *I) {
1983 touchAndErase(PredicateToUsers, I);
1986 // Mark users affected by a memory leader change.
1987 void NewGVN::markMemoryLeaderChangeTouched(CongruenceClass *CC) {
1988 for (auto M : CC->memory())
1989 markMemoryDefTouched(M);
1992 // Touch the instructions that need to be updated after a congruence class has a
1993 // leader change, and mark changed values.
1994 void NewGVN::markValueLeaderChangeTouched(CongruenceClass *CC) {
1995 for (auto M : *CC) {
1996 if (auto *I = dyn_cast<Instruction>(M))
1997 TouchedInstructions.set(InstrToDFSNum(I));
1998 LeaderChanges.insert(M);
2002 // Give a range of things that have instruction DFS numbers, this will return
2003 // the member of the range with the smallest dfs number.
2004 template <class T, class Range>
2005 T *NewGVN::getMinDFSOfRange(const Range &R) const {
2006 std::pair<T *, unsigned> MinDFS = {nullptr, ~0U};
2007 for (const auto X : R) {
2008 auto DFSNum = InstrToDFSNum(X);
2009 if (DFSNum < MinDFS.second)
2010 MinDFS = {X, DFSNum};
2012 return MinDFS.first;
2015 // This function returns the MemoryAccess that should be the next leader of
2016 // congruence class CC, under the assumption that the current leader is going to
2018 const MemoryAccess *NewGVN::getNextMemoryLeader(CongruenceClass *CC) const {
2019 // TODO: If this ends up to slow, we can maintain a next memory leader like we
2020 // do for regular leaders.
2021 // Make sure there will be a leader to find
2022 assert(!CC->definesNoMemory() && "Can't get next leader if there is none");
2023 if (CC->getStoreCount() > 0) {
2024 if (auto *NL = dyn_cast_or_null<StoreInst>(CC->getNextLeader().first))
2025 return getMemoryAccess(NL);
2026 // Find the store with the minimum DFS number.
2027 auto *V = getMinDFSOfRange<Value>(make_filter_range(
2028 *CC, [&](const Value *V) { return isa<StoreInst>(V); }));
2029 return getMemoryAccess(cast<StoreInst>(V));
2031 assert(CC->getStoreCount() == 0);
2033 // Given our assertion, hitting this part must mean
2034 // !OldClass->memory_empty()
2035 if (CC->memory_size() == 1)
2036 return *CC->memory_begin();
2037 return getMinDFSOfRange<const MemoryPhi>(CC->memory());
2040 // This function returns the next value leader of a congruence class, under the
2041 // assumption that the current leader is going away. This should end up being
2042 // the next most dominating member.
2043 Value *NewGVN::getNextValueLeader(CongruenceClass *CC) const {
2044 // We don't need to sort members if there is only 1, and we don't care about
2045 // sorting the TOP class because everything either gets out of it or is
2048 if (CC->size() == 1 || CC == TOPClass) {
2049 return *(CC->begin());
2050 } else if (CC->getNextLeader().first) {
2051 ++NumGVNAvoidedSortedLeaderChanges;
2052 return CC->getNextLeader().first;
2054 ++NumGVNSortedLeaderChanges;
2055 // NOTE: If this ends up to slow, we can maintain a dual structure for
2056 // member testing/insertion, or keep things mostly sorted, and sort only
2057 // here, or use SparseBitVector or ....
2058 return getMinDFSOfRange<Value>(*CC);
2062 // Move a MemoryAccess, currently in OldClass, to NewClass, including updates to
2063 // the memory members, etc for the move.
2065 // The invariants of this function are:
2067 // I must be moving to NewClass from OldClass The StoreCount of OldClass and
2068 // NewClass is expected to have been updated for I already if it is is a store.
2069 // The OldClass memory leader has not been updated yet if I was the leader.
2070 void NewGVN::moveMemoryToNewCongruenceClass(Instruction *I,
2071 MemoryAccess *InstMA,
2072 CongruenceClass *OldClass,
2073 CongruenceClass *NewClass) {
2074 // If the leader is I, and we had a represenative MemoryAccess, it should
2075 // be the MemoryAccess of OldClass.
2076 assert((!InstMA || !OldClass->getMemoryLeader() ||
2077 OldClass->getLeader() != I ||
2078 OldClass->getMemoryLeader() == InstMA) &&
2079 "Representative MemoryAccess mismatch");
2080 // First, see what happens to the new class
2081 if (!NewClass->getMemoryLeader()) {
2082 // Should be a new class, or a store becoming a leader of a new class.
2083 assert(NewClass->size() == 1 ||
2084 (isa<StoreInst>(I) && NewClass->getStoreCount() == 1));
2085 NewClass->setMemoryLeader(InstMA);
2086 // Mark it touched if we didn't just create a singleton
2087 DEBUG(dbgs() << "Memory class leader change for class " << NewClass->getID()
2088 << " due to new memory instruction becoming leader\n");
2089 markMemoryLeaderChangeTouched(NewClass);
2091 setMemoryClass(InstMA, NewClass);
2092 // Now, fixup the old class if necessary
2093 if (OldClass->getMemoryLeader() == InstMA) {
2094 if (!OldClass->definesNoMemory()) {
2095 OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
2096 DEBUG(dbgs() << "Memory class leader change for class "
2097 << OldClass->getID() << " to "
2098 << *OldClass->getMemoryLeader()
2099 << " due to removal of old leader " << *InstMA << "\n");
2100 markMemoryLeaderChangeTouched(OldClass);
2102 OldClass->setMemoryLeader(nullptr);
2106 // Move a value, currently in OldClass, to be part of NewClass
2107 // Update OldClass and NewClass for the move (including changing leaders, etc).
2108 void NewGVN::moveValueToNewCongruenceClass(Instruction *I, const Expression *E,
2109 CongruenceClass *OldClass,
2110 CongruenceClass *NewClass) {
2111 if (I == OldClass->getNextLeader().first)
2112 OldClass->resetNextLeader();
2115 NewClass->insert(I);
2117 if (NewClass->getLeader() != I)
2118 NewClass->addPossibleNextLeader({I, InstrToDFSNum(I)});
2119 // Handle our special casing of stores.
2120 if (auto *SI = dyn_cast<StoreInst>(I)) {
2121 OldClass->decStoreCount();
2122 // Okay, so when do we want to make a store a leader of a class?
2123 // If we have a store defined by an earlier load, we want the earlier load
2124 // to lead the class.
2125 // If we have a store defined by something else, we want the store to lead
2126 // the class so everything else gets the "something else" as a value.
2127 // If we have a store as the single member of the class, we want the store
2129 if (NewClass->getStoreCount() == 0 && !NewClass->getStoredValue()) {
2130 // If it's a store expression we are using, it means we are not equivalent
2131 // to something earlier.
2132 if (auto *SE = dyn_cast<StoreExpression>(E)) {
2133 NewClass->setStoredValue(SE->getStoredValue());
2134 markValueLeaderChangeTouched(NewClass);
2135 // Shift the new class leader to be the store
2136 DEBUG(dbgs() << "Changing leader of congruence class "
2137 << NewClass->getID() << " from " << *NewClass->getLeader()
2138 << " to " << *SI << " because store joined class\n");
2139 // If we changed the leader, we have to mark it changed because we don't
2140 // know what it will do to symbolic evlauation.
2141 NewClass->setLeader(SI);
2143 // We rely on the code below handling the MemoryAccess change.
2145 NewClass->incStoreCount();
2147 // True if there is no memory instructions left in a class that had memory
2148 // instructions before.
2150 // If it's not a memory use, set the MemoryAccess equivalence
2151 auto *InstMA = dyn_cast_or_null<MemoryDef>(getMemoryAccess(I));
2153 moveMemoryToNewCongruenceClass(I, InstMA, OldClass, NewClass);
2154 ValueToClass[I] = NewClass;
2155 // See if we destroyed the class or need to swap leaders.
2156 if (OldClass->empty() && OldClass != TOPClass) {
2157 if (OldClass->getDefiningExpr()) {
2158 DEBUG(dbgs() << "Erasing expression " << *OldClass->getDefiningExpr()
2159 << " from table\n");
2160 ExpressionToClass.erase(OldClass->getDefiningExpr());
2162 } else if (OldClass->getLeader() == I) {
2163 // When the leader changes, the value numbering of
2164 // everything may change due to symbolization changes, so we need to
2166 DEBUG(dbgs() << "Value class leader change for class " << OldClass->getID()
2168 ++NumGVNLeaderChanges;
2169 // Destroy the stored value if there are no more stores to represent it.
2170 // Note that this is basically clean up for the expression removal that
2171 // happens below. If we remove stores from a class, we may leave it as a
2172 // class of equivalent memory phis.
2173 if (OldClass->getStoreCount() == 0) {
2174 if (OldClass->getStoredValue())
2175 OldClass->setStoredValue(nullptr);
2177 OldClass->setLeader(getNextValueLeader(OldClass));
2178 OldClass->resetNextLeader();
2179 markValueLeaderChangeTouched(OldClass);
2183 // For a given expression, mark the phi of ops instructions that could have
2184 // changed as a result.
2185 void NewGVN::markPhiOfOpsChanged(const Expression *E) {
2186 touchAndErase(ExpressionToPhiOfOps, E);
2189 // Perform congruence finding on a given value numbering expression.
2190 void NewGVN::performCongruenceFinding(Instruction *I, const Expression *E) {
2191 // This is guaranteed to return something, since it will at least find
2194 CongruenceClass *IClass = ValueToClass.lookup(I);
2195 assert(IClass && "Should have found a IClass");
2196 // Dead classes should have been eliminated from the mapping.
2197 assert(!IClass->isDead() && "Found a dead class");
2199 CongruenceClass *EClass = nullptr;
2200 if (const auto *VE = dyn_cast<VariableExpression>(E)) {
2201 EClass = ValueToClass.lookup(VE->getVariableValue());
2202 } else if (isa<DeadExpression>(E)) {
2206 auto lookupResult = ExpressionToClass.insert({E, nullptr});
2208 // If it's not in the value table, create a new congruence class.
2209 if (lookupResult.second) {
2210 CongruenceClass *NewClass = createCongruenceClass(nullptr, E);
2211 auto place = lookupResult.first;
2212 place->second = NewClass;
2214 // Constants and variables should always be made the leader.
2215 if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2216 NewClass->setLeader(CE->getConstantValue());
2217 } else if (const auto *SE = dyn_cast<StoreExpression>(E)) {
2218 StoreInst *SI = SE->getStoreInst();
2219 NewClass->setLeader(SI);
2220 NewClass->setStoredValue(SE->getStoredValue());
2221 // The RepMemoryAccess field will be filled in properly by the
2222 // moveValueToNewCongruenceClass call.
2224 NewClass->setLeader(I);
2226 assert(!isa<VariableExpression>(E) &&
2227 "VariableExpression should have been handled already");
2230 DEBUG(dbgs() << "Created new congruence class for " << *I
2231 << " using expression " << *E << " at " << NewClass->getID()
2232 << " and leader " << *(NewClass->getLeader()));
2233 if (NewClass->getStoredValue())
2234 DEBUG(dbgs() << " and stored value " << *(NewClass->getStoredValue()));
2235 DEBUG(dbgs() << "\n");
2237 EClass = lookupResult.first->second;
2238 if (isa<ConstantExpression>(E))
2239 assert((isa<Constant>(EClass->getLeader()) ||
2240 (EClass->getStoredValue() &&
2241 isa<Constant>(EClass->getStoredValue()))) &&
2242 "Any class with a constant expression should have a "
2245 assert(EClass && "Somehow don't have an eclass");
2247 assert(!EClass->isDead() && "We accidentally looked up a dead class");
2250 bool ClassChanged = IClass != EClass;
2251 bool LeaderChanged = LeaderChanges.erase(I);
2252 if (ClassChanged || LeaderChanged) {
2253 DEBUG(dbgs() << "New class " << EClass->getID() << " for expression " << *E
2256 moveValueToNewCongruenceClass(I, E, IClass, EClass);
2257 markPhiOfOpsChanged(E);
2260 markUsersTouched(I);
2261 if (MemoryAccess *MA = getMemoryAccess(I))
2262 markMemoryUsersTouched(MA);
2263 if (auto *CI = dyn_cast<CmpInst>(I))
2264 markPredicateUsersTouched(CI);
2266 // If we changed the class of the store, we want to ensure nothing finds the
2267 // old store expression. In particular, loads do not compare against stored
2268 // value, so they will find old store expressions (and associated class
2269 // mappings) if we leave them in the table.
2270 if (ClassChanged && isa<StoreInst>(I)) {
2271 auto *OldE = ValueToExpression.lookup(I);
2272 // It could just be that the old class died. We don't want to erase it if we
2273 // just moved classes.
2274 if (OldE && isa<StoreExpression>(OldE) && *E != *OldE)
2275 ExpressionToClass.erase(OldE);
2277 ValueToExpression[I] = E;
2280 // Process the fact that Edge (from, to) is reachable, including marking
2281 // any newly reachable blocks and instructions for processing.
2282 void NewGVN::updateReachableEdge(BasicBlock *From, BasicBlock *To) {
2283 // Check if the Edge was reachable before.
2284 if (ReachableEdges.insert({From, To}).second) {
2285 // If this block wasn't reachable before, all instructions are touched.
2286 if (ReachableBlocks.insert(To).second) {
2287 DEBUG(dbgs() << "Block " << getBlockName(To) << " marked reachable\n");
2288 const auto &InstRange = BlockInstRange.lookup(To);
2289 TouchedInstructions.set(InstRange.first, InstRange.second);
2291 DEBUG(dbgs() << "Block " << getBlockName(To)
2292 << " was reachable, but new edge {" << getBlockName(From)
2293 << "," << getBlockName(To) << "} to it found\n");
2295 // We've made an edge reachable to an existing block, which may
2296 // impact predicates. Otherwise, only mark the phi nodes as touched, as
2297 // they are the only thing that depend on new edges. Anything using their
2298 // values will get propagated to if necessary.
2299 if (MemoryAccess *MemPhi = getMemoryAccess(To))
2300 TouchedInstructions.set(InstrToDFSNum(MemPhi));
2302 auto BI = To->begin();
2303 while (isa<PHINode>(BI)) {
2304 TouchedInstructions.set(InstrToDFSNum(&*BI));
2307 for_each_found(PHIOfOpsPHIs, To, [&](const PHINode *I) {
2308 TouchedInstructions.set(InstrToDFSNum(I));
2314 // Given a predicate condition (from a switch, cmp, or whatever) and a block,
2315 // see if we know some constant value for it already.
2316 Value *NewGVN::findConditionEquivalence(Value *Cond) const {
2317 auto Result = lookupOperandLeader(Cond);
2318 if (isa<Constant>(Result))
2323 // Process the outgoing edges of a block for reachability.
2324 void NewGVN::processOutgoingEdges(TerminatorInst *TI, BasicBlock *B) {
2325 // Evaluate reachability of terminator instruction.
2327 if ((BR = dyn_cast<BranchInst>(TI)) && BR->isConditional()) {
2328 Value *Cond = BR->getCondition();
2329 Value *CondEvaluated = findConditionEquivalence(Cond);
2330 if (!CondEvaluated) {
2331 if (auto *I = dyn_cast<Instruction>(Cond)) {
2332 const Expression *E = createExpression(I);
2333 if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2334 CondEvaluated = CE->getConstantValue();
2336 } else if (isa<ConstantInt>(Cond)) {
2337 CondEvaluated = Cond;
2341 BasicBlock *TrueSucc = BR->getSuccessor(0);
2342 BasicBlock *FalseSucc = BR->getSuccessor(1);
2343 if (CondEvaluated && (CI = dyn_cast<ConstantInt>(CondEvaluated))) {
2345 DEBUG(dbgs() << "Condition for Terminator " << *TI
2346 << " evaluated to true\n");
2347 updateReachableEdge(B, TrueSucc);
2348 } else if (CI->isZero()) {
2349 DEBUG(dbgs() << "Condition for Terminator " << *TI
2350 << " evaluated to false\n");
2351 updateReachableEdge(B, FalseSucc);
2354 updateReachableEdge(B, TrueSucc);
2355 updateReachableEdge(B, FalseSucc);
2357 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
2358 // For switches, propagate the case values into the case
2361 // Remember how many outgoing edges there are to every successor.
2362 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2364 Value *SwitchCond = SI->getCondition();
2365 Value *CondEvaluated = findConditionEquivalence(SwitchCond);
2366 // See if we were able to turn this switch statement into a constant.
2367 if (CondEvaluated && isa<ConstantInt>(CondEvaluated)) {
2368 auto *CondVal = cast<ConstantInt>(CondEvaluated);
2369 // We should be able to get case value for this.
2370 auto Case = *SI->findCaseValue(CondVal);
2371 if (Case.getCaseSuccessor() == SI->getDefaultDest()) {
2372 // We proved the value is outside of the range of the case.
2373 // We can't do anything other than mark the default dest as reachable,
2375 updateReachableEdge(B, SI->getDefaultDest());
2378 // Now get where it goes and mark it reachable.
2379 BasicBlock *TargetBlock = Case.getCaseSuccessor();
2380 updateReachableEdge(B, TargetBlock);
2382 for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) {
2383 BasicBlock *TargetBlock = SI->getSuccessor(i);
2384 ++SwitchEdges[TargetBlock];
2385 updateReachableEdge(B, TargetBlock);
2389 // Otherwise this is either unconditional, or a type we have no
2390 // idea about. Just mark successors as reachable.
2391 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
2392 BasicBlock *TargetBlock = TI->getSuccessor(i);
2393 updateReachableEdge(B, TargetBlock);
2396 // This also may be a memory defining terminator, in which case, set it
2397 // equivalent only to itself.
2399 auto *MA = getMemoryAccess(TI);
2400 if (MA && !isa<MemoryUse>(MA)) {
2401 auto *CC = ensureLeaderOfMemoryClass(MA);
2402 if (setMemoryClass(MA, CC))
2403 markMemoryUsersTouched(MA);
2408 void NewGVN::addPhiOfOps(PHINode *Op, BasicBlock *BB,
2409 Instruction *ExistingValue) {
2410 InstrDFS[Op] = InstrToDFSNum(ExistingValue);
2411 AllTempInstructions.insert(Op);
2412 PHIOfOpsPHIs[BB].push_back(Op);
2413 TempToBlock[Op] = BB;
2415 RealToTemp[ExistingValue] = Op;
2418 static bool okayForPHIOfOps(const Instruction *I) {
2419 return isa<BinaryOperator>(I) || isa<SelectInst>(I) || isa<CmpInst>(I) ||
2423 // When we see an instruction that is an op of phis, generate the equivalent phi
2426 NewGVN::makePossiblePhiOfOps(Instruction *I, bool HasBackedge,
2427 SmallPtrSetImpl<Value *> &Visited) {
2428 if (!okayForPHIOfOps(I))
2431 if (!Visited.insert(I).second)
2433 // For now, we require the instruction be cycle free because we don't
2434 // *always* create a phi of ops for instructions that could be done as phi
2435 // of ops, we only do it if we think it is useful. If we did do it all the
2436 // time, we could remove the cycle free check.
2437 if (!isCycleFree(I))
2440 unsigned IDFSNum = InstrToDFSNum(I);
2441 // Pretty much all of the instructions we can convert to phi of ops over a
2442 // backedge that are adds, are really induction variables, and those are
2443 // pretty much pointless to convert. This is very coarse-grained for a
2444 // test, so if we do find some value, we can change it later.
2445 // But otherwise, what can happen is we convert the induction variable from
2451 // i = phi (0, tmpphi)
2452 // tmpphi = phi(1, tmpphi+1)
2454 // Which we don't want to happen. We could just avoid this for all non-cycle
2455 // free phis, and we made go that route.
2456 if (HasBackedge && I->getOpcode() == Instruction::Add)
2459 SmallPtrSet<const Value *, 8> ProcessedPHIs;
2460 // TODO: We don't do phi translation on memory accesses because it's
2461 // complicated. For a load, we'd need to be able to simulate a new memoryuse,
2462 // which we don't have a good way of doing ATM.
2463 auto *MemAccess = getMemoryAccess(I);
2464 // If the memory operation is defined by a memory operation this block that
2465 // isn't a MemoryPhi, transforming the pointer backwards through a scalar phi
2466 // can't help, as it would still be killed by that memory operation.
2467 if (MemAccess && !isa<MemoryPhi>(MemAccess->getDefiningAccess()) &&
2468 MemAccess->getDefiningAccess()->getBlock() == I->getParent())
2471 // Convert op of phis to phi of ops
2472 for (auto &Op : I->operands()) {
2473 if (!isa<PHINode>(Op))
2475 auto *OpPHI = cast<PHINode>(Op);
2476 // No point in doing this for one-operand phis.
2477 if (OpPHI->getNumOperands() == 1)
2479 if (!DebugCounter::shouldExecute(PHIOfOpsCounter))
2481 SmallVector<std::pair<Value *, BasicBlock *>, 4> Ops;
2482 auto *PHIBlock = getBlockForValue(OpPHI);
2483 for (auto PredBB : OpPHI->blocks()) {
2484 Value *FoundVal = nullptr;
2485 // We could just skip unreachable edges entirely but it's tricky to do
2486 // with rewriting existing phi nodes.
2487 if (ReachableEdges.count({PredBB, PHIBlock})) {
2488 // Clone the instruction, create an expression from it, and see if we
2490 Instruction *ValueOp = I->clone();
2492 TempToMemory.insert({ValueOp, MemAccess});
2494 for (auto &Op : ValueOp->operands()) {
2495 Op = Op->DoPHITranslation(PHIBlock, PredBB);
2496 // When this operand changes, it could change whether there is a
2497 // leader for us or not.
2498 addAdditionalUsers(Op, I);
2500 // Make sure it's marked as a temporary instruction.
2501 AllTempInstructions.insert(ValueOp);
2502 // and make sure anything that tries to add it's DFS number is
2503 // redirected to the instruction we are making a phi of ops
2505 InstrDFS.insert({ValueOp, IDFSNum});
2506 const Expression *E = performSymbolicEvaluation(ValueOp, Visited);
2507 InstrDFS.erase(ValueOp);
2508 AllTempInstructions.erase(ValueOp);
2509 ValueOp->deleteValue();
2511 TempToMemory.erase(ValueOp);
2514 FoundVal = findPhiOfOpsLeader(E, PredBB);
2516 ExpressionToPhiOfOps[E].insert(I);
2519 if (auto *SI = dyn_cast<StoreInst>(FoundVal))
2520 FoundVal = SI->getValueOperand();
2522 DEBUG(dbgs() << "Skipping phi of ops operand for incoming block "
2523 << getBlockName(PredBB)
2524 << " because the block is unreachable\n");
2525 FoundVal = UndefValue::get(I->getType());
2528 Ops.push_back({FoundVal, PredBB});
2529 DEBUG(dbgs() << "Found phi of ops operand " << *FoundVal << " in "
2530 << getBlockName(PredBB) << "\n");
2532 auto *ValuePHI = RealToTemp.lookup(I);
2533 bool NewPHI = false;
2535 ValuePHI = PHINode::Create(I->getType(), OpPHI->getNumOperands());
2536 addPhiOfOps(ValuePHI, PHIBlock, I);
2538 NumGVNPHIOfOpsCreated++;
2541 for (auto PHIOp : Ops)
2542 ValuePHI->addIncoming(PHIOp.first, PHIOp.second);
2545 for (auto PHIOp : Ops) {
2546 ValuePHI->setIncomingValue(i, PHIOp.first);
2547 ValuePHI->setIncomingBlock(i, PHIOp.second);
2552 DEBUG(dbgs() << "Created phi of ops " << *ValuePHI << " for " << *I
2554 return performSymbolicEvaluation(ValuePHI, Visited);
2559 // The algorithm initially places the values of the routine in the TOP
2560 // congruence class. The leader of TOP is the undetermined value `undef`.
2561 // When the algorithm has finished, values still in TOP are unreachable.
2562 void NewGVN::initializeCongruenceClasses(Function &F) {
2563 NextCongruenceNum = 0;
2565 // Note that even though we use the live on entry def as a representative
2566 // MemoryAccess, it is *not* the same as the actual live on entry def. We
2567 // have no real equivalemnt to undef for MemoryAccesses, and so we really
2568 // should be checking whether the MemoryAccess is top if we want to know if it
2569 // is equivalent to everything. Otherwise, what this really signifies is that
2570 // the access "it reaches all the way back to the beginning of the function"
2572 // Initialize all other instructions to be in TOP class.
2573 TOPClass = createCongruenceClass(nullptr, nullptr);
2574 TOPClass->setMemoryLeader(MSSA->getLiveOnEntryDef());
2575 // The live on entry def gets put into it's own class
2576 MemoryAccessToClass[MSSA->getLiveOnEntryDef()] =
2577 createMemoryClass(MSSA->getLiveOnEntryDef());
2579 for (auto DTN : nodes(DT)) {
2580 BasicBlock *BB = DTN->getBlock();
2581 // All MemoryAccesses are equivalent to live on entry to start. They must
2582 // be initialized to something so that initial changes are noticed. For
2583 // the maximal answer, we initialize them all to be the same as
2585 auto *MemoryBlockDefs = MSSA->getBlockDefs(BB);
2586 if (MemoryBlockDefs)
2587 for (const auto &Def : *MemoryBlockDefs) {
2588 MemoryAccessToClass[&Def] = TOPClass;
2589 auto *MD = dyn_cast<MemoryDef>(&Def);
2590 // Insert the memory phis into the member list.
2592 const MemoryPhi *MP = cast<MemoryPhi>(&Def);
2593 TOPClass->memory_insert(MP);
2594 MemoryPhiState.insert({MP, MPS_TOP});
2597 if (MD && isa<StoreInst>(MD->getMemoryInst()))
2598 TOPClass->incStoreCount();
2600 for (auto &I : *BB) {
2601 // TODO: Move to helper
2602 if (isa<PHINode>(&I))
2603 for (auto *U : I.users())
2604 if (auto *UInst = dyn_cast<Instruction>(U))
2605 if (InstrToDFSNum(UInst) != 0 && okayForPHIOfOps(UInst))
2606 PHINodeUses.insert(UInst);
2607 // Don't insert void terminators into the class. We don't value number
2608 // them, and they just end up sitting in TOP.
2609 if (isa<TerminatorInst>(I) && I.getType()->isVoidTy())
2611 TOPClass->insert(&I);
2612 ValueToClass[&I] = TOPClass;
2616 // Initialize arguments to be in their own unique congruence classes
2617 for (auto &FA : F.args())
2618 createSingletonCongruenceClass(&FA);
2621 void NewGVN::cleanupTables() {
2622 for (unsigned i = 0, e = CongruenceClasses.size(); i != e; ++i) {
2623 DEBUG(dbgs() << "Congruence class " << CongruenceClasses[i]->getID()
2624 << " has " << CongruenceClasses[i]->size() << " members\n");
2625 // Make sure we delete the congruence class (probably worth switching to
2626 // a unique_ptr at some point.
2627 delete CongruenceClasses[i];
2628 CongruenceClasses[i] = nullptr;
2631 // Destroy the value expressions
2632 SmallVector<Instruction *, 8> TempInst(AllTempInstructions.begin(),
2633 AllTempInstructions.end());
2634 AllTempInstructions.clear();
2636 // We have to drop all references for everything first, so there are no uses
2637 // left as we delete them.
2638 for (auto *I : TempInst) {
2639 I->dropAllReferences();
2642 while (!TempInst.empty()) {
2643 auto *I = TempInst.back();
2644 TempInst.pop_back();
2648 ValueToClass.clear();
2649 ArgRecycler.clear(ExpressionAllocator);
2650 ExpressionAllocator.Reset();
2651 CongruenceClasses.clear();
2652 ExpressionToClass.clear();
2653 ValueToExpression.clear();
2655 AdditionalUsers.clear();
2656 ExpressionToPhiOfOps.clear();
2657 TempToBlock.clear();
2658 TempToMemory.clear();
2659 PHIOfOpsPHIs.clear();
2660 ReachableBlocks.clear();
2661 ReachableEdges.clear();
2663 ProcessedCount.clear();
2666 InstructionsToErase.clear();
2668 BlockInstRange.clear();
2669 TouchedInstructions.clear();
2670 MemoryAccessToClass.clear();
2671 PredicateToUsers.clear();
2672 MemoryToUsers.clear();
2675 // Assign local DFS number mapping to instructions, and leave space for Value
2677 std::pair<unsigned, unsigned> NewGVN::assignDFSNumbers(BasicBlock *B,
2679 unsigned End = Start;
2680 if (MemoryAccess *MemPhi = getMemoryAccess(B)) {
2681 InstrDFS[MemPhi] = End++;
2682 DFSToInstr.emplace_back(MemPhi);
2685 // Then the real block goes next.
2686 for (auto &I : *B) {
2687 // There's no need to call isInstructionTriviallyDead more than once on
2688 // an instruction. Therefore, once we know that an instruction is dead
2689 // we change its DFS number so that it doesn't get value numbered.
2690 if (isInstructionTriviallyDead(&I, TLI)) {
2692 DEBUG(dbgs() << "Skipping trivially dead instruction " << I << "\n");
2693 markInstructionForDeletion(&I);
2696 InstrDFS[&I] = End++;
2697 DFSToInstr.emplace_back(&I);
2700 // All of the range functions taken half-open ranges (open on the end side).
2701 // So we do not subtract one from count, because at this point it is one
2702 // greater than the last instruction.
2703 return std::make_pair(Start, End);
2706 void NewGVN::updateProcessedCount(const Value *V) {
2708 if (ProcessedCount.count(V) == 0) {
2709 ProcessedCount.insert({V, 1});
2711 ++ProcessedCount[V];
2712 assert(ProcessedCount[V] < 100 &&
2713 "Seem to have processed the same Value a lot");
2717 // Evaluate MemoryPhi nodes symbolically, just like PHI nodes
2718 void NewGVN::valueNumberMemoryPhi(MemoryPhi *MP) {
2719 // If all the arguments are the same, the MemoryPhi has the same value as the
2720 // argument. Filter out unreachable blocks and self phis from our operands.
2721 // TODO: We could do cycle-checking on the memory phis to allow valueizing for
2722 // self-phi checking.
2723 const BasicBlock *PHIBlock = MP->getBlock();
2724 auto Filtered = make_filter_range(MP->operands(), [&](const Use &U) {
2725 return cast<MemoryAccess>(U) != MP &&
2726 !isMemoryAccessTOP(cast<MemoryAccess>(U)) &&
2727 ReachableEdges.count({MP->getIncomingBlock(U), PHIBlock});
2729 // If all that is left is nothing, our memoryphi is undef. We keep it as
2730 // InitialClass. Note: The only case this should happen is if we have at
2731 // least one self-argument.
2732 if (Filtered.begin() == Filtered.end()) {
2733 if (setMemoryClass(MP, TOPClass))
2734 markMemoryUsersTouched(MP);
2738 // Transform the remaining operands into operand leaders.
2739 // FIXME: mapped_iterator should have a range version.
2740 auto LookupFunc = [&](const Use &U) {
2741 return lookupMemoryLeader(cast<MemoryAccess>(U));
2743 auto MappedBegin = map_iterator(Filtered.begin(), LookupFunc);
2744 auto MappedEnd = map_iterator(Filtered.end(), LookupFunc);
2746 // and now check if all the elements are equal.
2747 // Sadly, we can't use std::equals since these are random access iterators.
2748 const auto *AllSameValue = *MappedBegin;
2750 bool AllEqual = std::all_of(
2751 MappedBegin, MappedEnd,
2752 [&AllSameValue](const MemoryAccess *V) { return V == AllSameValue; });
2755 DEBUG(dbgs() << "Memory Phi value numbered to " << *AllSameValue << "\n");
2757 DEBUG(dbgs() << "Memory Phi value numbered to itself\n");
2758 // If it's equal to something, it's in that class. Otherwise, it has to be in
2759 // a class where it is the leader (other things may be equivalent to it, but
2760 // it needs to start off in its own class, which means it must have been the
2761 // leader, and it can't have stopped being the leader because it was never
2763 CongruenceClass *CC =
2764 AllEqual ? getMemoryClass(AllSameValue) : ensureLeaderOfMemoryClass(MP);
2765 auto OldState = MemoryPhiState.lookup(MP);
2766 assert(OldState != MPS_Invalid && "Invalid memory phi state");
2767 auto NewState = AllEqual ? MPS_Equivalent : MPS_Unique;
2768 MemoryPhiState[MP] = NewState;
2769 if (setMemoryClass(MP, CC) || OldState != NewState)
2770 markMemoryUsersTouched(MP);
2773 // Value number a single instruction, symbolically evaluating, performing
2774 // congruence finding, and updating mappings.
2775 void NewGVN::valueNumberInstruction(Instruction *I) {
2776 DEBUG(dbgs() << "Processing instruction " << *I << "\n");
2777 if (!I->isTerminator()) {
2778 const Expression *Symbolized = nullptr;
2779 SmallPtrSet<Value *, 2> Visited;
2780 if (DebugCounter::shouldExecute(VNCounter)) {
2781 Symbolized = performSymbolicEvaluation(I, Visited);
2782 // Make a phi of ops if necessary
2783 if (Symbolized && !isa<ConstantExpression>(Symbolized) &&
2784 !isa<VariableExpression>(Symbolized) && PHINodeUses.count(I)) {
2785 // FIXME: Backedge argument
2786 auto *PHIE = makePossiblePhiOfOps(I, false, Visited);
2792 // Mark the instruction as unused so we don't value number it again.
2795 // If we couldn't come up with a symbolic expression, use the unknown
2797 if (Symbolized == nullptr)
2798 Symbolized = createUnknownExpression(I);
2799 performCongruenceFinding(I, Symbolized);
2801 // Handle terminators that return values. All of them produce values we
2802 // don't currently understand. We don't place non-value producing
2803 // terminators in a class.
2804 if (!I->getType()->isVoidTy()) {
2805 auto *Symbolized = createUnknownExpression(I);
2806 performCongruenceFinding(I, Symbolized);
2808 processOutgoingEdges(dyn_cast<TerminatorInst>(I), I->getParent());
2812 // Check if there is a path, using single or equal argument phi nodes, from
2814 bool NewGVN::singleReachablePHIPath(
2815 SmallPtrSet<const MemoryAccess *, 8> &Visited, const MemoryAccess *First,
2816 const MemoryAccess *Second) const {
2817 if (First == Second)
2819 if (MSSA->isLiveOnEntryDef(First))
2822 // This is not perfect, but as we're just verifying here, we can live with
2823 // the loss of precision. The real solution would be that of doing strongly
2824 // connected component finding in this routine, and it's probably not worth
2825 // the complexity for the time being. So, we just keep a set of visited
2826 // MemoryAccess and return true when we hit a cycle.
2827 if (Visited.count(First))
2829 Visited.insert(First);
2831 const auto *EndDef = First;
2832 for (auto *ChainDef : optimized_def_chain(First)) {
2833 if (ChainDef == Second)
2835 if (MSSA->isLiveOnEntryDef(ChainDef))
2839 auto *MP = cast<MemoryPhi>(EndDef);
2840 auto ReachableOperandPred = [&](const Use &U) {
2841 return ReachableEdges.count({MP->getIncomingBlock(U), MP->getBlock()});
2843 auto FilteredPhiArgs =
2844 make_filter_range(MP->operands(), ReachableOperandPred);
2845 SmallVector<const Value *, 32> OperandList;
2846 std::copy(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
2847 std::back_inserter(OperandList));
2848 bool Okay = OperandList.size() == 1;
2851 std::equal(OperandList.begin(), OperandList.end(), OperandList.begin());
2853 return singleReachablePHIPath(Visited, cast<MemoryAccess>(OperandList[0]),
2858 // Verify the that the memory equivalence table makes sense relative to the
2859 // congruence classes. Note that this checking is not perfect, and is currently
2860 // subject to very rare false negatives. It is only useful for
2861 // testing/debugging.
2862 void NewGVN::verifyMemoryCongruency() const {
2864 // Verify that the memory table equivalence and memory member set match
2865 for (const auto *CC : CongruenceClasses) {
2866 if (CC == TOPClass || CC->isDead())
2868 if (CC->getStoreCount() != 0) {
2869 assert((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) &&
2870 "Any class with a store as a leader should have a "
2871 "representative stored value");
2872 assert(CC->getMemoryLeader() &&
2873 "Any congruence class with a store should have a "
2874 "representative access");
2877 if (CC->getMemoryLeader())
2878 assert(MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
2879 "Representative MemoryAccess does not appear to be reverse "
2881 for (auto M : CC->memory())
2882 assert(MemoryAccessToClass.lookup(M) == CC &&
2883 "Memory member does not appear to be reverse mapped properly");
2886 // Anything equivalent in the MemoryAccess table should be in the same
2887 // congruence class.
2889 // Filter out the unreachable and trivially dead entries, because they may
2890 // never have been updated if the instructions were not processed.
2891 auto ReachableAccessPred =
2892 [&](const std::pair<const MemoryAccess *, CongruenceClass *> Pair) {
2893 bool Result = ReachableBlocks.count(Pair.first->getBlock());
2894 if (!Result || MSSA->isLiveOnEntryDef(Pair.first) ||
2895 MemoryToDFSNum(Pair.first) == 0)
2897 if (auto *MemDef = dyn_cast<MemoryDef>(Pair.first))
2898 return !isInstructionTriviallyDead(MemDef->getMemoryInst());
2900 // We could have phi nodes which operands are all trivially dead,
2901 // so we don't process them.
2902 if (auto *MemPHI = dyn_cast<MemoryPhi>(Pair.first)) {
2903 for (auto &U : MemPHI->incoming_values()) {
2904 if (Instruction *I = dyn_cast<Instruction>(U.get())) {
2905 if (!isInstructionTriviallyDead(I))
2915 auto Filtered = make_filter_range(MemoryAccessToClass, ReachableAccessPred);
2916 for (auto KV : Filtered) {
2917 if (auto *FirstMUD = dyn_cast<MemoryUseOrDef>(KV.first)) {
2918 auto *SecondMUD = dyn_cast<MemoryUseOrDef>(KV.second->getMemoryLeader());
2919 if (FirstMUD && SecondMUD) {
2920 SmallPtrSet<const MemoryAccess *, 8> VisitedMAS;
2921 assert((singleReachablePHIPath(VisitedMAS, FirstMUD, SecondMUD) ||
2922 ValueToClass.lookup(FirstMUD->getMemoryInst()) ==
2923 ValueToClass.lookup(SecondMUD->getMemoryInst())) &&
2924 "The instructions for these memory operations should have "
2925 "been in the same congruence class or reachable through"
2926 "a single argument phi");
2928 } else if (auto *FirstMP = dyn_cast<MemoryPhi>(KV.first)) {
2929 // We can only sanely verify that MemoryDefs in the operand list all have
2931 auto ReachableOperandPred = [&](const Use &U) {
2932 return ReachableEdges.count(
2933 {FirstMP->getIncomingBlock(U), FirstMP->getBlock()}) &&
2937 // All arguments should in the same class, ignoring unreachable arguments
2938 auto FilteredPhiArgs =
2939 make_filter_range(FirstMP->operands(), ReachableOperandPred);
2940 SmallVector<const CongruenceClass *, 16> PhiOpClasses;
2941 std::transform(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
2942 std::back_inserter(PhiOpClasses), [&](const Use &U) {
2943 const MemoryDef *MD = cast<MemoryDef>(U);
2944 return ValueToClass.lookup(MD->getMemoryInst());
2946 assert(std::equal(PhiOpClasses.begin(), PhiOpClasses.end(),
2947 PhiOpClasses.begin()) &&
2948 "All MemoryPhi arguments should be in the same class");
2954 // Verify that the sparse propagation we did actually found the maximal fixpoint
2955 // We do this by storing the value to class mapping, touching all instructions,
2956 // and redoing the iteration to see if anything changed.
2957 void NewGVN::verifyIterationSettled(Function &F) {
2959 DEBUG(dbgs() << "Beginning iteration verification\n");
2960 if (DebugCounter::isCounterSet(VNCounter))
2961 DebugCounter::setCounterValue(VNCounter, StartingVNCounter);
2963 // Note that we have to store the actual classes, as we may change existing
2964 // classes during iteration. This is because our memory iteration propagation
2965 // is not perfect, and so may waste a little work. But it should generate
2966 // exactly the same congruence classes we have now, with different IDs.
2967 std::map<const Value *, CongruenceClass> BeforeIteration;
2969 for (auto &KV : ValueToClass) {
2970 if (auto *I = dyn_cast<Instruction>(KV.first))
2971 // Skip unused/dead instructions.
2972 if (InstrToDFSNum(I) == 0)
2974 BeforeIteration.insert({KV.first, *KV.second});
2977 TouchedInstructions.set();
2978 TouchedInstructions.reset(0);
2979 iterateTouchedInstructions();
2980 DenseSet<std::pair<const CongruenceClass *, const CongruenceClass *>>
2982 for (const auto &KV : ValueToClass) {
2983 if (auto *I = dyn_cast<Instruction>(KV.first))
2984 // Skip unused/dead instructions.
2985 if (InstrToDFSNum(I) == 0)
2987 // We could sink these uses, but i think this adds a bit of clarity here as
2988 // to what we are comparing.
2989 auto *BeforeCC = &BeforeIteration.find(KV.first)->second;
2990 auto *AfterCC = KV.second;
2991 // Note that the classes can't change at this point, so we memoize the set
2993 if (!EqualClasses.count({BeforeCC, AfterCC})) {
2994 assert(BeforeCC->isEquivalentTo(AfterCC) &&
2995 "Value number changed after main loop completed!");
2996 EqualClasses.insert({BeforeCC, AfterCC});
3002 // Verify that for each store expression in the expression to class mapping,
3003 // only the latest appears, and multiple ones do not appear.
3004 // Because loads do not use the stored value when doing equality with stores,
3005 // if we don't erase the old store expressions from the table, a load can find
3006 // a no-longer valid StoreExpression.
3007 void NewGVN::verifyStoreExpressions() const {
3009 DenseSet<std::pair<const Value *, const Value *>> StoreExpressionSet;
3010 for (const auto &KV : ExpressionToClass) {
3011 if (auto *SE = dyn_cast<StoreExpression>(KV.first)) {
3012 // Make sure a version that will conflict with loads is not already there
3014 StoreExpressionSet.insert({SE->getOperand(0), SE->getMemoryLeader()});
3015 assert(Res.second &&
3016 "Stored expression conflict exists in expression table");
3017 auto *ValueExpr = ValueToExpression.lookup(SE->getStoreInst());
3018 assert(ValueExpr && ValueExpr->equals(*SE) &&
3019 "StoreExpression in ExpressionToClass is not latest "
3020 "StoreExpression for value");
3026 // This is the main value numbering loop, it iterates over the initial touched
3027 // instruction set, propagating value numbers, marking things touched, etc,
3028 // until the set of touched instructions is completely empty.
3029 void NewGVN::iterateTouchedInstructions() {
3030 unsigned int Iterations = 0;
3031 // Figure out where touchedinstructions starts
3032 int FirstInstr = TouchedInstructions.find_first();
3033 // Nothing set, nothing to iterate, just return.
3034 if (FirstInstr == -1)
3036 const BasicBlock *LastBlock = getBlockForValue(InstrFromDFSNum(FirstInstr));
3037 while (TouchedInstructions.any()) {
3039 // Walk through all the instructions in all the blocks in RPO.
3040 // TODO: As we hit a new block, we should push and pop equalities into a
3041 // table lookupOperandLeader can use, to catch things PredicateInfo
3042 // might miss, like edge-only equivalences.
3043 for (unsigned InstrNum : TouchedInstructions.set_bits()) {
3045 // This instruction was found to be dead. We don't bother looking
3047 if (InstrNum == 0) {
3048 TouchedInstructions.reset(InstrNum);
3052 Value *V = InstrFromDFSNum(InstrNum);
3053 const BasicBlock *CurrBlock = getBlockForValue(V);
3055 // If we hit a new block, do reachability processing.
3056 if (CurrBlock != LastBlock) {
3057 LastBlock = CurrBlock;
3058 bool BlockReachable = ReachableBlocks.count(CurrBlock);
3059 const auto &CurrInstRange = BlockInstRange.lookup(CurrBlock);
3061 // If it's not reachable, erase any touched instructions and move on.
3062 if (!BlockReachable) {
3063 TouchedInstructions.reset(CurrInstRange.first, CurrInstRange.second);
3064 DEBUG(dbgs() << "Skipping instructions in block "
3065 << getBlockName(CurrBlock)
3066 << " because it is unreachable\n");
3069 updateProcessedCount(CurrBlock);
3072 if (auto *MP = dyn_cast<MemoryPhi>(V)) {
3073 DEBUG(dbgs() << "Processing MemoryPhi " << *MP << "\n");
3074 valueNumberMemoryPhi(MP);
3075 } else if (auto *I = dyn_cast<Instruction>(V)) {
3076 valueNumberInstruction(I);
3078 llvm_unreachable("Should have been a MemoryPhi or Instruction");
3080 updateProcessedCount(V);
3081 // Reset after processing (because we may mark ourselves as touched when
3082 // we propagate equalities).
3083 TouchedInstructions.reset(InstrNum);
3086 NumGVNMaxIterations = std::max(NumGVNMaxIterations.getValue(), Iterations);
3089 // This is the main transformation entry point.
3090 bool NewGVN::runGVN() {
3091 if (DebugCounter::isCounterSet(VNCounter))
3092 StartingVNCounter = DebugCounter::getCounterValue(VNCounter);
3093 bool Changed = false;
3094 NumFuncArgs = F.arg_size();
3095 MSSAWalker = MSSA->getWalker();
3096 SingletonDeadExpression = new (ExpressionAllocator) DeadExpression();
3098 // Count number of instructions for sizing of hash tables, and come
3099 // up with a global dfs numbering for instructions.
3100 unsigned ICount = 1;
3101 // Add an empty instruction to account for the fact that we start at 1
3102 DFSToInstr.emplace_back(nullptr);
3103 // Note: We want ideal RPO traversal of the blocks, which is not quite the
3104 // same as dominator tree order, particularly with regard whether backedges
3105 // get visited first or second, given a block with multiple successors.
3106 // If we visit in the wrong order, we will end up performing N times as many
3108 // The dominator tree does guarantee that, for a given dom tree node, it's
3109 // parent must occur before it in the RPO ordering. Thus, we only need to sort
3111 ReversePostOrderTraversal<Function *> RPOT(&F);
3112 unsigned Counter = 0;
3113 for (auto &B : RPOT) {
3114 auto *Node = DT->getNode(B);
3115 assert(Node && "RPO and Dominator tree should have same reachability");
3116 RPOOrdering[Node] = ++Counter;
3118 // Sort dominator tree children arrays into RPO.
3119 for (auto &B : RPOT) {
3120 auto *Node = DT->getNode(B);
3121 if (Node->getChildren().size() > 1)
3122 std::sort(Node->begin(), Node->end(),
3123 [&](const DomTreeNode *A, const DomTreeNode *B) {
3124 return RPOOrdering[A] < RPOOrdering[B];
3128 // Now a standard depth first ordering of the domtree is equivalent to RPO.
3129 for (auto DTN : depth_first(DT->getRootNode())) {
3130 BasicBlock *B = DTN->getBlock();
3131 const auto &BlockRange = assignDFSNumbers(B, ICount);
3132 BlockInstRange.insert({B, BlockRange});
3133 ICount += BlockRange.second - BlockRange.first;
3135 initializeCongruenceClasses(F);
3137 TouchedInstructions.resize(ICount);
3138 // Ensure we don't end up resizing the expressionToClass map, as
3139 // that can be quite expensive. At most, we have one expression per
3141 ExpressionToClass.reserve(ICount);
3143 // Initialize the touched instructions to include the entry block.
3144 const auto &InstRange = BlockInstRange.lookup(&F.getEntryBlock());
3145 TouchedInstructions.set(InstRange.first, InstRange.second);
3146 DEBUG(dbgs() << "Block " << getBlockName(&F.getEntryBlock())
3147 << " marked reachable\n");
3148 ReachableBlocks.insert(&F.getEntryBlock());
3150 iterateTouchedInstructions();
3151 verifyMemoryCongruency();
3152 verifyIterationSettled(F);
3153 verifyStoreExpressions();
3155 Changed |= eliminateInstructions(F);
3157 // Delete all instructions marked for deletion.
3158 for (Instruction *ToErase : InstructionsToErase) {
3159 if (!ToErase->use_empty())
3160 ToErase->replaceAllUsesWith(UndefValue::get(ToErase->getType()));
3162 if (ToErase->getParent())
3163 ToErase->eraseFromParent();
3166 // Delete all unreachable blocks.
3167 auto UnreachableBlockPred = [&](const BasicBlock &BB) {
3168 return !ReachableBlocks.count(&BB);
3171 for (auto &BB : make_filter_range(F, UnreachableBlockPred)) {
3172 DEBUG(dbgs() << "We believe block " << getBlockName(&BB)
3173 << " is unreachable\n");
3174 deleteInstructionsInBlock(&BB);
3182 struct NewGVN::ValueDFS {
3186 // Only one of Def and U will be set.
3187 // The bool in the Def tells us whether the Def is the stored value of a
3189 PointerIntPair<Value *, 1, bool> Def;
3191 bool operator<(const ValueDFS &Other) const {
3192 // It's not enough that any given field be less than - we have sets
3193 // of fields that need to be evaluated together to give a proper ordering.
3194 // For example, if you have;
3199 // We want the second to be less than the first, but if we just go field
3200 // by field, we will get to Val 0 < Val 50 and say the first is less than
3201 // the second. We only want it to be less than if the DFS orders are equal.
3203 // Each LLVM instruction only produces one value, and thus the lowest-level
3204 // differentiator that really matters for the stack (and what we use as as a
3205 // replacement) is the local dfs number.
3206 // Everything else in the structure is instruction level, and only affects
3207 // the order in which we will replace operands of a given instruction.
3209 // For a given instruction (IE things with equal dfsin, dfsout, localnum),
3210 // the order of replacement of uses does not matter.
3214 // When you hit b, you will have two valuedfs with the same dfsin, out, and
3216 // The .val will be the same as well.
3217 // The .u's will be different.
3218 // You will replace both, and it does not matter what order you replace them
3219 // in (IE whether you replace operand 2, then operand 1, or operand 1, then
3221 // Similarly for the case of same dfsin, dfsout, localnum, but different
3226 // in c, we will a valuedfs for a, and one for b,with everything the same
3228 // It does not matter what order we replace these operands in.
3229 // You will always end up with the same IR, and this is guaranteed.
3230 return std::tie(DFSIn, DFSOut, LocalNum, Def, U) <
3231 std::tie(Other.DFSIn, Other.DFSOut, Other.LocalNum, Other.Def,
3236 // This function converts the set of members for a congruence class from values,
3237 // to sets of defs and uses with associated DFS info. The total number of
3238 // reachable uses for each value is stored in UseCount, and instructions that
3240 // dead (have no non-dead uses) are stored in ProbablyDead.
3241 void NewGVN::convertClassToDFSOrdered(
3242 const CongruenceClass &Dense, SmallVectorImpl<ValueDFS> &DFSOrderedSet,
3243 DenseMap<const Value *, unsigned int> &UseCounts,
3244 SmallPtrSetImpl<Instruction *> &ProbablyDead) const {
3245 for (auto D : Dense) {
3246 // First add the value.
3247 BasicBlock *BB = getBlockForValue(D);
3248 // Constants are handled prior to ever calling this function, so
3249 // we should only be left with instructions as members.
3250 assert(BB && "Should have figured out a basic block for value");
3252 DomTreeNode *DomNode = DT->getNode(BB);
3253 VDDef.DFSIn = DomNode->getDFSNumIn();
3254 VDDef.DFSOut = DomNode->getDFSNumOut();
3255 // If it's a store, use the leader of the value operand, if it's always
3256 // available, or the value operand. TODO: We could do dominance checks to
3257 // find a dominating leader, but not worth it ATM.
3258 if (auto *SI = dyn_cast<StoreInst>(D)) {
3259 auto Leader = lookupOperandLeader(SI->getValueOperand());
3260 if (alwaysAvailable(Leader)) {
3261 VDDef.Def.setPointer(Leader);
3263 VDDef.Def.setPointer(SI->getValueOperand());
3264 VDDef.Def.setInt(true);
3267 VDDef.Def.setPointer(D);
3269 assert(isa<Instruction>(D) &&
3270 "The dense set member should always be an instruction");
3271 Instruction *Def = cast<Instruction>(D);
3272 VDDef.LocalNum = InstrToDFSNum(D);
3273 DFSOrderedSet.push_back(VDDef);
3274 // If there is a phi node equivalent, add it
3275 if (auto *PN = RealToTemp.lookup(Def)) {
3277 dyn_cast_or_null<PHIExpression>(ValueToExpression.lookup(Def));
3279 VDDef.Def.setInt(false);
3280 VDDef.Def.setPointer(PN);
3282 DFSOrderedSet.push_back(VDDef);
3286 unsigned int UseCount = 0;
3287 // Now add the uses.
3288 for (auto &U : Def->uses()) {
3289 if (auto *I = dyn_cast<Instruction>(U.getUser())) {
3290 // Don't try to replace into dead uses
3291 if (InstructionsToErase.count(I))
3294 // Put the phi node uses in the incoming block.
3296 if (auto *P = dyn_cast<PHINode>(I)) {
3297 IBlock = P->getIncomingBlock(U);
3298 // Make phi node users appear last in the incoming block
3300 VDUse.LocalNum = InstrDFS.size() + 1;
3302 IBlock = getBlockForValue(I);
3303 VDUse.LocalNum = InstrToDFSNum(I);
3306 // Skip uses in unreachable blocks, as we're going
3308 if (ReachableBlocks.count(IBlock) == 0)
3311 DomTreeNode *DomNode = DT->getNode(IBlock);
3312 VDUse.DFSIn = DomNode->getDFSNumIn();
3313 VDUse.DFSOut = DomNode->getDFSNumOut();
3316 DFSOrderedSet.emplace_back(VDUse);
3320 // If there are no uses, it's probably dead (but it may have side-effects,
3321 // so not definitely dead. Otherwise, store the number of uses so we can
3322 // track if it becomes dead later).
3324 ProbablyDead.insert(Def);
3326 UseCounts[Def] = UseCount;
3330 // This function converts the set of members for a congruence class from values,
3331 // to the set of defs for loads and stores, with associated DFS info.
3332 void NewGVN::convertClassToLoadsAndStores(
3333 const CongruenceClass &Dense,
3334 SmallVectorImpl<ValueDFS> &LoadsAndStores) const {
3335 for (auto D : Dense) {
3336 if (!isa<LoadInst>(D) && !isa<StoreInst>(D))
3339 BasicBlock *BB = getBlockForValue(D);
3341 DomTreeNode *DomNode = DT->getNode(BB);
3342 VD.DFSIn = DomNode->getDFSNumIn();
3343 VD.DFSOut = DomNode->getDFSNumOut();
3344 VD.Def.setPointer(D);
3346 // If it's an instruction, use the real local dfs number.
3347 if (auto *I = dyn_cast<Instruction>(D))
3348 VD.LocalNum = InstrToDFSNum(I);
3350 llvm_unreachable("Should have been an instruction");
3352 LoadsAndStores.emplace_back(VD);
3356 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
3357 auto *ReplInst = dyn_cast<Instruction>(Repl);
3361 // Patch the replacement so that it is not more restrictive than the value
3363 // Note that if 'I' is a load being replaced by some operation,
3364 // for example, by an arithmetic operation, then andIRFlags()
3365 // would just erase all math flags from the original arithmetic
3366 // operation, which is clearly not wanted and not needed.
3367 if (!isa<LoadInst>(I))
3368 ReplInst->andIRFlags(I);
3370 // FIXME: If both the original and replacement value are part of the
3371 // same control-flow region (meaning that the execution of one
3372 // guarantees the execution of the other), then we can combine the
3373 // noalias scopes here and do better than the general conservative
3374 // answer used in combineMetadata().
3376 // In general, GVN unifies expressions over different control-flow
3377 // regions, and so we need a conservative combination of the noalias
3379 static const unsigned KnownIDs[] = {
3380 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
3381 LLVMContext::MD_noalias, LLVMContext::MD_range,
3382 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
3383 LLVMContext::MD_invariant_group};
3384 combineMetadata(ReplInst, I, KnownIDs);
3387 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
3388 patchReplacementInstruction(I, Repl);
3389 I->replaceAllUsesWith(Repl);
3392 void NewGVN::deleteInstructionsInBlock(BasicBlock *BB) {
3393 DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
3394 ++NumGVNBlocksDeleted;
3396 // Delete the instructions backwards, as it has a reduced likelihood of having
3397 // to update as many def-use and use-def chains. Start after the terminator.
3398 auto StartPoint = BB->rbegin();
3400 // Note that we explicitly recalculate BB->rend() on each iteration,
3401 // as it may change when we remove the first instruction.
3402 for (BasicBlock::reverse_iterator I(StartPoint); I != BB->rend();) {
3403 Instruction &Inst = *I++;
3404 if (!Inst.use_empty())
3405 Inst.replaceAllUsesWith(UndefValue::get(Inst.getType()));
3406 if (isa<LandingPadInst>(Inst))
3409 Inst.eraseFromParent();
3410 ++NumGVNInstrDeleted;
3412 // Now insert something that simplifycfg will turn into an unreachable.
3413 Type *Int8Ty = Type::getInt8Ty(BB->getContext());
3414 new StoreInst(UndefValue::get(Int8Ty),
3415 Constant::getNullValue(Int8Ty->getPointerTo()),
3416 BB->getTerminator());
3419 void NewGVN::markInstructionForDeletion(Instruction *I) {
3420 DEBUG(dbgs() << "Marking " << *I << " for deletion\n");
3421 InstructionsToErase.insert(I);
3424 void NewGVN::replaceInstruction(Instruction *I, Value *V) {
3426 DEBUG(dbgs() << "Replacing " << *I << " with " << *V << "\n");
3427 patchAndReplaceAllUsesWith(I, V);
3428 // We save the actual erasing to avoid invalidating memory
3429 // dependencies until we are done with everything.
3430 markInstructionForDeletion(I);
3435 // This is a stack that contains both the value and dfs info of where
3436 // that value is valid.
3437 class ValueDFSStack {
3439 Value *back() const { return ValueStack.back(); }
3440 std::pair<int, int> dfs_back() const { return DFSStack.back(); }
3442 void push_back(Value *V, int DFSIn, int DFSOut) {
3443 ValueStack.emplace_back(V);
3444 DFSStack.emplace_back(DFSIn, DFSOut);
3446 bool empty() const { return DFSStack.empty(); }
3447 bool isInScope(int DFSIn, int DFSOut) const {
3450 return DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second;
3453 void popUntilDFSScope(int DFSIn, int DFSOut) {
3455 // These two should always be in sync at this point.
3456 assert(ValueStack.size() == DFSStack.size() &&
3457 "Mismatch between ValueStack and DFSStack");
3459 !DFSStack.empty() &&
3460 !(DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second)) {
3461 DFSStack.pop_back();
3462 ValueStack.pop_back();
3467 SmallVector<Value *, 8> ValueStack;
3468 SmallVector<std::pair<int, int>, 8> DFSStack;
3472 // Given a value and a basic block we are trying to see if it is available in,
3473 // see if the value has a leader available in that block.
3474 Value *NewGVN::findPhiOfOpsLeader(const Expression *E,
3475 const BasicBlock *BB) const {
3476 // It would already be constant if we could make it constant
3477 if (auto *CE = dyn_cast<ConstantExpression>(E))
3478 return CE->getConstantValue();
3479 if (auto *VE = dyn_cast<VariableExpression>(E))
3480 return VE->getVariableValue();
3482 auto *CC = ExpressionToClass.lookup(E);
3485 if (alwaysAvailable(CC->getLeader()))
3486 return CC->getLeader();
3488 for (auto Member : *CC) {
3489 auto *MemberInst = dyn_cast<Instruction>(Member);
3490 // Anything that isn't an instruction is always available.
3493 // If we are looking for something in the same block as the member, it must
3494 // be a leader because this function is looking for operands for a phi node.
3495 if (MemberInst->getParent() == BB ||
3496 DT->dominates(MemberInst->getParent(), BB)) {
3503 bool NewGVN::eliminateInstructions(Function &F) {
3504 // This is a non-standard eliminator. The normal way to eliminate is
3505 // to walk the dominator tree in order, keeping track of available
3506 // values, and eliminating them. However, this is mildly
3507 // pointless. It requires doing lookups on every instruction,
3508 // regardless of whether we will ever eliminate it. For
3509 // instructions part of most singleton congruence classes, we know we
3510 // will never eliminate them.
3512 // Instead, this eliminator looks at the congruence classes directly, sorts
3513 // them into a DFS ordering of the dominator tree, and then we just
3514 // perform elimination straight on the sets by walking the congruence
3515 // class member uses in order, and eliminate the ones dominated by the
3516 // last member. This is worst case O(E log E) where E = number of
3517 // instructions in a single congruence class. In theory, this is all
3518 // instructions. In practice, it is much faster, as most instructions are
3519 // either in singleton congruence classes or can't possibly be eliminated
3520 // anyway (if there are no overlapping DFS ranges in class).
3521 // When we find something not dominated, it becomes the new leader
3522 // for elimination purposes.
3523 // TODO: If we wanted to be faster, We could remove any members with no
3524 // overlapping ranges while sorting, as we will never eliminate anything
3525 // with those members, as they don't dominate anything else in our set.
3527 bool AnythingReplaced = false;
3529 // Since we are going to walk the domtree anyway, and we can't guarantee the
3530 // DFS numbers are updated, we compute some ourselves.
3531 DT->updateDFSNumbers();
3533 // Go through all of our phi nodes, and kill the arguments associated with
3534 // unreachable edges.
3535 auto ReplaceUnreachablePHIArgs = [&](PHINode &PHI, BasicBlock *BB) {
3536 for (auto &Operand : PHI.incoming_values())
3537 if (!ReachableEdges.count({PHI.getIncomingBlock(Operand), BB})) {
3538 DEBUG(dbgs() << "Replacing incoming value of " << PHI << " for block "
3539 << getBlockName(PHI.getIncomingBlock(Operand))
3540 << " with undef due to it being unreachable\n");
3541 Operand.set(UndefValue::get(PHI.getType()));
3544 SmallPtrSet<BasicBlock *, 8> BlocksWithPhis;
3546 if ((!B.empty() && isa<PHINode>(*B.begin())) ||
3547 (PHIOfOpsPHIs.find(&B) != PHIOfOpsPHIs.end()))
3548 BlocksWithPhis.insert(&B);
3549 DenseMap<const BasicBlock *, unsigned> ReachablePredCount;
3550 for (auto KV : ReachableEdges)
3551 ReachablePredCount[KV.getEnd()]++;
3552 for (auto *BB : BlocksWithPhis)
3553 // TODO: It would be faster to use getNumIncomingBlocks() on a phi node in
3554 // the block and subtract the pred count, but it's more complicated.
3555 if (ReachablePredCount.lookup(BB) !=
3556 std::distance(pred_begin(BB), pred_end(BB))) {
3557 for (auto II = BB->begin(); isa<PHINode>(II); ++II) {
3558 auto &PHI = cast<PHINode>(*II);
3559 ReplaceUnreachablePHIArgs(PHI, BB);
3561 for_each_found(PHIOfOpsPHIs, BB, [&](PHINode *PHI) {
3562 ReplaceUnreachablePHIArgs(*PHI, BB);
3566 // Map to store the use counts
3567 DenseMap<const Value *, unsigned int> UseCounts;
3568 for (auto *CC : reverse(CongruenceClasses)) {
3569 DEBUG(dbgs() << "Eliminating in congruence class " << CC->getID() << "\n");
3570 // Track the equivalent store info so we can decide whether to try
3571 // dead store elimination.
3572 SmallVector<ValueDFS, 8> PossibleDeadStores;
3573 SmallPtrSet<Instruction *, 8> ProbablyDead;
3574 if (CC->isDead() || CC->empty())
3576 // Everything still in the TOP class is unreachable or dead.
3577 if (CC == TOPClass) {
3578 for (auto M : *CC) {
3579 auto *VTE = ValueToExpression.lookup(M);
3580 if (VTE && isa<DeadExpression>(VTE))
3581 markInstructionForDeletion(cast<Instruction>(M));
3582 assert((!ReachableBlocks.count(cast<Instruction>(M)->getParent()) ||
3583 InstructionsToErase.count(cast<Instruction>(M))) &&
3584 "Everything in TOP should be unreachable or dead at this "
3590 assert(CC->getLeader() && "We should have had a leader");
3591 // If this is a leader that is always available, and it's a
3592 // constant or has no equivalences, just replace everything with
3593 // it. We then update the congruence class with whatever members
3596 CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
3597 if (alwaysAvailable(Leader)) {
3598 CongruenceClass::MemberSet MembersLeft;
3599 for (auto M : *CC) {
3601 // Void things have no uses we can replace.
3602 if (Member == Leader || !isa<Instruction>(Member) ||
3603 Member->getType()->isVoidTy()) {
3604 MembersLeft.insert(Member);
3607 DEBUG(dbgs() << "Found replacement " << *(Leader) << " for " << *Member
3609 auto *I = cast<Instruction>(Member);
3610 assert(Leader != I && "About to accidentally remove our leader");
3611 replaceInstruction(I, Leader);
3612 AnythingReplaced = true;
3614 CC->swap(MembersLeft);
3616 // If this is a singleton, we can skip it.
3617 if (CC->size() != 1 || RealToTemp.lookup(Leader)) {
3618 // This is a stack because equality replacement/etc may place
3619 // constants in the middle of the member list, and we want to use
3620 // those constant values in preference to the current leader, over
3621 // the scope of those constants.
3622 ValueDFSStack EliminationStack;
3624 // Convert the members to DFS ordered sets and then merge them.
3625 SmallVector<ValueDFS, 8> DFSOrderedSet;
3626 convertClassToDFSOrdered(*CC, DFSOrderedSet, UseCounts, ProbablyDead);
3628 // Sort the whole thing.
3629 std::sort(DFSOrderedSet.begin(), DFSOrderedSet.end());
3630 for (auto &VD : DFSOrderedSet) {
3631 int MemberDFSIn = VD.DFSIn;
3632 int MemberDFSOut = VD.DFSOut;
3633 Value *Def = VD.Def.getPointer();
3634 bool FromStore = VD.Def.getInt();
3636 // We ignore void things because we can't get a value from them.
3637 if (Def && Def->getType()->isVoidTy())
3639 auto *DefInst = dyn_cast_or_null<Instruction>(Def);
3640 if (DefInst && AllTempInstructions.count(DefInst)) {
3641 auto *PN = cast<PHINode>(DefInst);
3643 // If this is a value phi and that's the expression we used, insert
3644 // it into the program
3645 // remove from temp instruction list.
3646 AllTempInstructions.erase(PN);
3647 auto *DefBlock = getBlockForValue(Def);
3648 DEBUG(dbgs() << "Inserting fully real phi of ops" << *Def
3650 << getBlockName(getBlockForValue(Def)) << "\n");
3651 PN->insertBefore(&DefBlock->front());
3653 NumGVNPHIOfOpsEliminations++;
3656 if (EliminationStack.empty()) {
3657 DEBUG(dbgs() << "Elimination Stack is empty\n");
3659 DEBUG(dbgs() << "Elimination Stack Top DFS numbers are ("
3660 << EliminationStack.dfs_back().first << ","
3661 << EliminationStack.dfs_back().second << ")\n");
3664 DEBUG(dbgs() << "Current DFS numbers are (" << MemberDFSIn << ","
3665 << MemberDFSOut << ")\n");
3666 // First, we see if we are out of scope or empty. If so,
3667 // and there equivalences, we try to replace the top of
3668 // stack with equivalences (if it's on the stack, it must
3669 // not have been eliminated yet).
3670 // Then we synchronize to our current scope, by
3671 // popping until we are back within a DFS scope that
3672 // dominates the current member.
3673 // Then, what happens depends on a few factors
3674 // If the stack is now empty, we need to push
3675 // If we have a constant or a local equivalence we want to
3676 // start using, we also push.
3677 // Otherwise, we walk along, processing members who are
3678 // dominated by this scope, and eliminate them.
3679 bool ShouldPush = Def && EliminationStack.empty();
3681 !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut);
3683 if (OutOfScope || ShouldPush) {
3684 // Sync to our current scope.
3685 EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
3686 bool ShouldPush = Def && EliminationStack.empty();
3688 EliminationStack.push_back(Def, MemberDFSIn, MemberDFSOut);
3692 // Skip the Def's, we only want to eliminate on their uses. But mark
3693 // dominated defs as dead.
3695 // For anything in this case, what and how we value number
3696 // guarantees that any side-effets that would have occurred (ie
3697 // throwing, etc) can be proven to either still occur (because it's
3698 // dominated by something that has the same side-effects), or never
3699 // occur. Otherwise, we would not have been able to prove it value
3700 // equivalent to something else. For these things, we can just mark
3701 // it all dead. Note that this is different from the "ProbablyDead"
3702 // set, which may not be dominated by anything, and thus, are only
3703 // easy to prove dead if they are also side-effect free. Note that
3704 // because stores are put in terms of the stored value, we skip
3705 // stored values here. If the stored value is really dead, it will
3706 // still be marked for deletion when we process it in its own class.
3707 if (!EliminationStack.empty() && Def != EliminationStack.back() &&
3708 isa<Instruction>(Def) && !FromStore)
3709 markInstructionForDeletion(cast<Instruction>(Def));
3712 // At this point, we know it is a Use we are trying to possibly
3715 assert(isa<Instruction>(U->get()) &&
3716 "Current def should have been an instruction");
3717 assert(isa<Instruction>(U->getUser()) &&
3718 "Current user should have been an instruction");
3720 // If the thing we are replacing into is already marked to be dead,
3721 // this use is dead. Note that this is true regardless of whether
3722 // we have anything dominating the use or not. We do this here
3723 // because we are already walking all the uses anyway.
3724 Instruction *InstUse = cast<Instruction>(U->getUser());
3725 if (InstructionsToErase.count(InstUse)) {
3726 auto &UseCount = UseCounts[U->get()];
3727 if (--UseCount == 0) {
3728 ProbablyDead.insert(cast<Instruction>(U->get()));
3732 // If we get to this point, and the stack is empty we must have a use
3733 // with nothing we can use to eliminate this use, so just skip it.
3734 if (EliminationStack.empty())
3737 Value *DominatingLeader = EliminationStack.back();
3739 auto *II = dyn_cast<IntrinsicInst>(DominatingLeader);
3740 if (II && II->getIntrinsicID() == Intrinsic::ssa_copy)
3741 DominatingLeader = II->getOperand(0);
3743 // Don't replace our existing users with ourselves.
3744 if (U->get() == DominatingLeader)
3746 DEBUG(dbgs() << "Found replacement " << *DominatingLeader << " for "
3747 << *U->get() << " in " << *(U->getUser()) << "\n");
3749 // If we replaced something in an instruction, handle the patching of
3750 // metadata. Skip this if we are replacing predicateinfo with its
3751 // original operand, as we already know we can just drop it.
3752 auto *ReplacedInst = cast<Instruction>(U->get());
3753 auto *PI = PredInfo->getPredicateInfoFor(ReplacedInst);
3754 if (!PI || DominatingLeader != PI->OriginalOp)
3755 patchReplacementInstruction(ReplacedInst, DominatingLeader);
3756 U->set(DominatingLeader);
3757 // This is now a use of the dominating leader, which means if the
3758 // dominating leader was dead, it's now live!
3759 auto &LeaderUseCount = UseCounts[DominatingLeader];
3760 // It's about to be alive again.
3761 if (LeaderUseCount == 0 && isa<Instruction>(DominatingLeader))
3762 ProbablyDead.erase(cast<Instruction>(DominatingLeader));
3763 if (LeaderUseCount == 0 && II)
3764 ProbablyDead.insert(II);
3766 AnythingReplaced = true;
3771 // At this point, anything still in the ProbablyDead set is actually dead if
3772 // would be trivially dead.
3773 for (auto *I : ProbablyDead)
3774 if (wouldInstructionBeTriviallyDead(I))
3775 markInstructionForDeletion(I);
3777 // Cleanup the congruence class.
3778 CongruenceClass::MemberSet MembersLeft;
3779 for (auto *Member : *CC)
3780 if (!isa<Instruction>(Member) ||
3781 !InstructionsToErase.count(cast<Instruction>(Member)))
3782 MembersLeft.insert(Member);
3783 CC->swap(MembersLeft);
3785 // If we have possible dead stores to look at, try to eliminate them.
3786 if (CC->getStoreCount() > 0) {
3787 convertClassToLoadsAndStores(*CC, PossibleDeadStores);
3788 std::sort(PossibleDeadStores.begin(), PossibleDeadStores.end());
3789 ValueDFSStack EliminationStack;
3790 for (auto &VD : PossibleDeadStores) {
3791 int MemberDFSIn = VD.DFSIn;
3792 int MemberDFSOut = VD.DFSOut;
3793 Instruction *Member = cast<Instruction>(VD.Def.getPointer());
3794 if (EliminationStack.empty() ||
3795 !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut)) {
3796 // Sync to our current scope.
3797 EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
3798 if (EliminationStack.empty()) {
3799 EliminationStack.push_back(Member, MemberDFSIn, MemberDFSOut);
3803 // We already did load elimination, so nothing to do here.
3804 if (isa<LoadInst>(Member))
3806 assert(!EliminationStack.empty());
3807 Instruction *Leader = cast<Instruction>(EliminationStack.back());
3809 assert(DT->dominates(Leader->getParent(), Member->getParent()));
3810 // Member is dominater by Leader, and thus dead
3811 DEBUG(dbgs() << "Marking dead store " << *Member
3812 << " that is dominated by " << *Leader << "\n");
3813 markInstructionForDeletion(Member);
3819 return AnythingReplaced;
3822 // This function provides global ranking of operations so that we can place them
3823 // in a canonical order. Note that rank alone is not necessarily enough for a
3824 // complete ordering, as constants all have the same rank. However, generally,
3825 // we will simplify an operation with all constants so that it doesn't matter
3826 // what order they appear in.
3827 unsigned int NewGVN::getRank(const Value *V) const {
3828 // Prefer constants to undef to anything else
3829 // Undef is a constant, have to check it first.
3830 // Prefer smaller constants to constantexprs
3831 if (isa<ConstantExpr>(V))
3833 if (isa<UndefValue>(V))
3835 if (isa<Constant>(V))
3837 else if (auto *A = dyn_cast<Argument>(V))
3838 return 3 + A->getArgNo();
3840 // Need to shift the instruction DFS by number of arguments + 3 to account for
3841 // the constant and argument ranking above.
3842 unsigned Result = InstrToDFSNum(V);
3844 return 4 + NumFuncArgs + Result;
3845 // Unreachable or something else, just return a really large number.
3849 // This is a function that says whether two commutative operations should
3850 // have their order swapped when canonicalizing.
3851 bool NewGVN::shouldSwapOperands(const Value *A, const Value *B) const {
3852 // Because we only care about a total ordering, and don't rewrite expressions
3853 // in this order, we order by rank, which will give a strict weak ordering to
3854 // everything but constants, and then we order by pointer address.
3855 return std::make_pair(getRank(A), A) > std::make_pair(getRank(B), B);
3859 class NewGVNLegacyPass : public FunctionPass {
3861 static char ID; // Pass identification, replacement for typeid.
3862 NewGVNLegacyPass() : FunctionPass(ID) {
3863 initializeNewGVNLegacyPassPass(*PassRegistry::getPassRegistry());
3865 bool runOnFunction(Function &F) override;
3868 void getAnalysisUsage(AnalysisUsage &AU) const override {
3869 AU.addRequired<AssumptionCacheTracker>();
3870 AU.addRequired<DominatorTreeWrapperPass>();
3871 AU.addRequired<TargetLibraryInfoWrapperPass>();
3872 AU.addRequired<MemorySSAWrapperPass>();
3873 AU.addRequired<AAResultsWrapperPass>();
3874 AU.addPreserved<DominatorTreeWrapperPass>();
3875 AU.addPreserved<GlobalsAAWrapperPass>();
3880 bool NewGVNLegacyPass::runOnFunction(Function &F) {
3881 if (skipFunction(F))
3883 return NewGVN(F, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
3884 &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
3885 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
3886 &getAnalysis<AAResultsWrapperPass>().getAAResults(),
3887 &getAnalysis<MemorySSAWrapperPass>().getMSSA(),
3888 F.getParent()->getDataLayout())
3892 INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering",
3894 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3895 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
3896 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3897 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3898 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3899 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3900 INITIALIZE_PASS_END(NewGVNLegacyPass, "newgvn", "Global Value Numbering", false,
3903 char NewGVNLegacyPass::ID = 0;
3905 // createGVNPass - The public interface to this file.
3906 FunctionPass *llvm::createNewGVNPass() { return new NewGVNLegacyPass(); }
3908 PreservedAnalyses NewGVNPass::run(Function &F, AnalysisManager<Function> &AM) {
3909 // Apparently the order in which we get these results matter for
3910 // the old GVN (see Chandler's comment in GVN.cpp). I'll keep
3911 // the same order here, just in case.
3912 auto &AC = AM.getResult<AssumptionAnalysis>(F);
3913 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
3914 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
3915 auto &AA = AM.getResult<AAManager>(F);
3916 auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
3918 NewGVN(F, &DT, &AC, &TLI, &AA, &MSSA, F.getParent()->getDataLayout())
3921 return PreservedAnalyses::all();
3922 PreservedAnalyses PA;
3923 PA.preserve<DominatorTreeAnalysis>();
3924 PA.preserve<GlobalsAA>();