1 //===- NewGVN.cpp - Global Value Numbering Pass ---------------------------===//
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 //===----------------------------------------------------------------------===//
11 /// This file implements the new LLVM's Global Value Numbering pass.
12 /// GVN partitions values computed by a function into congruence classes.
13 /// Values ending up in the same congruence class are guaranteed to be the same
14 /// for every execution of the program. In that respect, congruency is a
15 /// compile-time approximation of equivalence of values at runtime.
16 /// The algorithm implemented here uses a sparse formulation and it's based
17 /// on the ideas described in the paper:
18 /// "A Sparse Algorithm for Predicated Global Value Numbering" from
21 /// A brief overview of the algorithm: The algorithm is essentially the same as
22 /// the standard RPO value numbering algorithm (a good reference is the paper
23 /// "SCC based value numbering" by L. Taylor Simpson) with one major difference:
24 /// The RPO algorithm proceeds, on every iteration, to process every reachable
25 /// block and every instruction in that block. This is because the standard RPO
26 /// algorithm does not track what things have the same value number, it only
27 /// tracks what the value number of a given operation is (the mapping is
28 /// operation -> value number). Thus, when a value number of an operation
29 /// changes, it must reprocess everything to ensure all uses of a value number
30 /// get updated properly. In constrast, the sparse algorithm we use *also*
31 /// tracks what operations have a given value number (IE it also tracks the
32 /// reverse mapping from value number -> operations with that value number), so
33 /// that it only needs to reprocess the instructions that are affected when
34 /// something's value number changes. The vast majority of complexity and code
35 /// in this file is devoted to tracking what value numbers could change for what
36 /// instructions when various things happen. The rest of the algorithm is
37 /// devoted to performing symbolic evaluation, forward propagation, and
38 /// simplification of operations based on the value numbers deduced so far
40 /// In order to make the GVN mostly-complete, we use a technique derived from
41 /// "Detection of Redundant Expressions: A Complete and Polynomial-time
42 /// Algorithm in SSA" by R.R. Pai. The source of incompleteness in most SSA
43 /// based GVN algorithms is related to their inability to detect equivalence
44 /// between phi of ops (IE phi(a+b, c+d)) and op of phis (phi(a,c) + phi(b, d)).
45 /// We resolve this issue by generating the equivalent "phi of ops" form for
46 /// each op of phis we see, in a way that only takes polynomial time to resolve.
48 /// We also do not perform elimination by using any published algorithm. All
49 /// published algorithms are O(Instructions). Instead, we use a technique that
50 /// is O(number of operations with the same value number), enabling us to skip
51 /// trying to eliminate things that have unique value numbers.
53 //===----------------------------------------------------------------------===//
55 #include "llvm/Transforms/Scalar/NewGVN.h"
56 #include "llvm/ADT/ArrayRef.h"
57 #include "llvm/ADT/BitVector.h"
58 #include "llvm/ADT/DenseMap.h"
59 #include "llvm/ADT/DenseMapInfo.h"
60 #include "llvm/ADT/DenseSet.h"
61 #include "llvm/ADT/DepthFirstIterator.h"
62 #include "llvm/ADT/GraphTraits.h"
63 #include "llvm/ADT/Hashing.h"
64 #include "llvm/ADT/PointerIntPair.h"
65 #include "llvm/ADT/PostOrderIterator.h"
66 #include "llvm/ADT/SmallPtrSet.h"
67 #include "llvm/ADT/SmallVector.h"
68 #include "llvm/ADT/SparseBitVector.h"
69 #include "llvm/ADT/Statistic.h"
70 #include "llvm/ADT/iterator_range.h"
71 #include "llvm/Analysis/AliasAnalysis.h"
72 #include "llvm/Analysis/AssumptionCache.h"
73 #include "llvm/Analysis/CFGPrinter.h"
74 #include "llvm/Analysis/ConstantFolding.h"
75 #include "llvm/Analysis/GlobalsModRef.h"
76 #include "llvm/Analysis/InstructionSimplify.h"
77 #include "llvm/Analysis/MemoryBuiltins.h"
78 #include "llvm/Analysis/MemorySSA.h"
79 #include "llvm/Analysis/TargetLibraryInfo.h"
80 #include "llvm/Analysis/Utils/Local.h"
81 #include "llvm/IR/Argument.h"
82 #include "llvm/IR/BasicBlock.h"
83 #include "llvm/IR/Constant.h"
84 #include "llvm/IR/Constants.h"
85 #include "llvm/IR/Dominators.h"
86 #include "llvm/IR/Function.h"
87 #include "llvm/IR/InstrTypes.h"
88 #include "llvm/IR/Instruction.h"
89 #include "llvm/IR/Instructions.h"
90 #include "llvm/IR/IntrinsicInst.h"
91 #include "llvm/IR/Intrinsics.h"
92 #include "llvm/IR/LLVMContext.h"
93 #include "llvm/IR/Type.h"
94 #include "llvm/IR/Use.h"
95 #include "llvm/IR/User.h"
96 #include "llvm/IR/Value.h"
97 #include "llvm/Pass.h"
98 #include "llvm/Support/Allocator.h"
99 #include "llvm/Support/ArrayRecycler.h"
100 #include "llvm/Support/Casting.h"
101 #include "llvm/Support/CommandLine.h"
102 #include "llvm/Support/Debug.h"
103 #include "llvm/Support/DebugCounter.h"
104 #include "llvm/Support/ErrorHandling.h"
105 #include "llvm/Support/PointerLikeTypeTraits.h"
106 #include "llvm/Support/raw_ostream.h"
107 #include "llvm/Transforms/Scalar.h"
108 #include "llvm/Transforms/Scalar/GVNExpression.h"
109 #include "llvm/Transforms/Utils/PredicateInfo.h"
110 #include "llvm/Transforms/Utils/VNCoercion.h"
123 using namespace llvm;
124 using namespace llvm::GVNExpression;
125 using namespace llvm::VNCoercion;
127 #define DEBUG_TYPE "newgvn"
129 STATISTIC(NumGVNInstrDeleted, "Number of instructions deleted");
130 STATISTIC(NumGVNBlocksDeleted, "Number of blocks deleted");
131 STATISTIC(NumGVNOpsSimplified, "Number of Expressions simplified");
132 STATISTIC(NumGVNPhisAllSame, "Number of PHIs whos arguments are all the same");
133 STATISTIC(NumGVNMaxIterations,
134 "Maximum Number of iterations it took to converge GVN");
135 STATISTIC(NumGVNLeaderChanges, "Number of leader changes");
136 STATISTIC(NumGVNSortedLeaderChanges, "Number of sorted leader changes");
137 STATISTIC(NumGVNAvoidedSortedLeaderChanges,
138 "Number of avoided sorted leader changes");
139 STATISTIC(NumGVNDeadStores, "Number of redundant/dead stores eliminated");
140 STATISTIC(NumGVNPHIOfOpsCreated, "Number of PHI of ops created");
141 STATISTIC(NumGVNPHIOfOpsEliminations,
142 "Number of things eliminated using PHI of ops");
143 DEBUG_COUNTER(VNCounter, "newgvn-vn",
144 "Controls which instructions are value numbered");
145 DEBUG_COUNTER(PHIOfOpsCounter, "newgvn-phi",
146 "Controls which instructions we create phi of ops for");
147 // Currently store defining access refinement is too slow due to basicaa being
148 // egregiously slow. This flag lets us keep it working while we work on this
150 static cl::opt<bool> EnableStoreRefinement("enable-store-refinement",
151 cl::init(false), cl::Hidden);
153 /// Currently, the generation "phi of ops" can result in correctness issues.
154 static cl::opt<bool> EnablePhiOfOps("enable-phi-of-ops", cl::init(true),
157 //===----------------------------------------------------------------------===//
159 //===----------------------------------------------------------------------===//
163 namespace GVNExpression {
165 Expression::~Expression() = default;
166 BasicExpression::~BasicExpression() = default;
167 CallExpression::~CallExpression() = default;
168 LoadExpression::~LoadExpression() = default;
169 StoreExpression::~StoreExpression() = default;
170 AggregateValueExpression::~AggregateValueExpression() = default;
171 PHIExpression::~PHIExpression() = default;
173 } // end namespace GVNExpression
174 } // end namespace llvm
178 // Tarjan's SCC finding algorithm with Nuutila's improvements
179 // SCCIterator is actually fairly complex for the simple thing we want.
180 // It also wants to hand us SCC's that are unrelated to the phi node we ask
181 // about, and have us process them there or risk redoing work.
182 // Graph traits over a filter iterator also doesn't work that well here.
183 // This SCC finder is specialized to walk use-def chains, and only follows
185 // not generic values (arguments, etc).
187 TarjanSCC() : Components(1) {}
189 void Start(const Instruction *Start) {
190 if (Root.lookup(Start) == 0)
194 const SmallPtrSetImpl<const Value *> &getComponentFor(const Value *V) const {
195 unsigned ComponentID = ValueToComponent.lookup(V);
197 assert(ComponentID > 0 &&
198 "Asking for a component for a value we never processed");
199 return Components[ComponentID];
203 void FindSCC(const Instruction *I) {
205 // Store the DFS Number we had before it possibly gets incremented.
206 unsigned int OurDFS = DFSNum;
207 for (auto &Op : I->operands()) {
208 if (auto *InstOp = dyn_cast<Instruction>(Op)) {
209 if (Root.lookup(Op) == 0)
211 if (!InComponent.count(Op))
212 Root[I] = std::min(Root.lookup(I), Root.lookup(Op));
215 // See if we really were the root of a component, by seeing if we still have
216 // our DFSNumber. If we do, we are the root of the component, and we have
217 // completed a component. If we do not, we are not the root of a component,
218 // and belong on the component stack.
219 if (Root.lookup(I) == OurDFS) {
220 unsigned ComponentID = Components.size();
221 Components.resize(Components.size() + 1);
222 auto &Component = Components.back();
224 DEBUG(dbgs() << "Component root is " << *I << "\n");
225 InComponent.insert(I);
226 ValueToComponent[I] = ComponentID;
227 // Pop a component off the stack and label it.
228 while (!Stack.empty() && Root.lookup(Stack.back()) >= OurDFS) {
229 auto *Member = Stack.back();
230 DEBUG(dbgs() << "Component member is " << *Member << "\n");
231 Component.insert(Member);
232 InComponent.insert(Member);
233 ValueToComponent[Member] = ComponentID;
237 // Part of a component, push to stack
242 unsigned int DFSNum = 1;
243 SmallPtrSet<const Value *, 8> InComponent;
244 DenseMap<const Value *, unsigned int> Root;
245 SmallVector<const Value *, 8> Stack;
247 // Store the components as vector of ptr sets, because we need the topo order
248 // of SCC's, but not individual member order
249 SmallVector<SmallPtrSet<const Value *, 8>, 8> Components;
251 DenseMap<const Value *, unsigned> ValueToComponent;
254 // Congruence classes represent the set of expressions/instructions
255 // that are all the same *during some scope in the function*.
256 // That is, because of the way we perform equality propagation, and
257 // because of memory value numbering, it is not correct to assume
258 // you can willy-nilly replace any member with any other at any
259 // point in the function.
261 // For any Value in the Member set, it is valid to replace any dominated member
264 // Every congruence class has a leader, and the leader is used to symbolize
265 // instructions in a canonical way (IE every operand of an instruction that is a
266 // member of the same congruence class will always be replaced with leader
267 // during symbolization). To simplify symbolization, we keep the leader as a
268 // constant if class can be proved to be a constant value. Otherwise, the
269 // leader is the member of the value set with the smallest DFS number. Each
270 // congruence class also has a defining expression, though the expression may be
271 // null. If it exists, it can be used for forward propagation and reassociation
274 // For memory, we also track a representative MemoryAccess, and a set of memory
275 // members for MemoryPhis (which have no real instructions). Note that for
276 // memory, it seems tempting to try to split the memory members into a
277 // MemoryCongruenceClass or something. Unfortunately, this does not work
278 // easily. The value numbering of a given memory expression depends on the
279 // leader of the memory congruence class, and the leader of memory congruence
280 // class depends on the value numbering of a given memory expression. This
281 // leads to wasted propagation, and in some cases, missed optimization. For
282 // example: If we had value numbered two stores together before, but now do not,
283 // we move them to a new value congruence class. This in turn will move at one
284 // of the memorydefs to a new memory congruence class. Which in turn, affects
285 // the value numbering of the stores we just value numbered (because the memory
286 // congruence class is part of the value number). So while theoretically
287 // possible to split them up, it turns out to be *incredibly* complicated to get
288 // it to work right, because of the interdependency. While structurally
289 // slightly messier, it is algorithmically much simpler and faster to do what we
290 // do here, and track them both at once in the same class.
291 // Note: The default iterators for this class iterate over values
292 class CongruenceClass {
294 using MemberType = Value;
295 using MemberSet = SmallPtrSet<MemberType *, 4>;
296 using MemoryMemberType = MemoryPhi;
297 using MemoryMemberSet = SmallPtrSet<const MemoryMemberType *, 2>;
299 explicit CongruenceClass(unsigned ID) : ID(ID) {}
300 CongruenceClass(unsigned ID, Value *Leader, const Expression *E)
301 : ID(ID), RepLeader(Leader), DefiningExpr(E) {}
303 unsigned getID() const { return ID; }
305 // True if this class has no members left. This is mainly used for assertion
306 // purposes, and for skipping empty classes.
307 bool isDead() const {
308 // If it's both dead from a value perspective, and dead from a memory
309 // perspective, it's really dead.
310 return empty() && memory_empty();
314 Value *getLeader() const { return RepLeader; }
315 void setLeader(Value *Leader) { RepLeader = Leader; }
316 const std::pair<Value *, unsigned int> &getNextLeader() const {
319 void resetNextLeader() { NextLeader = {nullptr, ~0}; }
320 void addPossibleNextLeader(std::pair<Value *, unsigned int> LeaderPair) {
321 if (LeaderPair.second < NextLeader.second)
322 NextLeader = LeaderPair;
325 Value *getStoredValue() const { return RepStoredValue; }
326 void setStoredValue(Value *Leader) { RepStoredValue = Leader; }
327 const MemoryAccess *getMemoryLeader() const { return RepMemoryAccess; }
328 void setMemoryLeader(const MemoryAccess *Leader) { RepMemoryAccess = Leader; }
330 // Forward propagation info
331 const Expression *getDefiningExpr() const { return DefiningExpr; }
334 bool empty() const { return Members.empty(); }
335 unsigned size() const { return Members.size(); }
336 MemberSet::const_iterator begin() const { return Members.begin(); }
337 MemberSet::const_iterator end() const { return Members.end(); }
338 void insert(MemberType *M) { Members.insert(M); }
339 void erase(MemberType *M) { Members.erase(M); }
340 void swap(MemberSet &Other) { Members.swap(Other); }
343 bool memory_empty() const { return MemoryMembers.empty(); }
344 unsigned memory_size() const { return MemoryMembers.size(); }
345 MemoryMemberSet::const_iterator memory_begin() const {
346 return MemoryMembers.begin();
348 MemoryMemberSet::const_iterator memory_end() const {
349 return MemoryMembers.end();
351 iterator_range<MemoryMemberSet::const_iterator> memory() const {
352 return make_range(memory_begin(), memory_end());
355 void memory_insert(const MemoryMemberType *M) { MemoryMembers.insert(M); }
356 void memory_erase(const MemoryMemberType *M) { MemoryMembers.erase(M); }
359 unsigned getStoreCount() const { return StoreCount; }
360 void incStoreCount() { ++StoreCount; }
361 void decStoreCount() {
362 assert(StoreCount != 0 && "Store count went negative");
366 // True if this class has no memory members.
367 bool definesNoMemory() const { return StoreCount == 0 && memory_empty(); }
369 // Return true if two congruence classes are equivalent to each other. This
371 // that every field but the ID number and the dead field are equivalent.
372 bool isEquivalentTo(const CongruenceClass *Other) const {
378 if (std::tie(StoreCount, RepLeader, RepStoredValue, RepMemoryAccess) !=
379 std::tie(Other->StoreCount, Other->RepLeader, Other->RepStoredValue,
380 Other->RepMemoryAccess))
382 if (DefiningExpr != Other->DefiningExpr)
383 if (!DefiningExpr || !Other->DefiningExpr ||
384 *DefiningExpr != *Other->DefiningExpr)
386 // We need some ordered set
387 std::set<Value *> AMembers(Members.begin(), Members.end());
388 std::set<Value *> BMembers(Members.begin(), Members.end());
389 return AMembers == BMembers;
395 // Representative leader.
396 Value *RepLeader = nullptr;
398 // The most dominating leader after our current leader, because the member set
399 // is not sorted and is expensive to keep sorted all the time.
400 std::pair<Value *, unsigned int> NextLeader = {nullptr, ~0U};
402 // If this is represented by a store, the value of the store.
403 Value *RepStoredValue = nullptr;
405 // If this class contains MemoryDefs or MemoryPhis, this is the leading memory
407 const MemoryAccess *RepMemoryAccess = nullptr;
409 // Defining Expression.
410 const Expression *DefiningExpr = nullptr;
412 // Actual members of this class.
415 // This is the set of MemoryPhis that exist in the class. MemoryDefs and
416 // MemoryUses have real instructions representing them, so we only need to
417 // track MemoryPhis here.
418 MemoryMemberSet MemoryMembers;
420 // Number of stores in this congruence class.
421 // This is used so we can detect store equivalence changes properly.
425 } // end anonymous namespace
429 struct ExactEqualsExpression {
432 explicit ExactEqualsExpression(const Expression &E) : E(E) {}
434 hash_code getComputedHash() const { return E.getComputedHash(); }
436 bool operator==(const Expression &Other) const {
437 return E.exactlyEquals(Other);
441 template <> struct DenseMapInfo<const Expression *> {
442 static const Expression *getEmptyKey() {
443 auto Val = static_cast<uintptr_t>(-1);
444 Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
445 return reinterpret_cast<const Expression *>(Val);
448 static const Expression *getTombstoneKey() {
449 auto Val = static_cast<uintptr_t>(~1U);
450 Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
451 return reinterpret_cast<const Expression *>(Val);
454 static unsigned getHashValue(const Expression *E) {
455 return E->getComputedHash();
458 static unsigned getHashValue(const ExactEqualsExpression &E) {
459 return E.getComputedHash();
462 static bool isEqual(const ExactEqualsExpression &LHS, const Expression *RHS) {
463 if (RHS == getTombstoneKey() || RHS == getEmptyKey())
468 static bool isEqual(const Expression *LHS, const Expression *RHS) {
471 if (LHS == getTombstoneKey() || RHS == getTombstoneKey() ||
472 LHS == getEmptyKey() || RHS == getEmptyKey())
474 // Compare hashes before equality. This is *not* what the hashtable does,
475 // since it is computing it modulo the number of buckets, whereas we are
476 // using the full hash keyspace. Since the hashes are precomputed, this
477 // check is *much* faster than equality.
478 if (LHS->getComputedHash() != RHS->getComputedHash())
484 } // end namespace llvm
491 const TargetLibraryInfo *TLI;
494 MemorySSAWalker *MSSAWalker;
495 const DataLayout &DL;
496 std::unique_ptr<PredicateInfo> PredInfo;
498 // These are the only two things the create* functions should have
499 // side-effects on due to allocating memory.
500 mutable BumpPtrAllocator ExpressionAllocator;
501 mutable ArrayRecycler<Value *> ArgRecycler;
502 mutable TarjanSCC SCCFinder;
503 const SimplifyQuery SQ;
505 // Number of function arguments, used by ranking
506 unsigned int NumFuncArgs;
508 // RPOOrdering of basic blocks
509 DenseMap<const DomTreeNode *, unsigned> RPOOrdering;
511 // Congruence class info.
513 // This class is called INITIAL in the paper. It is the class everything
514 // startsout in, and represents any value. Being an optimistic analysis,
515 // anything in the TOP class has the value TOP, which is indeterminate and
516 // equivalent to everything.
517 CongruenceClass *TOPClass;
518 std::vector<CongruenceClass *> CongruenceClasses;
519 unsigned NextCongruenceNum;
522 DenseMap<Value *, CongruenceClass *> ValueToClass;
523 DenseMap<Value *, const Expression *> ValueToExpression;
525 // Value PHI handling, used to make equivalence between phi(op, op) and
527 // These mappings just store various data that would normally be part of the
529 SmallPtrSet<const Instruction *, 8> PHINodeUses;
531 DenseMap<const Value *, bool> OpSafeForPHIOfOps;
533 // Map a temporary instruction we created to a parent block.
534 DenseMap<const Value *, BasicBlock *> TempToBlock;
536 // Map between the already in-program instructions and the temporary phis we
537 // created that they are known equivalent to.
538 DenseMap<const Value *, PHINode *> RealToTemp;
540 // In order to know when we should re-process instructions that have
541 // phi-of-ops, we track the set of expressions that they needed as
542 // leaders. When we discover new leaders for those expressions, we process the
543 // associated phi-of-op instructions again in case they have changed. The
544 // other way they may change is if they had leaders, and those leaders
545 // disappear. However, at the point they have leaders, there are uses of the
546 // relevant operands in the created phi node, and so they will get reprocessed
547 // through the normal user marking we perform.
548 mutable DenseMap<const Value *, SmallPtrSet<Value *, 2>> AdditionalUsers;
549 DenseMap<const Expression *, SmallPtrSet<Instruction *, 2>>
550 ExpressionToPhiOfOps;
552 // Map from temporary operation to MemoryAccess.
553 DenseMap<const Instruction *, MemoryUseOrDef *> TempToMemory;
555 // Set of all temporary instructions we created.
556 // Note: This will include instructions that were just created during value
557 // numbering. The way to test if something is using them is to check
559 DenseSet<Instruction *> AllTempInstructions;
561 // This is the set of instructions to revisit on a reachability change. At
562 // the end of the main iteration loop it will contain at least all the phi of
563 // ops instructions that will be changed to phis, as well as regular phis.
564 // During the iteration loop, it may contain other things, such as phi of ops
565 // instructions that used edge reachability to reach a result, and so need to
566 // be revisited when the edge changes, independent of whether the phi they
567 // depended on changes.
568 DenseMap<BasicBlock *, SparseBitVector<>> RevisitOnReachabilityChange;
570 // Mapping from predicate info we used to the instructions we used it with.
571 // In order to correctly ensure propagation, we must keep track of what
572 // comparisons we used, so that when the values of the comparisons change, we
573 // propagate the information to the places we used the comparison.
574 mutable DenseMap<const Value *, SmallPtrSet<Instruction *, 2>>
577 // the same reasoning as PredicateToUsers. When we skip MemoryAccesses for
578 // stores, we no longer can rely solely on the def-use chains of MemorySSA.
579 mutable DenseMap<const MemoryAccess *, SmallPtrSet<MemoryAccess *, 2>>
582 // A table storing which memorydefs/phis represent a memory state provably
583 // equivalent to another memory state.
584 // We could use the congruence class machinery, but the MemoryAccess's are
585 // abstract memory states, so they can only ever be equivalent to each other,
586 // and not to constants, etc.
587 DenseMap<const MemoryAccess *, CongruenceClass *> MemoryAccessToClass;
589 // We could, if we wanted, build MemoryPhiExpressions and
590 // MemoryVariableExpressions, etc, and value number them the same way we value
591 // number phi expressions. For the moment, this seems like overkill. They
592 // can only exist in one of three states: they can be TOP (equal to
593 // everything), Equivalent to something else, or unique. Because we do not
594 // create expressions for them, we need to simulate leader change not just
595 // when they change class, but when they change state. Note: We can do the
596 // same thing for phis, and avoid having phi expressions if we wanted, We
597 // should eventually unify in one direction or the other, so this is a little
598 // bit of an experiment in which turns out easier to maintain.
599 enum MemoryPhiState { MPS_Invalid, MPS_TOP, MPS_Equivalent, MPS_Unique };
600 DenseMap<const MemoryPhi *, MemoryPhiState> MemoryPhiState;
602 enum InstCycleState { ICS_Unknown, ICS_CycleFree, ICS_Cycle };
603 mutable DenseMap<const Instruction *, InstCycleState> InstCycleState;
605 // Expression to class mapping.
606 using ExpressionClassMap = DenseMap<const Expression *, CongruenceClass *>;
607 ExpressionClassMap ExpressionToClass;
609 // We have a single expression that represents currently DeadExpressions.
610 // For dead expressions we can prove will stay dead, we mark them with
611 // DFS number zero. However, it's possible in the case of phi nodes
612 // for us to assume/prove all arguments are dead during fixpointing.
613 // We use DeadExpression for that case.
614 DeadExpression *SingletonDeadExpression = nullptr;
616 // Which values have changed as a result of leader changes.
617 SmallPtrSet<Value *, 8> LeaderChanges;
619 // Reachability info.
620 using BlockEdge = BasicBlockEdge;
621 DenseSet<BlockEdge> ReachableEdges;
622 SmallPtrSet<const BasicBlock *, 8> ReachableBlocks;
624 // This is a bitvector because, on larger functions, we may have
625 // thousands of touched instructions at once (entire blocks,
626 // instructions with hundreds of uses, etc). Even with optimization
627 // for when we mark whole blocks as touched, when this was a
628 // SmallPtrSet or DenseSet, for some functions, we spent >20% of all
629 // the time in GVN just managing this list. The bitvector, on the
630 // other hand, efficiently supports test/set/clear of both
631 // individual and ranges, as well as "find next element" This
632 // enables us to use it as a worklist with essentially 0 cost.
633 BitVector TouchedInstructions;
635 DenseMap<const BasicBlock *, std::pair<unsigned, unsigned>> BlockInstRange;
638 // Debugging for how many times each block and instruction got processed.
639 DenseMap<const Value *, unsigned> ProcessedCount;
643 // This contains a mapping from Instructions to DFS numbers.
644 // The numbering starts at 1. An instruction with DFS number zero
645 // means that the instruction is dead.
646 DenseMap<const Value *, unsigned> InstrDFS;
648 // This contains the mapping DFS numbers to instructions.
649 SmallVector<Value *, 32> DFSToInstr;
652 SmallPtrSet<Instruction *, 8> InstructionsToErase;
655 NewGVN(Function &F, DominatorTree *DT, AssumptionCache *AC,
656 TargetLibraryInfo *TLI, AliasAnalysis *AA, MemorySSA *MSSA,
657 const DataLayout &DL)
658 : F(F), DT(DT), TLI(TLI), AA(AA), MSSA(MSSA), DL(DL),
659 PredInfo(make_unique<PredicateInfo>(F, *DT, *AC)), SQ(DL, TLI, DT, AC) {
665 // Expression handling.
666 const Expression *createExpression(Instruction *) const;
667 const Expression *createBinaryExpression(unsigned, Type *, Value *, Value *,
668 Instruction *) const;
670 // Our canonical form for phi arguments is a pair of incoming value, incoming
672 using ValPair = std::pair<Value *, BasicBlock *>;
674 PHIExpression *createPHIExpression(ArrayRef<ValPair>, const Instruction *,
675 BasicBlock *, bool &HasBackEdge,
676 bool &OriginalOpsConstant) const;
677 const DeadExpression *createDeadExpression() const;
678 const VariableExpression *createVariableExpression(Value *) const;
679 const ConstantExpression *createConstantExpression(Constant *) const;
680 const Expression *createVariableOrConstant(Value *V) const;
681 const UnknownExpression *createUnknownExpression(Instruction *) const;
682 const StoreExpression *createStoreExpression(StoreInst *,
683 const MemoryAccess *) const;
684 LoadExpression *createLoadExpression(Type *, Value *, LoadInst *,
685 const MemoryAccess *) const;
686 const CallExpression *createCallExpression(CallInst *,
687 const MemoryAccess *) const;
688 const AggregateValueExpression *
689 createAggregateValueExpression(Instruction *) const;
690 bool setBasicExpressionInfo(Instruction *, BasicExpression *) const;
692 // Congruence class handling.
693 CongruenceClass *createCongruenceClass(Value *Leader, const Expression *E) {
694 auto *result = new CongruenceClass(NextCongruenceNum++, Leader, E);
695 CongruenceClasses.emplace_back(result);
699 CongruenceClass *createMemoryClass(MemoryAccess *MA) {
700 auto *CC = createCongruenceClass(nullptr, nullptr);
701 CC->setMemoryLeader(MA);
705 CongruenceClass *ensureLeaderOfMemoryClass(MemoryAccess *MA) {
706 auto *CC = getMemoryClass(MA);
707 if (CC->getMemoryLeader() != MA)
708 CC = createMemoryClass(MA);
712 CongruenceClass *createSingletonCongruenceClass(Value *Member) {
713 CongruenceClass *CClass = createCongruenceClass(Member, nullptr);
714 CClass->insert(Member);
715 ValueToClass[Member] = CClass;
719 void initializeCongruenceClasses(Function &F);
720 const Expression *makePossiblePHIOfOps(Instruction *,
721 SmallPtrSetImpl<Value *> &);
722 Value *findLeaderForInst(Instruction *ValueOp,
723 SmallPtrSetImpl<Value *> &Visited,
724 MemoryAccess *MemAccess, Instruction *OrigInst,
726 bool OpIsSafeForPHIOfOpsHelper(Value *V, const BasicBlock *PHIBlock,
727 SmallPtrSetImpl<const Value *> &Visited,
728 SmallVectorImpl<Instruction *> &Worklist);
729 bool OpIsSafeForPHIOfOps(Value *Op, const BasicBlock *PHIBlock,
730 SmallPtrSetImpl<const Value *> &);
731 void addPhiOfOps(PHINode *Op, BasicBlock *BB, Instruction *ExistingValue);
732 void removePhiOfOps(Instruction *I, PHINode *PHITemp);
734 // Value number an Instruction or MemoryPhi.
735 void valueNumberMemoryPhi(MemoryPhi *);
736 void valueNumberInstruction(Instruction *);
738 // Symbolic evaluation.
739 const Expression *checkSimplificationResults(Expression *, Instruction *,
741 const Expression *performSymbolicEvaluation(Value *,
742 SmallPtrSetImpl<Value *> &) const;
743 const Expression *performSymbolicLoadCoercion(Type *, Value *, LoadInst *,
745 MemoryAccess *) const;
746 const Expression *performSymbolicLoadEvaluation(Instruction *) const;
747 const Expression *performSymbolicStoreEvaluation(Instruction *) const;
748 const Expression *performSymbolicCallEvaluation(Instruction *) const;
749 void sortPHIOps(MutableArrayRef<ValPair> Ops) const;
750 const Expression *performSymbolicPHIEvaluation(ArrayRef<ValPair>,
752 BasicBlock *PHIBlock) const;
753 const Expression *performSymbolicAggrValueEvaluation(Instruction *) const;
754 const Expression *performSymbolicCmpEvaluation(Instruction *) const;
755 const Expression *performSymbolicPredicateInfoEvaluation(Instruction *) const;
757 // Congruence finding.
758 bool someEquivalentDominates(const Instruction *, const Instruction *) const;
759 Value *lookupOperandLeader(Value *) const;
760 CongruenceClass *getClassForExpression(const Expression *E) const;
761 void performCongruenceFinding(Instruction *, const Expression *);
762 void moveValueToNewCongruenceClass(Instruction *, const Expression *,
763 CongruenceClass *, CongruenceClass *);
764 void moveMemoryToNewCongruenceClass(Instruction *, MemoryAccess *,
765 CongruenceClass *, CongruenceClass *);
766 Value *getNextValueLeader(CongruenceClass *) const;
767 const MemoryAccess *getNextMemoryLeader(CongruenceClass *) const;
768 bool setMemoryClass(const MemoryAccess *From, CongruenceClass *To);
769 CongruenceClass *getMemoryClass(const MemoryAccess *MA) const;
770 const MemoryAccess *lookupMemoryLeader(const MemoryAccess *) const;
771 bool isMemoryAccessTOP(const MemoryAccess *) const;
774 unsigned int getRank(const Value *) const;
775 bool shouldSwapOperands(const Value *, const Value *) const;
777 // Reachability handling.
778 void updateReachableEdge(BasicBlock *, BasicBlock *);
779 void processOutgoingEdges(TerminatorInst *, BasicBlock *);
780 Value *findConditionEquivalence(Value *) const;
784 void convertClassToDFSOrdered(const CongruenceClass &,
785 SmallVectorImpl<ValueDFS> &,
786 DenseMap<const Value *, unsigned int> &,
787 SmallPtrSetImpl<Instruction *> &) const;
788 void convertClassToLoadsAndStores(const CongruenceClass &,
789 SmallVectorImpl<ValueDFS> &) const;
791 bool eliminateInstructions(Function &);
792 void replaceInstruction(Instruction *, Value *);
793 void markInstructionForDeletion(Instruction *);
794 void deleteInstructionsInBlock(BasicBlock *);
795 Value *findPHIOfOpsLeader(const Expression *, const Instruction *,
796 const BasicBlock *) const;
798 // New instruction creation.
799 void handleNewInstruction(Instruction *) {}
801 // Various instruction touch utilities
802 template <typename Map, typename KeyType, typename Func>
803 void for_each_found(Map &, const KeyType &, Func);
804 template <typename Map, typename KeyType>
805 void touchAndErase(Map &, const KeyType &);
806 void markUsersTouched(Value *);
807 void markMemoryUsersTouched(const MemoryAccess *);
808 void markMemoryDefTouched(const MemoryAccess *);
809 void markPredicateUsersTouched(Instruction *);
810 void markValueLeaderChangeTouched(CongruenceClass *CC);
811 void markMemoryLeaderChangeTouched(CongruenceClass *CC);
812 void markPhiOfOpsChanged(const Expression *E);
813 void addPredicateUsers(const PredicateBase *, Instruction *) const;
814 void addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const;
815 void addAdditionalUsers(Value *To, Value *User) const;
817 // Main loop of value numbering
818 void iterateTouchedInstructions();
821 void cleanupTables();
822 std::pair<unsigned, unsigned> assignDFSNumbers(BasicBlock *, unsigned);
823 void updateProcessedCount(const Value *V);
824 void verifyMemoryCongruency() const;
825 void verifyIterationSettled(Function &F);
826 void verifyStoreExpressions() const;
827 bool singleReachablePHIPath(SmallPtrSet<const MemoryAccess *, 8> &,
828 const MemoryAccess *, const MemoryAccess *) const;
829 BasicBlock *getBlockForValue(Value *V) const;
830 void deleteExpression(const Expression *E) const;
831 MemoryUseOrDef *getMemoryAccess(const Instruction *) const;
832 MemoryAccess *getDefiningAccess(const MemoryAccess *) const;
833 MemoryPhi *getMemoryAccess(const BasicBlock *) const;
834 template <class T, class Range> T *getMinDFSOfRange(const Range &) const;
836 unsigned InstrToDFSNum(const Value *V) const {
837 assert(isa<Instruction>(V) && "This should not be used for MemoryAccesses");
838 return InstrDFS.lookup(V);
841 unsigned InstrToDFSNum(const MemoryAccess *MA) const {
842 return MemoryToDFSNum(MA);
845 Value *InstrFromDFSNum(unsigned DFSNum) { return DFSToInstr[DFSNum]; }
847 // Given a MemoryAccess, return the relevant instruction DFS number. Note:
848 // This deliberately takes a value so it can be used with Use's, which will
849 // auto-convert to Value's but not to MemoryAccess's.
850 unsigned MemoryToDFSNum(const Value *MA) const {
851 assert(isa<MemoryAccess>(MA) &&
852 "This should not be used with instructions");
853 return isa<MemoryUseOrDef>(MA)
854 ? InstrToDFSNum(cast<MemoryUseOrDef>(MA)->getMemoryInst())
855 : InstrDFS.lookup(MA);
858 bool isCycleFree(const Instruction *) const;
859 bool isBackedge(BasicBlock *From, BasicBlock *To) const;
861 // Debug counter info. When verifying, we have to reset the value numbering
862 // debug counter to the same state it started in to get the same results.
863 std::pair<int, int> StartingVNCounter;
866 } // end anonymous namespace
868 template <typename T>
869 static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS) {
870 if (!isa<LoadExpression>(RHS) && !isa<StoreExpression>(RHS))
872 return LHS.MemoryExpression::equals(RHS);
875 bool LoadExpression::equals(const Expression &Other) const {
876 return equalsLoadStoreHelper(*this, Other);
879 bool StoreExpression::equals(const Expression &Other) const {
880 if (!equalsLoadStoreHelper(*this, Other))
882 // Make sure that store vs store includes the value operand.
883 if (const auto *S = dyn_cast<StoreExpression>(&Other))
884 if (getStoredValue() != S->getStoredValue())
889 // Determine if the edge From->To is a backedge
890 bool NewGVN::isBackedge(BasicBlock *From, BasicBlock *To) const {
892 RPOOrdering.lookup(DT->getNode(From)) >=
893 RPOOrdering.lookup(DT->getNode(To));
897 static std::string getBlockName(const BasicBlock *B) {
898 return DOTGraphTraits<const Function *>::getSimpleNodeLabel(B, nullptr);
902 // Get a MemoryAccess for an instruction, fake or real.
903 MemoryUseOrDef *NewGVN::getMemoryAccess(const Instruction *I) const {
904 auto *Result = MSSA->getMemoryAccess(I);
905 return Result ? Result : TempToMemory.lookup(I);
908 // Get a MemoryPhi for a basic block. These are all real.
909 MemoryPhi *NewGVN::getMemoryAccess(const BasicBlock *BB) const {
910 return MSSA->getMemoryAccess(BB);
913 // Get the basic block from an instruction/memory value.
914 BasicBlock *NewGVN::getBlockForValue(Value *V) const {
915 if (auto *I = dyn_cast<Instruction>(V)) {
916 auto *Parent = I->getParent();
919 Parent = TempToBlock.lookup(V);
920 assert(Parent && "Every fake instruction should have a block");
924 auto *MP = dyn_cast<MemoryPhi>(V);
925 assert(MP && "Should have been an instruction or a MemoryPhi");
926 return MP->getBlock();
929 // Delete a definitely dead expression, so it can be reused by the expression
930 // allocator. Some of these are not in creation functions, so we have to accept
932 void NewGVN::deleteExpression(const Expression *E) const {
933 assert(isa<BasicExpression>(E));
934 auto *BE = cast<BasicExpression>(E);
935 const_cast<BasicExpression *>(BE)->deallocateOperands(ArgRecycler);
936 ExpressionAllocator.Deallocate(E);
939 // If V is a predicateinfo copy, get the thing it is a copy of.
940 static Value *getCopyOf(const Value *V) {
941 if (auto *II = dyn_cast<IntrinsicInst>(V))
942 if (II->getIntrinsicID() == Intrinsic::ssa_copy)
943 return II->getOperand(0);
947 // Return true if V is really PN, even accounting for predicateinfo copies.
948 static bool isCopyOfPHI(const Value *V, const PHINode *PN) {
949 return V == PN || getCopyOf(V) == PN;
952 static bool isCopyOfAPHI(const Value *V) {
953 auto *CO = getCopyOf(V);
954 return CO && isa<PHINode>(CO);
957 // Sort PHI Operands into a canonical order. What we use here is an RPO
958 // order. The BlockInstRange numbers are generated in an RPO walk of the basic
960 void NewGVN::sortPHIOps(MutableArrayRef<ValPair> Ops) const {
961 std::sort(Ops.begin(), Ops.end(), [&](const ValPair &P1, const ValPair &P2) {
962 return BlockInstRange.lookup(P1.second).first <
963 BlockInstRange.lookup(P2.second).first;
967 // Return true if V is a value that will always be available (IE can
968 // be placed anywhere) in the function. We don't do globals here
969 // because they are often worse to put in place.
970 static bool alwaysAvailable(Value *V) {
971 return isa<Constant>(V) || isa<Argument>(V);
974 // Create a PHIExpression from an array of {incoming edge, value} pairs. I is
975 // the original instruction we are creating a PHIExpression for (but may not be
976 // a phi node). We require, as an invariant, that all the PHIOperands in the
977 // same block are sorted the same way. sortPHIOps will sort them into a
979 PHIExpression *NewGVN::createPHIExpression(ArrayRef<ValPair> PHIOperands,
980 const Instruction *I,
981 BasicBlock *PHIBlock,
983 bool &OriginalOpsConstant) const {
984 unsigned NumOps = PHIOperands.size();
985 auto *E = new (ExpressionAllocator) PHIExpression(NumOps, PHIBlock);
987 E->allocateOperands(ArgRecycler, ExpressionAllocator);
988 E->setType(PHIOperands.begin()->first->getType());
989 E->setOpcode(Instruction::PHI);
991 // Filter out unreachable phi operands.
992 auto Filtered = make_filter_range(PHIOperands, [&](const ValPair &P) {
994 if (auto *PHIOp = dyn_cast<PHINode>(I))
995 if (isCopyOfPHI(P.first, PHIOp))
997 if (!ReachableEdges.count({BB, PHIBlock}))
999 // Things in TOPClass are equivalent to everything.
1000 if (ValueToClass.lookup(P.first) == TOPClass)
1002 OriginalOpsConstant = OriginalOpsConstant && isa<Constant>(P.first);
1003 HasBackedge = HasBackedge || isBackedge(BB, PHIBlock);
1004 return lookupOperandLeader(P.first) != I;
1006 std::transform(Filtered.begin(), Filtered.end(), op_inserter(E),
1007 [&](const ValPair &P) -> Value * {
1008 return lookupOperandLeader(P.first);
1013 // Set basic expression info (Arguments, type, opcode) for Expression
1014 // E from Instruction I in block B.
1015 bool NewGVN::setBasicExpressionInfo(Instruction *I, BasicExpression *E) const {
1016 bool AllConstant = true;
1017 if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
1018 E->setType(GEP->getSourceElementType());
1020 E->setType(I->getType());
1021 E->setOpcode(I->getOpcode());
1022 E->allocateOperands(ArgRecycler, ExpressionAllocator);
1024 // Transform the operand array into an operand leader array, and keep track of
1025 // whether all members are constant.
1026 std::transform(I->op_begin(), I->op_end(), op_inserter(E), [&](Value *O) {
1027 auto Operand = lookupOperandLeader(O);
1028 AllConstant = AllConstant && isa<Constant>(Operand);
1035 const Expression *NewGVN::createBinaryExpression(unsigned Opcode, Type *T,
1036 Value *Arg1, Value *Arg2,
1037 Instruction *I) const {
1038 auto *E = new (ExpressionAllocator) BasicExpression(2);
1041 E->setOpcode(Opcode);
1042 E->allocateOperands(ArgRecycler, ExpressionAllocator);
1043 if (Instruction::isCommutative(Opcode)) {
1044 // Ensure that commutative instructions that only differ by a permutation
1045 // of their operands get the same value number by sorting the operand value
1046 // numbers. Since all commutative instructions have two operands it is more
1047 // efficient to sort by hand rather than using, say, std::sort.
1048 if (shouldSwapOperands(Arg1, Arg2))
1049 std::swap(Arg1, Arg2);
1051 E->op_push_back(lookupOperandLeader(Arg1));
1052 E->op_push_back(lookupOperandLeader(Arg2));
1054 Value *V = SimplifyBinOp(Opcode, E->getOperand(0), E->getOperand(1), SQ);
1055 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1060 // Take a Value returned by simplification of Expression E/Instruction
1061 // I, and see if it resulted in a simpler expression. If so, return
1063 const Expression *NewGVN::checkSimplificationResults(Expression *E,
1068 if (auto *C = dyn_cast<Constant>(V)) {
1070 DEBUG(dbgs() << "Simplified " << *I << " to "
1071 << " constant " << *C << "\n");
1072 NumGVNOpsSimplified++;
1073 assert(isa<BasicExpression>(E) &&
1074 "We should always have had a basic expression here");
1075 deleteExpression(E);
1076 return createConstantExpression(C);
1077 } else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1079 DEBUG(dbgs() << "Simplified " << *I << " to "
1080 << " variable " << *V << "\n");
1081 deleteExpression(E);
1082 return createVariableExpression(V);
1085 CongruenceClass *CC = ValueToClass.lookup(V);
1087 if (CC->getLeader() && CC->getLeader() != I) {
1088 // Don't add temporary instructions to the user lists.
1089 if (!AllTempInstructions.count(I))
1090 addAdditionalUsers(V, I);
1091 return createVariableOrConstant(CC->getLeader());
1093 if (CC->getDefiningExpr()) {
1094 // If we simplified to something else, we need to communicate
1095 // that we're users of the value we simplified to.
1097 // Don't add temporary instructions to the user lists.
1098 if (!AllTempInstructions.count(I))
1099 addAdditionalUsers(V, I);
1103 DEBUG(dbgs() << "Simplified " << *I << " to "
1104 << " expression " << *CC->getDefiningExpr() << "\n");
1105 NumGVNOpsSimplified++;
1106 deleteExpression(E);
1107 return CC->getDefiningExpr();
1114 // Create a value expression from the instruction I, replacing operands with
1117 const Expression *NewGVN::createExpression(Instruction *I) const {
1118 auto *E = new (ExpressionAllocator) BasicExpression(I->getNumOperands());
1120 bool AllConstant = setBasicExpressionInfo(I, E);
1122 if (I->isCommutative()) {
1123 // Ensure that commutative instructions that only differ by a permutation
1124 // of their operands get the same value number by sorting the operand value
1125 // numbers. Since all commutative instructions have two operands it is more
1126 // efficient to sort by hand rather than using, say, std::sort.
1127 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
1128 if (shouldSwapOperands(E->getOperand(0), E->getOperand(1)))
1129 E->swapOperands(0, 1);
1131 // Perform simplification.
1132 if (auto *CI = dyn_cast<CmpInst>(I)) {
1133 // Sort the operand value numbers so x<y and y>x get the same value
1135 CmpInst::Predicate Predicate = CI->getPredicate();
1136 if (shouldSwapOperands(E->getOperand(0), E->getOperand(1))) {
1137 E->swapOperands(0, 1);
1138 Predicate = CmpInst::getSwappedPredicate(Predicate);
1140 E->setOpcode((CI->getOpcode() << 8) | Predicate);
1141 // TODO: 25% of our time is spent in SimplifyCmpInst with pointer operands
1142 assert(I->getOperand(0)->getType() == I->getOperand(1)->getType() &&
1143 "Wrong types on cmp instruction");
1144 assert((E->getOperand(0)->getType() == I->getOperand(0)->getType() &&
1145 E->getOperand(1)->getType() == I->getOperand(1)->getType()));
1147 SimplifyCmpInst(Predicate, E->getOperand(0), E->getOperand(1), SQ);
1148 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1150 } else if (isa<SelectInst>(I)) {
1151 if (isa<Constant>(E->getOperand(0)) ||
1152 E->getOperand(1) == E->getOperand(2)) {
1153 assert(E->getOperand(1)->getType() == I->getOperand(1)->getType() &&
1154 E->getOperand(2)->getType() == I->getOperand(2)->getType());
1155 Value *V = SimplifySelectInst(E->getOperand(0), E->getOperand(1),
1156 E->getOperand(2), SQ);
1157 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1160 } else if (I->isBinaryOp()) {
1162 SimplifyBinOp(E->getOpcode(), E->getOperand(0), E->getOperand(1), SQ);
1163 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1165 } else if (auto *BI = dyn_cast<BitCastInst>(I)) {
1167 SimplifyCastInst(BI->getOpcode(), BI->getOperand(0), BI->getType(), SQ);
1168 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1170 } else if (isa<GetElementPtrInst>(I)) {
1171 Value *V = SimplifyGEPInst(
1172 E->getType(), ArrayRef<Value *>(E->op_begin(), E->op_end()), SQ);
1173 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1175 } else if (AllConstant) {
1176 // We don't bother trying to simplify unless all of the operands
1178 // TODO: There are a lot of Simplify*'s we could call here, if we
1179 // wanted to. The original motivating case for this code was a
1180 // zext i1 false to i8, which we don't have an interface to
1181 // simplify (IE there is no SimplifyZExt).
1183 SmallVector<Constant *, 8> C;
1184 for (Value *Arg : E->operands())
1185 C.emplace_back(cast<Constant>(Arg));
1187 if (Value *V = ConstantFoldInstOperands(I, C, DL, TLI))
1188 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1194 const AggregateValueExpression *
1195 NewGVN::createAggregateValueExpression(Instruction *I) const {
1196 if (auto *II = dyn_cast<InsertValueInst>(I)) {
1197 auto *E = new (ExpressionAllocator)
1198 AggregateValueExpression(I->getNumOperands(), II->getNumIndices());
1199 setBasicExpressionInfo(I, E);
1200 E->allocateIntOperands(ExpressionAllocator);
1201 std::copy(II->idx_begin(), II->idx_end(), int_op_inserter(E));
1203 } else if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1204 auto *E = new (ExpressionAllocator)
1205 AggregateValueExpression(I->getNumOperands(), EI->getNumIndices());
1206 setBasicExpressionInfo(EI, E);
1207 E->allocateIntOperands(ExpressionAllocator);
1208 std::copy(EI->idx_begin(), EI->idx_end(), int_op_inserter(E));
1211 llvm_unreachable("Unhandled type of aggregate value operation");
1214 const DeadExpression *NewGVN::createDeadExpression() const {
1215 // DeadExpression has no arguments and all DeadExpression's are the same,
1216 // so we only need one of them.
1217 return SingletonDeadExpression;
1220 const VariableExpression *NewGVN::createVariableExpression(Value *V) const {
1221 auto *E = new (ExpressionAllocator) VariableExpression(V);
1222 E->setOpcode(V->getValueID());
1226 const Expression *NewGVN::createVariableOrConstant(Value *V) const {
1227 if (auto *C = dyn_cast<Constant>(V))
1228 return createConstantExpression(C);
1229 return createVariableExpression(V);
1232 const ConstantExpression *NewGVN::createConstantExpression(Constant *C) const {
1233 auto *E = new (ExpressionAllocator) ConstantExpression(C);
1234 E->setOpcode(C->getValueID());
1238 const UnknownExpression *NewGVN::createUnknownExpression(Instruction *I) const {
1239 auto *E = new (ExpressionAllocator) UnknownExpression(I);
1240 E->setOpcode(I->getOpcode());
1244 const CallExpression *
1245 NewGVN::createCallExpression(CallInst *CI, const MemoryAccess *MA) const {
1246 // FIXME: Add operand bundles for calls.
1248 new (ExpressionAllocator) CallExpression(CI->getNumOperands(), CI, MA);
1249 setBasicExpressionInfo(CI, E);
1253 // Return true if some equivalent of instruction Inst dominates instruction U.
1254 bool NewGVN::someEquivalentDominates(const Instruction *Inst,
1255 const Instruction *U) const {
1256 auto *CC = ValueToClass.lookup(Inst);
1257 // This must be an instruction because we are only called from phi nodes
1258 // in the case that the value it needs to check against is an instruction.
1260 // The most likely candiates for dominance are the leader and the next leader.
1261 // The leader or nextleader will dominate in all cases where there is an
1262 // equivalent that is higher up in the dom tree.
1263 // We can't *only* check them, however, because the
1264 // dominator tree could have an infinite number of non-dominating siblings
1265 // with instructions that are in the right congruence class.
1270 // Instruction U could be in H, with equivalents in every other sibling.
1271 // Depending on the rpo order picked, the leader could be the equivalent in
1272 // any of these siblings.
1275 if (alwaysAvailable(CC->getLeader()))
1277 if (DT->dominates(cast<Instruction>(CC->getLeader()), U))
1279 if (CC->getNextLeader().first &&
1280 DT->dominates(cast<Instruction>(CC->getNextLeader().first), U))
1282 return llvm::any_of(*CC, [&](const Value *Member) {
1283 return Member != CC->getLeader() &&
1284 DT->dominates(cast<Instruction>(Member), U);
1288 // See if we have a congruence class and leader for this operand, and if so,
1289 // return it. Otherwise, return the operand itself.
1290 Value *NewGVN::lookupOperandLeader(Value *V) const {
1291 CongruenceClass *CC = ValueToClass.lookup(V);
1293 // Everything in TOP is represented by undef, as it can be any value.
1294 // We do have to make sure we get the type right though, so we can't set the
1295 // RepLeader to undef.
1297 return UndefValue::get(V->getType());
1298 return CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
1304 const MemoryAccess *NewGVN::lookupMemoryLeader(const MemoryAccess *MA) const {
1305 auto *CC = getMemoryClass(MA);
1306 assert(CC->getMemoryLeader() &&
1307 "Every MemoryAccess should be mapped to a congruence class with a "
1308 "representative memory access");
1309 return CC->getMemoryLeader();
1312 // Return true if the MemoryAccess is really equivalent to everything. This is
1313 // equivalent to the lattice value "TOP" in most lattices. This is the initial
1314 // state of all MemoryAccesses.
1315 bool NewGVN::isMemoryAccessTOP(const MemoryAccess *MA) const {
1316 return getMemoryClass(MA) == TOPClass;
1319 LoadExpression *NewGVN::createLoadExpression(Type *LoadType, Value *PointerOp,
1321 const MemoryAccess *MA) const {
1323 new (ExpressionAllocator) LoadExpression(1, LI, lookupMemoryLeader(MA));
1324 E->allocateOperands(ArgRecycler, ExpressionAllocator);
1325 E->setType(LoadType);
1327 // Give store and loads same opcode so they value number together.
1329 E->op_push_back(PointerOp);
1331 E->setAlignment(LI->getAlignment());
1333 // TODO: Value number heap versions. We may be able to discover
1334 // things alias analysis can't on it's own (IE that a store and a
1335 // load have the same value, and thus, it isn't clobbering the load).
1339 const StoreExpression *
1340 NewGVN::createStoreExpression(StoreInst *SI, const MemoryAccess *MA) const {
1341 auto *StoredValueLeader = lookupOperandLeader(SI->getValueOperand());
1342 auto *E = new (ExpressionAllocator)
1343 StoreExpression(SI->getNumOperands(), SI, StoredValueLeader, MA);
1344 E->allocateOperands(ArgRecycler, ExpressionAllocator);
1345 E->setType(SI->getValueOperand()->getType());
1347 // Give store and loads same opcode so they value number together.
1349 E->op_push_back(lookupOperandLeader(SI->getPointerOperand()));
1351 // TODO: Value number heap versions. We may be able to discover
1352 // things alias analysis can't on it's own (IE that a store and a
1353 // load have the same value, and thus, it isn't clobbering the load).
1357 const Expression *NewGVN::performSymbolicStoreEvaluation(Instruction *I) const {
1358 // Unlike loads, we never try to eliminate stores, so we do not check if they
1359 // are simple and avoid value numbering them.
1360 auto *SI = cast<StoreInst>(I);
1361 auto *StoreAccess = getMemoryAccess(SI);
1362 // Get the expression, if any, for the RHS of the MemoryDef.
1363 const MemoryAccess *StoreRHS = StoreAccess->getDefiningAccess();
1364 if (EnableStoreRefinement)
1365 StoreRHS = MSSAWalker->getClobberingMemoryAccess(StoreAccess);
1366 // If we bypassed the use-def chains, make sure we add a use.
1367 StoreRHS = lookupMemoryLeader(StoreRHS);
1368 if (StoreRHS != StoreAccess->getDefiningAccess())
1369 addMemoryUsers(StoreRHS, StoreAccess);
1370 // If we are defined by ourselves, use the live on entry def.
1371 if (StoreRHS == StoreAccess)
1372 StoreRHS = MSSA->getLiveOnEntryDef();
1374 if (SI->isSimple()) {
1375 // See if we are defined by a previous store expression, it already has a
1376 // value, and it's the same value as our current store. FIXME: Right now, we
1377 // only do this for simple stores, we should expand to cover memcpys, etc.
1378 const auto *LastStore = createStoreExpression(SI, StoreRHS);
1379 const auto *LastCC = ExpressionToClass.lookup(LastStore);
1380 // We really want to check whether the expression we matched was a store. No
1381 // easy way to do that. However, we can check that the class we found has a
1382 // store, which, assuming the value numbering state is not corrupt, is
1383 // sufficient, because we must also be equivalent to that store's expression
1384 // for it to be in the same class as the load.
1385 if (LastCC && LastCC->getStoredValue() == LastStore->getStoredValue())
1387 // Also check if our value operand is defined by a load of the same memory
1388 // location, and the memory state is the same as it was then (otherwise, it
1389 // could have been overwritten later. See test32 in
1390 // transforms/DeadStoreElimination/simple.ll).
1391 if (auto *LI = dyn_cast<LoadInst>(LastStore->getStoredValue()))
1392 if ((lookupOperandLeader(LI->getPointerOperand()) ==
1393 LastStore->getOperand(0)) &&
1394 (lookupMemoryLeader(getMemoryAccess(LI)->getDefiningAccess()) ==
1397 deleteExpression(LastStore);
1400 // If the store is not equivalent to anything, value number it as a store that
1401 // produces a unique memory state (instead of using it's MemoryUse, we use
1403 return createStoreExpression(SI, StoreAccess);
1406 // See if we can extract the value of a loaded pointer from a load, a store, or
1407 // a memory instruction.
1409 NewGVN::performSymbolicLoadCoercion(Type *LoadType, Value *LoadPtr,
1410 LoadInst *LI, Instruction *DepInst,
1411 MemoryAccess *DefiningAccess) const {
1412 assert((!LI || LI->isSimple()) && "Not a simple load");
1413 if (auto *DepSI = dyn_cast<StoreInst>(DepInst)) {
1414 // Can't forward from non-atomic to atomic without violating memory model.
1415 // Also don't need to coerce if they are the same type, we will just
1417 if (LI->isAtomic() > DepSI->isAtomic() ||
1418 LoadType == DepSI->getValueOperand()->getType())
1420 int Offset = analyzeLoadFromClobberingStore(LoadType, LoadPtr, DepSI, DL);
1422 if (auto *C = dyn_cast<Constant>(
1423 lookupOperandLeader(DepSI->getValueOperand()))) {
1424 DEBUG(dbgs() << "Coercing load from store " << *DepSI << " to constant "
1426 return createConstantExpression(
1427 getConstantStoreValueForLoad(C, Offset, LoadType, DL));
1430 } else if (auto *DepLI = dyn_cast<LoadInst>(DepInst)) {
1431 // Can't forward from non-atomic to atomic without violating memory model.
1432 if (LI->isAtomic() > DepLI->isAtomic())
1434 int Offset = analyzeLoadFromClobberingLoad(LoadType, LoadPtr, DepLI, DL);
1436 // We can coerce a constant load into a load.
1437 if (auto *C = dyn_cast<Constant>(lookupOperandLeader(DepLI)))
1438 if (auto *PossibleConstant =
1439 getConstantLoadValueForLoad(C, Offset, LoadType, DL)) {
1440 DEBUG(dbgs() << "Coercing load from load " << *LI << " to constant "
1441 << *PossibleConstant << "\n");
1442 return createConstantExpression(PossibleConstant);
1445 } else if (auto *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
1446 int Offset = analyzeLoadFromClobberingMemInst(LoadType, LoadPtr, DepMI, DL);
1448 if (auto *PossibleConstant =
1449 getConstantMemInstValueForLoad(DepMI, Offset, LoadType, DL)) {
1450 DEBUG(dbgs() << "Coercing load from meminst " << *DepMI
1451 << " to constant " << *PossibleConstant << "\n");
1452 return createConstantExpression(PossibleConstant);
1457 // All of the below are only true if the loaded pointer is produced
1458 // by the dependent instruction.
1459 if (LoadPtr != lookupOperandLeader(DepInst) &&
1460 !AA->isMustAlias(LoadPtr, DepInst))
1462 // If this load really doesn't depend on anything, then we must be loading an
1463 // undef value. This can happen when loading for a fresh allocation with no
1464 // intervening stores, for example. Note that this is only true in the case
1465 // that the result of the allocation is pointer equal to the load ptr.
1466 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1467 return createConstantExpression(UndefValue::get(LoadType));
1469 // If this load occurs either right after a lifetime begin,
1470 // then the loaded value is undefined.
1471 else if (auto *II = dyn_cast<IntrinsicInst>(DepInst)) {
1472 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1473 return createConstantExpression(UndefValue::get(LoadType));
1475 // If this load follows a calloc (which zero initializes memory),
1476 // then the loaded value is zero
1477 else if (isCallocLikeFn(DepInst, TLI)) {
1478 return createConstantExpression(Constant::getNullValue(LoadType));
1484 const Expression *NewGVN::performSymbolicLoadEvaluation(Instruction *I) const {
1485 auto *LI = cast<LoadInst>(I);
1487 // We can eliminate in favor of non-simple loads, but we won't be able to
1488 // eliminate the loads themselves.
1489 if (!LI->isSimple())
1492 Value *LoadAddressLeader = lookupOperandLeader(LI->getPointerOperand());
1493 // Load of undef is undef.
1494 if (isa<UndefValue>(LoadAddressLeader))
1495 return createConstantExpression(UndefValue::get(LI->getType()));
1496 MemoryAccess *OriginalAccess = getMemoryAccess(I);
1497 MemoryAccess *DefiningAccess =
1498 MSSAWalker->getClobberingMemoryAccess(OriginalAccess);
1500 if (!MSSA->isLiveOnEntryDef(DefiningAccess)) {
1501 if (auto *MD = dyn_cast<MemoryDef>(DefiningAccess)) {
1502 Instruction *DefiningInst = MD->getMemoryInst();
1503 // If the defining instruction is not reachable, replace with undef.
1504 if (!ReachableBlocks.count(DefiningInst->getParent()))
1505 return createConstantExpression(UndefValue::get(LI->getType()));
1506 // This will handle stores and memory insts. We only do if it the
1507 // defining access has a different type, or it is a pointer produced by
1508 // certain memory operations that cause the memory to have a fixed value
1509 // (IE things like calloc).
1510 if (const auto *CoercionResult =
1511 performSymbolicLoadCoercion(LI->getType(), LoadAddressLeader, LI,
1512 DefiningInst, DefiningAccess))
1513 return CoercionResult;
1517 const auto *LE = createLoadExpression(LI->getType(), LoadAddressLeader, LI,
1519 // If our MemoryLeader is not our defining access, add a use to the
1520 // MemoryLeader, so that we get reprocessed when it changes.
1521 if (LE->getMemoryLeader() != DefiningAccess)
1522 addMemoryUsers(LE->getMemoryLeader(), OriginalAccess);
1527 NewGVN::performSymbolicPredicateInfoEvaluation(Instruction *I) const {
1528 auto *PI = PredInfo->getPredicateInfoFor(I);
1532 DEBUG(dbgs() << "Found predicate info from instruction !\n");
1534 auto *PWC = dyn_cast<PredicateWithCondition>(PI);
1538 auto *CopyOf = I->getOperand(0);
1539 auto *Cond = PWC->Condition;
1541 // If this a copy of the condition, it must be either true or false depending
1542 // on the predicate info type and edge.
1543 if (CopyOf == Cond) {
1544 // We should not need to add predicate users because the predicate info is
1545 // already a use of this operand.
1546 if (isa<PredicateAssume>(PI))
1547 return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1548 if (auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1549 if (PBranch->TrueEdge)
1550 return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1551 return createConstantExpression(ConstantInt::getFalse(Cond->getType()));
1553 if (auto *PSwitch = dyn_cast<PredicateSwitch>(PI))
1554 return createConstantExpression(cast<Constant>(PSwitch->CaseValue));
1557 // Not a copy of the condition, so see what the predicates tell us about this
1558 // value. First, though, we check to make sure the value is actually a copy
1559 // of one of the condition operands. It's possible, in certain cases, for it
1560 // to be a copy of a predicateinfo copy. In particular, if two branch
1561 // operations use the same condition, and one branch dominates the other, we
1562 // will end up with a copy of a copy. This is currently a small deficiency in
1563 // predicateinfo. What will end up happening here is that we will value
1564 // number both copies the same anyway.
1566 // Everything below relies on the condition being a comparison.
1567 auto *Cmp = dyn_cast<CmpInst>(Cond);
1571 if (CopyOf != Cmp->getOperand(0) && CopyOf != Cmp->getOperand(1)) {
1572 DEBUG(dbgs() << "Copy is not of any condition operands!\n");
1575 Value *FirstOp = lookupOperandLeader(Cmp->getOperand(0));
1576 Value *SecondOp = lookupOperandLeader(Cmp->getOperand(1));
1577 bool SwappedOps = false;
1579 if (shouldSwapOperands(FirstOp, SecondOp)) {
1580 std::swap(FirstOp, SecondOp);
1583 CmpInst::Predicate Predicate =
1584 SwappedOps ? Cmp->getSwappedPredicate() : Cmp->getPredicate();
1586 if (isa<PredicateAssume>(PI)) {
1587 // If the comparison is true when the operands are equal, then we know the
1588 // operands are equal, because assumes must always be true.
1589 if (CmpInst::isTrueWhenEqual(Predicate)) {
1590 addPredicateUsers(PI, I);
1591 addAdditionalUsers(Cmp->getOperand(0), I);
1592 return createVariableOrConstant(FirstOp);
1595 if (const auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1596 // If we are *not* a copy of the comparison, we may equal to the other
1597 // operand when the predicate implies something about equality of
1598 // operations. In particular, if the comparison is true/false when the
1599 // operands are equal, and we are on the right edge, we know this operation
1600 // is equal to something.
1601 if ((PBranch->TrueEdge && Predicate == CmpInst::ICMP_EQ) ||
1602 (!PBranch->TrueEdge && Predicate == CmpInst::ICMP_NE)) {
1603 addPredicateUsers(PI, I);
1604 addAdditionalUsers(SwappedOps ? Cmp->getOperand(1) : Cmp->getOperand(0),
1606 return createVariableOrConstant(FirstOp);
1608 // Handle the special case of floating point.
1609 if (((PBranch->TrueEdge && Predicate == CmpInst::FCMP_OEQ) ||
1610 (!PBranch->TrueEdge && Predicate == CmpInst::FCMP_UNE)) &&
1611 isa<ConstantFP>(FirstOp) && !cast<ConstantFP>(FirstOp)->isZero()) {
1612 addPredicateUsers(PI, I);
1613 addAdditionalUsers(SwappedOps ? Cmp->getOperand(1) : Cmp->getOperand(0),
1615 return createConstantExpression(cast<Constant>(FirstOp));
1621 // Evaluate read only and pure calls, and create an expression result.
1622 const Expression *NewGVN::performSymbolicCallEvaluation(Instruction *I) const {
1623 auto *CI = cast<CallInst>(I);
1624 if (auto *II = dyn_cast<IntrinsicInst>(I)) {
1625 // Instrinsics with the returned attribute are copies of arguments.
1626 if (auto *ReturnedValue = II->getReturnedArgOperand()) {
1627 if (II->getIntrinsicID() == Intrinsic::ssa_copy)
1628 if (const auto *Result = performSymbolicPredicateInfoEvaluation(I))
1630 return createVariableOrConstant(ReturnedValue);
1633 if (AA->doesNotAccessMemory(CI)) {
1634 return createCallExpression(CI, TOPClass->getMemoryLeader());
1635 } else if (AA->onlyReadsMemory(CI)) {
1636 MemoryAccess *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(CI);
1637 return createCallExpression(CI, DefiningAccess);
1642 // Retrieve the memory class for a given MemoryAccess.
1643 CongruenceClass *NewGVN::getMemoryClass(const MemoryAccess *MA) const {
1644 auto *Result = MemoryAccessToClass.lookup(MA);
1645 assert(Result && "Should have found memory class");
1649 // Update the MemoryAccess equivalence table to say that From is equal to To,
1650 // and return true if this is different from what already existed in the table.
1651 bool NewGVN::setMemoryClass(const MemoryAccess *From,
1652 CongruenceClass *NewClass) {
1654 "Every MemoryAccess should be getting mapped to a non-null class");
1655 DEBUG(dbgs() << "Setting " << *From);
1656 DEBUG(dbgs() << " equivalent to congruence class ");
1657 DEBUG(dbgs() << NewClass->getID() << " with current MemoryAccess leader ");
1658 DEBUG(dbgs() << *NewClass->getMemoryLeader() << "\n");
1660 auto LookupResult = MemoryAccessToClass.find(From);
1661 bool Changed = false;
1662 // If it's already in the table, see if the value changed.
1663 if (LookupResult != MemoryAccessToClass.end()) {
1664 auto *OldClass = LookupResult->second;
1665 if (OldClass != NewClass) {
1666 // If this is a phi, we have to handle memory member updates.
1667 if (auto *MP = dyn_cast<MemoryPhi>(From)) {
1668 OldClass->memory_erase(MP);
1669 NewClass->memory_insert(MP);
1670 // This may have killed the class if it had no non-memory members
1671 if (OldClass->getMemoryLeader() == From) {
1672 if (OldClass->definesNoMemory()) {
1673 OldClass->setMemoryLeader(nullptr);
1675 OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
1676 DEBUG(dbgs() << "Memory class leader change for class "
1677 << OldClass->getID() << " to "
1678 << *OldClass->getMemoryLeader()
1679 << " due to removal of a memory member " << *From
1681 markMemoryLeaderChangeTouched(OldClass);
1685 // It wasn't equivalent before, and now it is.
1686 LookupResult->second = NewClass;
1694 // Determine if a instruction is cycle-free. That means the values in the
1695 // instruction don't depend on any expressions that can change value as a result
1696 // of the instruction. For example, a non-cycle free instruction would be v =
1698 bool NewGVN::isCycleFree(const Instruction *I) const {
1699 // In order to compute cycle-freeness, we do SCC finding on the instruction,
1700 // and see what kind of SCC it ends up in. If it is a singleton, it is
1701 // cycle-free. If it is not in a singleton, it is only cycle free if the
1702 // other members are all phi nodes (as they do not compute anything, they are
1704 auto ICS = InstCycleState.lookup(I);
1705 if (ICS == ICS_Unknown) {
1707 auto &SCC = SCCFinder.getComponentFor(I);
1708 // It's cycle free if it's size 1 or the SCC is *only* phi nodes.
1709 if (SCC.size() == 1)
1710 InstCycleState.insert({I, ICS_CycleFree});
1712 bool AllPhis = llvm::all_of(SCC, [](const Value *V) {
1713 return isa<PHINode>(V) || isCopyOfAPHI(V);
1715 ICS = AllPhis ? ICS_CycleFree : ICS_Cycle;
1716 for (auto *Member : SCC)
1717 if (auto *MemberPhi = dyn_cast<PHINode>(Member))
1718 InstCycleState.insert({MemberPhi, ICS});
1721 if (ICS == ICS_Cycle)
1726 // Evaluate PHI nodes symbolically and create an expression result.
1728 NewGVN::performSymbolicPHIEvaluation(ArrayRef<ValPair> PHIOps,
1730 BasicBlock *PHIBlock) const {
1731 // True if one of the incoming phi edges is a backedge.
1732 bool HasBackedge = false;
1733 // All constant tracks the state of whether all the *original* phi operands
1734 // This is really shorthand for "this phi cannot cycle due to forward
1735 // change in value of the phi is guaranteed not to later change the value of
1736 // the phi. IE it can't be v = phi(undef, v+1)
1737 bool OriginalOpsConstant = true;
1738 auto *E = cast<PHIExpression>(createPHIExpression(
1739 PHIOps, I, PHIBlock, HasBackedge, OriginalOpsConstant));
1740 // We match the semantics of SimplifyPhiNode from InstructionSimplify here.
1741 // See if all arguments are the same.
1742 // We track if any were undef because they need special handling.
1743 bool HasUndef = false;
1744 auto Filtered = make_filter_range(E->operands(), [&](Value *Arg) {
1745 if (isa<UndefValue>(Arg)) {
1751 // If we are left with no operands, it's dead.
1752 if (Filtered.begin() == Filtered.end()) {
1753 // If it has undef at this point, it means there are no-non-undef arguments,
1754 // and thus, the value of the phi node must be undef.
1756 DEBUG(dbgs() << "PHI Node " << *I
1757 << " has no non-undef arguments, valuing it as undef\n");
1758 return createConstantExpression(UndefValue::get(I->getType()));
1761 DEBUG(dbgs() << "No arguments of PHI node " << *I << " are live\n");
1762 deleteExpression(E);
1763 return createDeadExpression();
1765 Value *AllSameValue = *(Filtered.begin());
1767 // Can't use std::equal here, sadly, because filter.begin moves.
1768 if (llvm::all_of(Filtered, [&](Value *Arg) { return Arg == AllSameValue; })) {
1769 // In LLVM's non-standard representation of phi nodes, it's possible to have
1770 // phi nodes with cycles (IE dependent on other phis that are .... dependent
1771 // on the original phi node), especially in weird CFG's where some arguments
1772 // are unreachable, or uninitialized along certain paths. This can cause
1773 // infinite loops during evaluation. We work around this by not trying to
1774 // really evaluate them independently, but instead using a variable
1775 // expression to say if one is equivalent to the other.
1776 // We also special case undef, so that if we have an undef, we can't use the
1777 // common value unless it dominates the phi block.
1779 // If we have undef and at least one other value, this is really a
1780 // multivalued phi, and we need to know if it's cycle free in order to
1781 // evaluate whether we can ignore the undef. The other parts of this are
1782 // just shortcuts. If there is no backedge, or all operands are
1783 // constants, it also must be cycle free.
1784 if (HasBackedge && !OriginalOpsConstant &&
1785 !isa<UndefValue>(AllSameValue) && !isCycleFree(I))
1788 // Only have to check for instructions
1789 if (auto *AllSameInst = dyn_cast<Instruction>(AllSameValue))
1790 if (!someEquivalentDominates(AllSameInst, I))
1793 // Can't simplify to something that comes later in the iteration.
1794 // Otherwise, when and if it changes congruence class, we will never catch
1795 // up. We will always be a class behind it.
1796 if (isa<Instruction>(AllSameValue) &&
1797 InstrToDFSNum(AllSameValue) > InstrToDFSNum(I))
1799 NumGVNPhisAllSame++;
1800 DEBUG(dbgs() << "Simplified PHI node " << *I << " to " << *AllSameValue
1802 deleteExpression(E);
1803 return createVariableOrConstant(AllSameValue);
1809 NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) const {
1810 if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1811 auto *II = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
1812 if (II && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
1813 unsigned Opcode = 0;
1814 // EI might be an extract from one of our recognised intrinsics. If it
1815 // is we'll synthesize a semantically equivalent expression instead on
1816 // an extract value expression.
1817 switch (II->getIntrinsicID()) {
1818 case Intrinsic::sadd_with_overflow:
1819 case Intrinsic::uadd_with_overflow:
1820 Opcode = Instruction::Add;
1822 case Intrinsic::ssub_with_overflow:
1823 case Intrinsic::usub_with_overflow:
1824 Opcode = Instruction::Sub;
1826 case Intrinsic::smul_with_overflow:
1827 case Intrinsic::umul_with_overflow:
1828 Opcode = Instruction::Mul;
1835 // Intrinsic recognized. Grab its args to finish building the
1837 assert(II->getNumArgOperands() == 2 &&
1838 "Expect two args for recognised intrinsics.");
1839 return createBinaryExpression(Opcode, EI->getType(),
1840 II->getArgOperand(0),
1841 II->getArgOperand(1), I);
1846 return createAggregateValueExpression(I);
1849 const Expression *NewGVN::performSymbolicCmpEvaluation(Instruction *I) const {
1850 assert(isa<CmpInst>(I) && "Expected a cmp instruction.");
1852 auto *CI = cast<CmpInst>(I);
1853 // See if our operands are equal to those of a previous predicate, and if so,
1854 // if it implies true or false.
1855 auto Op0 = lookupOperandLeader(CI->getOperand(0));
1856 auto Op1 = lookupOperandLeader(CI->getOperand(1));
1857 auto OurPredicate = CI->getPredicate();
1858 if (shouldSwapOperands(Op0, Op1)) {
1859 std::swap(Op0, Op1);
1860 OurPredicate = CI->getSwappedPredicate();
1863 // Avoid processing the same info twice.
1864 const PredicateBase *LastPredInfo = nullptr;
1865 // See if we know something about the comparison itself, like it is the target
1867 auto *CmpPI = PredInfo->getPredicateInfoFor(I);
1868 if (dyn_cast_or_null<PredicateAssume>(CmpPI))
1869 return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1872 // This condition does not depend on predicates, no need to add users
1873 if (CI->isTrueWhenEqual())
1874 return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1875 else if (CI->isFalseWhenEqual())
1876 return createConstantExpression(ConstantInt::getFalse(CI->getType()));
1879 // NOTE: Because we are comparing both operands here and below, and using
1880 // previous comparisons, we rely on fact that predicateinfo knows to mark
1881 // comparisons that use renamed operands as users of the earlier comparisons.
1882 // It is *not* enough to just mark predicateinfo renamed operands as users of
1883 // the earlier comparisons, because the *other* operand may have changed in a
1884 // previous iteration.
1887 // %b.0 = ssa.copy(%b)
1889 // icmp slt %c, %b.0
1891 // %c and %a may start out equal, and thus, the code below will say the second
1892 // %icmp is false. c may become equal to something else, and in that case the
1893 // %second icmp *must* be reexamined, but would not if only the renamed
1894 // %operands are considered users of the icmp.
1896 // *Currently* we only check one level of comparisons back, and only mark one
1897 // level back as touched when changes happen. If you modify this code to look
1898 // back farther through comparisons, you *must* mark the appropriate
1899 // comparisons as users in PredicateInfo.cpp, or you will cause bugs. See if
1900 // we know something just from the operands themselves
1902 // See if our operands have predicate info, so that we may be able to derive
1903 // something from a previous comparison.
1904 for (const auto &Op : CI->operands()) {
1905 auto *PI = PredInfo->getPredicateInfoFor(Op);
1906 if (const auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI)) {
1907 if (PI == LastPredInfo)
1910 // In phi of ops cases, we may have predicate info that we are evaluating
1911 // in a different context.
1912 if (!DT->dominates(PBranch->To, getBlockForValue(I)))
1914 // TODO: Along the false edge, we may know more things too, like
1916 // same operands is false.
1917 // TODO: We only handle actual comparison conditions below, not
1919 auto *BranchCond = dyn_cast<CmpInst>(PBranch->Condition);
1922 auto *BranchOp0 = lookupOperandLeader(BranchCond->getOperand(0));
1923 auto *BranchOp1 = lookupOperandLeader(BranchCond->getOperand(1));
1924 auto BranchPredicate = BranchCond->getPredicate();
1925 if (shouldSwapOperands(BranchOp0, BranchOp1)) {
1926 std::swap(BranchOp0, BranchOp1);
1927 BranchPredicate = BranchCond->getSwappedPredicate();
1929 if (BranchOp0 == Op0 && BranchOp1 == Op1) {
1930 if (PBranch->TrueEdge) {
1931 // If we know the previous predicate is true and we are in the true
1932 // edge then we may be implied true or false.
1933 if (CmpInst::isImpliedTrueByMatchingCmp(BranchPredicate,
1935 addPredicateUsers(PI, I);
1936 return createConstantExpression(
1937 ConstantInt::getTrue(CI->getType()));
1940 if (CmpInst::isImpliedFalseByMatchingCmp(BranchPredicate,
1942 addPredicateUsers(PI, I);
1943 return createConstantExpression(
1944 ConstantInt::getFalse(CI->getType()));
1947 // Just handle the ne and eq cases, where if we have the same
1948 // operands, we may know something.
1949 if (BranchPredicate == OurPredicate) {
1950 addPredicateUsers(PI, I);
1951 // Same predicate, same ops,we know it was false, so this is false.
1952 return createConstantExpression(
1953 ConstantInt::getFalse(CI->getType()));
1954 } else if (BranchPredicate ==
1955 CmpInst::getInversePredicate(OurPredicate)) {
1956 addPredicateUsers(PI, I);
1957 // Inverse predicate, we know the other was false, so this is true.
1958 return createConstantExpression(
1959 ConstantInt::getTrue(CI->getType()));
1965 // Create expression will take care of simplifyCmpInst
1966 return createExpression(I);
1969 // Substitute and symbolize the value before value numbering.
1971 NewGVN::performSymbolicEvaluation(Value *V,
1972 SmallPtrSetImpl<Value *> &Visited) const {
1973 const Expression *E = nullptr;
1974 if (auto *C = dyn_cast<Constant>(V))
1975 E = createConstantExpression(C);
1976 else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1977 E = createVariableExpression(V);
1979 // TODO: memory intrinsics.
1980 // TODO: Some day, we should do the forward propagation and reassociation
1981 // parts of the algorithm.
1982 auto *I = cast<Instruction>(V);
1983 switch (I->getOpcode()) {
1984 case Instruction::ExtractValue:
1985 case Instruction::InsertValue:
1986 E = performSymbolicAggrValueEvaluation(I);
1988 case Instruction::PHI: {
1989 SmallVector<ValPair, 3> Ops;
1990 auto *PN = cast<PHINode>(I);
1991 for (unsigned i = 0; i < PN->getNumOperands(); ++i)
1992 Ops.push_back({PN->getIncomingValue(i), PN->getIncomingBlock(i)});
1993 // Sort to ensure the invariant createPHIExpression requires is met.
1995 E = performSymbolicPHIEvaluation(Ops, I, getBlockForValue(I));
1997 case Instruction::Call:
1998 E = performSymbolicCallEvaluation(I);
2000 case Instruction::Store:
2001 E = performSymbolicStoreEvaluation(I);
2003 case Instruction::Load:
2004 E = performSymbolicLoadEvaluation(I);
2006 case Instruction::BitCast:
2007 E = createExpression(I);
2009 case Instruction::ICmp:
2010 case Instruction::FCmp:
2011 E = performSymbolicCmpEvaluation(I);
2013 case Instruction::Add:
2014 case Instruction::FAdd:
2015 case Instruction::Sub:
2016 case Instruction::FSub:
2017 case Instruction::Mul:
2018 case Instruction::FMul:
2019 case Instruction::UDiv:
2020 case Instruction::SDiv:
2021 case Instruction::FDiv:
2022 case Instruction::URem:
2023 case Instruction::SRem:
2024 case Instruction::FRem:
2025 case Instruction::Shl:
2026 case Instruction::LShr:
2027 case Instruction::AShr:
2028 case Instruction::And:
2029 case Instruction::Or:
2030 case Instruction::Xor:
2031 case Instruction::Trunc:
2032 case Instruction::ZExt:
2033 case Instruction::SExt:
2034 case Instruction::FPToUI:
2035 case Instruction::FPToSI:
2036 case Instruction::UIToFP:
2037 case Instruction::SIToFP:
2038 case Instruction::FPTrunc:
2039 case Instruction::FPExt:
2040 case Instruction::PtrToInt:
2041 case Instruction::IntToPtr:
2042 case Instruction::Select:
2043 case Instruction::ExtractElement:
2044 case Instruction::InsertElement:
2045 case Instruction::ShuffleVector:
2046 case Instruction::GetElementPtr:
2047 E = createExpression(I);
2056 // Look up a container in a map, and then call a function for each thing in the
2058 template <typename Map, typename KeyType, typename Func>
2059 void NewGVN::for_each_found(Map &M, const KeyType &Key, Func F) {
2060 const auto Result = M.find_as(Key);
2061 if (Result != M.end())
2062 for (typename Map::mapped_type::value_type Mapped : Result->second)
2066 // Look up a container of values/instructions in a map, and touch all the
2067 // instructions in the container. Then erase value from the map.
2068 template <typename Map, typename KeyType>
2069 void NewGVN::touchAndErase(Map &M, const KeyType &Key) {
2070 const auto Result = M.find_as(Key);
2071 if (Result != M.end()) {
2072 for (const typename Map::mapped_type::value_type Mapped : Result->second)
2073 TouchedInstructions.set(InstrToDFSNum(Mapped));
2078 void NewGVN::addAdditionalUsers(Value *To, Value *User) const {
2079 assert(User && To != User);
2080 if (isa<Instruction>(To))
2081 AdditionalUsers[To].insert(User);
2084 void NewGVN::markUsersTouched(Value *V) {
2085 // Now mark the users as touched.
2086 for (auto *User : V->users()) {
2087 assert(isa<Instruction>(User) && "Use of value not within an instruction?");
2088 TouchedInstructions.set(InstrToDFSNum(User));
2090 touchAndErase(AdditionalUsers, V);
2093 void NewGVN::addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const {
2094 DEBUG(dbgs() << "Adding memory user " << *U << " to " << *To << "\n");
2095 MemoryToUsers[To].insert(U);
2098 void NewGVN::markMemoryDefTouched(const MemoryAccess *MA) {
2099 TouchedInstructions.set(MemoryToDFSNum(MA));
2102 void NewGVN::markMemoryUsersTouched(const MemoryAccess *MA) {
2103 if (isa<MemoryUse>(MA))
2105 for (auto U : MA->users())
2106 TouchedInstructions.set(MemoryToDFSNum(U));
2107 touchAndErase(MemoryToUsers, MA);
2110 // Add I to the set of users of a given predicate.
2111 void NewGVN::addPredicateUsers(const PredicateBase *PB, Instruction *I) const {
2112 // Don't add temporary instructions to the user lists.
2113 if (AllTempInstructions.count(I))
2116 if (auto *PBranch = dyn_cast<PredicateBranch>(PB))
2117 PredicateToUsers[PBranch->Condition].insert(I);
2118 else if (auto *PAssume = dyn_cast<PredicateBranch>(PB))
2119 PredicateToUsers[PAssume->Condition].insert(I);
2122 // Touch all the predicates that depend on this instruction.
2123 void NewGVN::markPredicateUsersTouched(Instruction *I) {
2124 touchAndErase(PredicateToUsers, I);
2127 // Mark users affected by a memory leader change.
2128 void NewGVN::markMemoryLeaderChangeTouched(CongruenceClass *CC) {
2129 for (auto M : CC->memory())
2130 markMemoryDefTouched(M);
2133 // Touch the instructions that need to be updated after a congruence class has a
2134 // leader change, and mark changed values.
2135 void NewGVN::markValueLeaderChangeTouched(CongruenceClass *CC) {
2136 for (auto M : *CC) {
2137 if (auto *I = dyn_cast<Instruction>(M))
2138 TouchedInstructions.set(InstrToDFSNum(I));
2139 LeaderChanges.insert(M);
2143 // Give a range of things that have instruction DFS numbers, this will return
2144 // the member of the range with the smallest dfs number.
2145 template <class T, class Range>
2146 T *NewGVN::getMinDFSOfRange(const Range &R) const {
2147 std::pair<T *, unsigned> MinDFS = {nullptr, ~0U};
2148 for (const auto X : R) {
2149 auto DFSNum = InstrToDFSNum(X);
2150 if (DFSNum < MinDFS.second)
2151 MinDFS = {X, DFSNum};
2153 return MinDFS.first;
2156 // This function returns the MemoryAccess that should be the next leader of
2157 // congruence class CC, under the assumption that the current leader is going to
2159 const MemoryAccess *NewGVN::getNextMemoryLeader(CongruenceClass *CC) const {
2160 // TODO: If this ends up to slow, we can maintain a next memory leader like we
2161 // do for regular leaders.
2162 // Make sure there will be a leader to find.
2163 assert(!CC->definesNoMemory() && "Can't get next leader if there is none");
2164 if (CC->getStoreCount() > 0) {
2165 if (auto *NL = dyn_cast_or_null<StoreInst>(CC->getNextLeader().first))
2166 return getMemoryAccess(NL);
2167 // Find the store with the minimum DFS number.
2168 auto *V = getMinDFSOfRange<Value>(make_filter_range(
2169 *CC, [&](const Value *V) { return isa<StoreInst>(V); }));
2170 return getMemoryAccess(cast<StoreInst>(V));
2172 assert(CC->getStoreCount() == 0);
2174 // Given our assertion, hitting this part must mean
2175 // !OldClass->memory_empty()
2176 if (CC->memory_size() == 1)
2177 return *CC->memory_begin();
2178 return getMinDFSOfRange<const MemoryPhi>(CC->memory());
2181 // This function returns the next value leader of a congruence class, under the
2182 // assumption that the current leader is going away. This should end up being
2183 // the next most dominating member.
2184 Value *NewGVN::getNextValueLeader(CongruenceClass *CC) const {
2185 // We don't need to sort members if there is only 1, and we don't care about
2186 // sorting the TOP class because everything either gets out of it or is
2189 if (CC->size() == 1 || CC == TOPClass) {
2190 return *(CC->begin());
2191 } else if (CC->getNextLeader().first) {
2192 ++NumGVNAvoidedSortedLeaderChanges;
2193 return CC->getNextLeader().first;
2195 ++NumGVNSortedLeaderChanges;
2196 // NOTE: If this ends up to slow, we can maintain a dual structure for
2197 // member testing/insertion, or keep things mostly sorted, and sort only
2198 // here, or use SparseBitVector or ....
2199 return getMinDFSOfRange<Value>(*CC);
2203 // Move a MemoryAccess, currently in OldClass, to NewClass, including updates to
2204 // the memory members, etc for the move.
2206 // The invariants of this function are:
2208 // - I must be moving to NewClass from OldClass
2209 // - The StoreCount of OldClass and NewClass is expected to have been updated
2210 // for I already if it is a store.
2211 // - The OldClass memory leader has not been updated yet if I was the leader.
2212 void NewGVN::moveMemoryToNewCongruenceClass(Instruction *I,
2213 MemoryAccess *InstMA,
2214 CongruenceClass *OldClass,
2215 CongruenceClass *NewClass) {
2216 // If the leader is I, and we had a represenative MemoryAccess, it should
2217 // be the MemoryAccess of OldClass.
2218 assert((!InstMA || !OldClass->getMemoryLeader() ||
2219 OldClass->getLeader() != I ||
2220 MemoryAccessToClass.lookup(OldClass->getMemoryLeader()) ==
2221 MemoryAccessToClass.lookup(InstMA)) &&
2222 "Representative MemoryAccess mismatch");
2223 // First, see what happens to the new class
2224 if (!NewClass->getMemoryLeader()) {
2225 // Should be a new class, or a store becoming a leader of a new class.
2226 assert(NewClass->size() == 1 ||
2227 (isa<StoreInst>(I) && NewClass->getStoreCount() == 1));
2228 NewClass->setMemoryLeader(InstMA);
2229 // Mark it touched if we didn't just create a singleton
2230 DEBUG(dbgs() << "Memory class leader change for class " << NewClass->getID()
2231 << " due to new memory instruction becoming leader\n");
2232 markMemoryLeaderChangeTouched(NewClass);
2234 setMemoryClass(InstMA, NewClass);
2235 // Now, fixup the old class if necessary
2236 if (OldClass->getMemoryLeader() == InstMA) {
2237 if (!OldClass->definesNoMemory()) {
2238 OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
2239 DEBUG(dbgs() << "Memory class leader change for class "
2240 << OldClass->getID() << " to "
2241 << *OldClass->getMemoryLeader()
2242 << " due to removal of old leader " << *InstMA << "\n");
2243 markMemoryLeaderChangeTouched(OldClass);
2245 OldClass->setMemoryLeader(nullptr);
2249 // Move a value, currently in OldClass, to be part of NewClass
2250 // Update OldClass and NewClass for the move (including changing leaders, etc).
2251 void NewGVN::moveValueToNewCongruenceClass(Instruction *I, const Expression *E,
2252 CongruenceClass *OldClass,
2253 CongruenceClass *NewClass) {
2254 if (I == OldClass->getNextLeader().first)
2255 OldClass->resetNextLeader();
2258 NewClass->insert(I);
2260 if (NewClass->getLeader() != I)
2261 NewClass->addPossibleNextLeader({I, InstrToDFSNum(I)});
2262 // Handle our special casing of stores.
2263 if (auto *SI = dyn_cast<StoreInst>(I)) {
2264 OldClass->decStoreCount();
2265 // Okay, so when do we want to make a store a leader of a class?
2266 // If we have a store defined by an earlier load, we want the earlier load
2267 // to lead the class.
2268 // If we have a store defined by something else, we want the store to lead
2269 // the class so everything else gets the "something else" as a value.
2270 // If we have a store as the single member of the class, we want the store
2272 if (NewClass->getStoreCount() == 0 && !NewClass->getStoredValue()) {
2273 // If it's a store expression we are using, it means we are not equivalent
2274 // to something earlier.
2275 if (auto *SE = dyn_cast<StoreExpression>(E)) {
2276 NewClass->setStoredValue(SE->getStoredValue());
2277 markValueLeaderChangeTouched(NewClass);
2278 // Shift the new class leader to be the store
2279 DEBUG(dbgs() << "Changing leader of congruence class "
2280 << NewClass->getID() << " from " << *NewClass->getLeader()
2281 << " to " << *SI << " because store joined class\n");
2282 // If we changed the leader, we have to mark it changed because we don't
2283 // know what it will do to symbolic evaluation.
2284 NewClass->setLeader(SI);
2286 // We rely on the code below handling the MemoryAccess change.
2288 NewClass->incStoreCount();
2290 // True if there is no memory instructions left in a class that had memory
2291 // instructions before.
2293 // If it's not a memory use, set the MemoryAccess equivalence
2294 auto *InstMA = dyn_cast_or_null<MemoryDef>(getMemoryAccess(I));
2296 moveMemoryToNewCongruenceClass(I, InstMA, OldClass, NewClass);
2297 ValueToClass[I] = NewClass;
2298 // See if we destroyed the class or need to swap leaders.
2299 if (OldClass->empty() && OldClass != TOPClass) {
2300 if (OldClass->getDefiningExpr()) {
2301 DEBUG(dbgs() << "Erasing expression " << *OldClass->getDefiningExpr()
2302 << " from table\n");
2303 // We erase it as an exact expression to make sure we don't just erase an
2305 auto Iter = ExpressionToClass.find_as(
2306 ExactEqualsExpression(*OldClass->getDefiningExpr()));
2307 if (Iter != ExpressionToClass.end())
2308 ExpressionToClass.erase(Iter);
2309 #ifdef EXPENSIVE_CHECKS
2311 (*OldClass->getDefiningExpr() != *E || ExpressionToClass.lookup(E)) &&
2312 "We erased the expression we just inserted, which should not happen");
2315 } else if (OldClass->getLeader() == I) {
2316 // When the leader changes, the value numbering of
2317 // everything may change due to symbolization changes, so we need to
2319 DEBUG(dbgs() << "Value class leader change for class " << OldClass->getID()
2321 ++NumGVNLeaderChanges;
2322 // Destroy the stored value if there are no more stores to represent it.
2323 // Note that this is basically clean up for the expression removal that
2324 // happens below. If we remove stores from a class, we may leave it as a
2325 // class of equivalent memory phis.
2326 if (OldClass->getStoreCount() == 0) {
2327 if (OldClass->getStoredValue())
2328 OldClass->setStoredValue(nullptr);
2330 OldClass->setLeader(getNextValueLeader(OldClass));
2331 OldClass->resetNextLeader();
2332 markValueLeaderChangeTouched(OldClass);
2336 // For a given expression, mark the phi of ops instructions that could have
2337 // changed as a result.
2338 void NewGVN::markPhiOfOpsChanged(const Expression *E) {
2339 touchAndErase(ExpressionToPhiOfOps, E);
2342 // Perform congruence finding on a given value numbering expression.
2343 void NewGVN::performCongruenceFinding(Instruction *I, const Expression *E) {
2344 // This is guaranteed to return something, since it will at least find
2347 CongruenceClass *IClass = ValueToClass.lookup(I);
2348 assert(IClass && "Should have found a IClass");
2349 // Dead classes should have been eliminated from the mapping.
2350 assert(!IClass->isDead() && "Found a dead class");
2352 CongruenceClass *EClass = nullptr;
2353 if (const auto *VE = dyn_cast<VariableExpression>(E)) {
2354 EClass = ValueToClass.lookup(VE->getVariableValue());
2355 } else if (isa<DeadExpression>(E)) {
2359 auto lookupResult = ExpressionToClass.insert({E, nullptr});
2361 // If it's not in the value table, create a new congruence class.
2362 if (lookupResult.second) {
2363 CongruenceClass *NewClass = createCongruenceClass(nullptr, E);
2364 auto place = lookupResult.first;
2365 place->second = NewClass;
2367 // Constants and variables should always be made the leader.
2368 if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2369 NewClass->setLeader(CE->getConstantValue());
2370 } else if (const auto *SE = dyn_cast<StoreExpression>(E)) {
2371 StoreInst *SI = SE->getStoreInst();
2372 NewClass->setLeader(SI);
2373 NewClass->setStoredValue(SE->getStoredValue());
2374 // The RepMemoryAccess field will be filled in properly by the
2375 // moveValueToNewCongruenceClass call.
2377 NewClass->setLeader(I);
2379 assert(!isa<VariableExpression>(E) &&
2380 "VariableExpression should have been handled already");
2383 DEBUG(dbgs() << "Created new congruence class for " << *I
2384 << " using expression " << *E << " at " << NewClass->getID()
2385 << " and leader " << *(NewClass->getLeader()));
2386 if (NewClass->getStoredValue())
2387 DEBUG(dbgs() << " and stored value " << *(NewClass->getStoredValue()));
2388 DEBUG(dbgs() << "\n");
2390 EClass = lookupResult.first->second;
2391 if (isa<ConstantExpression>(E))
2392 assert((isa<Constant>(EClass->getLeader()) ||
2393 (EClass->getStoredValue() &&
2394 isa<Constant>(EClass->getStoredValue()))) &&
2395 "Any class with a constant expression should have a "
2398 assert(EClass && "Somehow don't have an eclass");
2400 assert(!EClass->isDead() && "We accidentally looked up a dead class");
2403 bool ClassChanged = IClass != EClass;
2404 bool LeaderChanged = LeaderChanges.erase(I);
2405 if (ClassChanged || LeaderChanged) {
2406 DEBUG(dbgs() << "New class " << EClass->getID() << " for expression " << *E
2409 moveValueToNewCongruenceClass(I, E, IClass, EClass);
2410 markPhiOfOpsChanged(E);
2413 markUsersTouched(I);
2414 if (MemoryAccess *MA = getMemoryAccess(I))
2415 markMemoryUsersTouched(MA);
2416 if (auto *CI = dyn_cast<CmpInst>(I))
2417 markPredicateUsersTouched(CI);
2419 // If we changed the class of the store, we want to ensure nothing finds the
2420 // old store expression. In particular, loads do not compare against stored
2421 // value, so they will find old store expressions (and associated class
2422 // mappings) if we leave them in the table.
2423 if (ClassChanged && isa<StoreInst>(I)) {
2424 auto *OldE = ValueToExpression.lookup(I);
2425 // It could just be that the old class died. We don't want to erase it if we
2426 // just moved classes.
2427 if (OldE && isa<StoreExpression>(OldE) && *E != *OldE) {
2428 // Erase this as an exact expression to ensure we don't erase expressions
2429 // equivalent to it.
2430 auto Iter = ExpressionToClass.find_as(ExactEqualsExpression(*OldE));
2431 if (Iter != ExpressionToClass.end())
2432 ExpressionToClass.erase(Iter);
2435 ValueToExpression[I] = E;
2438 // Process the fact that Edge (from, to) is reachable, including marking
2439 // any newly reachable blocks and instructions for processing.
2440 void NewGVN::updateReachableEdge(BasicBlock *From, BasicBlock *To) {
2441 // Check if the Edge was reachable before.
2442 if (ReachableEdges.insert({From, To}).second) {
2443 // If this block wasn't reachable before, all instructions are touched.
2444 if (ReachableBlocks.insert(To).second) {
2445 DEBUG(dbgs() << "Block " << getBlockName(To) << " marked reachable\n");
2446 const auto &InstRange = BlockInstRange.lookup(To);
2447 TouchedInstructions.set(InstRange.first, InstRange.second);
2449 DEBUG(dbgs() << "Block " << getBlockName(To)
2450 << " was reachable, but new edge {" << getBlockName(From)
2451 << "," << getBlockName(To) << "} to it found\n");
2453 // We've made an edge reachable to an existing block, which may
2454 // impact predicates. Otherwise, only mark the phi nodes as touched, as
2455 // they are the only thing that depend on new edges. Anything using their
2456 // values will get propagated to if necessary.
2457 if (MemoryAccess *MemPhi = getMemoryAccess(To))
2458 TouchedInstructions.set(InstrToDFSNum(MemPhi));
2460 // FIXME: We should just add a union op on a Bitvector and
2461 // SparseBitVector. We can do it word by word faster than we are doing it
2463 for (auto InstNum : RevisitOnReachabilityChange[To])
2464 TouchedInstructions.set(InstNum);
2469 // Given a predicate condition (from a switch, cmp, or whatever) and a block,
2470 // see if we know some constant value for it already.
2471 Value *NewGVN::findConditionEquivalence(Value *Cond) const {
2472 auto Result = lookupOperandLeader(Cond);
2473 return isa<Constant>(Result) ? Result : nullptr;
2476 // Process the outgoing edges of a block for reachability.
2477 void NewGVN::processOutgoingEdges(TerminatorInst *TI, BasicBlock *B) {
2478 // Evaluate reachability of terminator instruction.
2480 if ((BR = dyn_cast<BranchInst>(TI)) && BR->isConditional()) {
2481 Value *Cond = BR->getCondition();
2482 Value *CondEvaluated = findConditionEquivalence(Cond);
2483 if (!CondEvaluated) {
2484 if (auto *I = dyn_cast<Instruction>(Cond)) {
2485 const Expression *E = createExpression(I);
2486 if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2487 CondEvaluated = CE->getConstantValue();
2489 } else if (isa<ConstantInt>(Cond)) {
2490 CondEvaluated = Cond;
2494 BasicBlock *TrueSucc = BR->getSuccessor(0);
2495 BasicBlock *FalseSucc = BR->getSuccessor(1);
2496 if (CondEvaluated && (CI = dyn_cast<ConstantInt>(CondEvaluated))) {
2498 DEBUG(dbgs() << "Condition for Terminator " << *TI
2499 << " evaluated to true\n");
2500 updateReachableEdge(B, TrueSucc);
2501 } else if (CI->isZero()) {
2502 DEBUG(dbgs() << "Condition for Terminator " << *TI
2503 << " evaluated to false\n");
2504 updateReachableEdge(B, FalseSucc);
2507 updateReachableEdge(B, TrueSucc);
2508 updateReachableEdge(B, FalseSucc);
2510 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
2511 // For switches, propagate the case values into the case
2514 // Remember how many outgoing edges there are to every successor.
2515 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2517 Value *SwitchCond = SI->getCondition();
2518 Value *CondEvaluated = findConditionEquivalence(SwitchCond);
2519 // See if we were able to turn this switch statement into a constant.
2520 if (CondEvaluated && isa<ConstantInt>(CondEvaluated)) {
2521 auto *CondVal = cast<ConstantInt>(CondEvaluated);
2522 // We should be able to get case value for this.
2523 auto Case = *SI->findCaseValue(CondVal);
2524 if (Case.getCaseSuccessor() == SI->getDefaultDest()) {
2525 // We proved the value is outside of the range of the case.
2526 // We can't do anything other than mark the default dest as reachable,
2528 updateReachableEdge(B, SI->getDefaultDest());
2531 // Now get where it goes and mark it reachable.
2532 BasicBlock *TargetBlock = Case.getCaseSuccessor();
2533 updateReachableEdge(B, TargetBlock);
2535 for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) {
2536 BasicBlock *TargetBlock = SI->getSuccessor(i);
2537 ++SwitchEdges[TargetBlock];
2538 updateReachableEdge(B, TargetBlock);
2542 // Otherwise this is either unconditional, or a type we have no
2543 // idea about. Just mark successors as reachable.
2544 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
2545 BasicBlock *TargetBlock = TI->getSuccessor(i);
2546 updateReachableEdge(B, TargetBlock);
2549 // This also may be a memory defining terminator, in which case, set it
2550 // equivalent only to itself.
2552 auto *MA = getMemoryAccess(TI);
2553 if (MA && !isa<MemoryUse>(MA)) {
2554 auto *CC = ensureLeaderOfMemoryClass(MA);
2555 if (setMemoryClass(MA, CC))
2556 markMemoryUsersTouched(MA);
2561 // Remove the PHI of Ops PHI for I
2562 void NewGVN::removePhiOfOps(Instruction *I, PHINode *PHITemp) {
2563 InstrDFS.erase(PHITemp);
2564 // It's still a temp instruction. We keep it in the array so it gets erased.
2565 // However, it's no longer used by I, or in the block
2566 TempToBlock.erase(PHITemp);
2567 RealToTemp.erase(I);
2568 // We don't remove the users from the phi node uses. This wastes a little
2569 // time, but such is life. We could use two sets to track which were there
2570 // are the start of NewGVN, and which were added, but right nowt he cost of
2571 // tracking is more than the cost of checking for more phi of ops.
2574 // Add PHI Op in BB as a PHI of operations version of ExistingValue.
2575 void NewGVN::addPhiOfOps(PHINode *Op, BasicBlock *BB,
2576 Instruction *ExistingValue) {
2577 InstrDFS[Op] = InstrToDFSNum(ExistingValue);
2578 AllTempInstructions.insert(Op);
2579 TempToBlock[Op] = BB;
2580 RealToTemp[ExistingValue] = Op;
2581 // Add all users to phi node use, as they are now uses of the phi of ops phis
2582 // and may themselves be phi of ops.
2583 for (auto *U : ExistingValue->users())
2584 if (auto *UI = dyn_cast<Instruction>(U))
2585 PHINodeUses.insert(UI);
2588 static bool okayForPHIOfOps(const Instruction *I) {
2589 if (!EnablePhiOfOps)
2591 return isa<BinaryOperator>(I) || isa<SelectInst>(I) || isa<CmpInst>(I) ||
2595 bool NewGVN::OpIsSafeForPHIOfOpsHelper(
2596 Value *V, const BasicBlock *PHIBlock,
2597 SmallPtrSetImpl<const Value *> &Visited,
2598 SmallVectorImpl<Instruction *> &Worklist) {
2600 if (!isa<Instruction>(V))
2602 auto OISIt = OpSafeForPHIOfOps.find(V);
2603 if (OISIt != OpSafeForPHIOfOps.end())
2604 return OISIt->second;
2606 // Keep walking until we either dominate the phi block, or hit a phi, or run
2607 // out of things to check.
2608 if (DT->properlyDominates(getBlockForValue(V), PHIBlock)) {
2609 OpSafeForPHIOfOps.insert({V, true});
2612 // PHI in the same block.
2613 if (isa<PHINode>(V) && getBlockForValue(V) == PHIBlock) {
2614 OpSafeForPHIOfOps.insert({V, false});
2618 auto *OrigI = cast<Instruction>(V);
2619 for (auto *Op : OrigI->operand_values()) {
2620 if (!isa<Instruction>(Op))
2622 // Stop now if we find an unsafe operand.
2623 auto OISIt = OpSafeForPHIOfOps.find(OrigI);
2624 if (OISIt != OpSafeForPHIOfOps.end()) {
2625 if (!OISIt->second) {
2626 OpSafeForPHIOfOps.insert({V, false});
2631 if (!Visited.insert(Op).second)
2633 Worklist.push_back(cast<Instruction>(Op));
2638 // Return true if this operand will be safe to use for phi of ops.
2640 // The reason some operands are unsafe is that we are not trying to recursively
2641 // translate everything back through phi nodes. We actually expect some lookups
2642 // of expressions to fail. In particular, a lookup where the expression cannot
2643 // exist in the predecessor. This is true even if the expression, as shown, can
2644 // be determined to be constant.
2645 bool NewGVN::OpIsSafeForPHIOfOps(Value *V, const BasicBlock *PHIBlock,
2646 SmallPtrSetImpl<const Value *> &Visited) {
2647 SmallVector<Instruction *, 4> Worklist;
2648 if (!OpIsSafeForPHIOfOpsHelper(V, PHIBlock, Visited, Worklist))
2650 while (!Worklist.empty()) {
2651 auto *I = Worklist.pop_back_val();
2652 if (!OpIsSafeForPHIOfOpsHelper(I, PHIBlock, Visited, Worklist))
2655 OpSafeForPHIOfOps.insert({V, true});
2659 // Try to find a leader for instruction TransInst, which is a phi translated
2660 // version of something in our original program. Visited is used to ensure we
2661 // don't infinite loop during translations of cycles. OrigInst is the
2662 // instruction in the original program, and PredBB is the predecessor we
2663 // translated it through.
2664 Value *NewGVN::findLeaderForInst(Instruction *TransInst,
2665 SmallPtrSetImpl<Value *> &Visited,
2666 MemoryAccess *MemAccess, Instruction *OrigInst,
2667 BasicBlock *PredBB) {
2668 unsigned IDFSNum = InstrToDFSNum(OrigInst);
2669 // Make sure it's marked as a temporary instruction.
2670 AllTempInstructions.insert(TransInst);
2671 // and make sure anything that tries to add it's DFS number is
2672 // redirected to the instruction we are making a phi of ops
2674 TempToBlock.insert({TransInst, PredBB});
2675 InstrDFS.insert({TransInst, IDFSNum});
2677 const Expression *E = performSymbolicEvaluation(TransInst, Visited);
2678 InstrDFS.erase(TransInst);
2679 AllTempInstructions.erase(TransInst);
2680 TempToBlock.erase(TransInst);
2682 TempToMemory.erase(TransInst);
2685 auto *FoundVal = findPHIOfOpsLeader(E, OrigInst, PredBB);
2687 ExpressionToPhiOfOps[E].insert(OrigInst);
2688 DEBUG(dbgs() << "Cannot find phi of ops operand for " << *TransInst
2689 << " in block " << getBlockName(PredBB) << "\n");
2692 if (auto *SI = dyn_cast<StoreInst>(FoundVal))
2693 FoundVal = SI->getValueOperand();
2697 // When we see an instruction that is an op of phis, generate the equivalent phi
2700 NewGVN::makePossiblePHIOfOps(Instruction *I,
2701 SmallPtrSetImpl<Value *> &Visited) {
2702 if (!okayForPHIOfOps(I))
2705 if (!Visited.insert(I).second)
2707 // For now, we require the instruction be cycle free because we don't
2708 // *always* create a phi of ops for instructions that could be done as phi
2709 // of ops, we only do it if we think it is useful. If we did do it all the
2710 // time, we could remove the cycle free check.
2711 if (!isCycleFree(I))
2714 SmallPtrSet<const Value *, 8> ProcessedPHIs;
2715 // TODO: We don't do phi translation on memory accesses because it's
2716 // complicated. For a load, we'd need to be able to simulate a new memoryuse,
2717 // which we don't have a good way of doing ATM.
2718 auto *MemAccess = getMemoryAccess(I);
2719 // If the memory operation is defined by a memory operation this block that
2720 // isn't a MemoryPhi, transforming the pointer backwards through a scalar phi
2721 // can't help, as it would still be killed by that memory operation.
2722 if (MemAccess && !isa<MemoryPhi>(MemAccess->getDefiningAccess()) &&
2723 MemAccess->getDefiningAccess()->getBlock() == I->getParent())
2726 SmallPtrSet<const Value *, 10> VisitedOps;
2727 // Convert op of phis to phi of ops
2728 for (auto *Op : I->operand_values()) {
2729 if (!isa<PHINode>(Op)) {
2730 auto *ValuePHI = RealToTemp.lookup(Op);
2733 DEBUG(dbgs() << "Found possible dependent phi of ops\n");
2736 auto *OpPHI = cast<PHINode>(Op);
2737 // No point in doing this for one-operand phis.
2738 if (OpPHI->getNumOperands() == 1)
2740 if (!DebugCounter::shouldExecute(PHIOfOpsCounter))
2742 SmallVector<ValPair, 4> Ops;
2743 SmallPtrSet<Value *, 4> Deps;
2744 auto *PHIBlock = getBlockForValue(OpPHI);
2745 RevisitOnReachabilityChange[PHIBlock].reset(InstrToDFSNum(I));
2746 for (unsigned PredNum = 0; PredNum < OpPHI->getNumOperands(); ++PredNum) {
2747 auto *PredBB = OpPHI->getIncomingBlock(PredNum);
2748 Value *FoundVal = nullptr;
2749 // We could just skip unreachable edges entirely but it's tricky to do
2750 // with rewriting existing phi nodes.
2751 if (ReachableEdges.count({PredBB, PHIBlock})) {
2752 // Clone the instruction, create an expression from it that is
2753 // translated back into the predecessor, and see if we have a leader.
2754 Instruction *ValueOp = I->clone();
2756 TempToMemory.insert({ValueOp, MemAccess});
2757 bool SafeForPHIOfOps = true;
2759 for (auto &Op : ValueOp->operands()) {
2760 auto *OrigOp = &*Op;
2761 // When these operand changes, it could change whether there is a
2762 // leader for us or not, so we have to add additional users.
2763 if (isa<PHINode>(Op)) {
2764 Op = Op->DoPHITranslation(PHIBlock, PredBB);
2765 if (Op != OrigOp && Op != I)
2767 } else if (auto *ValuePHI = RealToTemp.lookup(Op)) {
2768 if (getBlockForValue(ValuePHI) == PHIBlock)
2769 Op = ValuePHI->getIncomingValueForBlock(PredBB);
2771 // If we phi-translated the op, it must be safe.
2774 (Op != OrigOp || OpIsSafeForPHIOfOps(Op, PHIBlock, VisitedOps));
2776 // FIXME: For those things that are not safe we could generate
2777 // expressions all the way down, and see if this comes out to a
2778 // constant. For anything where that is true, and unsafe, we should
2779 // have made a phi-of-ops (or value numbered it equivalent to something)
2780 // for the pieces already.
2781 FoundVal = !SafeForPHIOfOps ? nullptr
2782 : findLeaderForInst(ValueOp, Visited,
2783 MemAccess, I, PredBB);
2784 ValueOp->deleteValue();
2788 DEBUG(dbgs() << "Skipping phi of ops operand for incoming block "
2789 << getBlockName(PredBB)
2790 << " because the block is unreachable\n");
2791 FoundVal = UndefValue::get(I->getType());
2792 RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2795 Ops.push_back({FoundVal, PredBB});
2796 DEBUG(dbgs() << "Found phi of ops operand " << *FoundVal << " in "
2797 << getBlockName(PredBB) << "\n");
2799 for (auto Dep : Deps)
2800 addAdditionalUsers(Dep, I);
2802 auto *E = performSymbolicPHIEvaluation(Ops, I, PHIBlock);
2803 if (isa<ConstantExpression>(E) || isa<VariableExpression>(E)) {
2805 << "Not creating real PHI of ops because it simplified to existing "
2806 "value or constant\n");
2809 auto *ValuePHI = RealToTemp.lookup(I);
2810 bool NewPHI = false;
2813 PHINode::Create(I->getType(), OpPHI->getNumOperands(), "phiofops");
2814 addPhiOfOps(ValuePHI, PHIBlock, I);
2816 NumGVNPHIOfOpsCreated++;
2819 for (auto PHIOp : Ops)
2820 ValuePHI->addIncoming(PHIOp.first, PHIOp.second);
2822 TempToBlock[ValuePHI] = PHIBlock;
2824 for (auto PHIOp : Ops) {
2825 ValuePHI->setIncomingValue(i, PHIOp.first);
2826 ValuePHI->setIncomingBlock(i, PHIOp.second);
2830 RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2831 DEBUG(dbgs() << "Created phi of ops " << *ValuePHI << " for " << *I
2839 // The algorithm initially places the values of the routine in the TOP
2840 // congruence class. The leader of TOP is the undetermined value `undef`.
2841 // When the algorithm has finished, values still in TOP are unreachable.
2842 void NewGVN::initializeCongruenceClasses(Function &F) {
2843 NextCongruenceNum = 0;
2845 // Note that even though we use the live on entry def as a representative
2846 // MemoryAccess, it is *not* the same as the actual live on entry def. We
2847 // have no real equivalemnt to undef for MemoryAccesses, and so we really
2848 // should be checking whether the MemoryAccess is top if we want to know if it
2849 // is equivalent to everything. Otherwise, what this really signifies is that
2850 // the access "it reaches all the way back to the beginning of the function"
2852 // Initialize all other instructions to be in TOP class.
2853 TOPClass = createCongruenceClass(nullptr, nullptr);
2854 TOPClass->setMemoryLeader(MSSA->getLiveOnEntryDef());
2855 // The live on entry def gets put into it's own class
2856 MemoryAccessToClass[MSSA->getLiveOnEntryDef()] =
2857 createMemoryClass(MSSA->getLiveOnEntryDef());
2859 for (auto DTN : nodes(DT)) {
2860 BasicBlock *BB = DTN->getBlock();
2861 // All MemoryAccesses are equivalent to live on entry to start. They must
2862 // be initialized to something so that initial changes are noticed. For
2863 // the maximal answer, we initialize them all to be the same as
2865 auto *MemoryBlockDefs = MSSA->getBlockDefs(BB);
2866 if (MemoryBlockDefs)
2867 for (const auto &Def : *MemoryBlockDefs) {
2868 MemoryAccessToClass[&Def] = TOPClass;
2869 auto *MD = dyn_cast<MemoryDef>(&Def);
2870 // Insert the memory phis into the member list.
2872 const MemoryPhi *MP = cast<MemoryPhi>(&Def);
2873 TOPClass->memory_insert(MP);
2874 MemoryPhiState.insert({MP, MPS_TOP});
2877 if (MD && isa<StoreInst>(MD->getMemoryInst()))
2878 TOPClass->incStoreCount();
2881 // FIXME: This is trying to discover which instructions are uses of phi
2882 // nodes. We should move this into one of the myriad of places that walk
2883 // all the operands already.
2884 for (auto &I : *BB) {
2885 if (isa<PHINode>(&I))
2886 for (auto *U : I.users())
2887 if (auto *UInst = dyn_cast<Instruction>(U))
2888 if (InstrToDFSNum(UInst) != 0 && okayForPHIOfOps(UInst))
2889 PHINodeUses.insert(UInst);
2890 // Don't insert void terminators into the class. We don't value number
2891 // them, and they just end up sitting in TOP.
2892 if (isa<TerminatorInst>(I) && I.getType()->isVoidTy())
2894 TOPClass->insert(&I);
2895 ValueToClass[&I] = TOPClass;
2899 // Initialize arguments to be in their own unique congruence classes
2900 for (auto &FA : F.args())
2901 createSingletonCongruenceClass(&FA);
2904 void NewGVN::cleanupTables() {
2905 for (unsigned i = 0, e = CongruenceClasses.size(); i != e; ++i) {
2906 DEBUG(dbgs() << "Congruence class " << CongruenceClasses[i]->getID()
2907 << " has " << CongruenceClasses[i]->size() << " members\n");
2908 // Make sure we delete the congruence class (probably worth switching to
2909 // a unique_ptr at some point.
2910 delete CongruenceClasses[i];
2911 CongruenceClasses[i] = nullptr;
2914 // Destroy the value expressions
2915 SmallVector<Instruction *, 8> TempInst(AllTempInstructions.begin(),
2916 AllTempInstructions.end());
2917 AllTempInstructions.clear();
2919 // We have to drop all references for everything first, so there are no uses
2920 // left as we delete them.
2921 for (auto *I : TempInst) {
2922 I->dropAllReferences();
2925 while (!TempInst.empty()) {
2926 auto *I = TempInst.back();
2927 TempInst.pop_back();
2931 ValueToClass.clear();
2932 ArgRecycler.clear(ExpressionAllocator);
2933 ExpressionAllocator.Reset();
2934 CongruenceClasses.clear();
2935 ExpressionToClass.clear();
2936 ValueToExpression.clear();
2938 AdditionalUsers.clear();
2939 ExpressionToPhiOfOps.clear();
2940 TempToBlock.clear();
2941 TempToMemory.clear();
2942 PHINodeUses.clear();
2943 OpSafeForPHIOfOps.clear();
2944 ReachableBlocks.clear();
2945 ReachableEdges.clear();
2947 ProcessedCount.clear();
2950 InstructionsToErase.clear();
2952 BlockInstRange.clear();
2953 TouchedInstructions.clear();
2954 MemoryAccessToClass.clear();
2955 PredicateToUsers.clear();
2956 MemoryToUsers.clear();
2957 RevisitOnReachabilityChange.clear();
2960 // Assign local DFS number mapping to instructions, and leave space for Value
2962 std::pair<unsigned, unsigned> NewGVN::assignDFSNumbers(BasicBlock *B,
2964 unsigned End = Start;
2965 if (MemoryAccess *MemPhi = getMemoryAccess(B)) {
2966 InstrDFS[MemPhi] = End++;
2967 DFSToInstr.emplace_back(MemPhi);
2970 // Then the real block goes next.
2971 for (auto &I : *B) {
2972 // There's no need to call isInstructionTriviallyDead more than once on
2973 // an instruction. Therefore, once we know that an instruction is dead
2974 // we change its DFS number so that it doesn't get value numbered.
2975 if (isInstructionTriviallyDead(&I, TLI)) {
2977 DEBUG(dbgs() << "Skipping trivially dead instruction " << I << "\n");
2978 markInstructionForDeletion(&I);
2981 if (isa<PHINode>(&I))
2982 RevisitOnReachabilityChange[B].set(End);
2983 InstrDFS[&I] = End++;
2984 DFSToInstr.emplace_back(&I);
2987 // All of the range functions taken half-open ranges (open on the end side).
2988 // So we do not subtract one from count, because at this point it is one
2989 // greater than the last instruction.
2990 return std::make_pair(Start, End);
2993 void NewGVN::updateProcessedCount(const Value *V) {
2995 if (ProcessedCount.count(V) == 0) {
2996 ProcessedCount.insert({V, 1});
2998 ++ProcessedCount[V];
2999 assert(ProcessedCount[V] < 100 &&
3000 "Seem to have processed the same Value a lot");
3005 // Evaluate MemoryPhi nodes symbolically, just like PHI nodes
3006 void NewGVN::valueNumberMemoryPhi(MemoryPhi *MP) {
3007 // If all the arguments are the same, the MemoryPhi has the same value as the
3008 // argument. Filter out unreachable blocks and self phis from our operands.
3009 // TODO: We could do cycle-checking on the memory phis to allow valueizing for
3010 // self-phi checking.
3011 const BasicBlock *PHIBlock = MP->getBlock();
3012 auto Filtered = make_filter_range(MP->operands(), [&](const Use &U) {
3013 return cast<MemoryAccess>(U) != MP &&
3014 !isMemoryAccessTOP(cast<MemoryAccess>(U)) &&
3015 ReachableEdges.count({MP->getIncomingBlock(U), PHIBlock});
3017 // If all that is left is nothing, our memoryphi is undef. We keep it as
3018 // InitialClass. Note: The only case this should happen is if we have at
3019 // least one self-argument.
3020 if (Filtered.begin() == Filtered.end()) {
3021 if (setMemoryClass(MP, TOPClass))
3022 markMemoryUsersTouched(MP);
3026 // Transform the remaining operands into operand leaders.
3027 // FIXME: mapped_iterator should have a range version.
3028 auto LookupFunc = [&](const Use &U) {
3029 return lookupMemoryLeader(cast<MemoryAccess>(U));
3031 auto MappedBegin = map_iterator(Filtered.begin(), LookupFunc);
3032 auto MappedEnd = map_iterator(Filtered.end(), LookupFunc);
3034 // and now check if all the elements are equal.
3035 // Sadly, we can't use std::equals since these are random access iterators.
3036 const auto *AllSameValue = *MappedBegin;
3038 bool AllEqual = std::all_of(
3039 MappedBegin, MappedEnd,
3040 [&AllSameValue](const MemoryAccess *V) { return V == AllSameValue; });
3043 DEBUG(dbgs() << "Memory Phi value numbered to " << *AllSameValue << "\n");
3045 DEBUG(dbgs() << "Memory Phi value numbered to itself\n");
3046 // If it's equal to something, it's in that class. Otherwise, it has to be in
3047 // a class where it is the leader (other things may be equivalent to it, but
3048 // it needs to start off in its own class, which means it must have been the
3049 // leader, and it can't have stopped being the leader because it was never
3051 CongruenceClass *CC =
3052 AllEqual ? getMemoryClass(AllSameValue) : ensureLeaderOfMemoryClass(MP);
3053 auto OldState = MemoryPhiState.lookup(MP);
3054 assert(OldState != MPS_Invalid && "Invalid memory phi state");
3055 auto NewState = AllEqual ? MPS_Equivalent : MPS_Unique;
3056 MemoryPhiState[MP] = NewState;
3057 if (setMemoryClass(MP, CC) || OldState != NewState)
3058 markMemoryUsersTouched(MP);
3061 // Value number a single instruction, symbolically evaluating, performing
3062 // congruence finding, and updating mappings.
3063 void NewGVN::valueNumberInstruction(Instruction *I) {
3064 DEBUG(dbgs() << "Processing instruction " << *I << "\n");
3065 if (!I->isTerminator()) {
3066 const Expression *Symbolized = nullptr;
3067 SmallPtrSet<Value *, 2> Visited;
3068 if (DebugCounter::shouldExecute(VNCounter)) {
3069 Symbolized = performSymbolicEvaluation(I, Visited);
3070 // Make a phi of ops if necessary
3071 if (Symbolized && !isa<ConstantExpression>(Symbolized) &&
3072 !isa<VariableExpression>(Symbolized) && PHINodeUses.count(I)) {
3073 auto *PHIE = makePossiblePHIOfOps(I, Visited);
3074 // If we created a phi of ops, use it.
3075 // If we couldn't create one, make sure we don't leave one lying around
3078 } else if (auto *Op = RealToTemp.lookup(I)) {
3079 removePhiOfOps(I, Op);
3083 // Mark the instruction as unused so we don't value number it again.
3086 // If we couldn't come up with a symbolic expression, use the unknown
3088 if (Symbolized == nullptr)
3089 Symbolized = createUnknownExpression(I);
3090 performCongruenceFinding(I, Symbolized);
3092 // Handle terminators that return values. All of them produce values we
3093 // don't currently understand. We don't place non-value producing
3094 // terminators in a class.
3095 if (!I->getType()->isVoidTy()) {
3096 auto *Symbolized = createUnknownExpression(I);
3097 performCongruenceFinding(I, Symbolized);
3099 processOutgoingEdges(dyn_cast<TerminatorInst>(I), I->getParent());
3103 // Check if there is a path, using single or equal argument phi nodes, from
3105 bool NewGVN::singleReachablePHIPath(
3106 SmallPtrSet<const MemoryAccess *, 8> &Visited, const MemoryAccess *First,
3107 const MemoryAccess *Second) const {
3108 if (First == Second)
3110 if (MSSA->isLiveOnEntryDef(First))
3113 // This is not perfect, but as we're just verifying here, we can live with
3114 // the loss of precision. The real solution would be that of doing strongly
3115 // connected component finding in this routine, and it's probably not worth
3116 // the complexity for the time being. So, we just keep a set of visited
3117 // MemoryAccess and return true when we hit a cycle.
3118 if (Visited.count(First))
3120 Visited.insert(First);
3122 const auto *EndDef = First;
3123 for (auto *ChainDef : optimized_def_chain(First)) {
3124 if (ChainDef == Second)
3126 if (MSSA->isLiveOnEntryDef(ChainDef))
3130 auto *MP = cast<MemoryPhi>(EndDef);
3131 auto ReachableOperandPred = [&](const Use &U) {
3132 return ReachableEdges.count({MP->getIncomingBlock(U), MP->getBlock()});
3134 auto FilteredPhiArgs =
3135 make_filter_range(MP->operands(), ReachableOperandPred);
3136 SmallVector<const Value *, 32> OperandList;
3137 std::copy(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
3138 std::back_inserter(OperandList));
3139 bool Okay = OperandList.size() == 1;
3142 std::equal(OperandList.begin(), OperandList.end(), OperandList.begin());
3144 return singleReachablePHIPath(Visited, cast<MemoryAccess>(OperandList[0]),
3149 // Verify the that the memory equivalence table makes sense relative to the
3150 // congruence classes. Note that this checking is not perfect, and is currently
3151 // subject to very rare false negatives. It is only useful for
3152 // testing/debugging.
3153 void NewGVN::verifyMemoryCongruency() const {
3155 // Verify that the memory table equivalence and memory member set match
3156 for (const auto *CC : CongruenceClasses) {
3157 if (CC == TOPClass || CC->isDead())
3159 if (CC->getStoreCount() != 0) {
3160 assert((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) &&
3161 "Any class with a store as a leader should have a "
3162 "representative stored value");
3163 assert(CC->getMemoryLeader() &&
3164 "Any congruence class with a store should have a "
3165 "representative access");
3168 if (CC->getMemoryLeader())
3169 assert(MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
3170 "Representative MemoryAccess does not appear to be reverse "
3172 for (auto M : CC->memory())
3173 assert(MemoryAccessToClass.lookup(M) == CC &&
3174 "Memory member does not appear to be reverse mapped properly");
3177 // Anything equivalent in the MemoryAccess table should be in the same
3178 // congruence class.
3180 // Filter out the unreachable and trivially dead entries, because they may
3181 // never have been updated if the instructions were not processed.
3182 auto ReachableAccessPred =
3183 [&](const std::pair<const MemoryAccess *, CongruenceClass *> Pair) {
3184 bool Result = ReachableBlocks.count(Pair.first->getBlock());
3185 if (!Result || MSSA->isLiveOnEntryDef(Pair.first) ||
3186 MemoryToDFSNum(Pair.first) == 0)
3188 if (auto *MemDef = dyn_cast<MemoryDef>(Pair.first))
3189 return !isInstructionTriviallyDead(MemDef->getMemoryInst());
3191 // We could have phi nodes which operands are all trivially dead,
3192 // so we don't process them.
3193 if (auto *MemPHI = dyn_cast<MemoryPhi>(Pair.first)) {
3194 for (auto &U : MemPHI->incoming_values()) {
3195 if (auto *I = dyn_cast<Instruction>(&*U)) {
3196 if (!isInstructionTriviallyDead(I))
3206 auto Filtered = make_filter_range(MemoryAccessToClass, ReachableAccessPred);
3207 for (auto KV : Filtered) {
3208 if (auto *FirstMUD = dyn_cast<MemoryUseOrDef>(KV.first)) {
3209 auto *SecondMUD = dyn_cast<MemoryUseOrDef>(KV.second->getMemoryLeader());
3210 if (FirstMUD && SecondMUD) {
3211 SmallPtrSet<const MemoryAccess *, 8> VisitedMAS;
3212 assert((singleReachablePHIPath(VisitedMAS, FirstMUD, SecondMUD) ||
3213 ValueToClass.lookup(FirstMUD->getMemoryInst()) ==
3214 ValueToClass.lookup(SecondMUD->getMemoryInst())) &&
3215 "The instructions for these memory operations should have "
3216 "been in the same congruence class or reachable through"
3217 "a single argument phi");
3219 } else if (auto *FirstMP = dyn_cast<MemoryPhi>(KV.first)) {
3220 // We can only sanely verify that MemoryDefs in the operand list all have
3222 auto ReachableOperandPred = [&](const Use &U) {
3223 return ReachableEdges.count(
3224 {FirstMP->getIncomingBlock(U), FirstMP->getBlock()}) &&
3228 // All arguments should in the same class, ignoring unreachable arguments
3229 auto FilteredPhiArgs =
3230 make_filter_range(FirstMP->operands(), ReachableOperandPred);
3231 SmallVector<const CongruenceClass *, 16> PhiOpClasses;
3232 std::transform(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
3233 std::back_inserter(PhiOpClasses), [&](const Use &U) {
3234 const MemoryDef *MD = cast<MemoryDef>(U);
3235 return ValueToClass.lookup(MD->getMemoryInst());
3237 assert(std::equal(PhiOpClasses.begin(), PhiOpClasses.end(),
3238 PhiOpClasses.begin()) &&
3239 "All MemoryPhi arguments should be in the same class");
3245 // Verify that the sparse propagation we did actually found the maximal fixpoint
3246 // We do this by storing the value to class mapping, touching all instructions,
3247 // and redoing the iteration to see if anything changed.
3248 void NewGVN::verifyIterationSettled(Function &F) {
3250 DEBUG(dbgs() << "Beginning iteration verification\n");
3251 if (DebugCounter::isCounterSet(VNCounter))
3252 DebugCounter::setCounterValue(VNCounter, StartingVNCounter);
3254 // Note that we have to store the actual classes, as we may change existing
3255 // classes during iteration. This is because our memory iteration propagation
3256 // is not perfect, and so may waste a little work. But it should generate
3257 // exactly the same congruence classes we have now, with different IDs.
3258 std::map<const Value *, CongruenceClass> BeforeIteration;
3260 for (auto &KV : ValueToClass) {
3261 if (auto *I = dyn_cast<Instruction>(KV.first))
3262 // Skip unused/dead instructions.
3263 if (InstrToDFSNum(I) == 0)
3265 BeforeIteration.insert({KV.first, *KV.second});
3268 TouchedInstructions.set();
3269 TouchedInstructions.reset(0);
3270 iterateTouchedInstructions();
3271 DenseSet<std::pair<const CongruenceClass *, const CongruenceClass *>>
3273 for (const auto &KV : ValueToClass) {
3274 if (auto *I = dyn_cast<Instruction>(KV.first))
3275 // Skip unused/dead instructions.
3276 if (InstrToDFSNum(I) == 0)
3278 // We could sink these uses, but i think this adds a bit of clarity here as
3279 // to what we are comparing.
3280 auto *BeforeCC = &BeforeIteration.find(KV.first)->second;
3281 auto *AfterCC = KV.second;
3282 // Note that the classes can't change at this point, so we memoize the set
3284 if (!EqualClasses.count({BeforeCC, AfterCC})) {
3285 assert(BeforeCC->isEquivalentTo(AfterCC) &&
3286 "Value number changed after main loop completed!");
3287 EqualClasses.insert({BeforeCC, AfterCC});
3293 // Verify that for each store expression in the expression to class mapping,
3294 // only the latest appears, and multiple ones do not appear.
3295 // Because loads do not use the stored value when doing equality with stores,
3296 // if we don't erase the old store expressions from the table, a load can find
3297 // a no-longer valid StoreExpression.
3298 void NewGVN::verifyStoreExpressions() const {
3300 // This is the only use of this, and it's not worth defining a complicated
3301 // densemapinfo hash/equality function for it.
3303 std::pair<const Value *,
3304 std::tuple<const Value *, const CongruenceClass *, Value *>>>
3306 for (const auto &KV : ExpressionToClass) {
3307 if (auto *SE = dyn_cast<StoreExpression>(KV.first)) {
3308 // Make sure a version that will conflict with loads is not already there
3309 auto Res = StoreExpressionSet.insert(
3310 {SE->getOperand(0), std::make_tuple(SE->getMemoryLeader(), KV.second,
3311 SE->getStoredValue())});
3312 bool Okay = Res.second;
3313 // It's okay to have the same expression already in there if it is
3314 // identical in nature.
3315 // This can happen when the leader of the stored value changes over time.
3317 Okay = (std::get<1>(Res.first->second) == KV.second) &&
3318 (lookupOperandLeader(std::get<2>(Res.first->second)) ==
3319 lookupOperandLeader(SE->getStoredValue()));
3320 assert(Okay && "Stored expression conflict exists in expression table");
3321 auto *ValueExpr = ValueToExpression.lookup(SE->getStoreInst());
3322 assert(ValueExpr && ValueExpr->equals(*SE) &&
3323 "StoreExpression in ExpressionToClass is not latest "
3324 "StoreExpression for value");
3330 // This is the main value numbering loop, it iterates over the initial touched
3331 // instruction set, propagating value numbers, marking things touched, etc,
3332 // until the set of touched instructions is completely empty.
3333 void NewGVN::iterateTouchedInstructions() {
3334 unsigned int Iterations = 0;
3335 // Figure out where touchedinstructions starts
3336 int FirstInstr = TouchedInstructions.find_first();
3337 // Nothing set, nothing to iterate, just return.
3338 if (FirstInstr == -1)
3340 const BasicBlock *LastBlock = getBlockForValue(InstrFromDFSNum(FirstInstr));
3341 while (TouchedInstructions.any()) {
3343 // Walk through all the instructions in all the blocks in RPO.
3344 // TODO: As we hit a new block, we should push and pop equalities into a
3345 // table lookupOperandLeader can use, to catch things PredicateInfo
3346 // might miss, like edge-only equivalences.
3347 for (unsigned InstrNum : TouchedInstructions.set_bits()) {
3349 // This instruction was found to be dead. We don't bother looking
3351 if (InstrNum == 0) {
3352 TouchedInstructions.reset(InstrNum);
3356 Value *V = InstrFromDFSNum(InstrNum);
3357 const BasicBlock *CurrBlock = getBlockForValue(V);
3359 // If we hit a new block, do reachability processing.
3360 if (CurrBlock != LastBlock) {
3361 LastBlock = CurrBlock;
3362 bool BlockReachable = ReachableBlocks.count(CurrBlock);
3363 const auto &CurrInstRange = BlockInstRange.lookup(CurrBlock);
3365 // If it's not reachable, erase any touched instructions and move on.
3366 if (!BlockReachable) {
3367 TouchedInstructions.reset(CurrInstRange.first, CurrInstRange.second);
3368 DEBUG(dbgs() << "Skipping instructions in block "
3369 << getBlockName(CurrBlock)
3370 << " because it is unreachable\n");
3373 updateProcessedCount(CurrBlock);
3375 // Reset after processing (because we may mark ourselves as touched when
3376 // we propagate equalities).
3377 TouchedInstructions.reset(InstrNum);
3379 if (auto *MP = dyn_cast<MemoryPhi>(V)) {
3380 DEBUG(dbgs() << "Processing MemoryPhi " << *MP << "\n");
3381 valueNumberMemoryPhi(MP);
3382 } else if (auto *I = dyn_cast<Instruction>(V)) {
3383 valueNumberInstruction(I);
3385 llvm_unreachable("Should have been a MemoryPhi or Instruction");
3387 updateProcessedCount(V);
3390 NumGVNMaxIterations = std::max(NumGVNMaxIterations.getValue(), Iterations);
3393 // This is the main transformation entry point.
3394 bool NewGVN::runGVN() {
3395 if (DebugCounter::isCounterSet(VNCounter))
3396 StartingVNCounter = DebugCounter::getCounterValue(VNCounter);
3397 bool Changed = false;
3398 NumFuncArgs = F.arg_size();
3399 MSSAWalker = MSSA->getWalker();
3400 SingletonDeadExpression = new (ExpressionAllocator) DeadExpression();
3402 // Count number of instructions for sizing of hash tables, and come
3403 // up with a global dfs numbering for instructions.
3404 unsigned ICount = 1;
3405 // Add an empty instruction to account for the fact that we start at 1
3406 DFSToInstr.emplace_back(nullptr);
3407 // Note: We want ideal RPO traversal of the blocks, which is not quite the
3408 // same as dominator tree order, particularly with regard whether backedges
3409 // get visited first or second, given a block with multiple successors.
3410 // If we visit in the wrong order, we will end up performing N times as many
3412 // The dominator tree does guarantee that, for a given dom tree node, it's
3413 // parent must occur before it in the RPO ordering. Thus, we only need to sort
3415 ReversePostOrderTraversal<Function *> RPOT(&F);
3416 unsigned Counter = 0;
3417 for (auto &B : RPOT) {
3418 auto *Node = DT->getNode(B);
3419 assert(Node && "RPO and Dominator tree should have same reachability");
3420 RPOOrdering[Node] = ++Counter;
3422 // Sort dominator tree children arrays into RPO.
3423 for (auto &B : RPOT) {
3424 auto *Node = DT->getNode(B);
3425 if (Node->getChildren().size() > 1)
3426 std::sort(Node->begin(), Node->end(),
3427 [&](const DomTreeNode *A, const DomTreeNode *B) {
3428 return RPOOrdering[A] < RPOOrdering[B];
3432 // Now a standard depth first ordering of the domtree is equivalent to RPO.
3433 for (auto DTN : depth_first(DT->getRootNode())) {
3434 BasicBlock *B = DTN->getBlock();
3435 const auto &BlockRange = assignDFSNumbers(B, ICount);
3436 BlockInstRange.insert({B, BlockRange});
3437 ICount += BlockRange.second - BlockRange.first;
3439 initializeCongruenceClasses(F);
3441 TouchedInstructions.resize(ICount);
3442 // Ensure we don't end up resizing the expressionToClass map, as
3443 // that can be quite expensive. At most, we have one expression per
3445 ExpressionToClass.reserve(ICount);
3447 // Initialize the touched instructions to include the entry block.
3448 const auto &InstRange = BlockInstRange.lookup(&F.getEntryBlock());
3449 TouchedInstructions.set(InstRange.first, InstRange.second);
3450 DEBUG(dbgs() << "Block " << getBlockName(&F.getEntryBlock())
3451 << " marked reachable\n");
3452 ReachableBlocks.insert(&F.getEntryBlock());
3454 iterateTouchedInstructions();
3455 verifyMemoryCongruency();
3456 verifyIterationSettled(F);
3457 verifyStoreExpressions();
3459 Changed |= eliminateInstructions(F);
3461 // Delete all instructions marked for deletion.
3462 for (Instruction *ToErase : InstructionsToErase) {
3463 if (!ToErase->use_empty())
3464 ToErase->replaceAllUsesWith(UndefValue::get(ToErase->getType()));
3466 if (ToErase->getParent())
3467 ToErase->eraseFromParent();
3470 // Delete all unreachable blocks.
3471 auto UnreachableBlockPred = [&](const BasicBlock &BB) {
3472 return !ReachableBlocks.count(&BB);
3475 for (auto &BB : make_filter_range(F, UnreachableBlockPred)) {
3476 DEBUG(dbgs() << "We believe block " << getBlockName(&BB)
3477 << " is unreachable\n");
3478 deleteInstructionsInBlock(&BB);
3486 struct NewGVN::ValueDFS {
3491 // Only one of Def and U will be set.
3492 // The bool in the Def tells us whether the Def is the stored value of a
3494 PointerIntPair<Value *, 1, bool> Def;
3497 bool operator<(const ValueDFS &Other) const {
3498 // It's not enough that any given field be less than - we have sets
3499 // of fields that need to be evaluated together to give a proper ordering.
3500 // For example, if you have;
3505 // We want the second to be less than the first, but if we just go field
3506 // by field, we will get to Val 0 < Val 50 and say the first is less than
3507 // the second. We only want it to be less than if the DFS orders are equal.
3509 // Each LLVM instruction only produces one value, and thus the lowest-level
3510 // differentiator that really matters for the stack (and what we use as as a
3511 // replacement) is the local dfs number.
3512 // Everything else in the structure is instruction level, and only affects
3513 // the order in which we will replace operands of a given instruction.
3515 // For a given instruction (IE things with equal dfsin, dfsout, localnum),
3516 // the order of replacement of uses does not matter.
3520 // When you hit b, you will have two valuedfs with the same dfsin, out, and
3522 // The .val will be the same as well.
3523 // The .u's will be different.
3524 // You will replace both, and it does not matter what order you replace them
3525 // in (IE whether you replace operand 2, then operand 1, or operand 1, then
3527 // Similarly for the case of same dfsin, dfsout, localnum, but different
3532 // in c, we will a valuedfs for a, and one for b,with everything the same
3534 // It does not matter what order we replace these operands in.
3535 // You will always end up with the same IR, and this is guaranteed.
3536 return std::tie(DFSIn, DFSOut, LocalNum, Def, U) <
3537 std::tie(Other.DFSIn, Other.DFSOut, Other.LocalNum, Other.Def,
3542 // This function converts the set of members for a congruence class from values,
3543 // to sets of defs and uses with associated DFS info. The total number of
3544 // reachable uses for each value is stored in UseCount, and instructions that
3546 // dead (have no non-dead uses) are stored in ProbablyDead.
3547 void NewGVN::convertClassToDFSOrdered(
3548 const CongruenceClass &Dense, SmallVectorImpl<ValueDFS> &DFSOrderedSet,
3549 DenseMap<const Value *, unsigned int> &UseCounts,
3550 SmallPtrSetImpl<Instruction *> &ProbablyDead) const {
3551 for (auto D : Dense) {
3552 // First add the value.
3553 BasicBlock *BB = getBlockForValue(D);
3554 // Constants are handled prior to ever calling this function, so
3555 // we should only be left with instructions as members.
3556 assert(BB && "Should have figured out a basic block for value");
3558 DomTreeNode *DomNode = DT->getNode(BB);
3559 VDDef.DFSIn = DomNode->getDFSNumIn();
3560 VDDef.DFSOut = DomNode->getDFSNumOut();
3561 // If it's a store, use the leader of the value operand, if it's always
3562 // available, or the value operand. TODO: We could do dominance checks to
3563 // find a dominating leader, but not worth it ATM.
3564 if (auto *SI = dyn_cast<StoreInst>(D)) {
3565 auto Leader = lookupOperandLeader(SI->getValueOperand());
3566 if (alwaysAvailable(Leader)) {
3567 VDDef.Def.setPointer(Leader);
3569 VDDef.Def.setPointer(SI->getValueOperand());
3570 VDDef.Def.setInt(true);
3573 VDDef.Def.setPointer(D);
3575 assert(isa<Instruction>(D) &&
3576 "The dense set member should always be an instruction");
3577 Instruction *Def = cast<Instruction>(D);
3578 VDDef.LocalNum = InstrToDFSNum(D);
3579 DFSOrderedSet.push_back(VDDef);
3580 // If there is a phi node equivalent, add it
3581 if (auto *PN = RealToTemp.lookup(Def)) {
3583 dyn_cast_or_null<PHIExpression>(ValueToExpression.lookup(Def));
3585 VDDef.Def.setInt(false);
3586 VDDef.Def.setPointer(PN);
3588 DFSOrderedSet.push_back(VDDef);
3592 unsigned int UseCount = 0;
3593 // Now add the uses.
3594 for (auto &U : Def->uses()) {
3595 if (auto *I = dyn_cast<Instruction>(U.getUser())) {
3596 // Don't try to replace into dead uses
3597 if (InstructionsToErase.count(I))
3600 // Put the phi node uses in the incoming block.
3602 if (auto *P = dyn_cast<PHINode>(I)) {
3603 IBlock = P->getIncomingBlock(U);
3604 // Make phi node users appear last in the incoming block
3606 VDUse.LocalNum = InstrDFS.size() + 1;
3608 IBlock = getBlockForValue(I);
3609 VDUse.LocalNum = InstrToDFSNum(I);
3612 // Skip uses in unreachable blocks, as we're going
3614 if (ReachableBlocks.count(IBlock) == 0)
3617 DomTreeNode *DomNode = DT->getNode(IBlock);
3618 VDUse.DFSIn = DomNode->getDFSNumIn();
3619 VDUse.DFSOut = DomNode->getDFSNumOut();
3622 DFSOrderedSet.emplace_back(VDUse);
3626 // If there are no uses, it's probably dead (but it may have side-effects,
3627 // so not definitely dead. Otherwise, store the number of uses so we can
3628 // track if it becomes dead later).
3630 ProbablyDead.insert(Def);
3632 UseCounts[Def] = UseCount;
3636 // This function converts the set of members for a congruence class from values,
3637 // to the set of defs for loads and stores, with associated DFS info.
3638 void NewGVN::convertClassToLoadsAndStores(
3639 const CongruenceClass &Dense,
3640 SmallVectorImpl<ValueDFS> &LoadsAndStores) const {
3641 for (auto D : Dense) {
3642 if (!isa<LoadInst>(D) && !isa<StoreInst>(D))
3645 BasicBlock *BB = getBlockForValue(D);
3647 DomTreeNode *DomNode = DT->getNode(BB);
3648 VD.DFSIn = DomNode->getDFSNumIn();
3649 VD.DFSOut = DomNode->getDFSNumOut();
3650 VD.Def.setPointer(D);
3652 // If it's an instruction, use the real local dfs number.
3653 if (auto *I = dyn_cast<Instruction>(D))
3654 VD.LocalNum = InstrToDFSNum(I);
3656 llvm_unreachable("Should have been an instruction");
3658 LoadsAndStores.emplace_back(VD);
3662 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
3663 auto *ReplInst = dyn_cast<Instruction>(Repl);
3667 // Patch the replacement so that it is not more restrictive than the value
3669 // Note that if 'I' is a load being replaced by some operation,
3670 // for example, by an arithmetic operation, then andIRFlags()
3671 // would just erase all math flags from the original arithmetic
3672 // operation, which is clearly not wanted and not needed.
3673 if (!isa<LoadInst>(I))
3674 ReplInst->andIRFlags(I);
3676 // FIXME: If both the original and replacement value are part of the
3677 // same control-flow region (meaning that the execution of one
3678 // guarantees the execution of the other), then we can combine the
3679 // noalias scopes here and do better than the general conservative
3680 // answer used in combineMetadata().
3682 // In general, GVN unifies expressions over different control-flow
3683 // regions, and so we need a conservative combination of the noalias
3685 static const unsigned KnownIDs[] = {
3686 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
3687 LLVMContext::MD_noalias, LLVMContext::MD_range,
3688 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
3689 LLVMContext::MD_invariant_group};
3690 combineMetadata(ReplInst, I, KnownIDs);
3693 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
3694 patchReplacementInstruction(I, Repl);
3695 I->replaceAllUsesWith(Repl);
3698 void NewGVN::deleteInstructionsInBlock(BasicBlock *BB) {
3699 DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
3700 ++NumGVNBlocksDeleted;
3702 // Delete the instructions backwards, as it has a reduced likelihood of having
3703 // to update as many def-use and use-def chains. Start after the terminator.
3704 auto StartPoint = BB->rbegin();
3706 // Note that we explicitly recalculate BB->rend() on each iteration,
3707 // as it may change when we remove the first instruction.
3708 for (BasicBlock::reverse_iterator I(StartPoint); I != BB->rend();) {
3709 Instruction &Inst = *I++;
3710 if (!Inst.use_empty())
3711 Inst.replaceAllUsesWith(UndefValue::get(Inst.getType()));
3712 if (isa<LandingPadInst>(Inst))
3715 Inst.eraseFromParent();
3716 ++NumGVNInstrDeleted;
3718 // Now insert something that simplifycfg will turn into an unreachable.
3719 Type *Int8Ty = Type::getInt8Ty(BB->getContext());
3720 new StoreInst(UndefValue::get(Int8Ty),
3721 Constant::getNullValue(Int8Ty->getPointerTo()),
3722 BB->getTerminator());
3725 void NewGVN::markInstructionForDeletion(Instruction *I) {
3726 DEBUG(dbgs() << "Marking " << *I << " for deletion\n");
3727 InstructionsToErase.insert(I);
3730 void NewGVN::replaceInstruction(Instruction *I, Value *V) {
3731 DEBUG(dbgs() << "Replacing " << *I << " with " << *V << "\n");
3732 patchAndReplaceAllUsesWith(I, V);
3733 // We save the actual erasing to avoid invalidating memory
3734 // dependencies until we are done with everything.
3735 markInstructionForDeletion(I);
3740 // This is a stack that contains both the value and dfs info of where
3741 // that value is valid.
3742 class ValueDFSStack {
3744 Value *back() const { return ValueStack.back(); }
3745 std::pair<int, int> dfs_back() const { return DFSStack.back(); }
3747 void push_back(Value *V, int DFSIn, int DFSOut) {
3748 ValueStack.emplace_back(V);
3749 DFSStack.emplace_back(DFSIn, DFSOut);
3752 bool empty() const { return DFSStack.empty(); }
3754 bool isInScope(int DFSIn, int DFSOut) const {
3757 return DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second;
3760 void popUntilDFSScope(int DFSIn, int DFSOut) {
3762 // These two should always be in sync at this point.
3763 assert(ValueStack.size() == DFSStack.size() &&
3764 "Mismatch between ValueStack and DFSStack");
3766 !DFSStack.empty() &&
3767 !(DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second)) {
3768 DFSStack.pop_back();
3769 ValueStack.pop_back();
3774 SmallVector<Value *, 8> ValueStack;
3775 SmallVector<std::pair<int, int>, 8> DFSStack;
3778 } // end anonymous namespace
3780 // Given an expression, get the congruence class for it.
3781 CongruenceClass *NewGVN::getClassForExpression(const Expression *E) const {
3782 if (auto *VE = dyn_cast<VariableExpression>(E))
3783 return ValueToClass.lookup(VE->getVariableValue());
3784 else if (isa<DeadExpression>(E))
3786 return ExpressionToClass.lookup(E);
3789 // Given a value and a basic block we are trying to see if it is available in,
3790 // see if the value has a leader available in that block.
3791 Value *NewGVN::findPHIOfOpsLeader(const Expression *E,
3792 const Instruction *OrigInst,
3793 const BasicBlock *BB) const {
3794 // It would already be constant if we could make it constant
3795 if (auto *CE = dyn_cast<ConstantExpression>(E))
3796 return CE->getConstantValue();
3797 if (auto *VE = dyn_cast<VariableExpression>(E)) {
3798 auto *V = VE->getVariableValue();
3799 if (alwaysAvailable(V) || DT->dominates(getBlockForValue(V), BB))
3800 return VE->getVariableValue();
3803 auto *CC = getClassForExpression(E);
3806 if (alwaysAvailable(CC->getLeader()))
3807 return CC->getLeader();
3809 for (auto Member : *CC) {
3810 auto *MemberInst = dyn_cast<Instruction>(Member);
3811 if (MemberInst == OrigInst)
3813 // Anything that isn't an instruction is always available.
3816 if (DT->dominates(getBlockForValue(MemberInst), BB))
3822 bool NewGVN::eliminateInstructions(Function &F) {
3823 // This is a non-standard eliminator. The normal way to eliminate is
3824 // to walk the dominator tree in order, keeping track of available
3825 // values, and eliminating them. However, this is mildly
3826 // pointless. It requires doing lookups on every instruction,
3827 // regardless of whether we will ever eliminate it. For
3828 // instructions part of most singleton congruence classes, we know we
3829 // will never eliminate them.
3831 // Instead, this eliminator looks at the congruence classes directly, sorts
3832 // them into a DFS ordering of the dominator tree, and then we just
3833 // perform elimination straight on the sets by walking the congruence
3834 // class member uses in order, and eliminate the ones dominated by the
3835 // last member. This is worst case O(E log E) where E = number of
3836 // instructions in a single congruence class. In theory, this is all
3837 // instructions. In practice, it is much faster, as most instructions are
3838 // either in singleton congruence classes or can't possibly be eliminated
3839 // anyway (if there are no overlapping DFS ranges in class).
3840 // When we find something not dominated, it becomes the new leader
3841 // for elimination purposes.
3842 // TODO: If we wanted to be faster, We could remove any members with no
3843 // overlapping ranges while sorting, as we will never eliminate anything
3844 // with those members, as they don't dominate anything else in our set.
3846 bool AnythingReplaced = false;
3848 // Since we are going to walk the domtree anyway, and we can't guarantee the
3849 // DFS numbers are updated, we compute some ourselves.
3850 DT->updateDFSNumbers();
3852 // Go through all of our phi nodes, and kill the arguments associated with
3853 // unreachable edges.
3854 auto ReplaceUnreachablePHIArgs = [&](PHINode *PHI, BasicBlock *BB) {
3855 for (auto &Operand : PHI->incoming_values())
3856 if (!ReachableEdges.count({PHI->getIncomingBlock(Operand), BB})) {
3857 DEBUG(dbgs() << "Replacing incoming value of " << PHI << " for block "
3858 << getBlockName(PHI->getIncomingBlock(Operand))
3859 << " with undef due to it being unreachable\n");
3860 Operand.set(UndefValue::get(PHI->getType()));
3863 // Replace unreachable phi arguments.
3864 // At this point, RevisitOnReachabilityChange only contains:
3867 // 2. Temporaries that will convert to PHIs
3868 // 3. Operations that are affected by an unreachable edge but do not fit into
3870 // So it is a slight overshoot of what we want. We could make it exact by
3871 // using two SparseBitVectors per block.
3872 DenseMap<const BasicBlock *, unsigned> ReachablePredCount;
3873 for (auto &KV : ReachableEdges)
3874 ReachablePredCount[KV.getEnd()]++;
3875 for (auto &BBPair : RevisitOnReachabilityChange) {
3876 for (auto InstNum : BBPair.second) {
3877 auto *Inst = InstrFromDFSNum(InstNum);
3878 auto *PHI = dyn_cast<PHINode>(Inst);
3879 PHI = PHI ? PHI : dyn_cast_or_null<PHINode>(RealToTemp.lookup(Inst));
3882 auto *BB = BBPair.first;
3883 if (ReachablePredCount.lookup(BB) != PHI->getNumIncomingValues())
3884 ReplaceUnreachablePHIArgs(PHI, BB);
3888 // Map to store the use counts
3889 DenseMap<const Value *, unsigned int> UseCounts;
3890 for (auto *CC : reverse(CongruenceClasses)) {
3891 DEBUG(dbgs() << "Eliminating in congruence class " << CC->getID() << "\n");
3892 // Track the equivalent store info so we can decide whether to try
3893 // dead store elimination.
3894 SmallVector<ValueDFS, 8> PossibleDeadStores;
3895 SmallPtrSet<Instruction *, 8> ProbablyDead;
3896 if (CC->isDead() || CC->empty())
3898 // Everything still in the TOP class is unreachable or dead.
3899 if (CC == TOPClass) {
3900 for (auto M : *CC) {
3901 auto *VTE = ValueToExpression.lookup(M);
3902 if (VTE && isa<DeadExpression>(VTE))
3903 markInstructionForDeletion(cast<Instruction>(M));
3904 assert((!ReachableBlocks.count(cast<Instruction>(M)->getParent()) ||
3905 InstructionsToErase.count(cast<Instruction>(M))) &&
3906 "Everything in TOP should be unreachable or dead at this "
3912 assert(CC->getLeader() && "We should have had a leader");
3913 // If this is a leader that is always available, and it's a
3914 // constant or has no equivalences, just replace everything with
3915 // it. We then update the congruence class with whatever members
3918 CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
3919 if (alwaysAvailable(Leader)) {
3920 CongruenceClass::MemberSet MembersLeft;
3921 for (auto M : *CC) {
3923 // Void things have no uses we can replace.
3924 if (Member == Leader || !isa<Instruction>(Member) ||
3925 Member->getType()->isVoidTy()) {
3926 MembersLeft.insert(Member);
3929 DEBUG(dbgs() << "Found replacement " << *(Leader) << " for " << *Member
3931 auto *I = cast<Instruction>(Member);
3932 assert(Leader != I && "About to accidentally remove our leader");
3933 replaceInstruction(I, Leader);
3934 AnythingReplaced = true;
3936 CC->swap(MembersLeft);
3938 // If this is a singleton, we can skip it.
3939 if (CC->size() != 1 || RealToTemp.count(Leader)) {
3940 // This is a stack because equality replacement/etc may place
3941 // constants in the middle of the member list, and we want to use
3942 // those constant values in preference to the current leader, over
3943 // the scope of those constants.
3944 ValueDFSStack EliminationStack;
3946 // Convert the members to DFS ordered sets and then merge them.
3947 SmallVector<ValueDFS, 8> DFSOrderedSet;
3948 convertClassToDFSOrdered(*CC, DFSOrderedSet, UseCounts, ProbablyDead);
3950 // Sort the whole thing.
3951 std::sort(DFSOrderedSet.begin(), DFSOrderedSet.end());
3952 for (auto &VD : DFSOrderedSet) {
3953 int MemberDFSIn = VD.DFSIn;
3954 int MemberDFSOut = VD.DFSOut;
3955 Value *Def = VD.Def.getPointer();
3956 bool FromStore = VD.Def.getInt();
3958 // We ignore void things because we can't get a value from them.
3959 if (Def && Def->getType()->isVoidTy())
3961 auto *DefInst = dyn_cast_or_null<Instruction>(Def);
3962 if (DefInst && AllTempInstructions.count(DefInst)) {
3963 auto *PN = cast<PHINode>(DefInst);
3965 // If this is a value phi and that's the expression we used, insert
3966 // it into the program
3967 // remove from temp instruction list.
3968 AllTempInstructions.erase(PN);
3969 auto *DefBlock = getBlockForValue(Def);
3970 DEBUG(dbgs() << "Inserting fully real phi of ops" << *Def
3972 << getBlockName(getBlockForValue(Def)) << "\n");
3973 PN->insertBefore(&DefBlock->front());
3975 NumGVNPHIOfOpsEliminations++;
3978 if (EliminationStack.empty()) {
3979 DEBUG(dbgs() << "Elimination Stack is empty\n");
3981 DEBUG(dbgs() << "Elimination Stack Top DFS numbers are ("
3982 << EliminationStack.dfs_back().first << ","
3983 << EliminationStack.dfs_back().second << ")\n");
3986 DEBUG(dbgs() << "Current DFS numbers are (" << MemberDFSIn << ","
3987 << MemberDFSOut << ")\n");
3988 // First, we see if we are out of scope or empty. If so,
3989 // and there equivalences, we try to replace the top of
3990 // stack with equivalences (if it's on the stack, it must
3991 // not have been eliminated yet).
3992 // Then we synchronize to our current scope, by
3993 // popping until we are back within a DFS scope that
3994 // dominates the current member.
3995 // Then, what happens depends on a few factors
3996 // If the stack is now empty, we need to push
3997 // If we have a constant or a local equivalence we want to
3998 // start using, we also push.
3999 // Otherwise, we walk along, processing members who are
4000 // dominated by this scope, and eliminate them.
4001 bool ShouldPush = Def && EliminationStack.empty();
4003 !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut);
4005 if (OutOfScope || ShouldPush) {
4006 // Sync to our current scope.
4007 EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4008 bool ShouldPush = Def && EliminationStack.empty();
4010 EliminationStack.push_back(Def, MemberDFSIn, MemberDFSOut);
4014 // Skip the Def's, we only want to eliminate on their uses. But mark
4015 // dominated defs as dead.
4017 // For anything in this case, what and how we value number
4018 // guarantees that any side-effets that would have occurred (ie
4019 // throwing, etc) can be proven to either still occur (because it's
4020 // dominated by something that has the same side-effects), or never
4021 // occur. Otherwise, we would not have been able to prove it value
4022 // equivalent to something else. For these things, we can just mark
4023 // it all dead. Note that this is different from the "ProbablyDead"
4024 // set, which may not be dominated by anything, and thus, are only
4025 // easy to prove dead if they are also side-effect free. Note that
4026 // because stores are put in terms of the stored value, we skip
4027 // stored values here. If the stored value is really dead, it will
4028 // still be marked for deletion when we process it in its own class.
4029 if (!EliminationStack.empty() && Def != EliminationStack.back() &&
4030 isa<Instruction>(Def) && !FromStore)
4031 markInstructionForDeletion(cast<Instruction>(Def));
4034 // At this point, we know it is a Use we are trying to possibly
4037 assert(isa<Instruction>(U->get()) &&
4038 "Current def should have been an instruction");
4039 assert(isa<Instruction>(U->getUser()) &&
4040 "Current user should have been an instruction");
4042 // If the thing we are replacing into is already marked to be dead,
4043 // this use is dead. Note that this is true regardless of whether
4044 // we have anything dominating the use or not. We do this here
4045 // because we are already walking all the uses anyway.
4046 Instruction *InstUse = cast<Instruction>(U->getUser());
4047 if (InstructionsToErase.count(InstUse)) {
4048 auto &UseCount = UseCounts[U->get()];
4049 if (--UseCount == 0) {
4050 ProbablyDead.insert(cast<Instruction>(U->get()));
4054 // If we get to this point, and the stack is empty we must have a use
4055 // with nothing we can use to eliminate this use, so just skip it.
4056 if (EliminationStack.empty())
4059 Value *DominatingLeader = EliminationStack.back();
4061 auto *II = dyn_cast<IntrinsicInst>(DominatingLeader);
4062 bool isSSACopy = II && II->getIntrinsicID() == Intrinsic::ssa_copy;
4064 DominatingLeader = II->getOperand(0);
4066 // Don't replace our existing users with ourselves.
4067 if (U->get() == DominatingLeader)
4069 DEBUG(dbgs() << "Found replacement " << *DominatingLeader << " for "
4070 << *U->get() << " in " << *(U->getUser()) << "\n");
4072 // If we replaced something in an instruction, handle the patching of
4073 // metadata. Skip this if we are replacing predicateinfo with its
4074 // original operand, as we already know we can just drop it.
4075 auto *ReplacedInst = cast<Instruction>(U->get());
4076 auto *PI = PredInfo->getPredicateInfoFor(ReplacedInst);
4077 if (!PI || DominatingLeader != PI->OriginalOp)
4078 patchReplacementInstruction(ReplacedInst, DominatingLeader);
4079 U->set(DominatingLeader);
4080 // This is now a use of the dominating leader, which means if the
4081 // dominating leader was dead, it's now live!
4082 auto &LeaderUseCount = UseCounts[DominatingLeader];
4083 // It's about to be alive again.
4084 if (LeaderUseCount == 0 && isa<Instruction>(DominatingLeader))
4085 ProbablyDead.erase(cast<Instruction>(DominatingLeader));
4086 // Copy instructions, however, are still dead beacuse we use their
4087 // operand as the leader.
4088 if (LeaderUseCount == 0 && isSSACopy)
4089 ProbablyDead.insert(II);
4091 AnythingReplaced = true;
4096 // At this point, anything still in the ProbablyDead set is actually dead if
4097 // would be trivially dead.
4098 for (auto *I : ProbablyDead)
4099 if (wouldInstructionBeTriviallyDead(I))
4100 markInstructionForDeletion(I);
4102 // Cleanup the congruence class.
4103 CongruenceClass::MemberSet MembersLeft;
4104 for (auto *Member : *CC)
4105 if (!isa<Instruction>(Member) ||
4106 !InstructionsToErase.count(cast<Instruction>(Member)))
4107 MembersLeft.insert(Member);
4108 CC->swap(MembersLeft);
4110 // If we have possible dead stores to look at, try to eliminate them.
4111 if (CC->getStoreCount() > 0) {
4112 convertClassToLoadsAndStores(*CC, PossibleDeadStores);
4113 std::sort(PossibleDeadStores.begin(), PossibleDeadStores.end());
4114 ValueDFSStack EliminationStack;
4115 for (auto &VD : PossibleDeadStores) {
4116 int MemberDFSIn = VD.DFSIn;
4117 int MemberDFSOut = VD.DFSOut;
4118 Instruction *Member = cast<Instruction>(VD.Def.getPointer());
4119 if (EliminationStack.empty() ||
4120 !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut)) {
4121 // Sync to our current scope.
4122 EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4123 if (EliminationStack.empty()) {
4124 EliminationStack.push_back(Member, MemberDFSIn, MemberDFSOut);
4128 // We already did load elimination, so nothing to do here.
4129 if (isa<LoadInst>(Member))
4131 assert(!EliminationStack.empty());
4132 Instruction *Leader = cast<Instruction>(EliminationStack.back());
4134 assert(DT->dominates(Leader->getParent(), Member->getParent()));
4135 // Member is dominater by Leader, and thus dead
4136 DEBUG(dbgs() << "Marking dead store " << *Member
4137 << " that is dominated by " << *Leader << "\n");
4138 markInstructionForDeletion(Member);
4144 return AnythingReplaced;
4147 // This function provides global ranking of operations so that we can place them
4148 // in a canonical order. Note that rank alone is not necessarily enough for a
4149 // complete ordering, as constants all have the same rank. However, generally,
4150 // we will simplify an operation with all constants so that it doesn't matter
4151 // what order they appear in.
4152 unsigned int NewGVN::getRank(const Value *V) const {
4153 // Prefer constants to undef to anything else
4154 // Undef is a constant, have to check it first.
4155 // Prefer smaller constants to constantexprs
4156 if (isa<ConstantExpr>(V))
4158 if (isa<UndefValue>(V))
4160 if (isa<Constant>(V))
4162 else if (auto *A = dyn_cast<Argument>(V))
4163 return 3 + A->getArgNo();
4165 // Need to shift the instruction DFS by number of arguments + 3 to account for
4166 // the constant and argument ranking above.
4167 unsigned Result = InstrToDFSNum(V);
4169 return 4 + NumFuncArgs + Result;
4170 // Unreachable or something else, just return a really large number.
4174 // This is a function that says whether two commutative operations should
4175 // have their order swapped when canonicalizing.
4176 bool NewGVN::shouldSwapOperands(const Value *A, const Value *B) const {
4177 // Because we only care about a total ordering, and don't rewrite expressions
4178 // in this order, we order by rank, which will give a strict weak ordering to
4179 // everything but constants, and then we order by pointer address.
4180 return std::make_pair(getRank(A), A) > std::make_pair(getRank(B), B);
4185 class NewGVNLegacyPass : public FunctionPass {
4187 // Pass identification, replacement for typeid.
4190 NewGVNLegacyPass() : FunctionPass(ID) {
4191 initializeNewGVNLegacyPassPass(*PassRegistry::getPassRegistry());
4194 bool runOnFunction(Function &F) override;
4197 void getAnalysisUsage(AnalysisUsage &AU) const override {
4198 AU.addRequired<AssumptionCacheTracker>();
4199 AU.addRequired<DominatorTreeWrapperPass>();
4200 AU.addRequired<TargetLibraryInfoWrapperPass>();
4201 AU.addRequired<MemorySSAWrapperPass>();
4202 AU.addRequired<AAResultsWrapperPass>();
4203 AU.addPreserved<DominatorTreeWrapperPass>();
4204 AU.addPreserved<GlobalsAAWrapperPass>();
4208 } // end anonymous namespace
4210 bool NewGVNLegacyPass::runOnFunction(Function &F) {
4211 if (skipFunction(F))
4213 return NewGVN(F, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4214 &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
4215 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
4216 &getAnalysis<AAResultsWrapperPass>().getAAResults(),
4217 &getAnalysis<MemorySSAWrapperPass>().getMSSA(),
4218 F.getParent()->getDataLayout())
4222 char NewGVNLegacyPass::ID = 0;
4224 INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering",
4226 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4227 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
4228 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4229 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
4230 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
4231 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
4232 INITIALIZE_PASS_END(NewGVNLegacyPass, "newgvn", "Global Value Numbering", false,
4235 // createGVNPass - The public interface to this file.
4236 FunctionPass *llvm::createNewGVNPass() { return new NewGVNLegacyPass(); }
4238 PreservedAnalyses NewGVNPass::run(Function &F, AnalysisManager<Function> &AM) {
4239 // Apparently the order in which we get these results matter for
4240 // the old GVN (see Chandler's comment in GVN.cpp). I'll keep
4241 // the same order here, just in case.
4242 auto &AC = AM.getResult<AssumptionAnalysis>(F);
4243 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
4244 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
4245 auto &AA = AM.getResult<AAManager>(F);
4246 auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
4248 NewGVN(F, &DT, &AC, &TLI, &AA, &MSSA, F.getParent()->getDataLayout())
4251 return PreservedAnalyses::all();
4252 PreservedAnalyses PA;
4253 PA.preserve<DominatorTreeAnalysis>();
4254 PA.preserve<GlobalsAA>();