xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/NewGVN.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- NewGVN.cpp - Global Value Numbering Pass ---------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// This file implements the new LLVM's Global Value Numbering pass.
11 /// GVN partitions values computed by a function into congruence classes.
12 /// Values ending up in the same congruence class are guaranteed to be the same
13 /// for every execution of the program. In that respect, congruency is a
14 /// compile-time approximation of equivalence of values at runtime.
15 /// The algorithm implemented here uses a sparse formulation and it's based
16 /// on the ideas described in the paper:
17 /// "A Sparse Algorithm for Predicated Global Value Numbering" from
18 /// Karthik Gargi.
19 ///
20 /// A brief overview of the algorithm: The algorithm is essentially the same as
21 /// the standard RPO value numbering algorithm (a good reference is the paper
22 /// "SCC based value numbering" by L. Taylor Simpson) with one major difference:
23 /// The RPO algorithm proceeds, on every iteration, to process every reachable
24 /// block and every instruction in that block.  This is because the standard RPO
25 /// algorithm does not track what things have the same value number, it only
26 /// tracks what the value number of a given operation is (the mapping is
27 /// operation -> value number).  Thus, when a value number of an operation
28 /// changes, it must reprocess everything to ensure all uses of a value number
29 /// get updated properly.  In constrast, the sparse algorithm we use *also*
30 /// tracks what operations have a given value number (IE it also tracks the
31 /// reverse mapping from value number -> operations with that value number), so
32 /// that it only needs to reprocess the instructions that are affected when
33 /// something's value number changes.  The vast majority of complexity and code
34 /// in this file is devoted to tracking what value numbers could change for what
35 /// instructions when various things happen.  The rest of the algorithm is
36 /// devoted to performing symbolic evaluation, forward propagation, and
37 /// simplification of operations based on the value numbers deduced so far
38 ///
39 /// In order to make the GVN mostly-complete, we use a technique derived from
40 /// "Detection of Redundant Expressions: A Complete and Polynomial-time
41 /// Algorithm in SSA" by R.R. Pai.  The source of incompleteness in most SSA
42 /// based GVN algorithms is related to their inability to detect equivalence
43 /// between phi of ops (IE phi(a+b, c+d)) and op of phis (phi(a,c) + phi(b, d)).
44 /// We resolve this issue by generating the equivalent "phi of ops" form for
45 /// each op of phis we see, in a way that only takes polynomial time to resolve.
46 ///
47 /// We also do not perform elimination by using any published algorithm.  All
48 /// published algorithms are O(Instructions). Instead, we use a technique that
49 /// is O(number of operations with the same value number), enabling us to skip
50 /// trying to eliminate things that have unique value numbers.
51 //
52 //===----------------------------------------------------------------------===//
53 
54 #include "llvm/Transforms/Scalar/NewGVN.h"
55 #include "llvm/ADT/ArrayRef.h"
56 #include "llvm/ADT/BitVector.h"
57 #include "llvm/ADT/DenseMap.h"
58 #include "llvm/ADT/DenseMapInfo.h"
59 #include "llvm/ADT/DenseSet.h"
60 #include "llvm/ADT/DepthFirstIterator.h"
61 #include "llvm/ADT/GraphTraits.h"
62 #include "llvm/ADT/Hashing.h"
63 #include "llvm/ADT/PointerIntPair.h"
64 #include "llvm/ADT/PostOrderIterator.h"
65 #include "llvm/ADT/SetOperations.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/ValueTracking.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/PatternMatch.h"
92 #include "llvm/IR/Type.h"
93 #include "llvm/IR/Use.h"
94 #include "llvm/IR/User.h"
95 #include "llvm/IR/Value.h"
96 #include "llvm/Support/Allocator.h"
97 #include "llvm/Support/ArrayRecycler.h"
98 #include "llvm/Support/Casting.h"
99 #include "llvm/Support/CommandLine.h"
100 #include "llvm/Support/Debug.h"
101 #include "llvm/Support/DebugCounter.h"
102 #include "llvm/Support/ErrorHandling.h"
103 #include "llvm/Support/PointerLikeTypeTraits.h"
104 #include "llvm/Support/raw_ostream.h"
105 #include "llvm/Transforms/Scalar/GVNExpression.h"
106 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
107 #include "llvm/Transforms/Utils/Local.h"
108 #include "llvm/Transforms/Utils/PredicateInfo.h"
109 #include "llvm/Transforms/Utils/VNCoercion.h"
110 #include <algorithm>
111 #include <cassert>
112 #include <cstdint>
113 #include <iterator>
114 #include <map>
115 #include <memory>
116 #include <set>
117 #include <string>
118 #include <tuple>
119 #include <utility>
120 #include <vector>
121 
122 using namespace llvm;
123 using namespace llvm::GVNExpression;
124 using namespace llvm::VNCoercion;
125 using namespace llvm::PatternMatch;
126 
127 #define DEBUG_TYPE "newgvn"
128 
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
149 // issue.
150 static cl::opt<bool> EnableStoreRefinement("enable-store-refinement",
151                                            cl::init(false), cl::Hidden);
152 
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),
155                                     cl::Hidden);
156 
157 //===----------------------------------------------------------------------===//
158 //                                GVN Pass
159 //===----------------------------------------------------------------------===//
160 
161 // Anchor methods.
162 namespace llvm {
163 namespace GVNExpression {
164 
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;
172 
173 } // end namespace GVNExpression
174 } // end namespace llvm
175 
176 namespace {
177 
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
184 // instructions,
185 // not generic values (arguments, etc).
186 struct TarjanSCC {
TarjanSCC__anonb3555ec80111::TarjanSCC187   TarjanSCC() : Components(1) {}
188 
Start__anonb3555ec80111::TarjanSCC189   void Start(const Instruction *Start) {
190     if (Root.lookup(Start) == 0)
191       FindSCC(Start);
192   }
193 
getComponentFor__anonb3555ec80111::TarjanSCC194   const SmallPtrSetImpl<const Value *> &getComponentFor(const Value *V) const {
195     unsigned ComponentID = ValueToComponent.lookup(V);
196 
197     assert(ComponentID > 0 &&
198            "Asking for a component for a value we never processed");
199     return Components[ComponentID];
200   }
201 
202 private:
FindSCC__anonb3555ec80111::TarjanSCC203   void FindSCC(const Instruction *I) {
204     Root[I] = ++DFSNum;
205     // Store the DFS Number we had before it possibly gets incremented.
206     unsigned int OurDFS = DFSNum;
207     for (const auto &Op : I->operands()) {
208       if (auto *InstOp = dyn_cast<Instruction>(Op)) {
209         if (Root.lookup(Op) == 0)
210           FindSCC(InstOp);
211         if (!InComponent.count(Op))
212           Root[I] = std::min(Root.lookup(I), Root.lookup(Op));
213       }
214     }
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();
223       Component.insert(I);
224       LLVM_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         LLVM_DEBUG(dbgs() << "Component member is " << *Member << "\n");
231         Component.insert(Member);
232         InComponent.insert(Member);
233         ValueToComponent[Member] = ComponentID;
234         Stack.pop_back();
235       }
236     } else {
237       // Part of a component, push to stack
238       Stack.push_back(I);
239     }
240   }
241 
242   unsigned int DFSNum = 1;
243   SmallPtrSet<const Value *, 8> InComponent;
244   DenseMap<const Value *, unsigned int> Root;
245   SmallVector<const Value *, 8> Stack;
246 
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;
250 
251   DenseMap<const Value *, unsigned> ValueToComponent;
252 };
253 
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.
260 //
261 // For any Value in the Member set, it is valid to replace any dominated member
262 // with that Value.
263 //
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
272 // of values.
273 
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 {
293 public:
294   using MemberType = Value;
295   using MemberSet = SmallPtrSet<MemberType *, 4>;
296   using MemoryMemberType = MemoryPhi;
297   using MemoryMemberSet = SmallPtrSet<const MemoryMemberType *, 2>;
298 
CongruenceClass(unsigned ID)299   explicit CongruenceClass(unsigned ID) : ID(ID) {}
CongruenceClass(unsigned ID,Value * Leader,const Expression * E)300   CongruenceClass(unsigned ID, Value *Leader, const Expression *E)
301       : ID(ID), RepLeader(Leader), DefiningExpr(E) {}
302 
getID() const303   unsigned getID() const { return ID; }
304 
305   // True if this class has no members left.  This is mainly used for assertion
306   // purposes, and for skipping empty classes.
isDead() const307   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();
311   }
312 
313   // Leader functions
getLeader() const314   Value *getLeader() const { return RepLeader; }
setLeader(Value * Leader)315   void setLeader(Value *Leader) { RepLeader = Leader; }
getNextLeader() const316   const std::pair<Value *, unsigned int> &getNextLeader() const {
317     return NextLeader;
318   }
resetNextLeader()319   void resetNextLeader() { NextLeader = {nullptr, ~0}; }
addPossibleNextLeader(std::pair<Value *,unsigned int> LeaderPair)320   void addPossibleNextLeader(std::pair<Value *, unsigned int> LeaderPair) {
321     if (LeaderPair.second < NextLeader.second)
322       NextLeader = LeaderPair;
323   }
324 
getStoredValue() const325   Value *getStoredValue() const { return RepStoredValue; }
setStoredValue(Value * Leader)326   void setStoredValue(Value *Leader) { RepStoredValue = Leader; }
getMemoryLeader() const327   const MemoryAccess *getMemoryLeader() const { return RepMemoryAccess; }
setMemoryLeader(const MemoryAccess * Leader)328   void setMemoryLeader(const MemoryAccess *Leader) { RepMemoryAccess = Leader; }
329 
330   // Forward propagation info
getDefiningExpr() const331   const Expression *getDefiningExpr() const { return DefiningExpr; }
332 
333   // Value member set
empty() const334   bool empty() const { return Members.empty(); }
size() const335   unsigned size() const { return Members.size(); }
begin() const336   MemberSet::const_iterator begin() const { return Members.begin(); }
end() const337   MemberSet::const_iterator end() const { return Members.end(); }
insert(MemberType * M)338   void insert(MemberType *M) { Members.insert(M); }
erase(MemberType * M)339   void erase(MemberType *M) { Members.erase(M); }
swap(MemberSet & Other)340   void swap(MemberSet &Other) { Members.swap(Other); }
341 
342   // Memory member set
memory_empty() const343   bool memory_empty() const { return MemoryMembers.empty(); }
memory_size() const344   unsigned memory_size() const { return MemoryMembers.size(); }
memory_begin() const345   MemoryMemberSet::const_iterator memory_begin() const {
346     return MemoryMembers.begin();
347   }
memory_end() const348   MemoryMemberSet::const_iterator memory_end() const {
349     return MemoryMembers.end();
350   }
memory() const351   iterator_range<MemoryMemberSet::const_iterator> memory() const {
352     return make_range(memory_begin(), memory_end());
353   }
354 
memory_insert(const MemoryMemberType * M)355   void memory_insert(const MemoryMemberType *M) { MemoryMembers.insert(M); }
memory_erase(const MemoryMemberType * M)356   void memory_erase(const MemoryMemberType *M) { MemoryMembers.erase(M); }
357 
358   // Store count
getStoreCount() const359   unsigned getStoreCount() const { return StoreCount; }
incStoreCount()360   void incStoreCount() { ++StoreCount; }
decStoreCount()361   void decStoreCount() {
362     assert(StoreCount != 0 && "Store count went negative");
363     --StoreCount;
364   }
365 
366   // True if this class has no memory members.
definesNoMemory() const367   bool definesNoMemory() const { return StoreCount == 0 && memory_empty(); }
368 
369   // Return true if two congruence classes are equivalent to each other. This
370   // means that every field but the ID number and the dead field are equivalent.
isEquivalentTo(const CongruenceClass * Other) const371   bool isEquivalentTo(const CongruenceClass *Other) const {
372     if (!Other)
373       return false;
374     if (this == Other)
375       return true;
376 
377     if (std::tie(StoreCount, RepLeader, RepStoredValue, RepMemoryAccess) !=
378         std::tie(Other->StoreCount, Other->RepLeader, Other->RepStoredValue,
379                  Other->RepMemoryAccess))
380       return false;
381     if (DefiningExpr != Other->DefiningExpr)
382       if (!DefiningExpr || !Other->DefiningExpr ||
383           *DefiningExpr != *Other->DefiningExpr)
384         return false;
385 
386     if (Members.size() != Other->Members.size())
387       return false;
388 
389     return llvm::set_is_subset(Members, Other->Members);
390   }
391 
392 private:
393   unsigned ID;
394 
395   // Representative leader.
396   Value *RepLeader = nullptr;
397 
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};
401 
402   // If this is represented by a store, the value of the store.
403   Value *RepStoredValue = nullptr;
404 
405   // If this class contains MemoryDefs or MemoryPhis, this is the leading memory
406   // access.
407   const MemoryAccess *RepMemoryAccess = nullptr;
408 
409   // Defining Expression.
410   const Expression *DefiningExpr = nullptr;
411 
412   // Actual members of this class.
413   MemberSet Members;
414 
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;
419 
420   // Number of stores in this congruence class.
421   // This is used so we can detect store equivalence changes properly.
422   int StoreCount = 0;
423 };
424 
425 } // end anonymous namespace
426 
427 namespace llvm {
428 
429 struct ExactEqualsExpression {
430   const Expression &E;
431 
ExactEqualsExpressionllvm::ExactEqualsExpression432   explicit ExactEqualsExpression(const Expression &E) : E(E) {}
433 
getComputedHashllvm::ExactEqualsExpression434   hash_code getComputedHash() const { return E.getComputedHash(); }
435 
operator ==llvm::ExactEqualsExpression436   bool operator==(const Expression &Other) const {
437     return E.exactlyEquals(Other);
438   }
439 };
440 
441 template <> struct DenseMapInfo<const Expression *> {
getEmptyKeyllvm::DenseMapInfo442   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);
446   }
447 
getTombstoneKeyllvm::DenseMapInfo448   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);
452   }
453 
getHashValuellvm::DenseMapInfo454   static unsigned getHashValue(const Expression *E) {
455     return E->getComputedHash();
456   }
457 
getHashValuellvm::DenseMapInfo458   static unsigned getHashValue(const ExactEqualsExpression &E) {
459     return E.getComputedHash();
460   }
461 
isEqualllvm::DenseMapInfo462   static bool isEqual(const ExactEqualsExpression &LHS, const Expression *RHS) {
463     if (RHS == getTombstoneKey() || RHS == getEmptyKey())
464       return false;
465     return LHS == *RHS;
466   }
467 
isEqualllvm::DenseMapInfo468   static bool isEqual(const Expression *LHS, const Expression *RHS) {
469     if (LHS == RHS)
470       return true;
471     if (LHS == getTombstoneKey() || RHS == getTombstoneKey() ||
472         LHS == getEmptyKey() || RHS == getEmptyKey())
473       return false;
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())
479       return false;
480     return *LHS == *RHS;
481   }
482 };
483 
484 } // end namespace llvm
485 
486 namespace {
487 
488 class NewGVN {
489   Function &F;
490   DominatorTree *DT = nullptr;
491   const TargetLibraryInfo *TLI = nullptr;
492   AliasAnalysis *AA = nullptr;
493   MemorySSA *MSSA = nullptr;
494   MemorySSAWalker *MSSAWalker = nullptr;
495   AssumptionCache *AC = nullptr;
496   const DataLayout &DL;
497   std::unique_ptr<PredicateInfo> PredInfo;
498 
499   // These are the only two things the create* functions should have
500   // side-effects on due to allocating memory.
501   mutable BumpPtrAllocator ExpressionAllocator;
502   mutable ArrayRecycler<Value *> ArgRecycler;
503   mutable TarjanSCC SCCFinder;
504   const SimplifyQuery SQ;
505 
506   // Number of function arguments, used by ranking
507   unsigned int NumFuncArgs = 0;
508 
509   // RPOOrdering of basic blocks
510   DenseMap<const DomTreeNode *, unsigned> RPOOrdering;
511 
512   // Congruence class info.
513 
514   // This class is called INITIAL in the paper. It is the class everything
515   // startsout in, and represents any value. Being an optimistic analysis,
516   // anything in the TOP class has the value TOP, which is indeterminate and
517   // equivalent to everything.
518   CongruenceClass *TOPClass = nullptr;
519   std::vector<CongruenceClass *> CongruenceClasses;
520   unsigned NextCongruenceNum = 0;
521 
522   // Value Mappings.
523   DenseMap<Value *, CongruenceClass *> ValueToClass;
524   DenseMap<Value *, const Expression *> ValueToExpression;
525 
526   // Value PHI handling, used to make equivalence between phi(op, op) and
527   // op(phi, phi).
528   // These mappings just store various data that would normally be part of the
529   // IR.
530   SmallPtrSet<const Instruction *, 8> PHINodeUses;
531 
532   // The cached results, in general, are only valid for the specific block where
533   // they were computed. The unsigned part of the key is a unique block
534   // identifier
535   DenseMap<std::pair<const Value *, unsigned>, bool> OpSafeForPHIOfOps;
536   unsigned CacheIdx;
537 
538   // Map a temporary instruction we created to a parent block.
539   DenseMap<const Value *, BasicBlock *> TempToBlock;
540 
541   // Map between the already in-program instructions and the temporary phis we
542   // created that they are known equivalent to.
543   DenseMap<const Value *, PHINode *> RealToTemp;
544 
545   // In order to know when we should re-process instructions that have
546   // phi-of-ops, we track the set of expressions that they needed as
547   // leaders. When we discover new leaders for those expressions, we process the
548   // associated phi-of-op instructions again in case they have changed.  The
549   // other way they may change is if they had leaders, and those leaders
550   // disappear.  However, at the point they have leaders, there are uses of the
551   // relevant operands in the created phi node, and so they will get reprocessed
552   // through the normal user marking we perform.
553   mutable DenseMap<const Value *, SmallPtrSet<Value *, 2>> AdditionalUsers;
554   DenseMap<const Expression *, SmallPtrSet<Instruction *, 2>>
555       ExpressionToPhiOfOps;
556 
557   // Map from temporary operation to MemoryAccess.
558   DenseMap<const Instruction *, MemoryUseOrDef *> TempToMemory;
559 
560   // Set of all temporary instructions we created.
561   // Note: This will include instructions that were just created during value
562   // numbering.  The way to test if something is using them is to check
563   // RealToTemp.
564   DenseSet<Instruction *> AllTempInstructions;
565 
566   // This is the set of instructions to revisit on a reachability change.  At
567   // the end of the main iteration loop it will contain at least all the phi of
568   // ops instructions that will be changed to phis, as well as regular phis.
569   // During the iteration loop, it may contain other things, such as phi of ops
570   // instructions that used edge reachability to reach a result, and so need to
571   // be revisited when the edge changes, independent of whether the phi they
572   // depended on changes.
573   DenseMap<BasicBlock *, SparseBitVector<>> RevisitOnReachabilityChange;
574 
575   // Mapping from predicate info we used to the instructions we used it with.
576   // In order to correctly ensure propagation, we must keep track of what
577   // comparisons we used, so that when the values of the comparisons change, we
578   // propagate the information to the places we used the comparison.
579   mutable DenseMap<const Value *, SmallPtrSet<Instruction *, 2>>
580       PredicateToUsers;
581 
582   // the same reasoning as PredicateToUsers.  When we skip MemoryAccesses for
583   // stores, we no longer can rely solely on the def-use chains of MemorySSA.
584   mutable DenseMap<const MemoryAccess *, SmallPtrSet<MemoryAccess *, 2>>
585       MemoryToUsers;
586 
587   // A table storing which memorydefs/phis represent a memory state provably
588   // equivalent to another memory state.
589   // We could use the congruence class machinery, but the MemoryAccess's are
590   // abstract memory states, so they can only ever be equivalent to each other,
591   // and not to constants, etc.
592   DenseMap<const MemoryAccess *, CongruenceClass *> MemoryAccessToClass;
593 
594   // We could, if we wanted, build MemoryPhiExpressions and
595   // MemoryVariableExpressions, etc, and value number them the same way we value
596   // number phi expressions.  For the moment, this seems like overkill.  They
597   // can only exist in one of three states: they can be TOP (equal to
598   // everything), Equivalent to something else, or unique.  Because we do not
599   // create expressions for them, we need to simulate leader change not just
600   // when they change class, but when they change state.  Note: We can do the
601   // same thing for phis, and avoid having phi expressions if we wanted, We
602   // should eventually unify in one direction or the other, so this is a little
603   // bit of an experiment in which turns out easier to maintain.
604   enum MemoryPhiState { MPS_Invalid, MPS_TOP, MPS_Equivalent, MPS_Unique };
605   DenseMap<const MemoryPhi *, MemoryPhiState> MemoryPhiState;
606 
607   enum InstCycleState { ICS_Unknown, ICS_CycleFree, ICS_Cycle };
608   mutable DenseMap<const Instruction *, InstCycleState> InstCycleState;
609 
610   // Expression to class mapping.
611   using ExpressionClassMap = DenseMap<const Expression *, CongruenceClass *>;
612   ExpressionClassMap ExpressionToClass;
613 
614   // We have a single expression that represents currently DeadExpressions.
615   // For dead expressions we can prove will stay dead, we mark them with
616   // DFS number zero.  However, it's possible in the case of phi nodes
617   // for us to assume/prove all arguments are dead during fixpointing.
618   // We use DeadExpression for that case.
619   DeadExpression *SingletonDeadExpression = nullptr;
620 
621   // Which values have changed as a result of leader changes.
622   SmallPtrSet<Value *, 8> LeaderChanges;
623 
624   // Reachability info.
625   using BlockEdge = BasicBlockEdge;
626   DenseSet<BlockEdge> ReachableEdges;
627   SmallPtrSet<const BasicBlock *, 8> ReachableBlocks;
628 
629   // This is a bitvector because, on larger functions, we may have
630   // thousands of touched instructions at once (entire blocks,
631   // instructions with hundreds of uses, etc).  Even with optimization
632   // for when we mark whole blocks as touched, when this was a
633   // SmallPtrSet or DenseSet, for some functions, we spent >20% of all
634   // the time in GVN just managing this list.  The bitvector, on the
635   // other hand, efficiently supports test/set/clear of both
636   // individual and ranges, as well as "find next element" This
637   // enables us to use it as a worklist with essentially 0 cost.
638   BitVector TouchedInstructions;
639 
640   DenseMap<const BasicBlock *, std::pair<unsigned, unsigned>> BlockInstRange;
641   mutable DenseMap<const IntrinsicInst *, const Value *> IntrinsicInstPred;
642 
643 #ifndef NDEBUG
644   // Debugging for how many times each block and instruction got processed.
645   DenseMap<const Value *, unsigned> ProcessedCount;
646 #endif
647 
648   // DFS info.
649   // This contains a mapping from Instructions to DFS numbers.
650   // The numbering starts at 1. An instruction with DFS number zero
651   // means that the instruction is dead.
652   DenseMap<const Value *, unsigned> InstrDFS;
653 
654   // This contains the mapping DFS numbers to instructions.
655   SmallVector<Value *, 32> DFSToInstr;
656 
657   // Deletion info.
658   SmallPtrSet<Instruction *, 8> InstructionsToErase;
659 
660 public:
NewGVN(Function & F,DominatorTree * DT,AssumptionCache * AC,TargetLibraryInfo * TLI,AliasAnalysis * AA,MemorySSA * MSSA,const DataLayout & DL)661   NewGVN(Function &F, DominatorTree *DT, AssumptionCache *AC,
662          TargetLibraryInfo *TLI, AliasAnalysis *AA, MemorySSA *MSSA,
663          const DataLayout &DL)
664       : F(F), DT(DT), TLI(TLI), AA(AA), MSSA(MSSA), AC(AC), DL(DL),
665         PredInfo(std::make_unique<PredicateInfo>(F, *DT, *AC)),
666         SQ(DL, TLI, DT, AC, /*CtxI=*/nullptr, /*UseInstrInfo=*/false,
667            /*CanUseUndef=*/false) {}
668 
669   bool runGVN();
670 
671 private:
672   /// Helper struct return a Expression with an optional extra dependency.
673   struct ExprResult {
674     const Expression *Expr;
675     Value *ExtraDep;
676     const PredicateBase *PredDep;
677 
ExprResult__anonb3555ec80211::NewGVN::ExprResult678     ExprResult(const Expression *Expr, Value *ExtraDep = nullptr,
679                const PredicateBase *PredDep = nullptr)
680         : Expr(Expr), ExtraDep(ExtraDep), PredDep(PredDep) {}
681     ExprResult(const ExprResult &) = delete;
ExprResult__anonb3555ec80211::NewGVN::ExprResult682     ExprResult(ExprResult &&Other)
683         : Expr(Other.Expr), ExtraDep(Other.ExtraDep), PredDep(Other.PredDep) {
684       Other.Expr = nullptr;
685       Other.ExtraDep = nullptr;
686       Other.PredDep = nullptr;
687     }
688     ExprResult &operator=(const ExprResult &Other) = delete;
689     ExprResult &operator=(ExprResult &&Other) = delete;
690 
~ExprResult__anonb3555ec80211::NewGVN::ExprResult691     ~ExprResult() { assert(!ExtraDep && "unhandled ExtraDep"); }
692 
operator bool__anonb3555ec80211::NewGVN::ExprResult693     operator bool() const { return Expr; }
694 
none__anonb3555ec80211::NewGVN::ExprResult695     static ExprResult none() { return {nullptr, nullptr, nullptr}; }
some__anonb3555ec80211::NewGVN::ExprResult696     static ExprResult some(const Expression *Expr, Value *ExtraDep = nullptr) {
697       return {Expr, ExtraDep, nullptr};
698     }
some__anonb3555ec80211::NewGVN::ExprResult699     static ExprResult some(const Expression *Expr,
700                            const PredicateBase *PredDep) {
701       return {Expr, nullptr, PredDep};
702     }
some__anonb3555ec80211::NewGVN::ExprResult703     static ExprResult some(const Expression *Expr, Value *ExtraDep,
704                            const PredicateBase *PredDep) {
705       return {Expr, ExtraDep, PredDep};
706     }
707   };
708 
709   // Expression handling.
710   ExprResult createExpression(Instruction *) const;
711   const Expression *createBinaryExpression(unsigned, Type *, Value *, Value *,
712                                            Instruction *) const;
713 
714   // Our canonical form for phi arguments is a pair of incoming value, incoming
715   // basic block.
716   using ValPair = std::pair<Value *, BasicBlock *>;
717 
718   PHIExpression *createPHIExpression(ArrayRef<ValPair>, const Instruction *,
719                                      BasicBlock *, bool &HasBackEdge,
720                                      bool &OriginalOpsConstant) const;
721   const DeadExpression *createDeadExpression() const;
722   const VariableExpression *createVariableExpression(Value *) const;
723   const ConstantExpression *createConstantExpression(Constant *) const;
724   const Expression *createVariableOrConstant(Value *V) const;
725   const UnknownExpression *createUnknownExpression(Instruction *) const;
726   const StoreExpression *createStoreExpression(StoreInst *,
727                                                const MemoryAccess *) const;
728   LoadExpression *createLoadExpression(Type *, Value *, LoadInst *,
729                                        const MemoryAccess *) const;
730   const CallExpression *createCallExpression(CallInst *,
731                                              const MemoryAccess *) const;
732   const AggregateValueExpression *
733   createAggregateValueExpression(Instruction *) const;
734   bool setBasicExpressionInfo(Instruction *, BasicExpression *) const;
735 
736   // Congruence class handling.
createCongruenceClass(Value * Leader,const Expression * E)737   CongruenceClass *createCongruenceClass(Value *Leader, const Expression *E) {
738     auto *result = new CongruenceClass(NextCongruenceNum++, Leader, E);
739     CongruenceClasses.emplace_back(result);
740     return result;
741   }
742 
createMemoryClass(MemoryAccess * MA)743   CongruenceClass *createMemoryClass(MemoryAccess *MA) {
744     auto *CC = createCongruenceClass(nullptr, nullptr);
745     CC->setMemoryLeader(MA);
746     return CC;
747   }
748 
ensureLeaderOfMemoryClass(MemoryAccess * MA)749   CongruenceClass *ensureLeaderOfMemoryClass(MemoryAccess *MA) {
750     auto *CC = getMemoryClass(MA);
751     if (CC->getMemoryLeader() != MA)
752       CC = createMemoryClass(MA);
753     return CC;
754   }
755 
createSingletonCongruenceClass(Value * Member)756   CongruenceClass *createSingletonCongruenceClass(Value *Member) {
757     CongruenceClass *CClass = createCongruenceClass(Member, nullptr);
758     CClass->insert(Member);
759     ValueToClass[Member] = CClass;
760     return CClass;
761   }
762 
763   void initializeCongruenceClasses(Function &F);
764   const Expression *makePossiblePHIOfOps(Instruction *,
765                                          SmallPtrSetImpl<Value *> &);
766   Value *findLeaderForInst(Instruction *ValueOp,
767                            SmallPtrSetImpl<Value *> &Visited,
768                            MemoryAccess *MemAccess, Instruction *OrigInst,
769                            BasicBlock *PredBB);
770   bool OpIsSafeForPHIOfOps(Value *Op, const BasicBlock *PHIBlock,
771                            SmallPtrSetImpl<const Value *> &);
772   void addPhiOfOps(PHINode *Op, BasicBlock *BB, Instruction *ExistingValue);
773   void removePhiOfOps(Instruction *I, PHINode *PHITemp);
774 
775   // Value number an Instruction or MemoryPhi.
776   void valueNumberMemoryPhi(MemoryPhi *);
777   void valueNumberInstruction(Instruction *);
778 
779   // Symbolic evaluation.
780   ExprResult checkExprResults(Expression *, Instruction *, Value *) const;
781   ExprResult performSymbolicEvaluation(Instruction *,
782                                        SmallPtrSetImpl<Value *> &) const;
783   const Expression *performSymbolicLoadCoercion(Type *, Value *, LoadInst *,
784                                                 Instruction *,
785                                                 MemoryAccess *) const;
786   const Expression *performSymbolicLoadEvaluation(Instruction *) const;
787   const Expression *performSymbolicStoreEvaluation(Instruction *) const;
788   ExprResult performSymbolicCallEvaluation(Instruction *) const;
789   void sortPHIOps(MutableArrayRef<ValPair> Ops) const;
790   const Expression *performSymbolicPHIEvaluation(ArrayRef<ValPair>,
791                                                  Instruction *I,
792                                                  BasicBlock *PHIBlock) const;
793   const Expression *performSymbolicAggrValueEvaluation(Instruction *) const;
794   ExprResult performSymbolicCmpEvaluation(Instruction *) const;
795   ExprResult performSymbolicPredicateInfoEvaluation(IntrinsicInst *) const;
796 
797   // Congruence finding.
798   bool someEquivalentDominates(const Instruction *, const Instruction *) const;
799   Value *lookupOperandLeader(Value *) const;
800   CongruenceClass *getClassForExpression(const Expression *E) const;
801   void performCongruenceFinding(Instruction *, const Expression *);
802   void moveValueToNewCongruenceClass(Instruction *, const Expression *,
803                                      CongruenceClass *, CongruenceClass *);
804   void moveMemoryToNewCongruenceClass(Instruction *, MemoryAccess *,
805                                       CongruenceClass *, CongruenceClass *);
806   Value *getNextValueLeader(CongruenceClass *) const;
807   const MemoryAccess *getNextMemoryLeader(CongruenceClass *) const;
808   bool setMemoryClass(const MemoryAccess *From, CongruenceClass *To);
809   CongruenceClass *getMemoryClass(const MemoryAccess *MA) const;
810   const MemoryAccess *lookupMemoryLeader(const MemoryAccess *) const;
811   bool isMemoryAccessTOP(const MemoryAccess *) const;
812 
813   // Ranking
814   unsigned int getRank(const Value *) const;
815   bool shouldSwapOperands(const Value *, const Value *) const;
816   bool shouldSwapOperandsForIntrinsic(const Value *, const Value *,
817                                       const IntrinsicInst *I) const;
818 
819   // Reachability handling.
820   void updateReachableEdge(BasicBlock *, BasicBlock *);
821   void processOutgoingEdges(Instruction *, BasicBlock *);
822   Value *findConditionEquivalence(Value *) const;
823 
824   // Elimination.
825   struct ValueDFS;
826   void convertClassToDFSOrdered(const CongruenceClass &,
827                                 SmallVectorImpl<ValueDFS> &,
828                                 DenseMap<const Value *, unsigned int> &,
829                                 SmallPtrSetImpl<Instruction *> &) const;
830   void convertClassToLoadsAndStores(const CongruenceClass &,
831                                     SmallVectorImpl<ValueDFS> &) const;
832 
833   bool eliminateInstructions(Function &);
834   void replaceInstruction(Instruction *, Value *);
835   void markInstructionForDeletion(Instruction *);
836   void deleteInstructionsInBlock(BasicBlock *);
837   Value *findPHIOfOpsLeader(const Expression *, const Instruction *,
838                             const BasicBlock *) const;
839 
840   // Various instruction touch utilities
841   template <typename Map, typename KeyType>
842   void touchAndErase(Map &, const KeyType &);
843   void markUsersTouched(Value *);
844   void markMemoryUsersTouched(const MemoryAccess *);
845   void markMemoryDefTouched(const MemoryAccess *);
846   void markPredicateUsersTouched(Instruction *);
847   void markValueLeaderChangeTouched(CongruenceClass *CC);
848   void markMemoryLeaderChangeTouched(CongruenceClass *CC);
849   void markPhiOfOpsChanged(const Expression *E);
850   void addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const;
851   void addAdditionalUsers(Value *To, Value *User) const;
852   void addAdditionalUsers(ExprResult &Res, Instruction *User) const;
853 
854   // Main loop of value numbering
855   void iterateTouchedInstructions();
856 
857   // Utilities.
858   void cleanupTables();
859   std::pair<unsigned, unsigned> assignDFSNumbers(BasicBlock *, unsigned);
860   void updateProcessedCount(const Value *V);
861   void verifyMemoryCongruency() const;
862   void verifyIterationSettled(Function &F);
863   void verifyStoreExpressions() const;
864   bool singleReachablePHIPath(SmallPtrSet<const MemoryAccess *, 8> &,
865                               const MemoryAccess *, const MemoryAccess *) const;
866   BasicBlock *getBlockForValue(Value *V) const;
867   void deleteExpression(const Expression *E) const;
868   MemoryUseOrDef *getMemoryAccess(const Instruction *) const;
869   MemoryPhi *getMemoryAccess(const BasicBlock *) const;
870   template <class T, class Range> T *getMinDFSOfRange(const Range &) const;
871 
InstrToDFSNum(const Value * V) const872   unsigned InstrToDFSNum(const Value *V) const {
873     assert(isa<Instruction>(V) && "This should not be used for MemoryAccesses");
874     return InstrDFS.lookup(V);
875   }
876 
InstrToDFSNum(const MemoryAccess * MA) const877   unsigned InstrToDFSNum(const MemoryAccess *MA) const {
878     return MemoryToDFSNum(MA);
879   }
880 
InstrFromDFSNum(unsigned DFSNum)881   Value *InstrFromDFSNum(unsigned DFSNum) { return DFSToInstr[DFSNum]; }
882 
883   // Given a MemoryAccess, return the relevant instruction DFS number.  Note:
884   // This deliberately takes a value so it can be used with Use's, which will
885   // auto-convert to Value's but not to MemoryAccess's.
MemoryToDFSNum(const Value * MA) const886   unsigned MemoryToDFSNum(const Value *MA) const {
887     assert(isa<MemoryAccess>(MA) &&
888            "This should not be used with instructions");
889     return isa<MemoryUseOrDef>(MA)
890                ? InstrToDFSNum(cast<MemoryUseOrDef>(MA)->getMemoryInst())
891                : InstrDFS.lookup(MA);
892   }
893 
894   bool isCycleFree(const Instruction *) const;
895   bool isBackedge(BasicBlock *From, BasicBlock *To) const;
896 
897   // Debug counter info.  When verifying, we have to reset the value numbering
898   // debug counter to the same state it started in to get the same results.
899   DebugCounter::CounterState StartingVNCounter;
900 };
901 
902 } // end anonymous namespace
903 
904 template <typename T>
equalsLoadStoreHelper(const T & LHS,const Expression & RHS)905 static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS) {
906   if (!isa<LoadExpression>(RHS) && !isa<StoreExpression>(RHS))
907     return false;
908   return LHS.MemoryExpression::equals(RHS);
909 }
910 
equals(const Expression & Other) const911 bool LoadExpression::equals(const Expression &Other) const {
912   return equalsLoadStoreHelper(*this, Other);
913 }
914 
equals(const Expression & Other) const915 bool StoreExpression::equals(const Expression &Other) const {
916   if (!equalsLoadStoreHelper(*this, Other))
917     return false;
918   // Make sure that store vs store includes the value operand.
919   if (const auto *S = dyn_cast<StoreExpression>(&Other))
920     if (getStoredValue() != S->getStoredValue())
921       return false;
922   return true;
923 }
924 
925 // Determine if the edge From->To is a backedge
isBackedge(BasicBlock * From,BasicBlock * To) const926 bool NewGVN::isBackedge(BasicBlock *From, BasicBlock *To) const {
927   return From == To ||
928          RPOOrdering.lookup(DT->getNode(From)) >=
929              RPOOrdering.lookup(DT->getNode(To));
930 }
931 
932 #ifndef NDEBUG
getBlockName(const BasicBlock * B)933 static std::string getBlockName(const BasicBlock *B) {
934   return DOTGraphTraits<DOTFuncInfo *>::getSimpleNodeLabel(B, nullptr);
935 }
936 #endif
937 
938 // Get a MemoryAccess for an instruction, fake or real.
getMemoryAccess(const Instruction * I) const939 MemoryUseOrDef *NewGVN::getMemoryAccess(const Instruction *I) const {
940   auto *Result = MSSA->getMemoryAccess(I);
941   return Result ? Result : TempToMemory.lookup(I);
942 }
943 
944 // Get a MemoryPhi for a basic block. These are all real.
getMemoryAccess(const BasicBlock * BB) const945 MemoryPhi *NewGVN::getMemoryAccess(const BasicBlock *BB) const {
946   return MSSA->getMemoryAccess(BB);
947 }
948 
949 // Get the basic block from an instruction/memory value.
getBlockForValue(Value * V) const950 BasicBlock *NewGVN::getBlockForValue(Value *V) const {
951   if (auto *I = dyn_cast<Instruction>(V)) {
952     auto *Parent = I->getParent();
953     if (Parent)
954       return Parent;
955     Parent = TempToBlock.lookup(V);
956     assert(Parent && "Every fake instruction should have a block");
957     return Parent;
958   }
959 
960   auto *MP = dyn_cast<MemoryPhi>(V);
961   assert(MP && "Should have been an instruction or a MemoryPhi");
962   return MP->getBlock();
963 }
964 
965 // Delete a definitely dead expression, so it can be reused by the expression
966 // allocator.  Some of these are not in creation functions, so we have to accept
967 // const versions.
deleteExpression(const Expression * E) const968 void NewGVN::deleteExpression(const Expression *E) const {
969   assert(isa<BasicExpression>(E));
970   auto *BE = cast<BasicExpression>(E);
971   const_cast<BasicExpression *>(BE)->deallocateOperands(ArgRecycler);
972   ExpressionAllocator.Deallocate(E);
973 }
974 
975 // If V is a predicateinfo copy, get the thing it is a copy of.
getCopyOf(const Value * V)976 static Value *getCopyOf(const Value *V) {
977   if (auto *II = dyn_cast<IntrinsicInst>(V))
978     if (II->getIntrinsicID() == Intrinsic::ssa_copy)
979       return II->getOperand(0);
980   return nullptr;
981 }
982 
983 // Return true if V is really PN, even accounting for predicateinfo copies.
isCopyOfPHI(const Value * V,const PHINode * PN)984 static bool isCopyOfPHI(const Value *V, const PHINode *PN) {
985   return V == PN || getCopyOf(V) == PN;
986 }
987 
isCopyOfAPHI(const Value * V)988 static bool isCopyOfAPHI(const Value *V) {
989   auto *CO = getCopyOf(V);
990   return CO && isa<PHINode>(CO);
991 }
992 
993 // Sort PHI Operands into a canonical order.  What we use here is an RPO
994 // order. The BlockInstRange numbers are generated in an RPO walk of the basic
995 // blocks.
sortPHIOps(MutableArrayRef<ValPair> Ops) const996 void NewGVN::sortPHIOps(MutableArrayRef<ValPair> Ops) const {
997   llvm::sort(Ops, [&](const ValPair &P1, const ValPair &P2) {
998     return BlockInstRange.lookup(P1.second).first <
999            BlockInstRange.lookup(P2.second).first;
1000   });
1001 }
1002 
1003 // Return true if V is a value that will always be available (IE can
1004 // be placed anywhere) in the function.  We don't do globals here
1005 // because they are often worse to put in place.
alwaysAvailable(Value * V)1006 static bool alwaysAvailable(Value *V) {
1007   return isa<Constant>(V) || isa<Argument>(V);
1008 }
1009 
1010 // Create a PHIExpression from an array of {incoming edge, value} pairs.  I is
1011 // the original instruction we are creating a PHIExpression for (but may not be
1012 // a phi node). We require, as an invariant, that all the PHIOperands in the
1013 // same block are sorted the same way. sortPHIOps will sort them into a
1014 // canonical order.
createPHIExpression(ArrayRef<ValPair> PHIOperands,const Instruction * I,BasicBlock * PHIBlock,bool & HasBackedge,bool & OriginalOpsConstant) const1015 PHIExpression *NewGVN::createPHIExpression(ArrayRef<ValPair> PHIOperands,
1016                                            const Instruction *I,
1017                                            BasicBlock *PHIBlock,
1018                                            bool &HasBackedge,
1019                                            bool &OriginalOpsConstant) const {
1020   unsigned NumOps = PHIOperands.size();
1021   auto *E = new (ExpressionAllocator) PHIExpression(NumOps, PHIBlock);
1022 
1023   E->allocateOperands(ArgRecycler, ExpressionAllocator);
1024   E->setType(PHIOperands.begin()->first->getType());
1025   E->setOpcode(Instruction::PHI);
1026 
1027   // Filter out unreachable phi operands.
1028   auto Filtered = make_filter_range(PHIOperands, [&](const ValPair &P) {
1029     auto *BB = P.second;
1030     if (auto *PHIOp = dyn_cast<PHINode>(I))
1031       if (isCopyOfPHI(P.first, PHIOp))
1032         return false;
1033     if (!ReachableEdges.count({BB, PHIBlock}))
1034       return false;
1035     // Things in TOPClass are equivalent to everything.
1036     if (ValueToClass.lookup(P.first) == TOPClass)
1037       return false;
1038     OriginalOpsConstant = OriginalOpsConstant && isa<Constant>(P.first);
1039     HasBackedge = HasBackedge || isBackedge(BB, PHIBlock);
1040     return lookupOperandLeader(P.first) != I;
1041   });
1042   std::transform(Filtered.begin(), Filtered.end(), op_inserter(E),
1043                  [&](const ValPair &P) -> Value * {
1044                    return lookupOperandLeader(P.first);
1045                  });
1046   return E;
1047 }
1048 
1049 // Set basic expression info (Arguments, type, opcode) for Expression
1050 // E from Instruction I in block B.
setBasicExpressionInfo(Instruction * I,BasicExpression * E) const1051 bool NewGVN::setBasicExpressionInfo(Instruction *I, BasicExpression *E) const {
1052   bool AllConstant = true;
1053   if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
1054     E->setType(GEP->getSourceElementType());
1055   else
1056     E->setType(I->getType());
1057   E->setOpcode(I->getOpcode());
1058   E->allocateOperands(ArgRecycler, ExpressionAllocator);
1059 
1060   // Transform the operand array into an operand leader array, and keep track of
1061   // whether all members are constant.
1062   std::transform(I->op_begin(), I->op_end(), op_inserter(E), [&](Value *O) {
1063     auto Operand = lookupOperandLeader(O);
1064     AllConstant = AllConstant && isa<Constant>(Operand);
1065     return Operand;
1066   });
1067 
1068   return AllConstant;
1069 }
1070 
createBinaryExpression(unsigned Opcode,Type * T,Value * Arg1,Value * Arg2,Instruction * I) const1071 const Expression *NewGVN::createBinaryExpression(unsigned Opcode, Type *T,
1072                                                  Value *Arg1, Value *Arg2,
1073                                                  Instruction *I) const {
1074   auto *E = new (ExpressionAllocator) BasicExpression(2);
1075   // TODO: we need to remove context instruction after Value Tracking
1076   // can run without context instruction
1077   const SimplifyQuery Q = SQ.getWithInstruction(I);
1078 
1079   E->setType(T);
1080   E->setOpcode(Opcode);
1081   E->allocateOperands(ArgRecycler, ExpressionAllocator);
1082   if (Instruction::isCommutative(Opcode)) {
1083     // Ensure that commutative instructions that only differ by a permutation
1084     // of their operands get the same value number by sorting the operand value
1085     // numbers.  Since all commutative instructions have two operands it is more
1086     // efficient to sort by hand rather than using, say, std::sort.
1087     if (shouldSwapOperands(Arg1, Arg2))
1088       std::swap(Arg1, Arg2);
1089   }
1090   E->op_push_back(lookupOperandLeader(Arg1));
1091   E->op_push_back(lookupOperandLeader(Arg2));
1092 
1093   Value *V = simplifyBinOp(Opcode, E->getOperand(0), E->getOperand(1), Q);
1094   if (auto Simplified = checkExprResults(E, I, V)) {
1095     addAdditionalUsers(Simplified, I);
1096     return Simplified.Expr;
1097   }
1098   return E;
1099 }
1100 
1101 // Take a Value returned by simplification of Expression E/Instruction
1102 // I, and see if it resulted in a simpler expression. If so, return
1103 // that expression.
checkExprResults(Expression * E,Instruction * I,Value * V) const1104 NewGVN::ExprResult NewGVN::checkExprResults(Expression *E, Instruction *I,
1105                                             Value *V) const {
1106   if (!V)
1107     return ExprResult::none();
1108 
1109   if (auto *C = dyn_cast<Constant>(V)) {
1110     if (I)
1111       LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1112                         << " constant " << *C << "\n");
1113     NumGVNOpsSimplified++;
1114     assert(isa<BasicExpression>(E) &&
1115            "We should always have had a basic expression here");
1116     deleteExpression(E);
1117     return ExprResult::some(createConstantExpression(C));
1118   } else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1119     if (I)
1120       LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1121                         << " variable " << *V << "\n");
1122     deleteExpression(E);
1123     return ExprResult::some(createVariableExpression(V));
1124   }
1125 
1126   CongruenceClass *CC = ValueToClass.lookup(V);
1127   if (CC) {
1128     if (CC->getLeader() && CC->getLeader() != I) {
1129       return ExprResult::some(createVariableOrConstant(CC->getLeader()), V);
1130     }
1131     if (CC->getDefiningExpr()) {
1132       if (I)
1133         LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1134                           << " expression " << *CC->getDefiningExpr() << "\n");
1135       NumGVNOpsSimplified++;
1136       deleteExpression(E);
1137       return ExprResult::some(CC->getDefiningExpr(), V);
1138     }
1139   }
1140 
1141   return ExprResult::none();
1142 }
1143 
1144 // Create a value expression from the instruction I, replacing operands with
1145 // their leaders.
1146 
createExpression(Instruction * I) const1147 NewGVN::ExprResult NewGVN::createExpression(Instruction *I) const {
1148   auto *E = new (ExpressionAllocator) BasicExpression(I->getNumOperands());
1149   // TODO: we need to remove context instruction after Value Tracking
1150   // can run without context instruction
1151   const SimplifyQuery Q = SQ.getWithInstruction(I);
1152 
1153   bool AllConstant = setBasicExpressionInfo(I, E);
1154 
1155   if (I->isCommutative()) {
1156     // Ensure that commutative instructions that only differ by a permutation
1157     // of their operands get the same value number by sorting the operand value
1158     // numbers.  Since all commutative instructions have two operands it is more
1159     // efficient to sort by hand rather than using, say, std::sort.
1160     assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
1161     if (shouldSwapOperands(E->getOperand(0), E->getOperand(1)))
1162       E->swapOperands(0, 1);
1163   }
1164   // Perform simplification.
1165   if (auto *CI = dyn_cast<CmpInst>(I)) {
1166     // Sort the operand value numbers so x<y and y>x get the same value
1167     // number.
1168     CmpInst::Predicate Predicate = CI->getPredicate();
1169     if (shouldSwapOperands(E->getOperand(0), E->getOperand(1))) {
1170       E->swapOperands(0, 1);
1171       Predicate = CmpInst::getSwappedPredicate(Predicate);
1172     }
1173     E->setOpcode((CI->getOpcode() << 8) | Predicate);
1174     // TODO: 25% of our time is spent in simplifyCmpInst with pointer operands
1175     assert(I->getOperand(0)->getType() == I->getOperand(1)->getType() &&
1176            "Wrong types on cmp instruction");
1177     assert((E->getOperand(0)->getType() == I->getOperand(0)->getType() &&
1178             E->getOperand(1)->getType() == I->getOperand(1)->getType()));
1179     Value *V =
1180         simplifyCmpInst(Predicate, E->getOperand(0), E->getOperand(1), Q);
1181     if (auto Simplified = checkExprResults(E, I, V))
1182       return Simplified;
1183   } else if (isa<SelectInst>(I)) {
1184     if (isa<Constant>(E->getOperand(0)) ||
1185         E->getOperand(1) == E->getOperand(2)) {
1186       assert(E->getOperand(1)->getType() == I->getOperand(1)->getType() &&
1187              E->getOperand(2)->getType() == I->getOperand(2)->getType());
1188       Value *V = simplifySelectInst(E->getOperand(0), E->getOperand(1),
1189                                     E->getOperand(2), Q);
1190       if (auto Simplified = checkExprResults(E, I, V))
1191         return Simplified;
1192     }
1193   } else if (I->isBinaryOp()) {
1194     Value *V =
1195         simplifyBinOp(E->getOpcode(), E->getOperand(0), E->getOperand(1), Q);
1196     if (auto Simplified = checkExprResults(E, I, V))
1197       return Simplified;
1198   } else if (auto *CI = dyn_cast<CastInst>(I)) {
1199     Value *V =
1200         simplifyCastInst(CI->getOpcode(), E->getOperand(0), CI->getType(), Q);
1201     if (auto Simplified = checkExprResults(E, I, V))
1202       return Simplified;
1203   } else if (auto *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1204     Value *V = simplifyGEPInst(GEPI->getSourceElementType(), *E->op_begin(),
1205                                ArrayRef(std::next(E->op_begin()), E->op_end()),
1206                                GEPI->getNoWrapFlags(), Q);
1207     if (auto Simplified = checkExprResults(E, I, V))
1208       return Simplified;
1209   } else if (AllConstant) {
1210     // We don't bother trying to simplify unless all of the operands
1211     // were constant.
1212     // TODO: There are a lot of Simplify*'s we could call here, if we
1213     // wanted to.  The original motivating case for this code was a
1214     // zext i1 false to i8, which we don't have an interface to
1215     // simplify (IE there is no SimplifyZExt).
1216 
1217     SmallVector<Constant *, 8> C;
1218     for (Value *Arg : E->operands())
1219       C.emplace_back(cast<Constant>(Arg));
1220 
1221     if (Value *V = ConstantFoldInstOperands(I, C, DL, TLI))
1222       if (auto Simplified = checkExprResults(E, I, V))
1223         return Simplified;
1224   }
1225   return ExprResult::some(E);
1226 }
1227 
1228 const AggregateValueExpression *
createAggregateValueExpression(Instruction * I) const1229 NewGVN::createAggregateValueExpression(Instruction *I) const {
1230   if (auto *II = dyn_cast<InsertValueInst>(I)) {
1231     auto *E = new (ExpressionAllocator)
1232         AggregateValueExpression(I->getNumOperands(), II->getNumIndices());
1233     setBasicExpressionInfo(I, E);
1234     E->allocateIntOperands(ExpressionAllocator);
1235     std::copy(II->idx_begin(), II->idx_end(), int_op_inserter(E));
1236     return E;
1237   } else if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1238     auto *E = new (ExpressionAllocator)
1239         AggregateValueExpression(I->getNumOperands(), EI->getNumIndices());
1240     setBasicExpressionInfo(EI, E);
1241     E->allocateIntOperands(ExpressionAllocator);
1242     std::copy(EI->idx_begin(), EI->idx_end(), int_op_inserter(E));
1243     return E;
1244   }
1245   llvm_unreachable("Unhandled type of aggregate value operation");
1246 }
1247 
createDeadExpression() const1248 const DeadExpression *NewGVN::createDeadExpression() const {
1249   // DeadExpression has no arguments and all DeadExpression's are the same,
1250   // so we only need one of them.
1251   return SingletonDeadExpression;
1252 }
1253 
createVariableExpression(Value * V) const1254 const VariableExpression *NewGVN::createVariableExpression(Value *V) const {
1255   auto *E = new (ExpressionAllocator) VariableExpression(V);
1256   E->setOpcode(V->getValueID());
1257   return E;
1258 }
1259 
createVariableOrConstant(Value * V) const1260 const Expression *NewGVN::createVariableOrConstant(Value *V) const {
1261   if (auto *C = dyn_cast<Constant>(V))
1262     return createConstantExpression(C);
1263   return createVariableExpression(V);
1264 }
1265 
createConstantExpression(Constant * C) const1266 const ConstantExpression *NewGVN::createConstantExpression(Constant *C) const {
1267   auto *E = new (ExpressionAllocator) ConstantExpression(C);
1268   E->setOpcode(C->getValueID());
1269   return E;
1270 }
1271 
createUnknownExpression(Instruction * I) const1272 const UnknownExpression *NewGVN::createUnknownExpression(Instruction *I) const {
1273   auto *E = new (ExpressionAllocator) UnknownExpression(I);
1274   E->setOpcode(I->getOpcode());
1275   return E;
1276 }
1277 
1278 const CallExpression *
createCallExpression(CallInst * CI,const MemoryAccess * MA) const1279 NewGVN::createCallExpression(CallInst *CI, const MemoryAccess *MA) const {
1280   // FIXME: Add operand bundles for calls.
1281   auto *E =
1282       new (ExpressionAllocator) CallExpression(CI->getNumOperands(), CI, MA);
1283   setBasicExpressionInfo(CI, E);
1284   if (CI->isCommutative()) {
1285     // Ensure that commutative intrinsics that only differ by a permutation
1286     // of their operands get the same value number by sorting the operand value
1287     // numbers.
1288     assert(CI->getNumOperands() >= 2 && "Unsupported commutative intrinsic!");
1289     if (shouldSwapOperands(E->getOperand(0), E->getOperand(1)))
1290       E->swapOperands(0, 1);
1291   }
1292   return E;
1293 }
1294 
1295 // Return true if some equivalent of instruction Inst dominates instruction U.
someEquivalentDominates(const Instruction * Inst,const Instruction * U) const1296 bool NewGVN::someEquivalentDominates(const Instruction *Inst,
1297                                      const Instruction *U) const {
1298   auto *CC = ValueToClass.lookup(Inst);
1299    // This must be an instruction because we are only called from phi nodes
1300   // in the case that the value it needs to check against is an instruction.
1301 
1302   // The most likely candidates for dominance are the leader and the next leader.
1303   // The leader or nextleader will dominate in all cases where there is an
1304   // equivalent that is higher up in the dom tree.
1305   // We can't *only* check them, however, because the
1306   // dominator tree could have an infinite number of non-dominating siblings
1307   // with instructions that are in the right congruence class.
1308   //       A
1309   // B C D E F G
1310   // |
1311   // H
1312   // Instruction U could be in H,  with equivalents in every other sibling.
1313   // Depending on the rpo order picked, the leader could be the equivalent in
1314   // any of these siblings.
1315   if (!CC)
1316     return false;
1317   if (alwaysAvailable(CC->getLeader()))
1318     return true;
1319   if (DT->dominates(cast<Instruction>(CC->getLeader()), U))
1320     return true;
1321   if (CC->getNextLeader().first &&
1322       DT->dominates(cast<Instruction>(CC->getNextLeader().first), U))
1323     return true;
1324   return llvm::any_of(*CC, [&](const Value *Member) {
1325     return Member != CC->getLeader() &&
1326            DT->dominates(cast<Instruction>(Member), U);
1327   });
1328 }
1329 
1330 // See if we have a congruence class and leader for this operand, and if so,
1331 // return it. Otherwise, return the operand itself.
lookupOperandLeader(Value * V) const1332 Value *NewGVN::lookupOperandLeader(Value *V) const {
1333   CongruenceClass *CC = ValueToClass.lookup(V);
1334   if (CC) {
1335     // Everything in TOP is represented by poison, as it can be any value.
1336     // We do have to make sure we get the type right though, so we can't set the
1337     // RepLeader to poison.
1338     if (CC == TOPClass)
1339       return PoisonValue::get(V->getType());
1340     return CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
1341   }
1342 
1343   return V;
1344 }
1345 
lookupMemoryLeader(const MemoryAccess * MA) const1346 const MemoryAccess *NewGVN::lookupMemoryLeader(const MemoryAccess *MA) const {
1347   auto *CC = getMemoryClass(MA);
1348   assert(CC->getMemoryLeader() &&
1349          "Every MemoryAccess should be mapped to a congruence class with a "
1350          "representative memory access");
1351   return CC->getMemoryLeader();
1352 }
1353 
1354 // Return true if the MemoryAccess is really equivalent to everything. This is
1355 // equivalent to the lattice value "TOP" in most lattices.  This is the initial
1356 // state of all MemoryAccesses.
isMemoryAccessTOP(const MemoryAccess * MA) const1357 bool NewGVN::isMemoryAccessTOP(const MemoryAccess *MA) const {
1358   return getMemoryClass(MA) == TOPClass;
1359 }
1360 
createLoadExpression(Type * LoadType,Value * PointerOp,LoadInst * LI,const MemoryAccess * MA) const1361 LoadExpression *NewGVN::createLoadExpression(Type *LoadType, Value *PointerOp,
1362                                              LoadInst *LI,
1363                                              const MemoryAccess *MA) const {
1364   auto *E =
1365       new (ExpressionAllocator) LoadExpression(1, LI, lookupMemoryLeader(MA));
1366   E->allocateOperands(ArgRecycler, ExpressionAllocator);
1367   E->setType(LoadType);
1368 
1369   // Give store and loads same opcode so they value number together.
1370   E->setOpcode(0);
1371   E->op_push_back(PointerOp);
1372 
1373   // TODO: Value number heap versions. We may be able to discover
1374   // things alias analysis can't on it's own (IE that a store and a
1375   // load have the same value, and thus, it isn't clobbering the load).
1376   return E;
1377 }
1378 
1379 const StoreExpression *
createStoreExpression(StoreInst * SI,const MemoryAccess * MA) const1380 NewGVN::createStoreExpression(StoreInst *SI, const MemoryAccess *MA) const {
1381   auto *StoredValueLeader = lookupOperandLeader(SI->getValueOperand());
1382   auto *E = new (ExpressionAllocator)
1383       StoreExpression(SI->getNumOperands(), SI, StoredValueLeader, MA);
1384   E->allocateOperands(ArgRecycler, ExpressionAllocator);
1385   E->setType(SI->getValueOperand()->getType());
1386 
1387   // Give store and loads same opcode so they value number together.
1388   E->setOpcode(0);
1389   E->op_push_back(lookupOperandLeader(SI->getPointerOperand()));
1390 
1391   // TODO: Value number heap versions. We may be able to discover
1392   // things alias analysis can't on it's own (IE that a store and a
1393   // load have the same value, and thus, it isn't clobbering the load).
1394   return E;
1395 }
1396 
performSymbolicStoreEvaluation(Instruction * I) const1397 const Expression *NewGVN::performSymbolicStoreEvaluation(Instruction *I) const {
1398   // Unlike loads, we never try to eliminate stores, so we do not check if they
1399   // are simple and avoid value numbering them.
1400   auto *SI = cast<StoreInst>(I);
1401   auto *StoreAccess = getMemoryAccess(SI);
1402   // Get the expression, if any, for the RHS of the MemoryDef.
1403   const MemoryAccess *StoreRHS = StoreAccess->getDefiningAccess();
1404   if (EnableStoreRefinement)
1405     StoreRHS = MSSAWalker->getClobberingMemoryAccess(StoreAccess);
1406   // If we bypassed the use-def chains, make sure we add a use.
1407   StoreRHS = lookupMemoryLeader(StoreRHS);
1408   if (StoreRHS != StoreAccess->getDefiningAccess())
1409     addMemoryUsers(StoreRHS, StoreAccess);
1410   // If we are defined by ourselves, use the live on entry def.
1411   if (StoreRHS == StoreAccess)
1412     StoreRHS = MSSA->getLiveOnEntryDef();
1413 
1414   if (SI->isSimple()) {
1415     // See if we are defined by a previous store expression, it already has a
1416     // value, and it's the same value as our current store. FIXME: Right now, we
1417     // only do this for simple stores, we should expand to cover memcpys, etc.
1418     const auto *LastStore = createStoreExpression(SI, StoreRHS);
1419     const auto *LastCC = ExpressionToClass.lookup(LastStore);
1420     // We really want to check whether the expression we matched was a store. No
1421     // easy way to do that. However, we can check that the class we found has a
1422     // store, which, assuming the value numbering state is not corrupt, is
1423     // sufficient, because we must also be equivalent to that store's expression
1424     // for it to be in the same class as the load.
1425     if (LastCC && LastCC->getStoredValue() == LastStore->getStoredValue())
1426       return LastStore;
1427     // Also check if our value operand is defined by a load of the same memory
1428     // location, and the memory state is the same as it was then (otherwise, it
1429     // could have been overwritten later. See test32 in
1430     // transforms/DeadStoreElimination/simple.ll).
1431     if (auto *LI = dyn_cast<LoadInst>(LastStore->getStoredValue()))
1432       if ((lookupOperandLeader(LI->getPointerOperand()) ==
1433            LastStore->getOperand(0)) &&
1434           (lookupMemoryLeader(getMemoryAccess(LI)->getDefiningAccess()) ==
1435            StoreRHS))
1436         return LastStore;
1437     deleteExpression(LastStore);
1438   }
1439 
1440   // If the store is not equivalent to anything, value number it as a store that
1441   // produces a unique memory state (instead of using it's MemoryUse, we use
1442   // it's MemoryDef).
1443   return createStoreExpression(SI, StoreAccess);
1444 }
1445 
1446 // See if we can extract the value of a loaded pointer from a load, a store, or
1447 // a memory instruction.
1448 const Expression *
performSymbolicLoadCoercion(Type * LoadType,Value * LoadPtr,LoadInst * LI,Instruction * DepInst,MemoryAccess * DefiningAccess) const1449 NewGVN::performSymbolicLoadCoercion(Type *LoadType, Value *LoadPtr,
1450                                     LoadInst *LI, Instruction *DepInst,
1451                                     MemoryAccess *DefiningAccess) const {
1452   assert((!LI || LI->isSimple()) && "Not a simple load");
1453   if (auto *DepSI = dyn_cast<StoreInst>(DepInst)) {
1454     // Can't forward from non-atomic to atomic without violating memory model.
1455     // Also don't need to coerce if they are the same type, we will just
1456     // propagate.
1457     if (LI->isAtomic() > DepSI->isAtomic() ||
1458         LoadType == DepSI->getValueOperand()->getType())
1459       return nullptr;
1460     int Offset = analyzeLoadFromClobberingStore(LoadType, LoadPtr, DepSI, DL);
1461     if (Offset >= 0) {
1462       if (auto *C = dyn_cast<Constant>(
1463               lookupOperandLeader(DepSI->getValueOperand()))) {
1464         if (Constant *Res = getConstantValueForLoad(C, Offset, LoadType, DL)) {
1465           LLVM_DEBUG(dbgs() << "Coercing load from store " << *DepSI
1466                             << " to constant " << *Res << "\n");
1467           return createConstantExpression(Res);
1468         }
1469       }
1470     }
1471   } else if (auto *DepLI = dyn_cast<LoadInst>(DepInst)) {
1472     // Can't forward from non-atomic to atomic without violating memory model.
1473     if (LI->isAtomic() > DepLI->isAtomic())
1474       return nullptr;
1475     int Offset = analyzeLoadFromClobberingLoad(LoadType, LoadPtr, DepLI, DL);
1476     if (Offset >= 0) {
1477       // We can coerce a constant load into a load.
1478       if (auto *C = dyn_cast<Constant>(lookupOperandLeader(DepLI)))
1479         if (auto *PossibleConstant =
1480                 getConstantValueForLoad(C, Offset, LoadType, DL)) {
1481           LLVM_DEBUG(dbgs() << "Coercing load from load " << *LI
1482                             << " to constant " << *PossibleConstant << "\n");
1483           return createConstantExpression(PossibleConstant);
1484         }
1485     }
1486   } else if (auto *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
1487     int Offset = analyzeLoadFromClobberingMemInst(LoadType, LoadPtr, DepMI, DL);
1488     if (Offset >= 0) {
1489       if (auto *PossibleConstant =
1490               getConstantMemInstValueForLoad(DepMI, Offset, LoadType, DL)) {
1491         LLVM_DEBUG(dbgs() << "Coercing load from meminst " << *DepMI
1492                           << " to constant " << *PossibleConstant << "\n");
1493         return createConstantExpression(PossibleConstant);
1494       }
1495     }
1496   }
1497 
1498   // All of the below are only true if the loaded pointer is produced
1499   // by the dependent instruction.
1500   if (LoadPtr != lookupOperandLeader(DepInst) &&
1501       !AA->isMustAlias(LoadPtr, DepInst))
1502     return nullptr;
1503   // If this load really doesn't depend on anything, then we must be loading an
1504   // undef value.  This can happen when loading for a fresh allocation with no
1505   // intervening stores, for example.  Note that this is only true in the case
1506   // that the result of the allocation is pointer equal to the load ptr.
1507   if (isa<AllocaInst>(DepInst)) {
1508     return createConstantExpression(UndefValue::get(LoadType));
1509   }
1510   // If this load occurs either right after a lifetime begin,
1511   // then the loaded value is undefined.
1512   else if (auto *II = dyn_cast<IntrinsicInst>(DepInst)) {
1513     if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1514       return createConstantExpression(UndefValue::get(LoadType));
1515   } else if (auto *InitVal =
1516                  getInitialValueOfAllocation(DepInst, TLI, LoadType))
1517       return createConstantExpression(InitVal);
1518 
1519   return nullptr;
1520 }
1521 
performSymbolicLoadEvaluation(Instruction * I) const1522 const Expression *NewGVN::performSymbolicLoadEvaluation(Instruction *I) const {
1523   auto *LI = cast<LoadInst>(I);
1524 
1525   // We can eliminate in favor of non-simple loads, but we won't be able to
1526   // eliminate the loads themselves.
1527   if (!LI->isSimple())
1528     return nullptr;
1529 
1530   Value *LoadAddressLeader = lookupOperandLeader(LI->getPointerOperand());
1531   // Load of undef is UB.
1532   if (isa<UndefValue>(LoadAddressLeader))
1533     return createConstantExpression(PoisonValue::get(LI->getType()));
1534   MemoryAccess *OriginalAccess = getMemoryAccess(I);
1535   MemoryAccess *DefiningAccess =
1536       MSSAWalker->getClobberingMemoryAccess(OriginalAccess);
1537 
1538   if (!MSSA->isLiveOnEntryDef(DefiningAccess)) {
1539     if (auto *MD = dyn_cast<MemoryDef>(DefiningAccess)) {
1540       Instruction *DefiningInst = MD->getMemoryInst();
1541       // If the defining instruction is not reachable, replace with poison.
1542       if (!ReachableBlocks.count(DefiningInst->getParent()))
1543         return createConstantExpression(PoisonValue::get(LI->getType()));
1544       // This will handle stores and memory insts.  We only do if it the
1545       // defining access has a different type, or it is a pointer produced by
1546       // certain memory operations that cause the memory to have a fixed value
1547       // (IE things like calloc).
1548       if (const auto *CoercionResult =
1549               performSymbolicLoadCoercion(LI->getType(), LoadAddressLeader, LI,
1550                                           DefiningInst, DefiningAccess))
1551         return CoercionResult;
1552     }
1553   }
1554 
1555   const auto *LE = createLoadExpression(LI->getType(), LoadAddressLeader, LI,
1556                                         DefiningAccess);
1557   // If our MemoryLeader is not our defining access, add a use to the
1558   // MemoryLeader, so that we get reprocessed when it changes.
1559   if (LE->getMemoryLeader() != DefiningAccess)
1560     addMemoryUsers(LE->getMemoryLeader(), OriginalAccess);
1561   return LE;
1562 }
1563 
1564 NewGVN::ExprResult
performSymbolicPredicateInfoEvaluation(IntrinsicInst * I) const1565 NewGVN::performSymbolicPredicateInfoEvaluation(IntrinsicInst *I) const {
1566   auto *PI = PredInfo->getPredicateInfoFor(I);
1567   if (!PI)
1568     return ExprResult::none();
1569 
1570   LLVM_DEBUG(dbgs() << "Found predicate info from instruction !\n");
1571 
1572   const std::optional<PredicateConstraint> &Constraint = PI->getConstraint();
1573   if (!Constraint)
1574     return ExprResult::none();
1575 
1576   CmpInst::Predicate Predicate = Constraint->Predicate;
1577   Value *CmpOp0 = I->getOperand(0);
1578   Value *CmpOp1 = Constraint->OtherOp;
1579 
1580   Value *FirstOp = lookupOperandLeader(CmpOp0);
1581   Value *SecondOp = lookupOperandLeader(CmpOp1);
1582   Value *AdditionallyUsedValue = CmpOp0;
1583 
1584   // Sort the ops.
1585   if (shouldSwapOperandsForIntrinsic(FirstOp, SecondOp, I)) {
1586     std::swap(FirstOp, SecondOp);
1587     Predicate = CmpInst::getSwappedPredicate(Predicate);
1588     AdditionallyUsedValue = CmpOp1;
1589   }
1590 
1591   if (Predicate == CmpInst::ICMP_EQ)
1592     return ExprResult::some(createVariableOrConstant(FirstOp),
1593                             AdditionallyUsedValue, PI);
1594 
1595   // Handle the special case of floating point.
1596   if (Predicate == CmpInst::FCMP_OEQ && isa<ConstantFP>(FirstOp) &&
1597       !cast<ConstantFP>(FirstOp)->isZero())
1598     return ExprResult::some(createConstantExpression(cast<Constant>(FirstOp)),
1599                             AdditionallyUsedValue, PI);
1600 
1601   return ExprResult::none();
1602 }
1603 
1604 // Evaluate read only and pure calls, and create an expression result.
performSymbolicCallEvaluation(Instruction * I) const1605 NewGVN::ExprResult NewGVN::performSymbolicCallEvaluation(Instruction *I) const {
1606   auto *CI = cast<CallInst>(I);
1607   if (auto *II = dyn_cast<IntrinsicInst>(I)) {
1608     // Intrinsics with the returned attribute are copies of arguments.
1609     if (auto *ReturnedValue = II->getReturnedArgOperand()) {
1610       if (II->getIntrinsicID() == Intrinsic::ssa_copy)
1611         if (auto Res = performSymbolicPredicateInfoEvaluation(II))
1612           return Res;
1613       return ExprResult::some(createVariableOrConstant(ReturnedValue));
1614     }
1615   }
1616 
1617   // FIXME: Currently the calls which may access the thread id may
1618   // be considered as not accessing the memory. But this is
1619   // problematic for coroutines, since coroutines may resume in a
1620   // different thread. So we disable the optimization here for the
1621   // correctness. However, it may block many other correct
1622   // optimizations. Revert this one when we detect the memory
1623   // accessing kind more precisely.
1624   if (CI->getFunction()->isPresplitCoroutine())
1625     return ExprResult::none();
1626 
1627   // Do not combine convergent calls since they implicitly depend on the set of
1628   // threads that is currently executing, and they might be in different basic
1629   // blocks.
1630   if (CI->isConvergent())
1631     return ExprResult::none();
1632 
1633   if (AA->doesNotAccessMemory(CI)) {
1634     return ExprResult::some(
1635         createCallExpression(CI, TOPClass->getMemoryLeader()));
1636   } else if (AA->onlyReadsMemory(CI)) {
1637     if (auto *MA = MSSA->getMemoryAccess(CI)) {
1638       auto *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(MA);
1639       return ExprResult::some(createCallExpression(CI, DefiningAccess));
1640     } else // MSSA determined that CI does not access memory.
1641       return ExprResult::some(
1642           createCallExpression(CI, TOPClass->getMemoryLeader()));
1643   }
1644   return ExprResult::none();
1645 }
1646 
1647 // Retrieve the memory class for a given MemoryAccess.
getMemoryClass(const MemoryAccess * MA) const1648 CongruenceClass *NewGVN::getMemoryClass(const MemoryAccess *MA) const {
1649   auto *Result = MemoryAccessToClass.lookup(MA);
1650   assert(Result && "Should have found memory class");
1651   return Result;
1652 }
1653 
1654 // Update the MemoryAccess equivalence table to say that From is equal to To,
1655 // and return true if this is different from what already existed in the table.
setMemoryClass(const MemoryAccess * From,CongruenceClass * NewClass)1656 bool NewGVN::setMemoryClass(const MemoryAccess *From,
1657                             CongruenceClass *NewClass) {
1658   assert(NewClass &&
1659          "Every MemoryAccess should be getting mapped to a non-null class");
1660   LLVM_DEBUG(dbgs() << "Setting " << *From);
1661   LLVM_DEBUG(dbgs() << " equivalent to congruence class ");
1662   LLVM_DEBUG(dbgs() << NewClass->getID()
1663                     << " with current MemoryAccess leader ");
1664   LLVM_DEBUG(dbgs() << *NewClass->getMemoryLeader() << "\n");
1665 
1666   auto LookupResult = MemoryAccessToClass.find(From);
1667   bool Changed = false;
1668   // If it's already in the table, see if the value changed.
1669   if (LookupResult != MemoryAccessToClass.end()) {
1670     auto *OldClass = LookupResult->second;
1671     if (OldClass != NewClass) {
1672       // If this is a phi, we have to handle memory member updates.
1673       if (auto *MP = dyn_cast<MemoryPhi>(From)) {
1674         OldClass->memory_erase(MP);
1675         NewClass->memory_insert(MP);
1676         // This may have killed the class if it had no non-memory members
1677         if (OldClass->getMemoryLeader() == From) {
1678           if (OldClass->definesNoMemory()) {
1679             OldClass->setMemoryLeader(nullptr);
1680           } else {
1681             OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
1682             LLVM_DEBUG(dbgs() << "Memory class leader change for class "
1683                               << OldClass->getID() << " to "
1684                               << *OldClass->getMemoryLeader()
1685                               << " due to removal of a memory member " << *From
1686                               << "\n");
1687             markMemoryLeaderChangeTouched(OldClass);
1688           }
1689         }
1690       }
1691       // It wasn't equivalent before, and now it is.
1692       LookupResult->second = NewClass;
1693       Changed = true;
1694     }
1695   }
1696 
1697   return Changed;
1698 }
1699 
1700 // Determine if a instruction is cycle-free.  That means the values in the
1701 // instruction don't depend on any expressions that can change value as a result
1702 // of the instruction.  For example, a non-cycle free instruction would be v =
1703 // phi(0, v+1).
isCycleFree(const Instruction * I) const1704 bool NewGVN::isCycleFree(const Instruction *I) const {
1705   // In order to compute cycle-freeness, we do SCC finding on the instruction,
1706   // and see what kind of SCC it ends up in.  If it is a singleton, it is
1707   // cycle-free.  If it is not in a singleton, it is only cycle free if the
1708   // other members are all phi nodes (as they do not compute anything, they are
1709   // copies).
1710   auto ICS = InstCycleState.lookup(I);
1711   if (ICS == ICS_Unknown) {
1712     SCCFinder.Start(I);
1713     auto &SCC = SCCFinder.getComponentFor(I);
1714     // It's cycle free if it's size 1 or the SCC is *only* phi nodes.
1715     if (SCC.size() == 1)
1716       InstCycleState.insert({I, ICS_CycleFree});
1717     else {
1718       bool AllPhis = llvm::all_of(SCC, [](const Value *V) {
1719         return isa<PHINode>(V) || isCopyOfAPHI(V);
1720       });
1721       ICS = AllPhis ? ICS_CycleFree : ICS_Cycle;
1722       for (const auto *Member : SCC)
1723         if (auto *MemberPhi = dyn_cast<PHINode>(Member))
1724           InstCycleState.insert({MemberPhi, ICS});
1725     }
1726   }
1727   if (ICS == ICS_Cycle)
1728     return false;
1729   return true;
1730 }
1731 
1732 // Evaluate PHI nodes symbolically and create an expression result.
1733 const Expression *
performSymbolicPHIEvaluation(ArrayRef<ValPair> PHIOps,Instruction * I,BasicBlock * PHIBlock) const1734 NewGVN::performSymbolicPHIEvaluation(ArrayRef<ValPair> PHIOps,
1735                                      Instruction *I,
1736                                      BasicBlock *PHIBlock) const {
1737   // True if one of the incoming phi edges is a backedge.
1738   bool HasBackedge = false;
1739   // All constant tracks the state of whether all the *original* phi operands
1740   // This is really shorthand for "this phi cannot cycle due to forward
1741   // change in value of the phi is guaranteed not to later change the value of
1742   // the phi. IE it can't be v = phi(undef, v+1)
1743   bool OriginalOpsConstant = true;
1744   auto *E = cast<PHIExpression>(createPHIExpression(
1745       PHIOps, I, PHIBlock, HasBackedge, OriginalOpsConstant));
1746   // We match the semantics of SimplifyPhiNode from InstructionSimplify here.
1747   // See if all arguments are the same.
1748   // We track if any were undef because they need special handling.
1749   bool HasUndef = false, HasPoison = false;
1750   auto Filtered = make_filter_range(E->operands(), [&](Value *Arg) {
1751     if (isa<PoisonValue>(Arg)) {
1752       HasPoison = true;
1753       return false;
1754     }
1755     if (isa<UndefValue>(Arg)) {
1756       HasUndef = true;
1757       return false;
1758     }
1759     return true;
1760   });
1761   // If we are left with no operands, it's dead.
1762   if (Filtered.empty()) {
1763     // If it has undef or poison at this point, it means there are no-non-undef
1764     // arguments, and thus, the value of the phi node must be undef.
1765     if (HasUndef) {
1766       LLVM_DEBUG(
1767           dbgs() << "PHI Node " << *I
1768                  << " has no non-undef arguments, valuing it as undef\n");
1769       return createConstantExpression(UndefValue::get(I->getType()));
1770     }
1771     if (HasPoison) {
1772       LLVM_DEBUG(
1773           dbgs() << "PHI Node " << *I
1774                  << " has no non-poison arguments, valuing it as poison\n");
1775       return createConstantExpression(PoisonValue::get(I->getType()));
1776     }
1777 
1778     LLVM_DEBUG(dbgs() << "No arguments of PHI node " << *I << " are live\n");
1779     deleteExpression(E);
1780     return createDeadExpression();
1781   }
1782   Value *AllSameValue = *(Filtered.begin());
1783   ++Filtered.begin();
1784   // Can't use std::equal here, sadly, because filter.begin moves.
1785   if (llvm::all_of(Filtered, [&](Value *Arg) { return Arg == AllSameValue; })) {
1786     // Can't fold phi(undef, X) -> X unless X can't be poison (thus X is undef
1787     // in the worst case).
1788     if (HasUndef && !isGuaranteedNotToBePoison(AllSameValue, AC, nullptr, DT))
1789       return E;
1790 
1791     // In LLVM's non-standard representation of phi nodes, it's possible to have
1792     // phi nodes with cycles (IE dependent on other phis that are .... dependent
1793     // on the original phi node), especially in weird CFG's where some arguments
1794     // are unreachable, or uninitialized along certain paths.  This can cause
1795     // infinite loops during evaluation. We work around this by not trying to
1796     // really evaluate them independently, but instead using a variable
1797     // expression to say if one is equivalent to the other.
1798     // We also special case undef/poison, so that if we have an undef, we can't
1799     // use the common value unless it dominates the phi block.
1800     if (HasPoison || HasUndef) {
1801       // If we have undef and at least one other value, this is really a
1802       // multivalued phi, and we need to know if it's cycle free in order to
1803       // evaluate whether we can ignore the undef.  The other parts of this are
1804       // just shortcuts.  If there is no backedge, or all operands are
1805       // constants, it also must be cycle free.
1806       if (HasBackedge && !OriginalOpsConstant &&
1807           !isa<UndefValue>(AllSameValue) && !isCycleFree(I))
1808         return E;
1809 
1810       // Only have to check for instructions
1811       if (auto *AllSameInst = dyn_cast<Instruction>(AllSameValue))
1812         if (!someEquivalentDominates(AllSameInst, I))
1813           return E;
1814     }
1815     // Can't simplify to something that comes later in the iteration.
1816     // Otherwise, when and if it changes congruence class, we will never catch
1817     // up. We will always be a class behind it.
1818     if (isa<Instruction>(AllSameValue) &&
1819         InstrToDFSNum(AllSameValue) > InstrToDFSNum(I))
1820       return E;
1821     NumGVNPhisAllSame++;
1822     LLVM_DEBUG(dbgs() << "Simplified PHI node " << *I << " to " << *AllSameValue
1823                       << "\n");
1824     deleteExpression(E);
1825     return createVariableOrConstant(AllSameValue);
1826   }
1827   return E;
1828 }
1829 
1830 const Expression *
performSymbolicAggrValueEvaluation(Instruction * I) const1831 NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) const {
1832   if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1833     auto *WO = dyn_cast<WithOverflowInst>(EI->getAggregateOperand());
1834     if (WO && EI->getNumIndices() == 1 && *EI->idx_begin() == 0)
1835       // EI is an extract from one of our with.overflow intrinsics. Synthesize
1836       // a semantically equivalent expression instead of an extract value
1837       // expression.
1838       return createBinaryExpression(WO->getBinaryOp(), EI->getType(),
1839                                     WO->getLHS(), WO->getRHS(), I);
1840   }
1841 
1842   return createAggregateValueExpression(I);
1843 }
1844 
performSymbolicCmpEvaluation(Instruction * I) const1845 NewGVN::ExprResult NewGVN::performSymbolicCmpEvaluation(Instruction *I) const {
1846   assert(isa<CmpInst>(I) && "Expected a cmp instruction.");
1847 
1848   auto *CI = cast<CmpInst>(I);
1849   // See if our operands are equal to those of a previous predicate, and if so,
1850   // if it implies true or false.
1851   auto Op0 = lookupOperandLeader(CI->getOperand(0));
1852   auto Op1 = lookupOperandLeader(CI->getOperand(1));
1853   auto OurPredicate = CI->getPredicate();
1854   if (shouldSwapOperands(Op0, Op1)) {
1855     std::swap(Op0, Op1);
1856     OurPredicate = CI->getSwappedPredicate();
1857   }
1858 
1859   // Avoid processing the same info twice.
1860   const PredicateBase *LastPredInfo = nullptr;
1861   // See if we know something about the comparison itself, like it is the target
1862   // of an assume.
1863   auto *CmpPI = PredInfo->getPredicateInfoFor(I);
1864   if (isa_and_nonnull<PredicateAssume>(CmpPI))
1865     return ExprResult::some(
1866         createConstantExpression(ConstantInt::getTrue(CI->getType())));
1867 
1868   if (Op0 == Op1) {
1869     // This condition does not depend on predicates, no need to add users
1870     if (CI->isTrueWhenEqual())
1871       return ExprResult::some(
1872           createConstantExpression(ConstantInt::getTrue(CI->getType())));
1873     else if (CI->isFalseWhenEqual())
1874       return ExprResult::some(
1875           createConstantExpression(ConstantInt::getFalse(CI->getType())));
1876   }
1877 
1878   // NOTE: Because we are comparing both operands here and below, and using
1879   // previous comparisons, we rely on fact that predicateinfo knows to mark
1880   // comparisons that use renamed operands as users of the earlier comparisons.
1881   // It is *not* enough to just mark predicateinfo renamed operands as users of
1882   // the earlier comparisons, because the *other* operand may have changed in a
1883   // previous iteration.
1884   // Example:
1885   // icmp slt %a, %b
1886   // %b.0 = ssa.copy(%b)
1887   // false branch:
1888   // icmp slt %c, %b.0
1889 
1890   // %c and %a may start out equal, and thus, the code below will say the second
1891   // %icmp is false.  c may become equal to something else, and in that case the
1892   // %second icmp *must* be reexamined, but would not if only the renamed
1893   // %operands are considered users of the icmp.
1894 
1895   // *Currently* we only check one level of comparisons back, and only mark one
1896   // level back as touched when changes happen.  If you modify this code to look
1897   // back farther through comparisons, you *must* mark the appropriate
1898   // comparisons as users in PredicateInfo.cpp, or you will cause bugs.  See if
1899   // we know something just from the operands themselves
1900 
1901   // See if our operands have predicate info, so that we may be able to derive
1902   // something from a previous comparison.
1903   for (const auto &Op : CI->operands()) {
1904     auto *PI = PredInfo->getPredicateInfoFor(Op);
1905     if (const auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI)) {
1906       if (PI == LastPredInfo)
1907         continue;
1908       LastPredInfo = PI;
1909       // In phi of ops cases, we may have predicate info that we are evaluating
1910       // in a different context.
1911       if (!DT->dominates(PBranch->To, I->getParent()))
1912         continue;
1913       // TODO: Along the false edge, we may know more things too, like
1914       // icmp of
1915       // same operands is false.
1916       // TODO: We only handle actual comparison conditions below, not
1917       // and/or.
1918       auto *BranchCond = dyn_cast<CmpInst>(PBranch->Condition);
1919       if (!BranchCond)
1920         continue;
1921       auto *BranchOp0 = lookupOperandLeader(BranchCond->getOperand(0));
1922       auto *BranchOp1 = lookupOperandLeader(BranchCond->getOperand(1));
1923       auto BranchPredicate = BranchCond->getPredicate();
1924       if (shouldSwapOperands(BranchOp0, BranchOp1)) {
1925         std::swap(BranchOp0, BranchOp1);
1926         BranchPredicate = BranchCond->getSwappedPredicate();
1927       }
1928       if (BranchOp0 == Op0 && BranchOp1 == Op1) {
1929         if (PBranch->TrueEdge) {
1930           // If we know the previous predicate is true and we are in the true
1931           // edge then we may be implied true or false.
1932           if (CmpInst::isImpliedTrueByMatchingCmp(BranchPredicate,
1933                                                   OurPredicate)) {
1934             return ExprResult::some(
1935                 createConstantExpression(ConstantInt::getTrue(CI->getType())),
1936                 PI);
1937           }
1938 
1939           if (CmpInst::isImpliedFalseByMatchingCmp(BranchPredicate,
1940                                                    OurPredicate)) {
1941             return ExprResult::some(
1942                 createConstantExpression(ConstantInt::getFalse(CI->getType())),
1943                 PI);
1944           }
1945         } else {
1946           // Just handle the ne and eq cases, where if we have the same
1947           // operands, we may know something.
1948           if (BranchPredicate == OurPredicate) {
1949             // Same predicate, same ops,we know it was false, so this is false.
1950             return ExprResult::some(
1951                 createConstantExpression(ConstantInt::getFalse(CI->getType())),
1952                 PI);
1953           } else if (BranchPredicate ==
1954                      CmpInst::getInversePredicate(OurPredicate)) {
1955             // Inverse predicate, we know the other was false, so this is true.
1956             return ExprResult::some(
1957                 createConstantExpression(ConstantInt::getTrue(CI->getType())),
1958                 PI);
1959           }
1960         }
1961       }
1962     }
1963   }
1964   // Create expression will take care of simplifyCmpInst
1965   return createExpression(I);
1966 }
1967 
1968 // Substitute and symbolize the instruction before value numbering.
1969 NewGVN::ExprResult
performSymbolicEvaluation(Instruction * I,SmallPtrSetImpl<Value * > & Visited) const1970 NewGVN::performSymbolicEvaluation(Instruction *I,
1971                                   SmallPtrSetImpl<Value *> &Visited) const {
1972 
1973   const Expression *E = nullptr;
1974   // TODO: memory intrinsics.
1975   // TODO: Some day, we should do the forward propagation and reassociation
1976   // parts of the algorithm.
1977   switch (I->getOpcode()) {
1978   case Instruction::ExtractValue:
1979   case Instruction::InsertValue:
1980     E = performSymbolicAggrValueEvaluation(I);
1981     break;
1982   case Instruction::PHI: {
1983     SmallVector<ValPair, 3> Ops;
1984     auto *PN = cast<PHINode>(I);
1985     for (unsigned i = 0; i < PN->getNumOperands(); ++i)
1986       Ops.push_back({PN->getIncomingValue(i), PN->getIncomingBlock(i)});
1987     // Sort to ensure the invariant createPHIExpression requires is met.
1988     sortPHIOps(Ops);
1989     E = performSymbolicPHIEvaluation(Ops, I, getBlockForValue(I));
1990   } break;
1991   case Instruction::Call:
1992     return performSymbolicCallEvaluation(I);
1993     break;
1994   case Instruction::Store:
1995     E = performSymbolicStoreEvaluation(I);
1996     break;
1997   case Instruction::Load:
1998     E = performSymbolicLoadEvaluation(I);
1999     break;
2000   case Instruction::BitCast:
2001   case Instruction::AddrSpaceCast:
2002   case Instruction::Freeze:
2003     return createExpression(I);
2004     break;
2005   case Instruction::ICmp:
2006   case Instruction::FCmp:
2007     return performSymbolicCmpEvaluation(I);
2008     break;
2009   case Instruction::FNeg:
2010   case Instruction::Add:
2011   case Instruction::FAdd:
2012   case Instruction::Sub:
2013   case Instruction::FSub:
2014   case Instruction::Mul:
2015   case Instruction::FMul:
2016   case Instruction::UDiv:
2017   case Instruction::SDiv:
2018   case Instruction::FDiv:
2019   case Instruction::URem:
2020   case Instruction::SRem:
2021   case Instruction::FRem:
2022   case Instruction::Shl:
2023   case Instruction::LShr:
2024   case Instruction::AShr:
2025   case Instruction::And:
2026   case Instruction::Or:
2027   case Instruction::Xor:
2028   case Instruction::Trunc:
2029   case Instruction::ZExt:
2030   case Instruction::SExt:
2031   case Instruction::FPToUI:
2032   case Instruction::FPToSI:
2033   case Instruction::UIToFP:
2034   case Instruction::SIToFP:
2035   case Instruction::FPTrunc:
2036   case Instruction::FPExt:
2037   case Instruction::PtrToInt:
2038   case Instruction::IntToPtr:
2039   case Instruction::Select:
2040   case Instruction::ExtractElement:
2041   case Instruction::InsertElement:
2042   case Instruction::GetElementPtr:
2043     return createExpression(I);
2044     break;
2045   case Instruction::ShuffleVector:
2046     // FIXME: Add support for shufflevector to createExpression.
2047     return ExprResult::none();
2048   default:
2049     return ExprResult::none();
2050   }
2051   return ExprResult::some(E);
2052 }
2053 
2054 // Look up a container of values/instructions in a map, and touch all the
2055 // instructions in the container.  Then erase value from the map.
2056 template <typename Map, typename KeyType>
touchAndErase(Map & M,const KeyType & Key)2057 void NewGVN::touchAndErase(Map &M, const KeyType &Key) {
2058   const auto Result = M.find_as(Key);
2059   if (Result != M.end()) {
2060     for (const typename Map::mapped_type::value_type Mapped : Result->second)
2061       TouchedInstructions.set(InstrToDFSNum(Mapped));
2062     M.erase(Result);
2063   }
2064 }
2065 
addAdditionalUsers(Value * To,Value * User) const2066 void NewGVN::addAdditionalUsers(Value *To, Value *User) const {
2067   assert(User && To != User);
2068   if (isa<Instruction>(To))
2069     AdditionalUsers[To].insert(User);
2070 }
2071 
addAdditionalUsers(ExprResult & Res,Instruction * User) const2072 void NewGVN::addAdditionalUsers(ExprResult &Res, Instruction *User) const {
2073   if (Res.ExtraDep && Res.ExtraDep != User)
2074     addAdditionalUsers(Res.ExtraDep, User);
2075   Res.ExtraDep = nullptr;
2076 
2077   if (Res.PredDep) {
2078     if (const auto *PBranch = dyn_cast<PredicateBranch>(Res.PredDep))
2079       PredicateToUsers[PBranch->Condition].insert(User);
2080     else if (const auto *PAssume = dyn_cast<PredicateAssume>(Res.PredDep))
2081       PredicateToUsers[PAssume->Condition].insert(User);
2082   }
2083   Res.PredDep = nullptr;
2084 }
2085 
markUsersTouched(Value * V)2086 void NewGVN::markUsersTouched(Value *V) {
2087   // Now mark the users as touched.
2088   for (auto *User : V->users()) {
2089     assert(isa<Instruction>(User) && "Use of value not within an instruction?");
2090     TouchedInstructions.set(InstrToDFSNum(User));
2091   }
2092   touchAndErase(AdditionalUsers, V);
2093 }
2094 
addMemoryUsers(const MemoryAccess * To,MemoryAccess * U) const2095 void NewGVN::addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const {
2096   LLVM_DEBUG(dbgs() << "Adding memory user " << *U << " to " << *To << "\n");
2097   MemoryToUsers[To].insert(U);
2098 }
2099 
markMemoryDefTouched(const MemoryAccess * MA)2100 void NewGVN::markMemoryDefTouched(const MemoryAccess *MA) {
2101   TouchedInstructions.set(MemoryToDFSNum(MA));
2102 }
2103 
markMemoryUsersTouched(const MemoryAccess * MA)2104 void NewGVN::markMemoryUsersTouched(const MemoryAccess *MA) {
2105   if (isa<MemoryUse>(MA))
2106     return;
2107   for (const auto *U : MA->users())
2108     TouchedInstructions.set(MemoryToDFSNum(U));
2109   touchAndErase(MemoryToUsers, MA);
2110 }
2111 
2112 // Touch all the predicates that depend on this instruction.
markPredicateUsersTouched(Instruction * I)2113 void NewGVN::markPredicateUsersTouched(Instruction *I) {
2114   touchAndErase(PredicateToUsers, I);
2115 }
2116 
2117 // Mark users affected by a memory leader change.
markMemoryLeaderChangeTouched(CongruenceClass * CC)2118 void NewGVN::markMemoryLeaderChangeTouched(CongruenceClass *CC) {
2119   for (const auto *M : CC->memory())
2120     markMemoryDefTouched(M);
2121 }
2122 
2123 // Touch the instructions that need to be updated after a congruence class has a
2124 // leader change, and mark changed values.
markValueLeaderChangeTouched(CongruenceClass * CC)2125 void NewGVN::markValueLeaderChangeTouched(CongruenceClass *CC) {
2126   for (auto *M : *CC) {
2127     if (auto *I = dyn_cast<Instruction>(M))
2128       TouchedInstructions.set(InstrToDFSNum(I));
2129     LeaderChanges.insert(M);
2130   }
2131 }
2132 
2133 // Give a range of things that have instruction DFS numbers, this will return
2134 // the member of the range with the smallest dfs number.
2135 template <class T, class Range>
getMinDFSOfRange(const Range & R) const2136 T *NewGVN::getMinDFSOfRange(const Range &R) const {
2137   std::pair<T *, unsigned> MinDFS = {nullptr, ~0U};
2138   for (const auto X : R) {
2139     auto DFSNum = InstrToDFSNum(X);
2140     if (DFSNum < MinDFS.second)
2141       MinDFS = {X, DFSNum};
2142   }
2143   return MinDFS.first;
2144 }
2145 
2146 // This function returns the MemoryAccess that should be the next leader of
2147 // congruence class CC, under the assumption that the current leader is going to
2148 // disappear.
getNextMemoryLeader(CongruenceClass * CC) const2149 const MemoryAccess *NewGVN::getNextMemoryLeader(CongruenceClass *CC) const {
2150   // TODO: If this ends up to slow, we can maintain a next memory leader like we
2151   // do for regular leaders.
2152   // Make sure there will be a leader to find.
2153   assert(!CC->definesNoMemory() && "Can't get next leader if there is none");
2154   if (CC->getStoreCount() > 0) {
2155     if (auto *NL = dyn_cast_or_null<StoreInst>(CC->getNextLeader().first))
2156       return getMemoryAccess(NL);
2157     // Find the store with the minimum DFS number.
2158     auto *V = getMinDFSOfRange<Value>(make_filter_range(
2159         *CC, [&](const Value *V) { return isa<StoreInst>(V); }));
2160     return getMemoryAccess(cast<StoreInst>(V));
2161   }
2162   assert(CC->getStoreCount() == 0);
2163 
2164   // Given our assertion, hitting this part must mean
2165   // !OldClass->memory_empty()
2166   if (CC->memory_size() == 1)
2167     return *CC->memory_begin();
2168   return getMinDFSOfRange<const MemoryPhi>(CC->memory());
2169 }
2170 
2171 // This function returns the next value leader of a congruence class, under the
2172 // assumption that the current leader is going away.  This should end up being
2173 // the next most dominating member.
getNextValueLeader(CongruenceClass * CC) const2174 Value *NewGVN::getNextValueLeader(CongruenceClass *CC) const {
2175   // We don't need to sort members if there is only 1, and we don't care about
2176   // sorting the TOP class because everything either gets out of it or is
2177   // unreachable.
2178 
2179   if (CC->size() == 1 || CC == TOPClass) {
2180     return *(CC->begin());
2181   } else if (CC->getNextLeader().first) {
2182     ++NumGVNAvoidedSortedLeaderChanges;
2183     return CC->getNextLeader().first;
2184   } else {
2185     ++NumGVNSortedLeaderChanges;
2186     // NOTE: If this ends up to slow, we can maintain a dual structure for
2187     // member testing/insertion, or keep things mostly sorted, and sort only
2188     // here, or use SparseBitVector or ....
2189     return getMinDFSOfRange<Value>(*CC);
2190   }
2191 }
2192 
2193 // Move a MemoryAccess, currently in OldClass, to NewClass, including updates to
2194 // the memory members, etc for the move.
2195 //
2196 // The invariants of this function are:
2197 //
2198 // - I must be moving to NewClass from OldClass
2199 // - The StoreCount of OldClass and NewClass is expected to have been updated
2200 //   for I already if it is a store.
2201 // - The OldClass memory leader has not been updated yet if I was the leader.
moveMemoryToNewCongruenceClass(Instruction * I,MemoryAccess * InstMA,CongruenceClass * OldClass,CongruenceClass * NewClass)2202 void NewGVN::moveMemoryToNewCongruenceClass(Instruction *I,
2203                                             MemoryAccess *InstMA,
2204                                             CongruenceClass *OldClass,
2205                                             CongruenceClass *NewClass) {
2206   // If the leader is I, and we had a representative MemoryAccess, it should
2207   // be the MemoryAccess of OldClass.
2208   assert((!InstMA || !OldClass->getMemoryLeader() ||
2209           OldClass->getLeader() != I ||
2210           MemoryAccessToClass.lookup(OldClass->getMemoryLeader()) ==
2211               MemoryAccessToClass.lookup(InstMA)) &&
2212          "Representative MemoryAccess mismatch");
2213   // First, see what happens to the new class
2214   if (!NewClass->getMemoryLeader()) {
2215     // Should be a new class, or a store becoming a leader of a new class.
2216     assert(NewClass->size() == 1 ||
2217            (isa<StoreInst>(I) && NewClass->getStoreCount() == 1));
2218     NewClass->setMemoryLeader(InstMA);
2219     // Mark it touched if we didn't just create a singleton
2220     LLVM_DEBUG(dbgs() << "Memory class leader change for class "
2221                       << NewClass->getID()
2222                       << " due to new memory instruction becoming leader\n");
2223     markMemoryLeaderChangeTouched(NewClass);
2224   }
2225   setMemoryClass(InstMA, NewClass);
2226   // Now, fixup the old class if necessary
2227   if (OldClass->getMemoryLeader() == InstMA) {
2228     if (!OldClass->definesNoMemory()) {
2229       OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
2230       LLVM_DEBUG(dbgs() << "Memory class leader change for class "
2231                         << OldClass->getID() << " to "
2232                         << *OldClass->getMemoryLeader()
2233                         << " due to removal of old leader " << *InstMA << "\n");
2234       markMemoryLeaderChangeTouched(OldClass);
2235     } else
2236       OldClass->setMemoryLeader(nullptr);
2237   }
2238 }
2239 
2240 // Move a value, currently in OldClass, to be part of NewClass
2241 // Update OldClass and NewClass for the move (including changing leaders, etc).
moveValueToNewCongruenceClass(Instruction * I,const Expression * E,CongruenceClass * OldClass,CongruenceClass * NewClass)2242 void NewGVN::moveValueToNewCongruenceClass(Instruction *I, const Expression *E,
2243                                            CongruenceClass *OldClass,
2244                                            CongruenceClass *NewClass) {
2245   if (I == OldClass->getNextLeader().first)
2246     OldClass->resetNextLeader();
2247 
2248   OldClass->erase(I);
2249   NewClass->insert(I);
2250 
2251   if (NewClass->getLeader() != I)
2252     NewClass->addPossibleNextLeader({I, InstrToDFSNum(I)});
2253   // Handle our special casing of stores.
2254   if (auto *SI = dyn_cast<StoreInst>(I)) {
2255     OldClass->decStoreCount();
2256     // Okay, so when do we want to make a store a leader of a class?
2257     // If we have a store defined by an earlier load, we want the earlier load
2258     // to lead the class.
2259     // If we have a store defined by something else, we want the store to lead
2260     // the class so everything else gets the "something else" as a value.
2261     // If we have a store as the single member of the class, we want the store
2262     // as the leader
2263     if (NewClass->getStoreCount() == 0 && !NewClass->getStoredValue()) {
2264       // If it's a store expression we are using, it means we are not equivalent
2265       // to something earlier.
2266       if (auto *SE = dyn_cast<StoreExpression>(E)) {
2267         NewClass->setStoredValue(SE->getStoredValue());
2268         markValueLeaderChangeTouched(NewClass);
2269         // Shift the new class leader to be the store
2270         LLVM_DEBUG(dbgs() << "Changing leader of congruence class "
2271                           << NewClass->getID() << " from "
2272                           << *NewClass->getLeader() << " to  " << *SI
2273                           << " because store joined class\n");
2274         // If we changed the leader, we have to mark it changed because we don't
2275         // know what it will do to symbolic evaluation.
2276         NewClass->setLeader(SI);
2277       }
2278       // We rely on the code below handling the MemoryAccess change.
2279     }
2280     NewClass->incStoreCount();
2281   }
2282   // True if there is no memory instructions left in a class that had memory
2283   // instructions before.
2284 
2285   // If it's not a memory use, set the MemoryAccess equivalence
2286   auto *InstMA = dyn_cast_or_null<MemoryDef>(getMemoryAccess(I));
2287   if (InstMA)
2288     moveMemoryToNewCongruenceClass(I, InstMA, OldClass, NewClass);
2289   ValueToClass[I] = NewClass;
2290   // See if we destroyed the class or need to swap leaders.
2291   if (OldClass->empty() && OldClass != TOPClass) {
2292     if (OldClass->getDefiningExpr()) {
2293       LLVM_DEBUG(dbgs() << "Erasing expression " << *OldClass->getDefiningExpr()
2294                         << " from table\n");
2295       // We erase it as an exact expression to make sure we don't just erase an
2296       // equivalent one.
2297       auto Iter = ExpressionToClass.find_as(
2298           ExactEqualsExpression(*OldClass->getDefiningExpr()));
2299       if (Iter != ExpressionToClass.end())
2300         ExpressionToClass.erase(Iter);
2301 #ifdef EXPENSIVE_CHECKS
2302       assert(
2303           (*OldClass->getDefiningExpr() != *E || ExpressionToClass.lookup(E)) &&
2304           "We erased the expression we just inserted, which should not happen");
2305 #endif
2306     }
2307   } else if (OldClass->getLeader() == I) {
2308     // When the leader changes, the value numbering of
2309     // everything may change due to symbolization changes, so we need to
2310     // reprocess.
2311     LLVM_DEBUG(dbgs() << "Value class leader change for class "
2312                       << OldClass->getID() << "\n");
2313     ++NumGVNLeaderChanges;
2314     // Destroy the stored value if there are no more stores to represent it.
2315     // Note that this is basically clean up for the expression removal that
2316     // happens below.  If we remove stores from a class, we may leave it as a
2317     // class of equivalent memory phis.
2318     if (OldClass->getStoreCount() == 0) {
2319       if (OldClass->getStoredValue())
2320         OldClass->setStoredValue(nullptr);
2321     }
2322     OldClass->setLeader(getNextValueLeader(OldClass));
2323     OldClass->resetNextLeader();
2324     markValueLeaderChangeTouched(OldClass);
2325   }
2326 }
2327 
2328 // For a given expression, mark the phi of ops instructions that could have
2329 // changed as a result.
markPhiOfOpsChanged(const Expression * E)2330 void NewGVN::markPhiOfOpsChanged(const Expression *E) {
2331   touchAndErase(ExpressionToPhiOfOps, E);
2332 }
2333 
2334 // Perform congruence finding on a given value numbering expression.
performCongruenceFinding(Instruction * I,const Expression * E)2335 void NewGVN::performCongruenceFinding(Instruction *I, const Expression *E) {
2336   // This is guaranteed to return something, since it will at least find
2337   // TOP.
2338 
2339   CongruenceClass *IClass = ValueToClass.lookup(I);
2340   assert(IClass && "Should have found a IClass");
2341   // Dead classes should have been eliminated from the mapping.
2342   assert(!IClass->isDead() && "Found a dead class");
2343 
2344   CongruenceClass *EClass = nullptr;
2345   if (const auto *VE = dyn_cast<VariableExpression>(E)) {
2346     EClass = ValueToClass.lookup(VE->getVariableValue());
2347   } else if (isa<DeadExpression>(E)) {
2348     EClass = TOPClass;
2349   }
2350   if (!EClass) {
2351     auto lookupResult = ExpressionToClass.insert({E, nullptr});
2352 
2353     // If it's not in the value table, create a new congruence class.
2354     if (lookupResult.second) {
2355       CongruenceClass *NewClass = createCongruenceClass(nullptr, E);
2356       auto place = lookupResult.first;
2357       place->second = NewClass;
2358 
2359       // Constants and variables should always be made the leader.
2360       if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2361         NewClass->setLeader(CE->getConstantValue());
2362       } else if (const auto *SE = dyn_cast<StoreExpression>(E)) {
2363         StoreInst *SI = SE->getStoreInst();
2364         NewClass->setLeader(SI);
2365         NewClass->setStoredValue(SE->getStoredValue());
2366         // The RepMemoryAccess field will be filled in properly by the
2367         // moveValueToNewCongruenceClass call.
2368       } else {
2369         NewClass->setLeader(I);
2370       }
2371       assert(!isa<VariableExpression>(E) &&
2372              "VariableExpression should have been handled already");
2373 
2374       EClass = NewClass;
2375       LLVM_DEBUG(dbgs() << "Created new congruence class for " << *I
2376                         << " using expression " << *E << " at "
2377                         << NewClass->getID() << " and leader "
2378                         << *(NewClass->getLeader()));
2379       if (NewClass->getStoredValue())
2380         LLVM_DEBUG(dbgs() << " and stored value "
2381                           << *(NewClass->getStoredValue()));
2382       LLVM_DEBUG(dbgs() << "\n");
2383     } else {
2384       EClass = lookupResult.first->second;
2385       if (isa<ConstantExpression>(E))
2386         assert((isa<Constant>(EClass->getLeader()) ||
2387                 (EClass->getStoredValue() &&
2388                  isa<Constant>(EClass->getStoredValue()))) &&
2389                "Any class with a constant expression should have a "
2390                "constant leader");
2391 
2392       assert(EClass && "Somehow don't have an eclass");
2393 
2394       assert(!EClass->isDead() && "We accidentally looked up a dead class");
2395     }
2396   }
2397   bool ClassChanged = IClass != EClass;
2398   bool LeaderChanged = LeaderChanges.erase(I);
2399   if (ClassChanged || LeaderChanged) {
2400     LLVM_DEBUG(dbgs() << "New class " << EClass->getID() << " for expression "
2401                       << *E << "\n");
2402     if (ClassChanged) {
2403       moveValueToNewCongruenceClass(I, E, IClass, EClass);
2404       markPhiOfOpsChanged(E);
2405     }
2406 
2407     markUsersTouched(I);
2408     if (MemoryAccess *MA = getMemoryAccess(I))
2409       markMemoryUsersTouched(MA);
2410     if (auto *CI = dyn_cast<CmpInst>(I))
2411       markPredicateUsersTouched(CI);
2412   }
2413   // If we changed the class of the store, we want to ensure nothing finds the
2414   // old store expression.  In particular, loads do not compare against stored
2415   // value, so they will find old store expressions (and associated class
2416   // mappings) if we leave them in the table.
2417   if (ClassChanged && isa<StoreInst>(I)) {
2418     auto *OldE = ValueToExpression.lookup(I);
2419     // It could just be that the old class died. We don't want to erase it if we
2420     // just moved classes.
2421     if (OldE && isa<StoreExpression>(OldE) && *E != *OldE) {
2422       // Erase this as an exact expression to ensure we don't erase expressions
2423       // equivalent to it.
2424       auto Iter = ExpressionToClass.find_as(ExactEqualsExpression(*OldE));
2425       if (Iter != ExpressionToClass.end())
2426         ExpressionToClass.erase(Iter);
2427     }
2428   }
2429   ValueToExpression[I] = E;
2430 }
2431 
2432 // Process the fact that Edge (from, to) is reachable, including marking
2433 // any newly reachable blocks and instructions for processing.
updateReachableEdge(BasicBlock * From,BasicBlock * To)2434 void NewGVN::updateReachableEdge(BasicBlock *From, BasicBlock *To) {
2435   // Check if the Edge was reachable before.
2436   if (ReachableEdges.insert({From, To}).second) {
2437     // If this block wasn't reachable before, all instructions are touched.
2438     if (ReachableBlocks.insert(To).second) {
2439       LLVM_DEBUG(dbgs() << "Block " << getBlockName(To)
2440                         << " marked reachable\n");
2441       const auto &InstRange = BlockInstRange.lookup(To);
2442       TouchedInstructions.set(InstRange.first, InstRange.second);
2443     } else {
2444       LLVM_DEBUG(dbgs() << "Block " << getBlockName(To)
2445                         << " was reachable, but new edge {"
2446                         << getBlockName(From) << "," << getBlockName(To)
2447                         << "} to it found\n");
2448 
2449       // We've made an edge reachable to an existing block, which may
2450       // impact predicates. Otherwise, only mark the phi nodes as touched, as
2451       // they are the only thing that depend on new edges. Anything using their
2452       // values will get propagated to if necessary.
2453       if (MemoryAccess *MemPhi = getMemoryAccess(To))
2454         TouchedInstructions.set(InstrToDFSNum(MemPhi));
2455 
2456       // FIXME: We should just add a union op on a Bitvector and
2457       // SparseBitVector.  We can do it word by word faster than we are doing it
2458       // here.
2459       for (auto InstNum : RevisitOnReachabilityChange[To])
2460         TouchedInstructions.set(InstNum);
2461     }
2462   }
2463 }
2464 
2465 // Given a predicate condition (from a switch, cmp, or whatever) and a block,
2466 // see if we know some constant value for it already.
findConditionEquivalence(Value * Cond) const2467 Value *NewGVN::findConditionEquivalence(Value *Cond) const {
2468   auto Result = lookupOperandLeader(Cond);
2469   return isa<Constant>(Result) ? Result : nullptr;
2470 }
2471 
2472 // Process the outgoing edges of a block for reachability.
processOutgoingEdges(Instruction * TI,BasicBlock * B)2473 void NewGVN::processOutgoingEdges(Instruction *TI, BasicBlock *B) {
2474   // Evaluate reachability of terminator instruction.
2475   Value *Cond;
2476   BasicBlock *TrueSucc, *FalseSucc;
2477   if (match(TI, m_Br(m_Value(Cond), TrueSucc, FalseSucc))) {
2478     Value *CondEvaluated = findConditionEquivalence(Cond);
2479     if (!CondEvaluated) {
2480       if (auto *I = dyn_cast<Instruction>(Cond)) {
2481         SmallPtrSet<Value *, 4> Visited;
2482         auto Res = performSymbolicEvaluation(I, Visited);
2483         if (const auto *CE = dyn_cast_or_null<ConstantExpression>(Res.Expr)) {
2484           CondEvaluated = CE->getConstantValue();
2485           addAdditionalUsers(Res, I);
2486         } else {
2487           // Did not use simplification result, no need to add the extra
2488           // dependency.
2489           Res.ExtraDep = nullptr;
2490         }
2491       } else if (isa<ConstantInt>(Cond)) {
2492         CondEvaluated = Cond;
2493       }
2494     }
2495     ConstantInt *CI;
2496     if (CondEvaluated && (CI = dyn_cast<ConstantInt>(CondEvaluated))) {
2497       if (CI->isOne()) {
2498         LLVM_DEBUG(dbgs() << "Condition for Terminator " << *TI
2499                           << " evaluated to true\n");
2500         updateReachableEdge(B, TrueSucc);
2501       } else if (CI->isZero()) {
2502         LLVM_DEBUG(dbgs() << "Condition for Terminator " << *TI
2503                           << " evaluated to false\n");
2504         updateReachableEdge(B, FalseSucc);
2505       }
2506     } else {
2507       updateReachableEdge(B, TrueSucc);
2508       updateReachableEdge(B, FalseSucc);
2509     }
2510   } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
2511     // For switches, propagate the case values into the case
2512     // destinations.
2513 
2514     Value *SwitchCond = SI->getCondition();
2515     Value *CondEvaluated = findConditionEquivalence(SwitchCond);
2516     // See if we were able to turn this switch statement into a constant.
2517     if (CondEvaluated && isa<ConstantInt>(CondEvaluated)) {
2518       auto *CondVal = cast<ConstantInt>(CondEvaluated);
2519       // We should be able to get case value for this.
2520       auto Case = *SI->findCaseValue(CondVal);
2521       if (Case.getCaseSuccessor() == SI->getDefaultDest()) {
2522         // We proved the value is outside of the range of the case.
2523         // We can't do anything other than mark the default dest as reachable,
2524         // and go home.
2525         updateReachableEdge(B, SI->getDefaultDest());
2526         return;
2527       }
2528       // Now get where it goes and mark it reachable.
2529       BasicBlock *TargetBlock = Case.getCaseSuccessor();
2530       updateReachableEdge(B, TargetBlock);
2531     } else {
2532       for (BasicBlock *TargetBlock : successors(SI->getParent()))
2533         updateReachableEdge(B, TargetBlock);
2534     }
2535   } else {
2536     // Otherwise this is either unconditional, or a type we have no
2537     // idea about. Just mark successors as reachable.
2538     for (BasicBlock *TargetBlock : successors(TI->getParent()))
2539       updateReachableEdge(B, TargetBlock);
2540 
2541     // This also may be a memory defining terminator, in which case, set it
2542     // equivalent only to itself.
2543     //
2544     auto *MA = getMemoryAccess(TI);
2545     if (MA && !isa<MemoryUse>(MA)) {
2546       auto *CC = ensureLeaderOfMemoryClass(MA);
2547       if (setMemoryClass(MA, CC))
2548         markMemoryUsersTouched(MA);
2549     }
2550   }
2551 }
2552 
2553 // Remove the PHI of Ops PHI for I
removePhiOfOps(Instruction * I,PHINode * PHITemp)2554 void NewGVN::removePhiOfOps(Instruction *I, PHINode *PHITemp) {
2555   InstrDFS.erase(PHITemp);
2556   // It's still a temp instruction. We keep it in the array so it gets erased.
2557   // However, it's no longer used by I, or in the block
2558   TempToBlock.erase(PHITemp);
2559   RealToTemp.erase(I);
2560   // We don't remove the users from the phi node uses. This wastes a little
2561   // time, but such is life.  We could use two sets to track which were there
2562   // are the start of NewGVN, and which were added, but right nowt he cost of
2563   // tracking is more than the cost of checking for more phi of ops.
2564 }
2565 
2566 // Add PHI Op in BB as a PHI of operations version of ExistingValue.
addPhiOfOps(PHINode * Op,BasicBlock * BB,Instruction * ExistingValue)2567 void NewGVN::addPhiOfOps(PHINode *Op, BasicBlock *BB,
2568                          Instruction *ExistingValue) {
2569   InstrDFS[Op] = InstrToDFSNum(ExistingValue);
2570   AllTempInstructions.insert(Op);
2571   TempToBlock[Op] = BB;
2572   RealToTemp[ExistingValue] = Op;
2573   // Add all users to phi node use, as they are now uses of the phi of ops phis
2574   // and may themselves be phi of ops.
2575   for (auto *U : ExistingValue->users())
2576     if (auto *UI = dyn_cast<Instruction>(U))
2577       PHINodeUses.insert(UI);
2578 }
2579 
okayForPHIOfOps(const Instruction * I)2580 static bool okayForPHIOfOps(const Instruction *I) {
2581   if (!EnablePhiOfOps)
2582     return false;
2583   return isa<BinaryOperator>(I) || isa<SelectInst>(I) || isa<CmpInst>(I) ||
2584          isa<LoadInst>(I);
2585 }
2586 
2587 // Return true if this operand will be safe to use for phi of ops.
2588 //
2589 // The reason some operands are unsafe is that we are not trying to recursively
2590 // translate everything back through phi nodes.  We actually expect some lookups
2591 // of expressions to fail.  In particular, a lookup where the expression cannot
2592 // exist in the predecessor.  This is true even if the expression, as shown, can
2593 // be determined to be constant.
OpIsSafeForPHIOfOps(Value * V,const BasicBlock * PHIBlock,SmallPtrSetImpl<const Value * > & Visited)2594 bool NewGVN::OpIsSafeForPHIOfOps(Value *V, const BasicBlock *PHIBlock,
2595                                  SmallPtrSetImpl<const Value *> &Visited) {
2596   SmallVector<Value *, 4> Worklist;
2597   Worklist.push_back(V);
2598   while (!Worklist.empty()) {
2599     auto *I = Worklist.pop_back_val();
2600     if (!isa<Instruction>(I))
2601       continue;
2602 
2603     auto OISIt = OpSafeForPHIOfOps.find({I, CacheIdx});
2604     if (OISIt != OpSafeForPHIOfOps.end())
2605       return OISIt->second;
2606 
2607     // Keep walking until we either dominate the phi block, or hit a phi, or run
2608     // out of things to check.
2609     if (DT->properlyDominates(getBlockForValue(I), PHIBlock)) {
2610       OpSafeForPHIOfOps.insert({{I, CacheIdx}, true});
2611       continue;
2612     }
2613     // PHI in the same block.
2614     if (isa<PHINode>(I) && getBlockForValue(I) == PHIBlock) {
2615       OpSafeForPHIOfOps.insert({{I, CacheIdx}, false});
2616       return false;
2617     }
2618 
2619     auto *OrigI = cast<Instruction>(I);
2620     // When we hit an instruction that reads memory (load, call, etc), we must
2621     // consider any store that may happen in the loop. For now, we assume the
2622     // worst: there is a store in the loop that alias with this read.
2623     // The case where the load is outside the loop is already covered by the
2624     // dominator check above.
2625     // TODO: relax this condition
2626     if (OrigI->mayReadFromMemory())
2627       return false;
2628 
2629     // Check the operands of the current instruction.
2630     for (auto *Op : OrigI->operand_values()) {
2631       if (!isa<Instruction>(Op))
2632         continue;
2633       // Stop now if we find an unsafe operand.
2634       auto OISIt = OpSafeForPHIOfOps.find({OrigI, CacheIdx});
2635       if (OISIt != OpSafeForPHIOfOps.end()) {
2636         if (!OISIt->second) {
2637           OpSafeForPHIOfOps.insert({{I, CacheIdx}, false});
2638           return false;
2639         }
2640         continue;
2641       }
2642       if (!Visited.insert(Op).second)
2643         continue;
2644       Worklist.push_back(cast<Instruction>(Op));
2645     }
2646   }
2647   OpSafeForPHIOfOps.insert({{V, CacheIdx}, true});
2648   return true;
2649 }
2650 
2651 // Try to find a leader for instruction TransInst, which is a phi translated
2652 // version of something in our original program.  Visited is used to ensure we
2653 // don't infinite loop during translations of cycles.  OrigInst is the
2654 // instruction in the original program, and PredBB is the predecessor we
2655 // translated it through.
findLeaderForInst(Instruction * TransInst,SmallPtrSetImpl<Value * > & Visited,MemoryAccess * MemAccess,Instruction * OrigInst,BasicBlock * PredBB)2656 Value *NewGVN::findLeaderForInst(Instruction *TransInst,
2657                                  SmallPtrSetImpl<Value *> &Visited,
2658                                  MemoryAccess *MemAccess, Instruction *OrigInst,
2659                                  BasicBlock *PredBB) {
2660   unsigned IDFSNum = InstrToDFSNum(OrigInst);
2661   // Make sure it's marked as a temporary instruction.
2662   AllTempInstructions.insert(TransInst);
2663   // and make sure anything that tries to add it's DFS number is
2664   // redirected to the instruction we are making a phi of ops
2665   // for.
2666   TempToBlock.insert({TransInst, PredBB});
2667   InstrDFS.insert({TransInst, IDFSNum});
2668 
2669   auto Res = performSymbolicEvaluation(TransInst, Visited);
2670   const Expression *E = Res.Expr;
2671   addAdditionalUsers(Res, OrigInst);
2672   InstrDFS.erase(TransInst);
2673   AllTempInstructions.erase(TransInst);
2674   TempToBlock.erase(TransInst);
2675   if (MemAccess)
2676     TempToMemory.erase(TransInst);
2677   if (!E)
2678     return nullptr;
2679   auto *FoundVal = findPHIOfOpsLeader(E, OrigInst, PredBB);
2680   if (!FoundVal) {
2681     ExpressionToPhiOfOps[E].insert(OrigInst);
2682     LLVM_DEBUG(dbgs() << "Cannot find phi of ops operand for " << *TransInst
2683                       << " in block " << getBlockName(PredBB) << "\n");
2684     return nullptr;
2685   }
2686   if (auto *SI = dyn_cast<StoreInst>(FoundVal))
2687     FoundVal = SI->getValueOperand();
2688   return FoundVal;
2689 }
2690 
2691 // When we see an instruction that is an op of phis, generate the equivalent phi
2692 // of ops form.
2693 const Expression *
makePossiblePHIOfOps(Instruction * I,SmallPtrSetImpl<Value * > & Visited)2694 NewGVN::makePossiblePHIOfOps(Instruction *I,
2695                              SmallPtrSetImpl<Value *> &Visited) {
2696   if (!okayForPHIOfOps(I))
2697     return nullptr;
2698 
2699   if (!Visited.insert(I).second)
2700     return nullptr;
2701   // For now, we require the instruction be cycle free because we don't
2702   // *always* create a phi of ops for instructions that could be done as phi
2703   // of ops, we only do it if we think it is useful.  If we did do it all the
2704   // time, we could remove the cycle free check.
2705   if (!isCycleFree(I))
2706     return nullptr;
2707 
2708   SmallPtrSet<const Value *, 8> ProcessedPHIs;
2709   // TODO: We don't do phi translation on memory accesses because it's
2710   // complicated. For a load, we'd need to be able to simulate a new memoryuse,
2711   // which we don't have a good way of doing ATM.
2712   auto *MemAccess = getMemoryAccess(I);
2713   // If the memory operation is defined by a memory operation this block that
2714   // isn't a MemoryPhi, transforming the pointer backwards through a scalar phi
2715   // can't help, as it would still be killed by that memory operation.
2716   if (MemAccess && !isa<MemoryPhi>(MemAccess->getDefiningAccess()) &&
2717       MemAccess->getDefiningAccess()->getBlock() == I->getParent())
2718     return nullptr;
2719 
2720   // Convert op of phis to phi of ops
2721   SmallPtrSet<const Value *, 10> VisitedOps;
2722   SmallVector<Value *, 4> Ops(I->operand_values());
2723   BasicBlock *SamePHIBlock = nullptr;
2724   PHINode *OpPHI = nullptr;
2725   if (!DebugCounter::shouldExecute(PHIOfOpsCounter))
2726     return nullptr;
2727   for (auto *Op : Ops) {
2728     if (!isa<PHINode>(Op)) {
2729       auto *ValuePHI = RealToTemp.lookup(Op);
2730       if (!ValuePHI)
2731         continue;
2732       LLVM_DEBUG(dbgs() << "Found possible dependent phi of ops\n");
2733       Op = ValuePHI;
2734     }
2735     OpPHI = cast<PHINode>(Op);
2736     if (!SamePHIBlock) {
2737       SamePHIBlock = getBlockForValue(OpPHI);
2738     } else if (SamePHIBlock != getBlockForValue(OpPHI)) {
2739       LLVM_DEBUG(
2740           dbgs()
2741           << "PHIs for operands are not all in the same block, aborting\n");
2742       return nullptr;
2743     }
2744     // No point in doing this for one-operand phis.
2745     // Since all PHIs for operands must be in the same block, then they must
2746     // have the same number of operands so we can just abort.
2747     if (OpPHI->getNumOperands() == 1)
2748       return nullptr;
2749   }
2750 
2751   if (!OpPHI)
2752     return nullptr;
2753 
2754   SmallVector<ValPair, 4> PHIOps;
2755   SmallPtrSet<Value *, 4> Deps;
2756   auto *PHIBlock = getBlockForValue(OpPHI);
2757   RevisitOnReachabilityChange[PHIBlock].reset(InstrToDFSNum(I));
2758   for (unsigned PredNum = 0; PredNum < OpPHI->getNumOperands(); ++PredNum) {
2759     auto *PredBB = OpPHI->getIncomingBlock(PredNum);
2760     Value *FoundVal = nullptr;
2761     SmallPtrSet<Value *, 4> CurrentDeps;
2762     // We could just skip unreachable edges entirely but it's tricky to do
2763     // with rewriting existing phi nodes.
2764     if (ReachableEdges.count({PredBB, PHIBlock})) {
2765       // Clone the instruction, create an expression from it that is
2766       // translated back into the predecessor, and see if we have a leader.
2767       Instruction *ValueOp = I->clone();
2768       // Emit the temporal instruction in the predecessor basic block where the
2769       // corresponding value is defined.
2770       ValueOp->insertBefore(PredBB->getTerminator());
2771       if (MemAccess)
2772         TempToMemory.insert({ValueOp, MemAccess});
2773       bool SafeForPHIOfOps = true;
2774       VisitedOps.clear();
2775       for (auto &Op : ValueOp->operands()) {
2776         auto *OrigOp = &*Op;
2777         // When these operand changes, it could change whether there is a
2778         // leader for us or not, so we have to add additional users.
2779         if (isa<PHINode>(Op)) {
2780           Op = Op->DoPHITranslation(PHIBlock, PredBB);
2781           if (Op != OrigOp && Op != I)
2782             CurrentDeps.insert(Op);
2783         } else if (auto *ValuePHI = RealToTemp.lookup(Op)) {
2784           if (getBlockForValue(ValuePHI) == PHIBlock)
2785             Op = ValuePHI->getIncomingValueForBlock(PredBB);
2786         }
2787         // If we phi-translated the op, it must be safe.
2788         SafeForPHIOfOps =
2789             SafeForPHIOfOps &&
2790             (Op != OrigOp || OpIsSafeForPHIOfOps(Op, PHIBlock, VisitedOps));
2791       }
2792       // FIXME: For those things that are not safe we could generate
2793       // expressions all the way down, and see if this comes out to a
2794       // constant.  For anything where that is true, and unsafe, we should
2795       // have made a phi-of-ops (or value numbered it equivalent to something)
2796       // for the pieces already.
2797       FoundVal = !SafeForPHIOfOps ? nullptr
2798                                   : findLeaderForInst(ValueOp, Visited,
2799                                                       MemAccess, I, PredBB);
2800       ValueOp->eraseFromParent();
2801       if (!FoundVal) {
2802         // We failed to find a leader for the current ValueOp, but this might
2803         // change in case of the translated operands change.
2804         if (SafeForPHIOfOps)
2805           for (auto *Dep : CurrentDeps)
2806             addAdditionalUsers(Dep, I);
2807 
2808         return nullptr;
2809       }
2810       Deps.insert(CurrentDeps.begin(), CurrentDeps.end());
2811     } else {
2812       LLVM_DEBUG(dbgs() << "Skipping phi of ops operand for incoming block "
2813                         << getBlockName(PredBB)
2814                         << " because the block is unreachable\n");
2815       FoundVal = PoisonValue::get(I->getType());
2816       RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2817     }
2818 
2819     PHIOps.push_back({FoundVal, PredBB});
2820     LLVM_DEBUG(dbgs() << "Found phi of ops operand " << *FoundVal << " in "
2821                       << getBlockName(PredBB) << "\n");
2822   }
2823   for (auto *Dep : Deps)
2824     addAdditionalUsers(Dep, I);
2825   sortPHIOps(PHIOps);
2826   auto *E = performSymbolicPHIEvaluation(PHIOps, I, PHIBlock);
2827   if (isa<ConstantExpression>(E) || isa<VariableExpression>(E)) {
2828     LLVM_DEBUG(
2829         dbgs()
2830         << "Not creating real PHI of ops because it simplified to existing "
2831            "value or constant\n");
2832     // We have leaders for all operands, but do not create a real PHI node with
2833     // those leaders as operands, so the link between the operands and the
2834     // PHI-of-ops is not materialized in the IR. If any of those leaders
2835     // changes, the PHI-of-op may change also, so we need to add the operands as
2836     // additional users.
2837     for (auto &O : PHIOps)
2838       addAdditionalUsers(O.first, I);
2839 
2840     return E;
2841   }
2842   auto *ValuePHI = RealToTemp.lookup(I);
2843   bool NewPHI = false;
2844   if (!ValuePHI) {
2845     ValuePHI =
2846         PHINode::Create(I->getType(), OpPHI->getNumOperands(), "phiofops");
2847     addPhiOfOps(ValuePHI, PHIBlock, I);
2848     NewPHI = true;
2849     NumGVNPHIOfOpsCreated++;
2850   }
2851   if (NewPHI) {
2852     for (auto PHIOp : PHIOps)
2853       ValuePHI->addIncoming(PHIOp.first, PHIOp.second);
2854   } else {
2855     TempToBlock[ValuePHI] = PHIBlock;
2856     unsigned int i = 0;
2857     for (auto PHIOp : PHIOps) {
2858       ValuePHI->setIncomingValue(i, PHIOp.first);
2859       ValuePHI->setIncomingBlock(i, PHIOp.second);
2860       ++i;
2861     }
2862   }
2863   RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2864   LLVM_DEBUG(dbgs() << "Created phi of ops " << *ValuePHI << " for " << *I
2865                     << "\n");
2866 
2867   return E;
2868 }
2869 
2870 // The algorithm initially places the values of the routine in the TOP
2871 // congruence class. The leader of TOP is the undetermined value `poison`.
2872 // When the algorithm has finished, values still in TOP are unreachable.
initializeCongruenceClasses(Function & F)2873 void NewGVN::initializeCongruenceClasses(Function &F) {
2874   NextCongruenceNum = 0;
2875 
2876   // Note that even though we use the live on entry def as a representative
2877   // MemoryAccess, it is *not* the same as the actual live on entry def. We
2878   // have no real equivalent to poison for MemoryAccesses, and so we really
2879   // should be checking whether the MemoryAccess is top if we want to know if it
2880   // is equivalent to everything.  Otherwise, what this really signifies is that
2881   // the access "it reaches all the way back to the beginning of the function"
2882 
2883   // Initialize all other instructions to be in TOP class.
2884   TOPClass = createCongruenceClass(nullptr, nullptr);
2885   TOPClass->setMemoryLeader(MSSA->getLiveOnEntryDef());
2886   //  The live on entry def gets put into it's own class
2887   MemoryAccessToClass[MSSA->getLiveOnEntryDef()] =
2888       createMemoryClass(MSSA->getLiveOnEntryDef());
2889 
2890   for (auto *DTN : nodes(DT)) {
2891     BasicBlock *BB = DTN->getBlock();
2892     // All MemoryAccesses are equivalent to live on entry to start. They must
2893     // be initialized to something so that initial changes are noticed. For
2894     // the maximal answer, we initialize them all to be the same as
2895     // liveOnEntry.
2896     auto *MemoryBlockDefs = MSSA->getBlockDefs(BB);
2897     if (MemoryBlockDefs)
2898       for (const auto &Def : *MemoryBlockDefs) {
2899         MemoryAccessToClass[&Def] = TOPClass;
2900         auto *MD = dyn_cast<MemoryDef>(&Def);
2901         // Insert the memory phis into the member list.
2902         if (!MD) {
2903           const MemoryPhi *MP = cast<MemoryPhi>(&Def);
2904           TOPClass->memory_insert(MP);
2905           MemoryPhiState.insert({MP, MPS_TOP});
2906         }
2907 
2908         if (MD && isa<StoreInst>(MD->getMemoryInst()))
2909           TOPClass->incStoreCount();
2910       }
2911 
2912     // FIXME: This is trying to discover which instructions are uses of phi
2913     // nodes.  We should move this into one of the myriad of places that walk
2914     // all the operands already.
2915     for (auto &I : *BB) {
2916       if (isa<PHINode>(&I))
2917         for (auto *U : I.users())
2918           if (auto *UInst = dyn_cast<Instruction>(U))
2919             if (InstrToDFSNum(UInst) != 0 && okayForPHIOfOps(UInst))
2920               PHINodeUses.insert(UInst);
2921       // Don't insert void terminators into the class. We don't value number
2922       // them, and they just end up sitting in TOP.
2923       if (I.isTerminator() && I.getType()->isVoidTy())
2924         continue;
2925       TOPClass->insert(&I);
2926       ValueToClass[&I] = TOPClass;
2927     }
2928   }
2929 
2930   // Initialize arguments to be in their own unique congruence classes
2931   for (auto &FA : F.args())
2932     createSingletonCongruenceClass(&FA);
2933 }
2934 
cleanupTables()2935 void NewGVN::cleanupTables() {
2936   for (CongruenceClass *&CC : CongruenceClasses) {
2937     LLVM_DEBUG(dbgs() << "Congruence class " << CC->getID() << " has "
2938                       << CC->size() << " members\n");
2939     // Make sure we delete the congruence class (probably worth switching to
2940     // a unique_ptr at some point.
2941     delete CC;
2942     CC = nullptr;
2943   }
2944 
2945   // Destroy the value expressions
2946   SmallVector<Instruction *, 8> TempInst(AllTempInstructions.begin(),
2947                                          AllTempInstructions.end());
2948   AllTempInstructions.clear();
2949 
2950   // We have to drop all references for everything first, so there are no uses
2951   // left as we delete them.
2952   for (auto *I : TempInst) {
2953     I->dropAllReferences();
2954   }
2955 
2956   while (!TempInst.empty()) {
2957     auto *I = TempInst.pop_back_val();
2958     I->deleteValue();
2959   }
2960 
2961   ValueToClass.clear();
2962   ArgRecycler.clear(ExpressionAllocator);
2963   ExpressionAllocator.Reset();
2964   CongruenceClasses.clear();
2965   ExpressionToClass.clear();
2966   ValueToExpression.clear();
2967   RealToTemp.clear();
2968   AdditionalUsers.clear();
2969   ExpressionToPhiOfOps.clear();
2970   TempToBlock.clear();
2971   TempToMemory.clear();
2972   PHINodeUses.clear();
2973   OpSafeForPHIOfOps.clear();
2974   ReachableBlocks.clear();
2975   ReachableEdges.clear();
2976 #ifndef NDEBUG
2977   ProcessedCount.clear();
2978 #endif
2979   InstrDFS.clear();
2980   InstructionsToErase.clear();
2981   DFSToInstr.clear();
2982   BlockInstRange.clear();
2983   TouchedInstructions.clear();
2984   MemoryAccessToClass.clear();
2985   PredicateToUsers.clear();
2986   MemoryToUsers.clear();
2987   RevisitOnReachabilityChange.clear();
2988   IntrinsicInstPred.clear();
2989 }
2990 
2991 // Assign local DFS number mapping to instructions, and leave space for Value
2992 // PHI's.
assignDFSNumbers(BasicBlock * B,unsigned Start)2993 std::pair<unsigned, unsigned> NewGVN::assignDFSNumbers(BasicBlock *B,
2994                                                        unsigned Start) {
2995   unsigned End = Start;
2996   if (MemoryAccess *MemPhi = getMemoryAccess(B)) {
2997     InstrDFS[MemPhi] = End++;
2998     DFSToInstr.emplace_back(MemPhi);
2999   }
3000 
3001   // Then the real block goes next.
3002   for (auto &I : *B) {
3003     // There's no need to call isInstructionTriviallyDead more than once on
3004     // an instruction. Therefore, once we know that an instruction is dead
3005     // we change its DFS number so that it doesn't get value numbered.
3006     if (isInstructionTriviallyDead(&I, TLI)) {
3007       InstrDFS[&I] = 0;
3008       LLVM_DEBUG(dbgs() << "Skipping trivially dead instruction " << I << "\n");
3009       markInstructionForDeletion(&I);
3010       continue;
3011     }
3012     if (isa<PHINode>(&I))
3013       RevisitOnReachabilityChange[B].set(End);
3014     InstrDFS[&I] = End++;
3015     DFSToInstr.emplace_back(&I);
3016   }
3017 
3018   // All of the range functions taken half-open ranges (open on the end side).
3019   // So we do not subtract one from count, because at this point it is one
3020   // greater than the last instruction.
3021   return std::make_pair(Start, End);
3022 }
3023 
updateProcessedCount(const Value * V)3024 void NewGVN::updateProcessedCount(const Value *V) {
3025 #ifndef NDEBUG
3026   if (ProcessedCount.count(V) == 0) {
3027     ProcessedCount.insert({V, 1});
3028   } else {
3029     ++ProcessedCount[V];
3030     assert(ProcessedCount[V] < 100 &&
3031            "Seem to have processed the same Value a lot");
3032   }
3033 #endif
3034 }
3035 
3036 // Evaluate MemoryPhi nodes symbolically, just like PHI nodes
valueNumberMemoryPhi(MemoryPhi * MP)3037 void NewGVN::valueNumberMemoryPhi(MemoryPhi *MP) {
3038   // If all the arguments are the same, the MemoryPhi has the same value as the
3039   // argument.  Filter out unreachable blocks and self phis from our operands.
3040   // TODO: We could do cycle-checking on the memory phis to allow valueizing for
3041   // self-phi checking.
3042   const BasicBlock *PHIBlock = MP->getBlock();
3043   auto Filtered = make_filter_range(MP->operands(), [&](const Use &U) {
3044     return cast<MemoryAccess>(U) != MP &&
3045            !isMemoryAccessTOP(cast<MemoryAccess>(U)) &&
3046            ReachableEdges.count({MP->getIncomingBlock(U), PHIBlock});
3047   });
3048   // If all that is left is nothing, our memoryphi is poison. We keep it as
3049   // InitialClass.  Note: The only case this should happen is if we have at
3050   // least one self-argument.
3051   if (Filtered.begin() == Filtered.end()) {
3052     if (setMemoryClass(MP, TOPClass))
3053       markMemoryUsersTouched(MP);
3054     return;
3055   }
3056 
3057   // Transform the remaining operands into operand leaders.
3058   // FIXME: mapped_iterator should have a range version.
3059   auto LookupFunc = [&](const Use &U) {
3060     return lookupMemoryLeader(cast<MemoryAccess>(U));
3061   };
3062   auto MappedBegin = map_iterator(Filtered.begin(), LookupFunc);
3063   auto MappedEnd = map_iterator(Filtered.end(), LookupFunc);
3064 
3065   // and now check if all the elements are equal.
3066   // Sadly, we can't use std::equals since these are random access iterators.
3067   const auto *AllSameValue = *MappedBegin;
3068   ++MappedBegin;
3069   bool AllEqual = std::all_of(
3070       MappedBegin, MappedEnd,
3071       [&AllSameValue](const MemoryAccess *V) { return V == AllSameValue; });
3072 
3073   if (AllEqual)
3074     LLVM_DEBUG(dbgs() << "Memory Phi value numbered to " << *AllSameValue
3075                       << "\n");
3076   else
3077     LLVM_DEBUG(dbgs() << "Memory Phi value numbered to itself\n");
3078   // If it's equal to something, it's in that class. Otherwise, it has to be in
3079   // a class where it is the leader (other things may be equivalent to it, but
3080   // it needs to start off in its own class, which means it must have been the
3081   // leader, and it can't have stopped being the leader because it was never
3082   // removed).
3083   CongruenceClass *CC =
3084       AllEqual ? getMemoryClass(AllSameValue) : ensureLeaderOfMemoryClass(MP);
3085   auto OldState = MemoryPhiState.lookup(MP);
3086   assert(OldState != MPS_Invalid && "Invalid memory phi state");
3087   auto NewState = AllEqual ? MPS_Equivalent : MPS_Unique;
3088   MemoryPhiState[MP] = NewState;
3089   if (setMemoryClass(MP, CC) || OldState != NewState)
3090     markMemoryUsersTouched(MP);
3091 }
3092 
3093 // Value number a single instruction, symbolically evaluating, performing
3094 // congruence finding, and updating mappings.
valueNumberInstruction(Instruction * I)3095 void NewGVN::valueNumberInstruction(Instruction *I) {
3096   LLVM_DEBUG(dbgs() << "Processing instruction " << *I << "\n");
3097   if (!I->isTerminator()) {
3098     const Expression *Symbolized = nullptr;
3099     SmallPtrSet<Value *, 2> Visited;
3100     if (DebugCounter::shouldExecute(VNCounter)) {
3101       auto Res = performSymbolicEvaluation(I, Visited);
3102       Symbolized = Res.Expr;
3103       addAdditionalUsers(Res, I);
3104 
3105       // Make a phi of ops if necessary
3106       if (Symbolized && !isa<ConstantExpression>(Symbolized) &&
3107           !isa<VariableExpression>(Symbolized) && PHINodeUses.count(I)) {
3108         auto *PHIE = makePossiblePHIOfOps(I, Visited);
3109         // If we created a phi of ops, use it.
3110         // If we couldn't create one, make sure we don't leave one lying around
3111         if (PHIE) {
3112           Symbolized = PHIE;
3113         } else if (auto *Op = RealToTemp.lookup(I)) {
3114           removePhiOfOps(I, Op);
3115         }
3116       }
3117     } else {
3118       // Mark the instruction as unused so we don't value number it again.
3119       InstrDFS[I] = 0;
3120     }
3121     // If we couldn't come up with a symbolic expression, use the unknown
3122     // expression
3123     if (Symbolized == nullptr)
3124       Symbolized = createUnknownExpression(I);
3125     performCongruenceFinding(I, Symbolized);
3126   } else {
3127     // Handle terminators that return values. All of them produce values we
3128     // don't currently understand.  We don't place non-value producing
3129     // terminators in a class.
3130     if (!I->getType()->isVoidTy()) {
3131       auto *Symbolized = createUnknownExpression(I);
3132       performCongruenceFinding(I, Symbolized);
3133     }
3134     processOutgoingEdges(I, I->getParent());
3135   }
3136 }
3137 
3138 // Check if there is a path, using single or equal argument phi nodes, from
3139 // First to Second.
singleReachablePHIPath(SmallPtrSet<const MemoryAccess *,8> & Visited,const MemoryAccess * First,const MemoryAccess * Second) const3140 bool NewGVN::singleReachablePHIPath(
3141     SmallPtrSet<const MemoryAccess *, 8> &Visited, const MemoryAccess *First,
3142     const MemoryAccess *Second) const {
3143   if (First == Second)
3144     return true;
3145   if (MSSA->isLiveOnEntryDef(First))
3146     return false;
3147 
3148   // This is not perfect, but as we're just verifying here, we can live with
3149   // the loss of precision. The real solution would be that of doing strongly
3150   // connected component finding in this routine, and it's probably not worth
3151   // the complexity for the time being. So, we just keep a set of visited
3152   // MemoryAccess and return true when we hit a cycle.
3153   if (!Visited.insert(First).second)
3154     return true;
3155 
3156   const auto *EndDef = First;
3157   for (const auto *ChainDef : optimized_def_chain(First)) {
3158     if (ChainDef == Second)
3159       return true;
3160     if (MSSA->isLiveOnEntryDef(ChainDef))
3161       return false;
3162     EndDef = ChainDef;
3163   }
3164   auto *MP = cast<MemoryPhi>(EndDef);
3165   auto ReachableOperandPred = [&](const Use &U) {
3166     return ReachableEdges.count({MP->getIncomingBlock(U), MP->getBlock()});
3167   };
3168   auto FilteredPhiArgs =
3169       make_filter_range(MP->operands(), ReachableOperandPred);
3170   SmallVector<const Value *, 32> OperandList;
3171   llvm::copy(FilteredPhiArgs, std::back_inserter(OperandList));
3172   bool Okay = all_equal(OperandList);
3173   if (Okay)
3174     return singleReachablePHIPath(Visited, cast<MemoryAccess>(OperandList[0]),
3175                                   Second);
3176   return false;
3177 }
3178 
3179 // Verify the that the memory equivalence table makes sense relative to the
3180 // congruence classes.  Note that this checking is not perfect, and is currently
3181 // subject to very rare false negatives. It is only useful for
3182 // testing/debugging.
verifyMemoryCongruency() const3183 void NewGVN::verifyMemoryCongruency() const {
3184 #ifndef NDEBUG
3185   // Verify that the memory table equivalence and memory member set match
3186   for (const auto *CC : CongruenceClasses) {
3187     if (CC == TOPClass || CC->isDead())
3188       continue;
3189     if (CC->getStoreCount() != 0) {
3190       assert((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) &&
3191              "Any class with a store as a leader should have a "
3192              "representative stored value");
3193       assert(CC->getMemoryLeader() &&
3194              "Any congruence class with a store should have a "
3195              "representative access");
3196     }
3197 
3198     if (CC->getMemoryLeader())
3199       assert(MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
3200              "Representative MemoryAccess does not appear to be reverse "
3201              "mapped properly");
3202     for (const auto *M : CC->memory())
3203       assert(MemoryAccessToClass.lookup(M) == CC &&
3204              "Memory member does not appear to be reverse mapped properly");
3205   }
3206 
3207   // Anything equivalent in the MemoryAccess table should be in the same
3208   // congruence class.
3209 
3210   // Filter out the unreachable and trivially dead entries, because they may
3211   // never have been updated if the instructions were not processed.
3212   auto ReachableAccessPred =
3213       [&](const std::pair<const MemoryAccess *, CongruenceClass *> Pair) {
3214         bool Result = ReachableBlocks.count(Pair.first->getBlock());
3215         if (!Result || MSSA->isLiveOnEntryDef(Pair.first) ||
3216             MemoryToDFSNum(Pair.first) == 0)
3217           return false;
3218         if (auto *MemDef = dyn_cast<MemoryDef>(Pair.first))
3219           return !isInstructionTriviallyDead(MemDef->getMemoryInst());
3220 
3221         // We could have phi nodes which operands are all trivially dead,
3222         // so we don't process them.
3223         if (auto *MemPHI = dyn_cast<MemoryPhi>(Pair.first)) {
3224           for (const auto &U : MemPHI->incoming_values()) {
3225             if (auto *I = dyn_cast<Instruction>(&*U)) {
3226               if (!isInstructionTriviallyDead(I))
3227                 return true;
3228             }
3229           }
3230           return false;
3231         }
3232 
3233         return true;
3234       };
3235 
3236   auto Filtered = make_filter_range(MemoryAccessToClass, ReachableAccessPred);
3237   for (auto KV : Filtered) {
3238     if (auto *FirstMUD = dyn_cast<MemoryUseOrDef>(KV.first)) {
3239       auto *SecondMUD = dyn_cast<MemoryUseOrDef>(KV.second->getMemoryLeader());
3240       if (FirstMUD && SecondMUD) {
3241         SmallPtrSet<const MemoryAccess *, 8> VisitedMAS;
3242         assert((singleReachablePHIPath(VisitedMAS, FirstMUD, SecondMUD) ||
3243                 ValueToClass.lookup(FirstMUD->getMemoryInst()) ==
3244                     ValueToClass.lookup(SecondMUD->getMemoryInst())) &&
3245                "The instructions for these memory operations should have "
3246                "been in the same congruence class or reachable through"
3247                "a single argument phi");
3248       }
3249     } else if (auto *FirstMP = dyn_cast<MemoryPhi>(KV.first)) {
3250       // We can only sanely verify that MemoryDefs in the operand list all have
3251       // the same class.
3252       auto ReachableOperandPred = [&](const Use &U) {
3253         return ReachableEdges.count(
3254                    {FirstMP->getIncomingBlock(U), FirstMP->getBlock()}) &&
3255                isa<MemoryDef>(U);
3256 
3257       };
3258       // All arguments should in the same class, ignoring unreachable arguments
3259       auto FilteredPhiArgs =
3260           make_filter_range(FirstMP->operands(), ReachableOperandPred);
3261       SmallVector<const CongruenceClass *, 16> PhiOpClasses;
3262       std::transform(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
3263                      std::back_inserter(PhiOpClasses), [&](const Use &U) {
3264                        const MemoryDef *MD = cast<MemoryDef>(U);
3265                        return ValueToClass.lookup(MD->getMemoryInst());
3266                      });
3267       assert(all_equal(PhiOpClasses) &&
3268              "All MemoryPhi arguments should be in the same class");
3269     }
3270   }
3271 #endif
3272 }
3273 
3274 // Verify that the sparse propagation we did actually found the maximal fixpoint
3275 // We do this by storing the value to class mapping, touching all instructions,
3276 // and redoing the iteration to see if anything changed.
verifyIterationSettled(Function & F)3277 void NewGVN::verifyIterationSettled(Function &F) {
3278 #ifndef NDEBUG
3279   LLVM_DEBUG(dbgs() << "Beginning iteration verification\n");
3280   if (DebugCounter::isCounterSet(VNCounter))
3281     DebugCounter::setCounterState(VNCounter, StartingVNCounter);
3282 
3283   // Note that we have to store the actual classes, as we may change existing
3284   // classes during iteration.  This is because our memory iteration propagation
3285   // is not perfect, and so may waste a little work.  But it should generate
3286   // exactly the same congruence classes we have now, with different IDs.
3287   std::map<const Value *, CongruenceClass> BeforeIteration;
3288 
3289   for (auto &KV : ValueToClass) {
3290     if (auto *I = dyn_cast<Instruction>(KV.first))
3291       // Skip unused/dead instructions.
3292       if (InstrToDFSNum(I) == 0)
3293         continue;
3294     BeforeIteration.insert({KV.first, *KV.second});
3295   }
3296 
3297   TouchedInstructions.set();
3298   TouchedInstructions.reset(0);
3299   OpSafeForPHIOfOps.clear();
3300   CacheIdx = 0;
3301   iterateTouchedInstructions();
3302   DenseSet<std::pair<const CongruenceClass *, const CongruenceClass *>>
3303       EqualClasses;
3304   for (const auto &KV : ValueToClass) {
3305     if (auto *I = dyn_cast<Instruction>(KV.first))
3306       // Skip unused/dead instructions.
3307       if (InstrToDFSNum(I) == 0)
3308         continue;
3309     // We could sink these uses, but i think this adds a bit of clarity here as
3310     // to what we are comparing.
3311     auto *BeforeCC = &BeforeIteration.find(KV.first)->second;
3312     auto *AfterCC = KV.second;
3313     // Note that the classes can't change at this point, so we memoize the set
3314     // that are equal.
3315     if (!EqualClasses.count({BeforeCC, AfterCC})) {
3316       assert(BeforeCC->isEquivalentTo(AfterCC) &&
3317              "Value number changed after main loop completed!");
3318       EqualClasses.insert({BeforeCC, AfterCC});
3319     }
3320   }
3321 #endif
3322 }
3323 
3324 // Verify that for each store expression in the expression to class mapping,
3325 // only the latest appears, and multiple ones do not appear.
3326 // Because loads do not use the stored value when doing equality with stores,
3327 // if we don't erase the old store expressions from the table, a load can find
3328 // a no-longer valid StoreExpression.
verifyStoreExpressions() const3329 void NewGVN::verifyStoreExpressions() const {
3330 #ifndef NDEBUG
3331   // This is the only use of this, and it's not worth defining a complicated
3332   // densemapinfo hash/equality function for it.
3333   std::set<
3334       std::pair<const Value *,
3335                 std::tuple<const Value *, const CongruenceClass *, Value *>>>
3336       StoreExpressionSet;
3337   for (const auto &KV : ExpressionToClass) {
3338     if (auto *SE = dyn_cast<StoreExpression>(KV.first)) {
3339       // Make sure a version that will conflict with loads is not already there
3340       auto Res = StoreExpressionSet.insert(
3341           {SE->getOperand(0), std::make_tuple(SE->getMemoryLeader(), KV.second,
3342                                               SE->getStoredValue())});
3343       bool Okay = Res.second;
3344       // It's okay to have the same expression already in there if it is
3345       // identical in nature.
3346       // This can happen when the leader of the stored value changes over time.
3347       if (!Okay)
3348         Okay = (std::get<1>(Res.first->second) == KV.second) &&
3349                (lookupOperandLeader(std::get<2>(Res.first->second)) ==
3350                 lookupOperandLeader(SE->getStoredValue()));
3351       assert(Okay && "Stored expression conflict exists in expression table");
3352       auto *ValueExpr = ValueToExpression.lookup(SE->getStoreInst());
3353       assert(ValueExpr && ValueExpr->equals(*SE) &&
3354              "StoreExpression in ExpressionToClass is not latest "
3355              "StoreExpression for value");
3356     }
3357   }
3358 #endif
3359 }
3360 
3361 // This is the main value numbering loop, it iterates over the initial touched
3362 // instruction set, propagating value numbers, marking things touched, etc,
3363 // until the set of touched instructions is completely empty.
iterateTouchedInstructions()3364 void NewGVN::iterateTouchedInstructions() {
3365   uint64_t Iterations = 0;
3366   // Figure out where touchedinstructions starts
3367   int FirstInstr = TouchedInstructions.find_first();
3368   // Nothing set, nothing to iterate, just return.
3369   if (FirstInstr == -1)
3370     return;
3371   const BasicBlock *LastBlock = getBlockForValue(InstrFromDFSNum(FirstInstr));
3372   while (TouchedInstructions.any()) {
3373     ++Iterations;
3374     // Walk through all the instructions in all the blocks in RPO.
3375     // TODO: As we hit a new block, we should push and pop equalities into a
3376     // table lookupOperandLeader can use, to catch things PredicateInfo
3377     // might miss, like edge-only equivalences.
3378     for (unsigned InstrNum : TouchedInstructions.set_bits()) {
3379 
3380       // This instruction was found to be dead. We don't bother looking
3381       // at it again.
3382       if (InstrNum == 0) {
3383         TouchedInstructions.reset(InstrNum);
3384         continue;
3385       }
3386 
3387       Value *V = InstrFromDFSNum(InstrNum);
3388       const BasicBlock *CurrBlock = getBlockForValue(V);
3389 
3390       // If we hit a new block, do reachability processing.
3391       if (CurrBlock != LastBlock) {
3392         LastBlock = CurrBlock;
3393         bool BlockReachable = ReachableBlocks.count(CurrBlock);
3394         const auto &CurrInstRange = BlockInstRange.lookup(CurrBlock);
3395 
3396         // If it's not reachable, erase any touched instructions and move on.
3397         if (!BlockReachable) {
3398           TouchedInstructions.reset(CurrInstRange.first, CurrInstRange.second);
3399           LLVM_DEBUG(dbgs() << "Skipping instructions in block "
3400                             << getBlockName(CurrBlock)
3401                             << " because it is unreachable\n");
3402           continue;
3403         }
3404         // Use the appropriate cache for "OpIsSafeForPHIOfOps".
3405         CacheIdx = RPOOrdering.lookup(DT->getNode(CurrBlock)) - 1;
3406         updateProcessedCount(CurrBlock);
3407       }
3408       // Reset after processing (because we may mark ourselves as touched when
3409       // we propagate equalities).
3410       TouchedInstructions.reset(InstrNum);
3411 
3412       if (auto *MP = dyn_cast<MemoryPhi>(V)) {
3413         LLVM_DEBUG(dbgs() << "Processing MemoryPhi " << *MP << "\n");
3414         valueNumberMemoryPhi(MP);
3415       } else if (auto *I = dyn_cast<Instruction>(V)) {
3416         valueNumberInstruction(I);
3417       } else {
3418         llvm_unreachable("Should have been a MemoryPhi or Instruction");
3419       }
3420       updateProcessedCount(V);
3421     }
3422   }
3423   NumGVNMaxIterations = std::max(NumGVNMaxIterations.getValue(), Iterations);
3424 }
3425 
3426 // This is the main transformation entry point.
runGVN()3427 bool NewGVN::runGVN() {
3428   if (DebugCounter::isCounterSet(VNCounter))
3429     StartingVNCounter = DebugCounter::getCounterState(VNCounter);
3430   bool Changed = false;
3431   NumFuncArgs = F.arg_size();
3432   MSSAWalker = MSSA->getWalker();
3433   SingletonDeadExpression = new (ExpressionAllocator) DeadExpression();
3434 
3435   // Count number of instructions for sizing of hash tables, and come
3436   // up with a global dfs numbering for instructions.
3437   unsigned ICount = 1;
3438   // Add an empty instruction to account for the fact that we start at 1
3439   DFSToInstr.emplace_back(nullptr);
3440   // Note: We want ideal RPO traversal of the blocks, which is not quite the
3441   // same as dominator tree order, particularly with regard whether backedges
3442   // get visited first or second, given a block with multiple successors.
3443   // If we visit in the wrong order, we will end up performing N times as many
3444   // iterations.
3445   // The dominator tree does guarantee that, for a given dom tree node, it's
3446   // parent must occur before it in the RPO ordering. Thus, we only need to sort
3447   // the siblings.
3448   ReversePostOrderTraversal<Function *> RPOT(&F);
3449   unsigned Counter = 0;
3450   for (auto &B : RPOT) {
3451     auto *Node = DT->getNode(B);
3452     assert(Node && "RPO and Dominator tree should have same reachability");
3453     RPOOrdering[Node] = ++Counter;
3454   }
3455   // Sort dominator tree children arrays into RPO.
3456   for (auto &B : RPOT) {
3457     auto *Node = DT->getNode(B);
3458     if (Node->getNumChildren() > 1)
3459       llvm::sort(*Node, [&](const DomTreeNode *A, const DomTreeNode *B) {
3460         return RPOOrdering[A] < RPOOrdering[B];
3461       });
3462   }
3463 
3464   // Now a standard depth first ordering of the domtree is equivalent to RPO.
3465   for (auto *DTN : depth_first(DT->getRootNode())) {
3466     BasicBlock *B = DTN->getBlock();
3467     const auto &BlockRange = assignDFSNumbers(B, ICount);
3468     BlockInstRange.insert({B, BlockRange});
3469     ICount += BlockRange.second - BlockRange.first;
3470   }
3471   initializeCongruenceClasses(F);
3472 
3473   TouchedInstructions.resize(ICount);
3474   // Ensure we don't end up resizing the expressionToClass map, as
3475   // that can be quite expensive. At most, we have one expression per
3476   // instruction.
3477   ExpressionToClass.reserve(ICount);
3478 
3479   // Initialize the touched instructions to include the entry block.
3480   const auto &InstRange = BlockInstRange.lookup(&F.getEntryBlock());
3481   TouchedInstructions.set(InstRange.first, InstRange.second);
3482   LLVM_DEBUG(dbgs() << "Block " << getBlockName(&F.getEntryBlock())
3483                     << " marked reachable\n");
3484   ReachableBlocks.insert(&F.getEntryBlock());
3485   // Use index corresponding to entry block.
3486   CacheIdx = 0;
3487 
3488   iterateTouchedInstructions();
3489   verifyMemoryCongruency();
3490   verifyIterationSettled(F);
3491   verifyStoreExpressions();
3492 
3493   Changed |= eliminateInstructions(F);
3494 
3495   // Delete all instructions marked for deletion.
3496   for (Instruction *ToErase : InstructionsToErase) {
3497     if (!ToErase->use_empty())
3498       ToErase->replaceAllUsesWith(PoisonValue::get(ToErase->getType()));
3499 
3500     assert(ToErase->getParent() &&
3501            "BB containing ToErase deleted unexpectedly!");
3502     ToErase->eraseFromParent();
3503   }
3504   Changed |= !InstructionsToErase.empty();
3505 
3506   // Delete all unreachable blocks.
3507   auto UnreachableBlockPred = [&](const BasicBlock &BB) {
3508     return !ReachableBlocks.count(&BB);
3509   };
3510 
3511   for (auto &BB : make_filter_range(F, UnreachableBlockPred)) {
3512     LLVM_DEBUG(dbgs() << "We believe block " << getBlockName(&BB)
3513                       << " is unreachable\n");
3514     deleteInstructionsInBlock(&BB);
3515     Changed = true;
3516   }
3517 
3518   cleanupTables();
3519   return Changed;
3520 }
3521 
3522 struct NewGVN::ValueDFS {
3523   int DFSIn = 0;
3524   int DFSOut = 0;
3525   int LocalNum = 0;
3526 
3527   // Only one of Def and U will be set.
3528   // The bool in the Def tells us whether the Def is the stored value of a
3529   // store.
3530   PointerIntPair<Value *, 1, bool> Def;
3531   Use *U = nullptr;
3532 
operator <NewGVN::ValueDFS3533   bool operator<(const ValueDFS &Other) const {
3534     // It's not enough that any given field be less than - we have sets
3535     // of fields that need to be evaluated together to give a proper ordering.
3536     // For example, if you have;
3537     // DFS (1, 3)
3538     // Val 0
3539     // DFS (1, 2)
3540     // Val 50
3541     // We want the second to be less than the first, but if we just go field
3542     // by field, we will get to Val 0 < Val 50 and say the first is less than
3543     // the second. We only want it to be less than if the DFS orders are equal.
3544     //
3545     // Each LLVM instruction only produces one value, and thus the lowest-level
3546     // differentiator that really matters for the stack (and what we use as a
3547     // replacement) is the local dfs number.
3548     // Everything else in the structure is instruction level, and only affects
3549     // the order in which we will replace operands of a given instruction.
3550     //
3551     // For a given instruction (IE things with equal dfsin, dfsout, localnum),
3552     // the order of replacement of uses does not matter.
3553     // IE given,
3554     //  a = 5
3555     //  b = a + a
3556     // When you hit b, you will have two valuedfs with the same dfsin, out, and
3557     // localnum.
3558     // The .val will be the same as well.
3559     // The .u's will be different.
3560     // You will replace both, and it does not matter what order you replace them
3561     // in (IE whether you replace operand 2, then operand 1, or operand 1, then
3562     // operand 2).
3563     // Similarly for the case of same dfsin, dfsout, localnum, but different
3564     // .val's
3565     //  a = 5
3566     //  b  = 6
3567     //  c = a + b
3568     // in c, we will a valuedfs for a, and one for b,with everything the same
3569     // but .val  and .u.
3570     // It does not matter what order we replace these operands in.
3571     // You will always end up with the same IR, and this is guaranteed.
3572     return std::tie(DFSIn, DFSOut, LocalNum, Def, U) <
3573            std::tie(Other.DFSIn, Other.DFSOut, Other.LocalNum, Other.Def,
3574                     Other.U);
3575   }
3576 };
3577 
3578 // This function converts the set of members for a congruence class from values,
3579 // to sets of defs and uses with associated DFS info.  The total number of
3580 // reachable uses for each value is stored in UseCount, and instructions that
3581 // seem
3582 // dead (have no non-dead uses) are stored in ProbablyDead.
convertClassToDFSOrdered(const CongruenceClass & Dense,SmallVectorImpl<ValueDFS> & DFSOrderedSet,DenseMap<const Value *,unsigned int> & UseCounts,SmallPtrSetImpl<Instruction * > & ProbablyDead) const3583 void NewGVN::convertClassToDFSOrdered(
3584     const CongruenceClass &Dense, SmallVectorImpl<ValueDFS> &DFSOrderedSet,
3585     DenseMap<const Value *, unsigned int> &UseCounts,
3586     SmallPtrSetImpl<Instruction *> &ProbablyDead) const {
3587   for (auto *D : Dense) {
3588     // First add the value.
3589     BasicBlock *BB = getBlockForValue(D);
3590     // Constants are handled prior to ever calling this function, so
3591     // we should only be left with instructions as members.
3592     assert(BB && "Should have figured out a basic block for value");
3593     ValueDFS VDDef;
3594     DomTreeNode *DomNode = DT->getNode(BB);
3595     VDDef.DFSIn = DomNode->getDFSNumIn();
3596     VDDef.DFSOut = DomNode->getDFSNumOut();
3597     // If it's a store, use the leader of the value operand, if it's always
3598     // available, or the value operand.  TODO: We could do dominance checks to
3599     // find a dominating leader, but not worth it ATM.
3600     if (auto *SI = dyn_cast<StoreInst>(D)) {
3601       auto Leader = lookupOperandLeader(SI->getValueOperand());
3602       if (alwaysAvailable(Leader)) {
3603         VDDef.Def.setPointer(Leader);
3604       } else {
3605         VDDef.Def.setPointer(SI->getValueOperand());
3606         VDDef.Def.setInt(true);
3607       }
3608     } else {
3609       VDDef.Def.setPointer(D);
3610     }
3611     assert(isa<Instruction>(D) &&
3612            "The dense set member should always be an instruction");
3613     Instruction *Def = cast<Instruction>(D);
3614     VDDef.LocalNum = InstrToDFSNum(D);
3615     DFSOrderedSet.push_back(VDDef);
3616     // If there is a phi node equivalent, add it
3617     if (auto *PN = RealToTemp.lookup(Def)) {
3618       auto *PHIE =
3619           dyn_cast_or_null<PHIExpression>(ValueToExpression.lookup(Def));
3620       if (PHIE) {
3621         VDDef.Def.setInt(false);
3622         VDDef.Def.setPointer(PN);
3623         VDDef.LocalNum = 0;
3624         DFSOrderedSet.push_back(VDDef);
3625       }
3626     }
3627 
3628     unsigned int UseCount = 0;
3629     // Now add the uses.
3630     for (auto &U : Def->uses()) {
3631       if (auto *I = dyn_cast<Instruction>(U.getUser())) {
3632         // Don't try to replace into dead uses
3633         if (InstructionsToErase.count(I))
3634           continue;
3635         ValueDFS VDUse;
3636         // Put the phi node uses in the incoming block.
3637         BasicBlock *IBlock;
3638         if (auto *P = dyn_cast<PHINode>(I)) {
3639           IBlock = P->getIncomingBlock(U);
3640           // Make phi node users appear last in the incoming block
3641           // they are from.
3642           VDUse.LocalNum = InstrDFS.size() + 1;
3643         } else {
3644           IBlock = getBlockForValue(I);
3645           VDUse.LocalNum = InstrToDFSNum(I);
3646         }
3647 
3648         // Skip uses in unreachable blocks, as we're going
3649         // to delete them.
3650         if (!ReachableBlocks.contains(IBlock))
3651           continue;
3652 
3653         DomTreeNode *DomNode = DT->getNode(IBlock);
3654         VDUse.DFSIn = DomNode->getDFSNumIn();
3655         VDUse.DFSOut = DomNode->getDFSNumOut();
3656         VDUse.U = &U;
3657         ++UseCount;
3658         DFSOrderedSet.emplace_back(VDUse);
3659       }
3660     }
3661 
3662     // If there are no uses, it's probably dead (but it may have side-effects,
3663     // so not definitely dead. Otherwise, store the number of uses so we can
3664     // track if it becomes dead later).
3665     if (UseCount == 0)
3666       ProbablyDead.insert(Def);
3667     else
3668       UseCounts[Def] = UseCount;
3669   }
3670 }
3671 
3672 // This function converts the set of members for a congruence class from values,
3673 // to the set of defs for loads and stores, with associated DFS info.
convertClassToLoadsAndStores(const CongruenceClass & Dense,SmallVectorImpl<ValueDFS> & LoadsAndStores) const3674 void NewGVN::convertClassToLoadsAndStores(
3675     const CongruenceClass &Dense,
3676     SmallVectorImpl<ValueDFS> &LoadsAndStores) const {
3677   for (auto *D : Dense) {
3678     if (!isa<LoadInst>(D) && !isa<StoreInst>(D))
3679       continue;
3680 
3681     BasicBlock *BB = getBlockForValue(D);
3682     ValueDFS VD;
3683     DomTreeNode *DomNode = DT->getNode(BB);
3684     VD.DFSIn = DomNode->getDFSNumIn();
3685     VD.DFSOut = DomNode->getDFSNumOut();
3686     VD.Def.setPointer(D);
3687 
3688     // If it's an instruction, use the real local dfs number.
3689     if (auto *I = dyn_cast<Instruction>(D))
3690       VD.LocalNum = InstrToDFSNum(I);
3691     else
3692       llvm_unreachable("Should have been an instruction");
3693 
3694     LoadsAndStores.emplace_back(VD);
3695   }
3696 }
3697 
patchAndReplaceAllUsesWith(Instruction * I,Value * Repl)3698 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
3699   patchReplacementInstruction(I, Repl);
3700   I->replaceAllUsesWith(Repl);
3701 }
3702 
deleteInstructionsInBlock(BasicBlock * BB)3703 void NewGVN::deleteInstructionsInBlock(BasicBlock *BB) {
3704   LLVM_DEBUG(dbgs() << "  BasicBlock Dead:" << *BB);
3705   ++NumGVNBlocksDeleted;
3706 
3707   // Delete the instructions backwards, as it has a reduced likelihood of having
3708   // to update as many def-use and use-def chains. Start after the terminator.
3709   auto StartPoint = BB->rbegin();
3710   ++StartPoint;
3711   // Note that we explicitly recalculate BB->rend() on each iteration,
3712   // as it may change when we remove the first instruction.
3713   for (BasicBlock::reverse_iterator I(StartPoint); I != BB->rend();) {
3714     Instruction &Inst = *I++;
3715     if (!Inst.use_empty())
3716       Inst.replaceAllUsesWith(PoisonValue::get(Inst.getType()));
3717     if (isa<LandingPadInst>(Inst))
3718       continue;
3719     salvageKnowledge(&Inst, AC);
3720 
3721     Inst.eraseFromParent();
3722     ++NumGVNInstrDeleted;
3723   }
3724   // Now insert something that simplifycfg will turn into an unreachable.
3725   Type *Int8Ty = Type::getInt8Ty(BB->getContext());
3726   new StoreInst(
3727       PoisonValue::get(Int8Ty),
3728       Constant::getNullValue(PointerType::getUnqual(BB->getContext())),
3729       BB->getTerminator()->getIterator());
3730 }
3731 
markInstructionForDeletion(Instruction * I)3732 void NewGVN::markInstructionForDeletion(Instruction *I) {
3733   LLVM_DEBUG(dbgs() << "Marking " << *I << " for deletion\n");
3734   InstructionsToErase.insert(I);
3735 }
3736 
replaceInstruction(Instruction * I,Value * V)3737 void NewGVN::replaceInstruction(Instruction *I, Value *V) {
3738   LLVM_DEBUG(dbgs() << "Replacing " << *I << " with " << *V << "\n");
3739   patchAndReplaceAllUsesWith(I, V);
3740   // We save the actual erasing to avoid invalidating memory
3741   // dependencies until we are done with everything.
3742   markInstructionForDeletion(I);
3743 }
3744 
3745 namespace {
3746 
3747 // This is a stack that contains both the value and dfs info of where
3748 // that value is valid.
3749 class ValueDFSStack {
3750 public:
back() const3751   Value *back() const { return ValueStack.back(); }
dfs_back() const3752   std::pair<int, int> dfs_back() const { return DFSStack.back(); }
3753 
push_back(Value * V,int DFSIn,int DFSOut)3754   void push_back(Value *V, int DFSIn, int DFSOut) {
3755     ValueStack.emplace_back(V);
3756     DFSStack.emplace_back(DFSIn, DFSOut);
3757   }
3758 
empty() const3759   bool empty() const { return DFSStack.empty(); }
3760 
isInScope(int DFSIn,int DFSOut) const3761   bool isInScope(int DFSIn, int DFSOut) const {
3762     if (empty())
3763       return false;
3764     return DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second;
3765   }
3766 
popUntilDFSScope(int DFSIn,int DFSOut)3767   void popUntilDFSScope(int DFSIn, int DFSOut) {
3768 
3769     // These two should always be in sync at this point.
3770     assert(ValueStack.size() == DFSStack.size() &&
3771            "Mismatch between ValueStack and DFSStack");
3772     while (
3773         !DFSStack.empty() &&
3774         !(DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second)) {
3775       DFSStack.pop_back();
3776       ValueStack.pop_back();
3777     }
3778   }
3779 
3780 private:
3781   SmallVector<Value *, 8> ValueStack;
3782   SmallVector<std::pair<int, int>, 8> DFSStack;
3783 };
3784 
3785 } // end anonymous namespace
3786 
3787 // Given an expression, get the congruence class for it.
getClassForExpression(const Expression * E) const3788 CongruenceClass *NewGVN::getClassForExpression(const Expression *E) const {
3789   if (auto *VE = dyn_cast<VariableExpression>(E))
3790     return ValueToClass.lookup(VE->getVariableValue());
3791   else if (isa<DeadExpression>(E))
3792     return TOPClass;
3793   return ExpressionToClass.lookup(E);
3794 }
3795 
3796 // Given a value and a basic block we are trying to see if it is available in,
3797 // see if the value has a leader available in that block.
findPHIOfOpsLeader(const Expression * E,const Instruction * OrigInst,const BasicBlock * BB) const3798 Value *NewGVN::findPHIOfOpsLeader(const Expression *E,
3799                                   const Instruction *OrigInst,
3800                                   const BasicBlock *BB) const {
3801   // It would already be constant if we could make it constant
3802   if (auto *CE = dyn_cast<ConstantExpression>(E))
3803     return CE->getConstantValue();
3804   if (auto *VE = dyn_cast<VariableExpression>(E)) {
3805     auto *V = VE->getVariableValue();
3806     if (alwaysAvailable(V) || DT->dominates(getBlockForValue(V), BB))
3807       return VE->getVariableValue();
3808   }
3809 
3810   auto *CC = getClassForExpression(E);
3811   if (!CC)
3812     return nullptr;
3813   if (alwaysAvailable(CC->getLeader()))
3814     return CC->getLeader();
3815 
3816   for (auto *Member : *CC) {
3817     auto *MemberInst = dyn_cast<Instruction>(Member);
3818     if (MemberInst == OrigInst)
3819       continue;
3820     // Anything that isn't an instruction is always available.
3821     if (!MemberInst)
3822       return Member;
3823     if (DT->dominates(getBlockForValue(MemberInst), BB))
3824       return Member;
3825   }
3826   return nullptr;
3827 }
3828 
eliminateInstructions(Function & F)3829 bool NewGVN::eliminateInstructions(Function &F) {
3830   // This is a non-standard eliminator. The normal way to eliminate is
3831   // to walk the dominator tree in order, keeping track of available
3832   // values, and eliminating them.  However, this is mildly
3833   // pointless. It requires doing lookups on every instruction,
3834   // regardless of whether we will ever eliminate it.  For
3835   // instructions part of most singleton congruence classes, we know we
3836   // will never eliminate them.
3837 
3838   // Instead, this eliminator looks at the congruence classes directly, sorts
3839   // them into a DFS ordering of the dominator tree, and then we just
3840   // perform elimination straight on the sets by walking the congruence
3841   // class member uses in order, and eliminate the ones dominated by the
3842   // last member.   This is worst case O(E log E) where E = number of
3843   // instructions in a single congruence class.  In theory, this is all
3844   // instructions.   In practice, it is much faster, as most instructions are
3845   // either in singleton congruence classes or can't possibly be eliminated
3846   // anyway (if there are no overlapping DFS ranges in class).
3847   // When we find something not dominated, it becomes the new leader
3848   // for elimination purposes.
3849   // TODO: If we wanted to be faster, We could remove any members with no
3850   // overlapping ranges while sorting, as we will never eliminate anything
3851   // with those members, as they don't dominate anything else in our set.
3852 
3853   bool AnythingReplaced = false;
3854 
3855   // Since we are going to walk the domtree anyway, and we can't guarantee the
3856   // DFS numbers are updated, we compute some ourselves.
3857   DT->updateDFSNumbers();
3858 
3859   // Go through all of our phi nodes, and kill the arguments associated with
3860   // unreachable edges.
3861   auto ReplaceUnreachablePHIArgs = [&](PHINode *PHI, BasicBlock *BB) {
3862     for (auto &Operand : PHI->incoming_values())
3863       if (!ReachableEdges.count({PHI->getIncomingBlock(Operand), BB})) {
3864         LLVM_DEBUG(dbgs() << "Replacing incoming value of " << PHI
3865                           << " for block "
3866                           << getBlockName(PHI->getIncomingBlock(Operand))
3867                           << " with poison due to it being unreachable\n");
3868         Operand.set(PoisonValue::get(PHI->getType()));
3869       }
3870   };
3871   // Replace unreachable phi arguments.
3872   // At this point, RevisitOnReachabilityChange only contains:
3873   //
3874   // 1. PHIs
3875   // 2. Temporaries that will convert to PHIs
3876   // 3. Operations that are affected by an unreachable edge but do not fit into
3877   // 1 or 2 (rare).
3878   // So it is a slight overshoot of what we want. We could make it exact by
3879   // using two SparseBitVectors per block.
3880   DenseMap<const BasicBlock *, unsigned> ReachablePredCount;
3881   for (auto &KV : ReachableEdges)
3882     ReachablePredCount[KV.getEnd()]++;
3883   for (auto &BBPair : RevisitOnReachabilityChange) {
3884     for (auto InstNum : BBPair.second) {
3885       auto *Inst = InstrFromDFSNum(InstNum);
3886       auto *PHI = dyn_cast<PHINode>(Inst);
3887       PHI = PHI ? PHI : dyn_cast_or_null<PHINode>(RealToTemp.lookup(Inst));
3888       if (!PHI)
3889         continue;
3890       auto *BB = BBPair.first;
3891       if (ReachablePredCount.lookup(BB) != PHI->getNumIncomingValues())
3892         ReplaceUnreachablePHIArgs(PHI, BB);
3893     }
3894   }
3895 
3896   // Map to store the use counts
3897   DenseMap<const Value *, unsigned int> UseCounts;
3898   for (auto *CC : reverse(CongruenceClasses)) {
3899     LLVM_DEBUG(dbgs() << "Eliminating in congruence class " << CC->getID()
3900                       << "\n");
3901     // Track the equivalent store info so we can decide whether to try
3902     // dead store elimination.
3903     SmallVector<ValueDFS, 8> PossibleDeadStores;
3904     SmallPtrSet<Instruction *, 8> ProbablyDead;
3905     if (CC->isDead() || CC->empty())
3906       continue;
3907     // Everything still in the TOP class is unreachable or dead.
3908     if (CC == TOPClass) {
3909       for (auto *M : *CC) {
3910         auto *VTE = ValueToExpression.lookup(M);
3911         if (VTE && isa<DeadExpression>(VTE))
3912           markInstructionForDeletion(cast<Instruction>(M));
3913         assert((!ReachableBlocks.count(cast<Instruction>(M)->getParent()) ||
3914                 InstructionsToErase.count(cast<Instruction>(M))) &&
3915                "Everything in TOP should be unreachable or dead at this "
3916                "point");
3917       }
3918       continue;
3919     }
3920 
3921     assert(CC->getLeader() && "We should have had a leader");
3922     // If this is a leader that is always available, and it's a
3923     // constant or has no equivalences, just replace everything with
3924     // it. We then update the congruence class with whatever members
3925     // are left.
3926     Value *Leader =
3927         CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
3928     if (alwaysAvailable(Leader)) {
3929       CongruenceClass::MemberSet MembersLeft;
3930       for (auto *M : *CC) {
3931         Value *Member = M;
3932         // Void things have no uses we can replace.
3933         if (Member == Leader || !isa<Instruction>(Member) ||
3934             Member->getType()->isVoidTy()) {
3935           MembersLeft.insert(Member);
3936           continue;
3937         }
3938         LLVM_DEBUG(dbgs() << "Found replacement " << *(Leader) << " for "
3939                           << *Member << "\n");
3940         auto *I = cast<Instruction>(Member);
3941         assert(Leader != I && "About to accidentally remove our leader");
3942         replaceInstruction(I, Leader);
3943         AnythingReplaced = true;
3944       }
3945       CC->swap(MembersLeft);
3946     } else {
3947       // If this is a singleton, we can skip it.
3948       if (CC->size() != 1 || RealToTemp.count(Leader)) {
3949         // This is a stack because equality replacement/etc may place
3950         // constants in the middle of the member list, and we want to use
3951         // those constant values in preference to the current leader, over
3952         // the scope of those constants.
3953         ValueDFSStack EliminationStack;
3954 
3955         // Convert the members to DFS ordered sets and then merge them.
3956         SmallVector<ValueDFS, 8> DFSOrderedSet;
3957         convertClassToDFSOrdered(*CC, DFSOrderedSet, UseCounts, ProbablyDead);
3958 
3959         // Sort the whole thing.
3960         llvm::sort(DFSOrderedSet);
3961         for (auto &VD : DFSOrderedSet) {
3962           int MemberDFSIn = VD.DFSIn;
3963           int MemberDFSOut = VD.DFSOut;
3964           Value *Def = VD.Def.getPointer();
3965           bool FromStore = VD.Def.getInt();
3966           Use *U = VD.U;
3967           // We ignore void things because we can't get a value from them.
3968           if (Def && Def->getType()->isVoidTy())
3969             continue;
3970           auto *DefInst = dyn_cast_or_null<Instruction>(Def);
3971           if (DefInst && AllTempInstructions.count(DefInst)) {
3972             auto *PN = cast<PHINode>(DefInst);
3973 
3974             // If this is a value phi and that's the expression we used, insert
3975             // it into the program
3976             // remove from temp instruction list.
3977             AllTempInstructions.erase(PN);
3978             auto *DefBlock = getBlockForValue(Def);
3979             LLVM_DEBUG(dbgs() << "Inserting fully real phi of ops" << *Def
3980                               << " into block "
3981                               << getBlockName(getBlockForValue(Def)) << "\n");
3982             PN->insertBefore(&DefBlock->front());
3983             Def = PN;
3984             NumGVNPHIOfOpsEliminations++;
3985           }
3986 
3987           if (EliminationStack.empty()) {
3988             LLVM_DEBUG(dbgs() << "Elimination Stack is empty\n");
3989           } else {
3990             LLVM_DEBUG(dbgs() << "Elimination Stack Top DFS numbers are ("
3991                               << EliminationStack.dfs_back().first << ","
3992                               << EliminationStack.dfs_back().second << ")\n");
3993           }
3994 
3995           LLVM_DEBUG(dbgs() << "Current DFS numbers are (" << MemberDFSIn << ","
3996                             << MemberDFSOut << ")\n");
3997           // First, we see if we are out of scope or empty.  If so,
3998           // and there equivalences, we try to replace the top of
3999           // stack with equivalences (if it's on the stack, it must
4000           // not have been eliminated yet).
4001           // Then we synchronize to our current scope, by
4002           // popping until we are back within a DFS scope that
4003           // dominates the current member.
4004           // Then, what happens depends on a few factors
4005           // If the stack is now empty, we need to push
4006           // If we have a constant or a local equivalence we want to
4007           // start using, we also push.
4008           // Otherwise, we walk along, processing members who are
4009           // dominated by this scope, and eliminate them.
4010           bool ShouldPush = Def && EliminationStack.empty();
4011           bool OutOfScope =
4012               !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut);
4013 
4014           if (OutOfScope || ShouldPush) {
4015             // Sync to our current scope.
4016             EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4017             bool ShouldPush = Def && EliminationStack.empty();
4018             if (ShouldPush) {
4019               EliminationStack.push_back(Def, MemberDFSIn, MemberDFSOut);
4020             }
4021           }
4022 
4023           // Skip the Def's, we only want to eliminate on their uses.  But mark
4024           // dominated defs as dead.
4025           if (Def) {
4026             // For anything in this case, what and how we value number
4027             // guarantees that any side-effects that would have occurred (ie
4028             // throwing, etc) can be proven to either still occur (because it's
4029             // dominated by something that has the same side-effects), or never
4030             // occur.  Otherwise, we would not have been able to prove it value
4031             // equivalent to something else. For these things, we can just mark
4032             // it all dead.  Note that this is different from the "ProbablyDead"
4033             // set, which may not be dominated by anything, and thus, are only
4034             // easy to prove dead if they are also side-effect free. Note that
4035             // because stores are put in terms of the stored value, we skip
4036             // stored values here. If the stored value is really dead, it will
4037             // still be marked for deletion when we process it in its own class.
4038             auto *DefI = dyn_cast<Instruction>(Def);
4039             if (!EliminationStack.empty() && DefI && !FromStore) {
4040               Value *DominatingLeader = EliminationStack.back();
4041               if (DominatingLeader != Def) {
4042                 // Even if the instruction is removed, we still need to update
4043                 // flags/metadata due to downstreams users of the leader.
4044                 if (!match(DefI, m_Intrinsic<Intrinsic::ssa_copy>()))
4045                   patchReplacementInstruction(DefI, DominatingLeader);
4046 
4047                 markInstructionForDeletion(DefI);
4048               }
4049             }
4050             continue;
4051           }
4052           // At this point, we know it is a Use we are trying to possibly
4053           // replace.
4054 
4055           assert(isa<Instruction>(U->get()) &&
4056                  "Current def should have been an instruction");
4057           assert(isa<Instruction>(U->getUser()) &&
4058                  "Current user should have been an instruction");
4059 
4060           // If the thing we are replacing into is already marked to be dead,
4061           // this use is dead.  Note that this is true regardless of whether
4062           // we have anything dominating the use or not.  We do this here
4063           // because we are already walking all the uses anyway.
4064           Instruction *InstUse = cast<Instruction>(U->getUser());
4065           if (InstructionsToErase.count(InstUse)) {
4066             auto &UseCount = UseCounts[U->get()];
4067             if (--UseCount == 0) {
4068               ProbablyDead.insert(cast<Instruction>(U->get()));
4069             }
4070           }
4071 
4072           // If we get to this point, and the stack is empty we must have a use
4073           // with nothing we can use to eliminate this use, so just skip it.
4074           if (EliminationStack.empty())
4075             continue;
4076 
4077           Value *DominatingLeader = EliminationStack.back();
4078 
4079           auto *II = dyn_cast<IntrinsicInst>(DominatingLeader);
4080           bool isSSACopy = II && II->getIntrinsicID() == Intrinsic::ssa_copy;
4081           if (isSSACopy)
4082             DominatingLeader = II->getOperand(0);
4083 
4084           // Don't replace our existing users with ourselves.
4085           if (U->get() == DominatingLeader)
4086             continue;
4087           LLVM_DEBUG(dbgs()
4088                      << "Found replacement " << *DominatingLeader << " for "
4089                      << *U->get() << " in " << *(U->getUser()) << "\n");
4090 
4091           // If we replaced something in an instruction, handle the patching of
4092           // metadata.  Skip this if we are replacing predicateinfo with its
4093           // original operand, as we already know we can just drop it.
4094           auto *ReplacedInst = cast<Instruction>(U->get());
4095           auto *PI = PredInfo->getPredicateInfoFor(ReplacedInst);
4096           if (!PI || DominatingLeader != PI->OriginalOp)
4097             patchReplacementInstruction(ReplacedInst, DominatingLeader);
4098           U->set(DominatingLeader);
4099           // This is now a use of the dominating leader, which means if the
4100           // dominating leader was dead, it's now live!
4101           auto &LeaderUseCount = UseCounts[DominatingLeader];
4102           // It's about to be alive again.
4103           if (LeaderUseCount == 0 && isa<Instruction>(DominatingLeader))
4104             ProbablyDead.erase(cast<Instruction>(DominatingLeader));
4105           // For copy instructions, we use their operand as a leader,
4106           // which means we remove a user of the copy and it may become dead.
4107           if (isSSACopy) {
4108             auto It = UseCounts.find(II);
4109             if (It != UseCounts.end()) {
4110               unsigned &IIUseCount = It->second;
4111               if (--IIUseCount == 0)
4112                 ProbablyDead.insert(II);
4113             }
4114           }
4115           ++LeaderUseCount;
4116           AnythingReplaced = true;
4117         }
4118       }
4119     }
4120 
4121     // At this point, anything still in the ProbablyDead set is actually dead if
4122     // would be trivially dead.
4123     for (auto *I : ProbablyDead)
4124       if (wouldInstructionBeTriviallyDead(I))
4125         markInstructionForDeletion(I);
4126 
4127     // Cleanup the congruence class.
4128     CongruenceClass::MemberSet MembersLeft;
4129     for (auto *Member : *CC)
4130       if (!isa<Instruction>(Member) ||
4131           !InstructionsToErase.count(cast<Instruction>(Member)))
4132         MembersLeft.insert(Member);
4133     CC->swap(MembersLeft);
4134 
4135     // If we have possible dead stores to look at, try to eliminate them.
4136     if (CC->getStoreCount() > 0) {
4137       convertClassToLoadsAndStores(*CC, PossibleDeadStores);
4138       llvm::sort(PossibleDeadStores);
4139       ValueDFSStack EliminationStack;
4140       for (auto &VD : PossibleDeadStores) {
4141         int MemberDFSIn = VD.DFSIn;
4142         int MemberDFSOut = VD.DFSOut;
4143         Instruction *Member = cast<Instruction>(VD.Def.getPointer());
4144         if (EliminationStack.empty() ||
4145             !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut)) {
4146           // Sync to our current scope.
4147           EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4148           if (EliminationStack.empty()) {
4149             EliminationStack.push_back(Member, MemberDFSIn, MemberDFSOut);
4150             continue;
4151           }
4152         }
4153         // We already did load elimination, so nothing to do here.
4154         if (isa<LoadInst>(Member))
4155           continue;
4156         assert(!EliminationStack.empty());
4157         Instruction *Leader = cast<Instruction>(EliminationStack.back());
4158         (void)Leader;
4159         assert(DT->dominates(Leader->getParent(), Member->getParent()));
4160         // Member is dominater by Leader, and thus dead
4161         LLVM_DEBUG(dbgs() << "Marking dead store " << *Member
4162                           << " that is dominated by " << *Leader << "\n");
4163         markInstructionForDeletion(Member);
4164         CC->erase(Member);
4165         ++NumGVNDeadStores;
4166       }
4167     }
4168   }
4169   return AnythingReplaced;
4170 }
4171 
4172 // This function provides global ranking of operations so that we can place them
4173 // in a canonical order.  Note that rank alone is not necessarily enough for a
4174 // complete ordering, as constants all have the same rank.  However, generally,
4175 // we will simplify an operation with all constants so that it doesn't matter
4176 // what order they appear in.
getRank(const Value * V) const4177 unsigned int NewGVN::getRank(const Value *V) const {
4178   // Prefer constants to undef to anything else
4179   // Undef is a constant, have to check it first.
4180   // Prefer poison to undef as it's less defined.
4181   // Prefer smaller constants to constantexprs
4182   // Note that the order here matters because of class inheritance
4183   if (isa<ConstantExpr>(V))
4184     return 3;
4185   if (isa<PoisonValue>(V))
4186     return 1;
4187   if (isa<UndefValue>(V))
4188     return 2;
4189   if (isa<Constant>(V))
4190     return 0;
4191   if (auto *A = dyn_cast<Argument>(V))
4192     return 4 + A->getArgNo();
4193 
4194   // Need to shift the instruction DFS by number of arguments + 5 to account for
4195   // the constant and argument ranking above.
4196   unsigned Result = InstrToDFSNum(V);
4197   if (Result > 0)
4198     return 5 + NumFuncArgs + Result;
4199   // Unreachable or something else, just return a really large number.
4200   return ~0;
4201 }
4202 
4203 // This is a function that says whether two commutative operations should
4204 // have their order swapped when canonicalizing.
shouldSwapOperands(const Value * A,const Value * B) const4205 bool NewGVN::shouldSwapOperands(const Value *A, const Value *B) const {
4206   // Because we only care about a total ordering, and don't rewrite expressions
4207   // in this order, we order by rank, which will give a strict weak ordering to
4208   // everything but constants, and then we order by pointer address.
4209   return std::make_pair(getRank(A), A) > std::make_pair(getRank(B), B);
4210 }
4211 
shouldSwapOperandsForIntrinsic(const Value * A,const Value * B,const IntrinsicInst * I) const4212 bool NewGVN::shouldSwapOperandsForIntrinsic(const Value *A, const Value *B,
4213                                             const IntrinsicInst *I) const {
4214   auto LookupResult = IntrinsicInstPred.find(I);
4215   if (shouldSwapOperands(A, B)) {
4216     if (LookupResult == IntrinsicInstPred.end())
4217       IntrinsicInstPred.insert({I, B});
4218     else
4219       LookupResult->second = B;
4220     return true;
4221   }
4222 
4223   if (LookupResult != IntrinsicInstPred.end()) {
4224     auto *SeenPredicate = LookupResult->second;
4225     if (SeenPredicate) {
4226       if (SeenPredicate == B)
4227         return true;
4228       else
4229         LookupResult->second = nullptr;
4230     }
4231   }
4232   return false;
4233 }
4234 
run(Function & F,AnalysisManager<Function> & AM)4235 PreservedAnalyses NewGVNPass::run(Function &F, AnalysisManager<Function> &AM) {
4236   // Apparently the order in which we get these results matter for
4237   // the old GVN (see Chandler's comment in GVN.cpp). I'll keep
4238   // the same order here, just in case.
4239   auto &AC = AM.getResult<AssumptionAnalysis>(F);
4240   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
4241   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
4242   auto &AA = AM.getResult<AAManager>(F);
4243   auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
4244   bool Changed =
4245       NewGVN(F, &DT, &AC, &TLI, &AA, &MSSA, F.getDataLayout())
4246           .runGVN();
4247   if (!Changed)
4248     return PreservedAnalyses::all();
4249   PreservedAnalyses PA;
4250   PA.preserve<DominatorTreeAnalysis>();
4251   return PA;
4252 }
4253