LLVM  6.0.0svn
NewGVN.cpp
Go to the documentation of this file.
1 //===- NewGVN.cpp - Global Value Numbering Pass ---------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// This file implements the new LLVM's Global Value Numbering pass.
12 /// GVN partitions values computed by a function into congruence classes.
13 /// Values ending up in the same congruence class are guaranteed to be the same
14 /// for every execution of the program. In that respect, congruency is a
15 /// compile-time approximation of equivalence of values at runtime.
16 /// The algorithm implemented here uses a sparse formulation and it's based
17 /// on the ideas described in the paper:
18 /// "A Sparse Algorithm for Predicated Global Value Numbering" from
19 /// Karthik Gargi.
20 ///
21 /// A brief overview of the algorithm: The algorithm is essentially the same as
22 /// the standard RPO value numbering algorithm (a good reference is the paper
23 /// "SCC based value numbering" by L. Taylor Simpson) with one major difference:
24 /// The RPO algorithm proceeds, on every iteration, to process every reachable
25 /// block and every instruction in that block. This is because the standard RPO
26 /// algorithm does not track what things have the same value number, it only
27 /// tracks what the value number of a given operation is (the mapping is
28 /// operation -> value number). Thus, when a value number of an operation
29 /// changes, it must reprocess everything to ensure all uses of a value number
30 /// get updated properly. In constrast, the sparse algorithm we use *also*
31 /// tracks what operations have a given value number (IE it also tracks the
32 /// reverse mapping from value number -> operations with that value number), so
33 /// that it only needs to reprocess the instructions that are affected when
34 /// something's value number changes. The vast majority of complexity and code
35 /// in this file is devoted to tracking what value numbers could change for what
36 /// instructions when various things happen. The rest of the algorithm is
37 /// devoted to performing symbolic evaluation, forward propagation, and
38 /// simplification of operations based on the value numbers deduced so far
39 ///
40 /// In order to make the GVN mostly-complete, we use a technique derived from
41 /// "Detection of Redundant Expressions: A Complete and Polynomial-time
42 /// Algorithm in SSA" by R.R. Pai. The source of incompleteness in most SSA
43 /// based GVN algorithms is related to their inability to detect equivalence
44 /// between phi of ops (IE phi(a+b, c+d)) and op of phis (phi(a,c) + phi(b, d)).
45 /// We resolve this issue by generating the equivalent "phi of ops" form for
46 /// each op of phis we see, in a way that only takes polynomial time to resolve.
47 ///
48 /// We also do not perform elimination by using any published algorithm. All
49 /// published algorithms are O(Instructions). Instead, we use a technique that
50 /// is O(number of operations with the same value number), enabling us to skip
51 /// trying to eliminate things that have unique value numbers.
52 //
53 //===----------------------------------------------------------------------===//
54 
56 #include "llvm/ADT/ArrayRef.h"
57 #include "llvm/ADT/BitVector.h"
58 #include "llvm/ADT/DenseMap.h"
59 #include "llvm/ADT/DenseMapInfo.h"
60 #include "llvm/ADT/DenseSet.h"
62 #include "llvm/ADT/GraphTraits.h"
63 #include "llvm/ADT/Hashing.h"
66 #include "llvm/ADT/SmallPtrSet.h"
67 #include "llvm/ADT/SmallVector.h"
69 #include "llvm/ADT/Statistic.h"
80 #include "llvm/IR/Argument.h"
81 #include "llvm/IR/BasicBlock.h"
82 #include "llvm/IR/Constant.h"
83 #include "llvm/IR/Constants.h"
84 #include "llvm/IR/Dominators.h"
85 #include "llvm/IR/Function.h"
86 #include "llvm/IR/InstrTypes.h"
87 #include "llvm/IR/Instruction.h"
88 #include "llvm/IR/Instructions.h"
89 #include "llvm/IR/IntrinsicInst.h"
90 #include "llvm/IR/Intrinsics.h"
91 #include "llvm/IR/LLVMContext.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/Pass.h"
97 #include "llvm/Support/Allocator.h"
99 #include "llvm/Support/Casting.h"
101 #include "llvm/Support/Debug.h"
106 #include "llvm/Transforms/Scalar.h"
111 #include <algorithm>
112 #include <cassert>
113 #include <cstdint>
114 #include <iterator>
115 #include <map>
116 #include <memory>
117 #include <set>
118 #include <string>
119 #include <tuple>
120 #include <utility>
121 #include <vector>
122 
123 using namespace llvm;
124 using namespace llvm::GVNExpression;
125 using namespace llvm::VNCoercion;
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;
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 {
187  TarjanSCC() : Components(1) {}
188 
189  void Start(const Instruction *Start) {
190  if (Root.lookup(Start) == 0)
191  FindSCC(Start);
192  }
193 
194  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:
203  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 (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  DEBUG(dbgs() << "Component root is " << *I << "\n");
225  InComponent.insert(I);
226  ValueToComponent[I] = ComponentID;
227  // Pop a component off the stack and label it.
228  while (!Stack.empty() && Root.lookup(Stack.back()) >= OurDFS) {
229  auto *Member = Stack.back();
230  DEBUG(dbgs() << "Component member is " << *Member << "\n");
231  Component.insert(Member);
232  InComponent.insert(Member);
233  ValueToComponent[Member] = ComponentID;
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;
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
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 
299  explicit CongruenceClass(unsigned ID) : ID(ID) {}
300  CongruenceClass(unsigned ID, Value *Leader, const Expression *E)
301  : ID(ID), RepLeader(Leader), DefiningExpr(E) {}
302 
303  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.
307  bool isDead() const {
308  // If it's both dead from a value perspective, and dead from a memory
309  // perspective, it's really dead.
310  return empty() && memory_empty();
311  }
312 
313  // Leader functions
314  Value *getLeader() const { return RepLeader; }
315  void setLeader(Value *Leader) { RepLeader = Leader; }
316  const std::pair<Value *, unsigned int> &getNextLeader() const {
317  return NextLeader;
318  }
319  void resetNextLeader() { NextLeader = {nullptr, ~0}; }
320  void addPossibleNextLeader(std::pair<Value *, unsigned int> LeaderPair) {
321  if (LeaderPair.second < NextLeader.second)
322  NextLeader = LeaderPair;
323  }
324 
325  Value *getStoredValue() const { return RepStoredValue; }
326  void setStoredValue(Value *Leader) { RepStoredValue = Leader; }
327  const MemoryAccess *getMemoryLeader() const { return RepMemoryAccess; }
328  void setMemoryLeader(const MemoryAccess *Leader) { RepMemoryAccess = Leader; }
329 
330  // Forward propagation info
331  const Expression *getDefiningExpr() const { return DefiningExpr; }
332 
333  // Value member set
334  bool empty() const { return Members.empty(); }
335  unsigned size() const { return Members.size(); }
336  MemberSet::const_iterator begin() const { return Members.begin(); }
337  MemberSet::const_iterator end() const { return Members.end(); }
338  void insert(MemberType *M) { Members.insert(M); }
339  void erase(MemberType *M) { Members.erase(M); }
340  void swap(MemberSet &Other) { Members.swap(Other); }
341 
342  // Memory member set
343  bool memory_empty() const { return MemoryMembers.empty(); }
344  unsigned memory_size() const { return MemoryMembers.size(); }
345  MemoryMemberSet::const_iterator memory_begin() const {
346  return MemoryMembers.begin();
347  }
348  MemoryMemberSet::const_iterator memory_end() const {
349  return MemoryMembers.end();
350  }
352  return make_range(memory_begin(), memory_end());
353  }
354 
355  void memory_insert(const MemoryMemberType *M) { MemoryMembers.insert(M); }
356  void memory_erase(const MemoryMemberType *M) { MemoryMembers.erase(M); }
357 
358  // Store count
359  unsigned getStoreCount() const { return StoreCount; }
360  void incStoreCount() { ++StoreCount; }
361  void decStoreCount() {
362  assert(StoreCount != 0 && "Store count went negative");
363  --StoreCount;
364  }
365 
366  // True if this class has no memory members.
367  bool definesNoMemory() const { return StoreCount == 0 && memory_empty(); }
368 
369  // Return true if two congruence classes are equivalent to each other. This
370  // means
371  // that every field but the ID number and the dead field are equivalent.
372  bool isEquivalentTo(const CongruenceClass *Other) const {
373  if (!Other)
374  return false;
375  if (this == Other)
376  return true;
377 
378  if (std::tie(StoreCount, RepLeader, RepStoredValue, RepMemoryAccess) !=
379  std::tie(Other->StoreCount, Other->RepLeader, Other->RepStoredValue,
380  Other->RepMemoryAccess))
381  return false;
382  if (DefiningExpr != Other->DefiningExpr)
383  if (!DefiningExpr || !Other->DefiningExpr ||
384  *DefiningExpr != *Other->DefiningExpr)
385  return false;
386  // We need some ordered set
387  std::set<Value *> AMembers(Members.begin(), Members.end());
388  std::set<Value *> BMembers(Members.begin(), Members.end());
389  return AMembers == BMembers;
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 
430  const Expression &E;
431 
432  explicit ExactEqualsExpression(const Expression &E) : E(E) {}
433 
434  hash_code getComputedHash() const { return E.getComputedHash(); }
435 
436  bool operator==(const Expression &Other) const {
437  return E.exactlyEquals(Other);
438  }
439 };
440 
441 template <> struct DenseMapInfo<const Expression *> {
442  static const Expression *getEmptyKey() {
443  auto Val = static_cast<uintptr_t>(-1);
444  Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
445  return reinterpret_cast<const Expression *>(Val);
446  }
447 
448  static const Expression *getTombstoneKey() {
449  auto Val = static_cast<uintptr_t>(~1U);
450  Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
451  return reinterpret_cast<const Expression *>(Val);
452  }
453 
454  static unsigned getHashValue(const Expression *E) {
455  return E->getComputedHash();
456  }
457 
458  static unsigned getHashValue(const ExactEqualsExpression &E) {
459  return E.getComputedHash();
460  }
461 
462  static bool isEqual(const ExactEqualsExpression &LHS, const Expression *RHS) {
463  if (RHS == getTombstoneKey() || RHS == getEmptyKey())
464  return false;
465  return LHS == *RHS;
466  }
467 
468  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;
491  const TargetLibraryInfo *TLI;
492  AliasAnalysis *AA;
493  MemorySSA *MSSA;
494  MemorySSAWalker *MSSAWalker;
495  const DataLayout &DL;
496  std::unique_ptr<PredicateInfo> PredInfo;
497 
498  // These are the only two things the create* functions should have
499  // side-effects on due to allocating memory.
500  mutable BumpPtrAllocator ExpressionAllocator;
501  mutable ArrayRecycler<Value *> ArgRecycler;
502  mutable TarjanSCC SCCFinder;
503  const SimplifyQuery SQ;
504 
505  // Number of function arguments, used by ranking
506  unsigned int NumFuncArgs;
507 
508  // RPOOrdering of basic blocks
510 
511  // Congruence class info.
512 
513  // This class is called INITIAL in the paper. It is the class everything
514  // startsout in, and represents any value. Being an optimistic analysis,
515  // anything in the TOP class has the value TOP, which is indeterminate and
516  // equivalent to everything.
517  CongruenceClass *TOPClass;
518  std::vector<CongruenceClass *> CongruenceClasses;
519  unsigned NextCongruenceNum;
520 
521  // Value Mappings.
523  DenseMap<Value *, const Expression *> ValueToExpression;
524 
525  // Value PHI handling, used to make equivalence between phi(op, op) and
526  // op(phi, phi).
527  // These mappings just store various data that would normally be part of the
528  // IR.
530 
531  DenseMap<const Value *, bool> OpSafeForPHIOfOps;
532 
533  // Map a temporary instruction we created to a parent block.
535 
536  // Map between the already in-program instructions and the temporary phis we
537  // created that they are known equivalent to.
539 
540  // In order to know when we should re-process instructions that have
541  // phi-of-ops, we track the set of expressions that they needed as
542  // leaders. When we discover new leaders for those expressions, we process the
543  // associated phi-of-op instructions again in case they have changed. The
544  // other way they may change is if they had leaders, and those leaders
545  // disappear. However, at the point they have leaders, there are uses of the
546  // relevant operands in the created phi node, and so they will get reprocessed
547  // through the normal user marking we perform.
548  mutable DenseMap<const Value *, SmallPtrSet<Value *, 2>> AdditionalUsers;
550  ExpressionToPhiOfOps;
551 
552  // Map from temporary operation to MemoryAccess.
554 
555  // Set of all temporary instructions we created.
556  // Note: This will include instructions that were just created during value
557  // numbering. The way to test if something is using them is to check
558  // RealToTemp.
559  DenseSet<Instruction *> AllTempInstructions;
560 
561  // This is the set of instructions to revisit on a reachability change. At
562  // the end of the main iteration loop it will contain at least all the phi of
563  // ops instructions that will be changed to phis, as well as regular phis.
564  // During the iteration loop, it may contain other things, such as phi of ops
565  // instructions that used edge reachability to reach a result, and so need to
566  // be revisited when the edge changes, independent of whether the phi they
567  // depended on changes.
568  DenseMap<BasicBlock *, SparseBitVector<>> RevisitOnReachabilityChange;
569 
570  // Mapping from predicate info we used to the instructions we used it with.
571  // In order to correctly ensure propagation, we must keep track of what
572  // comparisons we used, so that when the values of the comparisons change, we
573  // propagate the information to the places we used the comparison.
575  PredicateToUsers;
576 
577  // the same reasoning as PredicateToUsers. When we skip MemoryAccesses for
578  // stores, we no longer can rely solely on the def-use chains of MemorySSA.
580  MemoryToUsers;
581 
582  // A table storing which memorydefs/phis represent a memory state provably
583  // equivalent to another memory state.
584  // We could use the congruence class machinery, but the MemoryAccess's are
585  // abstract memory states, so they can only ever be equivalent to each other,
586  // and not to constants, etc.
588 
589  // We could, if we wanted, build MemoryPhiExpressions and
590  // MemoryVariableExpressions, etc, and value number them the same way we value
591  // number phi expressions. For the moment, this seems like overkill. They
592  // can only exist in one of three states: they can be TOP (equal to
593  // everything), Equivalent to something else, or unique. Because we do not
594  // create expressions for them, we need to simulate leader change not just
595  // when they change class, but when they change state. Note: We can do the
596  // same thing for phis, and avoid having phi expressions if we wanted, We
597  // should eventually unify in one direction or the other, so this is a little
598  // bit of an experiment in which turns out easier to maintain.
599  enum MemoryPhiState { MPS_Invalid, MPS_TOP, MPS_Equivalent, MPS_Unique };
601 
602  enum InstCycleState { ICS_Unknown, ICS_CycleFree, ICS_Cycle };
603  mutable DenseMap<const Instruction *, InstCycleState> InstCycleState;
604 
605  // Expression to class mapping.
606  using ExpressionClassMap = DenseMap<const Expression *, CongruenceClass *>;
607  ExpressionClassMap ExpressionToClass;
608 
609  // We have a single expression that represents currently DeadExpressions.
610  // For dead expressions we can prove will stay dead, we mark them with
611  // DFS number zero. However, it's possible in the case of phi nodes
612  // for us to assume/prove all arguments are dead during fixpointing.
613  // We use DeadExpression for that case.
614  DeadExpression *SingletonDeadExpression = nullptr;
615 
616  // Which values have changed as a result of leader changes.
617  SmallPtrSet<Value *, 8> LeaderChanges;
618 
619  // Reachability info.
620  using BlockEdge = BasicBlockEdge;
621  DenseSet<BlockEdge> ReachableEdges;
622  SmallPtrSet<const BasicBlock *, 8> ReachableBlocks;
623 
624  // This is a bitvector because, on larger functions, we may have
625  // thousands of touched instructions at once (entire blocks,
626  // instructions with hundreds of uses, etc). Even with optimization
627  // for when we mark whole blocks as touched, when this was a
628  // SmallPtrSet or DenseSet, for some functions, we spent >20% of all
629  // the time in GVN just managing this list. The bitvector, on the
630  // other hand, efficiently supports test/set/clear of both
631  // individual and ranges, as well as "find next element" This
632  // enables us to use it as a worklist with essentially 0 cost.
633  BitVector TouchedInstructions;
634 
636 
637 #ifndef NDEBUG
638  // Debugging for how many times each block and instruction got processed.
639  DenseMap<const Value *, unsigned> ProcessedCount;
640 #endif
641 
642  // DFS info.
643  // This contains a mapping from Instructions to DFS numbers.
644  // The numbering starts at 1. An instruction with DFS number zero
645  // means that the instruction is dead.
647 
648  // This contains the mapping DFS numbers to instructions.
649  SmallVector<Value *, 32> DFSToInstr;
650 
651  // Deletion info.
652  SmallPtrSet<Instruction *, 8> InstructionsToErase;
653 
654 public:
655  NewGVN(Function &F, DominatorTree *DT, AssumptionCache *AC,
656  TargetLibraryInfo *TLI, AliasAnalysis *AA, MemorySSA *MSSA,
657  const DataLayout &DL)
658  : F(F), DT(DT), TLI(TLI), AA(AA), MSSA(MSSA), DL(DL),
659  PredInfo(make_unique<PredicateInfo>(F, *DT, *AC)), SQ(DL, TLI, DT, AC) {
660  }
661 
662  bool runGVN();
663 
664 private:
665  // Expression handling.
666  const Expression *createExpression(Instruction *) const;
667  const Expression *createBinaryExpression(unsigned, Type *, Value *, Value *,
668  Instruction *) const;
669 
670  // Our canonical form for phi arguments is a pair of incoming value, incoming
671  // basic block.
672  using ValPair = std::pair<Value *, BasicBlock *>;
673 
674  PHIExpression *createPHIExpression(ArrayRef<ValPair>, const Instruction *,
675  BasicBlock *, bool &HasBackEdge,
676  bool &OriginalOpsConstant) const;
677  const DeadExpression *createDeadExpression() const;
678  const VariableExpression *createVariableExpression(Value *) const;
679  const ConstantExpression *createConstantExpression(Constant *) const;
680  const Expression *createVariableOrConstant(Value *V) const;
681  const UnknownExpression *createUnknownExpression(Instruction *) const;
682  const StoreExpression *createStoreExpression(StoreInst *,
683  const MemoryAccess *) const;
684  LoadExpression *createLoadExpression(Type *, Value *, LoadInst *,
685  const MemoryAccess *) const;
686  const CallExpression *createCallExpression(CallInst *,
687  const MemoryAccess *) const;
689  createAggregateValueExpression(Instruction *) const;
690  bool setBasicExpressionInfo(Instruction *, BasicExpression *) const;
691 
692  // Congruence class handling.
693  CongruenceClass *createCongruenceClass(Value *Leader, const Expression *E) {
694  auto *result = new CongruenceClass(NextCongruenceNum++, Leader, E);
695  CongruenceClasses.emplace_back(result);
696  return result;
697  }
698 
699  CongruenceClass *createMemoryClass(MemoryAccess *MA) {
700  auto *CC = createCongruenceClass(nullptr, nullptr);
701  CC->setMemoryLeader(MA);
702  return CC;
703  }
704 
705  CongruenceClass *ensureLeaderOfMemoryClass(MemoryAccess *MA) {
706  auto *CC = getMemoryClass(MA);
707  if (CC->getMemoryLeader() != MA)
708  CC = createMemoryClass(MA);
709  return CC;
710  }
711 
712  CongruenceClass *createSingletonCongruenceClass(Value *Member) {
713  CongruenceClass *CClass = createCongruenceClass(Member, nullptr);
714  CClass->insert(Member);
715  ValueToClass[Member] = CClass;
716  return CClass;
717  }
718 
719  void initializeCongruenceClasses(Function &F);
720  const Expression *makePossiblePHIOfOps(Instruction *,
722  Value *findLeaderForInst(Instruction *ValueOp,
723  SmallPtrSetImpl<Value *> &Visited,
724  MemoryAccess *MemAccess, Instruction *OrigInst,
725  BasicBlock *PredBB);
726  bool OpIsSafeForPHIOfOpsHelper(Value *V, const BasicBlock *PHIBlock,
729  bool OpIsSafeForPHIOfOps(Value *Op, const BasicBlock *PHIBlock,
731  void addPhiOfOps(PHINode *Op, BasicBlock *BB, Instruction *ExistingValue);
732  void removePhiOfOps(Instruction *I, PHINode *PHITemp);
733 
734  // Value number an Instruction or MemoryPhi.
735  void valueNumberMemoryPhi(MemoryPhi *);
736  void valueNumberInstruction(Instruction *);
737 
738  // Symbolic evaluation.
739  const Expression *checkSimplificationResults(Expression *, Instruction *,
740  Value *) const;
741  const Expression *performSymbolicEvaluation(Value *,
742  SmallPtrSetImpl<Value *> &) const;
743  const Expression *performSymbolicLoadCoercion(Type *, Value *, LoadInst *,
744  Instruction *,
745  MemoryAccess *) const;
746  const Expression *performSymbolicLoadEvaluation(Instruction *) const;
747  const Expression *performSymbolicStoreEvaluation(Instruction *) const;
748  const Expression *performSymbolicCallEvaluation(Instruction *) const;
749  void sortPHIOps(MutableArrayRef<ValPair> Ops) const;
750  const Expression *performSymbolicPHIEvaluation(ArrayRef<ValPair>,
751  Instruction *I,
752  BasicBlock *PHIBlock) const;
753  const Expression *performSymbolicAggrValueEvaluation(Instruction *) const;
754  const Expression *performSymbolicCmpEvaluation(Instruction *) const;
755  const Expression *performSymbolicPredicateInfoEvaluation(Instruction *) const;
756 
757  // Congruence finding.
758  bool someEquivalentDominates(const Instruction *, const Instruction *) const;
759  Value *lookupOperandLeader(Value *) const;
760  CongruenceClass *getClassForExpression(const Expression *E) const;
761  void performCongruenceFinding(Instruction *, const Expression *);
762  void moveValueToNewCongruenceClass(Instruction *, const Expression *,
763  CongruenceClass *, CongruenceClass *);
764  void moveMemoryToNewCongruenceClass(Instruction *, MemoryAccess *,
765  CongruenceClass *, CongruenceClass *);
766  Value *getNextValueLeader(CongruenceClass *) const;
767  const MemoryAccess *getNextMemoryLeader(CongruenceClass *) const;
768  bool setMemoryClass(const MemoryAccess *From, CongruenceClass *To);
769  CongruenceClass *getMemoryClass(const MemoryAccess *MA) const;
770  const MemoryAccess *lookupMemoryLeader(const MemoryAccess *) const;
771  bool isMemoryAccessTOP(const MemoryAccess *) const;
772 
773  // Ranking
774  unsigned int getRank(const Value *) const;
775  bool shouldSwapOperands(const Value *, const Value *) const;
776 
777  // Reachability handling.
778  void updateReachableEdge(BasicBlock *, BasicBlock *);
779  void processOutgoingEdges(TerminatorInst *, BasicBlock *);
780  Value *findConditionEquivalence(Value *) const;
781 
782  // Elimination.
783  struct ValueDFS;
784  void convertClassToDFSOrdered(const CongruenceClass &,
788  void convertClassToLoadsAndStores(const CongruenceClass &,
789  SmallVectorImpl<ValueDFS> &) const;
790 
791  bool eliminateInstructions(Function &);
792  void replaceInstruction(Instruction *, Value *);
793  void markInstructionForDeletion(Instruction *);
794  void deleteInstructionsInBlock(BasicBlock *);
795  Value *findPHIOfOpsLeader(const Expression *, const Instruction *,
796  const BasicBlock *) const;
797 
798  // New instruction creation.
799  void handleNewInstruction(Instruction *) {}
800 
801  // Various instruction touch utilities
802  template <typename Map, typename KeyType, typename Func>
803  void for_each_found(Map &, const KeyType &, Func);
804  template <typename Map, typename KeyType>
805  void touchAndErase(Map &, const KeyType &);
806  void markUsersTouched(Value *);
807  void markMemoryUsersTouched(const MemoryAccess *);
808  void markMemoryDefTouched(const MemoryAccess *);
809  void markPredicateUsersTouched(Instruction *);
810  void markValueLeaderChangeTouched(CongruenceClass *CC);
811  void markMemoryLeaderChangeTouched(CongruenceClass *CC);
812  void markPhiOfOpsChanged(const Expression *E);
813  void addPredicateUsers(const PredicateBase *, Instruction *) const;
814  void addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const;
815  void addAdditionalUsers(Value *To, Value *User) const;
816 
817  // Main loop of value numbering
818  void iterateTouchedInstructions();
819 
820  // Utilities.
821  void cleanupTables();
822  std::pair<unsigned, unsigned> assignDFSNumbers(BasicBlock *, unsigned);
823  void updateProcessedCount(const Value *V);
824  void verifyMemoryCongruency() const;
825  void verifyIterationSettled(Function &F);
826  void verifyStoreExpressions() const;
827  bool singleReachablePHIPath(SmallPtrSet<const MemoryAccess *, 8> &,
828  const MemoryAccess *, const MemoryAccess *) const;
829  BasicBlock *getBlockForValue(Value *V) const;
830  void deleteExpression(const Expression *E) const;
831  MemoryUseOrDef *getMemoryAccess(const Instruction *) const;
832  MemoryAccess *getDefiningAccess(const MemoryAccess *) const;
833  MemoryPhi *getMemoryAccess(const BasicBlock *) const;
834  template <class T, class Range> T *getMinDFSOfRange(const Range &) const;
835 
836  unsigned InstrToDFSNum(const Value *V) const {
837  assert(isa<Instruction>(V) && "This should not be used for MemoryAccesses");
838  return InstrDFS.lookup(V);
839  }
840 
841  unsigned InstrToDFSNum(const MemoryAccess *MA) const {
842  return MemoryToDFSNum(MA);
843  }
844 
845  Value *InstrFromDFSNum(unsigned DFSNum) { return DFSToInstr[DFSNum]; }
846 
847  // Given a MemoryAccess, return the relevant instruction DFS number. Note:
848  // This deliberately takes a value so it can be used with Use's, which will
849  // auto-convert to Value's but not to MemoryAccess's.
850  unsigned MemoryToDFSNum(const Value *MA) const {
851  assert(isa<MemoryAccess>(MA) &&
852  "This should not be used with instructions");
853  return isa<MemoryUseOrDef>(MA)
854  ? InstrToDFSNum(cast<MemoryUseOrDef>(MA)->getMemoryInst())
855  : InstrDFS.lookup(MA);
856  }
857 
858  bool isCycleFree(const Instruction *) const;
859  bool isBackedge(BasicBlock *From, BasicBlock *To) const;
860 
861  // Debug counter info. When verifying, we have to reset the value numbering
862  // debug counter to the same state it started in to get the same results.
863  std::pair<int, int> StartingVNCounter;
864 };
865 
866 } // end anonymous namespace
867 
868 template <typename T>
869 static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS) {
870  if (!isa<LoadExpression>(RHS) && !isa<StoreExpression>(RHS))
871  return false;
872  return LHS.MemoryExpression::equals(RHS);
873 }
874 
875 bool LoadExpression::equals(const Expression &Other) const {
876  return equalsLoadStoreHelper(*this, Other);
877 }
878 
879 bool StoreExpression::equals(const Expression &Other) const {
880  if (!equalsLoadStoreHelper(*this, Other))
881  return false;
882  // Make sure that store vs store includes the value operand.
883  if (const auto *S = dyn_cast<StoreExpression>(&Other))
884  if (getStoredValue() != S->getStoredValue())
885  return false;
886  return true;
887 }
888 
889 // Determine if the edge From->To is a backedge
890 bool NewGVN::isBackedge(BasicBlock *From, BasicBlock *To) const {
891  return From == To ||
892  RPOOrdering.lookup(DT->getNode(From)) >=
893  RPOOrdering.lookup(DT->getNode(To));
894 }
895 
896 #ifndef NDEBUG
897 static std::string getBlockName(const BasicBlock *B) {
899 }
900 #endif
901 
902 // Get a MemoryAccess for an instruction, fake or real.
903 MemoryUseOrDef *NewGVN::getMemoryAccess(const Instruction *I) const {
904  auto *Result = MSSA->getMemoryAccess(I);
905  return Result ? Result : TempToMemory.lookup(I);
906 }
907 
908 // Get a MemoryPhi for a basic block. These are all real.
909 MemoryPhi *NewGVN::getMemoryAccess(const BasicBlock *BB) const {
910  return MSSA->getMemoryAccess(BB);
911 }
912 
913 // Get the basic block from an instruction/memory value.
914 BasicBlock *NewGVN::getBlockForValue(Value *V) const {
915  if (auto *I = dyn_cast<Instruction>(V)) {
916  auto *Parent = I->getParent();
917  if (Parent)
918  return Parent;
919  Parent = TempToBlock.lookup(V);
920  assert(Parent && "Every fake instruction should have a block");
921  return Parent;
922  }
923 
924  auto *MP = dyn_cast<MemoryPhi>(V);
925  assert(MP && "Should have been an instruction or a MemoryPhi");
926  return MP->getBlock();
927 }
928 
929 // Delete a definitely dead expression, so it can be reused by the expression
930 // allocator. Some of these are not in creation functions, so we have to accept
931 // const versions.
932 void NewGVN::deleteExpression(const Expression *E) const {
933  assert(isa<BasicExpression>(E));
934  auto *BE = cast<BasicExpression>(E);
935  const_cast<BasicExpression *>(BE)->deallocateOperands(ArgRecycler);
936  ExpressionAllocator.Deallocate(E);
937 }
938 
939 // If V is a predicateinfo copy, get the thing it is a copy of.
940 static Value *getCopyOf(const Value *V) {
941  if (auto *II = dyn_cast<IntrinsicInst>(V))
942  if (II->getIntrinsicID() == Intrinsic::ssa_copy)
943  return II->getOperand(0);
944  return nullptr;
945 }
946 
947 // Return true if V is really PN, even accounting for predicateinfo copies.
948 static bool isCopyOfPHI(const Value *V, const PHINode *PN) {
949  return V == PN || getCopyOf(V) == PN;
950 }
951 
952 static bool isCopyOfAPHI(const Value *V) {
953  auto *CO = getCopyOf(V);
954  return CO && isa<PHINode>(CO);
955 }
956 
957 // Sort PHI Operands into a canonical order. What we use here is an RPO
958 // order. The BlockInstRange numbers are generated in an RPO walk of the basic
959 // blocks.
960 void NewGVN::sortPHIOps(MutableArrayRef<ValPair> Ops) const {
961  std::sort(Ops.begin(), Ops.end(), [&](const ValPair &P1, const ValPair &P2) {
962  return BlockInstRange.lookup(P1.second).first <
963  BlockInstRange.lookup(P2.second).first;
964  });
965 }
966 
967 // Return true if V is a value that will always be available (IE can
968 // be placed anywhere) in the function. We don't do globals here
969 // because they are often worse to put in place.
970 static bool alwaysAvailable(Value *V) {
971  return isa<Constant>(V) || isa<Argument>(V);
972 }
973 
974 // Create a PHIExpression from an array of {incoming edge, value} pairs. I is
975 // the original instruction we are creating a PHIExpression for (but may not be
976 // a phi node). We require, as an invariant, that all the PHIOperands in the
977 // same block are sorted the same way. sortPHIOps will sort them into a
978 // canonical order.
979 PHIExpression *NewGVN::createPHIExpression(ArrayRef<ValPair> PHIOperands,
980  const Instruction *I,
981  BasicBlock *PHIBlock,
982  bool &HasBackedge,
983  bool &OriginalOpsConstant) const {
984  unsigned NumOps = PHIOperands.size();
985  auto *E = new (ExpressionAllocator) PHIExpression(NumOps, PHIBlock);
986 
987  E->allocateOperands(ArgRecycler, ExpressionAllocator);
988  E->setType(PHIOperands.begin()->first->getType());
989  E->setOpcode(Instruction::PHI);
990 
991  // Filter out unreachable phi operands.
992  auto Filtered = make_filter_range(PHIOperands, [&](const ValPair &P) {
993  auto *BB = P.second;
994  if (auto *PHIOp = dyn_cast<PHINode>(I))
995  if (isCopyOfPHI(P.first, PHIOp))
996  return false;
997  if (!ReachableEdges.count({BB, PHIBlock}))
998  return false;
999  // Things in TOPClass are equivalent to everything.
1000  if (ValueToClass.lookup(P.first) == TOPClass)
1001  return false;
1002  OriginalOpsConstant = OriginalOpsConstant && isa<Constant>(P.first);
1003  HasBackedge = HasBackedge || isBackedge(BB, PHIBlock);
1004  return lookupOperandLeader(P.first) != I;
1005  });
1006  std::transform(Filtered.begin(), Filtered.end(), op_inserter(E),
1007  [&](const ValPair &P) -> Value * {
1008  return lookupOperandLeader(P.first);
1009  });
1010  return E;
1011 }
1012 
1013 // Set basic expression info (Arguments, type, opcode) for Expression
1014 // E from Instruction I in block B.
1015 bool NewGVN::setBasicExpressionInfo(Instruction *I, BasicExpression *E) const {
1016  bool AllConstant = true;
1017  if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
1018  E->setType(GEP->getSourceElementType());
1019  else
1020  E->setType(I->getType());
1021  E->setOpcode(I->getOpcode());
1022  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1023 
1024  // Transform the operand array into an operand leader array, and keep track of
1025  // whether all members are constant.
1026  std::transform(I->op_begin(), I->op_end(), op_inserter(E), [&](Value *O) {
1027  auto Operand = lookupOperandLeader(O);
1028  AllConstant = AllConstant && isa<Constant>(Operand);
1029  return Operand;
1030  });
1031 
1032  return AllConstant;
1033 }
1034 
1035 const Expression *NewGVN::createBinaryExpression(unsigned Opcode, Type *T,
1036  Value *Arg1, Value *Arg2,
1037  Instruction *I) const {
1038  auto *E = new (ExpressionAllocator) BasicExpression(2);
1039 
1040  E->setType(T);
1041  E->setOpcode(Opcode);
1042  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1043  if (Instruction::isCommutative(Opcode)) {
1044  // Ensure that commutative instructions that only differ by a permutation
1045  // of their operands get the same value number by sorting the operand value
1046  // numbers. Since all commutative instructions have two operands it is more
1047  // efficient to sort by hand rather than using, say, std::sort.
1048  if (shouldSwapOperands(Arg1, Arg2))
1049  std::swap(Arg1, Arg2);
1050  }
1051  E->op_push_back(lookupOperandLeader(Arg1));
1052  E->op_push_back(lookupOperandLeader(Arg2));
1053 
1054  Value *V = SimplifyBinOp(Opcode, E->getOperand(0), E->getOperand(1), SQ);
1055  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1056  return SimplifiedE;
1057  return E;
1058 }
1059 
1060 // Take a Value returned by simplification of Expression E/Instruction
1061 // I, and see if it resulted in a simpler expression. If so, return
1062 // that expression.
1063 const Expression *NewGVN::checkSimplificationResults(Expression *E,
1064  Instruction *I,
1065  Value *V) const {
1066  if (!V)
1067  return nullptr;
1068  if (auto *C = dyn_cast<Constant>(V)) {
1069  if (I)
1070  DEBUG(dbgs() << "Simplified " << *I << " to "
1071  << " constant " << *C << "\n");
1072  NumGVNOpsSimplified++;
1073  assert(isa<BasicExpression>(E) &&
1074  "We should always have had a basic expression here");
1075  deleteExpression(E);
1076  return createConstantExpression(C);
1077  } else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1078  if (I)
1079  DEBUG(dbgs() << "Simplified " << *I << " to "
1080  << " variable " << *V << "\n");
1081  deleteExpression(E);
1082  return createVariableExpression(V);
1083  }
1084 
1085  CongruenceClass *CC = ValueToClass.lookup(V);
1086  if (CC) {
1087  if (CC->getLeader() && CC->getLeader() != I) {
1088  // Don't add temporary instructions to the user lists.
1089  if (!AllTempInstructions.count(I))
1090  addAdditionalUsers(V, I);
1091  return createVariableOrConstant(CC->getLeader());
1092  }
1093  if (CC->getDefiningExpr()) {
1094  // If we simplified to something else, we need to communicate
1095  // that we're users of the value we simplified to.
1096  if (I != V) {
1097  // Don't add temporary instructions to the user lists.
1098  if (!AllTempInstructions.count(I))
1099  addAdditionalUsers(V, I);
1100  }
1101 
1102  if (I)
1103  DEBUG(dbgs() << "Simplified " << *I << " to "
1104  << " expression " << *CC->getDefiningExpr() << "\n");
1105  NumGVNOpsSimplified++;
1106  deleteExpression(E);
1107  return CC->getDefiningExpr();
1108  }
1109  }
1110 
1111  return nullptr;
1112 }
1113 
1114 // Create a value expression from the instruction I, replacing operands with
1115 // their leaders.
1116 
1117 const Expression *NewGVN::createExpression(Instruction *I) const {
1118  auto *E = new (ExpressionAllocator) BasicExpression(I->getNumOperands());
1119 
1120  bool AllConstant = setBasicExpressionInfo(I, E);
1121 
1122  if (I->isCommutative()) {
1123  // Ensure that commutative instructions that only differ by a permutation
1124  // of their operands get the same value number by sorting the operand value
1125  // numbers. Since all commutative instructions have two operands it is more
1126  // efficient to sort by hand rather than using, say, std::sort.
1127  assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
1128  if (shouldSwapOperands(E->getOperand(0), E->getOperand(1)))
1129  E->swapOperands(0, 1);
1130  }
1131  // Perform simplification.
1132  if (auto *CI = dyn_cast<CmpInst>(I)) {
1133  // Sort the operand value numbers so x<y and y>x get the same value
1134  // number.
1135  CmpInst::Predicate Predicate = CI->getPredicate();
1136  if (shouldSwapOperands(E->getOperand(0), E->getOperand(1))) {
1137  E->swapOperands(0, 1);
1138  Predicate = CmpInst::getSwappedPredicate(Predicate);
1139  }
1140  E->setOpcode((CI->getOpcode() << 8) | Predicate);
1141  // TODO: 25% of our time is spent in SimplifyCmpInst with pointer operands
1142  assert(I->getOperand(0)->getType() == I->getOperand(1)->getType() &&
1143  "Wrong types on cmp instruction");
1144  assert((E->getOperand(0)->getType() == I->getOperand(0)->getType() &&
1145  E->getOperand(1)->getType() == I->getOperand(1)->getType()));
1146  Value *V =
1147  SimplifyCmpInst(Predicate, E->getOperand(0), E->getOperand(1), SQ);
1148  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1149  return SimplifiedE;
1150  } else if (isa<SelectInst>(I)) {
1151  if (isa<Constant>(E->getOperand(0)) ||
1152  E->getOperand(1) == E->getOperand(2)) {
1153  assert(E->getOperand(1)->getType() == I->getOperand(1)->getType() &&
1154  E->getOperand(2)->getType() == I->getOperand(2)->getType());
1155  Value *V = SimplifySelectInst(E->getOperand(0), E->getOperand(1),
1156  E->getOperand(2), SQ);
1157  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1158  return SimplifiedE;
1159  }
1160  } else if (I->isBinaryOp()) {
1161  Value *V =
1162  SimplifyBinOp(E->getOpcode(), E->getOperand(0), E->getOperand(1), SQ);
1163  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1164  return SimplifiedE;
1165  } else if (auto *BI = dyn_cast<BitCastInst>(I)) {
1166  Value *V =
1167  SimplifyCastInst(BI->getOpcode(), BI->getOperand(0), BI->getType(), SQ);
1168  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1169  return SimplifiedE;
1170  } else if (isa<GetElementPtrInst>(I)) {
1171  Value *V = SimplifyGEPInst(
1172  E->getType(), ArrayRef<Value *>(E->op_begin(), E->op_end()), SQ);
1173  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1174  return SimplifiedE;
1175  } else if (AllConstant) {
1176  // We don't bother trying to simplify unless all of the operands
1177  // were constant.
1178  // TODO: There are a lot of Simplify*'s we could call here, if we
1179  // wanted to. The original motivating case for this code was a
1180  // zext i1 false to i8, which we don't have an interface to
1181  // simplify (IE there is no SimplifyZExt).
1182 
1184  for (Value *Arg : E->operands())
1185  C.emplace_back(cast<Constant>(Arg));
1186 
1187  if (Value *V = ConstantFoldInstOperands(I, C, DL, TLI))
1188  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1189  return SimplifiedE;
1190  }
1191  return E;
1192 }
1193 
1195 NewGVN::createAggregateValueExpression(Instruction *I) const {
1196  if (auto *II = dyn_cast<InsertValueInst>(I)) {
1197  auto *E = new (ExpressionAllocator)
1198  AggregateValueExpression(I->getNumOperands(), II->getNumIndices());
1199  setBasicExpressionInfo(I, E);
1200  E->allocateIntOperands(ExpressionAllocator);
1201  std::copy(II->idx_begin(), II->idx_end(), int_op_inserter(E));
1202  return E;
1203  } else if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1204  auto *E = new (ExpressionAllocator)
1205  AggregateValueExpression(I->getNumOperands(), EI->getNumIndices());
1206  setBasicExpressionInfo(EI, E);
1207  E->allocateIntOperands(ExpressionAllocator);
1208  std::copy(EI->idx_begin(), EI->idx_end(), int_op_inserter(E));
1209  return E;
1210  }
1211  llvm_unreachable("Unhandled type of aggregate value operation");
1212 }
1213 
1214 const DeadExpression *NewGVN::createDeadExpression() const {
1215  // DeadExpression has no arguments and all DeadExpression's are the same,
1216  // so we only need one of them.
1217  return SingletonDeadExpression;
1218 }
1219 
1220 const VariableExpression *NewGVN::createVariableExpression(Value *V) const {
1221  auto *E = new (ExpressionAllocator) VariableExpression(V);
1222  E->setOpcode(V->getValueID());
1223  return E;
1224 }
1225 
1226 const Expression *NewGVN::createVariableOrConstant(Value *V) const {
1227  if (auto *C = dyn_cast<Constant>(V))
1228  return createConstantExpression(C);
1229  return createVariableExpression(V);
1230 }
1231 
1232 const ConstantExpression *NewGVN::createConstantExpression(Constant *C) const {
1233  auto *E = new (ExpressionAllocator) ConstantExpression(C);
1234  E->setOpcode(C->getValueID());
1235  return E;
1236 }
1237 
1238 const UnknownExpression *NewGVN::createUnknownExpression(Instruction *I) const {
1239  auto *E = new (ExpressionAllocator) UnknownExpression(I);
1240  E->setOpcode(I->getOpcode());
1241  return E;
1242 }
1243 
1244 const CallExpression *
1245 NewGVN::createCallExpression(CallInst *CI, const MemoryAccess *MA) const {
1246  // FIXME: Add operand bundles for calls.
1247  auto *E =
1248  new (ExpressionAllocator) CallExpression(CI->getNumOperands(), CI, MA);
1249  setBasicExpressionInfo(CI, E);
1250  return E;
1251 }
1252 
1253 // Return true if some equivalent of instruction Inst dominates instruction U.
1254 bool NewGVN::someEquivalentDominates(const Instruction *Inst,
1255  const Instruction *U) const {
1256  auto *CC = ValueToClass.lookup(Inst);
1257  // This must be an instruction because we are only called from phi nodes
1258  // in the case that the value it needs to check against is an instruction.
1259 
1260  // The most likely candiates for dominance are the leader and the next leader.
1261  // The leader or nextleader will dominate in all cases where there is an
1262  // equivalent that is higher up in the dom tree.
1263  // We can't *only* check them, however, because the
1264  // dominator tree could have an infinite number of non-dominating siblings
1265  // with instructions that are in the right congruence class.
1266  // A
1267  // B C D E F G
1268  // |
1269  // H
1270  // Instruction U could be in H, with equivalents in every other sibling.
1271  // Depending on the rpo order picked, the leader could be the equivalent in
1272  // any of these siblings.
1273  if (!CC)
1274  return false;
1275  if (alwaysAvailable(CC->getLeader()))
1276  return true;
1277  if (DT->dominates(cast<Instruction>(CC->getLeader()), U))
1278  return true;
1279  if (CC->getNextLeader().first &&
1280  DT->dominates(cast<Instruction>(CC->getNextLeader().first), U))
1281  return true;
1282  return llvm::any_of(*CC, [&](const Value *Member) {
1283  return Member != CC->getLeader() &&
1284  DT->dominates(cast<Instruction>(Member), U);
1285  });
1286 }
1287 
1288 // See if we have a congruence class and leader for this operand, and if so,
1289 // return it. Otherwise, return the operand itself.
1290 Value *NewGVN::lookupOperandLeader(Value *V) const {
1291  CongruenceClass *CC = ValueToClass.lookup(V);
1292  if (CC) {
1293  // Everything in TOP is represented by undef, as it can be any value.
1294  // We do have to make sure we get the type right though, so we can't set the
1295  // RepLeader to undef.
1296  if (CC == TOPClass)
1297  return UndefValue::get(V->getType());
1298  return CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
1299  }
1300 
1301  return V;
1302 }
1303 
1304 const MemoryAccess *NewGVN::lookupMemoryLeader(const MemoryAccess *MA) const {
1305  auto *CC = getMemoryClass(MA);
1306  assert(CC->getMemoryLeader() &&
1307  "Every MemoryAccess should be mapped to a congruence class with a "
1308  "representative memory access");
1309  return CC->getMemoryLeader();
1310 }
1311 
1312 // Return true if the MemoryAccess is really equivalent to everything. This is
1313 // equivalent to the lattice value "TOP" in most lattices. This is the initial
1314 // state of all MemoryAccesses.
1315 bool NewGVN::isMemoryAccessTOP(const MemoryAccess *MA) const {
1316  return getMemoryClass(MA) == TOPClass;
1317 }
1318 
1319 LoadExpression *NewGVN::createLoadExpression(Type *LoadType, Value *PointerOp,
1320  LoadInst *LI,
1321  const MemoryAccess *MA) const {
1322  auto *E =
1323  new (ExpressionAllocator) LoadExpression(1, LI, lookupMemoryLeader(MA));
1324  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1325  E->setType(LoadType);
1326 
1327  // Give store and loads same opcode so they value number together.
1328  E->setOpcode(0);
1329  E->op_push_back(PointerOp);
1330  if (LI)
1331  E->setAlignment(LI->getAlignment());
1332 
1333  // TODO: Value number heap versions. We may be able to discover
1334  // things alias analysis can't on it's own (IE that a store and a
1335  // load have the same value, and thus, it isn't clobbering the load).
1336  return E;
1337 }
1338 
1339 const StoreExpression *
1340 NewGVN::createStoreExpression(StoreInst *SI, const MemoryAccess *MA) const {
1341  auto *StoredValueLeader = lookupOperandLeader(SI->getValueOperand());
1342  auto *E = new (ExpressionAllocator)
1343  StoreExpression(SI->getNumOperands(), SI, StoredValueLeader, MA);
1344  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1345  E->setType(SI->getValueOperand()->getType());
1346 
1347  // Give store and loads same opcode so they value number together.
1348  E->setOpcode(0);
1349  E->op_push_back(lookupOperandLeader(SI->getPointerOperand()));
1350 
1351  // TODO: Value number heap versions. We may be able to discover
1352  // things alias analysis can't on it's own (IE that a store and a
1353  // load have the same value, and thus, it isn't clobbering the load).
1354  return E;
1355 }
1356 
1357 const Expression *NewGVN::performSymbolicStoreEvaluation(Instruction *I) const {
1358  // Unlike loads, we never try to eliminate stores, so we do not check if they
1359  // are simple and avoid value numbering them.
1360  auto *SI = cast<StoreInst>(I);
1361  auto *StoreAccess = getMemoryAccess(SI);
1362  // Get the expression, if any, for the RHS of the MemoryDef.
1363  const MemoryAccess *StoreRHS = StoreAccess->getDefiningAccess();
1365  StoreRHS = MSSAWalker->getClobberingMemoryAccess(StoreAccess);
1366  // If we bypassed the use-def chains, make sure we add a use.
1367  StoreRHS = lookupMemoryLeader(StoreRHS);
1368  if (StoreRHS != StoreAccess->getDefiningAccess())
1369  addMemoryUsers(StoreRHS, StoreAccess);
1370  // If we are defined by ourselves, use the live on entry def.
1371  if (StoreRHS == StoreAccess)
1372  StoreRHS = MSSA->getLiveOnEntryDef();
1373 
1374  if (SI->isSimple()) {
1375  // See if we are defined by a previous store expression, it already has a
1376  // value, and it's the same value as our current store. FIXME: Right now, we
1377  // only do this for simple stores, we should expand to cover memcpys, etc.
1378  const auto *LastStore = createStoreExpression(SI, StoreRHS);
1379  const auto *LastCC = ExpressionToClass.lookup(LastStore);
1380  // We really want to check whether the expression we matched was a store. No
1381  // easy way to do that. However, we can check that the class we found has a
1382  // store, which, assuming the value numbering state is not corrupt, is
1383  // sufficient, because we must also be equivalent to that store's expression
1384  // for it to be in the same class as the load.
1385  if (LastCC && LastCC->getStoredValue() == LastStore->getStoredValue())
1386  return LastStore;
1387  // Also check if our value operand is defined by a load of the same memory
1388  // location, and the memory state is the same as it was then (otherwise, it
1389  // could have been overwritten later. See test32 in
1390  // transforms/DeadStoreElimination/simple.ll).
1391  if (auto *LI = dyn_cast<LoadInst>(LastStore->getStoredValue()))
1392  if ((lookupOperandLeader(LI->getPointerOperand()) ==
1393  LastStore->getOperand(0)) &&
1394  (lookupMemoryLeader(getMemoryAccess(LI)->getDefiningAccess()) ==
1395  StoreRHS))
1396  return LastStore;
1397  deleteExpression(LastStore);
1398  }
1399 
1400  // If the store is not equivalent to anything, value number it as a store that
1401  // produces a unique memory state (instead of using it's MemoryUse, we use
1402  // it's MemoryDef).
1403  return createStoreExpression(SI, StoreAccess);
1404 }
1405 
1406 // See if we can extract the value of a loaded pointer from a load, a store, or
1407 // a memory instruction.
1408 const Expression *
1409 NewGVN::performSymbolicLoadCoercion(Type *LoadType, Value *LoadPtr,
1410  LoadInst *LI, Instruction *DepInst,
1411  MemoryAccess *DefiningAccess) const {
1412  assert((!LI || LI->isSimple()) && "Not a simple load");
1413  if (auto *DepSI = dyn_cast<StoreInst>(DepInst)) {
1414  // Can't forward from non-atomic to atomic without violating memory model.
1415  // Also don't need to coerce if they are the same type, we will just
1416  // propagate.
1417  if (LI->isAtomic() > DepSI->isAtomic() ||
1418  LoadType == DepSI->getValueOperand()->getType())
1419  return nullptr;
1420  int Offset = analyzeLoadFromClobberingStore(LoadType, LoadPtr, DepSI, DL);
1421  if (Offset >= 0) {
1422  if (auto *C = dyn_cast<Constant>(
1423  lookupOperandLeader(DepSI->getValueOperand()))) {
1424  DEBUG(dbgs() << "Coercing load from store " << *DepSI << " to constant "
1425  << *C << "\n");
1426  return createConstantExpression(
1427  getConstantStoreValueForLoad(C, Offset, LoadType, DL));
1428  }
1429  }
1430  } else if (auto *DepLI = dyn_cast<LoadInst>(DepInst)) {
1431  // Can't forward from non-atomic to atomic without violating memory model.
1432  if (LI->isAtomic() > DepLI->isAtomic())
1433  return nullptr;
1434  int Offset = analyzeLoadFromClobberingLoad(LoadType, LoadPtr, DepLI, DL);
1435  if (Offset >= 0) {
1436  // We can coerce a constant load into a load.
1437  if (auto *C = dyn_cast<Constant>(lookupOperandLeader(DepLI)))
1438  if (auto *PossibleConstant =
1439  getConstantLoadValueForLoad(C, Offset, LoadType, DL)) {
1440  DEBUG(dbgs() << "Coercing load from load " << *LI << " to constant "
1441  << *PossibleConstant << "\n");
1442  return createConstantExpression(PossibleConstant);
1443  }
1444  }
1445  } else if (auto *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
1446  int Offset = analyzeLoadFromClobberingMemInst(LoadType, LoadPtr, DepMI, DL);
1447  if (Offset >= 0) {
1448  if (auto *PossibleConstant =
1449  getConstantMemInstValueForLoad(DepMI, Offset, LoadType, DL)) {
1450  DEBUG(dbgs() << "Coercing load from meminst " << *DepMI
1451  << " to constant " << *PossibleConstant << "\n");
1452  return createConstantExpression(PossibleConstant);
1453  }
1454  }
1455  }
1456 
1457  // All of the below are only true if the loaded pointer is produced
1458  // by the dependent instruction.
1459  if (LoadPtr != lookupOperandLeader(DepInst) &&
1460  !AA->isMustAlias(LoadPtr, DepInst))
1461  return nullptr;
1462  // If this load really doesn't depend on anything, then we must be loading an
1463  // undef value. This can happen when loading for a fresh allocation with no
1464  // intervening stores, for example. Note that this is only true in the case
1465  // that the result of the allocation is pointer equal to the load ptr.
1466  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1467  return createConstantExpression(UndefValue::get(LoadType));
1468  }
1469  // If this load occurs either right after a lifetime begin,
1470  // then the loaded value is undefined.
1471  else if (auto *II = dyn_cast<IntrinsicInst>(DepInst)) {
1472  if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1473  return createConstantExpression(UndefValue::get(LoadType));
1474  }
1475  // If this load follows a calloc (which zero initializes memory),
1476  // then the loaded value is zero
1477  else if (isCallocLikeFn(DepInst, TLI)) {
1478  return createConstantExpression(Constant::getNullValue(LoadType));
1479  }
1480 
1481  return nullptr;
1482 }
1483 
1484 const Expression *NewGVN::performSymbolicLoadEvaluation(Instruction *I) const {
1485  auto *LI = cast<LoadInst>(I);
1486 
1487  // We can eliminate in favor of non-simple loads, but we won't be able to
1488  // eliminate the loads themselves.
1489  if (!LI->isSimple())
1490  return nullptr;
1491 
1492  Value *LoadAddressLeader = lookupOperandLeader(LI->getPointerOperand());
1493  // Load of undef is undef.
1494  if (isa<UndefValue>(LoadAddressLeader))
1495  return createConstantExpression(UndefValue::get(LI->getType()));
1496  MemoryAccess *OriginalAccess = getMemoryAccess(I);
1497  MemoryAccess *DefiningAccess =
1498  MSSAWalker->getClobberingMemoryAccess(OriginalAccess);
1499 
1500  if (!MSSA->isLiveOnEntryDef(DefiningAccess)) {
1501  if (auto *MD = dyn_cast<MemoryDef>(DefiningAccess)) {
1502  Instruction *DefiningInst = MD->getMemoryInst();
1503  // If the defining instruction is not reachable, replace with undef.
1504  if (!ReachableBlocks.count(DefiningInst->getParent()))
1505  return createConstantExpression(UndefValue::get(LI->getType()));
1506  // This will handle stores and memory insts. We only do if it the
1507  // defining access has a different type, or it is a pointer produced by
1508  // certain memory operations that cause the memory to have a fixed value
1509  // (IE things like calloc).
1510  if (const auto *CoercionResult =
1511  performSymbolicLoadCoercion(LI->getType(), LoadAddressLeader, LI,
1512  DefiningInst, DefiningAccess))
1513  return CoercionResult;
1514  }
1515  }
1516 
1517  const auto *LE = createLoadExpression(LI->getType(), LoadAddressLeader, LI,
1518  DefiningAccess);
1519  // If our MemoryLeader is not our defining access, add a use to the
1520  // MemoryLeader, so that we get reprocessed when it changes.
1521  if (LE->getMemoryLeader() != DefiningAccess)
1522  addMemoryUsers(LE->getMemoryLeader(), OriginalAccess);
1523  return LE;
1524 }
1525 
1526 const Expression *
1527 NewGVN::performSymbolicPredicateInfoEvaluation(Instruction *I) const {
1528  auto *PI = PredInfo->getPredicateInfoFor(I);
1529  if (!PI)
1530  return nullptr;
1531 
1532  DEBUG(dbgs() << "Found predicate info from instruction !\n");
1533 
1534  auto *PWC = dyn_cast<PredicateWithCondition>(PI);
1535  if (!PWC)
1536  return nullptr;
1537 
1538  auto *CopyOf = I->getOperand(0);
1539  auto *Cond = PWC->Condition;
1540 
1541  // If this a copy of the condition, it must be either true or false depending
1542  // on the predicate info type and edge.
1543  if (CopyOf == Cond) {
1544  // We should not need to add predicate users because the predicate info is
1545  // already a use of this operand.
1546  if (isa<PredicateAssume>(PI))
1547  return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1548  if (auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1549  if (PBranch->TrueEdge)
1550  return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1551  return createConstantExpression(ConstantInt::getFalse(Cond->getType()));
1552  }
1553  if (auto *PSwitch = dyn_cast<PredicateSwitch>(PI))
1554  return createConstantExpression(cast<Constant>(PSwitch->CaseValue));
1555  }
1556 
1557  // Not a copy of the condition, so see what the predicates tell us about this
1558  // value. First, though, we check to make sure the value is actually a copy
1559  // of one of the condition operands. It's possible, in certain cases, for it
1560  // to be a copy of a predicateinfo copy. In particular, if two branch
1561  // operations use the same condition, and one branch dominates the other, we
1562  // will end up with a copy of a copy. This is currently a small deficiency in
1563  // predicateinfo. What will end up happening here is that we will value
1564  // number both copies the same anyway.
1565 
1566  // Everything below relies on the condition being a comparison.
1567  auto *Cmp = dyn_cast<CmpInst>(Cond);
1568  if (!Cmp)
1569  return nullptr;
1570 
1571  if (CopyOf != Cmp->getOperand(0) && CopyOf != Cmp->getOperand(1)) {
1572  DEBUG(dbgs() << "Copy is not of any condition operands!\n");
1573  return nullptr;
1574  }
1575  Value *FirstOp = lookupOperandLeader(Cmp->getOperand(0));
1576  Value *SecondOp = lookupOperandLeader(Cmp->getOperand(1));
1577  bool SwappedOps = false;
1578  // Sort the ops.
1579  if (shouldSwapOperands(FirstOp, SecondOp)) {
1580  std::swap(FirstOp, SecondOp);
1581  SwappedOps = true;
1582  }
1584  SwappedOps ? Cmp->getSwappedPredicate() : Cmp->getPredicate();
1585 
1586  if (isa<PredicateAssume>(PI)) {
1587  // If the comparison is true when the operands are equal, then we know the
1588  // operands are equal, because assumes must always be true.
1589  if (CmpInst::isTrueWhenEqual(Predicate)) {
1590  addPredicateUsers(PI, I);
1591  addAdditionalUsers(Cmp->getOperand(0), I);
1592  return createVariableOrConstant(FirstOp);
1593  }
1594  }
1595  if (const auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1596  // If we are *not* a copy of the comparison, we may equal to the other
1597  // operand when the predicate implies something about equality of
1598  // operations. In particular, if the comparison is true/false when the
1599  // operands are equal, and we are on the right edge, we know this operation
1600  // is equal to something.
1601  if ((PBranch->TrueEdge && Predicate == CmpInst::ICMP_EQ) ||
1602  (!PBranch->TrueEdge && Predicate == CmpInst::ICMP_NE)) {
1603  addPredicateUsers(PI, I);
1604  addAdditionalUsers(SwappedOps ? Cmp->getOperand(1) : Cmp->getOperand(0),
1605  I);
1606  return createVariableOrConstant(FirstOp);
1607  }
1608  // Handle the special case of floating point.
1609  if (((PBranch->TrueEdge && Predicate == CmpInst::FCMP_OEQ) ||
1610  (!PBranch->TrueEdge && Predicate == CmpInst::FCMP_UNE)) &&
1611  isa<ConstantFP>(FirstOp) && !cast<ConstantFP>(FirstOp)->isZero()) {
1612  addPredicateUsers(PI, I);
1613  addAdditionalUsers(SwappedOps ? Cmp->getOperand(1) : Cmp->getOperand(0),
1614  I);
1615  return createConstantExpression(cast<Constant>(FirstOp));
1616  }
1617  }
1618  return nullptr;
1619 }
1620 
1621 // Evaluate read only and pure calls, and create an expression result.
1622 const Expression *NewGVN::performSymbolicCallEvaluation(Instruction *I) const {
1623  auto *CI = cast<CallInst>(I);
1624  if (auto *II = dyn_cast<IntrinsicInst>(I)) {
1625  // Instrinsics with the returned attribute are copies of arguments.
1626  if (auto *ReturnedValue = II->getReturnedArgOperand()) {
1627  if (II->getIntrinsicID() == Intrinsic::ssa_copy)
1628  if (const auto *Result = performSymbolicPredicateInfoEvaluation(I))
1629  return Result;
1630  return createVariableOrConstant(ReturnedValue);
1631  }
1632  }
1633  if (AA->doesNotAccessMemory(CI)) {
1634  return createCallExpression(CI, TOPClass->getMemoryLeader());
1635  } else if (AA->onlyReadsMemory(CI)) {
1636  MemoryAccess *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(CI);
1637  return createCallExpression(CI, DefiningAccess);
1638  }
1639  return nullptr;
1640 }
1641 
1642 // Retrieve the memory class for a given MemoryAccess.
1643 CongruenceClass *NewGVN::getMemoryClass(const MemoryAccess *MA) const {
1644  auto *Result = MemoryAccessToClass.lookup(MA);
1645  assert(Result && "Should have found memory class");
1646  return Result;
1647 }
1648 
1649 // Update the MemoryAccess equivalence table to say that From is equal to To,
1650 // and return true if this is different from what already existed in the table.
1651 bool NewGVN::setMemoryClass(const MemoryAccess *From,
1652  CongruenceClass *NewClass) {
1653  assert(NewClass &&
1654  "Every MemoryAccess should be getting mapped to a non-null class");
1655  DEBUG(dbgs() << "Setting " << *From);
1656  DEBUG(dbgs() << " equivalent to congruence class ");
1657  DEBUG(dbgs() << NewClass->getID() << " with current MemoryAccess leader ");
1658  DEBUG(dbgs() << *NewClass->getMemoryLeader() << "\n");
1659 
1660  auto LookupResult = MemoryAccessToClass.find(From);
1661  bool Changed = false;
1662  // If it's already in the table, see if the value changed.
1663  if (LookupResult != MemoryAccessToClass.end()) {
1664  auto *OldClass = LookupResult->second;
1665  if (OldClass != NewClass) {
1666  // If this is a phi, we have to handle memory member updates.
1667  if (auto *MP = dyn_cast<MemoryPhi>(From)) {
1668  OldClass->memory_erase(MP);
1669  NewClass->memory_insert(MP);
1670  // This may have killed the class if it had no non-memory members
1671  if (OldClass->getMemoryLeader() == From) {
1672  if (OldClass->definesNoMemory()) {
1673  OldClass->setMemoryLeader(nullptr);
1674  } else {
1675  OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
1676  DEBUG(dbgs() << "Memory class leader change for class "
1677  << OldClass->getID() << " to "
1678  << *OldClass->getMemoryLeader()
1679  << " due to removal of a memory member " << *From
1680  << "\n");
1681  markMemoryLeaderChangeTouched(OldClass);
1682  }
1683  }
1684  }
1685  // It wasn't equivalent before, and now it is.
1686  LookupResult->second = NewClass;
1687  Changed = true;
1688  }
1689  }
1690 
1691  return Changed;
1692 }
1693 
1694 // Determine if a instruction is cycle-free. That means the values in the
1695 // instruction don't depend on any expressions that can change value as a result
1696 // of the instruction. For example, a non-cycle free instruction would be v =
1697 // phi(0, v+1).
1698 bool NewGVN::isCycleFree(const Instruction *I) const {
1699  // In order to compute cycle-freeness, we do SCC finding on the instruction,
1700  // and see what kind of SCC it ends up in. If it is a singleton, it is
1701  // cycle-free. If it is not in a singleton, it is only cycle free if the
1702  // other members are all phi nodes (as they do not compute anything, they are
1703  // copies).
1704  auto ICS = InstCycleState.lookup(I);
1705  if (ICS == ICS_Unknown) {
1706  SCCFinder.Start(I);
1707  auto &SCC = SCCFinder.getComponentFor(I);
1708  // It's cycle free if it's size 1 or or the SCC is *only* phi nodes.
1709  if (SCC.size() == 1)
1710  InstCycleState.insert({I, ICS_CycleFree});
1711  else {
1712  bool AllPhis = llvm::all_of(SCC, [](const Value *V) {
1713  return isa<PHINode>(V) || isCopyOfAPHI(V);
1714  });
1715  ICS = AllPhis ? ICS_CycleFree : ICS_Cycle;
1716  for (auto *Member : SCC)
1717  if (auto *MemberPhi = dyn_cast<PHINode>(Member))
1718  InstCycleState.insert({MemberPhi, ICS});
1719  }
1720  }
1721  if (ICS == ICS_Cycle)
1722  return false;
1723  return true;
1724 }
1725 
1726 // Evaluate PHI nodes symbolically and create an expression result.
1727 const Expression *
1728 NewGVN::performSymbolicPHIEvaluation(ArrayRef<ValPair> PHIOps,
1729  Instruction *I,
1730  BasicBlock *PHIBlock) const {
1731  // True if one of the incoming phi edges is a backedge.
1732  bool HasBackedge = false;
1733  // All constant tracks the state of whether all the *original* phi operands
1734  // This is really shorthand for "this phi cannot cycle due to forward
1735  // change in value of the phi is guaranteed not to later change the value of
1736  // the phi. IE it can't be v = phi(undef, v+1)
1737  bool OriginalOpsConstant = true;
1738  auto *E = cast<PHIExpression>(createPHIExpression(
1739  PHIOps, I, PHIBlock, HasBackedge, OriginalOpsConstant));
1740  // We match the semantics of SimplifyPhiNode from InstructionSimplify here.
1741  // See if all arguments are the same.
1742  // We track if any were undef because they need special handling.
1743  bool HasUndef = false;
1744  auto Filtered = make_filter_range(E->operands(), [&](Value *Arg) {
1745  if (isa<UndefValue>(Arg)) {
1746  HasUndef = true;
1747  return false;
1748  }
1749  return true;
1750  });
1751  // If we are left with no operands, it's dead.
1752  if (Filtered.begin() == Filtered.end()) {
1753  // If it has undef at this point, it means there are no-non-undef arguments,
1754  // and thus, the value of the phi node must be undef.
1755  if (HasUndef) {
1756  DEBUG(dbgs() << "PHI Node " << *I
1757  << " has no non-undef arguments, valuing it as undef\n");
1758  return createConstantExpression(UndefValue::get(I->getType()));
1759  }
1760 
1761  DEBUG(dbgs() << "No arguments of PHI node " << *I << " are live\n");
1762  deleteExpression(E);
1763  return createDeadExpression();
1764  }
1765  Value *AllSameValue = *(Filtered.begin());
1766  ++Filtered.begin();
1767  // Can't use std::equal here, sadly, because filter.begin moves.
1768  if (llvm::all_of(Filtered, [&](Value *Arg) { return Arg == AllSameValue; })) {
1769  // In LLVM's non-standard representation of phi nodes, it's possible to have
1770  // phi nodes with cycles (IE dependent on other phis that are .... dependent
1771  // on the original phi node), especially in weird CFG's where some arguments
1772  // are unreachable, or uninitialized along certain paths. This can cause
1773  // infinite loops during evaluation. We work around this by not trying to
1774  // really evaluate them independently, but instead using a variable
1775  // expression to say if one is equivalent to the other.
1776  // We also special case undef, so that if we have an undef, we can't use the
1777  // common value unless it dominates the phi block.
1778  if (HasUndef) {
1779  // If we have undef and at least one other value, this is really a
1780  // multivalued phi, and we need to know if it's cycle free in order to
1781  // evaluate whether we can ignore the undef. The other parts of this are
1782  // just shortcuts. If there is no backedge, or all operands are
1783  // constants, it also must be cycle free.
1784  if (HasBackedge && !OriginalOpsConstant &&
1785  !isa<UndefValue>(AllSameValue) && !isCycleFree(I))
1786  return E;
1787 
1788  // Only have to check for instructions
1789  if (auto *AllSameInst = dyn_cast<Instruction>(AllSameValue))
1790  if (!someEquivalentDominates(AllSameInst, I))
1791  return E;
1792  }
1793  // Can't simplify to something that comes later in the iteration.
1794  // Otherwise, when and if it changes congruence class, we will never catch
1795  // up. We will always be a class behind it.
1796  if (isa<Instruction>(AllSameValue) &&
1797  InstrToDFSNum(AllSameValue) > InstrToDFSNum(I))
1798  return E;
1799  NumGVNPhisAllSame++;
1800  DEBUG(dbgs() << "Simplified PHI node " << *I << " to " << *AllSameValue
1801  << "\n");
1802  deleteExpression(E);
1803  return createVariableOrConstant(AllSameValue);
1804  }
1805  return E;
1806 }
1807 
1808 const Expression *
1809 NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) const {
1810  if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1811  auto *II = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
1812  if (II && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
1813  unsigned Opcode = 0;
1814  // EI might be an extract from one of our recognised intrinsics. If it
1815  // is we'll synthesize a semantically equivalent expression instead on
1816  // an extract value expression.
1817  switch (II->getIntrinsicID()) {
1818  case Intrinsic::sadd_with_overflow:
1819  case Intrinsic::uadd_with_overflow:
1820  Opcode = Instruction::Add;
1821  break;
1822  case Intrinsic::ssub_with_overflow:
1823  case Intrinsic::usub_with_overflow:
1824  Opcode = Instruction::Sub;
1825  break;
1826  case Intrinsic::smul_with_overflow:
1827  case Intrinsic::umul_with_overflow:
1828  Opcode = Instruction::Mul;
1829  break;
1830  default:
1831  break;
1832  }
1833 
1834  if (Opcode != 0) {
1835  // Intrinsic recognized. Grab its args to finish building the
1836  // expression.
1837  assert(II->getNumArgOperands() == 2 &&
1838  "Expect two args for recognised intrinsics.");
1839  return createBinaryExpression(Opcode, EI->getType(),
1840  II->getArgOperand(0),
1841  II->getArgOperand(1), I);
1842  }
1843  }
1844  }
1845 
1846  return createAggregateValueExpression(I);
1847 }
1848 
1849 const Expression *NewGVN::performSymbolicCmpEvaluation(Instruction *I) const {
1850  assert(isa<CmpInst>(I) && "Expected a cmp instruction.");
1851 
1852  auto *CI = cast<CmpInst>(I);
1853  // See if our operands are equal to those of a previous predicate, and if so,
1854  // if it implies true or false.
1855  auto Op0 = lookupOperandLeader(CI->getOperand(0));
1856  auto Op1 = lookupOperandLeader(CI->getOperand(1));
1857  auto OurPredicate = CI->getPredicate();
1858  if (shouldSwapOperands(Op0, Op1)) {
1859  std::swap(Op0, Op1);
1860  OurPredicate = CI->getSwappedPredicate();
1861  }
1862 
1863  // Avoid processing the same info twice.
1864  const PredicateBase *LastPredInfo = nullptr;
1865  // See if we know something about the comparison itself, like it is the target
1866  // of an assume.
1867  auto *CmpPI = PredInfo->getPredicateInfoFor(I);
1868  if (dyn_cast_or_null<PredicateAssume>(CmpPI))
1869  return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1870 
1871  if (Op0 == Op1) {
1872  // This condition does not depend on predicates, no need to add users
1873  if (CI->isTrueWhenEqual())
1874  return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1875  else if (CI->isFalseWhenEqual())
1876  return createConstantExpression(ConstantInt::getFalse(CI->getType()));
1877  }
1878 
1879  // NOTE: Because we are comparing both operands here and below, and using
1880  // previous comparisons, we rely on fact that predicateinfo knows to mark
1881  // comparisons that use renamed operands as users of the earlier comparisons.
1882  // It is *not* enough to just mark predicateinfo renamed operands as users of
1883  // the earlier comparisons, because the *other* operand may have changed in a
1884  // previous iteration.
1885  // Example:
1886  // icmp slt %a, %b
1887  // %b.0 = ssa.copy(%b)
1888  // false branch:
1889  // icmp slt %c, %b.0
1890 
1891  // %c and %a may start out equal, and thus, the code below will say the second
1892  // %icmp is false. c may become equal to something else, and in that case the
1893  // %second icmp *must* be reexamined, but would not if only the renamed
1894  // %operands are considered users of the icmp.
1895 
1896  // *Currently* we only check one level of comparisons back, and only mark one
1897  // level back as touched when changes happen. If you modify this code to look
1898  // back farther through comparisons, you *must* mark the appropriate
1899  // comparisons as users in PredicateInfo.cpp, or you will cause bugs. See if
1900  // we know something just from the operands themselves
1901 
1902  // See if our operands have predicate info, so that we may be able to derive
1903  // something from a previous comparison.
1904  for (const auto &Op : CI->operands()) {
1905  auto *PI = PredInfo->getPredicateInfoFor(Op);
1906  if (const auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI)) {
1907  if (PI == LastPredInfo)
1908  continue;
1909  LastPredInfo = PI;
1910  // In phi of ops cases, we may have predicate info that we are evaluating
1911  // in a different context.
1912  if (!DT->dominates(PBranch->To, getBlockForValue(I)))
1913  continue;
1914  // TODO: Along the false edge, we may know more things too, like
1915  // icmp of
1916  // same operands is false.
1917  // TODO: We only handle actual comparison conditions below, not
1918  // and/or.
1919  auto *BranchCond = dyn_cast<CmpInst>(PBranch->Condition);
1920  if (!BranchCond)
1921  continue;
1922  auto *BranchOp0 = lookupOperandLeader(BranchCond->getOperand(0));
1923  auto *BranchOp1 = lookupOperandLeader(BranchCond->getOperand(1));
1924  auto BranchPredicate = BranchCond->getPredicate();
1925  if (shouldSwapOperands(BranchOp0, BranchOp1)) {
1926  std::swap(BranchOp0, BranchOp1);
1927  BranchPredicate = BranchCond->getSwappedPredicate();
1928  }
1929  if (BranchOp0 == Op0 && BranchOp1 == Op1) {
1930  if (PBranch->TrueEdge) {
1931  // If we know the previous predicate is true and we are in the true
1932  // edge then we may be implied true or false.
1933  if (CmpInst::isImpliedTrueByMatchingCmp(BranchPredicate,
1934  OurPredicate)) {
1935  addPredicateUsers(PI, I);
1936  return createConstantExpression(
1937  ConstantInt::getTrue(CI->getType()));
1938  }
1939 
1940  if (CmpInst::isImpliedFalseByMatchingCmp(BranchPredicate,
1941  OurPredicate)) {
1942  addPredicateUsers(PI, I);
1943  return createConstantExpression(
1945  }
1946  } else {
1947  // Just handle the ne and eq cases, where if we have the same
1948  // operands, we may know something.
1949  if (BranchPredicate == OurPredicate) {
1950  addPredicateUsers(PI, I);
1951  // Same predicate, same ops,we know it was false, so this is false.
1952  return createConstantExpression(
1954  } else if (BranchPredicate ==
1955  CmpInst::getInversePredicate(OurPredicate)) {
1956  addPredicateUsers(PI, I);
1957  // Inverse predicate, we know the other was false, so this is true.
1958  return createConstantExpression(
1959  ConstantInt::getTrue(CI->getType()));
1960  }
1961  }
1962  }
1963  }
1964  }
1965  // Create expression will take care of simplifyCmpInst
1966  return createExpression(I);
1967 }
1968 
1969 // Substitute and symbolize the value before value numbering.
1970 const Expression *
1971 NewGVN::performSymbolicEvaluation(Value *V,
1972  SmallPtrSetImpl<Value *> &Visited) const {
1973  const Expression *E = nullptr;
1974  if (auto *C = dyn_cast<Constant>(V))
1975  E = createConstantExpression(C);
1976  else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1977  E = createVariableExpression(V);
1978  } else {
1979  // TODO: memory intrinsics.
1980  // TODO: Some day, we should do the forward propagation and reassociation
1981  // parts of the algorithm.
1982  auto *I = cast<Instruction>(V);
1983  switch (I->getOpcode()) {
1984  case Instruction::ExtractValue:
1985  case Instruction::InsertValue:
1986  E = performSymbolicAggrValueEvaluation(I);
1987  break;
1988  case Instruction::PHI: {
1990  auto *PN = cast<PHINode>(I);
1991  for (unsigned i = 0; i < PN->getNumOperands(); ++i)
1992  Ops.push_back({PN->getIncomingValue(i), PN->getIncomingBlock(i)});
1993  // Sort to ensure the invariant createPHIExpression requires is met.
1994  sortPHIOps(Ops);
1995  E = performSymbolicPHIEvaluation(Ops, I, getBlockForValue(I));
1996  } break;
1997  case Instruction::Call:
1998  E = performSymbolicCallEvaluation(I);
1999  break;
2000  case Instruction::Store:
2001  E = performSymbolicStoreEvaluation(I);
2002  break;
2003  case Instruction::Load:
2004  E = performSymbolicLoadEvaluation(I);
2005  break;
2006  case Instruction::BitCast:
2007  E = createExpression(I);
2008  break;
2009  case Instruction::ICmp:
2010  case Instruction::FCmp:
2011  E = performSymbolicCmpEvaluation(I);
2012  break;
2013  case Instruction::Add:
2014  case Instruction::FAdd:
2015  case Instruction::Sub:
2016  case Instruction::FSub:
2017  case Instruction::Mul:
2018  case Instruction::FMul:
2019  case Instruction::UDiv:
2020  case Instruction::SDiv:
2021  case Instruction::FDiv:
2022  case Instruction::URem:
2023  case Instruction::SRem:
2024  case Instruction::FRem:
2025  case Instruction::Shl:
2026  case Instruction::LShr:
2027  case Instruction::AShr:
2028  case Instruction::And:
2029  case Instruction::Or:
2030  case Instruction::Xor:
2031  case Instruction::Trunc:
2032  case Instruction::ZExt:
2033  case Instruction::SExt:
2034  case Instruction::FPToUI:
2035  case Instruction::FPToSI:
2036  case Instruction::UIToFP:
2037  case Instruction::SIToFP:
2038  case Instruction::FPTrunc:
2039  case Instruction::FPExt:
2040  case Instruction::PtrToInt:
2041  case Instruction::IntToPtr:
2042  case Instruction::Select:
2043  case Instruction::ExtractElement:
2044  case Instruction::InsertElement:
2045  case Instruction::ShuffleVector:
2046  case Instruction::GetElementPtr:
2047  E = createExpression(I);
2048  break;
2049  default:
2050  return nullptr;
2051  }
2052  }
2053  return E;
2054 }
2055 
2056 // Look up a container in a map, and then call a function for each thing in the
2057 // found container.
2058 template <typename Map, typename KeyType, typename Func>
2059 void NewGVN::for_each_found(Map &M, const KeyType &Key, Func F) {
2060  const auto Result = M.find_as(Key);
2061  if (Result != M.end())
2062  for (typename Map::mapped_type::value_type Mapped : Result->second)
2063  F(Mapped);
2064 }
2065 
2066 // Look up a container of values/instructions in a map, and touch all the
2067 // instructions in the container. Then erase value from the map.
2068 template <typename Map, typename KeyType>
2069 void NewGVN::touchAndErase(Map &M, const KeyType &Key) {
2070  const auto Result = M.find_as(Key);
2071  if (Result != M.end()) {
2072  for (const typename Map::mapped_type::value_type Mapped : Result->second)
2073  TouchedInstructions.set(InstrToDFSNum(Mapped));
2074  M.erase(Result);
2075  }
2076 }
2077 
2078 void NewGVN::addAdditionalUsers(Value *To, Value *User) const {
2079  assert(User && To != User);
2080  if (isa<Instruction>(To))
2081  AdditionalUsers[To].insert(User);
2082 }
2083 
2084 void NewGVN::markUsersTouched(Value *V) {
2085  // Now mark the users as touched.
2086  for (auto *User : V->users()) {
2087  assert(isa<Instruction>(User) && "Use of value not within an instruction?");
2088  TouchedInstructions.set(InstrToDFSNum(User));
2089  }
2090  touchAndErase(AdditionalUsers, V);
2091 }
2092 
2093 void NewGVN::addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const {
2094  DEBUG(dbgs() << "Adding memory user " << *U << " to " << *To << "\n");
2095  MemoryToUsers[To].insert(U);
2096 }
2097 
2098 void NewGVN::markMemoryDefTouched(const MemoryAccess *MA) {
2099  TouchedInstructions.set(MemoryToDFSNum(MA));
2100 }
2101 
2102 void NewGVN::markMemoryUsersTouched(const MemoryAccess *MA) {
2103  if (isa<MemoryUse>(MA))
2104  return;
2105  for (auto U : MA->users())
2106  TouchedInstructions.set(MemoryToDFSNum(U));
2107  touchAndErase(MemoryToUsers, MA);
2108 }
2109 
2110 // Add I to the set of users of a given predicate.
2111 void NewGVN::addPredicateUsers(const PredicateBase *PB, Instruction *I) const {
2112  // Don't add temporary instructions to the user lists.
2113  if (AllTempInstructions.count(I))
2114  return;
2115 
2116  if (auto *PBranch = dyn_cast<PredicateBranch>(PB))
2117  PredicateToUsers[PBranch->Condition].insert(I);
2118  else if (auto *PAssume = dyn_cast<PredicateBranch>(PB))
2119  PredicateToUsers[PAssume->Condition].insert(I);
2120 }
2121 
2122 // Touch all the predicates that depend on this instruction.
2123 void NewGVN::markPredicateUsersTouched(Instruction *I) {
2124  touchAndErase(PredicateToUsers, I);
2125 }
2126 
2127 // Mark users affected by a memory leader change.
2128 void NewGVN::markMemoryLeaderChangeTouched(CongruenceClass *CC) {
2129  for (auto M : CC->memory())
2130  markMemoryDefTouched(M);
2131 }
2132 
2133 // Touch the instructions that need to be updated after a congruence class has a
2134 // leader change, and mark changed values.
2135 void NewGVN::markValueLeaderChangeTouched(CongruenceClass *CC) {
2136  for (auto M : *CC) {
2137  if (auto *I = dyn_cast<Instruction>(M))
2138  TouchedInstructions.set(InstrToDFSNum(I));
2139  LeaderChanges.insert(M);
2140  }
2141 }
2142 
2143 // Give a range of things that have instruction DFS numbers, this will return
2144 // the member of the range with the smallest dfs number.
2145 template <class T, class Range>
2146 T *NewGVN::getMinDFSOfRange(const Range &R) const {
2147  std::pair<T *, unsigned> MinDFS = {nullptr, ~0U};
2148  for (const auto X : R) {
2149  auto DFSNum = InstrToDFSNum(X);
2150  if (DFSNum < MinDFS.second)
2151  MinDFS = {X, DFSNum};
2152  }
2153  return MinDFS.first;
2154 }
2155 
2156 // This function returns the MemoryAccess that should be the next leader of
2157 // congruence class CC, under the assumption that the current leader is going to
2158 // disappear.
2159 const MemoryAccess *NewGVN::getNextMemoryLeader(CongruenceClass *CC) const {
2160  // TODO: If this ends up to slow, we can maintain a next memory leader like we
2161  // do for regular leaders.
2162  // Make sure there will be a leader to find.
2163  assert(!CC->definesNoMemory() && "Can't get next leader if there is none");
2164  if (CC->getStoreCount() > 0) {
2165  if (auto *NL = dyn_cast_or_null<StoreInst>(CC->getNextLeader().first))
2166  return getMemoryAccess(NL);
2167  // Find the store with the minimum DFS number.
2168  auto *V = getMinDFSOfRange<Value>(make_filter_range(
2169  *CC, [&](const Value *V) { return isa<StoreInst>(V); }));
2170  return getMemoryAccess(cast<StoreInst>(V));
2171  }
2172  assert(CC->getStoreCount() == 0);
2173 
2174  // Given our assertion, hitting this part must mean
2175  // !OldClass->memory_empty()
2176  if (CC->memory_size() == 1)
2177  return *CC->memory_begin();
2178  return getMinDFSOfRange<const MemoryPhi>(CC->memory());
2179 }
2180 
2181 // This function returns the next value leader of a congruence class, under the
2182 // assumption that the current leader is going away. This should end up being
2183 // the next most dominating member.
2184 Value *NewGVN::getNextValueLeader(CongruenceClass *CC) const {
2185  // We don't need to sort members if there is only 1, and we don't care about
2186  // sorting the TOP class because everything either gets out of it or is
2187  // unreachable.
2188 
2189  if (CC->size() == 1 || CC == TOPClass) {
2190  return *(CC->begin());
2191  } else if (CC->getNextLeader().first) {
2192  ++NumGVNAvoidedSortedLeaderChanges;
2193  return CC->getNextLeader().first;
2194  } else {
2195  ++NumGVNSortedLeaderChanges;
2196  // NOTE: If this ends up to slow, we can maintain a dual structure for
2197  // member testing/insertion, or keep things mostly sorted, and sort only
2198  // here, or use SparseBitVector or ....
2199  return getMinDFSOfRange<Value>(*CC);
2200  }
2201 }
2202 
2203 // Move a MemoryAccess, currently in OldClass, to NewClass, including updates to
2204 // the memory members, etc for the move.
2205 //
2206 // The invariants of this function are:
2207 //
2208 // - I must be moving to NewClass from OldClass
2209 // - The StoreCount of OldClass and NewClass is expected to have been updated
2210 // for I already if it is is a store.
2211 // - The OldClass memory leader has not been updated yet if I was the leader.
2212 void NewGVN::moveMemoryToNewCongruenceClass(Instruction *I,
2213  MemoryAccess *InstMA,
2214  CongruenceClass *OldClass,
2215  CongruenceClass *NewClass) {
2216  // If the leader is I, and we had a represenative MemoryAccess, it should
2217  // be the MemoryAccess of OldClass.
2218  assert((!InstMA || !OldClass->getMemoryLeader() ||
2219  OldClass->getLeader() != I ||
2220  MemoryAccessToClass.lookup(OldClass->getMemoryLeader()) ==
2221  MemoryAccessToClass.lookup(InstMA)) &&
2222  "Representative MemoryAccess mismatch");
2223  // First, see what happens to the new class
2224  if (!NewClass->getMemoryLeader()) {
2225  // Should be a new class, or a store becoming a leader of a new class.
2226  assert(NewClass->size() == 1 ||
2227  (isa<StoreInst>(I) && NewClass->getStoreCount() == 1));
2228  NewClass->setMemoryLeader(InstMA);
2229  // Mark it touched if we didn't just create a singleton
2230  DEBUG(dbgs() << "Memory class leader change for class " << NewClass->getID()
2231  << " due to new memory instruction becoming leader\n");
2232  markMemoryLeaderChangeTouched(NewClass);
2233  }
2234  setMemoryClass(InstMA, NewClass);
2235  // Now, fixup the old class if necessary
2236  if (OldClass->getMemoryLeader() == InstMA) {
2237  if (!OldClass->definesNoMemory()) {
2238  OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
2239  DEBUG(dbgs() << "Memory class leader change for class "
2240  << OldClass->getID() << " to "
2241  << *OldClass->getMemoryLeader()
2242  << " due to removal of old leader " << *InstMA << "\n");
2243  markMemoryLeaderChangeTouched(OldClass);
2244  } else
2245  OldClass->setMemoryLeader(nullptr);
2246  }
2247 }
2248 
2249 // Move a value, currently in OldClass, to be part of NewClass
2250 // Update OldClass and NewClass for the move (including changing leaders, etc).
2251 void NewGVN::moveValueToNewCongruenceClass(Instruction *I, const Expression *E,
2252  CongruenceClass *OldClass,
2253  CongruenceClass *NewClass) {
2254  if (I == OldClass->getNextLeader().first)
2255  OldClass->resetNextLeader();
2256 
2257  OldClass->erase(I);
2258  NewClass->insert(I);
2259 
2260  if (NewClass->getLeader() != I)
2261  NewClass->addPossibleNextLeader({I, InstrToDFSNum(I)});
2262  // Handle our special casing of stores.
2263  if (auto *SI = dyn_cast<StoreInst>(I)) {
2264  OldClass->decStoreCount();
2265  // Okay, so when do we want to make a store a leader of a class?
2266  // If we have a store defined by an earlier load, we want the earlier load
2267  // to lead the class.
2268  // If we have a store defined by something else, we want the store to lead
2269  // the class so everything else gets the "something else" as a value.
2270  // If we have a store as the single member of the class, we want the store
2271  // as the leader
2272  if (NewClass->getStoreCount() == 0 && !NewClass->getStoredValue()) {
2273  // If it's a store expression we are using, it means we are not equivalent
2274  // to something earlier.
2275  if (auto *SE = dyn_cast<StoreExpression>(E)) {
2276  NewClass->setStoredValue(SE->getStoredValue());
2277  markValueLeaderChangeTouched(NewClass);
2278  // Shift the new class leader to be the store
2279  DEBUG(dbgs() << "Changing leader of congruence class "
2280  << NewClass->getID() << " from " << *NewClass->getLeader()
2281  << " to " << *SI << " because store joined class\n");
2282  // If we changed the leader, we have to mark it changed because we don't
2283  // know what it will do to symbolic evaluation.
2284  NewClass->setLeader(SI);
2285  }
2286  // We rely on the code below handling the MemoryAccess change.
2287  }
2288  NewClass->incStoreCount();
2289  }
2290  // True if there is no memory instructions left in a class that had memory
2291  // instructions before.
2292 
2293  // If it's not a memory use, set the MemoryAccess equivalence
2294  auto *InstMA = dyn_cast_or_null<MemoryDef>(getMemoryAccess(I));
2295  if (InstMA)
2296  moveMemoryToNewCongruenceClass(I, InstMA, OldClass, NewClass);
2297  ValueToClass[I] = NewClass;
2298  // See if we destroyed the class or need to swap leaders.
2299  if (OldClass->empty() && OldClass != TOPClass) {
2300  if (OldClass->getDefiningExpr()) {
2301  DEBUG(dbgs() << "Erasing expression " << *OldClass->getDefiningExpr()
2302  << " from table\n");
2303  // We erase it as an exact expression to make sure we don't just erase an
2304  // equivalent one.
2305  auto Iter = ExpressionToClass.find_as(
2306  ExactEqualsExpression(*OldClass->getDefiningExpr()));
2307  if (Iter != ExpressionToClass.end())
2308  ExpressionToClass.erase(Iter);
2309 #ifdef EXPENSIVE_CHECKS
2310  assert(
2311  (*OldClass->getDefiningExpr() != *E || ExpressionToClass.lookup(E)) &&
2312  "We erased the expression we just inserted, which should not happen");
2313 #endif
2314  }
2315  } else if (OldClass->getLeader() == I) {
2316  // When the leader changes, the value numbering of
2317  // everything may change due to symbolization changes, so we need to
2318  // reprocess.
2319  DEBUG(dbgs() << "Value class leader change for class " << OldClass->getID()
2320  << "\n");
2321  ++NumGVNLeaderChanges;
2322  // Destroy the stored value if there are no more stores to represent it.
2323  // Note that this is basically clean up for the expression removal that
2324  // happens below. If we remove stores from a class, we may leave it as a
2325  // class of equivalent memory phis.
2326  if (OldClass->getStoreCount() == 0) {
2327  if (OldClass->getStoredValue())
2328  OldClass->setStoredValue(nullptr);
2329  }
2330  OldClass->setLeader(getNextValueLeader(OldClass));
2331  OldClass->resetNextLeader();
2332  markValueLeaderChangeTouched(OldClass);
2333  }
2334 }
2335 
2336 // For a given expression, mark the phi of ops instructions that could have
2337 // changed as a result.
2338 void NewGVN::markPhiOfOpsChanged(const Expression *E) {
2339  touchAndErase(ExpressionToPhiOfOps, E);
2340 }
2341 
2342 // Perform congruence finding on a given value numbering expression.
2343 void NewGVN::performCongruenceFinding(Instruction *I, const Expression *E) {
2344  // This is guaranteed to return something, since it will at least find
2345  // TOP.
2346 
2347  CongruenceClass *IClass = ValueToClass.lookup(I);
2348  assert(IClass && "Should have found a IClass");
2349  // Dead classes should have been eliminated from the mapping.
2350  assert(!IClass->isDead() && "Found a dead class");
2351 
2352  CongruenceClass *EClass = nullptr;
2353  if (const auto *VE = dyn_cast<VariableExpression>(E)) {
2354  EClass = ValueToClass.lookup(VE->getVariableValue());
2355  } else if (isa<DeadExpression>(E)) {
2356  EClass = TOPClass;
2357  }
2358  if (!EClass) {
2359  auto lookupResult = ExpressionToClass.insert({E, nullptr});
2360 
2361  // If it's not in the value table, create a new congruence class.
2362  if (lookupResult.second) {
2363  CongruenceClass *NewClass = createCongruenceClass(nullptr, E);
2364  auto place = lookupResult.first;
2365  place->second = NewClass;
2366 
2367  // Constants and variables should always be made the leader.
2368  if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2369  NewClass->setLeader(CE->getConstantValue());
2370  } else if (const auto *SE = dyn_cast<StoreExpression>(E)) {
2371  StoreInst *SI = SE->getStoreInst();
2372  NewClass->setLeader(SI);
2373  NewClass->setStoredValue(SE->getStoredValue());
2374  // The RepMemoryAccess field will be filled in properly by the
2375  // moveValueToNewCongruenceClass call.
2376  } else {
2377  NewClass->setLeader(I);
2378  }
2379  assert(!isa<VariableExpression>(E) &&
2380  "VariableExpression should have been handled already");
2381 
2382  EClass = NewClass;
2383  DEBUG(dbgs() << "Created new congruence class for " << *I
2384  << " using expression " << *E << " at " << NewClass->getID()
2385  << " and leader " << *(NewClass->getLeader()));
2386  if (NewClass->getStoredValue())
2387  DEBUG(dbgs() << " and stored value " << *(NewClass->getStoredValue()));
2388  DEBUG(dbgs() << "\n");
2389  } else {
2390  EClass = lookupResult.first->second;
2391  if (isa<ConstantExpression>(E))
2392  assert((isa<Constant>(EClass->getLeader()) ||
2393  (EClass->getStoredValue() &&
2394  isa<Constant>(EClass->getStoredValue()))) &&
2395  "Any class with a constant expression should have a "
2396  "constant leader");
2397 
2398  assert(EClass && "Somehow don't have an eclass");
2399 
2400  assert(!EClass->isDead() && "We accidentally looked up a dead class");
2401  }
2402  }
2403  bool ClassChanged = IClass != EClass;
2404  bool LeaderChanged = LeaderChanges.erase(I);
2405  if (ClassChanged || LeaderChanged) {
2406  DEBUG(dbgs() << "New class " << EClass->getID() << " for expression " << *E
2407  << "\n");
2408  if (ClassChanged) {
2409  moveValueToNewCongruenceClass(I, E, IClass, EClass);
2410  markPhiOfOpsChanged(E);
2411  }
2412 
2413  markUsersTouched(I);
2414  if (MemoryAccess *MA = getMemoryAccess(I))
2415  markMemoryUsersTouched(MA);
2416  if (auto *CI = dyn_cast<CmpInst>(I))
2417  markPredicateUsersTouched(CI);
2418  }
2419  // If we changed the class of the store, we want to ensure nothing finds the
2420  // old store expression. In particular, loads do not compare against stored
2421  // value, so they will find old store expressions (and associated class
2422  // mappings) if we leave them in the table.
2423  if (ClassChanged && isa<StoreInst>(I)) {
2424  auto *OldE = ValueToExpression.lookup(I);
2425  // It could just be that the old class died. We don't want to erase it if we
2426  // just moved classes.
2427  if (OldE && isa<StoreExpression>(OldE) && *E != *OldE) {
2428  // Erase this as an exact expression to ensure we don't erase expressions
2429  // equivalent to it.
2430  auto Iter = ExpressionToClass.find_as(ExactEqualsExpression(*OldE));
2431  if (Iter != ExpressionToClass.end())
2432  ExpressionToClass.erase(Iter);
2433  }
2434  }
2435  ValueToExpression[I] = E;
2436 }
2437 
2438 // Process the fact that Edge (from, to) is reachable, including marking
2439 // any newly reachable blocks and instructions for processing.
2440 void NewGVN::updateReachableEdge(BasicBlock *From, BasicBlock *To) {
2441  // Check if the Edge was reachable before.
2442  if (ReachableEdges.insert({From, To}).second) {
2443  // If this block wasn't reachable before, all instructions are touched.
2444  if (ReachableBlocks.insert(To).second) {
2445  DEBUG(dbgs() << "Block " << getBlockName(To) << " marked reachable\n");
2446  const auto &InstRange = BlockInstRange.lookup(To);
2447  TouchedInstructions.set(InstRange.first, InstRange.second);
2448  } else {
2449  DEBUG(dbgs() << "Block " << getBlockName(To)
2450  << " was reachable, but new edge {" << getBlockName(From)
2451  << "," << getBlockName(To) << "} to it found\n");
2452 
2453  // We've made an edge reachable to an existing block, which may
2454  // impact predicates. Otherwise, only mark the phi nodes as touched, as
2455  // they are the only thing that depend on new edges. Anything using their
2456  // values will get propagated to if necessary.
2457  if (MemoryAccess *MemPhi = getMemoryAccess(To))
2458  TouchedInstructions.set(InstrToDFSNum(MemPhi));
2459 
2460  // FIXME: We should just add a union op on a Bitvector and
2461  // SparseBitVector. We can do it word by word faster than we are doing it
2462  // here.
2463  for (auto InstNum : RevisitOnReachabilityChange[To])
2464  TouchedInstructions.set(InstNum);
2465  }
2466  }
2467 }
2468 
2469 // Given a predicate condition (from a switch, cmp, or whatever) and a block,
2470 // see if we know some constant value for it already.
2471 Value *NewGVN::findConditionEquivalence(Value *Cond) const {
2472  auto Result = lookupOperandLeader(Cond);
2473  return isa<Constant>(Result) ? Result : nullptr;
2474 }
2475 
2476 // Process the outgoing edges of a block for reachability.
2477 void NewGVN::processOutgoingEdges(TerminatorInst *TI, BasicBlock *B) {
2478  // Evaluate reachability of terminator instruction.
2479  BranchInst *BR;
2480  if ((BR = dyn_cast<BranchInst>(TI)) && BR->isConditional()) {
2481  Value *Cond = BR->getCondition();
2482  Value *CondEvaluated = findConditionEquivalence(Cond);
2483  if (!CondEvaluated) {
2484  if (auto *I = dyn_cast<Instruction>(Cond)) {
2485  const Expression *E = createExpression(I);
2486  if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2487  CondEvaluated = CE->getConstantValue();
2488  }
2489  } else if (isa<ConstantInt>(Cond)) {
2490  CondEvaluated = Cond;
2491  }
2492  }
2493  ConstantInt *CI;
2494  BasicBlock *TrueSucc = BR->getSuccessor(0);
2495  BasicBlock *FalseSucc = BR->getSuccessor(1);
2496  if (CondEvaluated && (CI = dyn_cast<ConstantInt>(CondEvaluated))) {
2497  if (CI->isOne()) {
2498  DEBUG(dbgs() << "Condition for Terminator " << *TI
2499  << " evaluated to true\n");
2500  updateReachableEdge(B, TrueSucc);
2501  } else if (CI->isZero()) {
2502  DEBUG(dbgs() << "Condition for Terminator " << *TI
2503  << " evaluated to false\n");
2504  updateReachableEdge(B, FalseSucc);
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  // Remember how many outgoing edges there are to every successor.
2516 
2517  Value *SwitchCond = SI->getCondition();
2518  Value *CondEvaluated = findConditionEquivalence(SwitchCond);
2519  // See if we were able to turn this switch statement into a constant.
2520  if (CondEvaluated && isa<ConstantInt>(CondEvaluated)) {
2521  auto *CondVal = cast<ConstantInt>(CondEvaluated);
2522  // We should be able to get case value for this.
2523  auto Case = *SI->findCaseValue(CondVal);
2524  if (Case.getCaseSuccessor() == SI->getDefaultDest()) {
2525  // We proved the value is outside of the range of the case.
2526  // We can't do anything other than mark the default dest as reachable,
2527  // and go home.
2528  updateReachableEdge(B, SI->getDefaultDest());
2529  return;
2530  }
2531  // Now get where it goes and mark it reachable.
2532  BasicBlock *TargetBlock = Case.getCaseSuccessor();
2533  updateReachableEdge(B, TargetBlock);
2534  } else {
2535  for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) {
2536  BasicBlock *TargetBlock = SI->getSuccessor(i);
2537  ++SwitchEdges[TargetBlock];
2538  updateReachableEdge(B, TargetBlock);
2539  }
2540  }
2541  } else {
2542  // Otherwise this is either unconditional, or a type we have no
2543  // idea about. Just mark successors as reachable.
2544  for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
2545  BasicBlock *TargetBlock = TI->getSuccessor(i);
2546  updateReachableEdge(B, TargetBlock);
2547  }
2548 
2549  // This also may be a memory defining terminator, in which case, set it
2550  // equivalent only to itself.
2551  //
2552  auto *MA = getMemoryAccess(TI);
2553  if (MA && !isa<MemoryUse>(MA)) {
2554  auto *CC = ensureLeaderOfMemoryClass(MA);
2555  if (setMemoryClass(MA, CC))
2556  markMemoryUsersTouched(MA);
2557  }
2558  }
2559 }
2560 
2561 // Remove the PHI of Ops PHI for I
2562 void NewGVN::removePhiOfOps(Instruction *I, PHINode *PHITemp) {
2563  InstrDFS.erase(PHITemp);
2564  // It's still a temp instruction. We keep it in the array so it gets erased.
2565  // However, it's no longer used by I, or in the block
2566  TempToBlock.erase(PHITemp);
2567  RealToTemp.erase(I);
2568  // We don't remove the users from the phi node uses. This wastes a little
2569  // time, but such is life. We could use two sets to track which were there
2570  // are the start of NewGVN, and which were added, but right nowt he cost of
2571  // tracking is more than the cost of checking for more phi of ops.
2572 }
2573 
2574 // Add PHI Op in BB as a PHI of operations version of ExistingValue.
2575 void NewGVN::addPhiOfOps(PHINode *Op, BasicBlock *BB,
2576  Instruction *ExistingValue) {
2577  InstrDFS[Op] = InstrToDFSNum(ExistingValue);
2578  AllTempInstructions.insert(Op);
2579  TempToBlock[Op] = BB;
2580  RealToTemp[ExistingValue] = Op;
2581  // Add all users to phi node use, as they are now uses of the phi of ops phis
2582  // and may themselves be phi of ops.
2583  for (auto *U : ExistingValue->users())
2584  if (auto *UI = dyn_cast<Instruction>(U))
2585  PHINodeUses.insert(UI);
2586 }
2587 
2588 static bool okayForPHIOfOps(const Instruction *I) {
2589  if (!EnablePhiOfOps)
2590  return false;
2591  return isa<BinaryOperator>(I) || isa<SelectInst>(I) || isa<CmpInst>(I) ||
2592  isa<LoadInst>(I);
2593 }
2594 
2595 bool NewGVN::OpIsSafeForPHIOfOpsHelper(
2596  Value *V, const BasicBlock *PHIBlock,
2598  SmallVectorImpl<Instruction *> &Worklist) {
2599 
2600  if (!isa<Instruction>(V))
2601  return true;
2602  auto OISIt = OpSafeForPHIOfOps.find(V);
2603  if (OISIt != OpSafeForPHIOfOps.end())
2604  return OISIt->second;
2605 
2606  // Keep walking until we either dominate the phi block, or hit a phi, or run
2607  // out of things to check.
2608  if (DT->properlyDominates(getBlockForValue(V), PHIBlock)) {
2609  OpSafeForPHIOfOps.insert({V, true});
2610  return true;
2611  }
2612  // PHI in the same block.
2613  if (isa<PHINode>(V) && getBlockForValue(V) == PHIBlock) {
2614  OpSafeForPHIOfOps.insert({V, false});
2615  return false;
2616  }
2617 
2618  auto *OrigI = cast<Instruction>(V);
2619  for (auto *Op : OrigI->operand_values()) {
2620  if (!isa<Instruction>(Op))
2621  continue;
2622  // Stop now if we find an unsafe operand.
2623  auto OISIt = OpSafeForPHIOfOps.find(OrigI);
2624  if (OISIt != OpSafeForPHIOfOps.end()) {
2625  if (!OISIt->second) {
2626  OpSafeForPHIOfOps.insert({V, false});
2627  return false;
2628  }
2629  continue;
2630  }
2631  if (!Visited.insert(Op).second)
2632  continue;
2633  Worklist.push_back(cast<Instruction>(Op));
2634  }
2635  return true;
2636 }
2637 
2638 // Return true if this operand will be safe to use for phi of ops.
2639 //
2640 // The reason some operands are unsafe is that we are not trying to recursively
2641 // translate everything back through phi nodes. We actually expect some lookups
2642 // of expressions to fail. In particular, a lookup where the expression cannot
2643 // exist in the predecessor. This is true even if the expression, as shown, can
2644 // be determined to be constant.
2645 bool NewGVN::OpIsSafeForPHIOfOps(Value *V, const BasicBlock *PHIBlock,
2646  SmallPtrSetImpl<const Value *> &Visited) {
2648  if (!OpIsSafeForPHIOfOpsHelper(V, PHIBlock, Visited, Worklist))
2649  return false;
2650  while (!Worklist.empty()) {
2651  auto *I = Worklist.pop_back_val();
2652  if (!OpIsSafeForPHIOfOpsHelper(I, PHIBlock, Visited, Worklist))
2653  return false;
2654  }
2655  OpSafeForPHIOfOps.insert({V, true});
2656  return true;
2657 }
2658 
2659 // Try to find a leader for instruction TransInst, which is a phi translated
2660 // version of something in our original program. Visited is used to ensure we
2661 // don't infinite loop during translations of cycles. OrigInst is the
2662 // instruction in the original program, and PredBB is the predecessor we
2663 // translated it through.
2664 Value *NewGVN::findLeaderForInst(Instruction *TransInst,
2665  SmallPtrSetImpl<Value *> &Visited,
2666  MemoryAccess *MemAccess, Instruction *OrigInst,
2667  BasicBlock *PredBB) {
2668  unsigned IDFSNum = InstrToDFSNum(OrigInst);
2669  // Make sure it's marked as a temporary instruction.
2670  AllTempInstructions.insert(TransInst);
2671  // and make sure anything that tries to add it's DFS number is
2672  // redirected to the instruction we are making a phi of ops
2673  // for.
2674  TempToBlock.insert({TransInst, PredBB});
2675  InstrDFS.insert({TransInst, IDFSNum});
2676 
2677  const Expression *E = performSymbolicEvaluation(TransInst, Visited);
2678  InstrDFS.erase(TransInst);
2679  AllTempInstructions.erase(TransInst);
2680  TempToBlock.erase(TransInst);
2681  if (MemAccess)
2682  TempToMemory.erase(TransInst);
2683  if (!E)
2684  return nullptr;
2685  auto *FoundVal = findPHIOfOpsLeader(E, OrigInst, PredBB);
2686  if (!FoundVal) {
2687  ExpressionToPhiOfOps[E].insert(OrigInst);
2688  DEBUG(dbgs() << "Cannot find phi of ops operand for " << *TransInst
2689  << " in block " << getBlockName(PredBB) << "\n");
2690  return nullptr;
2691  }
2692  if (auto *SI = dyn_cast<StoreInst>(FoundVal))
2693  FoundVal = SI->getValueOperand();
2694  return FoundVal;
2695 }
2696 
2697 // When we see an instruction that is an op of phis, generate the equivalent phi
2698 // of ops form.
2699 const Expression *
2700 NewGVN::makePossiblePHIOfOps(Instruction *I,
2701  SmallPtrSetImpl<Value *> &Visited) {
2702  if (!okayForPHIOfOps(I))
2703  return nullptr;
2704 
2705  if (!Visited.insert(I).second)
2706  return nullptr;
2707  // For now, we require the instruction be cycle free because we don't
2708  // *always* create a phi of ops for instructions that could be done as phi
2709  // of ops, we only do it if we think it is useful. If we did do it all the
2710  // time, we could remove the cycle free check.
2711  if (!isCycleFree(I))
2712  return nullptr;
2713 
2714  SmallPtrSet<const Value *, 8> ProcessedPHIs;
2715  // TODO: We don't do phi translation on memory accesses because it's
2716  // complicated. For a load, we'd need to be able to simulate a new memoryuse,
2717  // which we don't have a good way of doing ATM.
2718  auto *MemAccess = getMemoryAccess(I);
2719  // If the memory operation is defined by a memory operation this block that
2720  // isn't a MemoryPhi, transforming the pointer backwards through a scalar phi
2721  // can't help, as it would still be killed by that memory operation.
2722  if (MemAccess && !isa<MemoryPhi>(MemAccess->getDefiningAccess()) &&
2723  MemAccess->getDefiningAccess()->getBlock() == I->getParent())
2724  return nullptr;
2725 
2726  SmallPtrSet<const Value *, 10> VisitedOps;
2727  // Convert op of phis to phi of ops
2728  for (auto *Op : I->operand_values()) {
2729  if (!isa<PHINode>(Op)) {
2730  auto *ValuePHI = RealToTemp.lookup(Op);
2731  if (!ValuePHI)
2732  continue;
2733  DEBUG(dbgs() << "Found possible dependent phi of ops\n");
2734  Op = ValuePHI;
2735  }
2736  auto *OpPHI = cast<PHINode>(Op);
2737  // No point in doing this for one-operand phis.
2738  if (OpPHI->getNumOperands() == 1)
2739  continue;
2740  if (!DebugCounter::shouldExecute(PHIOfOpsCounter))
2741  return nullptr;
2744  auto *PHIBlock = getBlockForValue(OpPHI);
2745  RevisitOnReachabilityChange[PHIBlock].reset(InstrToDFSNum(I));
2746  for (unsigned PredNum = 0; PredNum < OpPHI->getNumOperands(); ++PredNum) {
2747  auto *PredBB = OpPHI->getIncomingBlock(PredNum);
2748  Value *FoundVal = nullptr;
2749  // We could just skip unreachable edges entirely but it's tricky to do
2750  // with rewriting existing phi nodes.
2751  if (ReachableEdges.count({PredBB, PHIBlock})) {
2752  // Clone the instruction, create an expression from it that is
2753  // translated back into the predecessor, and see if we have a leader.
2754  Instruction *ValueOp = I->clone();
2755  if (MemAccess)
2756  TempToMemory.insert({ValueOp, MemAccess});
2757  bool SafeForPHIOfOps = true;
2758  VisitedOps.clear();
2759  for (auto &Op : ValueOp->operands()) {
2760  auto *OrigOp = &*Op;
2761  // When these operand changes, it could change whether there is a
2762  // leader for us or not, so we have to add additional users.
2763  if (isa<PHINode>(Op)) {
2764  Op = Op->DoPHITranslation(PHIBlock, PredBB);
2765  if (Op != OrigOp && Op != I)
2766  Deps.insert(Op);
2767  } else if (auto *ValuePHI = RealToTemp.lookup(Op)) {
2768  if (getBlockForValue(ValuePHI) == PHIBlock)
2769  Op = ValuePHI->getIncomingValueForBlock(PredBB);
2770  }
2771  // If we phi-translated the op, it must be safe.
2772  SafeForPHIOfOps =
2773  SafeForPHIOfOps &&
2774  (Op != OrigOp || OpIsSafeForPHIOfOps(Op, PHIBlock, VisitedOps));
2775  }
2776  // FIXME: For those things that are not safe we could generate
2777  // expressions all the way down, and see if this comes out to a
2778  // constant. For anything where that is true, and unsafe, we should
2779  // have made a phi-of-ops (or value numbered it equivalent to something)
2780  // for the pieces already.
2781  FoundVal = !SafeForPHIOfOps ? nullptr
2782  : findLeaderForInst(ValueOp, Visited,
2783  MemAccess, I, PredBB);
2784  ValueOp->deleteValue();
2785  if (!FoundVal)
2786  return nullptr;
2787  } else {
2788  DEBUG(dbgs() << "Skipping phi of ops operand for incoming block "
2789  << getBlockName(PredBB)
2790  << " because the block is unreachable\n");
2791  FoundVal = UndefValue::get(I->getType());
2792  RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2793  }
2794 
2795  Ops.push_back({FoundVal, PredBB});
2796  DEBUG(dbgs() << "Found phi of ops operand " << *FoundVal << " in "
2797  << getBlockName(PredBB) << "\n");
2798  }
2799  for (auto Dep : Deps)
2800  addAdditionalUsers(Dep, I);
2801  sortPHIOps(Ops);
2802  auto *E = performSymbolicPHIEvaluation(Ops, I, PHIBlock);
2803  if (isa<ConstantExpression>(E) || isa<VariableExpression>(E)) {
2804  DEBUG(dbgs()
2805  << "Not creating real PHI of ops because it simplified to existing "
2806  "value or constant\n");
2807  return E;
2808  }
2809  auto *ValuePHI = RealToTemp.lookup(I);
2810  bool NewPHI = false;
2811  if (!ValuePHI) {
2812  ValuePHI =
2813  PHINode::Create(I->getType(), OpPHI->getNumOperands(), "phiofops");
2814  addPhiOfOps(ValuePHI, PHIBlock, I);
2815  NewPHI = true;
2816  NumGVNPHIOfOpsCreated++;
2817  }
2818  if (NewPHI) {
2819  for (auto PHIOp : Ops)
2820  ValuePHI->addIncoming(PHIOp.first, PHIOp.second);
2821  } else {
2822  unsigned int i = 0;
2823  for (auto PHIOp : Ops) {
2824  ValuePHI->setIncomingValue(i, PHIOp.first);
2825  ValuePHI->setIncomingBlock(i, PHIOp.second);
2826  ++i;
2827  }
2828  }
2829  RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2830  DEBUG(dbgs() << "Created phi of ops " << *ValuePHI << " for " << *I
2831  << "\n");
2832 
2833  return E;
2834  }
2835  return nullptr;
2836 }
2837 
2838 // The algorithm initially places the values of the routine in the TOP
2839 // congruence class. The leader of TOP is the undetermined value `undef`.
2840 // When the algorithm has finished, values still in TOP are unreachable.
2841 void NewGVN::initializeCongruenceClasses(Function &F) {
2842  NextCongruenceNum = 0;
2843 
2844  // Note that even though we use the live on entry def as a representative
2845  // MemoryAccess, it is *not* the same as the actual live on entry def. We
2846  // have no real equivalemnt to undef for MemoryAccesses, and so we really
2847  // should be checking whether the MemoryAccess is top if we want to know if it
2848  // is equivalent to everything. Otherwise, what this really signifies is that
2849  // the access "it reaches all the way back to the beginning of the function"
2850 
2851  // Initialize all other instructions to be in TOP class.
2852  TOPClass = createCongruenceClass(nullptr, nullptr);
2853  TOPClass->setMemoryLeader(MSSA->getLiveOnEntryDef());
2854  // The live on entry def gets put into it's own class
2855  MemoryAccessToClass[MSSA->getLiveOnEntryDef()] =
2856  createMemoryClass(MSSA->getLiveOnEntryDef());
2857 
2858  for (auto DTN : nodes(DT)) {
2859  BasicBlock *BB = DTN->getBlock();
2860  // All MemoryAccesses are equivalent to live on entry to start. They must
2861  // be initialized to something so that initial changes are noticed. For
2862  // the maximal answer, we initialize them all to be the same as
2863  // liveOnEntry.
2864  auto *MemoryBlockDefs = MSSA->getBlockDefs(BB);
2865  if (MemoryBlockDefs)
2866  for (const auto &Def : *MemoryBlockDefs) {
2867  MemoryAccessToClass[&Def] = TOPClass;
2868  auto *MD = dyn_cast<MemoryDef>(&Def);
2869  // Insert the memory phis into the member list.
2870  if (!MD) {
2871  const MemoryPhi *MP = cast<MemoryPhi>(&Def);
2872  TOPClass->memory_insert(MP);
2873  MemoryPhiState.insert({MP, MPS_TOP});
2874  }
2875 
2876  if (MD && isa<StoreInst>(MD->getMemoryInst()))
2877  TOPClass->incStoreCount();
2878  }
2879 
2880  // FIXME: This is trying to discover which instructions are uses of phi
2881  // nodes. We should move this into one of the myriad of places that walk
2882  // all the operands already.
2883  for (auto &I : *BB) {
2884  if (isa<PHINode>(&I))
2885  for (auto *U : I.users())
2886  if (auto *UInst = dyn_cast<Instruction>(U))
2887  if (InstrToDFSNum(UInst) != 0 && okayForPHIOfOps(UInst))
2888  PHINodeUses.insert(UInst);
2889  // Don't insert void terminators into the class. We don't value number
2890  // them, and they just end up sitting in TOP.
2891  if (isa<TerminatorInst>(I) && I.getType()->isVoidTy())
2892  continue;
2893  TOPClass->insert(&I);
2894  ValueToClass[&I] = TOPClass;
2895  }
2896  }
2897 
2898  // Initialize arguments to be in their own unique congruence classes
2899  for (auto &FA : F.args())
2900  createSingletonCongruenceClass(&FA);
2901 }
2902 
2903 void NewGVN::cleanupTables() {
2904  for (unsigned i = 0, e = CongruenceClasses.size(); i != e; ++i) {
2905  DEBUG(dbgs() << "Congruence class " << CongruenceClasses[i]->getID()
2906  << " has " << CongruenceClasses[i]->size() << " members\n");
2907  // Make sure we delete the congruence class (probably worth switching to
2908  // a unique_ptr at some point.
2909  delete CongruenceClasses[i];
2910  CongruenceClasses[i] = nullptr;
2911  }
2912 
2913  // Destroy the value expressions
2914  SmallVector<Instruction *, 8> TempInst(AllTempInstructions.begin(),
2915  AllTempInstructions.end());
2916  AllTempInstructions.clear();
2917 
2918  // We have to drop all references for everything first, so there are no uses
2919  // left as we delete them.
2920  for (auto *I : TempInst) {
2921  I->dropAllReferences();
2922  }
2923 
2924  while (!TempInst.empty()) {
2925  auto *I = TempInst.back();
2926  TempInst.pop_back();
2927  I->deleteValue();
2928  }
2929 
2930  ValueToClass.clear();
2931  ArgRecycler.clear(ExpressionAllocator);
2932  ExpressionAllocator.Reset();
2933  CongruenceClasses.clear();
2934  ExpressionToClass.clear();
2935  ValueToExpression.clear();
2936  RealToTemp.clear();
2937  AdditionalUsers.clear();
2938  ExpressionToPhiOfOps.clear();
2939  TempToBlock.clear();
2940  TempToMemory.clear();
2941  PHINodeUses.clear();
2942  OpSafeForPHIOfOps.clear();
2943  ReachableBlocks.clear();
2944  ReachableEdges.clear();
2945 #ifndef NDEBUG
2946  ProcessedCount.clear();
2947 #endif
2948  InstrDFS.clear();
2949  InstructionsToErase.clear();
2950  DFSToInstr.clear();
2951  BlockInstRange.clear();
2952  TouchedInstructions.clear();
2953  MemoryAccessToClass.clear();
2954  PredicateToUsers.clear();
2955  MemoryToUsers.clear();
2956  RevisitOnReachabilityChange.clear();
2957 }
2958 
2959 // Assign local DFS number mapping to instructions, and leave space for Value
2960 // PHI's.
2961 std::pair<unsigned, unsigned> NewGVN::assignDFSNumbers(BasicBlock *B,
2962  unsigned Start) {
2963  unsigned End = Start;
2964  if (MemoryAccess *MemPhi = getMemoryAccess(B)) {
2965  InstrDFS[MemPhi] = End++;
2966  DFSToInstr.emplace_back(MemPhi);
2967  }
2968 
2969  // Then the real block goes next.
2970  for (auto &I : *B) {
2971  // There's no need to call isInstructionTriviallyDead more than once on
2972  // an instruction. Therefore, once we know that an instruction is dead
2973  // we change its DFS number so that it doesn't get value numbered.
2974  if (isInstructionTriviallyDead(&I, TLI)) {
2975  InstrDFS[&I] = 0;
2976  DEBUG(dbgs() << "Skipping trivially dead instruction " << I << "\n");
2977  markInstructionForDeletion(&I);
2978  continue;
2979  }
2980  if (isa<PHINode>(&I))
2981  RevisitOnReachabilityChange[B].set(End);
2982  InstrDFS[&I] = End++;
2983  DFSToInstr.emplace_back(&I);
2984  }
2985 
2986  // All of the range functions taken half-open ranges (open on the end side).
2987  // So we do not subtract one from count, because at this point it is one
2988  // greater than the last instruction.
2989  return std::make_pair(Start, End);
2990 }
2991 
2992 void NewGVN::updateProcessedCount(const Value *V) {
2993 #ifndef NDEBUG
2994  if (ProcessedCount.count(V) == 0) {
2995  ProcessedCount.insert({V, 1});
2996  } else {
2997  ++ProcessedCount[V];
2998  assert(ProcessedCount[V] < 100 &&
2999  "Seem to have processed the same Value a lot");
3000  }
3001 #endif
3002 }
3003 
3004 // Evaluate MemoryPhi nodes symbolically, just like PHI nodes
3005 void NewGVN::valueNumberMemoryPhi(MemoryPhi *MP) {
3006  // If all the arguments are the same, the MemoryPhi has the same value as the
3007  // argument. Filter out unreachable blocks and self phis from our operands.
3008  // TODO: We could do cycle-checking on the memory phis to allow valueizing for
3009  // self-phi checking.
3010  const BasicBlock *PHIBlock = MP->getBlock();
3011  auto Filtered = make_filter_range(MP->operands(), [&](const Use &U) {
3012  return cast<MemoryAccess>(U) != MP &&
3013  !isMemoryAccessTOP(cast<MemoryAccess>(U)) &&
3014  ReachableEdges.count({MP->getIncomingBlock(U), PHIBlock});
3015  });
3016  // If all that is left is nothing, our memoryphi is undef. We keep it as
3017  // InitialClass. Note: The only case this should happen is if we have at
3018  // least one self-argument.
3019  if (Filtered.begin() == Filtered.end()) {
3020  if (setMemoryClass(MP, TOPClass))
3021  markMemoryUsersTouched(MP);
3022  return;
3023  }
3024 
3025  // Transform the remaining operands into operand leaders.
3026  // FIXME: mapped_iterator should have a range version.
3027  auto LookupFunc = [&](const Use &U) {
3028  return lookupMemoryLeader(cast<MemoryAccess>(U));
3029  };
3030  auto MappedBegin = map_iterator(Filtered.begin(), LookupFunc);
3031  auto MappedEnd = map_iterator(Filtered.end(), LookupFunc);
3032 
3033  // and now check if all the elements are equal.
3034  // Sadly, we can't use std::equals since these are random access iterators.
3035  const auto *AllSameValue = *MappedBegin;
3036  ++MappedBegin;
3037  bool AllEqual = std::all_of(
3038  MappedBegin, MappedEnd,
3039  [&AllSameValue](const MemoryAccess *V) { return V == AllSameValue; });
3040 
3041  if (AllEqual)
3042  DEBUG(dbgs() << "Memory Phi value numbered to " << *AllSameValue << "\n");
3043  else
3044  DEBUG(dbgs() << "Memory Phi value numbered to itself\n");
3045  // If it's equal to something, it's in that class. Otherwise, it has to be in
3046  // a class where it is the leader (other things may be equivalent to it, but
3047  // it needs to start off in its own class, which means it must have been the
3048  // leader, and it can't have stopped being the leader because it was never
3049  // removed).
3050  CongruenceClass *CC =
3051  AllEqual ? getMemoryClass(AllSameValue) : ensureLeaderOfMemoryClass(MP);
3052  auto OldState = MemoryPhiState.lookup(MP);
3053  assert(OldState != MPS_Invalid && "Invalid memory phi state");
3054  auto NewState = AllEqual ? MPS_Equivalent : MPS_Unique;
3055  MemoryPhiState[MP] = NewState;
3056  if (setMemoryClass(MP, CC) || OldState != NewState)
3057  markMemoryUsersTouched(MP);
3058 }
3059 
3060 // Value number a single instruction, symbolically evaluating, performing
3061 // congruence finding, and updating mappings.
3062 void NewGVN::valueNumberInstruction(Instruction *I) {
3063  DEBUG(dbgs() << "Processing instruction " << *I << "\n");
3064  if (!I->isTerminator()) {
3065  const Expression *Symbolized = nullptr;
3066  SmallPtrSet<Value *, 2> Visited;
3067  if (DebugCounter::shouldExecute(VNCounter)) {
3068  Symbolized = performSymbolicEvaluation(I, Visited);
3069  // Make a phi of ops if necessary
3070  if (Symbolized && !isa<ConstantExpression>(Symbolized) &&
3071  !isa<VariableExpression>(Symbolized) && PHINodeUses.count(I)) {
3072  auto *PHIE = makePossiblePHIOfOps(I, Visited);
3073  // If we created a phi of ops, use it.
3074  // If we couldn't create one, make sure we don't leave one lying around
3075  if (PHIE) {
3076  Symbolized = PHIE;
3077  } else if (auto *Op = RealToTemp.lookup(I)) {
3078  removePhiOfOps(I, Op);
3079  }
3080  }
3081  } else {
3082  // Mark the instruction as unused so we don't value number it again.
3083  InstrDFS[I] = 0;
3084  }
3085  // If we couldn't come up with a symbolic expression, use the unknown
3086  // expression
3087  if (Symbolized == nullptr)
3088  Symbolized = createUnknownExpression(I);
3089  performCongruenceFinding(I, Symbolized);
3090  } else {
3091  // Handle terminators that return values. All of them produce values we
3092  // don't currently understand. We don't place non-value producing
3093  // terminators in a class.
3094  if (!I->getType()->isVoidTy()) {
3095  auto *Symbolized = createUnknownExpression(I);
3096  performCongruenceFinding(I, Symbolized);
3097  }
3098  processOutgoingEdges(dyn_cast<TerminatorInst>(I), I->getParent());
3099  }
3100 }
3101 
3102 // Check if there is a path, using single or equal argument phi nodes, from
3103 // First to Second.
3104 bool NewGVN::singleReachablePHIPath(
3105  SmallPtrSet<const MemoryAccess *, 8> &Visited, const MemoryAccess *First,
3106  const MemoryAccess *Second) const {
3107  if (First == Second)
3108  return true;
3109  if (MSSA->isLiveOnEntryDef(First))
3110  return false;
3111 
3112  // This is not perfect, but as we're just verifying here, we can live with
3113  // the loss of precision. The real solution would be that of doing strongly
3114  // connected component finding in this routine, and it's probably not worth
3115  // the complexity for the time being. So, we just keep a set of visited
3116  // MemoryAccess and return true when we hit a cycle.
3117  if (Visited.count(First))
3118  return true;
3119  Visited.insert(First);
3120 
3121  const auto *EndDef = First;
3122  for (auto *ChainDef : optimized_def_chain(First)) {
3123  if (ChainDef == Second)
3124  return true;
3125  if (MSSA->isLiveOnEntryDef(ChainDef))
3126  return false;
3127  EndDef = ChainDef;
3128  }
3129  auto *MP = cast<MemoryPhi>(EndDef);
3130  auto ReachableOperandPred = [&](const Use &U) {
3131  return ReachableEdges.count({MP->getIncomingBlock(U), MP->getBlock()});
3132  };
3133  auto FilteredPhiArgs =
3134  make_filter_range(MP->operands(), ReachableOperandPred);
3135  SmallVector<const Value *, 32> OperandList;
3136  std::copy(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
3137  std::back_inserter(OperandList));
3138  bool Okay = OperandList.size() == 1;
3139  if (!Okay)
3140  Okay =
3141  std::equal(OperandList.begin(), OperandList.end(), OperandList.begin());
3142  if (Okay)
3143  return singleReachablePHIPath(Visited, cast<MemoryAccess>(OperandList[0]),
3144  Second);
3145  return false;
3146 }
3147 
3148 // Verify the that the memory equivalence table makes sense relative to the
3149 // congruence classes. Note that this checking is not perfect, and is currently
3150 // subject to very rare false negatives. It is only useful for
3151 // testing/debugging.
3152 void NewGVN::verifyMemoryCongruency() const {
3153 #ifndef NDEBUG
3154  // Verify that the memory table equivalence and memory member set match
3155  for (const auto *CC : CongruenceClasses) {
3156  if (CC == TOPClass || CC->isDead())
3157  continue;
3158  if (CC->getStoreCount() != 0) {
3159  assert((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) &&
3160  "Any class with a store as a leader should have a "
3161  "representative stored value");
3162  assert(CC->getMemoryLeader() &&
3163  "Any congruence class with a store should have a "
3164  "representative access");
3165  }
3166 
3167  if (CC->getMemoryLeader())
3168  assert(MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
3169  "Representative MemoryAccess does not appear to be reverse "
3170  "mapped properly");
3171  for (auto M : CC->memory())
3172  assert(MemoryAccessToClass.lookup(M) == CC &&
3173  "Memory member does not appear to be reverse mapped properly");
3174  }
3175 
3176  // Anything equivalent in the MemoryAccess table should be in the same
3177  // congruence class.
3178 
3179  // Filter out the unreachable and trivially dead entries, because they may
3180  // never have been updated if the instructions were not processed.
3181  auto ReachableAccessPred =
3182  [&](const std::pair<const MemoryAccess *, CongruenceClass *> Pair) {
3183  bool Result = ReachableBlocks.count(Pair.first->getBlock());
3184  if (!Result || MSSA->isLiveOnEntryDef(Pair.first) ||
3185  MemoryToDFSNum(Pair.first) == 0)
3186  return false;
3187  if (auto *MemDef = dyn_cast<MemoryDef>(Pair.first))
3188  return !isInstructionTriviallyDead(MemDef->getMemoryInst());
3189 
3190  // We could have phi nodes which operands are all trivially dead,
3191  // so we don't process them.
3192  if (auto *MemPHI = dyn_cast<MemoryPhi>(Pair.first)) {
3193  for (auto &U : MemPHI->incoming_values()) {
3194  if (auto *I = dyn_cast<Instruction>(&*U)) {
3196  return true;
3197  }
3198  }
3199  return false;
3200  }
3201 
3202  return true;
3203  };
3204 
3205  auto Filtered = make_filter_range(MemoryAccessToClass, ReachableAccessPred);
3206  for (auto KV : Filtered) {
3207  if (auto *FirstMUD = dyn_cast<MemoryUseOrDef>(KV.first)) {
3208  auto *SecondMUD = dyn_cast<MemoryUseOrDef>(KV.second->getMemoryLeader());
3209  if (FirstMUD && SecondMUD) {
3211  assert((singleReachablePHIPath(VisitedMAS, FirstMUD, SecondMUD) ||
3212  ValueToClass.lookup(FirstMUD->getMemoryInst()) ==
3213  ValueToClass.lookup(SecondMUD->getMemoryInst())) &&
3214  "The instructions for these memory operations should have "
3215  "been in the same congruence class or reachable through"
3216  "a single argument phi");
3217  }
3218  } else if (auto *FirstMP = dyn_cast<MemoryPhi>(KV.first)) {
3219  // We can only sanely verify that MemoryDefs in the operand list all have
3220  // the same class.
3221  auto ReachableOperandPred = [&](const Use &U) {
3222  return ReachableEdges.count(
3223  {FirstMP->getIncomingBlock(U), FirstMP->getBlock()}) &&
3224  isa<MemoryDef>(U);
3225 
3226  };
3227  // All arguments should in the same class, ignoring unreachable arguments
3228  auto FilteredPhiArgs =
3229  make_filter_range(FirstMP->operands(), ReachableOperandPred);
3231  std::transform(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
3232  std::back_inserter(PhiOpClasses), [&](const Use &U) {
3233  const MemoryDef *MD = cast<MemoryDef>(U);
3234  return ValueToClass.lookup(MD->getMemoryInst());
3235  });
3236  assert(std::equal(PhiOpClasses.begin(), PhiOpClasses.end(),
3237  PhiOpClasses.begin()) &&
3238  "All MemoryPhi arguments should be in the same class");
3239  }
3240  }
3241 #endif
3242 }
3243 
3244 // Verify that the sparse propagation we did actually found the maximal fixpoint
3245 // We do this by storing the value to class mapping, touching all instructions,
3246 // and redoing the iteration to see if anything changed.
3247 void NewGVN::verifyIterationSettled(Function &F) {
3248 #ifndef NDEBUG
3249  DEBUG(dbgs() << "Beginning iteration verification\n");
3250  if (DebugCounter::isCounterSet(VNCounter))
3251  DebugCounter::setCounterValue(VNCounter, StartingVNCounter);
3252 
3253  // Note that we have to store the actual classes, as we may change existing
3254  // classes during iteration. This is because our memory iteration propagation
3255  // is not perfect, and so may waste a little work. But it should generate
3256  // exactly the same congruence classes we have now, with different IDs.
3257  std::map<const Value *, CongruenceClass> BeforeIteration;
3258 
3259  for (auto &KV : ValueToClass) {
3260  if (auto *I = dyn_cast<Instruction>(KV.first))
3261  // Skip unused/dead instructions.
3262  if (InstrToDFSNum(I) == 0)
3263  continue;
3264  BeforeIteration.insert({KV.first, *KV.second});
3265  }
3266 
3267  TouchedInstructions.set();
3268  TouchedInstructions.reset(0);
3269  iterateTouchedInstructions();
3271  EqualClasses;
3272  for (const auto &KV : ValueToClass) {
3273  if (auto *I = dyn_cast<Instruction>(KV.first))
3274  // Skip unused/dead instructions.
3275  if (InstrToDFSNum(I) == 0)
3276  continue;
3277  // We could sink these uses, but i think this adds a bit of clarity here as
3278  // to what we are comparing.
3279  auto *BeforeCC = &BeforeIteration.find(KV.first)->second;
3280  auto *AfterCC = KV.second;
3281  // Note that the classes can't change at this point, so we memoize the set
3282  // that are equal.
3283  if (!EqualClasses.count({BeforeCC, AfterCC})) {
3284  assert(BeforeCC->isEquivalentTo(AfterCC) &&
3285  "Value number changed after main loop completed!");
3286  EqualClasses.insert({BeforeCC, AfterCC});
3287  }
3288  }
3289 #endif
3290 }
3291 
3292 // Verify that for each store expression in the expression to class mapping,
3293 // only the latest appears, and multiple ones do not appear.
3294 // Because loads do not use the stored value when doing equality with stores,
3295 // if we don't erase the old store expressions from the table, a load can find
3296 // a no-longer valid StoreExpression.
3297 void NewGVN::verifyStoreExpressions() const {
3298 #ifndef NDEBUG
3299  // This is the only use of this, and it's not worth defining a complicated
3300  // densemapinfo hash/equality function for it.
3301  std::set<
3302  std::pair<const Value *,
3303  std::tuple<const Value *, const CongruenceClass *, Value *>>>
3304  StoreExpressionSet;
3305  for (const auto &KV : ExpressionToClass) {
3306  if (auto *SE = dyn_cast<StoreExpression>(KV.first)) {
3307  // Make sure a version that will conflict with loads is not already there
3308  auto Res = StoreExpressionSet.insert(
3309  {SE->getOperand(0), std::make_tuple(SE->getMemoryLeader(), KV.second,
3310  SE->getStoredValue())});
3311  bool Okay = Res.second;
3312  // It's okay to have the same expression already in there if it is
3313  // identical in nature.
3314  // This can happen when the leader of the stored value changes over time.
3315  if (!Okay)
3316  Okay = (std::get<1>(Res.first->second) == KV.second) &&
3317  (lookupOperandLeader(std::get<2>(Res.first->second)) ==
3318  lookupOperandLeader(SE->getStoredValue()));
3319  assert(Okay && "Stored expression conflict exists in expression table");
3320  auto *ValueExpr = ValueToExpression.lookup(SE->getStoreInst());
3321  assert(ValueExpr && ValueExpr->equals(*SE) &&
3322  "StoreExpression in ExpressionToClass is not latest "
3323  "StoreExpression for value");
3324  }
3325  }
3326 #endif
3327 }
3328 
3329 // This is the main value numbering loop, it iterates over the initial touched
3330 // instruction set, propagating value numbers, marking things touched, etc,
3331 // until the set of touched instructions is completely empty.
3332 void NewGVN::iterateTouchedInstructions() {
3333  unsigned int Iterations = 0;
3334  // Figure out where touchedinstructions starts
3335  int FirstInstr = TouchedInstructions.find_first();
3336  // Nothing set, nothing to iterate, just return.
3337  if (FirstInstr == -1)
3338  return;
3339  const BasicBlock *LastBlock = getBlockForValue(InstrFromDFSNum(FirstInstr));
3340  while (TouchedInstructions.any()) {
3341  ++Iterations;
3342  // Walk through all the instructions in all the blocks in RPO.
3343  // TODO: As we hit a new block, we should push and pop equalities into a
3344  // table lookupOperandLeader can use, to catch things PredicateInfo
3345  // might miss, like edge-only equivalences.
3346  for (unsigned InstrNum : TouchedInstructions.set_bits()) {
3347 
3348  // This instruction was found to be dead. We don't bother looking
3349  // at it again.
3350  if (InstrNum == 0) {
3351  TouchedInstructions.reset(InstrNum);
3352  continue;
3353  }
3354 
3355  Value *V = InstrFromDFSNum(InstrNum);
3356  const BasicBlock *CurrBlock = getBlockForValue(V);
3357 
3358  // If we hit a new block, do reachability processing.
3359  if (CurrBlock != LastBlock) {
3360  LastBlock = CurrBlock;
3361  bool BlockReachable = ReachableBlocks.count(CurrBlock);
3362  const auto &CurrInstRange = BlockInstRange.lookup(CurrBlock);
3363 
3364  // If it's not reachable, erase any touched instructions and move on.
3365  if (!BlockReachable) {
3366  TouchedInstructions.reset(CurrInstRange.first, CurrInstRange.second);
3367  DEBUG(dbgs() << "Skipping instructions in block "
3368  << getBlockName(CurrBlock)
3369  << " because it is unreachable\n");
3370  continue;
3371  }
3372  updateProcessedCount(CurrBlock);
3373  }
3374  // Reset after processing (because we may mark ourselves as touched when
3375  // we propagate equalities).
3376  TouchedInstructions.reset(InstrNum);
3377 
3378  if (auto *MP = dyn_cast<MemoryPhi>(V)) {
3379  DEBUG(dbgs() << "Processing MemoryPhi " << *MP << "\n");
3380  valueNumberMemoryPhi(MP);
3381  } else if (auto *I = dyn_cast<Instruction>(V)) {
3382  valueNumberInstruction(I);
3383  } else {
3384  llvm_unreachable("Should have been a MemoryPhi or Instruction");
3385  }
3386  updateProcessedCount(V);
3387  }
3388  }
3389  NumGVNMaxIterations = std::max(NumGVNMaxIterations.getValue(), Iterations);
3390 }
3391 
3392 // This is the main transformation entry point.
3393 bool NewGVN::runGVN() {
3394  if (DebugCounter::isCounterSet(VNCounter))
3395  StartingVNCounter = DebugCounter::getCounterValue(VNCounter);
3396  bool Changed = false;
3397  NumFuncArgs = F.arg_size();
3398  MSSAWalker = MSSA->getWalker();
3399  SingletonDeadExpression = new (ExpressionAllocator) DeadExpression();
3400 
3401  // Count number of instructions for sizing of hash tables, and come
3402  // up with a global dfs numbering for instructions.
3403  unsigned ICount = 1;
3404  // Add an empty instruction to account for the fact that we start at 1
3405  DFSToInstr.emplace_back(nullptr);
3406  // Note: We want ideal RPO traversal of the blocks, which is not quite the
3407  // same as dominator tree order, particularly with regard whether backedges
3408  // get visited first or second, given a block with multiple successors.
3409  // If we visit in the wrong order, we will end up performing N times as many
3410  // iterations.
3411  // The dominator tree does guarantee that, for a given dom tree node, it's
3412  // parent must occur before it in the RPO ordering. Thus, we only need to sort
3413  // the siblings.
3415  unsigned Counter = 0;
3416  for (auto &B : RPOT) {
3417  auto *Node = DT->getNode(B);
3418  assert(Node && "RPO and Dominator tree should have same reachability");
3419  RPOOrdering[Node] = ++Counter;
3420  }
3421  // Sort dominator tree children arrays into RPO.
3422  for (auto &B : RPOT) {
3423  auto *Node = DT->getNode(B);
3424  if (Node->getChildren().size() > 1)
3425  std::sort(Node->begin(), Node->end(),
3426  [&](const DomTreeNode *A, const DomTreeNode *B) {
3427  return RPOOrdering[A] < RPOOrdering[B];
3428  });
3429  }
3430 
3431  // Now a standard depth first ordering of the domtree is equivalent to RPO.
3432  for (auto DTN : depth_first(DT->getRootNode())) {
3433  BasicBlock *B = DTN->getBlock();
3434  const auto &BlockRange = assignDFSNumbers(B, ICount);
3435  BlockInstRange.insert({B, BlockRange});
3436  ICount += BlockRange.second - BlockRange.first;
3437  }
3438  initializeCongruenceClasses(F);
3439 
3440  TouchedInstructions.resize(ICount);
3441  // Ensure we don't end up resizing the expressionToClass map, as
3442  // that can be quite expensive. At most, we have one expression per
3443  // instruction.
3444  ExpressionToClass.reserve(ICount);
3445 
3446  // Initialize the touched instructions to include the entry block.
3447  const auto &InstRange = BlockInstRange.lookup(&F.getEntryBlock());
3448  TouchedInstructions.set(InstRange.first, InstRange.second);
3449  DEBUG(dbgs() << "Block " << getBlockName(&F.getEntryBlock())
3450  << " marked reachable\n");
3451  ReachableBlocks.insert(&F.getEntryBlock());
3452 
3453  iterateTouchedInstructions();
3454  verifyMemoryCongruency();
3455  verifyIterationSettled(F);
3456  verifyStoreExpressions();
3457 
3458  Changed |= eliminateInstructions(F);
3459 
3460  // Delete all instructions marked for deletion.
3461  for (Instruction *ToErase : InstructionsToErase) {
3462  if (!ToErase->use_empty())
3463  ToErase->replaceAllUsesWith(UndefValue::get(ToErase->getType()));
3464 
3465  if (ToErase->getParent())
3466  ToErase->eraseFromParent();
3467  }
3468 
3469  // Delete all unreachable blocks.
3470  auto UnreachableBlockPred = [&](const BasicBlock &BB) {
3471  return !ReachableBlocks.count(&BB);
3472  };
3473 
3474  for (auto &BB : make_filter_range(F, UnreachableBlockPred)) {
3475  DEBUG(dbgs() << "We believe block " << getBlockName(&BB)
3476  << " is unreachable\n");
3477  deleteInstructionsInBlock(&BB);
3478  Changed = true;
3479  }
3480 
3481  cleanupTables();
3482  return Changed;
3483 }
3484 
3486  int DFSIn = 0;
3487  int DFSOut = 0;
3488  int LocalNum = 0;
3489 
3490  // Only one of Def and U will be set.
3491  // The bool in the Def tells us whether the Def is the stored value of a
3492  // store.
3494  Use *U = nullptr;
3495 
3496  bool operator<(const ValueDFS &Other) const {
3497  // It's not enough that any given field be less than - we have sets
3498  // of fields that need to be evaluated together to give a proper ordering.
3499  // For example, if you have;
3500  // DFS (1, 3)
3501  // Val 0
3502  // DFS (1, 2)
3503  // Val 50
3504  // We want the second to be less than the first, but if we just go field
3505  // by field, we will get to Val 0 < Val 50 and say the first is less than
3506  // the second. We only want it to be less than if the DFS orders are equal.
3507  //
3508  // Each LLVM instruction only produces one value, and thus the lowest-level
3509  // differentiator that really matters for the stack (and what we use as as a
3510  // replacement) is the local dfs number.
3511  // Everything else in the structure is instruction level, and only affects
3512  // the order in which we will replace operands of a given instruction.
3513  //
3514  // For a given instruction (IE things with equal dfsin, dfsout, localnum),
3515  // the order of replacement of uses does not matter.
3516  // IE given,
3517  // a = 5
3518  // b = a + a
3519  // When you hit b, you will have two valuedfs with the same dfsin, out, and
3520  // localnum.
3521  // The .val will be the same as well.
3522  // The .u's will be different.
3523  // You will replace both, and it does not matter what order you replace them
3524  // in (IE whether you replace operand 2, then operand 1, or operand 1, then
3525  // operand 2).
3526  // Similarly for the case of same dfsin, dfsout, localnum, but different
3527  // .val's
3528  // a = 5
3529  // b = 6
3530  // c = a + b
3531  // in c, we will a valuedfs for a, and one for b,with everything the same
3532  // but .val and .u.
3533  // It does not matter what order we replace these operands in.
3534  // You will always end up with the same IR, and this is guaranteed.
3535  return std::tie(DFSIn, DFSOut, LocalNum, Def, U) <
3536  std::tie(Other.DFSIn, Other.DFSOut, Other.LocalNum, Other.Def,
3537  Other.U);
3538  }
3539 };
3540 
3541 // This function converts the set of members for a congruence class from values,
3542 // to sets of defs and uses with associated DFS info. The total number of
3543 // reachable uses for each value is stored in UseCount, and instructions that
3544 // seem
3545 // dead (have no non-dead uses) are stored in ProbablyDead.
3546 void NewGVN::convertClassToDFSOrdered(
3547  const CongruenceClass &Dense, SmallVectorImpl<ValueDFS> &DFSOrderedSet,
3549  SmallPtrSetImpl<Instruction *> &ProbablyDead) const {
3550  for (auto D : Dense) {
3551  // First add the value.
3552  BasicBlock *BB = getBlockForValue(D);
3553  // Constants are handled prior to ever calling this function, so
3554  // we should only be left with instructions as members.
3555  assert(BB && "Should have figured out a basic block for value");
3556  ValueDFS VDDef;
3557  DomTreeNode *DomNode = DT->getNode(BB);
3558  VDDef.DFSIn = DomNode->getDFSNumIn();
3559  VDDef.DFSOut = DomNode->getDFSNumOut();
3560  // If it's a store, use the leader of the value operand, if it's always
3561  // available, or the value operand. TODO: We could do dominance checks to
3562  // find a dominating leader, but not worth it ATM.
3563  if (auto *SI = dyn_cast<StoreInst>(D)) {
3564  auto Leader = lookupOperandLeader(SI->getValueOperand());
3565  if (alwaysAvailable(Leader)) {
3566  VDDef.Def.setPointer(Leader);
3567  } else {
3568  VDDef.Def.setPointer(SI->getValueOperand());
3569  VDDef.Def.setInt(true);
3570  }
3571  } else {
3572  VDDef.Def.setPointer(D);
3573  }
3574  assert(isa<Instruction>(D) &&
3575  "The dense set member should always be an instruction");
3576  Instruction *Def = cast<Instruction>(D);
3577  VDDef.LocalNum = InstrToDFSNum(D);
3578  DFSOrderedSet.push_back(VDDef);
3579  // If there is a phi node equivalent, add it
3580  if (auto *PN = RealToTemp.lookup(Def)) {
3581  auto *PHIE =
3582  dyn_cast_or_null<PHIExpression>(ValueToExpression.lookup(Def));
3583  if (PHIE) {
3584  VDDef.Def.setInt(false);
3585  VDDef.Def.setPointer(PN);
3586  VDDef.LocalNum = 0;
3587  DFSOrderedSet.push_back(VDDef);
3588  }
3589  }
3590 
3591  unsigned int UseCount = 0;
3592  // Now add the uses.
3593  for (auto &U : Def->uses()) {
3594  if (auto *I = dyn_cast<Instruction>(U.getUser())) {
3595  // Don't try to replace into dead uses
3596  if (InstructionsToErase.count(I))
3597  continue;
3598  ValueDFS VDUse;
3599  // Put the phi node uses in the incoming block.
3600  BasicBlock *IBlock;
3601  if (auto *P = dyn_cast<PHINode>(I)) {
3602  IBlock = P->getIncomingBlock(U);
3603  // Make phi node users appear last in the incoming block
3604  // they are from.
3605  VDUse.LocalNum = InstrDFS.size() + 1;
3606  } else {
3607  IBlock = getBlockForValue(I);
3608  VDUse.LocalNum = InstrToDFSNum(I);
3609  }
3610 
3611  // Skip uses in unreachable blocks, as we're going
3612  // to delete them.
3613  if (ReachableBlocks.count(IBlock) == 0)
3614  continue;
3615 
3616  DomTreeNode *DomNode = DT->getNode(IBlock);
3617  VDUse.DFSIn = DomNode->getDFSNumIn();
3618  VDUse.DFSOut = DomNode->getDFSNumOut();
3619  VDUse.U = &U;
3620  ++UseCount;
3621  DFSOrderedSet.emplace_back(VDUse);
3622  }
3623  }
3624 
3625  // If there are no uses, it's probably dead (but it may have side-effects,
3626  // so not definitely dead. Otherwise, store the number of uses so we can
3627  // track if it becomes dead later).
3628  if (UseCount == 0)
3629  ProbablyDead.insert(Def);
3630  else
3631  UseCounts[Def] = UseCount;
3632  }
3633 }
3634 
3635 // This function converts the set of members for a congruence class from values,
3636 // to the set of defs for loads and stores, with associated DFS info.
3637 void NewGVN::convertClassToLoadsAndStores(
3638  const CongruenceClass &Dense,
3639  SmallVectorImpl<ValueDFS> &LoadsAndStores) const {
3640  for (auto D : Dense) {
3641  if (!isa<LoadInst>(D) && !isa<StoreInst>(D))
3642  continue;
3643 
3644  BasicBlock *BB = getBlockForValue(D);
3645  ValueDFS VD;
3646  DomTreeNode *DomNode = DT->getNode(BB);
3647  VD.DFSIn = DomNode->getDFSNumIn();
3648  VD.DFSOut = DomNode->getDFSNumOut();
3649  VD.Def.setPointer(D);
3650 
3651  // If it's an instruction, use the real local dfs number.
3652  if (auto *I = dyn_cast<Instruction>(D))
3653  VD.LocalNum = InstrToDFSNum(I);
3654  else
3655  llvm_unreachable("Should have been an instruction");
3656 
3657  LoadsAndStores.emplace_back(VD);
3658  }
3659 }
3660 
3662  auto *ReplInst = dyn_cast<Instruction>(Repl);
3663  if (!ReplInst)
3664  return;
3665 
3666  // Patch the replacement so that it is not more restrictive than the value
3667  // being replaced.
3668  // Note that if 'I' is a load being replaced by some operation,
3669  // for example, by an arithmetic operation, then andIRFlags()
3670  // would just erase all math flags from the original arithmetic
3671  // operation, which is clearly not wanted and not needed.
3672  if (!isa<LoadInst>(I))
3673  ReplInst->andIRFlags(I);
3674 
3675  // FIXME: If both the original and replacement value are part of the
3676  // same control-flow region (meaning that the execution of one
3677  // guarantees the execution of the other), then we can combine the
3678  // noalias scopes here and do better than the general conservative
3679  // answer used in combineMetadata().
3680 
3681  // In general, GVN unifies expressions over different control-flow
3682  // regions, and so we need a conservative combination of the noalias
3683  // scopes.
3684  static const unsigned KnownIDs[] = {
3689  combineMetadata(ReplInst, I, KnownIDs);
3690 }
3691 
3693  patchReplacementInstruction(I, Repl);
3694  I->replaceAllUsesWith(Repl);
3695 }
3696 
3697 void NewGVN::deleteInstructionsInBlock(BasicBlock *BB) {
3698  DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
3699  ++NumGVNBlocksDeleted;
3700 
3701  // Delete the instructions backwards, as it has a reduced likelihood of having
3702  // to update as many def-use and use-def chains. Start after the terminator.
3703  auto StartPoint = BB->rbegin();
3704  ++StartPoint;
3705  // Note that we explicitly recalculate BB->rend() on each iteration,
3706  // as it may change when we remove the first instruction.
3707  for (BasicBlock::reverse_iterator I(StartPoint); I != BB->rend();) {
3708  Instruction &Inst = *I++;
3709  if (!Inst.use_empty())
3711  if (isa<LandingPadInst>(Inst))
3712  continue;
3713 
3714  Inst.eraseFromParent();
3715  ++NumGVNInstrDeleted;
3716  }
3717  // Now insert something that simplifycfg will turn into an unreachable.
3718  Type *Int8Ty = Type::getInt8Ty(BB->getContext());
3719  new StoreInst(UndefValue::get(Int8Ty),
3721  BB->getTerminator());
3722 }
3723 
3724 void NewGVN::markInstructionForDeletion(Instruction *I) {
3725  DEBUG(dbgs() << "Marking " << *I << " for deletion\n");
3726  InstructionsToErase.insert(I);
3727 }
3728 
3729 void NewGVN::replaceInstruction(Instruction *I, Value *V) {
3730  DEBUG(dbgs() << "Replacing " << *I << " with " << *V << "\n");
3732  // We save the actual erasing to avoid invalidating memory
3733  // dependencies until we are done with everything.
3734  markInstructionForDeletion(I);
3735 }
3736 
3737 namespace {
3738 
3739 // This is a stack that contains both the value and dfs info of where
3740 // that value is valid.
3741 class ValueDFSStack {
3742 public:
3743  Value *back() const { return ValueStack.back(); }
3744  std::pair<int, int> dfs_back() const { return DFSStack.back(); }
3745 
3746  void push_back(Value *V, int DFSIn, int DFSOut) {
3747  ValueStack.emplace_back(V);
3748  DFSStack.emplace_back(DFSIn, DFSOut);
3749  }
3750 
3751  bool empty() const { return DFSStack.empty(); }
3752 
3753  bool isInScope(int DFSIn, int DFSOut) const {
3754  if (empty())
3755  return false;
3756  return DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second;
3757  }
3758 
3759  void popUntilDFSScope(int DFSIn, int DFSOut) {
3760 
3761  // These two should always be in sync at this point.
3762  assert(ValueStack.size() == DFSStack.size() &&
3763  "Mismatch between ValueStack and DFSStack");
3764  while (
3765  !DFSStack.empty() &&
3766  !(DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second)) {
3767  DFSStack.pop_back();
3768  ValueStack.pop_back();
3769  }
3770  }
3771 
3772 private:
3773  SmallVector<Value *, 8> ValueStack;
3774  SmallVector<std::pair<int, int>, 8> DFSStack;
3775 };
3776 
3777 } // end anonymous namespace
3778 
3779 // Given an expression, get the congruence class for it.
3780 CongruenceClass *NewGVN::getClassForExpression(const Expression *E) const {
3781  if (auto *VE = dyn_cast<VariableExpression>(E))
3782  return ValueToClass.lookup(VE->getVariableValue());
3783  else if (isa<DeadExpression>(E))
3784  return TOPClass;
3785  return ExpressionToClass.lookup(E);
3786 }
3787 
3788 // Given a value and a basic block we are trying to see if it is available in,
3789 // see if the value has a leader available in that block.
3790 Value *NewGVN::findPHIOfOpsLeader(const Expression *E,
3791  const Instruction *OrigInst,
3792  const BasicBlock *BB) const {
3793  // It would already be constant if we could make it constant
3794  if (auto *CE = dyn_cast<ConstantExpression>(E))
3795  return CE->getConstantValue();
3796  if (auto *VE = dyn_cast<VariableExpression>(E)) {
3797  auto *V = VE->getVariableValue();
3798  if (alwaysAvailable(V) || DT->dominates(getBlockForValue(V), BB))
3799  return VE->getVariableValue();
3800  }
3801 
3802  auto *CC = getClassForExpression(E);
3803  if (!CC)
3804  return nullptr;
3805  if (alwaysAvailable(CC->getLeader()))
3806  return CC->getLeader();
3807 
3808  for (auto Member : *CC) {
3809  auto *MemberInst = dyn_cast<Instruction>(Member);
3810  if (MemberInst == OrigInst)
3811  continue;
3812  // Anything that isn't an instruction is always available.
3813  if (!MemberInst)
3814  return Member;
3815  if (DT->dominates(getBlockForValue(MemberInst), BB))
3816  return Member;
3817  }
3818  return nullptr;
3819 }
3820 
3821 bool NewGVN::eliminateInstructions(Function &F) {
3822  // This is a non-standard eliminator. The normal way to eliminate is
3823  // to walk the dominator tree in order, keeping track of available
3824  // values, and eliminating them. However, this is mildly
3825  // pointless. It requires doing lookups on every instruction,
3826  // regardless of whether we will ever eliminate it. For
3827  // instructions part of most singleton congruence classes, we know we
3828  // will never eliminate them.
3829 
3830  // Instead, this eliminator looks at the congruence classes directly, sorts
3831  // them into a DFS ordering of the dominator tree, and then we just
3832  // perform elimination straight on the sets by walking the congruence
3833  // class member uses in order, and eliminate the ones dominated by the
3834  // last member. This is worst case O(E log E) where E = number of
3835  // instructions in a single congruence class. In theory, this is all
3836  // instructions. In practice, it is much faster, as most instructions are
3837  // either in singleton congruence classes or can't possibly be eliminated
3838  // anyway (if there are no overlapping DFS ranges in class).
3839  // When we find something not dominated, it becomes the new leader
3840  // for elimination purposes.
3841  // TODO: If we wanted to be faster, We could remove any members with no
3842  // overlapping ranges while sorting, as we will never eliminate anything
3843  // with those members, as they don't dominate anything else in our set.
3844 
3845  bool AnythingReplaced = false;
3846 
3847  // Since we are going to walk the domtree anyway, and we can't guarantee the
3848  // DFS numbers are updated, we compute some ourselves.
3849  DT->updateDFSNumbers();
3850 
3851  // Go through all of our phi nodes, and kill the arguments associated with
3852  // unreachable edges.
3853  auto ReplaceUnreachablePHIArgs = [&](PHINode *PHI, BasicBlock *BB) {
3854  for (auto &Operand : PHI->incoming_values())
3855  if (!ReachableEdges.count({PHI->getIncomingBlock(Operand), BB})) {
3856  DEBUG(dbgs() << "Replacing incoming value of " << PHI << " for block "
3857  << getBlockName(PHI->getIncomingBlock(Operand))
3858  << " with undef due to it being unreachable\n");
3859  Operand.set(UndefValue::get(PHI->getType()));
3860  }
3861  };
3862  // Replace unreachable phi arguments.
3863  // At this point, RevisitOnReachabilityChange only contains:
3864  //
3865  // 1. PHIs
3866  // 2. Temporaries that will convert to PHIs
3867  // 3. Operations that are affected by an unreachable edge but do not fit into
3868  // 1 or 2 (rare).
3869  // So it is a slight overshoot of what we want. We could make it exact by
3870  // using two SparseBitVectors per block.
3871  DenseMap<const BasicBlock *, unsigned> ReachablePredCount;
3872  for (auto &KV : ReachableEdges)
3873  ReachablePredCount[KV.getEnd()]++;
3874  for (auto &BBPair : RevisitOnReachabilityChange) {
3875  for (auto InstNum : BBPair.second) {
3876  auto *Inst = InstrFromDFSNum(InstNum);
3877  auto *PHI = dyn_cast<PHINode>(Inst);
3878  PHI = PHI ? PHI : dyn_cast_or_null<PHINode>(RealToTemp.lookup(Inst));
3879  if (!PHI)
3880  continue;
3881  auto *BB = BBPair.first;
3882  if (ReachablePredCount.lookup(BB) != PHI->getNumIncomingValues())
3883  ReplaceUnreachablePHIArgs(PHI, BB);
3884  }
3885  }
3886 
3887  // Map to store the use counts
3889  for (auto *CC : reverse(CongruenceClasses)) {
3890  DEBUG(dbgs() << "Eliminating in congruence class " << CC->getID() << "\n");
3891  // Track the equivalent store info so we can decide whether to try
3892  // dead store elimination.
3893  SmallVector<ValueDFS, 8> PossibleDeadStores;
3894  SmallPtrSet<Instruction *, 8> ProbablyDead;
3895  if (CC->isDead() || CC->empty())
3896  continue;
3897  // Everything still in the TOP class is unreachable or dead.
3898  if (CC == TOPClass) {
3899  for (auto M : *CC) {
3900  auto *VTE = ValueToExpression.lookup(M);
3901  if (VTE && isa<DeadExpression>(VTE))
3902  markInstructionForDeletion(cast<Instruction>(M));
3903  assert((!ReachableBlocks.count(cast<Instruction>(M)->getParent()) ||
3904  InstructionsToErase.count(cast<Instruction>(M))) &&
3905  "Everything in TOP should be unreachable or dead at this "
3906  "point");
3907  }
3908  continue;
3909  }
3910 
3911  assert(CC->getLeader() && "We should have had a leader");
3912  // If this is a leader that is always available, and it's a
3913  // constant or has no equivalences, just replace everything with
3914  // it. We then update the congruence class with whatever members
3915  // are left.
3916  Value *Leader =
3917  CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
3918  if (alwaysAvailable(Leader)) {
3919  CongruenceClass::MemberSet MembersLeft;
3920  for (auto M : *CC) {
3921  Value *Member = M;
3922  // Void things have no uses we can replace.
3923  if (Member == Leader || !isa<Instruction>(Member) ||
3924  Member->getType()->isVoidTy()) {
3925  MembersLeft.insert(Member);
3926  continue;
3927  }
3928  DEBUG(dbgs() << "Found replacement " << *(Leader) << " for " << *Member
3929  << "\n");
3930  auto *I = cast<Instruction>(Member);
3931  assert(Leader != I && "About to accidentally remove our leader");
3932  replaceInstruction(I, Leader);
3933  AnythingReplaced = true;
3934  }
3935  CC->swap(MembersLeft);
3936  } else {
3937  // If this is a singleton, we can skip it.
3938  if (CC->size() != 1 || RealToTemp.count(Leader)) {
3939  // This is a stack because equality replacement/etc may place
3940  // constants in the middle of the member list, and we want to use
3941  // those constant values in preference to the current leader, over
3942  // the scope of those constants.
3943  ValueDFSStack EliminationStack;
3944 
3945  // Convert the members to DFS ordered sets and then merge them.
3946  SmallVector<ValueDFS, 8> DFSOrderedSet;
3947  convertClassToDFSOrdered(*CC, DFSOrderedSet, UseCounts, ProbablyDead);
3948 
3949  // Sort the whole thing.
3950  std::sort(DFSOrderedSet.begin(), DFSOrderedSet.end());
3951  for (auto &VD : DFSOrderedSet) {
3952  int MemberDFSIn = VD.DFSIn;
3953  int MemberDFSOut = VD.DFSOut;
3954  Value *Def = VD.Def.getPointer();
3955  bool FromStore = VD.Def.getInt();
3956  Use *U = VD.U;
3957  // We ignore void things because we can't get a value from them.
3958  if (Def && Def->getType()->isVoidTy())
3959  continue;
3960  auto *DefInst = dyn_cast_or_null<Instruction>(Def);
3961  if (DefInst && AllTempInstructions.count(DefInst)) {
3962  auto *PN = cast<PHINode>(DefInst);
3963 
3964  // If this is a value phi and that's the expression we used, insert
3965  // it into the program
3966  // remove from temp instruction list.
3967  AllTempInstructions.erase(PN);
3968  auto *DefBlock = getBlockForValue(Def);
3969  DEBUG(dbgs() << "Inserting fully real phi of ops" << *Def
3970  << " into block "
3971  << getBlockName(getBlockForValue(Def)) << "\n");
3972  PN->insertBefore(&DefBlock->front());
3973  Def = PN;
3974  NumGVNPHIOfOpsEliminations++;
3975  }
3976 
3977  if (EliminationStack.empty()) {
3978  DEBUG(dbgs() << "Elimination Stack is empty\n");
3979  } else {
3980  DEBUG(dbgs() << "Elimination Stack Top DFS numbers are ("
3981  << EliminationStack.dfs_back().first << ","
3982  << EliminationStack.dfs_back().second << ")\n");
3983  }
3984 
3985  DEBUG(dbgs() << "Current DFS numbers are (" << MemberDFSIn << ","
3986  << MemberDFSOut << ")\n");
3987  // First, we see if we are out of scope or empty. If so,
3988  // and there equivalences, we try to replace the top of
3989  // stack with equivalences (if it's on the stack, it must
3990  // not have been eliminated yet).
3991  // Then we synchronize to our current scope, by
3992  // popping until we are back within a DFS scope that
3993  // dominates the current member.
3994  // Then, what happens depends on a few factors
3995  // If the stack is now empty, we need to push
3996  // If we have a constant or a local equivalence we want to
3997  // start using, we also push.
3998  // Otherwise, we walk along, processing members who are
3999  // dominated by this scope, and eliminate them.
4000  bool ShouldPush = Def && EliminationStack.empty();
4001  bool OutOfScope =
4002  !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut);
4003 
4004  if (OutOfScope || ShouldPush) {
4005  // Sync to our current scope.
4006  EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4007  bool ShouldPush = Def && EliminationStack.empty();
4008  if (ShouldPush) {
4009  EliminationStack.push_back(Def, MemberDFSIn, MemberDFSOut);
4010  }
4011  }
4012 
4013  // Skip the Def's, we only want to eliminate on their uses. But mark
4014  // dominated defs as dead.
4015  if (Def) {
4016  // For anything in this case, what and how we value number
4017  // guarantees that any side-effets that would have occurred (ie
4018  // throwing, etc) can be proven to either still occur (because it's
4019  // dominated by something that has the same side-effects), or never
4020  // occur. Otherwise, we would not have been able to prove it value
4021  // equivalent to something else. For these things, we can just mark
4022  // it all dead. Note that this is different from the "ProbablyDead"
4023  // set, which may not be dominated by anything, and thus, are only
4024  // easy to prove dead if they are also side-effect free. Note that
4025  // because stores are put in terms of the stored value, we skip
4026  // stored values here. If the stored value is really dead, it will
4027  // still be marked for deletion when we process it in its own class.
4028  if (!EliminationStack.empty() && Def != EliminationStack.back() &&
4029  isa<Instruction>(Def) && !FromStore)
4030  markInstructionForDeletion(cast<Instruction>(Def));
4031  continue;
4032  }
4033  // At this point, we know it is a Use we are trying to possibly
4034  // replace.
4035 
4036  assert(isa<Instruction>(U->get()) &&
4037  "Current def should have been an instruction");
4038  assert(isa<Instruction>(U->getUser()) &&
4039  "Current user should have been an instruction");
4040 
4041  // If the thing we are replacing into is already marked to be dead,
4042  // this use is dead. Note that this is true regardless of whether
4043  // we have anything dominating the use or not. We do this here
4044  // because we are already walking all the uses anyway.
4045  Instruction *InstUse = cast<Instruction>(U->getUser());
4046  if (InstructionsToErase.count(InstUse)) {
4047  auto &UseCount = UseCounts[U->get()];
4048  if (--UseCount == 0) {
4049  ProbablyDead.insert(cast<Instruction>(U->get()));
4050  }
4051  }
4052 
4053  // If we get to this point, and the stack is empty we must have a use
4054  // with nothing we can use to eliminate this use, so just skip it.
4055  if (EliminationStack.empty())
4056  continue;
4057 
4058  Value *DominatingLeader = EliminationStack.back();
4059 
4060  auto *II = dyn_cast<IntrinsicInst>(DominatingLeader);
4061  if (II && II->getIntrinsicID() == Intrinsic::ssa_copy)
4062  DominatingLeader = II->getOperand(0);
4063 
4064  // Don't replace our existing users with ourselves.
4065  if (U->get() == DominatingLeader)
4066  continue;
4067  DEBUG(dbgs() << "Found replacement " << *DominatingLeader << " for "
4068  << *U->get() << " in " << *(U->getUser()) << "\n");
4069 
4070  // If we replaced something in an instruction, handle the patching of
4071  // metadata. Skip this if we are replacing predicateinfo with its
4072  // original operand, as we already know we can just drop it.
4073  auto *ReplacedInst = cast<Instruction>(U->get());
4074  auto *PI = PredInfo->getPredicateInfoFor(ReplacedInst);
4075  if (!PI || DominatingLeader != PI->OriginalOp)
4076  patchReplacementInstruction(ReplacedInst, DominatingLeader);
4077  U->set(DominatingLeader);
4078  // This is now a use of the dominating leader, which means if the
4079  // dominating leader was dead, it's now live!
4080  auto &LeaderUseCount = UseCounts[DominatingLeader];
4081  // It's about to be alive again.
4082  if (LeaderUseCount == 0 && isa<Instruction>(DominatingLeader))
4083  ProbablyDead.erase(cast<Instruction>(DominatingLeader));
4084  if (LeaderUseCount == 0 && II)
4085  ProbablyDead.insert(II);
4086  ++LeaderUseCount;
4087  AnythingReplaced = true;
4088  }
4089  }
4090  }
4091 
4092  // At this point, anything still in the ProbablyDead set is actually dead if
4093  // would be trivially dead.
4094  for (auto *I : ProbablyDead)
4096  markInstructionForDeletion(I);
4097 
4098  // Cleanup the congruence class.
4099  CongruenceClass::MemberSet MembersLeft;
4100  for (auto *Member : *CC)
4101  if (!isa<Instruction>(Member) ||
4102  !InstructionsToErase.count(cast<Instruction>(Member)))
4103  MembersLeft.insert(Member);
4104  CC->swap(MembersLeft);
4105 
4106  // If we have possible dead stores to look at, try to eliminate them.
4107  if (CC->getStoreCount() > 0) {
4108  convertClassToLoadsAndStores(*CC, PossibleDeadStores);
4109  std::sort(PossibleDeadStores.begin(), PossibleDeadStores.end());
4110  ValueDFSStack EliminationStack;
4111  for (auto &VD : PossibleDeadStores) {
4112  int MemberDFSIn = VD.DFSIn;
4113  int MemberDFSOut = VD.DFSOut;
4114  Instruction *Member = cast<Instruction>(VD.Def.getPointer());
4115  if (EliminationStack.empty() ||
4116  !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut)) {
4117  // Sync to our current scope.
4118  EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4119  if (EliminationStack.empty()) {
4120  EliminationStack.push_back(Member, MemberDFSIn, MemberDFSOut);
4121  continue;
4122  }
4123  }
4124  // We already did load elimination, so nothing to do here.
4125  if (isa<LoadInst>(Member))
4126  continue;
4127  assert(!EliminationStack.empty());
4128  Instruction *Leader = cast<Instruction>(EliminationStack.back());
4129  (void)Leader;
4130  assert(DT->dominates(Leader->getParent(), Member->getParent()));
4131  // Member is dominater by Leader, and thus dead
4132  DEBUG(dbgs() << "Marking dead store " << *Member
4133  << " that is dominated by " << *Leader << "\n");
4134  markInstructionForDeletion(Member);
4135  CC->erase(Member);
4136  ++NumGVNDeadStores;
4137  }
4138  }
4139  }
4140  return AnythingReplaced;
4141 }
4142 
4143 // This function provides global ranking of operations so that we can place them
4144 // in a canonical order. Note that rank alone is not necessarily enough for a
4145 // complete ordering, as constants all have the same rank. However, generally,
4146 // we will simplify an operation with all constants so that it doesn't matter
4147 // what order they appear in.
4148 unsigned int NewGVN::getRank(const Value *V) const {
4149  // Prefer constants to undef to anything else
4150  // Undef is a constant, have to check it first.
4151  // Prefer smaller constants to constantexprs
4152  if (isa<ConstantExpr>(V))
4153  return 2;
4154  if (isa<UndefValue>(V))
4155  return 1;
4156  if (isa<Constant>(V))
4157  return 0;
4158  else if (auto *A = dyn_cast<Argument>(V))
4159  return 3 + A->getArgNo();
4160 
4161  // Need to shift the instruction DFS by number of arguments + 3 to account for
4162  // the constant and argument ranking above.
4163  unsigned Result = InstrToDFSNum(V);
4164  if (Result > 0)
4165  return 4 + NumFuncArgs + Result;
4166  // Unreachable or something else, just return a really large number.
4167  return ~0;
4168 }
4169 
4170 // This is a function that says whether two commutative operations should
4171 // have their order swapped when canonicalizing.
4172 bool NewGVN::shouldSwapOperands(const Value *A, const Value *B) const {
4173  // Because we only care about a total ordering, and don't rewrite expressions
4174  // in this order, we order by rank, which will give a strict weak ordering to
4175  // everything but constants, and then we order by pointer address.
4176  return std::make_pair(getRank(A), A) > std::make_pair(getRank(B), B);
4177 }
4178 
4179 namespace {
4180 
4181 class NewGVNLegacyPass : public FunctionPass {
4182 public:
4183  // Pass identification, replacement for typeid.
4184  static char ID;
4185 
4186  NewGVNLegacyPass() : FunctionPass(ID) {
4188  }
4189 
4190  bool runOnFunction(Function &F) override;
4191 
4192 private:
4193  void getAnalysisUsage(AnalysisUsage &AU) const override {
4201  }
4202 };
4203 
4204 } // end anonymous namespace
4205 
4207  if (skipFunction(F))
4208  return false;
4209  return NewGVN(F, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4210  &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
4211  &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
4212  &getAnalysis<AAResultsWrapperPass>().getAAResults(),
4213  &getAnalysis<MemorySSAWrapperPass>().getMSSA(),
4214  F.getParent()->getDataLayout())
4215  .runGVN();
4216 }
4217 
4218 char NewGVNLegacyPass::ID = 0;
4219 
4220 INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering",
4221  false, false)
4228 INITIALIZE_PASS_END(NewGVNLegacyPass, "newgvn", "Global Value Numbering", false,
4229  false)
4230 
4231 // createGVNPass - The public interface to this file.
4232 FunctionPass *llvm::createNewGVNPass() { return new NewGVNLegacyPass(); }
4233 
4235  // Apparently the order in which we get these results matter for
4236  // the old GVN (see Chandler's comment in GVN.cpp). I'll keep
4237  // the same order here, just in case.
4238  auto &AC = AM.getResult<AssumptionAnalysis>(F);
4239  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
4240  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
4241  auto &AA = AM.getResult<AAManager>(F);
4242  auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
4243  bool Changed =
4244  NewGVN(F, &DT, &AC, &TLI, &AA, &MSSA, F.getParent()->getDataLayout())
4245  .runGVN();
4246  if (!Changed)
4247  return PreservedAnalyses::all();
4248  PreservedAnalyses PA;
4250  PA.preserve<GlobalsAA>();
4251  return PA;
4252 }
Legacy wrapper pass to provide the GlobalsAAResult object.
Global Value Numbering
Definition: NewGVN.cpp:4228
void initializeNewGVNLegacyPassPass(PassRegistry &)
uint64_t CallInst * C
Value * getValueOperand()
Definition: Instructions.h:395
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:69
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:109
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:245
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:523
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:843
void setInt(IntType IntVal)
bool isSimple() const
Definition: Instructions.h:262
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
iterator_range< use_iterator > uses()
Definition: Value.h:356
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
int analyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, LoadInst *DepLI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the load at De...
Definition: VNCoercion.cpp:219
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
unsigned getValueID() const
Return an ID for the concrete type of this object.
Definition: Value.h:465
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:236
void dropAllReferences()
Drop all references to operands.
Definition: User.h:279
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS)
Definition: NewGVN.cpp:869
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
bool isAtomic() const
Return true if this instruction has an AtomicOrdering of unordered or higher.
static bool isEqual(const ExactEqualsExpression &LHS, const Expression *RHS)
Definition: NewGVN.cpp:462
This is the interface for a simple mod/ref and alias analysis over globals.
BasicBlock * getSuccessor(unsigned idx) const
Return the specified successor.
iterator begin() const
Definition: ArrayRef.h:137
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
static bool okayForPHIOfOps(const Instruction *I)
Definition: NewGVN.cpp:2588
Implements a dense probed hash-table based set.
Definition: DenseSet.h:221
This provides a very simple, boring adaptor for a begin and end iterator into a range type...
This class represents a function call, abstracting a target machine&#39;s calling convention.
An immutable pass that tracks lazily created AssumptionCache objects.
A cache of .assume calls within a function.
Represents a read-write access to memory, whether it is a must-alias, or a may-alias.
Definition: MemorySSA.h:350
bool isTerminator() const
Definition: Instruction.h:128
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:869
Recycle small arrays allocated from a BumpPtrAllocator.
Definition: ArrayRecycler.h:29
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:95
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:813
BasicBlock * getSuccessor(unsigned i) const
STATISTIC(NumFunctions, "Total number of functions")
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:238
F(f)
reverse_iterator rend()
Definition: BasicBlock.h:259
An instruction for reading from memory.
Definition: Instructions.h:164
reverse_iterator rbegin()
Definition: BasicBlock.h:257
Hexagon Common GEP
Value * getCondition() const
This defines the Use class.
Constant * getConstantMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, Type *LoadTy, const DataLayout &DL)
Definition: VNCoercion.cpp:483
const Value * DoPHITranslation(const BasicBlock *CurBB, const BasicBlock *PredBB) const
Translate PHI node to its predecessor from the given basic block.
Definition: Value.cpp:720
ExactEqualsExpression(const Expression &E)
Definition: NewGVN.cpp:432
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:33
op_iterator op_begin()
Definition: User.h:214
This file defines the MallocAllocator and BumpPtrAllocator interfaces.
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:207
static bool isImpliedFalseByMatchingCmp(Predicate Pred1, Predicate Pred2)
Determine if Pred1 implies Pred2 is false when two compares have matching operands.
hash_code getComputedHash() const
Definition: GVNExpression.h:92
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
Legacy analysis pass which computes MemorySSA.
Definition: MemorySSA.h:850
void setPointer(PointerTy PtrVal)
static cl::opt< bool > EnablePhiOfOps("enable-phi-of-ops", cl::init(true), cl::Hidden)
Currently, the generation "phi of ops" can result in correctness issues.
Predicate getInversePredicate() const
For example, EQ -> NE, UGT -> ULE, SLT -> SGE, OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc.
Definition: InstrTypes.h:951
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:361
Value * SimplifyGEPInst(Type *SrcTy, ArrayRef< Value *> Ops, const SimplifyQuery &Q)
Given operands for a GetElementPtrInst, fold the result or return null.
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:639
hash_code getComputedHash() const
Definition: NewGVN.cpp:434
int analyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, MemIntrinsic *DepMI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the memory int...
Definition: VNCoercion.cpp:251
Constant * getConstantStoreValueForLoad(Constant *SrcVal, unsigned Offset, Type *LoadTy, const DataLayout &DL)
Definition: VNCoercion.cpp:351
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
Encapsulates MemorySSA, including all data associated with memory accesses.
Definition: MemorySSA.h:612
static bool isImpliedTrueByMatchingCmp(Predicate Pred1, Predicate Pred2)
Determine if Pred1 implies Pred2 is true when two compares have matching operands.
PointerIntPair< Value *, 1, bool > Def
Definition: NewGVN.cpp:3493
friend const_iterator begin(StringRef path, Style style)
Get begin iterator over path.
Definition: Path.cpp:236
bool equals(const Expression &Other) const override
Definition: NewGVN.cpp:875
This file provides an implementation of debug counters.
static unsigned getHashValue(const ExactEqualsExpression &E)
Definition: NewGVN.cpp:458
auto reverse(ContainerTy &&C, typename std::enable_if< has_rbegin< ContainerTy >::value >::type *=nullptr) -> decltype(make_range(C.rbegin(), C.rend()))
Definition: STLExtras.h:232
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
Definition: Constants.h:201
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
Key
PAL metadata keys.
Value * SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, const SimplifyQuery &Q)
Given operands for a CmpInst, fold the result or return null.
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
static int getID(struct InternalInstruction *insn, const void *miiArg)
void andIRFlags(const Value *V)
Logical &#39;and&#39; of any supported wrapping, exact, and fast-math flags of V and this instruction...
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:33
static bool isCounterSet(unsigned ID)
Definition: DebugCounter.h:106
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
An instruction for storing to memory.
Definition: Instructions.h:306
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:430
This is the generic walker interface for walkers of MemorySSA.
Definition: MemorySSA.h:881
const Expression & E
Definition: NewGVN.cpp:430
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:140
static const Expression * getEmptyKey()
Definition: NewGVN.cpp:442
Value * getOperand(unsigned i) const
Definition: User.h:154
int analyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, StoreInst *DepSI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the store at D...
Definition: VNCoercion.cpp:202
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:141
const BasicBlock & getEntryBlock() const
Definition: Function.h:572
unsigned getDFSNumIn() const
getDFSNumIn/getDFSNumOut - These return the DFS visitation order for nodes in the dominator tree...
static bool runOnFunction(Function &F, bool PostInlining)
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
friend const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:245
Value * SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, const SimplifyQuery &Q)
Given operands for a SelectInst, fold the result or return null.
Control flow instructions. These all have token chains.
Definition: ISDOpcodes.h:597
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
Subclasses of this class are all able to terminate a basic block.
Definition: InstrTypes.h:54
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
MutableArrayRef - Represent a mutable reference to an array (0 or more elements consecutively in memo...
Definition: ArrayRef.h:291
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
Allocate memory in an ever growing pool, as if by bump-pointer.
Definition: Allocator.h:138
Conditional or Unconditional Branch instruction.
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:149
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:116
Value * getIncomingValueForBlock(const BasicBlock *BB) const
Constant * getConstantLoadValueForLoad(Constant *SrcVal, unsigned Offset, Type *LoadTy, const DataLayout &DL)
Definition: VNCoercion.cpp:409
This file contains the declarations for the subclasses of Constant, which represent the different fla...
mapped_iterator< ItTy, FuncTy > map_iterator(ItTy I, FuncTy F)
Definition: STLExtras.h:205
A manager for alias analyses.
static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl)
Definition: NewGVN.cpp:3692
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:371
static void setCounterValue(unsigned ID, const std::pair< int, int > &Val)
Definition: DebugCounter.h:119
Value * SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty, const SimplifyQuery &Q)
Given operands for a CastInst, fold the result or return null.
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:216
iterator_range< def_chain_iterator< T, true > > optimized_def_chain(T MA)
Definition: MemorySSA.h:1195
bool any_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:820
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:853
bool isBinaryOp() const
Definition: Instruction.h:129
static const unsigned End
This file provides the interface for LLVM&#39;s Global Value Numbering pass.
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
size_t arg_size() const
Definition: Function.h:630
op_range operands()
Definition: User.h:222
Value * getPointerOperand()
Definition: Instructions.h:270
unsigned getDFSNumOut() const
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:382
static bool shouldExecute(unsigned CounterName)
Definition: DebugCounter.h:72
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1320
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
static const Expression * getTombstoneKey()
Definition: NewGVN.cpp:448
This file implements the PredicateInfo analysis, which creates an Extended SSA form for operations us...
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
static bool isCopyOfAPHI(const Value *V)
Definition: NewGVN.cpp:952
static std::pair< int, int > getCounterValue(unsigned ID)
Definition: DebugCounter.h:111
DOTGraphTraits - Template class that can be specialized to customize how graphs are converted to &#39;dot...
A function analysis which provides an AssumptionCache.
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:176
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
void setOpcode(unsigned opcode)
bool erase(PtrType Ptr)
erase - If the set contains the specified pointer, remove it and return true, otherwise return false...
Definition: SmallPtrSet.h:378
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:864
Provides information about what library functions are available for the current target.
An analysis that produces MemorySSA for a function.
Definition: MemorySSA.h:814
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:385
PreservedAnalyses run(Function &F, AnalysisManager< Function > &AM)
Run the pass over the function.
Definition: NewGVN.cpp:4234
BasicBlock * getBlock() const
Definition: MemorySSA.h:156
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
bool isConditional() const
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:516
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:451
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:923
bool isTrueWhenEqual() const
This is just a convenience.
Definition: InstrTypes.h:1019
A range adaptor for a pair of iterators.
static bool alwaysAvailable(Value *V)
Definition: NewGVN.cpp:970
Class that has the common methods + fields of memory uses/defs.
Definition: MemorySSA.h:236
iterator_range< user_iterator > users()
Definition: Value.h:401
BasicBlock * getIncomingBlock(unsigned I) const
Return incoming basic block number i.
Definition: MemorySSA.h:493
iterator begin() const
Definition: ArrayRef.h:331
An opaque object representing a hash code.
Definition: Hashing.h:72
bool isMallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates uninitialized memory (such ...
Instruction * getMemoryInst() const
Get the instruction that this MemoryUse represents.
Definition: MemorySSA.h:243
bool operator<(const ValueDFS &Other) const
Definition: NewGVN.cpp:3496
amdgpu Simplify well known AMD library false Value Value * Arg
Predicate getPredicate() const
Return the predicate for this instruction.
Definition: InstrTypes.h:927
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:120
iterator_range< typename GraphTraits< GraphType >::nodes_iterator > nodes(const GraphType &G)
Definition: GraphTraits.h:89
static Value * getCopyOf(const Value *V)
Definition: NewGVN.cpp:940
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:226
void emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:656
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
bool isCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates zero-filled memory (such as...
#define I(x, y, z)
Definition: MD5.cpp:58
static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")
iterator find(const_arg_type_t< ValueT > V)
Definition: DenseSet.h:165
bool isZero() const
This is just a convenience method to make client code smaller for a common code.
Definition: Constants.h:193
iterator_range< value_op_iterator > operand_values()
Definition: User.h:246
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
Constant * ConstantFoldInstOperands(Instruction *I, ArrayRef< Constant *> Ops, const DataLayout &DL, const TargetLibraryInfo *TLI=nullptr)
ConstantFoldInstOperands - Attempt to constant fold an instruction with the specified operands...
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
size_type count(const_arg_type_t< ValueT > V) const
Return 1 if the specified key is in the set, 0 otherwise.
Definition: DenseSet.h:91
iterator_range< filter_iterator< detail::IterOfRange< RangeT >, PredicateT > > make_filter_range(RangeT &&Range, PredicateT Pred)
Convenience function that takes a range of elements and a predicate, and return a new filter_iterator...
Definition: STLExtras.h:332
static std::string getBlockName(const BasicBlock *B)
Definition: NewGVN.cpp:897
ValueT lookup(const_arg_type_t< KeyT > Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: DenseMap.h:181
OutputIt transform(R &&Range, OutputIt d_first, UnaryPredicate P)
Wrapper function around std::transform to apply a function to a range and store the result elsewhere...
Definition: STLExtras.h:890
Analysis pass providing the TargetLibraryInfo.
iterator_range< df_iterator< T > > depth_first(const T &G)
Value * SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, const SimplifyQuery &Q)
Given operands for a BinaryOperator, fold the result or return null.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
unsigned getNumSuccessors() const
Return the number of successors that this terminator has.
bool equals(const Expression &Other) const override
Definition: NewGVN.cpp:879
static bool isEqual(const Expression *LHS, const Expression *RHS)
Definition: NewGVN.cpp:468
static bool isCopyOfPHI(const Value *V, const PHINode *PN)
Definition: NewGVN.cpp:948
aarch64 promote const
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:856
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:556
bool isInstructionTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction is not used, and the instruction has no side ef...
Definition: Local.cpp:324
LLVM Value Representation.
Definition: Value.h:73
Value * getOperand(unsigned N) const
The header file for the GVN pass that contains expression handling classes.
iterator end() const
Definition: ArrayRef.h:332
bool wouldInstructionBeTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction would have no side effects if it was not used...
Definition: Local.cpp:331
#define DEBUG(X)
Definition: Debug.h:118
static cl::opt< bool > EnableStoreRefinement("enable-store-refinement", cl::init(false), cl::Hidden)
This file exposes an interface to building/using memory SSA to walk memory instructions using a use/d...
FunctionPass * createNewGVNPass()
Definition: NewGVN.cpp:4232
Predicate getSwappedPredicate() const
For example, EQ->EQ, SLE->SGE, ULT->UGT, OEQ->OEQ, ULE->UGE, OLT->OGT, etc.
Definition: InstrTypes.h:967
A container for analyses that lazily runs them and caches their results.
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:267
void sort(Policy policy, RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp=Comparator())
Definition: Parallel.h:199
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
bool isSimple() const
Definition: Instructions.h:387
Represents phi nodes for memory accesses.
Definition: MemorySSA.h:431
const TerminatorInst * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:120
static unsigned getHashValue(const Expression *E)
Definition: NewGVN.cpp:454
virtual bool exactlyEquals(const Expression &Other) const
op_range incoming_values()
Value * getPointerOperand()
Definition: Instructions.h:398
DEBUG_COUNTER(VNCounter, "newgvn-vn", "Controls which instructions are value numbered")
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:174
void combineMetadata(Instruction *K, const Instruction *J, ArrayRef< unsigned > KnownIDs)
Combine the metadata of two instructions so that K can replace J.
Definition: Local.cpp:1766
INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering", false, false) INITIALIZE_PASS_END(NewGVNLegacyPass
static void patchReplacementInstruction(Instruction *I, Value *Repl)
Definition: NewGVN.cpp:3661
bool use_empty() const
Definition: Value.h:328
void allocateOperands(RecyclerType &Recycler, BumpPtrAllocator &Allocator)
iterator_range< arg_iterator > args()
Definition: Function.h:621
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:66
newgvn
Definition: NewGVN.cpp:4228
bool operator==(const Expression &Other) const
Definition: NewGVN.cpp:436