LLVM  9.0.0svn
NewGVN.cpp
Go to the documentation of this file.
1 //===- NewGVN.cpp - Global Value Numbering Pass ---------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// This file implements the new LLVM's Global Value Numbering pass.
11 /// GVN partitions values computed by a function into congruence classes.
12 /// Values ending up in the same congruence class are guaranteed to be the same
13 /// for every execution of the program. In that respect, congruency is a
14 /// compile-time approximation of equivalence of values at runtime.
15 /// The algorithm implemented here uses a sparse formulation and it's based
16 /// on the ideas described in the paper:
17 /// "A Sparse Algorithm for Predicated Global Value Numbering" from
18 /// Karthik Gargi.
19 ///
20 /// A brief overview of the algorithm: The algorithm is essentially the same as
21 /// the standard RPO value numbering algorithm (a good reference is the paper
22 /// "SCC based value numbering" by L. Taylor Simpson) with one major difference:
23 /// The RPO algorithm proceeds, on every iteration, to process every reachable
24 /// block and every instruction in that block. This is because the standard RPO
25 /// algorithm does not track what things have the same value number, it only
26 /// tracks what the value number of a given operation is (the mapping is
27 /// operation -> value number). Thus, when a value number of an operation
28 /// changes, it must reprocess everything to ensure all uses of a value number
29 /// get updated properly. In constrast, the sparse algorithm we use *also*
30 /// tracks what operations have a given value number (IE it also tracks the
31 /// reverse mapping from value number -> operations with that value number), so
32 /// that it only needs to reprocess the instructions that are affected when
33 /// something's value number changes. The vast majority of complexity and code
34 /// in this file is devoted to tracking what value numbers could change for what
35 /// instructions when various things happen. The rest of the algorithm is
36 /// devoted to performing symbolic evaluation, forward propagation, and
37 /// simplification of operations based on the value numbers deduced so far
38 ///
39 /// In order to make the GVN mostly-complete, we use a technique derived from
40 /// "Detection of Redundant Expressions: A Complete and Polynomial-time
41 /// Algorithm in SSA" by R.R. Pai. The source of incompleteness in most SSA
42 /// based GVN algorithms is related to their inability to detect equivalence
43 /// between phi of ops (IE phi(a+b, c+d)) and op of phis (phi(a,c) + phi(b, d)).
44 /// We resolve this issue by generating the equivalent "phi of ops" form for
45 /// each op of phis we see, in a way that only takes polynomial time to resolve.
46 ///
47 /// We also do not perform elimination by using any published algorithm. All
48 /// published algorithms are O(Instructions). Instead, we use a technique that
49 /// is O(number of operations with the same value number), enabling us to skip
50 /// trying to eliminate things that have unique value numbers.
51 //
52 //===----------------------------------------------------------------------===//
53 
55 #include "llvm/ADT/ArrayRef.h"
56 #include "llvm/ADT/BitVector.h"
57 #include "llvm/ADT/DenseMap.h"
58 #include "llvm/ADT/DenseMapInfo.h"
59 #include "llvm/ADT/DenseSet.h"
61 #include "llvm/ADT/GraphTraits.h"
62 #include "llvm/ADT/Hashing.h"
65 #include "llvm/ADT/SmallPtrSet.h"
66 #include "llvm/ADT/SmallVector.h"
68 #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"
110 #include <algorithm>
111 #include <cassert>
112 #include <cstdint>
113 #include <iterator>
114 #include <map>
115 #include <memory>
116 #include <set>
117 #include <string>
118 #include <tuple>
119 #include <utility>
120 #include <vector>
121 
122 using namespace llvm;
123 using namespace llvm::GVNExpression;
124 using namespace llvm::VNCoercion;
125 
126 #define DEBUG_TYPE "newgvn"
127 
128 STATISTIC(NumGVNInstrDeleted, "Number of instructions deleted");
129 STATISTIC(NumGVNBlocksDeleted, "Number of blocks deleted");
130 STATISTIC(NumGVNOpsSimplified, "Number of Expressions simplified");
131 STATISTIC(NumGVNPhisAllSame, "Number of PHIs whos arguments are all the same");
132 STATISTIC(NumGVNMaxIterations,
133  "Maximum Number of iterations it took to converge GVN");
134 STATISTIC(NumGVNLeaderChanges, "Number of leader changes");
135 STATISTIC(NumGVNSortedLeaderChanges, "Number of sorted leader changes");
136 STATISTIC(NumGVNAvoidedSortedLeaderChanges,
137  "Number of avoided sorted leader changes");
138 STATISTIC(NumGVNDeadStores, "Number of redundant/dead stores eliminated");
139 STATISTIC(NumGVNPHIOfOpsCreated, "Number of PHI of ops created");
140 STATISTIC(NumGVNPHIOfOpsEliminations,
141  "Number of things eliminated using PHI of ops");
142 DEBUG_COUNTER(VNCounter, "newgvn-vn",
143  "Controls which instructions are value numbered");
144 DEBUG_COUNTER(PHIOfOpsCounter, "newgvn-phi",
145  "Controls which instructions we create phi of ops for");
146 // Currently store defining access refinement is too slow due to basicaa being
147 // egregiously slow. This flag lets us keep it working while we work on this
148 // issue.
149 static cl::opt<bool> EnableStoreRefinement("enable-store-refinement",
150  cl::init(false), cl::Hidden);
151 
152 /// Currently, the generation "phi of ops" can result in correctness issues.
153 static cl::opt<bool> EnablePhiOfOps("enable-phi-of-ops", cl::init(true),
154  cl::Hidden);
155 
156 //===----------------------------------------------------------------------===//
157 // GVN Pass
158 //===----------------------------------------------------------------------===//
159 
160 // Anchor methods.
161 namespace llvm {
162 namespace GVNExpression {
163 
164 Expression::~Expression() = default;
171 
172 } // end namespace GVNExpression
173 } // end namespace llvm
174 
175 namespace {
176 
177 // Tarjan's SCC finding algorithm with Nuutila's improvements
178 // SCCIterator is actually fairly complex for the simple thing we want.
179 // It also wants to hand us SCC's that are unrelated to the phi node we ask
180 // about, and have us process them there or risk redoing work.
181 // Graph traits over a filter iterator also doesn't work that well here.
182 // This SCC finder is specialized to walk use-def chains, and only follows
183 // instructions,
184 // not generic values (arguments, etc).
185 struct TarjanSCC {
186  TarjanSCC() : Components(1) {}
187 
188  void Start(const Instruction *Start) {
189  if (Root.lookup(Start) == 0)
190  FindSCC(Start);
191  }
192 
193  const SmallPtrSetImpl<const Value *> &getComponentFor(const Value *V) const {
194  unsigned ComponentID = ValueToComponent.lookup(V);
195 
196  assert(ComponentID > 0 &&
197  "Asking for a component for a value we never processed");
198  return Components[ComponentID];
199  }
200 
201 private:
202  void FindSCC(const Instruction *I) {
203  Root[I] = ++DFSNum;
204  // Store the DFS Number we had before it possibly gets incremented.
205  unsigned int OurDFS = DFSNum;
206  for (auto &Op : I->operands()) {
207  if (auto *InstOp = dyn_cast<Instruction>(Op)) {
208  if (Root.lookup(Op) == 0)
209  FindSCC(InstOp);
210  if (!InComponent.count(Op))
211  Root[I] = std::min(Root.lookup(I), Root.lookup(Op));
212  }
213  }
214  // See if we really were the root of a component, by seeing if we still have
215  // our DFSNumber. If we do, we are the root of the component, and we have
216  // completed a component. If we do not, we are not the root of a component,
217  // and belong on the component stack.
218  if (Root.lookup(I) == OurDFS) {
219  unsigned ComponentID = Components.size();
220  Components.resize(Components.size() + 1);
221  auto &Component = Components.back();
222  Component.insert(I);
223  LLVM_DEBUG(dbgs() << "Component root is " << *I << "\n");
224  InComponent.insert(I);
225  ValueToComponent[I] = ComponentID;
226  // Pop a component off the stack and label it.
227  while (!Stack.empty() && Root.lookup(Stack.back()) >= OurDFS) {
228  auto *Member = Stack.back();
229  LLVM_DEBUG(dbgs() << "Component member is " << *Member << "\n");
230  Component.insert(Member);
231  InComponent.insert(Member);
232  ValueToComponent[Member] = ComponentID;
233  Stack.pop_back();
234  }
235  } else {
236  // Part of a component, push to stack
237  Stack.push_back(I);
238  }
239  }
240 
241  unsigned int DFSNum = 1;
242  SmallPtrSet<const Value *, 8> InComponent;
245 
246  // Store the components as vector of ptr sets, because we need the topo order
247  // of SCC's, but not individual member order
249 
250  DenseMap<const Value *, unsigned> ValueToComponent;
251 };
252 
253 // Congruence classes represent the set of expressions/instructions
254 // that are all the same *during some scope in the function*.
255 // That is, because of the way we perform equality propagation, and
256 // because of memory value numbering, it is not correct to assume
257 // you can willy-nilly replace any member with any other at any
258 // point in the function.
259 //
260 // For any Value in the Member set, it is valid to replace any dominated member
261 // with that Value.
262 //
263 // Every congruence class has a leader, and the leader is used to symbolize
264 // instructions in a canonical way (IE every operand of an instruction that is a
265 // member of the same congruence class will always be replaced with leader
266 // during symbolization). To simplify symbolization, we keep the leader as a
267 // constant if class can be proved to be a constant value. Otherwise, the
268 // leader is the member of the value set with the smallest DFS number. Each
269 // congruence class also has a defining expression, though the expression may be
270 // null. If it exists, it can be used for forward propagation and reassociation
271 // of values.
272 
273 // For memory, we also track a representative MemoryAccess, and a set of memory
274 // members for MemoryPhis (which have no real instructions). Note that for
275 // memory, it seems tempting to try to split the memory members into a
276 // MemoryCongruenceClass or something. Unfortunately, this does not work
277 // easily. The value numbering of a given memory expression depends on the
278 // leader of the memory congruence class, and the leader of memory congruence
279 // class depends on the value numbering of a given memory expression. This
280 // leads to wasted propagation, and in some cases, missed optimization. For
281 // example: If we had value numbered two stores together before, but now do not,
282 // we move them to a new value congruence class. This in turn will move at one
283 // of the memorydefs to a new memory congruence class. Which in turn, affects
284 // the value numbering of the stores we just value numbered (because the memory
285 // congruence class is part of the value number). So while theoretically
286 // possible to split them up, it turns out to be *incredibly* complicated to get
287 // it to work right, because of the interdependency. While structurally
288 // slightly messier, it is algorithmically much simpler and faster to do what we
289 // do here, and track them both at once in the same class.
290 // Note: The default iterators for this class iterate over values
291 class CongruenceClass {
292 public:
293  using MemberType = Value;
294  using MemberSet = SmallPtrSet<MemberType *, 4>;
295  using MemoryMemberType = MemoryPhi;
296  using MemoryMemberSet = SmallPtrSet<const MemoryMemberType *, 2>;
297 
298  explicit CongruenceClass(unsigned ID) : ID(ID) {}
299  CongruenceClass(unsigned ID, Value *Leader, const Expression *E)
300  : ID(ID), RepLeader(Leader), DefiningExpr(E) {}
301 
302  unsigned getID() const { return ID; }
303 
304  // True if this class has no members left. This is mainly used for assertion
305  // purposes, and for skipping empty classes.
306  bool isDead() const {
307  // If it's both dead from a value perspective, and dead from a memory
308  // perspective, it's really dead.
309  return empty() && memory_empty();
310  }
311 
312  // Leader functions
313  Value *getLeader() const { return RepLeader; }
314  void setLeader(Value *Leader) { RepLeader = Leader; }
315  const std::pair<Value *, unsigned int> &getNextLeader() const {
316  return NextLeader;
317  }
318  void resetNextLeader() { NextLeader = {nullptr, ~0}; }
319  void addPossibleNextLeader(std::pair<Value *, unsigned int> LeaderPair) {
320  if (LeaderPair.second < NextLeader.second)
321  NextLeader = LeaderPair;
322  }
323 
324  Value *getStoredValue() const { return RepStoredValue; }
325  void setStoredValue(Value *Leader) { RepStoredValue = Leader; }
326  const MemoryAccess *getMemoryLeader() const { return RepMemoryAccess; }
327  void setMemoryLeader(const MemoryAccess *Leader) { RepMemoryAccess = Leader; }
328 
329  // Forward propagation info
330  const Expression *getDefiningExpr() const { return DefiningExpr; }
331 
332  // Value member set
333  bool empty() const { return Members.empty(); }
334  unsigned size() const { return Members.size(); }
335  MemberSet::const_iterator begin() const { return Members.begin(); }
336  MemberSet::const_iterator end() const { return Members.end(); }
337  void insert(MemberType *M) { Members.insert(M); }
338  void erase(MemberType *M) { Members.erase(M); }
339  void swap(MemberSet &Other) { Members.swap(Other); }
340 
341  // Memory member set
342  bool memory_empty() const { return MemoryMembers.empty(); }
343  unsigned memory_size() const { return MemoryMembers.size(); }
344  MemoryMemberSet::const_iterator memory_begin() const {
345  return MemoryMembers.begin();
346  }
347  MemoryMemberSet::const_iterator memory_end() const {
348  return MemoryMembers.end();
349  }
351  return make_range(memory_begin(), memory_end());
352  }
353 
354  void memory_insert(const MemoryMemberType *M) { MemoryMembers.insert(M); }
355  void memory_erase(const MemoryMemberType *M) { MemoryMembers.erase(M); }
356 
357  // Store count
358  unsigned getStoreCount() const { return StoreCount; }
359  void incStoreCount() { ++StoreCount; }
360  void decStoreCount() {
361  assert(StoreCount != 0 && "Store count went negative");
362  --StoreCount;
363  }
364 
365  // True if this class has no memory members.
366  bool definesNoMemory() const { return StoreCount == 0 && memory_empty(); }
367 
368  // Return true if two congruence classes are equivalent to each other. This
369  // means that every field but the ID number and the dead field are equivalent.
370  bool isEquivalentTo(const CongruenceClass *Other) const {
371  if (!Other)
372  return false;
373  if (this == Other)
374  return true;
375 
376  if (std::tie(StoreCount, RepLeader, RepStoredValue, RepMemoryAccess) !=
377  std::tie(Other->StoreCount, Other->RepLeader, Other->RepStoredValue,
378  Other->RepMemoryAccess))
379  return false;
380  if (DefiningExpr != Other->DefiningExpr)
381  if (!DefiningExpr || !Other->DefiningExpr ||
382  *DefiningExpr != *Other->DefiningExpr)
383  return false;
384 
385  if (Members.size() != Other->Members.size())
386  return false;
387 
388  return all_of(Members,
389  [&](const Value *V) { return Other->Members.count(V); });
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)),
660  SQ(DL, TLI, DT, AC, /*CtxI=*/nullptr, /*UseInstrInfo=*/false) {}
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(Instruction *, 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  int64_t 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  llvm::sort(Ops, [&](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  LLVM_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  LLVM_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  // If we simplified to something else, we need to communicate
1089  // that we're users of the value we simplified to.
1090  if (I != V) {
1091  // Don't add temporary instructions to the user lists.
1092  if (!AllTempInstructions.count(I))
1093  addAdditionalUsers(V, I);
1094  }
1095  return createVariableOrConstant(CC->getLeader());
1096  }
1097  if (CC->getDefiningExpr()) {
1098  // If we simplified to something else, we need to communicate
1099  // that we're users of the value we simplified to.
1100  if (I != V) {
1101  // Don't add temporary instructions to the user lists.
1102  if (!AllTempInstructions.count(I))
1103  addAdditionalUsers(V, I);
1104  }
1105 
1106  if (I)
1107  LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1108  << " expression " << *CC->getDefiningExpr() << "\n");
1109  NumGVNOpsSimplified++;
1110  deleteExpression(E);
1111  return CC->getDefiningExpr();
1112  }
1113  }
1114 
1115  return nullptr;
1116 }
1117 
1118 // Create a value expression from the instruction I, replacing operands with
1119 // their leaders.
1120 
1121 const Expression *NewGVN::createExpression(Instruction *I) const {
1122  auto *E = new (ExpressionAllocator) BasicExpression(I->getNumOperands());
1123 
1124  bool AllConstant = setBasicExpressionInfo(I, E);
1125 
1126  if (I->isCommutative()) {
1127  // Ensure that commutative instructions that only differ by a permutation
1128  // of their operands get the same value number by sorting the operand value
1129  // numbers. Since all commutative instructions have two operands it is more
1130  // efficient to sort by hand rather than using, say, std::sort.
1131  assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
1132  if (shouldSwapOperands(E->getOperand(0), E->getOperand(1)))
1133  E->swapOperands(0, 1);
1134  }
1135  // Perform simplification.
1136  if (auto *CI = dyn_cast<CmpInst>(I)) {
1137  // Sort the operand value numbers so x<y and y>x get the same value
1138  // number.
1139  CmpInst::Predicate Predicate = CI->getPredicate();
1140  if (shouldSwapOperands(E->getOperand(0), E->getOperand(1))) {
1141  E->swapOperands(0, 1);
1142  Predicate = CmpInst::getSwappedPredicate(Predicate);
1143  }
1144  E->setOpcode((CI->getOpcode() << 8) | Predicate);
1145  // TODO: 25% of our time is spent in SimplifyCmpInst with pointer operands
1146  assert(I->getOperand(0)->getType() == I->getOperand(1)->getType() &&
1147  "Wrong types on cmp instruction");
1148  assert((E->getOperand(0)->getType() == I->getOperand(0)->getType() &&
1149  E->getOperand(1)->getType() == I->getOperand(1)->getType()));
1150  Value *V =
1151  SimplifyCmpInst(Predicate, E->getOperand(0), E->getOperand(1), SQ);
1152  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1153  return SimplifiedE;
1154  } else if (isa<SelectInst>(I)) {
1155  if (isa<Constant>(E->getOperand(0)) ||
1156  E->getOperand(1) == E->getOperand(2)) {
1157  assert(E->getOperand(1)->getType() == I->getOperand(1)->getType() &&
1158  E->getOperand(2)->getType() == I->getOperand(2)->getType());
1159  Value *V = SimplifySelectInst(E->getOperand(0), E->getOperand(1),
1160  E->getOperand(2), SQ);
1161  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1162  return SimplifiedE;
1163  }
1164  } else if (I->isBinaryOp()) {
1165  Value *V =
1166  SimplifyBinOp(E->getOpcode(), E->getOperand(0), E->getOperand(1), SQ);
1167  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1168  return SimplifiedE;
1169  } else if (auto *BI = dyn_cast<BitCastInst>(I)) {
1170  Value *V =
1171  SimplifyCastInst(BI->getOpcode(), BI->getOperand(0), BI->getType(), SQ);
1172  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1173  return SimplifiedE;
1174  } else if (isa<GetElementPtrInst>(I)) {
1175  Value *V = SimplifyGEPInst(
1176  E->getType(), ArrayRef<Value *>(E->op_begin(), E->op_end()), SQ);
1177  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1178  return SimplifiedE;
1179  } else if (AllConstant) {
1180  // We don't bother trying to simplify unless all of the operands
1181  // were constant.
1182  // TODO: There are a lot of Simplify*'s we could call here, if we
1183  // wanted to. The original motivating case for this code was a
1184  // zext i1 false to i8, which we don't have an interface to
1185  // simplify (IE there is no SimplifyZExt).
1186 
1188  for (Value *Arg : E->operands())
1189  C.emplace_back(cast<Constant>(Arg));
1190 
1191  if (Value *V = ConstantFoldInstOperands(I, C, DL, TLI))
1192  if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
1193  return SimplifiedE;
1194  }
1195  return E;
1196 }
1197 
1199 NewGVN::createAggregateValueExpression(Instruction *I) const {
1200  if (auto *II = dyn_cast<InsertValueInst>(I)) {
1201  auto *E = new (ExpressionAllocator)
1202  AggregateValueExpression(I->getNumOperands(), II->getNumIndices());
1203  setBasicExpressionInfo(I, E);
1204  E->allocateIntOperands(ExpressionAllocator);
1205  std::copy(II->idx_begin(), II->idx_end(), int_op_inserter(E));
1206  return E;
1207  } else if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1208  auto *E = new (ExpressionAllocator)
1209  AggregateValueExpression(I->getNumOperands(), EI->getNumIndices());
1210  setBasicExpressionInfo(EI, E);
1211  E->allocateIntOperands(ExpressionAllocator);
1212  std::copy(EI->idx_begin(), EI->idx_end(), int_op_inserter(E));
1213  return E;
1214  }
1215  llvm_unreachable("Unhandled type of aggregate value operation");
1216 }
1217 
1218 const DeadExpression *NewGVN::createDeadExpression() const {
1219  // DeadExpression has no arguments and all DeadExpression's are the same,
1220  // so we only need one of them.
1221  return SingletonDeadExpression;
1222 }
1223 
1224 const VariableExpression *NewGVN::createVariableExpression(Value *V) const {
1225  auto *E = new (ExpressionAllocator) VariableExpression(V);
1226  E->setOpcode(V->getValueID());
1227  return E;
1228 }
1229 
1230 const Expression *NewGVN::createVariableOrConstant(Value *V) const {
1231  if (auto *C = dyn_cast<Constant>(V))
1232  return createConstantExpression(C);
1233  return createVariableExpression(V);
1234 }
1235 
1236 const ConstantExpression *NewGVN::createConstantExpression(Constant *C) const {
1237  auto *E = new (ExpressionAllocator) ConstantExpression(C);
1238  E->setOpcode(C->getValueID());
1239  return E;
1240 }
1241 
1242 const UnknownExpression *NewGVN::createUnknownExpression(Instruction *I) const {
1243  auto *E = new (ExpressionAllocator) UnknownExpression(I);
1244  E->setOpcode(I->getOpcode());
1245  return E;
1246 }
1247 
1248 const CallExpression *
1249 NewGVN::createCallExpression(CallInst *CI, const MemoryAccess *MA) const {
1250  // FIXME: Add operand bundles for calls.
1251  auto *E =
1252  new (ExpressionAllocator) CallExpression(CI->getNumOperands(), CI, MA);
1253  setBasicExpressionInfo(CI, E);
1254  return E;
1255 }
1256 
1257 // Return true if some equivalent of instruction Inst dominates instruction U.
1258 bool NewGVN::someEquivalentDominates(const Instruction *Inst,
1259  const Instruction *U) const {
1260  auto *CC = ValueToClass.lookup(Inst);
1261  // This must be an instruction because we are only called from phi nodes
1262  // in the case that the value it needs to check against is an instruction.
1263 
1264  // The most likely candidates for dominance are the leader and the next leader.
1265  // The leader or nextleader will dominate in all cases where there is an
1266  // equivalent that is higher up in the dom tree.
1267  // We can't *only* check them, however, because the
1268  // dominator tree could have an infinite number of non-dominating siblings
1269  // with instructions that are in the right congruence class.
1270  // A
1271  // B C D E F G
1272  // |
1273  // H
1274  // Instruction U could be in H, with equivalents in every other sibling.
1275  // Depending on the rpo order picked, the leader could be the equivalent in
1276  // any of these siblings.
1277  if (!CC)
1278  return false;
1279  if (alwaysAvailable(CC->getLeader()))
1280  return true;
1281  if (DT->dominates(cast<Instruction>(CC->getLeader()), U))
1282  return true;
1283  if (CC->getNextLeader().first &&
1284  DT->dominates(cast<Instruction>(CC->getNextLeader().first), U))
1285  return true;
1286  return llvm::any_of(*CC, [&](const Value *Member) {
1287  return Member != CC->getLeader() &&
1288  DT->dominates(cast<Instruction>(Member), U);
1289  });
1290 }
1291 
1292 // See if we have a congruence class and leader for this operand, and if so,
1293 // return it. Otherwise, return the operand itself.
1294 Value *NewGVN::lookupOperandLeader(Value *V) const {
1295  CongruenceClass *CC = ValueToClass.lookup(V);
1296  if (CC) {
1297  // Everything in TOP is represented by undef, as it can be any value.
1298  // We do have to make sure we get the type right though, so we can't set the
1299  // RepLeader to undef.
1300  if (CC == TOPClass)
1301  return UndefValue::get(V->getType());
1302  return CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
1303  }
1304 
1305  return V;
1306 }
1307 
1308 const MemoryAccess *NewGVN::lookupMemoryLeader(const MemoryAccess *MA) const {
1309  auto *CC = getMemoryClass(MA);
1310  assert(CC->getMemoryLeader() &&
1311  "Every MemoryAccess should be mapped to a congruence class with a "
1312  "representative memory access");
1313  return CC->getMemoryLeader();
1314 }
1315 
1316 // Return true if the MemoryAccess is really equivalent to everything. This is
1317 // equivalent to the lattice value "TOP" in most lattices. This is the initial
1318 // state of all MemoryAccesses.
1319 bool NewGVN::isMemoryAccessTOP(const MemoryAccess *MA) const {
1320  return getMemoryClass(MA) == TOPClass;
1321 }
1322 
1323 LoadExpression *NewGVN::createLoadExpression(Type *LoadType, Value *PointerOp,
1324  LoadInst *LI,
1325  const MemoryAccess *MA) const {
1326  auto *E =
1327  new (ExpressionAllocator) LoadExpression(1, LI, lookupMemoryLeader(MA));
1328  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1329  E->setType(LoadType);
1330 
1331  // Give store and loads same opcode so they value number together.
1332  E->setOpcode(0);
1333  E->op_push_back(PointerOp);
1334  if (LI)
1335  E->setAlignment(LI->getAlignment());
1336 
1337  // TODO: Value number heap versions. We may be able to discover
1338  // things alias analysis can't on it's own (IE that a store and a
1339  // load have the same value, and thus, it isn't clobbering the load).
1340  return E;
1341 }
1342 
1343 const StoreExpression *
1344 NewGVN::createStoreExpression(StoreInst *SI, const MemoryAccess *MA) const {
1345  auto *StoredValueLeader = lookupOperandLeader(SI->getValueOperand());
1346  auto *E = new (ExpressionAllocator)
1347  StoreExpression(SI->getNumOperands(), SI, StoredValueLeader, MA);
1348  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1349  E->setType(SI->getValueOperand()->getType());
1350 
1351  // Give store and loads same opcode so they value number together.
1352  E->setOpcode(0);
1353  E->op_push_back(lookupOperandLeader(SI->getPointerOperand()));
1354 
1355  // TODO: Value number heap versions. We may be able to discover
1356  // things alias analysis can't on it's own (IE that a store and a
1357  // load have the same value, and thus, it isn't clobbering the load).
1358  return E;
1359 }
1360 
1361 const Expression *NewGVN::performSymbolicStoreEvaluation(Instruction *I) const {
1362  // Unlike loads, we never try to eliminate stores, so we do not check if they
1363  // are simple and avoid value numbering them.
1364  auto *SI = cast<StoreInst>(I);
1365  auto *StoreAccess = getMemoryAccess(SI);
1366  // Get the expression, if any, for the RHS of the MemoryDef.
1367  const MemoryAccess *StoreRHS = StoreAccess->getDefiningAccess();
1369  StoreRHS = MSSAWalker->getClobberingMemoryAccess(StoreAccess);
1370  // If we bypassed the use-def chains, make sure we add a use.
1371  StoreRHS = lookupMemoryLeader(StoreRHS);
1372  if (StoreRHS != StoreAccess->getDefiningAccess())
1373  addMemoryUsers(StoreRHS, StoreAccess);
1374  // If we are defined by ourselves, use the live on entry def.
1375  if (StoreRHS == StoreAccess)
1376  StoreRHS = MSSA->getLiveOnEntryDef();
1377 
1378  if (SI->isSimple()) {
1379  // See if we are defined by a previous store expression, it already has a
1380  // value, and it's the same value as our current store. FIXME: Right now, we
1381  // only do this for simple stores, we should expand to cover memcpys, etc.
1382  const auto *LastStore = createStoreExpression(SI, StoreRHS);
1383  const auto *LastCC = ExpressionToClass.lookup(LastStore);
1384  // We really want to check whether the expression we matched was a store. No
1385  // easy way to do that. However, we can check that the class we found has a
1386  // store, which, assuming the value numbering state is not corrupt, is
1387  // sufficient, because we must also be equivalent to that store's expression
1388  // for it to be in the same class as the load.
1389  if (LastCC && LastCC->getStoredValue() == LastStore->getStoredValue())
1390  return LastStore;
1391  // Also check if our value operand is defined by a load of the same memory
1392  // location, and the memory state is the same as it was then (otherwise, it
1393  // could have been overwritten later. See test32 in
1394  // transforms/DeadStoreElimination/simple.ll).
1395  if (auto *LI = dyn_cast<LoadInst>(LastStore->getStoredValue()))
1396  if ((lookupOperandLeader(LI->getPointerOperand()) ==
1397  LastStore->getOperand(0)) &&
1398  (lookupMemoryLeader(getMemoryAccess(LI)->getDefiningAccess()) ==
1399  StoreRHS))
1400  return LastStore;
1401  deleteExpression(LastStore);
1402  }
1403 
1404  // If the store is not equivalent to anything, value number it as a store that
1405  // produces a unique memory state (instead of using it's MemoryUse, we use
1406  // it's MemoryDef).
1407  return createStoreExpression(SI, StoreAccess);
1408 }
1409 
1410 // See if we can extract the value of a loaded pointer from a load, a store, or
1411 // a memory instruction.
1412 const Expression *
1413 NewGVN::performSymbolicLoadCoercion(Type *LoadType, Value *LoadPtr,
1414  LoadInst *LI, Instruction *DepInst,
1415  MemoryAccess *DefiningAccess) const {
1416  assert((!LI || LI->isSimple()) && "Not a simple load");
1417  if (auto *DepSI = dyn_cast<StoreInst>(DepInst)) {
1418  // Can't forward from non-atomic to atomic without violating memory model.
1419  // Also don't need to coerce if they are the same type, we will just
1420  // propagate.
1421  if (LI->isAtomic() > DepSI->isAtomic() ||
1422  LoadType == DepSI->getValueOperand()->getType())
1423  return nullptr;
1424  int Offset = analyzeLoadFromClobberingStore(LoadType, LoadPtr, DepSI, DL);
1425  if (Offset >= 0) {
1426  if (auto *C = dyn_cast<Constant>(
1427  lookupOperandLeader(DepSI->getValueOperand()))) {
1428  LLVM_DEBUG(dbgs() << "Coercing load from store " << *DepSI
1429  << " to constant " << *C << "\n");
1430  return createConstantExpression(
1431  getConstantStoreValueForLoad(C, Offset, LoadType, DL));
1432  }
1433  }
1434  } else if (auto *DepLI = dyn_cast<LoadInst>(DepInst)) {
1435  // Can't forward from non-atomic to atomic without violating memory model.
1436  if (LI->isAtomic() > DepLI->isAtomic())
1437  return nullptr;
1438  int Offset = analyzeLoadFromClobberingLoad(LoadType, LoadPtr, DepLI, DL);
1439  if (Offset >= 0) {
1440  // We can coerce a constant load into a load.
1441  if (auto *C = dyn_cast<Constant>(lookupOperandLeader(DepLI)))
1442  if (auto *PossibleConstant =
1443  getConstantLoadValueForLoad(C, Offset, LoadType, DL)) {
1444  LLVM_DEBUG(dbgs() << "Coercing load from load " << *LI
1445  << " to constant " << *PossibleConstant << "\n");
1446  return createConstantExpression(PossibleConstant);
1447  }
1448  }
1449  } else if (auto *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
1450  int Offset = analyzeLoadFromClobberingMemInst(LoadType, LoadPtr, DepMI, DL);
1451  if (Offset >= 0) {
1452  if (auto *PossibleConstant =
1453  getConstantMemInstValueForLoad(DepMI, Offset, LoadType, DL)) {
1454  LLVM_DEBUG(dbgs() << "Coercing load from meminst " << *DepMI
1455  << " to constant " << *PossibleConstant << "\n");
1456  return createConstantExpression(PossibleConstant);
1457  }
1458  }
1459  }
1460 
1461  // All of the below are only true if the loaded pointer is produced
1462  // by the dependent instruction.
1463  if (LoadPtr != lookupOperandLeader(DepInst) &&
1464  !AA->isMustAlias(LoadPtr, DepInst))
1465  return nullptr;
1466  // If this load really doesn't depend on anything, then we must be loading an
1467  // undef value. This can happen when loading for a fresh allocation with no
1468  // intervening stores, for example. Note that this is only true in the case
1469  // that the result of the allocation is pointer equal to the load ptr.
1470  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1471  return createConstantExpression(UndefValue::get(LoadType));
1472  }
1473  // If this load occurs either right after a lifetime begin,
1474  // then the loaded value is undefined.
1475  else if (auto *II = dyn_cast<IntrinsicInst>(DepInst)) {
1476  if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1477  return createConstantExpression(UndefValue::get(LoadType));
1478  }
1479  // If this load follows a calloc (which zero initializes memory),
1480  // then the loaded value is zero
1481  else if (isCallocLikeFn(DepInst, TLI)) {
1482  return createConstantExpression(Constant::getNullValue(LoadType));
1483  }
1484 
1485  return nullptr;
1486 }
1487 
1488 const Expression *NewGVN::performSymbolicLoadEvaluation(Instruction *I) const {
1489  auto *LI = cast<LoadInst>(I);
1490 
1491  // We can eliminate in favor of non-simple loads, but we won't be able to
1492  // eliminate the loads themselves.
1493  if (!LI->isSimple())
1494  return nullptr;
1495 
1496  Value *LoadAddressLeader = lookupOperandLeader(LI->getPointerOperand());
1497  // Load of undef is undef.
1498  if (isa<UndefValue>(LoadAddressLeader))
1499  return createConstantExpression(UndefValue::get(LI->getType()));
1500  MemoryAccess *OriginalAccess = getMemoryAccess(I);
1501  MemoryAccess *DefiningAccess =
1502  MSSAWalker->getClobberingMemoryAccess(OriginalAccess);
1503 
1504  if (!MSSA->isLiveOnEntryDef(DefiningAccess)) {
1505  if (auto *MD = dyn_cast<MemoryDef>(DefiningAccess)) {
1506  Instruction *DefiningInst = MD->getMemoryInst();
1507  // If the defining instruction is not reachable, replace with undef.
1508  if (!ReachableBlocks.count(DefiningInst->getParent()))
1509  return createConstantExpression(UndefValue::get(LI->getType()));
1510  // This will handle stores and memory insts. We only do if it the
1511  // defining access has a different type, or it is a pointer produced by
1512  // certain memory operations that cause the memory to have a fixed value
1513  // (IE things like calloc).
1514  if (const auto *CoercionResult =
1515  performSymbolicLoadCoercion(LI->getType(), LoadAddressLeader, LI,
1516  DefiningInst, DefiningAccess))
1517  return CoercionResult;
1518  }
1519  }
1520 
1521  const auto *LE = createLoadExpression(LI->getType(), LoadAddressLeader, LI,
1522  DefiningAccess);
1523  // If our MemoryLeader is not our defining access, add a use to the
1524  // MemoryLeader, so that we get reprocessed when it changes.
1525  if (LE->getMemoryLeader() != DefiningAccess)
1526  addMemoryUsers(LE->getMemoryLeader(), OriginalAccess);
1527  return LE;
1528 }
1529 
1530 const Expression *
1531 NewGVN::performSymbolicPredicateInfoEvaluation(Instruction *I) const {
1532  auto *PI = PredInfo->getPredicateInfoFor(I);
1533  if (!PI)
1534  return nullptr;
1535 
1536  LLVM_DEBUG(dbgs() << "Found predicate info from instruction !\n");
1537 
1538  auto *PWC = dyn_cast<PredicateWithCondition>(PI);
1539  if (!PWC)
1540  return nullptr;
1541 
1542  auto *CopyOf = I->getOperand(0);
1543  auto *Cond = PWC->Condition;
1544 
1545  // If this a copy of the condition, it must be either true or false depending
1546  // on the predicate info type and edge.
1547  if (CopyOf == Cond) {
1548  // We should not need to add predicate users because the predicate info is
1549  // already a use of this operand.
1550  if (isa<PredicateAssume>(PI))
1551  return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1552  if (auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1553  if (PBranch->TrueEdge)
1554  return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1555  return createConstantExpression(ConstantInt::getFalse(Cond->getType()));
1556  }
1557  if (auto *PSwitch = dyn_cast<PredicateSwitch>(PI))
1558  return createConstantExpression(cast<Constant>(PSwitch->CaseValue));
1559  }
1560 
1561  // Not a copy of the condition, so see what the predicates tell us about this
1562  // value. First, though, we check to make sure the value is actually a copy
1563  // of one of the condition operands. It's possible, in certain cases, for it
1564  // to be a copy of a predicateinfo copy. In particular, if two branch
1565  // operations use the same condition, and one branch dominates the other, we
1566  // will end up with a copy of a copy. This is currently a small deficiency in
1567  // predicateinfo. What will end up happening here is that we will value
1568  // number both copies the same anyway.
1569 
1570  // Everything below relies on the condition being a comparison.
1571  auto *Cmp = dyn_cast<CmpInst>(Cond);
1572  if (!Cmp)
1573  return nullptr;
1574 
1575  if (CopyOf != Cmp->getOperand(0) && CopyOf != Cmp->getOperand(1)) {
1576  LLVM_DEBUG(dbgs() << "Copy is not of any condition operands!\n");
1577  return nullptr;
1578  }
1579  Value *FirstOp = lookupOperandLeader(Cmp->getOperand(0));
1580  Value *SecondOp = lookupOperandLeader(Cmp->getOperand(1));
1581  bool SwappedOps = false;
1582  // Sort the ops.
1583  if (shouldSwapOperands(FirstOp, SecondOp)) {
1584  std::swap(FirstOp, SecondOp);
1585  SwappedOps = true;
1586  }
1588  SwappedOps ? Cmp->getSwappedPredicate() : Cmp->getPredicate();
1589 
1590  if (isa<PredicateAssume>(PI)) {
1591  // If we assume the operands are equal, then they are equal.
1592  if (Predicate == CmpInst::ICMP_EQ) {
1593  addPredicateUsers(PI, I);
1594  addAdditionalUsers(SwappedOps ? Cmp->getOperand(1) : Cmp->getOperand(0),
1595  I);
1596  return createVariableOrConstant(FirstOp);
1597  }
1598  }
1599  if (const auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1600  // If we are *not* a copy of the comparison, we may equal to the other
1601  // operand when the predicate implies something about equality of
1602  // operations. In particular, if the comparison is true/false when the
1603  // operands are equal, and we are on the right edge, we know this operation
1604  // is equal to something.
1605  if ((PBranch->TrueEdge && Predicate == CmpInst::ICMP_EQ) ||
1606  (!PBranch->TrueEdge && Predicate == CmpInst::ICMP_NE)) {
1607  addPredicateUsers(PI, I);
1608  addAdditionalUsers(SwappedOps ? Cmp->getOperand(1) : Cmp->getOperand(0),
1609  I);
1610  return createVariableOrConstant(FirstOp);
1611  }
1612  // Handle the special case of floating point.
1613  if (((PBranch->TrueEdge && Predicate == CmpInst::FCMP_OEQ) ||
1614  (!PBranch->TrueEdge && Predicate == CmpInst::FCMP_UNE)) &&
1615  isa<ConstantFP>(FirstOp) && !cast<ConstantFP>(FirstOp)->isZero()) {
1616  addPredicateUsers(PI, I);
1617  addAdditionalUsers(SwappedOps ? Cmp->getOperand(1) : Cmp->getOperand(0),
1618  I);
1619  return createConstantExpression(cast<Constant>(FirstOp));
1620  }
1621  }
1622  return nullptr;
1623 }
1624 
1625 // Evaluate read only and pure calls, and create an expression result.
1626 const Expression *NewGVN::performSymbolicCallEvaluation(Instruction *I) const {
1627  auto *CI = cast<CallInst>(I);
1628  if (auto *II = dyn_cast<IntrinsicInst>(I)) {
1629  // Intrinsics with the returned attribute are copies of arguments.
1630  if (auto *ReturnedValue = II->getReturnedArgOperand()) {
1631  if (II->getIntrinsicID() == Intrinsic::ssa_copy)
1632  if (const auto *Result = performSymbolicPredicateInfoEvaluation(I))
1633  return Result;
1634  return createVariableOrConstant(ReturnedValue);
1635  }
1636  }
1637  if (AA->doesNotAccessMemory(CI)) {
1638  return createCallExpression(CI, TOPClass->getMemoryLeader());
1639  } else if (AA->onlyReadsMemory(CI)) {
1640  MemoryAccess *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(CI);
1641  return createCallExpression(CI, DefiningAccess);
1642  }
1643  return nullptr;
1644 }
1645 
1646 // Retrieve the memory class for a given MemoryAccess.
1647 CongruenceClass *NewGVN::getMemoryClass(const MemoryAccess *MA) const {
1648  auto *Result = MemoryAccessToClass.lookup(MA);
1649  assert(Result && "Should have found memory class");
1650  return Result;
1651 }
1652 
1653 // Update the MemoryAccess equivalence table to say that From is equal to To,
1654 // and return true if this is different from what already existed in the table.
1655 bool NewGVN::setMemoryClass(const MemoryAccess *From,
1656  CongruenceClass *NewClass) {
1657  assert(NewClass &&
1658  "Every MemoryAccess should be getting mapped to a non-null class");
1659  LLVM_DEBUG(dbgs() << "Setting " << *From);
1660  LLVM_DEBUG(dbgs() << " equivalent to congruence class ");
1661  LLVM_DEBUG(dbgs() << NewClass->getID()
1662  << " with current MemoryAccess leader ");
1663  LLVM_DEBUG(dbgs() << *NewClass->getMemoryLeader() << "\n");
1664 
1665  auto LookupResult = MemoryAccessToClass.find(From);
1666  bool Changed = false;
1667  // If it's already in the table, see if the value changed.
1668  if (LookupResult != MemoryAccessToClass.end()) {
1669  auto *OldClass = LookupResult->second;
1670  if (OldClass != NewClass) {
1671  // If this is a phi, we have to handle memory member updates.
1672  if (auto *MP = dyn_cast<MemoryPhi>(From)) {
1673  OldClass->memory_erase(MP);
1674  NewClass->memory_insert(MP);
1675  // This may have killed the class if it had no non-memory members
1676  if (OldClass->getMemoryLeader() == From) {
1677  if (OldClass->definesNoMemory()) {
1678  OldClass->setMemoryLeader(nullptr);
1679  } else {
1680  OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
1681  LLVM_DEBUG(dbgs() << "Memory class leader change for class "
1682  << OldClass->getID() << " to "
1683  << *OldClass->getMemoryLeader()
1684  << " due to removal of a memory member " << *From
1685  << "\n");
1686  markMemoryLeaderChangeTouched(OldClass);
1687  }
1688  }
1689  }
1690  // It wasn't equivalent before, and now it is.
1691  LookupResult->second = NewClass;
1692  Changed = true;
1693  }
1694  }
1695 
1696  return Changed;
1697 }
1698 
1699 // Determine if a instruction is cycle-free. That means the values in the
1700 // instruction don't depend on any expressions that can change value as a result
1701 // of the instruction. For example, a non-cycle free instruction would be v =
1702 // phi(0, v+1).
1703 bool NewGVN::isCycleFree(const Instruction *I) const {
1704  // In order to compute cycle-freeness, we do SCC finding on the instruction,
1705  // and see what kind of SCC it ends up in. If it is a singleton, it is
1706  // cycle-free. If it is not in a singleton, it is only cycle free if the
1707  // other members are all phi nodes (as they do not compute anything, they are
1708  // copies).
1709  auto ICS = InstCycleState.lookup(I);
1710  if (ICS == ICS_Unknown) {
1711  SCCFinder.Start(I);
1712  auto &SCC = SCCFinder.getComponentFor(I);
1713  // It's cycle free if it's size 1 or the SCC is *only* phi nodes.
1714  if (SCC.size() == 1)
1715  InstCycleState.insert({I, ICS_CycleFree});
1716  else {
1717  bool AllPhis = llvm::all_of(SCC, [](const Value *V) {
1718  return isa<PHINode>(V) || isCopyOfAPHI(V);
1719  });
1720  ICS = AllPhis ? ICS_CycleFree : ICS_Cycle;
1721  for (auto *Member : SCC)
1722  if (auto *MemberPhi = dyn_cast<PHINode>(Member))
1723  InstCycleState.insert({MemberPhi, ICS});
1724  }
1725  }
1726  if (ICS == ICS_Cycle)
1727  return false;
1728  return true;
1729 }
1730 
1731 // Evaluate PHI nodes symbolically and create an expression result.
1732 const Expression *
1733 NewGVN::performSymbolicPHIEvaluation(ArrayRef<ValPair> PHIOps,
1734  Instruction *I,
1735  BasicBlock *PHIBlock) const {
1736  // True if one of the incoming phi edges is a backedge.
1737  bool HasBackedge = false;
1738  // All constant tracks the state of whether all the *original* phi operands
1739  // This is really shorthand for "this phi cannot cycle due to forward
1740  // change in value of the phi is guaranteed not to later change the value of
1741  // the phi. IE it can't be v = phi(undef, v+1)
1742  bool OriginalOpsConstant = true;
1743  auto *E = cast<PHIExpression>(createPHIExpression(
1744  PHIOps, I, PHIBlock, HasBackedge, OriginalOpsConstant));
1745  // We match the semantics of SimplifyPhiNode from InstructionSimplify here.
1746  // See if all arguments are the same.
1747  // We track if any were undef because they need special handling.
1748  bool HasUndef = false;
1749  auto Filtered = make_filter_range(E->operands(), [&](Value *Arg) {
1750  if (isa<UndefValue>(Arg)) {
1751  HasUndef = true;
1752  return false;
1753  }
1754  return true;
1755  });
1756  // If we are left with no operands, it's dead.
1757  if (empty(Filtered)) {
1758  // If it has undef at this point, it means there are no-non-undef arguments,
1759  // and thus, the value of the phi node must be undef.
1760  if (HasUndef) {
1761  LLVM_DEBUG(
1762  dbgs() << "PHI Node " << *I
1763  << " has no non-undef arguments, valuing it as undef\n");
1764  return createConstantExpression(UndefValue::get(I->getType()));
1765  }
1766 
1767  LLVM_DEBUG(dbgs() << "No arguments of PHI node " << *I << " are live\n");
1768  deleteExpression(E);
1769  return createDeadExpression();
1770  }
1771  Value *AllSameValue = *(Filtered.begin());
1772  ++Filtered.begin();
1773  // Can't use std::equal here, sadly, because filter.begin moves.
1774  if (llvm::all_of(Filtered, [&](Value *Arg) { return Arg == AllSameValue; })) {
1775  // In LLVM's non-standard representation of phi nodes, it's possible to have
1776  // phi nodes with cycles (IE dependent on other phis that are .... dependent
1777  // on the original phi node), especially in weird CFG's where some arguments
1778  // are unreachable, or uninitialized along certain paths. This can cause
1779  // infinite loops during evaluation. We work around this by not trying to
1780  // really evaluate them independently, but instead using a variable
1781  // expression to say if one is equivalent to the other.
1782  // We also special case undef, so that if we have an undef, we can't use the
1783  // common value unless it dominates the phi block.
1784  if (HasUndef) {
1785  // If we have undef and at least one other value, this is really a
1786  // multivalued phi, and we need to know if it's cycle free in order to
1787  // evaluate whether we can ignore the undef. The other parts of this are
1788  // just shortcuts. If there is no backedge, or all operands are
1789  // constants, it also must be cycle free.
1790  if (HasBackedge && !OriginalOpsConstant &&
1791  !isa<UndefValue>(AllSameValue) && !isCycleFree(I))
1792  return E;
1793 
1794  // Only have to check for instructions
1795  if (auto *AllSameInst = dyn_cast<Instruction>(AllSameValue))
1796  if (!someEquivalentDominates(AllSameInst, I))
1797  return E;
1798  }
1799  // Can't simplify to something that comes later in the iteration.
1800  // Otherwise, when and if it changes congruence class, we will never catch
1801  // up. We will always be a class behind it.
1802  if (isa<Instruction>(AllSameValue) &&
1803  InstrToDFSNum(AllSameValue) > InstrToDFSNum(I))
1804  return E;
1805  NumGVNPhisAllSame++;
1806  LLVM_DEBUG(dbgs() << "Simplified PHI node " << *I << " to " << *AllSameValue
1807  << "\n");
1808  deleteExpression(E);
1809  return createVariableOrConstant(AllSameValue);
1810  }
1811  return E;
1812 }
1813 
1814 const Expression *
1815 NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) const {
1816  if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1817  auto *II = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
1818  if (II && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
1819  unsigned Opcode = 0;
1820  // EI might be an extract from one of our recognised intrinsics. If it
1821  // is we'll synthesize a semantically equivalent expression instead on
1822  // an extract value expression.
1823  switch (II->getIntrinsicID()) {
1824  case Intrinsic::sadd_with_overflow:
1825  case Intrinsic::uadd_with_overflow:
1826  Opcode = Instruction::Add;
1827  break;
1828  case Intrinsic::ssub_with_overflow:
1829  case Intrinsic::usub_with_overflow:
1830  Opcode = Instruction::Sub;
1831  break;
1832  case Intrinsic::smul_with_overflow:
1833  case Intrinsic::umul_with_overflow:
1834  Opcode = Instruction::Mul;
1835  break;
1836  default:
1837  break;
1838  }
1839 
1840  if (Opcode != 0) {
1841  // Intrinsic recognized. Grab its args to finish building the
1842  // expression.
1843  assert(II->getNumArgOperands() == 2 &&
1844  "Expect two args for recognised intrinsics.");
1845  return createBinaryExpression(Opcode, EI->getType(),
1846  II->getArgOperand(0),
1847  II->getArgOperand(1), I);
1848  }
1849  }
1850  }
1851 
1852  return createAggregateValueExpression(I);
1853 }
1854 
1855 const Expression *NewGVN::performSymbolicCmpEvaluation(Instruction *I) const {
1856  assert(isa<CmpInst>(I) && "Expected a cmp instruction.");
1857 
1858  auto *CI = cast<CmpInst>(I);
1859  // See if our operands are equal to those of a previous predicate, and if so,
1860  // if it implies true or false.
1861  auto Op0 = lookupOperandLeader(CI->getOperand(0));
1862  auto Op1 = lookupOperandLeader(CI->getOperand(1));
1863  auto OurPredicate = CI->getPredicate();
1864  if (shouldSwapOperands(Op0, Op1)) {
1865  std::swap(Op0, Op1);
1866  OurPredicate = CI->getSwappedPredicate();
1867  }
1868 
1869  // Avoid processing the same info twice.
1870  const PredicateBase *LastPredInfo = nullptr;
1871  // See if we know something about the comparison itself, like it is the target
1872  // of an assume.
1873  auto *CmpPI = PredInfo->getPredicateInfoFor(I);
1874  if (dyn_cast_or_null<PredicateAssume>(CmpPI))
1875  return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1876 
1877  if (Op0 == Op1) {
1878  // This condition does not depend on predicates, no need to add users
1879  if (CI->isTrueWhenEqual())
1880  return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1881  else if (CI->isFalseWhenEqual())
1882  return createConstantExpression(ConstantInt::getFalse(CI->getType()));
1883  }
1884 
1885  // NOTE: Because we are comparing both operands here and below, and using
1886  // previous comparisons, we rely on fact that predicateinfo knows to mark
1887  // comparisons that use renamed operands as users of the earlier comparisons.
1888  // It is *not* enough to just mark predicateinfo renamed operands as users of
1889  // the earlier comparisons, because the *other* operand may have changed in a
1890  // previous iteration.
1891  // Example:
1892  // icmp slt %a, %b
1893  // %b.0 = ssa.copy(%b)
1894  // false branch:
1895  // icmp slt %c, %b.0
1896 
1897  // %c and %a may start out equal, and thus, the code below will say the second
1898  // %icmp is false. c may become equal to something else, and in that case the
1899  // %second icmp *must* be reexamined, but would not if only the renamed
1900  // %operands are considered users of the icmp.
1901 
1902  // *Currently* we only check one level of comparisons back, and only mark one
1903  // level back as touched when changes happen. If you modify this code to look
1904  // back farther through comparisons, you *must* mark the appropriate
1905  // comparisons as users in PredicateInfo.cpp, or you will cause bugs. See if
1906  // we know something just from the operands themselves
1907 
1908  // See if our operands have predicate info, so that we may be able to derive
1909  // something from a previous comparison.
1910  for (const auto &Op : CI->operands()) {
1911  auto *PI = PredInfo->getPredicateInfoFor(Op);
1912  if (const auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI)) {
1913  if (PI == LastPredInfo)
1914  continue;
1915  LastPredInfo = PI;
1916  // In phi of ops cases, we may have predicate info that we are evaluating
1917  // in a different context.
1918  if (!DT->dominates(PBranch->To, getBlockForValue(I)))
1919  continue;
1920  // TODO: Along the false edge, we may know more things too, like
1921  // icmp of
1922  // same operands is false.
1923  // TODO: We only handle actual comparison conditions below, not
1924  // and/or.
1925  auto *BranchCond = dyn_cast<CmpInst>(PBranch->Condition);
1926  if (!BranchCond)
1927  continue;
1928  auto *BranchOp0 = lookupOperandLeader(BranchCond->getOperand(0));
1929  auto *BranchOp1 = lookupOperandLeader(BranchCond->getOperand(1));
1930  auto BranchPredicate = BranchCond->getPredicate();
1931  if (shouldSwapOperands(BranchOp0, BranchOp1)) {
1932  std::swap(BranchOp0, BranchOp1);
1933  BranchPredicate = BranchCond->getSwappedPredicate();
1934  }
1935  if (BranchOp0 == Op0 && BranchOp1 == Op1) {
1936  if (PBranch->TrueEdge) {
1937  // If we know the previous predicate is true and we are in the true
1938  // edge then we may be implied true or false.
1939  if (CmpInst::isImpliedTrueByMatchingCmp(BranchPredicate,
1940  OurPredicate)) {
1941  addPredicateUsers(PI, I);
1942  return createConstantExpression(
1943  ConstantInt::getTrue(CI->getType()));
1944  }
1945 
1946  if (CmpInst::isImpliedFalseByMatchingCmp(BranchPredicate,
1947  OurPredicate)) {
1948  addPredicateUsers(PI, I);
1949  return createConstantExpression(
1951  }
1952  } else {
1953  // Just handle the ne and eq cases, where if we have the same
1954  // operands, we may know something.
1955  if (BranchPredicate == OurPredicate) {
1956  addPredicateUsers(PI, I);
1957  // Same predicate, same ops,we know it was false, so this is false.
1958  return createConstantExpression(
1960  } else if (BranchPredicate ==
1961  CmpInst::getInversePredicate(OurPredicate)) {
1962  addPredicateUsers(PI, I);
1963  // Inverse predicate, we know the other was false, so this is true.
1964  return createConstantExpression(
1965  ConstantInt::getTrue(CI->getType()));
1966  }
1967  }
1968  }
1969  }
1970  }
1971  // Create expression will take care of simplifyCmpInst
1972  return createExpression(I);
1973 }
1974 
1975 // Substitute and symbolize the value before value numbering.
1976 const Expression *
1977 NewGVN::performSymbolicEvaluation(Value *V,
1978  SmallPtrSetImpl<Value *> &Visited) const {
1979  const Expression *E = nullptr;
1980  if (auto *C = dyn_cast<Constant>(V))
1981  E = createConstantExpression(C);
1982  else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1983  E = createVariableExpression(V);
1984  } else {
1985  // TODO: memory intrinsics.
1986  // TODO: Some day, we should do the forward propagation and reassociation
1987  // parts of the algorithm.
1988  auto *I = cast<Instruction>(V);
1989  switch (I->getOpcode()) {
1990  case Instruction::ExtractValue:
1991  case Instruction::InsertValue:
1992  E = performSymbolicAggrValueEvaluation(I);
1993  break;
1994  case Instruction::PHI: {
1996  auto *PN = cast<PHINode>(I);
1997  for (unsigned i = 0; i < PN->getNumOperands(); ++i)
1998  Ops.push_back({PN->getIncomingValue(i), PN->getIncomingBlock(i)});
1999  // Sort to ensure the invariant createPHIExpression requires is met.
2000  sortPHIOps(Ops);
2001  E = performSymbolicPHIEvaluation(Ops, I, getBlockForValue(I));
2002  } break;
2003  case Instruction::Call:
2004  E = performSymbolicCallEvaluation(I);
2005  break;
2006  case Instruction::Store:
2007  E = performSymbolicStoreEvaluation(I);
2008  break;
2009  case Instruction::Load:
2010  E = performSymbolicLoadEvaluation(I);
2011  break;
2012  case Instruction::BitCast:
2013  E = createExpression(I);
2014  break;
2015  case Instruction::ICmp:
2016  case Instruction::FCmp:
2017  E = performSymbolicCmpEvaluation(I);
2018  break;
2019  case Instruction::Add:
2020  case Instruction::FAdd:
2021  case Instruction::Sub:
2022  case Instruction::FSub:
2023  case Instruction::Mul:
2024  case Instruction::FMul:
2025  case Instruction::UDiv:
2026  case Instruction::SDiv:
2027  case Instruction::FDiv:
2028  case Instruction::URem:
2029  case Instruction::SRem:
2030  case Instruction::FRem:
2031  case Instruction::Shl:
2032  case Instruction::LShr:
2033  case Instruction::AShr:
2034  case Instruction::And:
2035  case Instruction::Or:
2036  case Instruction::Xor:
2037  case Instruction::Trunc:
2038  case Instruction::ZExt:
2039  case Instruction::SExt:
2040  case Instruction::FPToUI:
2041  case Instruction::FPToSI:
2042  case Instruction::UIToFP:
2043  case Instruction::SIToFP:
2044  case Instruction::FPTrunc:
2045  case Instruction::FPExt:
2046  case Instruction::PtrToInt:
2047  case Instruction::IntToPtr:
2048  case Instruction::Select:
2049  case Instruction::ExtractElement:
2050  case Instruction::InsertElement:
2051  case Instruction::ShuffleVector:
2052  case Instruction::GetElementPtr:
2053  E = createExpression(I);
2054  break;
2055  default:
2056  return nullptr;
2057  }
2058  }
2059  return E;
2060 }
2061 
2062 // Look up a container in a map, and then call a function for each thing in the
2063 // found container.
2064 template <typename Map, typename KeyType, typename Func>
2065 void NewGVN::for_each_found(Map &M, const KeyType &Key, Func F) {
2066  const auto Result = M.find_as(Key);
2067  if (Result != M.end())
2068  for (typename Map::mapped_type::value_type Mapped : Result->second)
2069  F(Mapped);
2070 }
2071 
2072 // Look up a container of values/instructions in a map, and touch all the
2073 // instructions in the container. Then erase value from the map.
2074 template <typename Map, typename KeyType>
2075 void NewGVN::touchAndErase(Map &M, const KeyType &Key) {
2076  const auto Result = M.find_as(Key);
2077  if (Result != M.end()) {
2078  for (const typename Map::mapped_type::value_type Mapped : Result->second)
2079  TouchedInstructions.set(InstrToDFSNum(Mapped));
2080  M.erase(Result);
2081  }
2082 }
2083 
2084 void NewGVN::addAdditionalUsers(Value *To, Value *User) const {
2085  assert(User && To != User);
2086  if (isa<Instruction>(To))
2087  AdditionalUsers[To].insert(User);
2088 }
2089 
2090 void NewGVN::markUsersTouched(Value *V) {
2091  // Now mark the users as touched.
2092  for (auto *User : V->users()) {
2093  assert(isa<Instruction>(User) && "Use of value not within an instruction?");
2094  TouchedInstructions.set(InstrToDFSNum(User));
2095  }
2096  touchAndErase(AdditionalUsers, V);
2097 }
2098 
2099 void NewGVN::addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const {
2100  LLVM_DEBUG(dbgs() << "Adding memory user " << *U << " to " << *To << "\n");
2101  MemoryToUsers[To].insert(U);
2102 }
2103 
2104 void NewGVN::markMemoryDefTouched(const MemoryAccess *MA) {
2105  TouchedInstructions.set(MemoryToDFSNum(MA));
2106 }
2107 
2108 void NewGVN::markMemoryUsersTouched(const MemoryAccess *MA) {
2109  if (isa<MemoryUse>(MA))
2110  return;
2111  for (auto U : MA->users())
2112  TouchedInstructions.set(MemoryToDFSNum(U));
2113  touchAndErase(MemoryToUsers, MA);
2114 }
2115 
2116 // Add I to the set of users of a given predicate.
2117 void NewGVN::addPredicateUsers(const PredicateBase *PB, Instruction *I) const {
2118  // Don't add temporary instructions to the user lists.
2119  if (AllTempInstructions.count(I))
2120  return;
2121 
2122  if (auto *PBranch = dyn_cast<PredicateBranch>(PB))
2123  PredicateToUsers[PBranch->Condition].insert(I);
2124  else if (auto *PAssume = dyn_cast<PredicateBranch>(PB))
2125  PredicateToUsers[PAssume->Condition].insert(I);
2126 }
2127 
2128 // Touch all the predicates that depend on this instruction.
2129 void NewGVN::markPredicateUsersTouched(Instruction *I) {
2130  touchAndErase(PredicateToUsers, I);
2131 }
2132 
2133 // Mark users affected by a memory leader change.
2134 void NewGVN::markMemoryLeaderChangeTouched(CongruenceClass *CC) {
2135  for (auto M : CC->memory())
2136  markMemoryDefTouched(M);
2137 }
2138 
2139 // Touch the instructions that need to be updated after a congruence class has a
2140 // leader change, and mark changed values.
2141 void NewGVN::markValueLeaderChangeTouched(CongruenceClass *CC) {
2142  for (auto M : *CC) {
2143  if (auto *I = dyn_cast<Instruction>(M))
2144  TouchedInstructions.set(InstrToDFSNum(I));
2145  LeaderChanges.insert(M);
2146  }
2147 }
2148 
2149 // Give a range of things that have instruction DFS numbers, this will return
2150 // the member of the range with the smallest dfs number.
2151 template <class T, class Range>
2152 T *NewGVN::getMinDFSOfRange(const Range &R) const {
2153  std::pair<T *, unsigned> MinDFS = {nullptr, ~0U};
2154  for (const auto X : R) {
2155  auto DFSNum = InstrToDFSNum(X);
2156  if (DFSNum < MinDFS.second)
2157  MinDFS = {X, DFSNum};
2158  }
2159  return MinDFS.first;
2160 }
2161 
2162 // This function returns the MemoryAccess that should be the next leader of
2163 // congruence class CC, under the assumption that the current leader is going to
2164 // disappear.
2165 const MemoryAccess *NewGVN::getNextMemoryLeader(CongruenceClass *CC) const {
2166  // TODO: If this ends up to slow, we can maintain a next memory leader like we
2167  // do for regular leaders.
2168  // Make sure there will be a leader to find.
2169  assert(!CC->definesNoMemory() && "Can't get next leader if there is none");
2170  if (CC->getStoreCount() > 0) {
2171  if (auto *NL = dyn_cast_or_null<StoreInst>(CC->getNextLeader().first))
2172  return getMemoryAccess(NL);
2173  // Find the store with the minimum DFS number.
2174  auto *V = getMinDFSOfRange<Value>(make_filter_range(
2175  *CC, [&](const Value *V) { return isa<StoreInst>(V); }));
2176  return getMemoryAccess(cast<StoreInst>(V));
2177  }
2178  assert(CC->getStoreCount() == 0);
2179 
2180  // Given our assertion, hitting this part must mean
2181  // !OldClass->memory_empty()
2182  if (CC->memory_size() == 1)
2183  return *CC->memory_begin();
2184  return getMinDFSOfRange<const MemoryPhi>(CC->memory());
2185 }
2186 
2187 // This function returns the next value leader of a congruence class, under the
2188 // assumption that the current leader is going away. This should end up being
2189 // the next most dominating member.
2190 Value *NewGVN::getNextValueLeader(CongruenceClass *CC) const {
2191  // We don't need to sort members if there is only 1, and we don't care about
2192  // sorting the TOP class because everything either gets out of it or is
2193  // unreachable.
2194 
2195  if (CC->size() == 1 || CC == TOPClass) {
2196  return *(CC->begin());
2197  } else if (CC->getNextLeader().first) {
2198  ++NumGVNAvoidedSortedLeaderChanges;
2199  return CC->getNextLeader().first;
2200  } else {
2201  ++NumGVNSortedLeaderChanges;
2202  // NOTE: If this ends up to slow, we can maintain a dual structure for
2203  // member testing/insertion, or keep things mostly sorted, and sort only
2204  // here, or use SparseBitVector or ....
2205  return getMinDFSOfRange<Value>(*CC);
2206  }
2207 }
2208 
2209 // Move a MemoryAccess, currently in OldClass, to NewClass, including updates to
2210 // the memory members, etc for the move.
2211 //
2212 // The invariants of this function are:
2213 //
2214 // - I must be moving to NewClass from OldClass
2215 // - The StoreCount of OldClass and NewClass is expected to have been updated
2216 // for I already if it is a store.
2217 // - The OldClass memory leader has not been updated yet if I was the leader.
2218 void NewGVN::moveMemoryToNewCongruenceClass(Instruction *I,
2219  MemoryAccess *InstMA,
2220  CongruenceClass *OldClass,
2221  CongruenceClass *NewClass) {
2222  // If the leader is I, and we had a representative MemoryAccess, it should
2223  // be the MemoryAccess of OldClass.
2224  assert((!InstMA || !OldClass->getMemoryLeader() ||
2225  OldClass->getLeader() != I ||
2226  MemoryAccessToClass.lookup(OldClass->getMemoryLeader()) ==
2227  MemoryAccessToClass.lookup(InstMA)) &&
2228  "Representative MemoryAccess mismatch");
2229  // First, see what happens to the new class
2230  if (!NewClass->getMemoryLeader()) {
2231  // Should be a new class, or a store becoming a leader of a new class.
2232  assert(NewClass->size() == 1 ||
2233  (isa<StoreInst>(I) && NewClass->getStoreCount() == 1));
2234  NewClass->setMemoryLeader(InstMA);
2235  // Mark it touched if we didn't just create a singleton
2236  LLVM_DEBUG(dbgs() << "Memory class leader change for class "
2237  << NewClass->getID()
2238  << " due to new memory instruction becoming leader\n");
2239  markMemoryLeaderChangeTouched(NewClass);
2240  }
2241  setMemoryClass(InstMA, NewClass);
2242  // Now, fixup the old class if necessary
2243  if (OldClass->getMemoryLeader() == InstMA) {
2244  if (!OldClass->definesNoMemory()) {
2245  OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
2246  LLVM_DEBUG(dbgs() << "Memory class leader change for class "
2247  << OldClass->getID() << " to "
2248  << *OldClass->getMemoryLeader()
2249  << " due to removal of old leader " << *InstMA << "\n");
2250  markMemoryLeaderChangeTouched(OldClass);
2251  } else
2252  OldClass->setMemoryLeader(nullptr);
2253  }
2254 }
2255 
2256 // Move a value, currently in OldClass, to be part of NewClass
2257 // Update OldClass and NewClass for the move (including changing leaders, etc).
2258 void NewGVN::moveValueToNewCongruenceClass(Instruction *I, const Expression *E,
2259  CongruenceClass *OldClass,
2260  CongruenceClass *NewClass) {
2261  if (I == OldClass->getNextLeader().first)
2262  OldClass->resetNextLeader();
2263 
2264  OldClass->erase(I);
2265  NewClass->insert(I);
2266 
2267  if (NewClass->getLeader() != I)
2268  NewClass->addPossibleNextLeader({I, InstrToDFSNum(I)});
2269  // Handle our special casing of stores.
2270  if (auto *SI = dyn_cast<StoreInst>(I)) {
2271  OldClass->decStoreCount();
2272  // Okay, so when do we want to make a store a leader of a class?
2273  // If we have a store defined by an earlier load, we want the earlier load
2274  // to lead the class.
2275  // If we have a store defined by something else, we want the store to lead
2276  // the class so everything else gets the "something else" as a value.
2277  // If we have a store as the single member of the class, we want the store
2278  // as the leader
2279  if (NewClass->getStoreCount() == 0 && !NewClass->getStoredValue()) {
2280  // If it's a store expression we are using, it means we are not equivalent
2281  // to something earlier.
2282  if (auto *SE = dyn_cast<StoreExpression>(E)) {
2283  NewClass->setStoredValue(SE->getStoredValue());
2284  markValueLeaderChangeTouched(NewClass);
2285  // Shift the new class leader to be the store
2286  LLVM_DEBUG(dbgs() << "Changing leader of congruence class "
2287  << NewClass->getID() << " from "
2288  << *NewClass->getLeader() << " to " << *SI
2289  << " because store joined class\n");
2290  // If we changed the leader, we have to mark it changed because we don't
2291  // know what it will do to symbolic evaluation.
2292  NewClass->setLeader(SI);
2293  }
2294  // We rely on the code below handling the MemoryAccess change.
2295  }
2296  NewClass->incStoreCount();
2297  }
2298  // True if there is no memory instructions left in a class that had memory
2299  // instructions before.
2300 
2301  // If it's not a memory use, set the MemoryAccess equivalence
2302  auto *InstMA = dyn_cast_or_null<MemoryDef>(getMemoryAccess(I));
2303  if (InstMA)
2304  moveMemoryToNewCongruenceClass(I, InstMA, OldClass, NewClass);
2305  ValueToClass[I] = NewClass;
2306  // See if we destroyed the class or need to swap leaders.
2307  if (OldClass->empty() && OldClass != TOPClass) {
2308  if (OldClass->getDefiningExpr()) {
2309  LLVM_DEBUG(dbgs() << "Erasing expression " << *OldClass->getDefiningExpr()
2310  << " from table\n");
2311  // We erase it as an exact expression to make sure we don't just erase an
2312  // equivalent one.
2313  auto Iter = ExpressionToClass.find_as(
2314  ExactEqualsExpression(*OldClass->getDefiningExpr()));
2315  if (Iter != ExpressionToClass.end())
2316  ExpressionToClass.erase(Iter);
2317 #ifdef EXPENSIVE_CHECKS
2318  assert(
2319  (*OldClass->getDefiningExpr() != *E || ExpressionToClass.lookup(E)) &&
2320  "We erased the expression we just inserted, which should not happen");
2321 #endif
2322  }
2323  } else if (OldClass->getLeader() == I) {
2324  // When the leader changes, the value numbering of
2325  // everything may change due to symbolization changes, so we need to
2326  // reprocess.
2327  LLVM_DEBUG(dbgs() << "Value class leader change for class "
2328  << OldClass->getID() << "\n");
2329  ++NumGVNLeaderChanges;
2330  // Destroy the stored value if there are no more stores to represent it.
2331  // Note that this is basically clean up for the expression removal that
2332  // happens below. If we remove stores from a class, we may leave it as a
2333  // class of equivalent memory phis.
2334  if (OldClass->getStoreCount() == 0) {
2335  if (OldClass->getStoredValue())
2336  OldClass->setStoredValue(nullptr);
2337  }
2338  OldClass->setLeader(getNextValueLeader(OldClass));
2339  OldClass->resetNextLeader();
2340  markValueLeaderChangeTouched(OldClass);
2341  }
2342 }
2343 
2344 // For a given expression, mark the phi of ops instructions that could have
2345 // changed as a result.
2346 void NewGVN::markPhiOfOpsChanged(const Expression *E) {
2347  touchAndErase(ExpressionToPhiOfOps, E);
2348 }
2349 
2350 // Perform congruence finding on a given value numbering expression.
2351 void NewGVN::performCongruenceFinding(Instruction *I, const Expression *E) {
2352  // This is guaranteed to return something, since it will at least find
2353  // TOP.
2354 
2355  CongruenceClass *IClass = ValueToClass.lookup(I);
2356  assert(IClass && "Should have found a IClass");
2357  // Dead classes should have been eliminated from the mapping.
2358  assert(!IClass->isDead() && "Found a dead class");
2359 
2360  CongruenceClass *EClass = nullptr;
2361  if (const auto *VE = dyn_cast<VariableExpression>(E)) {
2362  EClass = ValueToClass.lookup(VE->getVariableValue());
2363  } else if (isa<DeadExpression>(E)) {
2364  EClass = TOPClass;
2365  }
2366  if (!EClass) {
2367  auto lookupResult = ExpressionToClass.insert({E, nullptr});
2368 
2369  // If it's not in the value table, create a new congruence class.
2370  if (lookupResult.second) {
2371  CongruenceClass *NewClass = createCongruenceClass(nullptr, E);
2372  auto place = lookupResult.first;
2373  place->second = NewClass;
2374 
2375  // Constants and variables should always be made the leader.
2376  if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2377  NewClass->setLeader(CE->getConstantValue());
2378  } else if (const auto *SE = dyn_cast<StoreExpression>(E)) {
2379  StoreInst *SI = SE->getStoreInst();
2380  NewClass->setLeader(SI);
2381  NewClass->setStoredValue(SE->getStoredValue());
2382  // The RepMemoryAccess field will be filled in properly by the
2383  // moveValueToNewCongruenceClass call.
2384  } else {
2385  NewClass->setLeader(I);
2386  }
2387  assert(!isa<VariableExpression>(E) &&
2388  "VariableExpression should have been handled already");
2389 
2390  EClass = NewClass;
2391  LLVM_DEBUG(dbgs() << "Created new congruence class for " << *I
2392  << " using expression " << *E << " at "
2393  << NewClass->getID() << " and leader "
2394  << *(NewClass->getLeader()));
2395  if (NewClass->getStoredValue())
2396  LLVM_DEBUG(dbgs() << " and stored value "
2397  << *(NewClass->getStoredValue()));
2398  LLVM_DEBUG(dbgs() << "\n");
2399  } else {
2400  EClass = lookupResult.first->second;
2401  if (isa<ConstantExpression>(E))
2402  assert((isa<Constant>(EClass->getLeader()) ||
2403  (EClass->getStoredValue() &&
2404  isa<Constant>(EClass->getStoredValue()))) &&
2405  "Any class with a constant expression should have a "
2406  "constant leader");
2407 
2408  assert(EClass && "Somehow don't have an eclass");
2409 
2410  assert(!EClass->isDead() && "We accidentally looked up a dead class");
2411  }
2412  }
2413  bool ClassChanged = IClass != EClass;
2414  bool LeaderChanged = LeaderChanges.erase(I);
2415  if (ClassChanged || LeaderChanged) {
2416  LLVM_DEBUG(dbgs() << "New class " << EClass->getID() << " for expression "
2417  << *E << "\n");
2418  if (ClassChanged) {
2419  moveValueToNewCongruenceClass(I, E, IClass, EClass);
2420  markPhiOfOpsChanged(E);
2421  }
2422 
2423  markUsersTouched(I);
2424  if (MemoryAccess *MA = getMemoryAccess(I))
2425  markMemoryUsersTouched(MA);
2426  if (auto *CI = dyn_cast<CmpInst>(I))
2427  markPredicateUsersTouched(CI);
2428  }
2429  // If we changed the class of the store, we want to ensure nothing finds the
2430  // old store expression. In particular, loads do not compare against stored
2431  // value, so they will find old store expressions (and associated class
2432  // mappings) if we leave them in the table.
2433  if (ClassChanged && isa<StoreInst>(I)) {
2434  auto *OldE = ValueToExpression.lookup(I);
2435  // It could just be that the old class died. We don't want to erase it if we
2436  // just moved classes.
2437  if (OldE && isa<StoreExpression>(OldE) && *E != *OldE) {
2438  // Erase this as an exact expression to ensure we don't erase expressions
2439  // equivalent to it.
2440  auto Iter = ExpressionToClass.find_as(ExactEqualsExpression(*OldE));
2441  if (Iter != ExpressionToClass.end())
2442  ExpressionToClass.erase(Iter);
2443  }
2444  }
2445  ValueToExpression[I] = E;
2446 }
2447 
2448 // Process the fact that Edge (from, to) is reachable, including marking
2449 // any newly reachable blocks and instructions for processing.
2450 void NewGVN::updateReachableEdge(BasicBlock *From, BasicBlock *To) {
2451  // Check if the Edge was reachable before.
2452  if (ReachableEdges.insert({From, To}).second) {
2453  // If this block wasn't reachable before, all instructions are touched.
2454  if (ReachableBlocks.insert(To).second) {
2455  LLVM_DEBUG(dbgs() << "Block " << getBlockName(To)
2456  << " marked reachable\n");
2457  const auto &InstRange = BlockInstRange.lookup(To);
2458  TouchedInstructions.set(InstRange.first, InstRange.second);
2459  } else {
2460  LLVM_DEBUG(dbgs() << "Block " << getBlockName(To)
2461  << " was reachable, but new edge {"
2462  << getBlockName(From) << "," << getBlockName(To)
2463  << "} to it found\n");
2464 
2465  // We've made an edge reachable to an existing block, which may
2466  // impact predicates. Otherwise, only mark the phi nodes as touched, as
2467  // they are the only thing that depend on new edges. Anything using their
2468  // values will get propagated to if necessary.
2469  if (MemoryAccess *MemPhi = getMemoryAccess(To))
2470  TouchedInstructions.set(InstrToDFSNum(MemPhi));
2471 
2472  // FIXME: We should just add a union op on a Bitvector and
2473  // SparseBitVector. We can do it word by word faster than we are doing it
2474  // here.
2475  for (auto InstNum : RevisitOnReachabilityChange[To])
2476  TouchedInstructions.set(InstNum);
2477  }
2478  }
2479 }
2480 
2481 // Given a predicate condition (from a switch, cmp, or whatever) and a block,
2482 // see if we know some constant value for it already.
2483 Value *NewGVN::findConditionEquivalence(Value *Cond) const {
2484  auto Result = lookupOperandLeader(Cond);
2485  return isa<Constant>(Result) ? Result : nullptr;
2486 }
2487 
2488 // Process the outgoing edges of a block for reachability.
2489 void NewGVN::processOutgoingEdges(Instruction *TI, BasicBlock *B) {
2490  // Evaluate reachability of terminator instruction.
2491  BranchInst *BR;
2492  if ((BR = dyn_cast<BranchInst>(TI)) && BR->isConditional()) {
2493  Value *Cond = BR->getCondition();
2494  Value *CondEvaluated = findConditionEquivalence(Cond);
2495  if (!CondEvaluated) {
2496  if (auto *I = dyn_cast<Instruction>(Cond)) {
2497  const Expression *E = createExpression(I);
2498  if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2499  CondEvaluated = CE->getConstantValue();
2500  }
2501  } else if (isa<ConstantInt>(Cond)) {
2502  CondEvaluated = Cond;
2503  }
2504  }
2505  ConstantInt *CI;
2506  BasicBlock *TrueSucc = BR->getSuccessor(0);
2507  BasicBlock *FalseSucc = BR->getSuccessor(1);
2508  if (CondEvaluated && (CI = dyn_cast<ConstantInt>(CondEvaluated))) {
2509  if (CI->isOne()) {
2510  LLVM_DEBUG(dbgs() << "Condition for Terminator " << *TI
2511  << " evaluated to true\n");
2512  updateReachableEdge(B, TrueSucc);
2513  } else if (CI->isZero()) {
2514  LLVM_DEBUG(dbgs() << "Condition for Terminator " << *TI
2515  << " evaluated to false\n");
2516  updateReachableEdge(B, FalseSucc);
2517  }
2518  } else {
2519  updateReachableEdge(B, TrueSucc);
2520  updateReachableEdge(B, FalseSucc);
2521  }
2522  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
2523  // For switches, propagate the case values into the case
2524  // destinations.
2525 
2526  // Remember how many outgoing edges there are to every successor.
2528 
2529  Value *SwitchCond = SI->getCondition();
2530  Value *CondEvaluated = findConditionEquivalence(SwitchCond);
2531  // See if we were able to turn this switch statement into a constant.
2532  if (CondEvaluated && isa<ConstantInt>(CondEvaluated)) {
2533  auto *CondVal = cast<ConstantInt>(CondEvaluated);
2534  // We should be able to get case value for this.
2535  auto Case = *SI->findCaseValue(CondVal);
2536  if (Case.getCaseSuccessor() == SI->getDefaultDest()) {
2537  // We proved the value is outside of the range of the case.
2538  // We can't do anything other than mark the default dest as reachable,
2539  // and go home.
2540  updateReachableEdge(B, SI->getDefaultDest());
2541  return;
2542  }
2543  // Now get where it goes and mark it reachable.
2544  BasicBlock *TargetBlock = Case.getCaseSuccessor();
2545  updateReachableEdge(B, TargetBlock);
2546  } else {
2547  for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) {
2548  BasicBlock *TargetBlock = SI->getSuccessor(i);
2549  ++SwitchEdges[TargetBlock];
2550  updateReachableEdge(B, TargetBlock);
2551  }
2552  }
2553  } else {
2554  // Otherwise this is either unconditional, or a type we have no
2555  // idea about. Just mark successors as reachable.
2556  for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
2557  BasicBlock *TargetBlock = TI->getSuccessor(i);
2558  updateReachableEdge(B, TargetBlock);
2559  }
2560 
2561  // This also may be a memory defining terminator, in which case, set it
2562  // equivalent only to itself.
2563  //
2564  auto *MA = getMemoryAccess(TI);
2565  if (MA && !isa<MemoryUse>(MA)) {
2566  auto *CC = ensureLeaderOfMemoryClass(MA);
2567  if (setMemoryClass(MA, CC))
2568  markMemoryUsersTouched(MA);
2569  }
2570  }
2571 }
2572 
2573 // Remove the PHI of Ops PHI for I
2574 void NewGVN::removePhiOfOps(Instruction *I, PHINode *PHITemp) {
2575  InstrDFS.erase(PHITemp);
2576  // It's still a temp instruction. We keep it in the array so it gets erased.
2577  // However, it's no longer used by I, or in the block
2578  TempToBlock.erase(PHITemp);
2579  RealToTemp.erase(I);
2580  // We don't remove the users from the phi node uses. This wastes a little
2581  // time, but such is life. We could use two sets to track which were there
2582  // are the start of NewGVN, and which were added, but right nowt he cost of
2583  // tracking is more than the cost of checking for more phi of ops.
2584 }
2585 
2586 // Add PHI Op in BB as a PHI of operations version of ExistingValue.
2587 void NewGVN::addPhiOfOps(PHINode *Op, BasicBlock *BB,
2588  Instruction *ExistingValue) {
2589  InstrDFS[Op] = InstrToDFSNum(ExistingValue);
2590  AllTempInstructions.insert(Op);
2591  TempToBlock[Op] = BB;
2592  RealToTemp[ExistingValue] = Op;
2593  // Add all users to phi node use, as they are now uses of the phi of ops phis
2594  // and may themselves be phi of ops.
2595  for (auto *U : ExistingValue->users())
2596  if (auto *UI = dyn_cast<Instruction>(U))
2597  PHINodeUses.insert(UI);
2598 }
2599 
2600 static bool okayForPHIOfOps(const Instruction *I) {
2601  if (!EnablePhiOfOps)
2602  return false;
2603  return isa<BinaryOperator>(I) || isa<SelectInst>(I) || isa<CmpInst>(I) ||
2604  isa<LoadInst>(I);
2605 }
2606 
2607 bool NewGVN::OpIsSafeForPHIOfOpsHelper(
2608  Value *V, const BasicBlock *PHIBlock,
2610  SmallVectorImpl<Instruction *> &Worklist) {
2611 
2612  if (!isa<Instruction>(V))
2613  return true;
2614  auto OISIt = OpSafeForPHIOfOps.find(V);
2615  if (OISIt != OpSafeForPHIOfOps.end())
2616  return OISIt->second;
2617 
2618  // Keep walking until we either dominate the phi block, or hit a phi, or run
2619  // out of things to check.
2620  if (DT->properlyDominates(getBlockForValue(V), PHIBlock)) {
2621  OpSafeForPHIOfOps.insert({V, true});
2622  return true;
2623  }
2624  // PHI in the same block.
2625  if (isa<PHINode>(V) && getBlockForValue(V) == PHIBlock) {
2626  OpSafeForPHIOfOps.insert({V, false});
2627  return false;
2628  }
2629 
2630  auto *OrigI = cast<Instruction>(V);
2631  for (auto *Op : OrigI->operand_values()) {
2632  if (!isa<Instruction>(Op))
2633  continue;
2634  // Stop now if we find an unsafe operand.
2635  auto OISIt = OpSafeForPHIOfOps.find(OrigI);
2636  if (OISIt != OpSafeForPHIOfOps.end()) {
2637  if (!OISIt->second) {
2638  OpSafeForPHIOfOps.insert({V, false});
2639  return false;
2640  }
2641  continue;
2642  }
2643  if (!Visited.insert(Op).second)
2644  continue;
2645  Worklist.push_back(cast<Instruction>(Op));
2646  }
2647  return true;
2648 }
2649 
2650 // Return true if this operand will be safe to use for phi of ops.
2651 //
2652 // The reason some operands are unsafe is that we are not trying to recursively
2653 // translate everything back through phi nodes. We actually expect some lookups
2654 // of expressions to fail. In particular, a lookup where the expression cannot
2655 // exist in the predecessor. This is true even if the expression, as shown, can
2656 // be determined to be constant.
2657 bool NewGVN::OpIsSafeForPHIOfOps(Value *V, const BasicBlock *PHIBlock,
2658  SmallPtrSetImpl<const Value *> &Visited) {
2660  if (!OpIsSafeForPHIOfOpsHelper(V, PHIBlock, Visited, Worklist))
2661  return false;
2662  while (!Worklist.empty()) {
2663  auto *I = Worklist.pop_back_val();
2664  if (!OpIsSafeForPHIOfOpsHelper(I, PHIBlock, Visited, Worklist))
2665  return false;
2666  }
2667  OpSafeForPHIOfOps.insert({V, true});
2668  return true;
2669 }
2670 
2671 // Try to find a leader for instruction TransInst, which is a phi translated
2672 // version of something in our original program. Visited is used to ensure we
2673 // don't infinite loop during translations of cycles. OrigInst is the
2674 // instruction in the original program, and PredBB is the predecessor we
2675 // translated it through.
2676 Value *NewGVN::findLeaderForInst(Instruction *TransInst,
2677  SmallPtrSetImpl<Value *> &Visited,
2678  MemoryAccess *MemAccess, Instruction *OrigInst,
2679  BasicBlock *PredBB) {
2680  unsigned IDFSNum = InstrToDFSNum(OrigInst);
2681  // Make sure it's marked as a temporary instruction.
2682  AllTempInstructions.insert(TransInst);
2683  // and make sure anything that tries to add it's DFS number is
2684  // redirected to the instruction we are making a phi of ops
2685  // for.
2686  TempToBlock.insert({TransInst, PredBB});
2687  InstrDFS.insert({TransInst, IDFSNum});
2688 
2689  const Expression *E = performSymbolicEvaluation(TransInst, Visited);
2690  InstrDFS.erase(TransInst);
2691  AllTempInstructions.erase(TransInst);
2692  TempToBlock.erase(TransInst);
2693  if (MemAccess)
2694  TempToMemory.erase(TransInst);
2695  if (!E)
2696  return nullptr;
2697  auto *FoundVal = findPHIOfOpsLeader(E, OrigInst, PredBB);
2698  if (!FoundVal) {
2699  ExpressionToPhiOfOps[E].insert(OrigInst);
2700  LLVM_DEBUG(dbgs() << "Cannot find phi of ops operand for " << *TransInst
2701  << " in block " << getBlockName(PredBB) << "\n");
2702  return nullptr;
2703  }
2704  if (auto *SI = dyn_cast<StoreInst>(FoundVal))
2705  FoundVal = SI->getValueOperand();
2706  return FoundVal;
2707 }
2708 
2709 // When we see an instruction that is an op of phis, generate the equivalent phi
2710 // of ops form.
2711 const Expression *
2712 NewGVN::makePossiblePHIOfOps(Instruction *I,
2713  SmallPtrSetImpl<Value *> &Visited) {
2714  if (!okayForPHIOfOps(I))
2715  return nullptr;
2716 
2717  if (!Visited.insert(I).second)
2718  return nullptr;
2719  // For now, we require the instruction be cycle free because we don't
2720  // *always* create a phi of ops for instructions that could be done as phi
2721  // of ops, we only do it if we think it is useful. If we did do it all the
2722  // time, we could remove the cycle free check.
2723  if (!isCycleFree(I))
2724  return nullptr;
2725 
2726  SmallPtrSet<const Value *, 8> ProcessedPHIs;
2727  // TODO: We don't do phi translation on memory accesses because it's
2728  // complicated. For a load, we'd need to be able to simulate a new memoryuse,
2729  // which we don't have a good way of doing ATM.
2730  auto *MemAccess = getMemoryAccess(I);
2731  // If the memory operation is defined by a memory operation this block that
2732  // isn't a MemoryPhi, transforming the pointer backwards through a scalar phi
2733  // can't help, as it would still be killed by that memory operation.
2734  if (MemAccess && !isa<MemoryPhi>(MemAccess->getDefiningAccess()) &&
2735  MemAccess->getDefiningAccess()->getBlock() == I->getParent())
2736  return nullptr;
2737 
2738  // Convert op of phis to phi of ops
2739  SmallPtrSet<const Value *, 10> VisitedOps;
2741  BasicBlock *SamePHIBlock = nullptr;
2742  PHINode *OpPHI = nullptr;
2743  if (!DebugCounter::shouldExecute(PHIOfOpsCounter))
2744  return nullptr;
2745  for (auto *Op : Ops) {
2746  if (!isa<PHINode>(Op)) {
2747  auto *ValuePHI = RealToTemp.lookup(Op);
2748  if (!ValuePHI)
2749  continue;
2750  LLVM_DEBUG(dbgs() << "Found possible dependent phi of ops\n");
2751  Op = ValuePHI;
2752  }
2753  OpPHI = cast<PHINode>(Op);
2754  if (!SamePHIBlock) {
2755  SamePHIBlock = getBlockForValue(OpPHI);
2756  } else if (SamePHIBlock != getBlockForValue(OpPHI)) {
2757  LLVM_DEBUG(
2758  dbgs()
2759  << "PHIs for operands are not all in the same block, aborting\n");
2760  return nullptr;
2761  }
2762  // No point in doing this for one-operand phis.
2763  if (OpPHI->getNumOperands() == 1) {
2764  OpPHI = nullptr;
2765  continue;
2766  }
2767  }
2768 
2769  if (!OpPHI)
2770  return nullptr;
2771 
2772  SmallVector<ValPair, 4> PHIOps;
2774  auto *PHIBlock = getBlockForValue(OpPHI);
2775  RevisitOnReachabilityChange[PHIBlock].reset(InstrToDFSNum(I));
2776  for (unsigned PredNum = 0; PredNum < OpPHI->getNumOperands(); ++PredNum) {
2777  auto *PredBB = OpPHI->getIncomingBlock(PredNum);
2778  Value *FoundVal = nullptr;
2779  SmallPtrSet<Value *, 4> CurrentDeps;
2780  // We could just skip unreachable edges entirely but it's tricky to do
2781  // with rewriting existing phi nodes.
2782  if (ReachableEdges.count({PredBB, PHIBlock})) {
2783  // Clone the instruction, create an expression from it that is
2784  // translated back into the predecessor, and see if we have a leader.
2785  Instruction *ValueOp = I->clone();
2786  if (MemAccess)
2787  TempToMemory.insert({ValueOp, MemAccess});
2788  bool SafeForPHIOfOps = true;
2789  VisitedOps.clear();
2790  for (auto &Op : ValueOp->operands()) {
2791  auto *OrigOp = &*Op;
2792  // When these operand changes, it could change whether there is a
2793  // leader for us or not, so we have to add additional users.
2794  if (isa<PHINode>(Op)) {
2795  Op = Op->DoPHITranslation(PHIBlock, PredBB);
2796  if (Op != OrigOp && Op != I)
2797  CurrentDeps.insert(Op);
2798  } else if (auto *ValuePHI = RealToTemp.lookup(Op)) {
2799  if (getBlockForValue(ValuePHI) == PHIBlock)
2800  Op = ValuePHI->getIncomingValueForBlock(PredBB);
2801  }
2802  // If we phi-translated the op, it must be safe.
2803  SafeForPHIOfOps =
2804  SafeForPHIOfOps &&
2805  (Op != OrigOp || OpIsSafeForPHIOfOps(Op, PHIBlock, VisitedOps));
2806  }
2807  // FIXME: For those things that are not safe we could generate
2808  // expressions all the way down, and see if this comes out to a
2809  // constant. For anything where that is true, and unsafe, we should
2810  // have made a phi-of-ops (or value numbered it equivalent to something)
2811  // for the pieces already.
2812  FoundVal = !SafeForPHIOfOps ? nullptr
2813  : findLeaderForInst(ValueOp, Visited,
2814  MemAccess, I, PredBB);
2815  ValueOp->deleteValue();
2816  if (!FoundVal) {
2817  // We failed to find a leader for the current ValueOp, but this might
2818  // change in case of the translated operands change.
2819  if (SafeForPHIOfOps)
2820  for (auto Dep : CurrentDeps)
2821  addAdditionalUsers(Dep, I);
2822 
2823  return nullptr;
2824  }
2825  Deps.insert(CurrentDeps.begin(), CurrentDeps.end());
2826  } else {
2827  LLVM_DEBUG(dbgs() << "Skipping phi of ops operand for incoming block "
2828  << getBlockName(PredBB)
2829  << " because the block is unreachable\n");
2830  FoundVal = UndefValue::get(I->getType());
2831  RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2832  }
2833 
2834  PHIOps.push_back({FoundVal, PredBB});
2835  LLVM_DEBUG(dbgs() << "Found phi of ops operand " << *FoundVal << " in "
2836  << getBlockName(PredBB) << "\n");
2837  }
2838  for (auto Dep : Deps)
2839  addAdditionalUsers(Dep, I);
2840  sortPHIOps(PHIOps);
2841  auto *E = performSymbolicPHIEvaluation(PHIOps, I, PHIBlock);
2842  if (isa<ConstantExpression>(E) || isa<VariableExpression>(E)) {
2843  LLVM_DEBUG(
2844  dbgs()
2845  << "Not creating real PHI of ops because it simplified to existing "
2846  "value or constant\n");
2847  return E;
2848  }
2849  auto *ValuePHI = RealToTemp.lookup(I);
2850  bool NewPHI = false;
2851  if (!ValuePHI) {
2852  ValuePHI =
2853  PHINode::Create(I->getType(), OpPHI->getNumOperands(), "phiofops");
2854  addPhiOfOps(ValuePHI, PHIBlock, I);
2855  NewPHI = true;
2856  NumGVNPHIOfOpsCreated++;
2857  }
2858  if (NewPHI) {
2859  for (auto PHIOp : PHIOps)
2860  ValuePHI->addIncoming(PHIOp.first, PHIOp.second);
2861  } else {
2862  TempToBlock[ValuePHI] = PHIBlock;
2863  unsigned int i = 0;
2864  for (auto PHIOp : PHIOps) {
2865  ValuePHI->setIncomingValue(i, PHIOp.first);
2866  ValuePHI->setIncomingBlock(i, PHIOp.second);
2867  ++i;
2868  }
2869  }
2870  RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2871  LLVM_DEBUG(dbgs() << "Created phi of ops " << *ValuePHI << " for " << *I
2872  << "\n");
2873 
2874  return E;
2875 }
2876 
2877 // The algorithm initially places the values of the routine in the TOP
2878 // congruence class. The leader of TOP is the undetermined value `undef`.
2879 // When the algorithm has finished, values still in TOP are unreachable.
2880 void NewGVN::initializeCongruenceClasses(Function &F) {
2881  NextCongruenceNum = 0;
2882 
2883  // Note that even though we use the live on entry def as a representative
2884  // MemoryAccess, it is *not* the same as the actual live on entry def. We
2885  // have no real equivalemnt to undef for MemoryAccesses, and so we really
2886  // should be checking whether the MemoryAccess is top if we want to know if it
2887  // is equivalent to everything. Otherwise, what this really signifies is that
2888  // the access "it reaches all the way back to the beginning of the function"
2889 
2890  // Initialize all other instructions to be in TOP class.
2891  TOPClass = createCongruenceClass(nullptr, nullptr);
2892  TOPClass->setMemoryLeader(MSSA->getLiveOnEntryDef());
2893  // The live on entry def gets put into it's own class
2894  MemoryAccessToClass[MSSA->getLiveOnEntryDef()] =
2895  createMemoryClass(MSSA->getLiveOnEntryDef());
2896 
2897  for (auto DTN : nodes(DT)) {
2898  BasicBlock *BB = DTN->getBlock();
2899  // All MemoryAccesses are equivalent to live on entry to start. They must
2900  // be initialized to something so that initial changes are noticed. For
2901  // the maximal answer, we initialize them all to be the same as
2902  // liveOnEntry.
2903  auto *MemoryBlockDefs = MSSA->getBlockDefs(BB);
2904  if (MemoryBlockDefs)
2905  for (const auto &Def : *MemoryBlockDefs) {
2906  MemoryAccessToClass[&Def] = TOPClass;
2907  auto *MD = dyn_cast<MemoryDef>(&Def);
2908  // Insert the memory phis into the member list.
2909  if (!MD) {
2910  const MemoryPhi *MP = cast<MemoryPhi>(&Def);
2911  TOPClass->memory_insert(MP);
2912  MemoryPhiState.insert({MP, MPS_TOP});
2913  }
2914 
2915  if (MD && isa<StoreInst>(MD->getMemoryInst()))
2916  TOPClass->incStoreCount();
2917  }
2918 
2919  // FIXME: This is trying to discover which instructions are uses of phi
2920  // nodes. We should move this into one of the myriad of places that walk
2921  // all the operands already.
2922  for (auto &I : *BB) {
2923  if (isa<PHINode>(&I))
2924  for (auto *U : I.users())
2925  if (auto *UInst = dyn_cast<Instruction>(U))
2926  if (InstrToDFSNum(UInst) != 0 && okayForPHIOfOps(UInst))
2927  PHINodeUses.insert(UInst);
2928  // Don't insert void terminators into the class. We don't value number
2929  // them, and they just end up sitting in TOP.
2930  if (I.isTerminator() && I.getType()->isVoidTy())
2931  continue;
2932  TOPClass->insert(&I);
2933  ValueToClass[&I] = TOPClass;
2934  }
2935  }
2936 
2937  // Initialize arguments to be in their own unique congruence classes
2938  for (auto &FA : F.args())
2939  createSingletonCongruenceClass(&FA);
2940 }
2941 
2942 void NewGVN::cleanupTables() {
2943  for (unsigned i = 0, e = CongruenceClasses.size(); i != e; ++i) {
2944  LLVM_DEBUG(dbgs() << "Congruence class " << CongruenceClasses[i]->getID()
2945  << " has " << CongruenceClasses[i]->size()
2946  << " members\n");
2947  // Make sure we delete the congruence class (probably worth switching to
2948  // a unique_ptr at some point.
2949  delete CongruenceClasses[i];
2950  CongruenceClasses[i] = nullptr;
2951  }
2952 
2953  // Destroy the value expressions
2954  SmallVector<Instruction *, 8> TempInst(AllTempInstructions.begin(),
2955  AllTempInstructions.end());
2956  AllTempInstructions.clear();
2957 
2958  // We have to drop all references for everything first, so there are no uses
2959  // left as we delete them.
2960  for (auto *I : TempInst) {
2961  I->dropAllReferences();
2962  }
2963 
2964  while (!TempInst.empty()) {
2965  auto *I = TempInst.back();
2966  TempInst.pop_back();
2967  I->deleteValue();
2968  }
2969 
2970  ValueToClass.clear();
2971  ArgRecycler.clear(ExpressionAllocator);
2972  ExpressionAllocator.Reset();
2973  CongruenceClasses.clear();
2974  ExpressionToClass.clear();
2975  ValueToExpression.clear();
2976  RealToTemp.clear();
2977  AdditionalUsers.clear();
2978  ExpressionToPhiOfOps.clear();
2979  TempToBlock.clear();
2980  TempToMemory.clear();
2981  PHINodeUses.clear();
2982  OpSafeForPHIOfOps.clear();
2983  ReachableBlocks.clear();
2984  ReachableEdges.clear();
2985 #ifndef NDEBUG
2986  ProcessedCount.clear();
2987 #endif
2988  InstrDFS.clear();
2989  InstructionsToErase.clear();
2990  DFSToInstr.clear();
2991  BlockInstRange.clear();
2992  TouchedInstructions.clear();
2993  MemoryAccessToClass.clear();
2994  PredicateToUsers.clear();
2995  MemoryToUsers.clear();
2996  RevisitOnReachabilityChange.clear();
2997 }
2998 
2999 // Assign local DFS number mapping to instructions, and leave space for Value
3000 // PHI's.
3001 std::pair<unsigned, unsigned> NewGVN::assignDFSNumbers(BasicBlock *B,
3002  unsigned Start) {
3003  unsigned End = Start;
3004  if (MemoryAccess *MemPhi = getMemoryAccess(B)) {
3005  InstrDFS[MemPhi] = End++;
3006  DFSToInstr.emplace_back(MemPhi);
3007  }
3008 
3009  // Then the real block goes next.
3010  for (auto &I : *B) {
3011  // There's no need to call isInstructionTriviallyDead more than once on
3012  // an instruction. Therefore, once we know that an instruction is dead
3013  // we change its DFS number so that it doesn't get value numbered.
3014  if (isInstructionTriviallyDead(&I, TLI)) {
3015  InstrDFS[&I] = 0;
3016  LLVM_DEBUG(dbgs() << "Skipping trivially dead instruction " << I << "\n");
3017  markInstructionForDeletion(&I);
3018  continue;
3019  }
3020  if (isa<PHINode>(&I))
3021  RevisitOnReachabilityChange[B].set(End);
3022  InstrDFS[&I] = End++;
3023  DFSToInstr.emplace_back(&I);
3024  }
3025 
3026  // All of the range functions taken half-open ranges (open on the end side).
3027  // So we do not subtract one from count, because at this point it is one
3028  // greater than the last instruction.
3029  return std::make_pair(Start, End);
3030 }
3031 
3032 void NewGVN::updateProcessedCount(const Value *V) {
3033 #ifndef NDEBUG
3034  if (ProcessedCount.count(V) == 0) {
3035  ProcessedCount.insert({V, 1});
3036  } else {
3037  ++ProcessedCount[V];
3038  assert(ProcessedCount[V] < 100 &&
3039  "Seem to have processed the same Value a lot");
3040  }
3041 #endif
3042 }
3043 
3044 // Evaluate MemoryPhi nodes symbolically, just like PHI nodes
3045 void NewGVN::valueNumberMemoryPhi(MemoryPhi *MP) {
3046  // If all the arguments are the same, the MemoryPhi has the same value as the
3047  // argument. Filter out unreachable blocks and self phis from our operands.
3048  // TODO: We could do cycle-checking on the memory phis to allow valueizing for
3049  // self-phi checking.
3050  const BasicBlock *PHIBlock = MP->getBlock();
3051  auto Filtered = make_filter_range(MP->operands(), [&](const Use &U) {
3052  return cast<MemoryAccess>(U) != MP &&
3053  !isMemoryAccessTOP(cast<MemoryAccess>(U)) &&
3054  ReachableEdges.count({MP->getIncomingBlock(U), PHIBlock});
3055  });
3056  // If all that is left is nothing, our memoryphi is undef. We keep it as
3057  // InitialClass. Note: The only case this should happen is if we have at
3058  // least one self-argument.
3059  if (Filtered.begin() == Filtered.end()) {
3060  if (setMemoryClass(MP, TOPClass))
3061  markMemoryUsersTouched(MP);
3062  return;
3063  }
3064 
3065  // Transform the remaining operands into operand leaders.
3066  // FIXME: mapped_iterator should have a range version.
3067  auto LookupFunc = [&](const Use &U) {
3068  return lookupMemoryLeader(cast<MemoryAccess>(U));
3069  };
3070  auto MappedBegin = map_iterator(Filtered.begin(), LookupFunc);
3071  auto MappedEnd = map_iterator(Filtered.end(), LookupFunc);
3072 
3073  // and now check if all the elements are equal.
3074  // Sadly, we can't use std::equals since these are random access iterators.
3075  const auto *AllSameValue = *MappedBegin;
3076  ++MappedBegin;
3077  bool AllEqual = std::all_of(
3078  MappedBegin, MappedEnd,
3079  [&AllSameValue](const MemoryAccess *V) { return V == AllSameValue; });
3080 
3081  if (AllEqual)
3082  LLVM_DEBUG(dbgs() << "Memory Phi value numbered to " << *AllSameValue
3083  << "\n");
3084  else
3085  LLVM_DEBUG(dbgs() << "Memory Phi value numbered to itself\n");
3086  // If it's equal to something, it's in that class. Otherwise, it has to be in
3087  // a class where it is the leader (other things may be equivalent to it, but
3088  // it needs to start off in its own class, which means it must have been the
3089  // leader, and it can't have stopped being the leader because it was never
3090  // removed).
3091  CongruenceClass *CC =
3092  AllEqual ? getMemoryClass(AllSameValue) : ensureLeaderOfMemoryClass(MP);
3093  auto OldState = MemoryPhiState.lookup(MP);
3094  assert(OldState != MPS_Invalid && "Invalid memory phi state");
3095  auto NewState = AllEqual ? MPS_Equivalent : MPS_Unique;
3096  MemoryPhiState[MP] = NewState;
3097  if (setMemoryClass(MP, CC) || OldState != NewState)
3098  markMemoryUsersTouched(MP);
3099 }
3100 
3101 // Value number a single instruction, symbolically evaluating, performing
3102 // congruence finding, and updating mappings.
3103 void NewGVN::valueNumberInstruction(Instruction *I) {
3104  LLVM_DEBUG(dbgs() << "Processing instruction " << *I << "\n");
3105  if (!I->isTerminator()) {
3106  const Expression *Symbolized = nullptr;
3107  SmallPtrSet<Value *, 2> Visited;
3108  if (DebugCounter::shouldExecute(VNCounter)) {
3109  Symbolized = performSymbolicEvaluation(I, Visited);
3110  // Make a phi of ops if necessary
3111  if (Symbolized && !isa<ConstantExpression>(Symbolized) &&
3112  !isa<VariableExpression>(Symbolized) && PHINodeUses.count(I)) {
3113  auto *PHIE = makePossiblePHIOfOps(I, Visited);
3114  // If we created a phi of ops, use it.
3115  // If we couldn't create one, make sure we don't leave one lying around
3116  if (PHIE) {
3117  Symbolized = PHIE;
3118  } else if (auto *Op = RealToTemp.lookup(I)) {
3119  removePhiOfOps(I, Op);
3120  }
3121  }
3122  } else {
3123  // Mark the instruction as unused so we don't value number it again.
3124  InstrDFS[I] = 0;
3125  }
3126  // If we couldn't come up with a symbolic expression, use the unknown
3127  // expression
3128  if (Symbolized == nullptr)
3129  Symbolized = createUnknownExpression(I);
3130  performCongruenceFinding(I, Symbolized);
3131  } else {
3132  // Handle terminators that return values. All of them produce values we
3133  // don't currently understand. We don't place non-value producing
3134  // terminators in a class.
3135  if (!I->getType()->isVoidTy()) {
3136  auto *Symbolized = createUnknownExpression(I);
3137  performCongruenceFinding(I, Symbolized);
3138  }
3139  processOutgoingEdges(I, I->getParent());
3140  }
3141 }
3142 
3143 // Check if there is a path, using single or equal argument phi nodes, from
3144 // First to Second.
3145 bool NewGVN::singleReachablePHIPath(
3146  SmallPtrSet<const MemoryAccess *, 8> &Visited, const MemoryAccess *First,
3147  const MemoryAccess *Second) const {
3148  if (First == Second)
3149  return true;
3150  if (MSSA->isLiveOnEntryDef(First))
3151  return false;
3152 
3153  // This is not perfect, but as we're just verifying here, we can live with
3154  // the loss of precision. The real solution would be that of doing strongly
3155  // connected component finding in this routine, and it's probably not worth
3156  // the complexity for the time being. So, we just keep a set of visited
3157  // MemoryAccess and return true when we hit a cycle.
3158  if (Visited.count(First))
3159  return true;
3160  Visited.insert(First);
3161 
3162  const auto *EndDef = First;
3163  for (auto *ChainDef : optimized_def_chain(First)) {
3164  if (ChainDef == Second)
3165  return true;
3166  if (MSSA->isLiveOnEntryDef(ChainDef))
3167  return false;
3168  EndDef = ChainDef;
3169  }
3170  auto *MP = cast<MemoryPhi>(EndDef);
3171  auto ReachableOperandPred = [&](const Use &U) {
3172  return ReachableEdges.count({MP->getIncomingBlock(U), MP->getBlock()});
3173  };
3174  auto FilteredPhiArgs =
3175  make_filter_range(MP->operands(), ReachableOperandPred);
3176  SmallVector<const Value *, 32> OperandList;
3177  llvm::copy(FilteredPhiArgs, std::back_inserter(OperandList));
3178  bool Okay = is_splat(OperandList);
3179  if (Okay)
3180  return singleReachablePHIPath(Visited, cast<MemoryAccess>(OperandList[0]),
3181  Second);
3182  return false;
3183 }
3184 
3185 // Verify the that the memory equivalence table makes sense relative to the
3186 // congruence classes. Note that this checking is not perfect, and is currently
3187 // subject to very rare false negatives. It is only useful for
3188 // testing/debugging.
3189 void NewGVN::verifyMemoryCongruency() const {
3190 #ifndef NDEBUG
3191  // Verify that the memory table equivalence and memory member set match
3192  for (const auto *CC : CongruenceClasses) {
3193  if (CC == TOPClass || CC->isDead())
3194  continue;
3195  if (CC->getStoreCount() != 0) {
3196  assert((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) &&
3197  "Any class with a store as a leader should have a "
3198  "representative stored value");
3199  assert(CC->getMemoryLeader() &&
3200  "Any congruence class with a store should have a "
3201  "representative access");
3202  }
3203 
3204  if (CC->getMemoryLeader())
3205  assert(MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
3206  "Representative MemoryAccess does not appear to be reverse "
3207  "mapped properly");
3208  for (auto M : CC->memory())
3209  assert(MemoryAccessToClass.lookup(M) == CC &&
3210  "Memory member does not appear to be reverse mapped properly");
3211  }
3212 
3213  // Anything equivalent in the MemoryAccess table should be in the same
3214  // congruence class.
3215 
3216  // Filter out the unreachable and trivially dead entries, because they may
3217  // never have been updated if the instructions were not processed.
3218  auto ReachableAccessPred =
3219  [&](const std::pair<const MemoryAccess *, CongruenceClass *> Pair) {
3220  bool Result = ReachableBlocks.count(Pair.first->getBlock());
3221  if (!Result || MSSA->isLiveOnEntryDef(Pair.first) ||
3222  MemoryToDFSNum(Pair.first) == 0)
3223  return false;
3224  if (auto *MemDef = dyn_cast<MemoryDef>(Pair.first))
3225  return !isInstructionTriviallyDead(MemDef->getMemoryInst());
3226 
3227  // We could have phi nodes which operands are all trivially dead,
3228  // so we don't process them.
3229  if (auto *MemPHI = dyn_cast<MemoryPhi>(Pair.first)) {
3230  for (auto &U : MemPHI->incoming_values()) {
3231  if (auto *I = dyn_cast<Instruction>(&*U)) {
3233  return true;
3234  }
3235  }
3236  return false;
3237  }
3238 
3239  return true;
3240  };
3241 
3242  auto Filtered = make_filter_range(MemoryAccessToClass, ReachableAccessPred);
3243  for (auto KV : Filtered) {
3244  if (auto *FirstMUD = dyn_cast<MemoryUseOrDef>(KV.first)) {
3245  auto *SecondMUD = dyn_cast<MemoryUseOrDef>(KV.second->getMemoryLeader());
3246  if (FirstMUD && SecondMUD) {
3248  assert((singleReachablePHIPath(VisitedMAS, FirstMUD, SecondMUD) ||
3249  ValueToClass.lookup(FirstMUD->getMemoryInst()) ==
3250  ValueToClass.lookup(SecondMUD->getMemoryInst())) &&
3251  "The instructions for these memory operations should have "
3252  "been in the same congruence class or reachable through"
3253  "a single argument phi");
3254  }
3255  } else if (auto *FirstMP = dyn_cast<MemoryPhi>(KV.first)) {
3256  // We can only sanely verify that MemoryDefs in the operand list all have
3257  // the same class.
3258  auto ReachableOperandPred = [&](const Use &U) {
3259  return ReachableEdges.count(
3260  {FirstMP->getIncomingBlock(U), FirstMP->getBlock()}) &&
3261  isa<MemoryDef>(U);
3262 
3263  };
3264  // All arguments should in the same class, ignoring unreachable arguments
3265  auto FilteredPhiArgs =
3266  make_filter_range(FirstMP->operands(), ReachableOperandPred);
3268  std::transform(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
3269  std::back_inserter(PhiOpClasses), [&](const Use &U) {
3270  const MemoryDef *MD = cast<MemoryDef>(U);
3271  return ValueToClass.lookup(MD->getMemoryInst());
3272  });
3273  assert(is_splat(PhiOpClasses) &&
3274  "All MemoryPhi arguments should be in the same class");
3275  }
3276  }
3277 #endif
3278 }
3279 
3280 // Verify that the sparse propagation we did actually found the maximal fixpoint
3281 // We do this by storing the value to class mapping, touching all instructions,
3282 // and redoing the iteration to see if anything changed.
3283 void NewGVN::verifyIterationSettled(Function &F) {
3284 #ifndef NDEBUG
3285  LLVM_DEBUG(dbgs() << "Beginning iteration verification\n");
3286  if (DebugCounter::isCounterSet(VNCounter))
3287  DebugCounter::setCounterValue(VNCounter, StartingVNCounter);
3288 
3289  // Note that we have to store the actual classes, as we may change existing
3290  // classes during iteration. This is because our memory iteration propagation
3291  // is not perfect, and so may waste a little work. But it should generate
3292  // exactly the same congruence classes we have now, with different IDs.
3293  std::map<const Value *, CongruenceClass> BeforeIteration;
3294 
3295  for (auto &KV : ValueToClass) {
3296  if (auto *I = dyn_cast<Instruction>(KV.first))
3297  // Skip unused/dead instructions.
3298  if (InstrToDFSNum(I) == 0)
3299  continue;
3300  BeforeIteration.insert({KV.first, *KV.second});
3301  }
3302 
3303  TouchedInstructions.set();
3304  TouchedInstructions.reset(0);
3305  iterateTouchedInstructions();
3307  EqualClasses;
3308  for (const auto &KV : ValueToClass) {
3309  if (auto *I = dyn_cast<Instruction>(KV.first))
3310  // Skip unused/dead instructions.
3311  if (InstrToDFSNum(I) == 0)
3312  continue;
3313  // We could sink these uses, but i think this adds a bit of clarity here as
3314  // to what we are comparing.
3315  auto *BeforeCC = &BeforeIteration.find(KV.first)->second;
3316  auto *AfterCC = KV.second;
3317  // Note that the classes can't change at this point, so we memoize the set
3318  // that are equal.
3319  if (!EqualClasses.count({BeforeCC, AfterCC})) {
3320  assert(BeforeCC->isEquivalentTo(AfterCC) &&
3321  "Value number changed after main loop completed!");
3322  EqualClasses.insert({BeforeCC, AfterCC});
3323  }
3324  }
3325 #endif
3326 }
3327 
3328 // Verify that for each store expression in the expression to class mapping,
3329 // only the latest appears, and multiple ones do not appear.
3330 // Because loads do not use the stored value when doing equality with stores,
3331 // if we don't erase the old store expressions from the table, a load can find
3332 // a no-longer valid StoreExpression.
3333 void NewGVN::verifyStoreExpressions() const {
3334 #ifndef NDEBUG
3335  // This is the only use of this, and it's not worth defining a complicated
3336  // densemapinfo hash/equality function for it.
3337  std::set<
3338  std::pair<const Value *,
3339  std::tuple<const Value *, const CongruenceClass *, Value *>>>
3340  StoreExpressionSet;
3341  for (const auto &KV : ExpressionToClass) {
3342  if (auto *SE = dyn_cast<StoreExpression>(KV.first)) {
3343  // Make sure a version that will conflict with loads is not already there
3344  auto Res = StoreExpressionSet.insert(
3345  {SE->getOperand(0), std::make_tuple(SE->getMemoryLeader(), KV.second,
3346  SE->getStoredValue())});
3347  bool Okay = Res.second;
3348  // It's okay to have the same expression already in there if it is
3349  // identical in nature.
3350  // This can happen when the leader of the stored value changes over time.
3351  if (!Okay)
3352  Okay = (std::get<1>(Res.first->second) == KV.second) &&
3353  (lookupOperandLeader(std::get<2>(Res.first->second)) ==
3354  lookupOperandLeader(SE->getStoredValue()));
3355  assert(Okay && "Stored expression conflict exists in expression table");
3356  auto *ValueExpr = ValueToExpression.lookup(SE->getStoreInst());
3357  assert(ValueExpr && ValueExpr->equals(*SE) &&
3358  "StoreExpression in ExpressionToClass is not latest "
3359  "StoreExpression for value");
3360  }
3361  }
3362 #endif
3363 }
3364 
3365 // This is the main value numbering loop, it iterates over the initial touched
3366 // instruction set, propagating value numbers, marking things touched, etc,
3367 // until the set of touched instructions is completely empty.
3368 void NewGVN::iterateTouchedInstructions() {
3369  unsigned int Iterations = 0;
3370  // Figure out where touchedinstructions starts
3371  int FirstInstr = TouchedInstructions.find_first();
3372  // Nothing set, nothing to iterate, just return.
3373  if (FirstInstr == -1)
3374  return;
3375  const BasicBlock *LastBlock = getBlockForValue(InstrFromDFSNum(FirstInstr));
3376  while (TouchedInstructions.any()) {
3377  ++Iterations;
3378  // Walk through all the instructions in all the blocks in RPO.
3379  // TODO: As we hit a new block, we should push and pop equalities into a
3380  // table lookupOperandLeader can use, to catch things PredicateInfo
3381  // might miss, like edge-only equivalences.
3382  for (unsigned InstrNum : TouchedInstructions.set_bits()) {
3383 
3384  // This instruction was found to be dead. We don't bother looking
3385  // at it again.
3386  if (InstrNum == 0) {
3387  TouchedInstructions.reset(InstrNum);
3388  continue;
3389  }
3390 
3391  Value *V = InstrFromDFSNum(InstrNum);
3392  const BasicBlock *CurrBlock = getBlockForValue(V);
3393 
3394  // If we hit a new block, do reachability processing.
3395  if (CurrBlock != LastBlock) {
3396  LastBlock = CurrBlock;
3397  bool BlockReachable = ReachableBlocks.count(CurrBlock);
3398  const auto &CurrInstRange = BlockInstRange.lookup(CurrBlock);
3399 
3400  // If it's not reachable, erase any touched instructions and move on.
3401  if (!BlockReachable) {
3402  TouchedInstructions.reset(CurrInstRange.first, CurrInstRange.second);
3403  LLVM_DEBUG(dbgs() << "Skipping instructions in block "
3404  << getBlockName(CurrBlock)
3405  << " because it is unreachable\n");
3406  continue;
3407  }
3408  updateProcessedCount(CurrBlock);
3409  }
3410  // Reset after processing (because we may mark ourselves as touched when
3411  // we propagate equalities).
3412  TouchedInstructions.reset(InstrNum);
3413 
3414  if (auto *MP = dyn_cast<MemoryPhi>(V)) {
3415  LLVM_DEBUG(dbgs() << "Processing MemoryPhi " << *MP << "\n");
3416  valueNumberMemoryPhi(MP);
3417  } else if (auto *I = dyn_cast<Instruction>(V)) {
3418  valueNumberInstruction(I);
3419  } else {
3420  llvm_unreachable("Should have been a MemoryPhi or Instruction");
3421  }
3422  updateProcessedCount(V);
3423  }
3424  }
3425  NumGVNMaxIterations = std::max(NumGVNMaxIterations.getValue(), Iterations);
3426 }
3427 
3428 // This is the main transformation entry point.
3429 bool NewGVN::runGVN() {
3430  if (DebugCounter::isCounterSet(VNCounter))
3431  StartingVNCounter = DebugCounter::getCounterValue(VNCounter);
3432  bool Changed = false;
3433  NumFuncArgs = F.arg_size();
3434  MSSAWalker = MSSA->getWalker();
3435  SingletonDeadExpression = new (ExpressionAllocator) DeadExpression();
3436 
3437  // Count number of instructions for sizing of hash tables, and come
3438  // up with a global dfs numbering for instructions.
3439  unsigned ICount = 1;
3440  // Add an empty instruction to account for the fact that we start at 1
3441  DFSToInstr.emplace_back(nullptr);
3442  // Note: We want ideal RPO traversal of the blocks, which is not quite the
3443  // same as dominator tree order, particularly with regard whether backedges
3444  // get visited first or second, given a block with multiple successors.
3445  // If we visit in the wrong order, we will end up performing N times as many
3446  // iterations.
3447  // The dominator tree does guarantee that, for a given dom tree node, it's
3448  // parent must occur before it in the RPO ordering. Thus, we only need to sort
3449  // the siblings.
3451  unsigned Counter = 0;
3452  for (auto &B : RPOT) {
3453  auto *Node = DT->getNode(B);
3454  assert(Node && "RPO and Dominator tree should have same reachability");
3455  RPOOrdering[Node] = ++Counter;
3456  }
3457  // Sort dominator tree children arrays into RPO.
3458  for (auto &B : RPOT) {
3459  auto *Node = DT->getNode(B);
3460  if (Node->getChildren().size() > 1)
3461  llvm::sort(Node->begin(), Node->end(),
3462  [&](const DomTreeNode *A, const DomTreeNode *B) {
3463  return RPOOrdering[A] < RPOOrdering[B];
3464  });
3465  }
3466 
3467  // Now a standard depth first ordering of the domtree is equivalent to RPO.
3468  for (auto DTN : depth_first(DT->getRootNode())) {
3469  BasicBlock *B = DTN->getBlock();
3470  const auto &BlockRange = assignDFSNumbers(B, ICount);
3471  BlockInstRange.insert({B, BlockRange});
3472  ICount += BlockRange.second - BlockRange.first;
3473  }
3474  initializeCongruenceClasses(F);
3475 
3476  TouchedInstructions.resize(ICount);
3477  // Ensure we don't end up resizing the expressionToClass map, as
3478  // that can be quite expensive. At most, we have one expression per
3479  // instruction.
3480  ExpressionToClass.reserve(ICount);
3481 
3482  // Initialize the touched instructions to include the entry block.
3483  const auto &InstRange = BlockInstRange.lookup(&F.getEntryBlock());
3484  TouchedInstructions.set(InstRange.first, InstRange.second);
3485  LLVM_DEBUG(dbgs() << "Block " << getBlockName(&F.getEntryBlock())
3486  << " marked reachable\n");
3487  ReachableBlocks.insert(&F.getEntryBlock());
3488 
3489  iterateTouchedInstructions();
3490  verifyMemoryCongruency();
3491  verifyIterationSettled(F);
3492  verifyStoreExpressions();
3493 
3494  Changed |= eliminateInstructions(F);
3495 
3496  // Delete all instructions marked for deletion.
3497  for (Instruction *ToErase : InstructionsToErase) {
3498  if (!ToErase->use_empty())
3499  ToErase->replaceAllUsesWith(UndefValue::get(ToErase->getType()));
3500 
3501  assert(ToErase->getParent() &&
3502  "BB containing ToErase deleted unexpectedly!");
3503  ToErase->eraseFromParent();
3504  }
3505  Changed |= !InstructionsToErase.empty();
3506 
3507  // Delete all unreachable blocks.
3508  auto UnreachableBlockPred = [&](const BasicBlock &BB) {
3509  return !ReachableBlocks.count(&BB);
3510  };
3511 
3512  for (auto &BB : make_filter_range(F, UnreachableBlockPred)) {
3513  LLVM_DEBUG(dbgs() << "We believe block " << getBlockName(&BB)
3514  << " is unreachable\n");
3515  deleteInstructionsInBlock(&BB);
3516  Changed = true;
3517  }
3518 
3519  cleanupTables();
3520  return Changed;
3521 }
3522 
3524  int DFSIn = 0;
3525  int DFSOut = 0;
3526  int LocalNum = 0;
3527 
3528  // Only one of Def and U will be set.
3529  // The bool in the Def tells us whether the Def is the stored value of a
3530  // store.
3532  Use *U = nullptr;
3533 
3534  bool operator<(const ValueDFS &Other) const {
3535  // It's not enough that any given field be less than - we have sets
3536  // of fields that need to be evaluated together to give a proper ordering.
3537  // For example, if you have;
3538  // DFS (1, 3)
3539  // Val 0
3540  // DFS (1, 2)
3541  // Val 50
3542  // We want the second to be less than the first, but if we just go field
3543  // by field, we will get to Val 0 < Val 50 and say the first is less than
3544  // the second. We only want it to be less than if the DFS orders are equal.
3545  //
3546  // Each LLVM instruction only produces one value, and thus the lowest-level
3547  // differentiator that really matters for the stack (and what we use as as a
3548  // replacement) is the local dfs number.
3549  // Everything else in the structure is instruction level, and only affects
3550  // the order in which we will replace operands of a given instruction.
3551  //
3552  // For a given instruction (IE things with equal dfsin, dfsout, localnum),
3553  // the order of replacement of uses does not matter.
3554  // IE given,
3555  // a = 5
3556  // b = a + a
3557  // When you hit b, you will have two valuedfs with the same dfsin, out, and
3558  // localnum.
3559  // The .val will be the same as well.
3560  // The .u's will be different.
3561  // You will replace both, and it does not matter what order you replace them
3562  // in (IE whether you replace operand 2, then operand 1, or operand 1, then
3563  // operand 2).
3564  // Similarly for the case of same dfsin, dfsout, localnum, but different
3565  // .val's
3566  // a = 5
3567  // b = 6
3568  // c = a + b
3569  // in c, we will a valuedfs for a, and one for b,with everything the same
3570  // but .val and .u.
3571  // It does not matter what order we replace these operands in.
3572  // You will always end up with the same IR, and this is guaranteed.
3573  return std::tie(DFSIn, DFSOut, LocalNum, Def, U) <
3574  std::tie(Other.DFSIn, Other.DFSOut, Other.LocalNum, Other.Def,
3575  Other.U);
3576  }
3577 };
3578 
3579 // This function converts the set of members for a congruence class from values,
3580 // to sets of defs and uses with associated DFS info. The total number of
3581 // reachable uses for each value is stored in UseCount, and instructions that
3582 // seem
3583 // dead (have no non-dead uses) are stored in ProbablyDead.
3584 void NewGVN::convertClassToDFSOrdered(
3585  const CongruenceClass &Dense, SmallVectorImpl<ValueDFS> &DFSOrderedSet,
3587  SmallPtrSetImpl<Instruction *> &ProbablyDead) const {
3588  for (auto D : Dense) {
3589  // First add the value.
3590  BasicBlock *BB = getBlockForValue(D);
3591  // Constants are handled prior to ever calling this function, so
3592  // we should only be left with instructions as members.
3593  assert(BB && "Should have figured out a basic block for value");
3594  ValueDFS VDDef;
3595  DomTreeNode *DomNode = DT->getNode(BB);
3596  VDDef.DFSIn = DomNode->getDFSNumIn();
3597  VDDef.DFSOut = DomNode->getDFSNumOut();
3598  // If it's a store, use the leader of the value operand, if it's always
3599  // available, or the value operand. TODO: We could do dominance checks to
3600  // find a dominating leader, but not worth it ATM.
3601  if (auto *SI = dyn_cast<StoreInst>(D)) {
3602  auto Leader = lookupOperandLeader(SI->getValueOperand());
3603  if (alwaysAvailable(Leader)) {
3604  VDDef.Def.setPointer(Leader);
3605  } else {
3606  VDDef.Def.setPointer(SI->getValueOperand());
3607  VDDef.Def.setInt(true);
3608  }
3609  } else {
3610  VDDef.Def.setPointer(D);
3611  }
3612  assert(isa<Instruction>(D) &&
3613  "The dense set member should always be an instruction");
3614  Instruction *Def = cast<Instruction>(D);
3615  VDDef.LocalNum = InstrToDFSNum(D);
3616  DFSOrderedSet.push_back(VDDef);
3617  // If there is a phi node equivalent, add it
3618  if (auto *PN = RealToTemp.lookup(Def)) {
3619  auto *PHIE =
3620  dyn_cast_or_null<PHIExpression>(ValueToExpression.lookup(Def));
3621  if (PHIE) {
3622  VDDef.Def.setInt(false);
3623  VDDef.Def.setPointer(PN);
3624  VDDef.LocalNum = 0;
3625  DFSOrderedSet.push_back(VDDef);
3626  }
3627  }
3628 
3629  unsigned int UseCount = 0;
3630  // Now add the uses.
3631  for (auto &U : Def->uses()) {
3632  if (auto *I = dyn_cast<Instruction>(U.getUser())) {
3633  // Don't try to replace into dead uses
3634  if (InstructionsToErase.count(I))
3635  continue;
3636  ValueDFS VDUse;
3637  // Put the phi node uses in the incoming block.
3638  BasicBlock *IBlock;
3639  if (auto *P = dyn_cast<PHINode>(I)) {
3640  IBlock = P->getIncomingBlock(U);
3641  // Make phi node users appear last in the incoming block
3642  // they are from.
3643  VDUse.LocalNum = InstrDFS.size() + 1;
3644  } else {
3645  IBlock = getBlockForValue(I);
3646  VDUse.LocalNum = InstrToDFSNum(I);
3647  }
3648 
3649  // Skip uses in unreachable blocks, as we're going
3650  // to delete them.
3651  if (ReachableBlocks.count(IBlock) == 0)
3652  continue;
3653 
3654  DomTreeNode *DomNode = DT->getNode(IBlock);
3655  VDUse.DFSIn = DomNode->getDFSNumIn();
3656  VDUse.DFSOut = DomNode->getDFSNumOut();
3657  VDUse.U = &U;
3658  ++UseCount;
3659  DFSOrderedSet.emplace_back(VDUse);
3660  }
3661  }
3662 
3663  // If there are no uses, it's probably dead (but it may have side-effects,
3664  // so not definitely dead. Otherwise, store the number of uses so we can
3665  // track if it becomes dead later).
3666  if (UseCount == 0)
3667  ProbablyDead.insert(Def);
3668  else
3669  UseCounts[Def] = UseCount;
3670  }
3671 }
3672 
3673 // This function converts the set of members for a congruence class from values,
3674 // to the set of defs for loads and stores, with associated DFS info.
3675 void NewGVN::convertClassToLoadsAndStores(
3676  const CongruenceClass &Dense,
3677  SmallVectorImpl<ValueDFS> &LoadsAndStores) const {
3678  for (auto D : Dense) {
3679  if (!isa<LoadInst>(D) && !isa<StoreInst>(D))
3680  continue;
3681 
3682  BasicBlock *BB = getBlockForValue(D);
3683  ValueDFS VD;
3684  DomTreeNode *DomNode = DT->getNode(BB);
3685  VD.DFSIn = DomNode->getDFSNumIn();
3686  VD.DFSOut = DomNode->getDFSNumOut();
3687  VD.Def.setPointer(D);
3688 
3689  // If it's an instruction, use the real local dfs number.
3690  if (auto *I = dyn_cast<Instruction>(D))
3691  VD.LocalNum = InstrToDFSNum(I);
3692  else
3693  llvm_unreachable("Should have been an instruction");
3694 
3695  LoadsAndStores.emplace_back(VD);
3696  }
3697 }
3698 
3700  patchReplacementInstruction(I, Repl);
3701  I->replaceAllUsesWith(Repl);
3702 }
3703 
3704 void NewGVN::deleteInstructionsInBlock(BasicBlock *BB) {
3705  LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
3706  ++NumGVNBlocksDeleted;
3707 
3708  // Delete the instructions backwards, as it has a reduced likelihood of having
3709  // to update as many def-use and use-def chains. Start after the terminator.
3710  auto StartPoint = BB->rbegin();
3711  ++StartPoint;
3712  // Note that we explicitly recalculate BB->rend() on each iteration,
3713  // as it may change when we remove the first instruction.
3714  for (BasicBlock::reverse_iterator I(StartPoint); I != BB->rend();) {
3715  Instruction &Inst = *I++;
3716  if (!Inst.use_empty())
3718  if (isa<LandingPadInst>(Inst))
3719  continue;
3720 
3721  Inst.eraseFromParent();
3722  ++NumGVNInstrDeleted;
3723  }
3724  // Now insert something that simplifycfg will turn into an unreachable.
3725  Type *Int8Ty = Type::getInt8Ty(BB->getContext());
3726  new StoreInst(UndefValue::get(Int8Ty),
3728  BB->getTerminator());
3729 }
3730 
3731 void NewGVN::markInstructionForDeletion(Instruction *I) {
3732  LLVM_DEBUG(dbgs() << "Marking " << *I << " for deletion\n");
3733  InstructionsToErase.insert(I);
3734 }
3735 
3736 void NewGVN::replaceInstruction(Instruction *I, Value *V) {
3737  LLVM_DEBUG(dbgs() << "Replacing " << *I << " with " << *V << "\n");
3739  // We save the actual erasing to avoid invalidating memory
3740  // dependencies until we are done with everything.
3741  markInstructionForDeletion(I);
3742 }
3743 
3744 namespace {
3745 
3746 // This is a stack that contains both the value and dfs info of where
3747 // that value is valid.
3748 class ValueDFSStack {
3749 public:
3750  Value *back() const { return ValueStack.back(); }
3751  std::pair<int, int> dfs_back() const { return DFSStack.back(); }
3752 
3753  void push_back(Value *V, int DFSIn, int DFSOut) {
3754  ValueStack.emplace_back(V);
3755  DFSStack.emplace_back(DFSIn, DFSOut);
3756  }
3757 
3758  bool empty() const { return DFSStack.empty(); }
3759 
3760  bool isInScope(int DFSIn, int DFSOut) const {
3761  if (empty())
3762  return false;
3763  return DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second;
3764  }
3765 
3766  void popUntilDFSScope(int DFSIn, int DFSOut) {
3767 
3768  // These two should always be in sync at this point.
3769  assert(ValueStack.size() == DFSStack.size() &&
3770  "Mismatch between ValueStack and DFSStack");
3771  while (
3772  !DFSStack.empty() &&
3773  !(DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second)) {
3774  DFSStack.pop_back();
3775  ValueStack.pop_back();
3776  }
3777  }
3778 
3779 private:
3780  SmallVector<Value *, 8> ValueStack;
3781  SmallVector<std::pair<int, int>, 8> DFSStack;
3782 };
3783 
3784 } // end anonymous namespace
3785 
3786 // Given an expression, get the congruence class for it.
3787 CongruenceClass *NewGVN::getClassForExpression(const Expression *E) const {
3788  if (auto *VE = dyn_cast<VariableExpression>(E))
3789  return ValueToClass.lookup(VE->getVariableValue());
3790  else if (isa<DeadExpression>(E))
3791  return TOPClass;
3792  return ExpressionToClass.lookup(E);
3793 }
3794 
3795 // Given a value and a basic block we are trying to see if it is available in,
3796 // see if the value has a leader available in that block.
3797 Value *NewGVN::findPHIOfOpsLeader(const Expression *E,
3798  const Instruction *OrigInst,
3799  const BasicBlock *BB) const {
3800  // It would already be constant if we could make it constant
3801  if (auto *CE = dyn_cast<ConstantExpression>(E))
3802  return CE->getConstantValue();
3803  if (auto *VE = dyn_cast<VariableExpression>(E)) {
3804  auto *V = VE->getVariableValue();
3805  if (alwaysAvailable(V) || DT->dominates(getBlockForValue(V), BB))
3806  return VE->getVariableValue();
3807  }
3808 
3809  auto *CC = getClassForExpression(E);
3810  if (!CC)
3811  return nullptr;
3812  if (alwaysAvailable(CC->getLeader()))
3813  return CC->getLeader();
3814 
3815  for (auto Member : *CC) {
3816  auto *MemberInst = dyn_cast<Instruction>(Member);
3817  if (MemberInst == OrigInst)
3818  continue;
3819  // Anything that isn't an instruction is always available.
3820  if (!MemberInst)
3821  return Member;
3822  if (DT->dominates(getBlockForValue(MemberInst), BB))
3823  return Member;
3824  }
3825  return nullptr;
3826 }
3827 
3828 bool NewGVN::eliminateInstructions(Function &F) {
3829  // This is a non-standard eliminator. The normal way to eliminate is
3830  // to walk the dominator tree in order, keeping track of available
3831  // values, and eliminating them. However, this is mildly
3832  // pointless. It requires doing lookups on every instruction,
3833  // regardless of whether we will ever eliminate it. For
3834  // instructions part of most singleton congruence classes, we know we
3835  // will never eliminate them.
3836 
3837  // Instead, this eliminator looks at the congruence classes directly, sorts
3838  // them into a DFS ordering of the dominator tree, and then we just
3839  // perform elimination straight on the sets by walking the congruence
3840  // class member uses in order, and eliminate the ones dominated by the
3841  // last member. This is worst case O(E log E) where E = number of
3842  // instructions in a single congruence class. In theory, this is all
3843  // instructions. In practice, it is much faster, as most instructions are
3844  // either in singleton congruence classes or can't possibly be eliminated
3845  // anyway (if there are no overlapping DFS ranges in class).
3846  // When we find something not dominated, it becomes the new leader
3847  // for elimination purposes.
3848  // TODO: If we wanted to be faster, We could remove any members with no
3849  // overlapping ranges while sorting, as we will never eliminate anything
3850  // with those members, as they don't dominate anything else in our set.
3851 
3852  bool AnythingReplaced = false;
3853 
3854  // Since we are going to walk the domtree anyway, and we can't guarantee the
3855  // DFS numbers are updated, we compute some ourselves.
3856  DT->updateDFSNumbers();
3857 
3858  // Go through all of our phi nodes, and kill the arguments associated with
3859  // unreachable edges.
3860  auto ReplaceUnreachablePHIArgs = [&](PHINode *PHI, BasicBlock *BB) {
3861  for (auto &Operand : PHI->incoming_values())
3862  if (!ReachableEdges.count({PHI->getIncomingBlock(Operand), BB})) {
3863  LLVM_DEBUG(dbgs() << "Replacing incoming value of " << PHI
3864  << " for block "
3865  << getBlockName(PHI->getIncomingBlock(Operand))
3866  << " with undef due to it being unreachable\n");
3867  Operand.set(UndefValue::get(PHI->getType()));
3868  }
3869  };
3870  // Replace unreachable phi arguments.
3871  // At this point, RevisitOnReachabilityChange only contains:
3872  //
3873  // 1. PHIs
3874  // 2. Temporaries that will convert to PHIs
3875  // 3. Operations that are affected by an unreachable edge but do not fit into
3876  // 1 or 2 (rare).
3877  // So it is a slight overshoot of what we want. We could make it exact by
3878  // using two SparseBitVectors per block.
3879  DenseMap<const BasicBlock *, unsigned> ReachablePredCount;
3880  for (auto &KV : ReachableEdges)
3881  ReachablePredCount[KV.getEnd()]++;
3882  for (auto &BBPair : RevisitOnReachabilityChange) {
3883  for (auto InstNum : BBPair.second) {
3884  auto *Inst = InstrFromDFSNum(InstNum);
3885  auto *PHI = dyn_cast<PHINode>(Inst);
3886  PHI = PHI ? PHI : dyn_cast_or_null<PHINode>(RealToTemp.lookup(Inst));
3887  if (!PHI)
3888  continue;
3889  auto *BB = BBPair.first;
3890  if (ReachablePredCount.lookup(BB) != PHI->getNumIncomingValues())
3891  ReplaceUnreachablePHIArgs(PHI, BB);
3892  }
3893  }
3894 
3895  // Map to store the use counts
3897  for (auto *CC : reverse(CongruenceClasses)) {
3898  LLVM_DEBUG(dbgs() << "Eliminating in congruence class " << CC->getID()
3899  << "\n");
3900  // Track the equivalent store info so we can decide whether to try
3901  // dead store elimination.
3902  SmallVector<ValueDFS, 8> PossibleDeadStores;
3903  SmallPtrSet<Instruction *, 8> ProbablyDead;
3904  if (CC->isDead() || CC->empty())
3905  continue;
3906  // Everything still in the TOP class is unreachable or dead.
3907  if (CC == TOPClass) {
3908  for (auto M : *CC) {
3909  auto *VTE = ValueToExpression.lookup(M);
3910  if (VTE && isa<DeadExpression>(VTE))
3911  markInstructionForDeletion(cast<Instruction>(M));
3912  assert((!ReachableBlocks.count(cast<Instruction>(M)->getParent()) ||
3913  InstructionsToErase.count(cast<Instruction>(M))) &&
3914  "Everything in TOP should be unreachable or dead at this "
3915  "point");
3916  }
3917  continue;
3918  }
3919 
3920  assert(CC->getLeader() && "We should have had a leader");
3921  // If this is a leader that is always available, and it's a
3922  // constant or has no equivalences, just replace everything with
3923  // it. We then update the congruence class with whatever members
3924  // are left.
3925  Value *Leader =
3926  CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
3927  if (alwaysAvailable(Leader)) {
3928  CongruenceClass::MemberSet MembersLeft;
3929  for (auto M : *CC) {
3930  Value *Member = M;
3931  // Void things have no uses we can replace.
3932  if (Member == Leader || !isa<Instruction>(Member) ||
3933  Member->getType()->isVoidTy()) {
3934  MembersLeft.insert(Member);
3935  continue;
3936  }
3937  LLVM_DEBUG(dbgs() << "Found replacement " << *(Leader) << " for "
3938  << *Member << "\n");
3939  auto *I = cast<Instruction>(Member);
3940  assert(Leader != I && "About to accidentally remove our leader");
3941  replaceInstruction(I, Leader);
3942  AnythingReplaced = true;
3943  }
3944  CC->swap(MembersLeft);
3945  } else {
3946  // If this is a singleton, we can skip it.
3947  if (CC->size() != 1 || RealToTemp.count(Leader)) {
3948  // This is a stack because equality replacement/etc may place
3949  // constants in the middle of the member list, and we want to use
3950  // those constant values in preference to the current leader, over
3951  // the scope of those constants.
3952  ValueDFSStack EliminationStack;
3953 
3954  // Convert the members to DFS ordered sets and then merge them.
3955  SmallVector<ValueDFS, 8> DFSOrderedSet;
3956  convertClassToDFSOrdered(*CC, DFSOrderedSet, UseCounts, ProbablyDead);
3957 
3958  // Sort the whole thing.
3959  llvm::sort(DFSOrderedSet);
3960  for (auto &VD : DFSOrderedSet) {
3961  int MemberDFSIn = VD.DFSIn;
3962  int MemberDFSOut = VD.DFSOut;
3963  Value *Def = VD.Def.getPointer();
3964  bool FromStore = VD.Def.getInt();
3965  Use *U = VD.U;
3966  // We ignore void things because we can't get a value from them.
3967  if (Def && Def->getType()->isVoidTy())
3968  continue;
3969  auto *DefInst = dyn_cast_or_null<Instruction>(Def);
3970  if (DefInst && AllTempInstructions.count(DefInst)) {
3971  auto *PN = cast<PHINode>(DefInst);
3972 
3973  // If this is a value phi and that's the expression we used, insert
3974  // it into the program
3975  // remove from temp instruction list.
3976  AllTempInstructions.erase(PN);
3977  auto *DefBlock = getBlockForValue(Def);
3978  LLVM_DEBUG(dbgs() << "Inserting fully real phi of ops" << *Def
3979  << " into block "
3980  << getBlockName(getBlockForValue(Def)) << "\n");
3981  PN->insertBefore(&DefBlock->front());
3982  Def = PN;
3983  NumGVNPHIOfOpsEliminations++;
3984  }
3985 
3986  if (EliminationStack.empty()) {
3987  LLVM_DEBUG(dbgs() << "Elimination Stack is empty\n");
3988  } else {
3989  LLVM_DEBUG(dbgs() << "Elimination Stack Top DFS numbers are ("
3990  << EliminationStack.dfs_back().first << ","
3991  << EliminationStack.dfs_back().second << ")\n");
3992  }
3993 
3994  LLVM_DEBUG(dbgs() << "Current DFS numbers are (" << MemberDFSIn << ","
3995  << MemberDFSOut << ")\n");
3996  // First, we see if we are out of scope or empty. If so,
3997  // and there equivalences, we try to replace the top of
3998  // stack with equivalences (if it's on the stack, it must
3999  // not have been eliminated yet).
4000  // Then we synchronize to our current scope, by
4001  // popping until we are back within a DFS scope that
4002  // dominates the current member.
4003  // Then, what happens depends on a few factors
4004  // If the stack is now empty, we need to push
4005  // If we have a constant or a local equivalence we want to
4006  // start using, we also push.
4007  // Otherwise, we walk along, processing members who are
4008  // dominated by this scope, and eliminate them.
4009  bool ShouldPush = Def && EliminationStack.empty();
4010  bool OutOfScope =
4011  !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut);
4012 
4013  if (OutOfScope || ShouldPush) {
4014  // Sync to our current scope.
4015  EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4016  bool ShouldPush = Def && EliminationStack.empty();
4017  if (ShouldPush) {
4018  EliminationStack.push_back(Def, MemberDFSIn, MemberDFSOut);
4019  }
4020  }
4021 
4022  // Skip the Def's, we only want to eliminate on their uses. But mark
4023  // dominated defs as dead.
4024  if (Def) {
4025  // For anything in this case, what and how we value number
4026  // guarantees that any side-effets that would have occurred (ie
4027  // throwing, etc) can be proven to either still occur (because it's
4028  // dominated by something that has the same side-effects), or never
4029  // occur. Otherwise, we would not have been able to prove it value
4030  // equivalent to something else. For these things, we can just mark
4031  // it all dead. Note that this is different from the "ProbablyDead"
4032  // set, which may not be dominated by anything, and thus, are only
4033  // easy to prove dead if they are also side-effect free. Note that
4034  // because stores are put in terms of the stored value, we skip
4035  // stored values here. If the stored value is really dead, it will
4036  // still be marked for deletion when we process it in its own class.
4037  if (!EliminationStack.empty() && Def != EliminationStack.back() &&
4038  isa<Instruction>(Def) && !FromStore)
4039  markInstructionForDeletion(cast<Instruction>(Def));
4040  continue;
4041  }
4042  // At this point, we know it is a Use we are trying to possibly
4043  // replace.
4044 
4045  assert(isa<Instruction>(U->get()) &&
4046  "Current def should have been an instruction");
4047  assert(isa<Instruction>(U->getUser()) &&
4048  "Current user should have been an instruction");
4049 
4050  // If the thing we are replacing into is already marked to be dead,
4051  // this use is dead. Note that this is true regardless of whether
4052  // we have anything dominating the use or not. We do this here
4053  // because we are already walking all the uses anyway.
4054  Instruction *InstUse = cast<Instruction>(U->getUser());
4055  if (InstructionsToErase.count(InstUse)) {
4056  auto &UseCount = UseCounts[U->get()];
4057  if (--UseCount == 0) {
4058  ProbablyDead.insert(cast<Instruction>(U->get()));
4059  }
4060  }
4061 
4062  // If we get to this point, and the stack is empty we must have a use
4063  // with nothing we can use to eliminate this use, so just skip it.
4064  if (EliminationStack.empty())
4065  continue;
4066 
4067  Value *DominatingLeader = EliminationStack.back();
4068 
4069  auto *II = dyn_cast<IntrinsicInst>(DominatingLeader);
4070  bool isSSACopy = II && II->getIntrinsicID() == Intrinsic::ssa_copy;
4071  if (isSSACopy)
4072  DominatingLeader = II->getOperand(0);
4073 
4074  // Don't replace our existing users with ourselves.
4075  if (U->get() == DominatingLeader)
4076  continue;
4077  LLVM_DEBUG(dbgs()
4078  << "Found replacement " << *DominatingLeader << " for "
4079  << *U->get() << " in " << *(U->getUser()) << "\n");
4080 
4081  // If we replaced something in an instruction, handle the patching of
4082  // metadata. Skip this if we are replacing predicateinfo with its
4083  // original operand, as we already know we can just drop it.
4084  auto *ReplacedInst = cast<Instruction>(U->get());
4085  auto *PI = PredInfo->getPredicateInfoFor(ReplacedInst);
4086  if (!PI || DominatingLeader != PI->OriginalOp)
4087  patchReplacementInstruction(ReplacedInst, DominatingLeader);
4088  U->set(DominatingLeader);
4089  // This is now a use of the dominating leader, which means if the
4090  // dominating leader was dead, it's now live!
4091  auto &LeaderUseCount = UseCounts[DominatingLeader];
4092  // It's about to be alive again.
4093  if (LeaderUseCount == 0 && isa<Instruction>(DominatingLeader))
4094  ProbablyDead.erase(cast<Instruction>(DominatingLeader));
4095  // For copy instructions, we use their operand as a leader,
4096  // which means we remove a user of the copy and it may become dead.
4097  if (isSSACopy) {
4098  unsigned &IIUseCount = UseCounts[II];
4099  if (--IIUseCount == 0)
4100  ProbablyDead.insert(II);
4101  }
4102  ++LeaderUseCount;
4103  AnythingReplaced = true;
4104  }
4105  }
4106  }
4107 
4108  // At this point, anything still in the ProbablyDead set is actually dead if
4109  // would be trivially dead.
4110  for (auto *I : ProbablyDead)
4112  markInstructionForDeletion(I);
4113 
4114  // Cleanup the congruence class.
4115  CongruenceClass::MemberSet MembersLeft;
4116  for (auto *Member : *CC)
4117  if (!isa<Instruction>(Member) ||
4118  !InstructionsToErase.count(cast<Instruction>(Member)))
4119  MembersLeft.insert(Member);
4120  CC->swap(MembersLeft);
4121 
4122  // If we have possible dead stores to look at, try to eliminate them.
4123  if (CC->getStoreCount() > 0) {
4124  convertClassToLoadsAndStores(*CC, PossibleDeadStores);
4125  llvm::sort(PossibleDeadStores);
4126  ValueDFSStack EliminationStack;
4127  for (auto &VD : PossibleDeadStores) {
4128  int MemberDFSIn = VD.DFSIn;
4129  int MemberDFSOut = VD.DFSOut;
4130  Instruction *Member = cast<Instruction>(VD.Def.getPointer());
4131  if (EliminationStack.empty() ||
4132  !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut)) {
4133  // Sync to our current scope.
4134  EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4135  if (EliminationStack.empty()) {
4136  EliminationStack.push_back(Member, MemberDFSIn, MemberDFSOut);
4137  continue;
4138  }
4139  }
4140  // We already did load elimination, so nothing to do here.
4141  if (isa<LoadInst>(Member))
4142  continue;
4143  assert(!EliminationStack.empty());
4144  Instruction *Leader = cast<Instruction>(EliminationStack.back());
4145  (void)Leader;
4146  assert(DT->dominates(Leader->getParent(), Member->getParent()));
4147  // Member is dominater by Leader, and thus dead
4148  LLVM_DEBUG(dbgs() << "Marking dead store " << *Member
4149  << " that is dominated by " << *Leader << "\n");
4150  markInstructionForDeletion(Member);
4151  CC->erase(Member);
4152  ++NumGVNDeadStores;
4153  }
4154  }
4155  }
4156  return AnythingReplaced;
4157 }
4158 
4159 // This function provides global ranking of operations so that we can place them
4160 // in a canonical order. Note that rank alone is not necessarily enough for a
4161 // complete ordering, as constants all have the same rank. However, generally,
4162 // we will simplify an operation with all constants so that it doesn't matter
4163 // what order they appear in.
4164 unsigned int NewGVN::getRank(const Value *V) const {
4165  // Prefer constants to undef to anything else
4166  // Undef is a constant, have to check it first.
4167  // Prefer smaller constants to constantexprs
4168  if (isa<ConstantExpr>(V))
4169  return 2;
4170  if (isa<UndefValue>(V))
4171  return 1;
4172  if (isa<Constant>(V))
4173  return 0;
4174  else if (auto *A = dyn_cast<Argument>(V))
4175  return 3 + A->getArgNo();
4176 
4177  // Need to shift the instruction DFS by number of arguments + 3 to account for
4178  // the constant and argument ranking above.
4179  unsigned Result = InstrToDFSNum(V);
4180  if (Result > 0)
4181  return 4 + NumFuncArgs + Result;
4182  // Unreachable or something else, just return a really large number.
4183  return ~0;
4184 }
4185 
4186 // This is a function that says whether two commutative operations should
4187 // have their order swapped when canonicalizing.
4188 bool NewGVN::shouldSwapOperands(const Value *A, const Value *B) const {
4189  // Because we only care about a total ordering, and don't rewrite expressions
4190  // in this order, we order by rank, which will give a strict weak ordering to
4191  // everything but constants, and then we order by pointer address.
4192  return std::make_pair(getRank(A), A) > std::make_pair(getRank(B), B);
4193 }
4194 
4195 namespace {
4196 
4197 class NewGVNLegacyPass : public FunctionPass {
4198 public:
4199  // Pass identification, replacement for typeid.
4200  static char ID;
4201 
4202  NewGVNLegacyPass() : FunctionPass(ID) {
4204  }
4205 
4206  bool runOnFunction(Function &F) override;
4207 
4208 private:
4209  void getAnalysisUsage(AnalysisUsage &AU) const override {
4217  }
4218 };
4219 
4220 } // end anonymous namespace
4221 
4223  if (skipFunction(F))
4224  return false;
4225  return NewGVN(F, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4226  &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
4227  &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
4228  &getAnalysis<AAResultsWrapperPass>().getAAResults(),
4229  &getAnalysis<MemorySSAWrapperPass>().getMSSA(),
4230  F.getParent()->getDataLayout())
4231  .runGVN();
4232 }
4233 
4234 char NewGVNLegacyPass::ID = 0;
4235 
4236 INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering",
4237  false, false)
4244 INITIALIZE_PASS_END(NewGVNLegacyPass, "newgvn", "Global Value Numbering", false,
4245  false)
4246 
4247 // createGVNPass - The public interface to this file.
4248 FunctionPass *llvm::createNewGVNPass() { return new NewGVNLegacyPass(); }
4249 
4251  // Apparently the order in which we get these results matter for
4252  // the old GVN (see Chandler's comment in GVN.cpp). I'll keep
4253  // the same order here, just in case.
4254  auto &AC = AM.getResult<AssumptionAnalysis>(F);
4255  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
4256  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
4257  auto &AA = AM.getResult<AAManager>(F);
4258  auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
4259  bool Changed =
4260  NewGVN(F, &DT, &AC, &TLI, &AA, &MSSA, F.getParent()->getDataLayout())
4261  .runGVN();
4262  if (!Changed)
4263  return PreservedAnalyses::all();
4264  PreservedAnalyses PA;
4266  PA.preserve<GlobalsAA>();
4267  return PA;
4268 }
Legacy wrapper pass to provide the GlobalsAAResult object.
Global Value Numbering
Definition: NewGVN.cpp:4244
void initializeNewGVNLegacyPassPass(PassRegistry &)
uint64_t CallInst * C
Value * getValueOperand()
Definition: Instructions.h:409
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:67
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:258
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:594
reference emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:645
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:636
void setInt(IntType IntVal)
bool isSimple() const
Definition: Instructions.h:276
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
iterator_range< use_iterator > uses()
Definition: Value.h:354
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:242
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:463
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:249
void dropAllReferences()
Drop all references to operands.
Definition: User.h:294
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:769
static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS)
Definition: NewGVN.cpp:869
This class represents lattice values for constants.
Definition: AllocatorList.h:23
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.
iterator begin() const
Definition: ArrayRef.h:136
BasicBlock * getSuccessor(unsigned Idx) const
Return the specified successor. This instruction must be a terminator.
static bool okayForPHIOfOps(const Instruction *I)
Definition: NewGVN.cpp:2600
Implements a dense probed hash-table based set.
Definition: DenseSet.h:249
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 @llvm.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:372
bool isTerminator() const
Definition: Instruction.h:128
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:662
Recycle small arrays allocated from a BumpPtrAllocator.
Definition: ArrayRecycler.h:28
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:98
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:1185
BasicBlock * getSuccessor(unsigned i) const
STATISTIC(NumFunctions, "Total number of functions")
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:230
F(f)
reverse_iterator rend()
Definition: BasicBlock.h:275
An instruction for reading from memory.
Definition: Instructions.h:167
reverse_iterator rbegin()
Definition: BasicBlock.h:273
Hexagon Common GEP
Value * getCondition() const
const Instruction * 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:137
This defines the Use class.
Constant * getConstantMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, Type *LoadTy, const DataLayout &DL)
Definition: VNCoercion.cpp:523
const Value * DoPHITranslation(const BasicBlock *CurBB, const BasicBlock *PredBB) const
Translate PHI node to its predecessor from the given basic block.
Definition: Value.cpp:701
ExactEqualsExpression(const Expression &E)
Definition: NewGVN.cpp:432
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:32
op_iterator op_begin()
Definition: User.h:229
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:274
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:91
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:50
Legacy analysis pass which computes MemorySSA.
Definition: MemorySSA.h:951
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.
void patchReplacementInstruction(Instruction *I, Value *Repl)
Patch the replacement so that it is not more restrictive than the value being replaced.
Definition: Local.cpp:2379
Predicate getInversePredicate() const
For example, EQ -> NE, UGT -> ULE, SLT -> SGE, OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc.
Definition: InstrTypes.h:745
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:369
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:55
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:651
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:279
Constant * getConstantStoreValueForLoad(Constant *SrcVal, unsigned Offset, Type *LoadTy, const DataLayout &DL)
Definition: VNCoercion.cpp:391
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:41
Encapsulates MemorySSA, including all data associated with memory accesses.
Definition: MemorySSA.h:700
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:3531
friend const_iterator begin(StringRef path, Style style)
Get begin iterator over path.
Definition: Path.cpp:249
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:266
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
Definition: Constants.h:200
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:244
static int getID(struct InternalInstruction *insn, const void *miiArg)
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:32
static bool isCounterSet(unsigned ID)
Definition: DebugCounter.h:101
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:320
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:429
This is the generic walker interface for walkers of MemorySSA.
Definition: MemorySSA.h:982
const Expression & E
Definition: NewGVN.cpp:430
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:144
static const Expression * getEmptyKey()
Definition: NewGVN.cpp:442
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Value * getOperand(unsigned i) const
Definition: User.h:169
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:214
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:140
const BasicBlock & getEntryBlock() const
Definition: Function.h:639
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:423
friend const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:258
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:630
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
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:290
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
Allocate memory in an ever growing pool, as if by bump-pointer.
Definition: Allocator.h:140
Conditional or Unconditional Branch instruction.
bool is_splat(R &&Range)
Wrapper function around std::equal to detect if all elements in a container are same.
Definition: STLExtras.h:1306
static void setCounterValue(unsigned ID, int64_t Count)
Definition: DebugCounter.h:114
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:148
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:41
Value * getIncomingValueForBlock(const BasicBlock *BB) const
Constant * getConstantLoadValueForLoad(Constant *SrcVal, unsigned Offset, Type *LoadTy, const DataLayout &DL)
Definition: VNCoercion.cpp:449
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:239
A manager for alias analyses.
static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl)
Definition: NewGVN.cpp:3699
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:370
static int64_t getCounterValue(unsigned ID)
Definition: DebugCounter.h:106
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:231
iterator_range< def_chain_iterator< T, true > > optimized_def_chain(T MA)
Definition: MemorySSA.h:1296
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:1192
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:646
bool isBinaryOp() const
Definition: Instruction.h:130
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:284
amdgpu Simplify well known AMD library false FunctionCallee Value * Arg
size_t arg_size() const
Definition: Function.h:697
op_range operands()
Definition: User.h:237
Value * getPointerOperand()
Definition: Instructions.h:284
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:381
static bool shouldExecute(unsigned CounterName)
Definition: DebugCounter.h:73
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1424
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
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:1115
constexpr bool empty(const T &RangeOrContainer)
Test whether RangeOrContainer is empty. Similar to C++17 std::empty.
Definition: STLExtras.h:209
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:50
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:191
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
auto size(R &&Range, typename std::enable_if< std::is_same< typename std::iterator_traits< decltype(Range.begin())>::iterator_category, std::random_access_iterator_tag >::value, void >::type *=nullptr) -> decltype(std::distance(Range.begin(), Range.end()))
Get the size of a range.
Definition: STLExtras.h:1166
void setOpcode(unsigned opcode)
BlockVerifier::State From
bool erase(PtrType Ptr)
erase - If the set contains the specified pointer, remove it and return true, otherwise return false...
Definition: SmallPtrSet.h:377
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:841
Provides information about what library functions are available for the current target.
An analysis that produces MemorySSA for a function.
Definition: MemorySSA.h:915
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:374
PreservedAnalyses run(Function &F, AnalysisManager< Function > &AM)
Run the pass over the function.
Definition: NewGVN.cpp:4250
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:587
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:491
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:940
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:244
iterator_range< user_iterator > users()
Definition: Value.h:399
BasicBlock * getIncomingBlock(unsigned I) const
Return incoming basic block number i.
Definition: MemorySSA.h:540
An opaque object representing a hash code.
Definition: Hashing.h:71
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:251
bool operator<(const ValueDFS &Other) const
Definition: NewGVN.cpp:3534
Predicate getPredicate() const
Return the predicate for this instruction.
Definition: InstrTypes.h:721
iterator_range< typename GraphTraits< GraphType >::nodes_iterator > nodes(const GraphType &G)
Definition: GraphTraits.h:108
static Value * getCopyOf(const Value *V)
Definition: NewGVN.cpp:940
iterator begin() const
Definition: SmallPtrSet.h:396
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:240
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
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:192
iterator_range< value_op_iterator > operand_values()
Definition: User.h:261
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:322
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:427
static std::string getBlockName(const BasicBlock *B)
Definition: NewGVN.cpp:897
iterator end() const
Definition: SmallPtrSet.h:401
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:211
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:1267
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())
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:649
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:565
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:349
LLVM Value Representation.
Definition: Value.h:72
Value * getOperand(unsigned N) const
The header file for the GVN pass that contains expression handling classes.
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:356
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:4248
Predicate getSwappedPredicate() const
For example, EQ->EQ, SLE->SGE, ULT->UGT, OEQ->OEQ, ULE->UGE, OLT->OGT, etc.
Definition: InstrTypes.h:761
A container for analyses that lazily runs them and caches their results.
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:259
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
bool isSimple() const
Definition: Instructions.h:401
Represents phi nodes for memory accesses.
Definition: MemorySSA.h:478
static unsigned getHashValue(const Expression *E)
Definition: NewGVN.cpp:454
virtual bool exactlyEquals(const Expression &Other) const
#define LLVM_DEBUG(X)
Definition: Debug.h:122
op_range incoming_values()
OutputIt copy(R &&Range, OutputIt Out)
Definition: STLExtras.h:1237
Value * getPointerOperand()
Definition: Instructions.h:412
DEBUG_COUNTER(VNCounter, "newgvn-vn", "Controls which instructions are value numbered")
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:173
INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering", false, false) INITIALIZE_PASS_END(NewGVNLegacyPass
bool use_empty() const
Definition: Value.h:322
void allocateOperands(RecyclerType &Recycler, BumpPtrAllocator &Allocator)
iterator_range< arg_iterator > args()
Definition: Function.h:688
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:43
const BasicBlock * getParent() const
Definition: Instruction.h:66
newgvn
Definition: NewGVN.cpp:4244
bool operator==(const Expression &Other) const
Definition: NewGVN.cpp:436