LLVM  15.0.0git
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/SetOperations.h"
66 #include "llvm/ADT/SmallPtrSet.h"
67 #include "llvm/ADT/SmallVector.h"
69 #include "llvm/ADT/Statistic.h"
81 #include "llvm/IR/Argument.h"
82 #include "llvm/IR/BasicBlock.h"
83 #include "llvm/IR/Constant.h"
84 #include "llvm/IR/Constants.h"
85 #include "llvm/IR/Dominators.h"
86 #include "llvm/IR/Function.h"
87 #include "llvm/IR/InstrTypes.h"
88 #include "llvm/IR/Instruction.h"
89 #include "llvm/IR/Instructions.h"
90 #include "llvm/IR/IntrinsicInst.h"
91 #include "llvm/IR/PatternMatch.h"
92 #include "llvm/IR/Type.h"
93 #include "llvm/IR/Use.h"
94 #include "llvm/IR/User.h"
95 #include "llvm/IR/Value.h"
96 #include "llvm/InitializePasses.h"
97 #include "llvm/Pass.h"
98 #include "llvm/Support/Allocator.h"
100 #include "llvm/Support/Casting.h"
102 #include "llvm/Support/Debug.h"
107 #include "llvm/Transforms/Scalar.h"
113 #include <algorithm>
114 #include <cassert>
115 #include <cstdint>
116 #include <iterator>
117 #include <map>
118 #include <memory>
119 #include <set>
120 #include <string>
121 #include <tuple>
122 #include <utility>
123 #include <vector>
124 
125 using namespace llvm;
126 using namespace llvm::GVNExpression;
127 using namespace llvm::VNCoercion;
128 using namespace llvm::PatternMatch;
129 
130 #define DEBUG_TYPE "newgvn"
131 
132 STATISTIC(NumGVNInstrDeleted, "Number of instructions deleted");
133 STATISTIC(NumGVNBlocksDeleted, "Number of blocks deleted");
134 STATISTIC(NumGVNOpsSimplified, "Number of Expressions simplified");
135 STATISTIC(NumGVNPhisAllSame, "Number of PHIs whos arguments are all the same");
136 STATISTIC(NumGVNMaxIterations,
137  "Maximum Number of iterations it took to converge GVN");
138 STATISTIC(NumGVNLeaderChanges, "Number of leader changes");
139 STATISTIC(NumGVNSortedLeaderChanges, "Number of sorted leader changes");
140 STATISTIC(NumGVNAvoidedSortedLeaderChanges,
141  "Number of avoided sorted leader changes");
142 STATISTIC(NumGVNDeadStores, "Number of redundant/dead stores eliminated");
143 STATISTIC(NumGVNPHIOfOpsCreated, "Number of PHI of ops created");
144 STATISTIC(NumGVNPHIOfOpsEliminations,
145  "Number of things eliminated using PHI of ops");
146 DEBUG_COUNTER(VNCounter, "newgvn-vn",
147  "Controls which instructions are value numbered");
148 DEBUG_COUNTER(PHIOfOpsCounter, "newgvn-phi",
149  "Controls which instructions we create phi of ops for");
150 // Currently store defining access refinement is too slow due to basicaa being
151 // egregiously slow. This flag lets us keep it working while we work on this
152 // issue.
153 static cl::opt<bool> EnableStoreRefinement("enable-store-refinement",
154  cl::init(false), cl::Hidden);
155 
156 /// Currently, the generation "phi of ops" can result in correctness issues.
157 static cl::opt<bool> EnablePhiOfOps("enable-phi-of-ops", cl::init(true),
158  cl::Hidden);
159 
160 //===----------------------------------------------------------------------===//
161 // GVN Pass
162 //===----------------------------------------------------------------------===//
163 
164 // Anchor methods.
165 namespace llvm {
166 namespace GVNExpression {
167 
168 Expression::~Expression() = default;
175 
176 } // end namespace GVNExpression
177 } // end namespace llvm
178 
179 namespace {
180 
181 // Tarjan's SCC finding algorithm with Nuutila's improvements
182 // SCCIterator is actually fairly complex for the simple thing we want.
183 // It also wants to hand us SCC's that are unrelated to the phi node we ask
184 // about, and have us process them there or risk redoing work.
185 // Graph traits over a filter iterator also doesn't work that well here.
186 // This SCC finder is specialized to walk use-def chains, and only follows
187 // instructions,
188 // not generic values (arguments, etc).
189 struct TarjanSCC {
190  TarjanSCC() : Components(1) {}
191 
192  void Start(const Instruction *Start) {
193  if (Root.lookup(Start) == 0)
194  FindSCC(Start);
195  }
196 
197  const SmallPtrSetImpl<const Value *> &getComponentFor(const Value *V) const {
198  unsigned ComponentID = ValueToComponent.lookup(V);
199 
200  assert(ComponentID > 0 &&
201  "Asking for a component for a value we never processed");
202  return Components[ComponentID];
203  }
204 
205 private:
206  void FindSCC(const Instruction *I) {
207  Root[I] = ++DFSNum;
208  // Store the DFS Number we had before it possibly gets incremented.
209  unsigned int OurDFS = DFSNum;
210  for (auto &Op : I->operands()) {
211  if (auto *InstOp = dyn_cast<Instruction>(Op)) {
212  if (Root.lookup(Op) == 0)
213  FindSCC(InstOp);
214  if (!InComponent.count(Op))
215  Root[I] = std::min(Root.lookup(I), Root.lookup(Op));
216  }
217  }
218  // See if we really were the root of a component, by seeing if we still have
219  // our DFSNumber. If we do, we are the root of the component, and we have
220  // completed a component. If we do not, we are not the root of a component,
221  // and belong on the component stack.
222  if (Root.lookup(I) == OurDFS) {
223  unsigned ComponentID = Components.size();
224  Components.resize(Components.size() + 1);
225  auto &Component = Components.back();
226  Component.insert(I);
227  LLVM_DEBUG(dbgs() << "Component root is " << *I << "\n");
228  InComponent.insert(I);
229  ValueToComponent[I] = ComponentID;
230  // Pop a component off the stack and label it.
231  while (!Stack.empty() && Root.lookup(Stack.back()) >= OurDFS) {
232  auto *Member = Stack.back();
233  LLVM_DEBUG(dbgs() << "Component member is " << *Member << "\n");
234  Component.insert(Member);
235  InComponent.insert(Member);
236  ValueToComponent[Member] = ComponentID;
237  Stack.pop_back();
238  }
239  } else {
240  // Part of a component, push to stack
241  Stack.push_back(I);
242  }
243  }
244 
245  unsigned int DFSNum = 1;
246  SmallPtrSet<const Value *, 8> InComponent;
249 
250  // Store the components as vector of ptr sets, because we need the topo order
251  // of SCC's, but not individual member order
253 
254  DenseMap<const Value *, unsigned> ValueToComponent;
255 };
256 
257 // Congruence classes represent the set of expressions/instructions
258 // that are all the same *during some scope in the function*.
259 // That is, because of the way we perform equality propagation, and
260 // because of memory value numbering, it is not correct to assume
261 // you can willy-nilly replace any member with any other at any
262 // point in the function.
263 //
264 // For any Value in the Member set, it is valid to replace any dominated member
265 // with that Value.
266 //
267 // Every congruence class has a leader, and the leader is used to symbolize
268 // instructions in a canonical way (IE every operand of an instruction that is a
269 // member of the same congruence class will always be replaced with leader
270 // during symbolization). To simplify symbolization, we keep the leader as a
271 // constant if class can be proved to be a constant value. Otherwise, the
272 // leader is the member of the value set with the smallest DFS number. Each
273 // congruence class also has a defining expression, though the expression may be
274 // null. If it exists, it can be used for forward propagation and reassociation
275 // of values.
276 
277 // For memory, we also track a representative MemoryAccess, and a set of memory
278 // members for MemoryPhis (which have no real instructions). Note that for
279 // memory, it seems tempting to try to split the memory members into a
280 // MemoryCongruenceClass or something. Unfortunately, this does not work
281 // easily. The value numbering of a given memory expression depends on the
282 // leader of the memory congruence class, and the leader of memory congruence
283 // class depends on the value numbering of a given memory expression. This
284 // leads to wasted propagation, and in some cases, missed optimization. For
285 // example: If we had value numbered two stores together before, but now do not,
286 // we move them to a new value congruence class. This in turn will move at one
287 // of the memorydefs to a new memory congruence class. Which in turn, affects
288 // the value numbering of the stores we just value numbered (because the memory
289 // congruence class is part of the value number). So while theoretically
290 // possible to split them up, it turns out to be *incredibly* complicated to get
291 // it to work right, because of the interdependency. While structurally
292 // slightly messier, it is algorithmically much simpler and faster to do what we
293 // do here, and track them both at once in the same class.
294 // Note: The default iterators for this class iterate over values
295 class CongruenceClass {
296 public:
297  using MemberType = Value;
298  using MemberSet = SmallPtrSet<MemberType *, 4>;
299  using MemoryMemberType = MemoryPhi;
300  using MemoryMemberSet = SmallPtrSet<const MemoryMemberType *, 2>;
301 
302  explicit CongruenceClass(unsigned ID) : ID(ID) {}
303  CongruenceClass(unsigned ID, Value *Leader, const Expression *E)
304  : ID(ID), RepLeader(Leader), DefiningExpr(E) {}
305 
306  unsigned getID() const { return ID; }
307 
308  // True if this class has no members left. This is mainly used for assertion
309  // purposes, and for skipping empty classes.
310  bool isDead() const {
311  // If it's both dead from a value perspective, and dead from a memory
312  // perspective, it's really dead.
313  return empty() && memory_empty();
314  }
315 
316  // Leader functions
317  Value *getLeader() const { return RepLeader; }
318  void setLeader(Value *Leader) { RepLeader = Leader; }
319  const std::pair<Value *, unsigned int> &getNextLeader() const {
320  return NextLeader;
321  }
322  void resetNextLeader() { NextLeader = {nullptr, ~0}; }
323  void addPossibleNextLeader(std::pair<Value *, unsigned int> LeaderPair) {
324  if (LeaderPair.second < NextLeader.second)
325  NextLeader = LeaderPair;
326  }
327 
328  Value *getStoredValue() const { return RepStoredValue; }
329  void setStoredValue(Value *Leader) { RepStoredValue = Leader; }
330  const MemoryAccess *getMemoryLeader() const { return RepMemoryAccess; }
331  void setMemoryLeader(const MemoryAccess *Leader) { RepMemoryAccess = Leader; }
332 
333  // Forward propagation info
334  const Expression *getDefiningExpr() const { return DefiningExpr; }
335 
336  // Value member set
337  bool empty() const { return Members.empty(); }
338  unsigned size() const { return Members.size(); }
339  MemberSet::const_iterator begin() const { return Members.begin(); }
340  MemberSet::const_iterator end() const { return Members.end(); }
341  void insert(MemberType *M) { Members.insert(M); }
342  void erase(MemberType *M) { Members.erase(M); }
343  void swap(MemberSet &Other) { Members.swap(Other); }
344 
345  // Memory member set
346  bool memory_empty() const { return MemoryMembers.empty(); }
347  unsigned memory_size() const { return MemoryMembers.size(); }
348  MemoryMemberSet::const_iterator memory_begin() const {
349  return MemoryMembers.begin();
350  }
351  MemoryMemberSet::const_iterator memory_end() const {
352  return MemoryMembers.end();
353  }
355  return make_range(memory_begin(), memory_end());
356  }
357 
358  void memory_insert(const MemoryMemberType *M) { MemoryMembers.insert(M); }
359  void memory_erase(const MemoryMemberType *M) { MemoryMembers.erase(M); }
360 
361  // Store count
362  unsigned getStoreCount() const { return StoreCount; }
363  void incStoreCount() { ++StoreCount; }
364  void decStoreCount() {
365  assert(StoreCount != 0 && "Store count went negative");
366  --StoreCount;
367  }
368 
369  // True if this class has no memory members.
370  bool definesNoMemory() const { return StoreCount == 0 && memory_empty(); }
371 
372  // Return true if two congruence classes are equivalent to each other. This
373  // means that every field but the ID number and the dead field are equivalent.
374  bool isEquivalentTo(const CongruenceClass *Other) const {
375  if (!Other)
376  return false;
377  if (this == Other)
378  return true;
379 
380  if (std::tie(StoreCount, RepLeader, RepStoredValue, RepMemoryAccess) !=
381  std::tie(Other->StoreCount, Other->RepLeader, Other->RepStoredValue,
382  Other->RepMemoryAccess))
383  return false;
384  if (DefiningExpr != Other->DefiningExpr)
385  if (!DefiningExpr || !Other->DefiningExpr ||
386  *DefiningExpr != *Other->DefiningExpr)
387  return false;
388 
389  if (Members.size() != Other->Members.size())
390  return false;
391 
392  return llvm::set_is_subset(Members, Other->Members);
393  }
394 
395 private:
396  unsigned ID;
397 
398  // Representative leader.
399  Value *RepLeader = nullptr;
400 
401  // The most dominating leader after our current leader, because the member set
402  // is not sorted and is expensive to keep sorted all the time.
403  std::pair<Value *, unsigned int> NextLeader = {nullptr, ~0U};
404 
405  // If this is represented by a store, the value of the store.
406  Value *RepStoredValue = nullptr;
407 
408  // If this class contains MemoryDefs or MemoryPhis, this is the leading memory
409  // access.
410  const MemoryAccess *RepMemoryAccess = nullptr;
411 
412  // Defining Expression.
413  const Expression *DefiningExpr = nullptr;
414 
415  // Actual members of this class.
416  MemberSet Members;
417 
418  // This is the set of MemoryPhis that exist in the class. MemoryDefs and
419  // MemoryUses have real instructions representing them, so we only need to
420  // track MemoryPhis here.
421  MemoryMemberSet MemoryMembers;
422 
423  // Number of stores in this congruence class.
424  // This is used so we can detect store equivalence changes properly.
425  int StoreCount = 0;
426 };
427 
428 } // end anonymous namespace
429 
430 namespace llvm {
431 
433  const Expression &E;
434 
435  explicit ExactEqualsExpression(const Expression &E) : E(E) {}
436 
437  hash_code getComputedHash() const { return E.getComputedHash(); }
438 
439  bool operator==(const Expression &Other) const {
440  return E.exactlyEquals(Other);
441  }
442 };
443 
444 template <> struct DenseMapInfo<const Expression *> {
445  static const Expression *getEmptyKey() {
446  auto Val = static_cast<uintptr_t>(-1);
447  Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
448  return reinterpret_cast<const Expression *>(Val);
449  }
450 
451  static const Expression *getTombstoneKey() {
452  auto Val = static_cast<uintptr_t>(~1U);
453  Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
454  return reinterpret_cast<const Expression *>(Val);
455  }
456 
457  static unsigned getHashValue(const Expression *E) {
458  return E->getComputedHash();
459  }
460 
461  static unsigned getHashValue(const ExactEqualsExpression &E) {
462  return E.getComputedHash();
463  }
464 
465  static bool isEqual(const ExactEqualsExpression &LHS, const Expression *RHS) {
466  if (RHS == getTombstoneKey() || RHS == getEmptyKey())
467  return false;
468  return LHS == *RHS;
469  }
470 
471  static bool isEqual(const Expression *LHS, const Expression *RHS) {
472  if (LHS == RHS)
473  return true;
474  if (LHS == getTombstoneKey() || RHS == getTombstoneKey() ||
475  LHS == getEmptyKey() || RHS == getEmptyKey())
476  return false;
477  // Compare hashes before equality. This is *not* what the hashtable does,
478  // since it is computing it modulo the number of buckets, whereas we are
479  // using the full hash keyspace. Since the hashes are precomputed, this
480  // check is *much* faster than equality.
481  if (LHS->getComputedHash() != RHS->getComputedHash())
482  return false;
483  return *LHS == *RHS;
484  }
485 };
486 
487 } // end namespace llvm
488 
489 namespace {
490 
491 class NewGVN {
492  Function &F;
493  DominatorTree *DT = nullptr;
494  const TargetLibraryInfo *TLI = nullptr;
495  AliasAnalysis *AA = nullptr;
496  MemorySSA *MSSA = nullptr;
497  MemorySSAWalker *MSSAWalker = nullptr;
498  AssumptionCache *AC = nullptr;
499  const DataLayout &DL;
500  std::unique_ptr<PredicateInfo> PredInfo;
501 
502  // These are the only two things the create* functions should have
503  // side-effects on due to allocating memory.
504  mutable BumpPtrAllocator ExpressionAllocator;
505  mutable ArrayRecycler<Value *> ArgRecycler;
506  mutable TarjanSCC SCCFinder;
507  const SimplifyQuery SQ;
508 
509  // Number of function arguments, used by ranking
510  unsigned int NumFuncArgs = 0;
511 
512  // RPOOrdering of basic blocks
514 
515  // Congruence class info.
516 
517  // This class is called INITIAL in the paper. It is the class everything
518  // startsout in, and represents any value. Being an optimistic analysis,
519  // anything in the TOP class has the value TOP, which is indeterminate and
520  // equivalent to everything.
521  CongruenceClass *TOPClass = nullptr;
522  std::vector<CongruenceClass *> CongruenceClasses;
523  unsigned NextCongruenceNum = 0;
524 
525  // Value Mappings.
527  DenseMap<Value *, const Expression *> ValueToExpression;
528 
529  // Value PHI handling, used to make equivalence between phi(op, op) and
530  // op(phi, phi).
531  // These mappings just store various data that would normally be part of the
532  // IR.
534 
535  DenseMap<const Value *, bool> OpSafeForPHIOfOps;
536 
537  // Map a temporary instruction we created to a parent block.
539 
540  // Map between the already in-program instructions and the temporary phis we
541  // created that they are known equivalent to.
543 
544  // In order to know when we should re-process instructions that have
545  // phi-of-ops, we track the set of expressions that they needed as
546  // leaders. When we discover new leaders for those expressions, we process the
547  // associated phi-of-op instructions again in case they have changed. The
548  // other way they may change is if they had leaders, and those leaders
549  // disappear. However, at the point they have leaders, there are uses of the
550  // relevant operands in the created phi node, and so they will get reprocessed
551  // through the normal user marking we perform.
552  mutable DenseMap<const Value *, SmallPtrSet<Value *, 2>> AdditionalUsers;
554  ExpressionToPhiOfOps;
555 
556  // Map from temporary operation to MemoryAccess.
558 
559  // Set of all temporary instructions we created.
560  // Note: This will include instructions that were just created during value
561  // numbering. The way to test if something is using them is to check
562  // RealToTemp.
563  DenseSet<Instruction *> AllTempInstructions;
564 
565  // This is the set of instructions to revisit on a reachability change. At
566  // the end of the main iteration loop it will contain at least all the phi of
567  // ops instructions that will be changed to phis, as well as regular phis.
568  // During the iteration loop, it may contain other things, such as phi of ops
569  // instructions that used edge reachability to reach a result, and so need to
570  // be revisited when the edge changes, independent of whether the phi they
571  // depended on changes.
572  DenseMap<BasicBlock *, SparseBitVector<>> RevisitOnReachabilityChange;
573 
574  // Mapping from predicate info we used to the instructions we used it with.
575  // In order to correctly ensure propagation, we must keep track of what
576  // comparisons we used, so that when the values of the comparisons change, we
577  // propagate the information to the places we used the comparison.
579  PredicateToUsers;
580 
581  // the same reasoning as PredicateToUsers. When we skip MemoryAccesses for
582  // stores, we no longer can rely solely on the def-use chains of MemorySSA.
584  MemoryToUsers;
585 
586  // A table storing which memorydefs/phis represent a memory state provably
587  // equivalent to another memory state.
588  // We could use the congruence class machinery, but the MemoryAccess's are
589  // abstract memory states, so they can only ever be equivalent to each other,
590  // and not to constants, etc.
592 
593  // We could, if we wanted, build MemoryPhiExpressions and
594  // MemoryVariableExpressions, etc, and value number them the same way we value
595  // number phi expressions. For the moment, this seems like overkill. They
596  // can only exist in one of three states: they can be TOP (equal to
597  // everything), Equivalent to something else, or unique. Because we do not
598  // create expressions for them, we need to simulate leader change not just
599  // when they change class, but when they change state. Note: We can do the
600  // same thing for phis, and avoid having phi expressions if we wanted, We
601  // should eventually unify in one direction or the other, so this is a little
602  // bit of an experiment in which turns out easier to maintain.
603  enum MemoryPhiState { MPS_Invalid, MPS_TOP, MPS_Equivalent, MPS_Unique };
605 
606  enum InstCycleState { ICS_Unknown, ICS_CycleFree, ICS_Cycle };
607  mutable DenseMap<const Instruction *, InstCycleState> InstCycleState;
608 
609  // Expression to class mapping.
610  using ExpressionClassMap = DenseMap<const Expression *, CongruenceClass *>;
611  ExpressionClassMap ExpressionToClass;
612 
613  // We have a single expression that represents currently DeadExpressions.
614  // For dead expressions we can prove will stay dead, we mark them with
615  // DFS number zero. However, it's possible in the case of phi nodes
616  // for us to assume/prove all arguments are dead during fixpointing.
617  // We use DeadExpression for that case.
618  DeadExpression *SingletonDeadExpression = nullptr;
619 
620  // Which values have changed as a result of leader changes.
621  SmallPtrSet<Value *, 8> LeaderChanges;
622 
623  // Reachability info.
624  using BlockEdge = BasicBlockEdge;
625  DenseSet<BlockEdge> ReachableEdges;
626  SmallPtrSet<const BasicBlock *, 8> ReachableBlocks;
627 
628  // This is a bitvector because, on larger functions, we may have
629  // thousands of touched instructions at once (entire blocks,
630  // instructions with hundreds of uses, etc). Even with optimization
631  // for when we mark whole blocks as touched, when this was a
632  // SmallPtrSet or DenseSet, for some functions, we spent >20% of all
633  // the time in GVN just managing this list. The bitvector, on the
634  // other hand, efficiently supports test/set/clear of both
635  // individual and ranges, as well as "find next element" This
636  // enables us to use it as a worklist with essentially 0 cost.
637  BitVector TouchedInstructions;
638 
640  mutable DenseMap<const IntrinsicInst *, const Value *> IntrinsicInstPred;
641 
642 #ifndef NDEBUG
643  // Debugging for how many times each block and instruction got processed.
644  DenseMap<const Value *, unsigned> ProcessedCount;
645 #endif
646 
647  // DFS info.
648  // This contains a mapping from Instructions to DFS numbers.
649  // The numbering starts at 1. An instruction with DFS number zero
650  // means that the instruction is dead.
652 
653  // This contains the mapping DFS numbers to instructions.
654  SmallVector<Value *, 32> DFSToInstr;
655 
656  // Deletion info.
657  SmallPtrSet<Instruction *, 8> InstructionsToErase;
658 
659 public:
660  NewGVN(Function &F, DominatorTree *DT, AssumptionCache *AC,
662  const DataLayout &DL)
663  : F(F), DT(DT), TLI(TLI), AA(AA), MSSA(MSSA), AC(AC), DL(DL),
664  PredInfo(std::make_unique<PredicateInfo>(F, *DT, *AC)),
665  SQ(DL, TLI, DT, AC, /*CtxI=*/nullptr, /*UseInstrInfo=*/false,
666  /*CanUseUndef=*/false) {}
667 
668  bool runGVN();
669 
670 private:
671  /// Helper struct return a Expression with an optional extra dependency.
672  struct ExprResult {
673  const Expression *Expr;
674  Value *ExtraDep;
675  const PredicateBase *PredDep;
676 
677  ExprResult(const Expression *Expr, Value *ExtraDep = nullptr,
678  const PredicateBase *PredDep = nullptr)
679  : Expr(Expr), ExtraDep(ExtraDep), PredDep(PredDep) {}
680  ExprResult(const ExprResult &) = delete;
681  ExprResult(ExprResult &&Other)
682  : Expr(Other.Expr), ExtraDep(Other.ExtraDep), PredDep(Other.PredDep) {
683  Other.Expr = nullptr;
684  Other.ExtraDep = nullptr;
685  Other.PredDep = nullptr;
686  }
687  ExprResult &operator=(const ExprResult &Other) = delete;
688  ExprResult &operator=(ExprResult &&Other) = delete;
689 
690  ~ExprResult() { assert(!ExtraDep && "unhandled ExtraDep"); }
691 
692  operator bool() const { return Expr; }
693 
694  static ExprResult none() { return {nullptr, nullptr, nullptr}; }
695  static ExprResult some(const Expression *Expr, Value *ExtraDep = nullptr) {
696  return {Expr, ExtraDep, nullptr};
697  }
698  static ExprResult some(const Expression *Expr,
699  const PredicateBase *PredDep) {
700  return {Expr, nullptr, PredDep};
701  }
702  static ExprResult some(const Expression *Expr, Value *ExtraDep,
703  const PredicateBase *PredDep) {
704  return {Expr, ExtraDep, PredDep};
705  }
706  };
707 
708  // Expression handling.
709  ExprResult createExpression(Instruction *) const;
710  const Expression *createBinaryExpression(unsigned, Type *, Value *, Value *,
711  Instruction *) const;
712 
713  // Our canonical form for phi arguments is a pair of incoming value, incoming
714  // basic block.
715  using ValPair = std::pair<Value *, BasicBlock *>;
716 
717  PHIExpression *createPHIExpression(ArrayRef<ValPair>, const Instruction *,
718  BasicBlock *, bool &HasBackEdge,
719  bool &OriginalOpsConstant) const;
720  const DeadExpression *createDeadExpression() const;
721  const VariableExpression *createVariableExpression(Value *) const;
722  const ConstantExpression *createConstantExpression(Constant *) const;
723  const Expression *createVariableOrConstant(Value *V) const;
724  const UnknownExpression *createUnknownExpression(Instruction *) const;
725  const StoreExpression *createStoreExpression(StoreInst *,
726  const MemoryAccess *) const;
727  LoadExpression *createLoadExpression(Type *, Value *, LoadInst *,
728  const MemoryAccess *) const;
729  const CallExpression *createCallExpression(CallInst *,
730  const MemoryAccess *) const;
732  createAggregateValueExpression(Instruction *) const;
733  bool setBasicExpressionInfo(Instruction *, BasicExpression *) const;
734 
735  // Congruence class handling.
736  CongruenceClass *createCongruenceClass(Value *Leader, const Expression *E) {
737  auto *result = new CongruenceClass(NextCongruenceNum++, Leader, E);
738  CongruenceClasses.emplace_back(result);
739  return result;
740  }
741 
742  CongruenceClass *createMemoryClass(MemoryAccess *MA) {
743  auto *CC = createCongruenceClass(nullptr, nullptr);
744  CC->setMemoryLeader(MA);
745  return CC;
746  }
747 
748  CongruenceClass *ensureLeaderOfMemoryClass(MemoryAccess *MA) {
749  auto *CC = getMemoryClass(MA);
750  if (CC->getMemoryLeader() != MA)
751  CC = createMemoryClass(MA);
752  return CC;
753  }
754 
755  CongruenceClass *createSingletonCongruenceClass(Value *Member) {
756  CongruenceClass *CClass = createCongruenceClass(Member, nullptr);
757  CClass->insert(Member);
758  ValueToClass[Member] = CClass;
759  return CClass;
760  }
761 
762  void initializeCongruenceClasses(Function &F);
763  const Expression *makePossiblePHIOfOps(Instruction *,
765  Value *findLeaderForInst(Instruction *ValueOp,
766  SmallPtrSetImpl<Value *> &Visited,
767  MemoryAccess *MemAccess, Instruction *OrigInst,
768  BasicBlock *PredBB);
769  bool OpIsSafeForPHIOfOpsHelper(Value *V, const BasicBlock *PHIBlock,
772  bool OpIsSafeForPHIOfOps(Value *Op, const BasicBlock *PHIBlock,
774  void addPhiOfOps(PHINode *Op, BasicBlock *BB, Instruction *ExistingValue);
775  void removePhiOfOps(Instruction *I, PHINode *PHITemp);
776 
777  // Value number an Instruction or MemoryPhi.
778  void valueNumberMemoryPhi(MemoryPhi *);
779  void valueNumberInstruction(Instruction *);
780 
781  // Symbolic evaluation.
782  ExprResult checkExprResults(Expression *, Instruction *, Value *) const;
783  ExprResult performSymbolicEvaluation(Value *,
784  SmallPtrSetImpl<Value *> &) const;
785  const Expression *performSymbolicLoadCoercion(Type *, Value *, LoadInst *,
786  Instruction *,
787  MemoryAccess *) const;
788  const Expression *performSymbolicLoadEvaluation(Instruction *) const;
789  const Expression *performSymbolicStoreEvaluation(Instruction *) const;
790  ExprResult performSymbolicCallEvaluation(Instruction *) const;
791  void sortPHIOps(MutableArrayRef<ValPair> Ops) const;
792  const Expression *performSymbolicPHIEvaluation(ArrayRef<ValPair>,
793  Instruction *I,
794  BasicBlock *PHIBlock) const;
795  const Expression *performSymbolicAggrValueEvaluation(Instruction *) const;
796  ExprResult performSymbolicCmpEvaluation(Instruction *) const;
797  ExprResult performSymbolicPredicateInfoEvaluation(IntrinsicInst *) const;
798 
799  // Congruence finding.
800  bool someEquivalentDominates(const Instruction *, const Instruction *) const;
801  Value *lookupOperandLeader(Value *) const;
802  CongruenceClass *getClassForExpression(const Expression *E) const;
803  void performCongruenceFinding(Instruction *, const Expression *);
804  void moveValueToNewCongruenceClass(Instruction *, const Expression *,
805  CongruenceClass *, CongruenceClass *);
806  void moveMemoryToNewCongruenceClass(Instruction *, MemoryAccess *,
807  CongruenceClass *, CongruenceClass *);
808  Value *getNextValueLeader(CongruenceClass *) const;
809  const MemoryAccess *getNextMemoryLeader(CongruenceClass *) const;
810  bool setMemoryClass(const MemoryAccess *From, CongruenceClass *To);
811  CongruenceClass *getMemoryClass(const MemoryAccess *MA) const;
812  const MemoryAccess *lookupMemoryLeader(const MemoryAccess *) const;
813  bool isMemoryAccessTOP(const MemoryAccess *) const;
814 
815  // Ranking
816  unsigned int getRank(const Value *) const;
817  bool shouldSwapOperands(const Value *, const Value *) const;
818  bool shouldSwapOperandsForIntrinsic(const Value *, const Value *,
819  const IntrinsicInst *I) const;
820 
821  // Reachability handling.
822  void updateReachableEdge(BasicBlock *, BasicBlock *);
823  void processOutgoingEdges(Instruction *, BasicBlock *);
824  Value *findConditionEquivalence(Value *) const;
825 
826  // Elimination.
827  struct ValueDFS;
828  void convertClassToDFSOrdered(const CongruenceClass &,
832  void convertClassToLoadsAndStores(const CongruenceClass &,
833  SmallVectorImpl<ValueDFS> &) const;
834 
835  bool eliminateInstructions(Function &);
836  void replaceInstruction(Instruction *, Value *);
837  void markInstructionForDeletion(Instruction *);
838  void deleteInstructionsInBlock(BasicBlock *);
839  Value *findPHIOfOpsLeader(const Expression *, const Instruction *,
840  const BasicBlock *) const;
841 
842  // Various instruction touch utilities
843  template <typename Map, typename KeyType>
844  void touchAndErase(Map &, const KeyType &);
845  void markUsersTouched(Value *);
846  void markMemoryUsersTouched(const MemoryAccess *);
847  void markMemoryDefTouched(const MemoryAccess *);
848  void markPredicateUsersTouched(Instruction *);
849  void markValueLeaderChangeTouched(CongruenceClass *CC);
850  void markMemoryLeaderChangeTouched(CongruenceClass *CC);
851  void markPhiOfOpsChanged(const Expression *E);
852  void addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const;
853  void addAdditionalUsers(Value *To, Value *User) const;
854  void addAdditionalUsers(ExprResult &Res, Instruction *User) const;
855 
856  // Main loop of value numbering
857  void iterateTouchedInstructions();
858 
859  // Utilities.
860  void cleanupTables();
861  std::pair<unsigned, unsigned> assignDFSNumbers(BasicBlock *, unsigned);
862  void updateProcessedCount(const Value *V);
863  void verifyMemoryCongruency() const;
864  void verifyIterationSettled(Function &F);
865  void verifyStoreExpressions() const;
866  bool singleReachablePHIPath(SmallPtrSet<const MemoryAccess *, 8> &,
867  const MemoryAccess *, const MemoryAccess *) const;
868  BasicBlock *getBlockForValue(Value *V) const;
869  void deleteExpression(const Expression *E) const;
870  MemoryUseOrDef *getMemoryAccess(const Instruction *) const;
871  MemoryPhi *getMemoryAccess(const BasicBlock *) const;
872  template <class T, class Range> T *getMinDFSOfRange(const Range &) const;
873 
874  unsigned InstrToDFSNum(const Value *V) const {
875  assert(isa<Instruction>(V) && "This should not be used for MemoryAccesses");
876  return InstrDFS.lookup(V);
877  }
878 
879  unsigned InstrToDFSNum(const MemoryAccess *MA) const {
880  return MemoryToDFSNum(MA);
881  }
882 
883  Value *InstrFromDFSNum(unsigned DFSNum) { return DFSToInstr[DFSNum]; }
884 
885  // Given a MemoryAccess, return the relevant instruction DFS number. Note:
886  // This deliberately takes a value so it can be used with Use's, which will
887  // auto-convert to Value's but not to MemoryAccess's.
888  unsigned MemoryToDFSNum(const Value *MA) const {
889  assert(isa<MemoryAccess>(MA) &&
890  "This should not be used with instructions");
891  return isa<MemoryUseOrDef>(MA)
892  ? InstrToDFSNum(cast<MemoryUseOrDef>(MA)->getMemoryInst())
893  : InstrDFS.lookup(MA);
894  }
895 
896  bool isCycleFree(const Instruction *) const;
897  bool isBackedge(BasicBlock *From, BasicBlock *To) const;
898 
899  // Debug counter info. When verifying, we have to reset the value numbering
900  // debug counter to the same state it started in to get the same results.
901  int64_t StartingVNCounter = 0;
902 };
903 
904 } // end anonymous namespace
905 
906 template <typename T>
907 static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS) {
908  if (!isa<LoadExpression>(RHS) && !isa<StoreExpression>(RHS))
909  return false;
910  return LHS.MemoryExpression::equals(RHS);
911 }
912 
914  return equalsLoadStoreHelper(*this, Other);
915 }
916 
918  if (!equalsLoadStoreHelper(*this, Other))
919  return false;
920  // Make sure that store vs store includes the value operand.
921  if (const auto *S = dyn_cast<StoreExpression>(&Other))
922  if (getStoredValue() != S->getStoredValue())
923  return false;
924  return true;
925 }
926 
927 // Determine if the edge From->To is a backedge
928 bool NewGVN::isBackedge(BasicBlock *From, BasicBlock *To) const {
929  return From == To ||
930  RPOOrdering.lookup(DT->getNode(From)) >=
931  RPOOrdering.lookup(DT->getNode(To));
932 }
933 
934 #ifndef NDEBUG
935 static std::string getBlockName(const BasicBlock *B) {
937 }
938 #endif
939 
940 // Get a MemoryAccess for an instruction, fake or real.
941 MemoryUseOrDef *NewGVN::getMemoryAccess(const Instruction *I) const {
942  auto *Result = MSSA->getMemoryAccess(I);
943  return Result ? Result : TempToMemory.lookup(I);
944 }
945 
946 // Get a MemoryPhi for a basic block. These are all real.
947 MemoryPhi *NewGVN::getMemoryAccess(const BasicBlock *BB) const {
948  return MSSA->getMemoryAccess(BB);
949 }
950 
951 // Get the basic block from an instruction/memory value.
952 BasicBlock *NewGVN::getBlockForValue(Value *V) const {
953  if (auto *I = dyn_cast<Instruction>(V)) {
954  auto *Parent = I->getParent();
955  if (Parent)
956  return Parent;
957  Parent = TempToBlock.lookup(V);
958  assert(Parent && "Every fake instruction should have a block");
959  return Parent;
960  }
961 
962  auto *MP = dyn_cast<MemoryPhi>(V);
963  assert(MP && "Should have been an instruction or a MemoryPhi");
964  return MP->getBlock();
965 }
966 
967 // Delete a definitely dead expression, so it can be reused by the expression
968 // allocator. Some of these are not in creation functions, so we have to accept
969 // const versions.
970 void NewGVN::deleteExpression(const Expression *E) const {
971  assert(isa<BasicExpression>(E));
972  auto *BE = cast<BasicExpression>(E);
973  const_cast<BasicExpression *>(BE)->deallocateOperands(ArgRecycler);
974  ExpressionAllocator.Deallocate(E);
975 }
976 
977 // If V is a predicateinfo copy, get the thing it is a copy of.
978 static Value *getCopyOf(const Value *V) {
979  if (auto *II = dyn_cast<IntrinsicInst>(V))
980  if (II->getIntrinsicID() == Intrinsic::ssa_copy)
981  return II->getOperand(0);
982  return nullptr;
983 }
984 
985 // Return true if V is really PN, even accounting for predicateinfo copies.
986 static bool isCopyOfPHI(const Value *V, const PHINode *PN) {
987  return V == PN || getCopyOf(V) == PN;
988 }
989 
990 static bool isCopyOfAPHI(const Value *V) {
991  auto *CO = getCopyOf(V);
992  return CO && isa<PHINode>(CO);
993 }
994 
995 // Sort PHI Operands into a canonical order. What we use here is an RPO
996 // order. The BlockInstRange numbers are generated in an RPO walk of the basic
997 // blocks.
998 void NewGVN::sortPHIOps(MutableArrayRef<ValPair> Ops) const {
999  llvm::sort(Ops, [&](const ValPair &P1, const ValPair &P2) {
1000  return BlockInstRange.lookup(P1.second).first <
1001  BlockInstRange.lookup(P2.second).first;
1002  });
1003 }
1004 
1005 // Return true if V is a value that will always be available (IE can
1006 // be placed anywhere) in the function. We don't do globals here
1007 // because they are often worse to put in place.
1008 static bool alwaysAvailable(Value *V) {
1009  return isa<Constant>(V) || isa<Argument>(V);
1010 }
1011 
1012 // Create a PHIExpression from an array of {incoming edge, value} pairs. I is
1013 // the original instruction we are creating a PHIExpression for (but may not be
1014 // a phi node). We require, as an invariant, that all the PHIOperands in the
1015 // same block are sorted the same way. sortPHIOps will sort them into a
1016 // canonical order.
1017 PHIExpression *NewGVN::createPHIExpression(ArrayRef<ValPair> PHIOperands,
1018  const Instruction *I,
1019  BasicBlock *PHIBlock,
1020  bool &HasBackedge,
1021  bool &OriginalOpsConstant) const {
1022  unsigned NumOps = PHIOperands.size();
1023  auto *E = new (ExpressionAllocator) PHIExpression(NumOps, PHIBlock);
1024 
1025  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1026  E->setType(PHIOperands.begin()->first->getType());
1027  E->setOpcode(Instruction::PHI);
1028 
1029  // Filter out unreachable phi operands.
1030  auto Filtered = make_filter_range(PHIOperands, [&](const ValPair &P) {
1031  auto *BB = P.second;
1032  if (auto *PHIOp = dyn_cast<PHINode>(I))
1033  if (isCopyOfPHI(P.first, PHIOp))
1034  return false;
1035  if (!ReachableEdges.count({BB, PHIBlock}))
1036  return false;
1037  // Things in TOPClass are equivalent to everything.
1038  if (ValueToClass.lookup(P.first) == TOPClass)
1039  return false;
1040  OriginalOpsConstant = OriginalOpsConstant && isa<Constant>(P.first);
1041  HasBackedge = HasBackedge || isBackedge(BB, PHIBlock);
1042  return lookupOperandLeader(P.first) != I;
1043  });
1044  std::transform(Filtered.begin(), Filtered.end(), op_inserter(E),
1045  [&](const ValPair &P) -> Value * {
1046  return lookupOperandLeader(P.first);
1047  });
1048  return E;
1049 }
1050 
1051 // Set basic expression info (Arguments, type, opcode) for Expression
1052 // E from Instruction I in block B.
1053 bool NewGVN::setBasicExpressionInfo(Instruction *I, BasicExpression *E) const {
1054  bool AllConstant = true;
1055  if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
1056  E->setType(GEP->getSourceElementType());
1057  else
1058  E->setType(I->getType());
1059  E->setOpcode(I->getOpcode());
1060  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1061 
1062  // Transform the operand array into an operand leader array, and keep track of
1063  // whether all members are constant.
1064  std::transform(I->op_begin(), I->op_end(), op_inserter(E), [&](Value *O) {
1065  auto Operand = lookupOperandLeader(O);
1066  AllConstant = AllConstant && isa<Constant>(Operand);
1067  return Operand;
1068  });
1069 
1070  return AllConstant;
1071 }
1072 
1073 const Expression *NewGVN::createBinaryExpression(unsigned Opcode, Type *T,
1074  Value *Arg1, Value *Arg2,
1075  Instruction *I) const {
1076  auto *E = new (ExpressionAllocator) BasicExpression(2);
1077 
1078  E->setType(T);
1079  E->setOpcode(Opcode);
1080  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1081  if (Instruction::isCommutative(Opcode)) {
1082  // Ensure that commutative instructions that only differ by a permutation
1083  // of their operands get the same value number by sorting the operand value
1084  // numbers. Since all commutative instructions have two operands it is more
1085  // efficient to sort by hand rather than using, say, std::sort.
1086  if (shouldSwapOperands(Arg1, Arg2))
1087  std::swap(Arg1, Arg2);
1088  }
1089  E->op_push_back(lookupOperandLeader(Arg1));
1090  E->op_push_back(lookupOperandLeader(Arg2));
1091 
1092  Value *V = SimplifyBinOp(Opcode, E->getOperand(0), E->getOperand(1), SQ);
1093  if (auto Simplified = checkExprResults(E, I, V)) {
1094  addAdditionalUsers(Simplified, I);
1095  return Simplified.Expr;
1096  }
1097  return E;
1098 }
1099 
1100 // Take a Value returned by simplification of Expression E/Instruction
1101 // I, and see if it resulted in a simpler expression. If so, return
1102 // that expression.
1103 NewGVN::ExprResult NewGVN::checkExprResults(Expression *E, Instruction *I,
1104  Value *V) const {
1105  if (!V)
1106  return ExprResult::none();
1107 
1108  if (auto *C = dyn_cast<Constant>(V)) {
1109  if (I)
1110  LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1111  << " constant " << *C << "\n");
1112  NumGVNOpsSimplified++;
1113  assert(isa<BasicExpression>(E) &&
1114  "We should always have had a basic expression here");
1115  deleteExpression(E);
1116  return ExprResult::some(createConstantExpression(C));
1117  } else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1118  if (I)
1119  LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1120  << " variable " << *V << "\n");
1121  deleteExpression(E);
1122  return ExprResult::some(createVariableExpression(V));
1123  }
1124 
1125  CongruenceClass *CC = ValueToClass.lookup(V);
1126  if (CC) {
1127  if (CC->getLeader() && CC->getLeader() != I) {
1128  return ExprResult::some(createVariableOrConstant(CC->getLeader()), V);
1129  }
1130  if (CC->getDefiningExpr()) {
1131  if (I)
1132  LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1133  << " expression " << *CC->getDefiningExpr() << "\n");
1134  NumGVNOpsSimplified++;
1135  deleteExpression(E);
1136  return ExprResult::some(CC->getDefiningExpr(), V);
1137  }
1138  }
1139 
1140  return ExprResult::none();
1141 }
1142 
1143 // Create a value expression from the instruction I, replacing operands with
1144 // their leaders.
1145 
1146 NewGVN::ExprResult NewGVN::createExpression(Instruction *I) const {
1147  auto *E = new (ExpressionAllocator) BasicExpression(I->getNumOperands());
1148 
1149  bool AllConstant = setBasicExpressionInfo(I, E);
1150 
1151  if (I->isCommutative()) {
1152  // Ensure that commutative instructions that only differ by a permutation
1153  // of their operands get the same value number by sorting the operand value
1154  // numbers. Since all commutative instructions have two operands it is more
1155  // efficient to sort by hand rather than using, say, std::sort.
1156  assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
1157  if (shouldSwapOperands(E->getOperand(0), E->getOperand(1)))
1158  E->swapOperands(0, 1);
1159  }
1160  // Perform simplification.
1161  if (auto *CI = dyn_cast<CmpInst>(I)) {
1162  // Sort the operand value numbers so x<y and y>x get the same value
1163  // number.
1164  CmpInst::Predicate Predicate = CI->getPredicate();
1165  if (shouldSwapOperands(E->getOperand(0), E->getOperand(1))) {
1166  E->swapOperands(0, 1);
1168  }
1169  E->setOpcode((CI->getOpcode() << 8) | Predicate);
1170  // TODO: 25% of our time is spent in SimplifyCmpInst with pointer operands
1171  assert(I->getOperand(0)->getType() == I->getOperand(1)->getType() &&
1172  "Wrong types on cmp instruction");
1173  assert((E->getOperand(0)->getType() == I->getOperand(0)->getType() &&
1174  E->getOperand(1)->getType() == I->getOperand(1)->getType()));
1175  Value *V =
1176  SimplifyCmpInst(Predicate, E->getOperand(0), E->getOperand(1), SQ);
1177  if (auto Simplified = checkExprResults(E, I, V))
1178  return Simplified;
1179  } else if (isa<SelectInst>(I)) {
1180  if (isa<Constant>(E->getOperand(0)) ||
1181  E->getOperand(1) == E->getOperand(2)) {
1182  assert(E->getOperand(1)->getType() == I->getOperand(1)->getType() &&
1183  E->getOperand(2)->getType() == I->getOperand(2)->getType());
1184  Value *V = SimplifySelectInst(E->getOperand(0), E->getOperand(1),
1185  E->getOperand(2), SQ);
1186  if (auto Simplified = checkExprResults(E, I, V))
1187  return Simplified;
1188  }
1189  } else if (I->isBinaryOp()) {
1190  Value *V =
1191  SimplifyBinOp(E->getOpcode(), E->getOperand(0), E->getOperand(1), SQ);
1192  if (auto Simplified = checkExprResults(E, I, V))
1193  return Simplified;
1194  } else if (auto *CI = dyn_cast<CastInst>(I)) {
1195  Value *V =
1196  SimplifyCastInst(CI->getOpcode(), E->getOperand(0), CI->getType(), SQ);
1197  if (auto Simplified = checkExprResults(E, I, V))
1198  return Simplified;
1199  } else if (auto *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1200  Value *V =
1201  SimplifyGEPInst(GEPI->getSourceElementType(), *E->op_begin(),
1202  makeArrayRef(std::next(E->op_begin()), E->op_end()),
1203  GEPI->isInBounds(), SQ);
1204  if (auto Simplified = checkExprResults(E, I, V))
1205  return Simplified;
1206  } else if (AllConstant) {
1207  // We don't bother trying to simplify unless all of the operands
1208  // were constant.
1209  // TODO: There are a lot of Simplify*'s we could call here, if we
1210  // wanted to. The original motivating case for this code was a
1211  // zext i1 false to i8, which we don't have an interface to
1212  // simplify (IE there is no SimplifyZExt).
1213 
1215  for (Value *Arg : E->operands())
1216  C.emplace_back(cast<Constant>(Arg));
1217 
1218  if (Value *V = ConstantFoldInstOperands(I, C, DL, TLI))
1219  if (auto Simplified = checkExprResults(E, I, V))
1220  return Simplified;
1221  }
1222  return ExprResult::some(E);
1223 }
1224 
1226 NewGVN::createAggregateValueExpression(Instruction *I) const {
1227  if (auto *II = dyn_cast<InsertValueInst>(I)) {
1228  auto *E = new (ExpressionAllocator)
1229  AggregateValueExpression(I->getNumOperands(), II->getNumIndices());
1230  setBasicExpressionInfo(I, E);
1231  E->allocateIntOperands(ExpressionAllocator);
1232  std::copy(II->idx_begin(), II->idx_end(), int_op_inserter(E));
1233  return E;
1234  } else if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1235  auto *E = new (ExpressionAllocator)
1236  AggregateValueExpression(I->getNumOperands(), EI->getNumIndices());
1237  setBasicExpressionInfo(EI, E);
1238  E->allocateIntOperands(ExpressionAllocator);
1239  std::copy(EI->idx_begin(), EI->idx_end(), int_op_inserter(E));
1240  return E;
1241  }
1242  llvm_unreachable("Unhandled type of aggregate value operation");
1243 }
1244 
1245 const DeadExpression *NewGVN::createDeadExpression() const {
1246  // DeadExpression has no arguments and all DeadExpression's are the same,
1247  // so we only need one of them.
1248  return SingletonDeadExpression;
1249 }
1250 
1251 const VariableExpression *NewGVN::createVariableExpression(Value *V) const {
1252  auto *E = new (ExpressionAllocator) VariableExpression(V);
1253  E->setOpcode(V->getValueID());
1254  return E;
1255 }
1256 
1257 const Expression *NewGVN::createVariableOrConstant(Value *V) const {
1258  if (auto *C = dyn_cast<Constant>(V))
1259  return createConstantExpression(C);
1260  return createVariableExpression(V);
1261 }
1262 
1263 const ConstantExpression *NewGVN::createConstantExpression(Constant *C) const {
1264  auto *E = new (ExpressionAllocator) ConstantExpression(C);
1265  E->setOpcode(C->getValueID());
1266  return E;
1267 }
1268 
1269 const UnknownExpression *NewGVN::createUnknownExpression(Instruction *I) const {
1270  auto *E = new (ExpressionAllocator) UnknownExpression(I);
1271  E->setOpcode(I->getOpcode());
1272  return E;
1273 }
1274 
1275 const CallExpression *
1276 NewGVN::createCallExpression(CallInst *CI, const MemoryAccess *MA) const {
1277  // FIXME: Add operand bundles for calls.
1278  // FIXME: Allow commutative matching for intrinsics.
1279  auto *E =
1280  new (ExpressionAllocator) CallExpression(CI->getNumOperands(), CI, MA);
1281  setBasicExpressionInfo(CI, E);
1282  return E;
1283 }
1284 
1285 // Return true if some equivalent of instruction Inst dominates instruction U.
1286 bool NewGVN::someEquivalentDominates(const Instruction *Inst,
1287  const Instruction *U) const {
1288  auto *CC = ValueToClass.lookup(Inst);
1289  // This must be an instruction because we are only called from phi nodes
1290  // in the case that the value it needs to check against is an instruction.
1291 
1292  // The most likely candidates for dominance are the leader and the next leader.
1293  // The leader or nextleader will dominate in all cases where there is an
1294  // equivalent that is higher up in the dom tree.
1295  // We can't *only* check them, however, because the
1296  // dominator tree could have an infinite number of non-dominating siblings
1297  // with instructions that are in the right congruence class.
1298  // A
1299  // B C D E F G
1300  // |
1301  // H
1302  // Instruction U could be in H, with equivalents in every other sibling.
1303  // Depending on the rpo order picked, the leader could be the equivalent in
1304  // any of these siblings.
1305  if (!CC)
1306  return false;
1307  if (alwaysAvailable(CC->getLeader()))
1308  return true;
1309  if (DT->dominates(cast<Instruction>(CC->getLeader()), U))
1310  return true;
1311  if (CC->getNextLeader().first &&
1312  DT->dominates(cast<Instruction>(CC->getNextLeader().first), U))
1313  return true;
1314  return llvm::any_of(*CC, [&](const Value *Member) {
1315  return Member != CC->getLeader() &&
1316  DT->dominates(cast<Instruction>(Member), U);
1317  });
1318 }
1319 
1320 // See if we have a congruence class and leader for this operand, and if so,
1321 // return it. Otherwise, return the operand itself.
1322 Value *NewGVN::lookupOperandLeader(Value *V) const {
1323  CongruenceClass *CC = ValueToClass.lookup(V);
1324  if (CC) {
1325  // Everything in TOP is represented by poison, as it can be any value.
1326  // We do have to make sure we get the type right though, so we can't set the
1327  // RepLeader to poison.
1328  if (CC == TOPClass)
1329  return PoisonValue::get(V->getType());
1330  return CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
1331  }
1332 
1333  return V;
1334 }
1335 
1336 const MemoryAccess *NewGVN::lookupMemoryLeader(const MemoryAccess *MA) const {
1337  auto *CC = getMemoryClass(MA);
1338  assert(CC->getMemoryLeader() &&
1339  "Every MemoryAccess should be mapped to a congruence class with a "
1340  "representative memory access");
1341  return CC->getMemoryLeader();
1342 }
1343 
1344 // Return true if the MemoryAccess is really equivalent to everything. This is
1345 // equivalent to the lattice value "TOP" in most lattices. This is the initial
1346 // state of all MemoryAccesses.
1347 bool NewGVN::isMemoryAccessTOP(const MemoryAccess *MA) const {
1348  return getMemoryClass(MA) == TOPClass;
1349 }
1350 
1351 LoadExpression *NewGVN::createLoadExpression(Type *LoadType, Value *PointerOp,
1352  LoadInst *LI,
1353  const MemoryAccess *MA) const {
1354  auto *E =
1355  new (ExpressionAllocator) LoadExpression(1, LI, lookupMemoryLeader(MA));
1356  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1357  E->setType(LoadType);
1358 
1359  // Give store and loads same opcode so they value number together.
1360  E->setOpcode(0);
1361  E->op_push_back(PointerOp);
1362 
1363  // TODO: Value number heap versions. We may be able to discover
1364  // things alias analysis can't on it's own (IE that a store and a
1365  // load have the same value, and thus, it isn't clobbering the load).
1366  return E;
1367 }
1368 
1369 const StoreExpression *
1370 NewGVN::createStoreExpression(StoreInst *SI, const MemoryAccess *MA) const {
1371  auto *StoredValueLeader = lookupOperandLeader(SI->getValueOperand());
1372  auto *E = new (ExpressionAllocator)
1373  StoreExpression(SI->getNumOperands(), SI, StoredValueLeader, MA);
1374  E->allocateOperands(ArgRecycler, ExpressionAllocator);
1375  E->setType(SI->getValueOperand()->getType());
1376 
1377  // Give store and loads same opcode so they value number together.
1378  E->setOpcode(0);
1379  E->op_push_back(lookupOperandLeader(SI->getPointerOperand()));
1380 
1381  // TODO: Value number heap versions. We may be able to discover
1382  // things alias analysis can't on it's own (IE that a store and a
1383  // load have the same value, and thus, it isn't clobbering the load).
1384  return E;
1385 }
1386 
1387 const Expression *NewGVN::performSymbolicStoreEvaluation(Instruction *I) const {
1388  // Unlike loads, we never try to eliminate stores, so we do not check if they
1389  // are simple and avoid value numbering them.
1390  auto *SI = cast<StoreInst>(I);
1391  auto *StoreAccess = getMemoryAccess(SI);
1392  // Get the expression, if any, for the RHS of the MemoryDef.
1393  const MemoryAccess *StoreRHS = StoreAccess->getDefiningAccess();
1395  StoreRHS = MSSAWalker->getClobberingMemoryAccess(StoreAccess);
1396  // If we bypassed the use-def chains, make sure we add a use.
1397  StoreRHS = lookupMemoryLeader(StoreRHS);
1398  if (StoreRHS != StoreAccess->getDefiningAccess())
1399  addMemoryUsers(StoreRHS, StoreAccess);
1400  // If we are defined by ourselves, use the live on entry def.
1401  if (StoreRHS == StoreAccess)
1402  StoreRHS = MSSA->getLiveOnEntryDef();
1403 
1404  if (SI->isSimple()) {
1405  // See if we are defined by a previous store expression, it already has a
1406  // value, and it's the same value as our current store. FIXME: Right now, we
1407  // only do this for simple stores, we should expand to cover memcpys, etc.
1408  const auto *LastStore = createStoreExpression(SI, StoreRHS);
1409  const auto *LastCC = ExpressionToClass.lookup(LastStore);
1410  // We really want to check whether the expression we matched was a store. No
1411  // easy way to do that. However, we can check that the class we found has a
1412  // store, which, assuming the value numbering state is not corrupt, is
1413  // sufficient, because we must also be equivalent to that store's expression
1414  // for it to be in the same class as the load.
1415  if (LastCC && LastCC->getStoredValue() == LastStore->getStoredValue())
1416  return LastStore;
1417  // Also check if our value operand is defined by a load of the same memory
1418  // location, and the memory state is the same as it was then (otherwise, it
1419  // could have been overwritten later. See test32 in
1420  // transforms/DeadStoreElimination/simple.ll).
1421  if (auto *LI = dyn_cast<LoadInst>(LastStore->getStoredValue()))
1422  if ((lookupOperandLeader(LI->getPointerOperand()) ==
1423  LastStore->getOperand(0)) &&
1424  (lookupMemoryLeader(getMemoryAccess(LI)->getDefiningAccess()) ==
1425  StoreRHS))
1426  return LastStore;
1427  deleteExpression(LastStore);
1428  }
1429 
1430  // If the store is not equivalent to anything, value number it as a store that
1431  // produces a unique memory state (instead of using it's MemoryUse, we use
1432  // it's MemoryDef).
1433  return createStoreExpression(SI, StoreAccess);
1434 }
1435 
1436 // See if we can extract the value of a loaded pointer from a load, a store, or
1437 // a memory instruction.
1438 const Expression *
1439 NewGVN::performSymbolicLoadCoercion(Type *LoadType, Value *LoadPtr,
1440  LoadInst *LI, Instruction *DepInst,
1441  MemoryAccess *DefiningAccess) const {
1442  assert((!LI || LI->isSimple()) && "Not a simple load");
1443  if (auto *DepSI = dyn_cast<StoreInst>(DepInst)) {
1444  // Can't forward from non-atomic to atomic without violating memory model.
1445  // Also don't need to coerce if they are the same type, we will just
1446  // propagate.
1447  if (LI->isAtomic() > DepSI->isAtomic() ||
1448  LoadType == DepSI->getValueOperand()->getType())
1449  return nullptr;
1450  int Offset = analyzeLoadFromClobberingStore(LoadType, LoadPtr, DepSI, DL);
1451  if (Offset >= 0) {
1452  if (auto *C = dyn_cast<Constant>(
1453  lookupOperandLeader(DepSI->getValueOperand()))) {
1454  LLVM_DEBUG(dbgs() << "Coercing load from store " << *DepSI
1455  << " to constant " << *C << "\n");
1456  return createConstantExpression(
1457  getConstantStoreValueForLoad(C, Offset, LoadType, DL));
1458  }
1459  }
1460  } else if (auto *DepLI = dyn_cast<LoadInst>(DepInst)) {
1461  // Can't forward from non-atomic to atomic without violating memory model.
1462  if (LI->isAtomic() > DepLI->isAtomic())
1463  return nullptr;
1464  int Offset = analyzeLoadFromClobberingLoad(LoadType, LoadPtr, DepLI, DL);
1465  if (Offset >= 0) {
1466  // We can coerce a constant load into a load.
1467  if (auto *C = dyn_cast<Constant>(lookupOperandLeader(DepLI)))
1468  if (auto *PossibleConstant =
1469  getConstantLoadValueForLoad(C, Offset, LoadType, DL)) {
1470  LLVM_DEBUG(dbgs() << "Coercing load from load " << *LI
1471  << " to constant " << *PossibleConstant << "\n");
1472  return createConstantExpression(PossibleConstant);
1473  }
1474  }
1475  } else if (auto *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
1476  int Offset = analyzeLoadFromClobberingMemInst(LoadType, LoadPtr, DepMI, DL);
1477  if (Offset >= 0) {
1478  if (auto *PossibleConstant =
1479  getConstantMemInstValueForLoad(DepMI, Offset, LoadType, DL)) {
1480  LLVM_DEBUG(dbgs() << "Coercing load from meminst " << *DepMI
1481  << " to constant " << *PossibleConstant << "\n");
1482  return createConstantExpression(PossibleConstant);
1483  }
1484  }
1485  }
1486 
1487  // All of the below are only true if the loaded pointer is produced
1488  // by the dependent instruction.
1489  if (LoadPtr != lookupOperandLeader(DepInst) &&
1490  !AA->isMustAlias(LoadPtr, DepInst))
1491  return nullptr;
1492  // If this load really doesn't depend on anything, then we must be loading an
1493  // undef value. This can happen when loading for a fresh allocation with no
1494  // intervening stores, for example. Note that this is only true in the case
1495  // that the result of the allocation is pointer equal to the load ptr.
1496  if (isa<AllocaInst>(DepInst)) {
1497  return createConstantExpression(UndefValue::get(LoadType));
1498  }
1499  // If this load occurs either right after a lifetime begin,
1500  // then the loaded value is undefined.
1501  else if (auto *II = dyn_cast<IntrinsicInst>(DepInst)) {
1502  if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1503  return createConstantExpression(UndefValue::get(LoadType));
1504  } else if (isAllocationFn(DepInst, TLI))
1505  if (auto *InitVal = getInitialValueOfAllocation(cast<CallBase>(DepInst),
1506  TLI, LoadType))
1507  return createConstantExpression(InitVal);
1508 
1509  return nullptr;
1510 }
1511 
1512 const Expression *NewGVN::performSymbolicLoadEvaluation(Instruction *I) const {
1513  auto *LI = cast<LoadInst>(I);
1514 
1515  // We can eliminate in favor of non-simple loads, but we won't be able to
1516  // eliminate the loads themselves.
1517  if (!LI->isSimple())
1518  return nullptr;
1519 
1520  Value *LoadAddressLeader = lookupOperandLeader(LI->getPointerOperand());
1521  // Load of undef is UB.
1522  if (isa<UndefValue>(LoadAddressLeader))
1523  return createConstantExpression(PoisonValue::get(LI->getType()));
1524  MemoryAccess *OriginalAccess = getMemoryAccess(I);
1525  MemoryAccess *DefiningAccess =
1526  MSSAWalker->getClobberingMemoryAccess(OriginalAccess);
1527 
1528  if (!MSSA->isLiveOnEntryDef(DefiningAccess)) {
1529  if (auto *MD = dyn_cast<MemoryDef>(DefiningAccess)) {
1530  Instruction *DefiningInst = MD->getMemoryInst();
1531  // If the defining instruction is not reachable, replace with poison.
1532  if (!ReachableBlocks.count(DefiningInst->getParent()))
1533  return createConstantExpression(PoisonValue::get(LI->getType()));
1534  // This will handle stores and memory insts. We only do if it the
1535  // defining access has a different type, or it is a pointer produced by
1536  // certain memory operations that cause the memory to have a fixed value
1537  // (IE things like calloc).
1538  if (const auto *CoercionResult =
1539  performSymbolicLoadCoercion(LI->getType(), LoadAddressLeader, LI,
1540  DefiningInst, DefiningAccess))
1541  return CoercionResult;
1542  }
1543  }
1544 
1545  const auto *LE = createLoadExpression(LI->getType(), LoadAddressLeader, LI,
1546  DefiningAccess);
1547  // If our MemoryLeader is not our defining access, add a use to the
1548  // MemoryLeader, so that we get reprocessed when it changes.
1549  if (LE->getMemoryLeader() != DefiningAccess)
1550  addMemoryUsers(LE->getMemoryLeader(), OriginalAccess);
1551  return LE;
1552 }
1553 
1554 NewGVN::ExprResult
1555 NewGVN::performSymbolicPredicateInfoEvaluation(IntrinsicInst *I) const {
1556  auto *PI = PredInfo->getPredicateInfoFor(I);
1557  if (!PI)
1558  return ExprResult::none();
1559 
1560  LLVM_DEBUG(dbgs() << "Found predicate info from instruction !\n");
1561 
1562  const Optional<PredicateConstraint> &Constraint = PI->getConstraint();
1563  if (!Constraint)
1564  return ExprResult::none();
1565 
1566  CmpInst::Predicate Predicate = Constraint->Predicate;
1567  Value *CmpOp0 = I->getOperand(0);
1568  Value *CmpOp1 = Constraint->OtherOp;
1569 
1570  Value *FirstOp = lookupOperandLeader(CmpOp0);
1571  Value *SecondOp = lookupOperandLeader(CmpOp1);
1572  Value *AdditionallyUsedValue = CmpOp0;
1573 
1574  // Sort the ops.
1575  if (shouldSwapOperandsForIntrinsic(FirstOp, SecondOp, I)) {
1576  std::swap(FirstOp, SecondOp);
1578  AdditionallyUsedValue = CmpOp1;
1579  }
1580 
1581  if (Predicate == CmpInst::ICMP_EQ)
1582  return ExprResult::some(createVariableOrConstant(FirstOp),
1583  AdditionallyUsedValue, PI);
1584 
1585  // Handle the special case of floating point.
1586  if (Predicate == CmpInst::FCMP_OEQ && isa<ConstantFP>(FirstOp) &&
1587  !cast<ConstantFP>(FirstOp)->isZero())
1588  return ExprResult::some(createConstantExpression(cast<Constant>(FirstOp)),
1589  AdditionallyUsedValue, PI);
1590 
1591  return ExprResult::none();
1592 }
1593 
1594 // Evaluate read only and pure calls, and create an expression result.
1595 NewGVN::ExprResult NewGVN::performSymbolicCallEvaluation(Instruction *I) const {
1596  auto *CI = cast<CallInst>(I);
1597  if (auto *II = dyn_cast<IntrinsicInst>(I)) {
1598  // Intrinsics with the returned attribute are copies of arguments.
1599  if (auto *ReturnedValue = II->getReturnedArgOperand()) {
1600  if (II->getIntrinsicID() == Intrinsic::ssa_copy)
1601  if (auto Res = performSymbolicPredicateInfoEvaluation(II))
1602  return Res;
1603  return ExprResult::some(createVariableOrConstant(ReturnedValue));
1604  }
1605  }
1606  if (AA->doesNotAccessMemory(CI)) {
1607  return ExprResult::some(
1608  createCallExpression(CI, TOPClass->getMemoryLeader()));
1609  } else if (AA->onlyReadsMemory(CI)) {
1610  if (auto *MA = MSSA->getMemoryAccess(CI)) {
1611  auto *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(MA);
1612  return ExprResult::some(createCallExpression(CI, DefiningAccess));
1613  } else // MSSA determined that CI does not access memory.
1614  return ExprResult::some(
1615  createCallExpression(CI, TOPClass->getMemoryLeader()));
1616  }
1617  return ExprResult::none();
1618 }
1619 
1620 // Retrieve the memory class for a given MemoryAccess.
1621 CongruenceClass *NewGVN::getMemoryClass(const MemoryAccess *MA) const {
1622  auto *Result = MemoryAccessToClass.lookup(MA);
1623  assert(Result && "Should have found memory class");
1624  return Result;
1625 }
1626 
1627 // Update the MemoryAccess equivalence table to say that From is equal to To,
1628 // and return true if this is different from what already existed in the table.
1629 bool NewGVN::setMemoryClass(const MemoryAccess *From,
1630  CongruenceClass *NewClass) {
1631  assert(NewClass &&
1632  "Every MemoryAccess should be getting mapped to a non-null class");
1633  LLVM_DEBUG(dbgs() << "Setting " << *From);
1634  LLVM_DEBUG(dbgs() << " equivalent to congruence class ");
1635  LLVM_DEBUG(dbgs() << NewClass->getID()
1636  << " with current MemoryAccess leader ");
1637  LLVM_DEBUG(dbgs() << *NewClass->getMemoryLeader() << "\n");
1638 
1639  auto LookupResult = MemoryAccessToClass.find(From);
1640  bool Changed = false;
1641  // If it's already in the table, see if the value changed.
1642  if (LookupResult != MemoryAccessToClass.end()) {
1643  auto *OldClass = LookupResult->second;
1644  if (OldClass != NewClass) {
1645  // If this is a phi, we have to handle memory member updates.
1646  if (auto *MP = dyn_cast<MemoryPhi>(From)) {
1647  OldClass->memory_erase(MP);
1648  NewClass->memory_insert(MP);
1649  // This may have killed the class if it had no non-memory members
1650  if (OldClass->getMemoryLeader() == From) {
1651  if (OldClass->definesNoMemory()) {
1652  OldClass->setMemoryLeader(nullptr);
1653  } else {
1654  OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
1655  LLVM_DEBUG(dbgs() << "Memory class leader change for class "
1656  << OldClass->getID() << " to "
1657  << *OldClass->getMemoryLeader()
1658  << " due to removal of a memory member " << *From
1659  << "\n");
1660  markMemoryLeaderChangeTouched(OldClass);
1661  }
1662  }
1663  }
1664  // It wasn't equivalent before, and now it is.
1665  LookupResult->second = NewClass;
1666  Changed = true;
1667  }
1668  }
1669 
1670  return Changed;
1671 }
1672 
1673 // Determine if a instruction is cycle-free. That means the values in the
1674 // instruction don't depend on any expressions that can change value as a result
1675 // of the instruction. For example, a non-cycle free instruction would be v =
1676 // phi(0, v+1).
1677 bool NewGVN::isCycleFree(const Instruction *I) const {
1678  // In order to compute cycle-freeness, we do SCC finding on the instruction,
1679  // and see what kind of SCC it ends up in. If it is a singleton, it is
1680  // cycle-free. If it is not in a singleton, it is only cycle free if the
1681  // other members are all phi nodes (as they do not compute anything, they are
1682  // copies).
1683  auto ICS = InstCycleState.lookup(I);
1684  if (ICS == ICS_Unknown) {
1685  SCCFinder.Start(I);
1686  auto &SCC = SCCFinder.getComponentFor(I);
1687  // It's cycle free if it's size 1 or the SCC is *only* phi nodes.
1688  if (SCC.size() == 1)
1689  InstCycleState.insert({I, ICS_CycleFree});
1690  else {
1691  bool AllPhis = llvm::all_of(SCC, [](const Value *V) {
1692  return isa<PHINode>(V) || isCopyOfAPHI(V);
1693  });
1694  ICS = AllPhis ? ICS_CycleFree : ICS_Cycle;
1695  for (auto *Member : SCC)
1696  if (auto *MemberPhi = dyn_cast<PHINode>(Member))
1697  InstCycleState.insert({MemberPhi, ICS});
1698  }
1699  }
1700  if (ICS == ICS_Cycle)
1701  return false;
1702  return true;
1703 }
1704 
1705 // Evaluate PHI nodes symbolically and create an expression result.
1706 const Expression *
1707 NewGVN::performSymbolicPHIEvaluation(ArrayRef<ValPair> PHIOps,
1708  Instruction *I,
1709  BasicBlock *PHIBlock) const {
1710  // True if one of the incoming phi edges is a backedge.
1711  bool HasBackedge = false;
1712  // All constant tracks the state of whether all the *original* phi operands
1713  // This is really shorthand for "this phi cannot cycle due to forward
1714  // change in value of the phi is guaranteed not to later change the value of
1715  // the phi. IE it can't be v = phi(undef, v+1)
1716  bool OriginalOpsConstant = true;
1717  auto *E = cast<PHIExpression>(createPHIExpression(
1718  PHIOps, I, PHIBlock, HasBackedge, OriginalOpsConstant));
1719  // We match the semantics of SimplifyPhiNode from InstructionSimplify here.
1720  // See if all arguments are the same.
1721  // We track if any were undef because they need special handling.
1722  bool HasUndef = false, HasPoison = false;
1723  auto Filtered = make_filter_range(E->operands(), [&](Value *Arg) {
1724  if (isa<PoisonValue>(Arg)) {
1725  HasPoison = true;
1726  return false;
1727  }
1728  if (isa<UndefValue>(Arg)) {
1729  HasUndef = true;
1730  return false;
1731  }
1732  return true;
1733  });
1734  // If we are left with no operands, it's dead.
1735  if (Filtered.empty()) {
1736  // If it has undef or poison at this point, it means there are no-non-undef
1737  // arguments, and thus, the value of the phi node must be undef.
1738  if (HasUndef) {
1739  LLVM_DEBUG(
1740  dbgs() << "PHI Node " << *I
1741  << " has no non-undef arguments, valuing it as undef\n");
1742  return createConstantExpression(UndefValue::get(I->getType()));
1743  }
1744  if (HasPoison) {
1745  LLVM_DEBUG(
1746  dbgs() << "PHI Node " << *I
1747  << " has no non-poison arguments, valuing it as poison\n");
1748  return createConstantExpression(PoisonValue::get(I->getType()));
1749  }
1750 
1751  LLVM_DEBUG(dbgs() << "No arguments of PHI node " << *I << " are live\n");
1752  deleteExpression(E);
1753  return createDeadExpression();
1754  }
1755  Value *AllSameValue = *(Filtered.begin());
1756  ++Filtered.begin();
1757  // Can't use std::equal here, sadly, because filter.begin moves.
1758  if (llvm::all_of(Filtered, [&](Value *Arg) { return Arg == AllSameValue; })) {
1759  // Can't fold phi(undef, X) -> X unless X can't be poison (thus X is undef
1760  // in the worst case).
1761  if (HasUndef && !isGuaranteedNotToBePoison(AllSameValue, AC, nullptr, DT))
1762  return E;
1763 
1764  // In LLVM's non-standard representation of phi nodes, it's possible to have
1765  // phi nodes with cycles (IE dependent on other phis that are .... dependent
1766  // on the original phi node), especially in weird CFG's where some arguments
1767  // are unreachable, or uninitialized along certain paths. This can cause
1768  // infinite loops during evaluation. We work around this by not trying to
1769  // really evaluate them independently, but instead using a variable
1770  // expression to say if one is equivalent to the other.
1771  // We also special case undef/poison, so that if we have an undef, we can't
1772  // use the common value unless it dominates the phi block.
1773  if (HasPoison || HasUndef) {
1774  // If we have undef and at least one other value, this is really a
1775  // multivalued phi, and we need to know if it's cycle free in order to
1776  // evaluate whether we can ignore the undef. The other parts of this are
1777  // just shortcuts. If there is no backedge, or all operands are
1778  // constants, it also must be cycle free.
1779  if (HasBackedge && !OriginalOpsConstant &&
1780  !isa<UndefValue>(AllSameValue) && !isCycleFree(I))
1781  return E;
1782 
1783  // Only have to check for instructions
1784  if (auto *AllSameInst = dyn_cast<Instruction>(AllSameValue))
1785  if (!someEquivalentDominates(AllSameInst, I))
1786  return E;
1787  }
1788  // Can't simplify to something that comes later in the iteration.
1789  // Otherwise, when and if it changes congruence class, we will never catch
1790  // up. We will always be a class behind it.
1791  if (isa<Instruction>(AllSameValue) &&
1792  InstrToDFSNum(AllSameValue) > InstrToDFSNum(I))
1793  return E;
1794  NumGVNPhisAllSame++;
1795  LLVM_DEBUG(dbgs() << "Simplified PHI node " << *I << " to " << *AllSameValue
1796  << "\n");
1797  deleteExpression(E);
1798  return createVariableOrConstant(AllSameValue);
1799  }
1800  return E;
1801 }
1802 
1803 const Expression *
1804 NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) const {
1805  if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1806  auto *WO = dyn_cast<WithOverflowInst>(EI->getAggregateOperand());
1807  if (WO && EI->getNumIndices() == 1 && *EI->idx_begin() == 0)
1808  // EI is an extract from one of our with.overflow intrinsics. Synthesize
1809  // a semantically equivalent expression instead of an extract value
1810  // expression.
1811  return createBinaryExpression(WO->getBinaryOp(), EI->getType(),
1812  WO->getLHS(), WO->getRHS(), I);
1813  }
1814 
1815  return createAggregateValueExpression(I);
1816 }
1817 
1818 NewGVN::ExprResult NewGVN::performSymbolicCmpEvaluation(Instruction *I) const {
1819  assert(isa<CmpInst>(I) && "Expected a cmp instruction.");
1820 
1821  auto *CI = cast<CmpInst>(I);
1822  // See if our operands are equal to those of a previous predicate, and if so,
1823  // if it implies true or false.
1824  auto Op0 = lookupOperandLeader(CI->getOperand(0));
1825  auto Op1 = lookupOperandLeader(CI->getOperand(1));
1826  auto OurPredicate = CI->getPredicate();
1827  if (shouldSwapOperands(Op0, Op1)) {
1828  std::swap(Op0, Op1);
1829  OurPredicate = CI->getSwappedPredicate();
1830  }
1831 
1832  // Avoid processing the same info twice.
1833  const PredicateBase *LastPredInfo = nullptr;
1834  // See if we know something about the comparison itself, like it is the target
1835  // of an assume.
1836  auto *CmpPI = PredInfo->getPredicateInfoFor(I);
1837  if (isa_and_nonnull<PredicateAssume>(CmpPI))
1838  return ExprResult::some(
1839  createConstantExpression(ConstantInt::getTrue(CI->getType())));
1840 
1841  if (Op0 == Op1) {
1842  // This condition does not depend on predicates, no need to add users
1843  if (CI->isTrueWhenEqual())
1844  return ExprResult::some(
1845  createConstantExpression(ConstantInt::getTrue(CI->getType())));
1846  else if (CI->isFalseWhenEqual())
1847  return ExprResult::some(
1848  createConstantExpression(ConstantInt::getFalse(CI->getType())));
1849  }
1850 
1851  // NOTE: Because we are comparing both operands here and below, and using
1852  // previous comparisons, we rely on fact that predicateinfo knows to mark
1853  // comparisons that use renamed operands as users of the earlier comparisons.
1854  // It is *not* enough to just mark predicateinfo renamed operands as users of
1855  // the earlier comparisons, because the *other* operand may have changed in a
1856  // previous iteration.
1857  // Example:
1858  // icmp slt %a, %b
1859  // %b.0 = ssa.copy(%b)
1860  // false branch:
1861  // icmp slt %c, %b.0
1862 
1863  // %c and %a may start out equal, and thus, the code below will say the second
1864  // %icmp is false. c may become equal to something else, and in that case the
1865  // %second icmp *must* be reexamined, but would not if only the renamed
1866  // %operands are considered users of the icmp.
1867 
1868  // *Currently* we only check one level of comparisons back, and only mark one
1869  // level back as touched when changes happen. If you modify this code to look
1870  // back farther through comparisons, you *must* mark the appropriate
1871  // comparisons as users in PredicateInfo.cpp, or you will cause bugs. See if
1872  // we know something just from the operands themselves
1873 
1874  // See if our operands have predicate info, so that we may be able to derive
1875  // something from a previous comparison.
1876  for (const auto &Op : CI->operands()) {
1877  auto *PI = PredInfo->getPredicateInfoFor(Op);
1878  if (const auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI)) {
1879  if (PI == LastPredInfo)
1880  continue;
1881  LastPredInfo = PI;
1882  // In phi of ops cases, we may have predicate info that we are evaluating
1883  // in a different context.
1884  if (!DT->dominates(PBranch->To, getBlockForValue(I)))
1885  continue;
1886  // TODO: Along the false edge, we may know more things too, like
1887  // icmp of
1888  // same operands is false.
1889  // TODO: We only handle actual comparison conditions below, not
1890  // and/or.
1891  auto *BranchCond = dyn_cast<CmpInst>(PBranch->Condition);
1892  if (!BranchCond)
1893  continue;
1894  auto *BranchOp0 = lookupOperandLeader(BranchCond->getOperand(0));
1895  auto *BranchOp1 = lookupOperandLeader(BranchCond->getOperand(1));
1896  auto BranchPredicate = BranchCond->getPredicate();
1897  if (shouldSwapOperands(BranchOp0, BranchOp1)) {
1898  std::swap(BranchOp0, BranchOp1);
1899  BranchPredicate = BranchCond->getSwappedPredicate();
1900  }
1901  if (BranchOp0 == Op0 && BranchOp1 == Op1) {
1902  if (PBranch->TrueEdge) {
1903  // If we know the previous predicate is true and we are in the true
1904  // edge then we may be implied true or false.
1905  if (CmpInst::isImpliedTrueByMatchingCmp(BranchPredicate,
1906  OurPredicate)) {
1907  return ExprResult::some(
1908  createConstantExpression(ConstantInt::getTrue(CI->getType())),
1909  PI);
1910  }
1911 
1912  if (CmpInst::isImpliedFalseByMatchingCmp(BranchPredicate,
1913  OurPredicate)) {
1914  return ExprResult::some(
1915  createConstantExpression(ConstantInt::getFalse(CI->getType())),
1916  PI);
1917  }
1918  } else {
1919  // Just handle the ne and eq cases, where if we have the same
1920  // operands, we may know something.
1921  if (BranchPredicate == OurPredicate) {
1922  // Same predicate, same ops,we know it was false, so this is false.
1923  return ExprResult::some(
1924  createConstantExpression(ConstantInt::getFalse(CI->getType())),
1925  PI);
1926  } else if (BranchPredicate ==
1927  CmpInst::getInversePredicate(OurPredicate)) {
1928  // Inverse predicate, we know the other was false, so this is true.
1929  return ExprResult::some(
1930  createConstantExpression(ConstantInt::getTrue(CI->getType())),
1931  PI);
1932  }
1933  }
1934  }
1935  }
1936  }
1937  // Create expression will take care of simplifyCmpInst
1938  return createExpression(I);
1939 }
1940 
1941 // Substitute and symbolize the value before value numbering.
1942 NewGVN::ExprResult
1943 NewGVN::performSymbolicEvaluation(Value *V,
1944  SmallPtrSetImpl<Value *> &Visited) const {
1945 
1946  const Expression *E = nullptr;
1947  if (auto *C = dyn_cast<Constant>(V))
1948  E = createConstantExpression(C);
1949  else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1950  E = createVariableExpression(V);
1951  } else {
1952  // TODO: memory intrinsics.
1953  // TODO: Some day, we should do the forward propagation and reassociation
1954  // parts of the algorithm.
1955  auto *I = cast<Instruction>(V);
1956  switch (I->getOpcode()) {
1957  case Instruction::ExtractValue:
1958  case Instruction::InsertValue:
1959  E = performSymbolicAggrValueEvaluation(I);
1960  break;
1961  case Instruction::PHI: {
1963  auto *PN = cast<PHINode>(I);
1964  for (unsigned i = 0; i < PN->getNumOperands(); ++i)
1965  Ops.push_back({PN->getIncomingValue(i), PN->getIncomingBlock(i)});
1966  // Sort to ensure the invariant createPHIExpression requires is met.
1967  sortPHIOps(Ops);
1968  E = performSymbolicPHIEvaluation(Ops, I, getBlockForValue(I));
1969  } break;
1970  case Instruction::Call:
1971  return performSymbolicCallEvaluation(I);
1972  break;
1973  case Instruction::Store:
1974  E = performSymbolicStoreEvaluation(I);
1975  break;
1976  case Instruction::Load:
1977  E = performSymbolicLoadEvaluation(I);
1978  break;
1979  case Instruction::BitCast:
1980  case Instruction::AddrSpaceCast:
1981  return createExpression(I);
1982  break;
1983  case Instruction::ICmp:
1984  case Instruction::FCmp:
1985  return performSymbolicCmpEvaluation(I);
1986  break;
1987  case Instruction::FNeg:
1988  case Instruction::Add:
1989  case Instruction::FAdd:
1990  case Instruction::Sub:
1991  case Instruction::FSub:
1992  case Instruction::Mul:
1993  case Instruction::FMul:
1994  case Instruction::UDiv:
1995  case Instruction::SDiv:
1996  case Instruction::FDiv:
1997  case Instruction::URem:
1998  case Instruction::SRem:
1999  case Instruction::FRem:
2000  case Instruction::Shl:
2001  case Instruction::LShr:
2002  case Instruction::AShr:
2003  case Instruction::And:
2004  case Instruction::Or:
2005  case Instruction::Xor:
2006  case Instruction::Trunc:
2007  case Instruction::ZExt:
2008  case Instruction::SExt:
2009  case Instruction::FPToUI:
2010  case Instruction::FPToSI:
2011  case Instruction::UIToFP:
2012  case Instruction::SIToFP:
2013  case Instruction::FPTrunc:
2014  case Instruction::FPExt:
2015  case Instruction::PtrToInt:
2016  case Instruction::IntToPtr:
2017  case Instruction::Select:
2018  case Instruction::ExtractElement:
2019  case Instruction::InsertElement:
2020  case Instruction::GetElementPtr:
2021  return createExpression(I);
2022  break;
2023  case Instruction::ShuffleVector:
2024  // FIXME: Add support for shufflevector to createExpression.
2025  return ExprResult::none();
2026  default:
2027  return ExprResult::none();
2028  }
2029  }
2030  return ExprResult::some(E);
2031 }
2032 
2033 // Look up a container of values/instructions in a map, and touch all the
2034 // instructions in the container. Then erase value from the map.
2035 template <typename Map, typename KeyType>
2036 void NewGVN::touchAndErase(Map &M, const KeyType &Key) {
2037  const auto Result = M.find_as(Key);
2038  if (Result != M.end()) {
2039  for (const typename Map::mapped_type::value_type Mapped : Result->second)
2040  TouchedInstructions.set(InstrToDFSNum(Mapped));
2041  M.erase(Result);
2042  }
2043 }
2044 
2045 void NewGVN::addAdditionalUsers(Value *To, Value *User) const {
2046  assert(User && To != User);
2047  if (isa<Instruction>(To))
2048  AdditionalUsers[To].insert(User);
2049 }
2050 
2051 void NewGVN::addAdditionalUsers(ExprResult &Res, Instruction *User) const {
2052  if (Res.ExtraDep && Res.ExtraDep != User)
2053  addAdditionalUsers(Res.ExtraDep, User);
2054  Res.ExtraDep = nullptr;
2055 
2056  if (Res.PredDep) {
2057  if (const auto *PBranch = dyn_cast<PredicateBranch>(Res.PredDep))
2058  PredicateToUsers[PBranch->Condition].insert(User);
2059  else if (const auto *PAssume = dyn_cast<PredicateAssume>(Res.PredDep))
2060  PredicateToUsers[PAssume->Condition].insert(User);
2061  }
2062  Res.PredDep = nullptr;
2063 }
2064 
2065 void NewGVN::markUsersTouched(Value *V) {
2066  // Now mark the users as touched.
2067  for (auto *User : V->users()) {
2068  assert(isa<Instruction>(User) && "Use of value not within an instruction?");
2069  TouchedInstructions.set(InstrToDFSNum(User));
2070  }
2071  touchAndErase(AdditionalUsers, V);
2072 }
2073 
2074 void NewGVN::addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const {
2075  LLVM_DEBUG(dbgs() << "Adding memory user " << *U << " to " << *To << "\n");
2076  MemoryToUsers[To].insert(U);
2077 }
2078 
2079 void NewGVN::markMemoryDefTouched(const MemoryAccess *MA) {
2080  TouchedInstructions.set(MemoryToDFSNum(MA));
2081 }
2082 
2083 void NewGVN::markMemoryUsersTouched(const MemoryAccess *MA) {
2084  if (isa<MemoryUse>(MA))
2085  return;
2086  for (auto U : MA->users())
2087  TouchedInstructions.set(MemoryToDFSNum(U));
2088  touchAndErase(MemoryToUsers, MA);
2089 }
2090 
2091 // Touch all the predicates that depend on this instruction.
2092 void NewGVN::markPredicateUsersTouched(Instruction *I) {
2093  touchAndErase(PredicateToUsers, I);
2094 }
2095 
2096 // Mark users affected by a memory leader change.
2097 void NewGVN::markMemoryLeaderChangeTouched(CongruenceClass *CC) {
2098  for (auto M : CC->memory())
2099  markMemoryDefTouched(M);
2100 }
2101 
2102 // Touch the instructions that need to be updated after a congruence class has a
2103 // leader change, and mark changed values.
2104 void NewGVN::markValueLeaderChangeTouched(CongruenceClass *CC) {
2105  for (auto M : *CC) {
2106  if (auto *I = dyn_cast<Instruction>(M))
2107  TouchedInstructions.set(InstrToDFSNum(I));
2108  LeaderChanges.insert(M);
2109  }
2110 }
2111 
2112 // Give a range of things that have instruction DFS numbers, this will return
2113 // the member of the range with the smallest dfs number.
2114 template <class T, class Range>
2115 T *NewGVN::getMinDFSOfRange(const Range &R) const {
2116  std::pair<T *, unsigned> MinDFS = {nullptr, ~0U};
2117  for (const auto X : R) {
2118  auto DFSNum = InstrToDFSNum(X);
2119  if (DFSNum < MinDFS.second)
2120  MinDFS = {X, DFSNum};
2121  }
2122  return MinDFS.first;
2123 }
2124 
2125 // This function returns the MemoryAccess that should be the next leader of
2126 // congruence class CC, under the assumption that the current leader is going to
2127 // disappear.
2128 const MemoryAccess *NewGVN::getNextMemoryLeader(CongruenceClass *CC) const {
2129  // TODO: If this ends up to slow, we can maintain a next memory leader like we
2130  // do for regular leaders.
2131  // Make sure there will be a leader to find.
2132  assert(!CC->definesNoMemory() && "Can't get next leader if there is none");
2133  if (CC->getStoreCount() > 0) {
2134  if (auto *NL = dyn_cast_or_null<StoreInst>(CC->getNextLeader().first))
2135  return getMemoryAccess(NL);
2136  // Find the store with the minimum DFS number.
2137  auto *V = getMinDFSOfRange<Value>(make_filter_range(
2138  *CC, [&](const Value *V) { return isa<StoreInst>(V); }));
2139  return getMemoryAccess(cast<StoreInst>(V));
2140  }
2141  assert(CC->getStoreCount() == 0);
2142 
2143  // Given our assertion, hitting this part must mean
2144  // !OldClass->memory_empty()
2145  if (CC->memory_size() == 1)
2146  return *CC->memory_begin();
2147  return getMinDFSOfRange<const MemoryPhi>(CC->memory());
2148 }
2149 
2150 // This function returns the next value leader of a congruence class, under the
2151 // assumption that the current leader is going away. This should end up being
2152 // the next most dominating member.
2153 Value *NewGVN::getNextValueLeader(CongruenceClass *CC) const {
2154  // We don't need to sort members if there is only 1, and we don't care about
2155  // sorting the TOP class because everything either gets out of it or is
2156  // unreachable.
2157 
2158  if (CC->size() == 1 || CC == TOPClass) {
2159  return *(CC->begin());
2160  } else if (CC->getNextLeader().first) {
2161  ++NumGVNAvoidedSortedLeaderChanges;
2162  return CC->getNextLeader().first;
2163  } else {
2164  ++NumGVNSortedLeaderChanges;
2165  // NOTE: If this ends up to slow, we can maintain a dual structure for
2166  // member testing/insertion, or keep things mostly sorted, and sort only
2167  // here, or use SparseBitVector or ....
2168  return getMinDFSOfRange<Value>(*CC);
2169  }
2170 }
2171 
2172 // Move a MemoryAccess, currently in OldClass, to NewClass, including updates to
2173 // the memory members, etc for the move.
2174 //
2175 // The invariants of this function are:
2176 //
2177 // - I must be moving to NewClass from OldClass
2178 // - The StoreCount of OldClass and NewClass is expected to have been updated
2179 // for I already if it is a store.
2180 // - The OldClass memory leader has not been updated yet if I was the leader.
2181 void NewGVN::moveMemoryToNewCongruenceClass(Instruction *I,
2182  MemoryAccess *InstMA,
2183  CongruenceClass *OldClass,
2184  CongruenceClass *NewClass) {
2185  // If the leader is I, and we had a representative MemoryAccess, it should
2186  // be the MemoryAccess of OldClass.
2187  assert((!InstMA || !OldClass->getMemoryLeader() ||
2188  OldClass->getLeader() != I ||
2189  MemoryAccessToClass.lookup(OldClass->getMemoryLeader()) ==
2190  MemoryAccessToClass.lookup(InstMA)) &&
2191  "Representative MemoryAccess mismatch");
2192  // First, see what happens to the new class
2193  if (!NewClass->getMemoryLeader()) {
2194  // Should be a new class, or a store becoming a leader of a new class.
2195  assert(NewClass->size() == 1 ||
2196  (isa<StoreInst>(I) && NewClass->getStoreCount() == 1));
2197  NewClass->setMemoryLeader(InstMA);
2198  // Mark it touched if we didn't just create a singleton
2199  LLVM_DEBUG(dbgs() << "Memory class leader change for class "
2200  << NewClass->getID()
2201  << " due to new memory instruction becoming leader\n");
2202  markMemoryLeaderChangeTouched(NewClass);
2203  }
2204  setMemoryClass(InstMA, NewClass);
2205  // Now, fixup the old class if necessary
2206  if (OldClass->getMemoryLeader() == InstMA) {
2207  if (!OldClass->definesNoMemory()) {
2208  OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
2209  LLVM_DEBUG(dbgs() << "Memory class leader change for class "
2210  << OldClass->getID() << " to "
2211  << *OldClass->getMemoryLeader()
2212  << " due to removal of old leader " << *InstMA << "\n");
2213  markMemoryLeaderChangeTouched(OldClass);
2214  } else
2215  OldClass->setMemoryLeader(nullptr);
2216  }
2217 }
2218 
2219 // Move a value, currently in OldClass, to be part of NewClass
2220 // Update OldClass and NewClass for the move (including changing leaders, etc).
2221 void NewGVN::moveValueToNewCongruenceClass(Instruction *I, const Expression *E,
2222  CongruenceClass *OldClass,
2223  CongruenceClass *NewClass) {
2224  if (I == OldClass->getNextLeader().first)
2225  OldClass->resetNextLeader();
2226 
2227  OldClass->erase(I);
2228  NewClass->insert(I);
2229 
2230  if (NewClass->getLeader() != I)
2231  NewClass->addPossibleNextLeader({I, InstrToDFSNum(I)});
2232  // Handle our special casing of stores.
2233  if (auto *SI = dyn_cast<StoreInst>(I)) {
2234  OldClass->decStoreCount();
2235  // Okay, so when do we want to make a store a leader of a class?
2236  // If we have a store defined by an earlier load, we want the earlier load
2237  // to lead the class.
2238  // If we have a store defined by something else, we want the store to lead
2239  // the class so everything else gets the "something else" as a value.
2240  // If we have a store as the single member of the class, we want the store
2241  // as the leader
2242  if (NewClass->getStoreCount() == 0 && !NewClass->getStoredValue()) {
2243  // If it's a store expression we are using, it means we are not equivalent
2244  // to something earlier.
2245  if (auto *SE = dyn_cast<StoreExpression>(E)) {
2246  NewClass->setStoredValue(SE->getStoredValue());
2247  markValueLeaderChangeTouched(NewClass);
2248  // Shift the new class leader to be the store
2249  LLVM_DEBUG(dbgs() << "Changing leader of congruence class "
2250  << NewClass->getID() << " from "
2251  << *NewClass->getLeader() << " to " << *SI
2252  << " because store joined class\n");
2253  // If we changed the leader, we have to mark it changed because we don't
2254  // know what it will do to symbolic evaluation.
2255  NewClass->setLeader(SI);
2256  }
2257  // We rely on the code below handling the MemoryAccess change.
2258  }
2259  NewClass->incStoreCount();
2260  }
2261  // True if there is no memory instructions left in a class that had memory
2262  // instructions before.
2263 
2264  // If it's not a memory use, set the MemoryAccess equivalence
2265  auto *InstMA = dyn_cast_or_null<MemoryDef>(getMemoryAccess(I));
2266  if (InstMA)
2267  moveMemoryToNewCongruenceClass(I, InstMA, OldClass, NewClass);
2268  ValueToClass[I] = NewClass;
2269  // See if we destroyed the class or need to swap leaders.
2270  if (OldClass->empty() && OldClass != TOPClass) {
2271  if (OldClass->getDefiningExpr()) {
2272  LLVM_DEBUG(dbgs() << "Erasing expression " << *OldClass->getDefiningExpr()
2273  << " from table\n");
2274  // We erase it as an exact expression to make sure we don't just erase an
2275  // equivalent one.
2276  auto Iter = ExpressionToClass.find_as(
2277  ExactEqualsExpression(*OldClass->getDefiningExpr()));
2278  if (Iter != ExpressionToClass.end())
2279  ExpressionToClass.erase(Iter);
2280 #ifdef EXPENSIVE_CHECKS
2281  assert(
2282  (*OldClass->getDefiningExpr() != *E || ExpressionToClass.lookup(E)) &&
2283  "We erased the expression we just inserted, which should not happen");
2284 #endif
2285  }
2286  } else if (OldClass->getLeader() == I) {
2287  // When the leader changes, the value numbering of
2288  // everything may change due to symbolization changes, so we need to
2289  // reprocess.
2290  LLVM_DEBUG(dbgs() << "Value class leader change for class "
2291  << OldClass->getID() << "\n");
2292  ++NumGVNLeaderChanges;
2293  // Destroy the stored value if there are no more stores to represent it.
2294  // Note that this is basically clean up for the expression removal that
2295  // happens below. If we remove stores from a class, we may leave it as a
2296  // class of equivalent memory phis.
2297  if (OldClass->getStoreCount() == 0) {
2298  if (OldClass->getStoredValue())
2299  OldClass->setStoredValue(nullptr);
2300  }
2301  OldClass->setLeader(getNextValueLeader(OldClass));
2302  OldClass->resetNextLeader();
2303  markValueLeaderChangeTouched(OldClass);
2304  }
2305 }
2306 
2307 // For a given expression, mark the phi of ops instructions that could have
2308 // changed as a result.
2309 void NewGVN::markPhiOfOpsChanged(const Expression *E) {
2310  touchAndErase(ExpressionToPhiOfOps, E);
2311 }
2312 
2313 // Perform congruence finding on a given value numbering expression.
2314 void NewGVN::performCongruenceFinding(Instruction *I, const Expression *E) {
2315  // This is guaranteed to return something, since it will at least find
2316  // TOP.
2317 
2318  CongruenceClass *IClass = ValueToClass.lookup(I);
2319  assert(IClass && "Should have found a IClass");
2320  // Dead classes should have been eliminated from the mapping.
2321  assert(!IClass->isDead() && "Found a dead class");
2322 
2323  CongruenceClass *EClass = nullptr;
2324  if (const auto *VE = dyn_cast<VariableExpression>(E)) {
2325  EClass = ValueToClass.lookup(VE->getVariableValue());
2326  } else if (isa<DeadExpression>(E)) {
2327  EClass = TOPClass;
2328  }
2329  if (!EClass) {
2330  auto lookupResult = ExpressionToClass.insert({E, nullptr});
2331 
2332  // If it's not in the value table, create a new congruence class.
2333  if (lookupResult.second) {
2334  CongruenceClass *NewClass = createCongruenceClass(nullptr, E);
2335  auto place = lookupResult.first;
2336  place->second = NewClass;
2337 
2338  // Constants and variables should always be made the leader.
2339  if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
2340  NewClass->setLeader(CE->getConstantValue());
2341  } else if (const auto *SE = dyn_cast<StoreExpression>(E)) {
2342  StoreInst *SI = SE->getStoreInst();
2343  NewClass->setLeader(SI);
2344  NewClass->setStoredValue(SE->getStoredValue());
2345  // The RepMemoryAccess field will be filled in properly by the
2346  // moveValueToNewCongruenceClass call.
2347  } else {
2348  NewClass->setLeader(I);
2349  }
2350  assert(!isa<VariableExpression>(E) &&
2351  "VariableExpression should have been handled already");
2352 
2353  EClass = NewClass;
2354  LLVM_DEBUG(dbgs() << "Created new congruence class for " << *I
2355  << " using expression " << *E << " at "
2356  << NewClass->getID() << " and leader "
2357  << *(NewClass->getLeader()));
2358  if (NewClass->getStoredValue())
2359  LLVM_DEBUG(dbgs() << " and stored value "
2360  << *(NewClass->getStoredValue()));
2361  LLVM_DEBUG(dbgs() << "\n");
2362  } else {
2363  EClass = lookupResult.first->second;
2364  if (isa<ConstantExpression>(E))
2365  assert((isa<Constant>(EClass->getLeader()) ||
2366  (EClass->getStoredValue() &&
2367  isa<Constant>(EClass->getStoredValue()))) &&
2368  "Any class with a constant expression should have a "
2369  "constant leader");
2370 
2371  assert(EClass && "Somehow don't have an eclass");
2372 
2373  assert(!EClass->isDead() && "We accidentally looked up a dead class");
2374  }
2375  }
2376  bool ClassChanged = IClass != EClass;
2377  bool LeaderChanged = LeaderChanges.erase(I);
2378  if (ClassChanged || LeaderChanged) {
2379  LLVM_DEBUG(dbgs() << "New class " << EClass->getID() << " for expression "
2380  << *E << "\n");
2381  if (ClassChanged) {
2382  moveValueToNewCongruenceClass(I, E, IClass, EClass);
2383  markPhiOfOpsChanged(E);
2384  }
2385 
2386  markUsersTouched(I);
2387  if (MemoryAccess *MA = getMemoryAccess(I))
2388  markMemoryUsersTouched(MA);
2389  if (auto *CI = dyn_cast<CmpInst>(I))
2390  markPredicateUsersTouched(CI);
2391  }
2392  // If we changed the class of the store, we want to ensure nothing finds the
2393  // old store expression. In particular, loads do not compare against stored
2394  // value, so they will find old store expressions (and associated class
2395  // mappings) if we leave them in the table.
2396  if (ClassChanged && isa<StoreInst>(I)) {
2397  auto *OldE = ValueToExpression.lookup(I);
2398  // It could just be that the old class died. We don't want to erase it if we
2399  // just moved classes.
2400  if (OldE && isa<StoreExpression>(OldE) && *E != *OldE) {
2401  // Erase this as an exact expression to ensure we don't erase expressions
2402  // equivalent to it.
2403  auto Iter = ExpressionToClass.find_as(ExactEqualsExpression(*OldE));
2404  if (Iter != ExpressionToClass.end())
2405  ExpressionToClass.erase(Iter);
2406  }
2407  }
2408  ValueToExpression[I] = E;
2409 }
2410 
2411 // Process the fact that Edge (from, to) is reachable, including marking
2412 // any newly reachable blocks and instructions for processing.
2413 void NewGVN::updateReachableEdge(BasicBlock *From, BasicBlock *To) {
2414  // Check if the Edge was reachable before.
2415  if (ReachableEdges.insert({From, To}).second) {
2416  // If this block wasn't reachable before, all instructions are touched.
2417  if (ReachableBlocks.insert(To).second) {
2418  LLVM_DEBUG(dbgs() << "Block " << getBlockName(To)
2419  << " marked reachable\n");
2420  const auto &InstRange = BlockInstRange.lookup(To);
2421  TouchedInstructions.set(InstRange.first, InstRange.second);
2422  } else {
2423  LLVM_DEBUG(dbgs() << "Block " << getBlockName(To)
2424  << " was reachable, but new edge {"
2425  << getBlockName(From) << "," << getBlockName(To)
2426  << "} to it found\n");
2427 
2428  // We've made an edge reachable to an existing block, which may
2429  // impact predicates. Otherwise, only mark the phi nodes as touched, as
2430  // they are the only thing that depend on new edges. Anything using their
2431  // values will get propagated to if necessary.
2432  if (MemoryAccess *MemPhi = getMemoryAccess(To))
2433  TouchedInstructions.set(InstrToDFSNum(MemPhi));
2434 
2435  // FIXME: We should just add a union op on a Bitvector and
2436  // SparseBitVector. We can do it word by word faster than we are doing it
2437  // here.
2438  for (auto InstNum : RevisitOnReachabilityChange[To])
2439  TouchedInstructions.set(InstNum);
2440  }
2441  }
2442 }
2443 
2444 // Given a predicate condition (from a switch, cmp, or whatever) and a block,
2445 // see if we know some constant value for it already.
2446 Value *NewGVN::findConditionEquivalence(Value *Cond) const {
2447  auto Result = lookupOperandLeader(Cond);
2448  return isa<Constant>(Result) ? Result : nullptr;
2449 }
2450 
2451 // Process the outgoing edges of a block for reachability.
2452 void NewGVN::processOutgoingEdges(Instruction *TI, BasicBlock *B) {
2453  // Evaluate reachability of terminator instruction.
2454  Value *Cond;
2455  BasicBlock *TrueSucc, *FalseSucc;
2456  if (match(TI, m_Br(m_Value(Cond), TrueSucc, FalseSucc))) {
2457  Value *CondEvaluated = findConditionEquivalence(Cond);
2458  if (!CondEvaluated) {
2459  if (auto *I = dyn_cast<Instruction>(Cond)) {
2460  SmallPtrSet<Value *, 4> Visited;
2461  auto Res = performSymbolicEvaluation(I, Visited);
2462  if (const auto *CE = dyn_cast_or_null<ConstantExpression>(Res.Expr)) {
2463  CondEvaluated = CE->getConstantValue();
2464  addAdditionalUsers(Res, I);
2465  } else {
2466  // Did not use simplification result, no need to add the extra
2467  // dependency.
2468  Res.ExtraDep = nullptr;
2469  }
2470  } else if (isa<ConstantInt>(Cond)) {
2471  CondEvaluated = Cond;
2472  }
2473  }
2474  ConstantInt *CI;
2475  if (CondEvaluated && (CI = dyn_cast<ConstantInt>(CondEvaluated))) {
2476  if (CI->isOne()) {
2477  LLVM_DEBUG(dbgs() << "Condition for Terminator " << *TI
2478  << " evaluated to true\n");
2479  updateReachableEdge(B, TrueSucc);
2480  } else if (CI->isZero()) {
2481  LLVM_DEBUG(dbgs() << "Condition for Terminator " << *TI
2482  << " evaluated to false\n");
2483  updateReachableEdge(B, FalseSucc);
2484  }
2485  } else {
2486  updateReachableEdge(B, TrueSucc);
2487  updateReachableEdge(B, FalseSucc);
2488  }
2489  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
2490  // For switches, propagate the case values into the case
2491  // destinations.
2492 
2493  Value *SwitchCond = SI->getCondition();
2494  Value *CondEvaluated = findConditionEquivalence(SwitchCond);
2495  // See if we were able to turn this switch statement into a constant.
2496  if (CondEvaluated && isa<ConstantInt>(CondEvaluated)) {
2497  auto *CondVal = cast<ConstantInt>(CondEvaluated);
2498  // We should be able to get case value for this.
2499  auto Case = *SI->findCaseValue(CondVal);
2500  if (Case.getCaseSuccessor() == SI->getDefaultDest()) {
2501  // We proved the value is outside of the range of the case.
2502  // We can't do anything other than mark the default dest as reachable,
2503  // and go home.
2504  updateReachableEdge(B, SI->getDefaultDest());
2505  return;
2506  }
2507  // Now get where it goes and mark it reachable.
2508  BasicBlock *TargetBlock = Case.getCaseSuccessor();
2509  updateReachableEdge(B, TargetBlock);
2510  } else {
2511  for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) {
2512  BasicBlock *TargetBlock = SI->getSuccessor(i);
2513  updateReachableEdge(B, TargetBlock);
2514  }
2515  }
2516  } else {
2517  // Otherwise this is either unconditional, or a type we have no
2518  // idea about. Just mark successors as reachable.
2519  for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
2520  BasicBlock *TargetBlock = TI->getSuccessor(i);
2521  updateReachableEdge(B, TargetBlock);
2522  }
2523 
2524  // This also may be a memory defining terminator, in which case, set it
2525  // equivalent only to itself.
2526  //
2527  auto *MA = getMemoryAccess(TI);
2528  if (MA && !isa<MemoryUse>(MA)) {
2529  auto *CC = ensureLeaderOfMemoryClass(MA);
2530  if (setMemoryClass(MA, CC))
2531  markMemoryUsersTouched(MA);
2532  }
2533  }
2534 }
2535 
2536 // Remove the PHI of Ops PHI for I
2537 void NewGVN::removePhiOfOps(Instruction *I, PHINode *PHITemp) {
2538  InstrDFS.erase(PHITemp);
2539  // It's still a temp instruction. We keep it in the array so it gets erased.
2540  // However, it's no longer used by I, or in the block
2541  TempToBlock.erase(PHITemp);
2542  RealToTemp.erase(I);
2543  // We don't remove the users from the phi node uses. This wastes a little
2544  // time, but such is life. We could use two sets to track which were there
2545  // are the start of NewGVN, and which were added, but right nowt he cost of
2546  // tracking is more than the cost of checking for more phi of ops.
2547 }
2548 
2549 // Add PHI Op in BB as a PHI of operations version of ExistingValue.
2550 void NewGVN::addPhiOfOps(PHINode *Op, BasicBlock *BB,
2551  Instruction *ExistingValue) {
2552  InstrDFS[Op] = InstrToDFSNum(ExistingValue);
2553  AllTempInstructions.insert(Op);
2554  TempToBlock[Op] = BB;
2555  RealToTemp[ExistingValue] = Op;
2556  // Add all users to phi node use, as they are now uses of the phi of ops phis
2557  // and may themselves be phi of ops.
2558  for (auto *U : ExistingValue->users())
2559  if (auto *UI = dyn_cast<Instruction>(U))
2560  PHINodeUses.insert(UI);
2561 }
2562 
2563 static bool okayForPHIOfOps(const Instruction *I) {
2564  if (!EnablePhiOfOps)
2565  return false;
2566  return isa<BinaryOperator>(I) || isa<SelectInst>(I) || isa<CmpInst>(I) ||
2567  isa<LoadInst>(I);
2568 }
2569 
2570 bool NewGVN::OpIsSafeForPHIOfOpsHelper(
2571  Value *V, const BasicBlock *PHIBlock,
2573  SmallVectorImpl<Instruction *> &Worklist) {
2574 
2575  if (!isa<Instruction>(V))
2576  return true;
2577  auto OISIt = OpSafeForPHIOfOps.find(V);
2578  if (OISIt != OpSafeForPHIOfOps.end())
2579  return OISIt->second;
2580 
2581  // Keep walking until we either dominate the phi block, or hit a phi, or run
2582  // out of things to check.
2583  if (DT->properlyDominates(getBlockForValue(V), PHIBlock)) {
2584  OpSafeForPHIOfOps.insert({V, true});
2585  return true;
2586  }
2587  // PHI in the same block.
2588  if (isa<PHINode>(V) && getBlockForValue(V) == PHIBlock) {
2589  OpSafeForPHIOfOps.insert({V, false});
2590  return false;
2591  }
2592 
2593  auto *OrigI = cast<Instruction>(V);
2594  // When we hit an instruction that reads memory (load, call, etc), we must
2595  // consider any store that may happen in the loop. For now, we assume the
2596  // worst: there is a store in the loop that alias with this read.
2597  // The case where the load is outside the loop is already covered by the
2598  // dominator check above.
2599  // TODO: relax this condition
2600  if (OrigI->mayReadFromMemory())
2601  return false;
2602 
2603  for (auto *Op : OrigI->operand_values()) {
2604  if (!isa<Instruction>(Op))
2605  continue;
2606  // Stop now if we find an unsafe operand.
2607  auto OISIt = OpSafeForPHIOfOps.find(OrigI);
2608  if (OISIt != OpSafeForPHIOfOps.end()) {
2609  if (!OISIt->second) {
2610  OpSafeForPHIOfOps.insert({V, false});
2611  return false;
2612  }
2613  continue;
2614  }
2615  if (!Visited.insert(Op).second)
2616  continue;
2617  Worklist.push_back(cast<Instruction>(Op));
2618  }
2619  return true;
2620 }
2621 
2622 // Return true if this operand will be safe to use for phi of ops.
2623 //
2624 // The reason some operands are unsafe is that we are not trying to recursively
2625 // translate everything back through phi nodes. We actually expect some lookups
2626 // of expressions to fail. In particular, a lookup where the expression cannot
2627 // exist in the predecessor. This is true even if the expression, as shown, can
2628 // be determined to be constant.
2629 bool NewGVN::OpIsSafeForPHIOfOps(Value *V, const BasicBlock *PHIBlock,
2630  SmallPtrSetImpl<const Value *> &Visited) {
2632  if (!OpIsSafeForPHIOfOpsHelper(V, PHIBlock, Visited, Worklist))
2633  return false;
2634  while (!Worklist.empty()) {
2635  auto *I = Worklist.pop_back_val();
2636  if (!OpIsSafeForPHIOfOpsHelper(I, PHIBlock, Visited, Worklist))
2637  return false;
2638  }
2639  OpSafeForPHIOfOps.insert({V, true});
2640  return true;
2641 }
2642 
2643 // Try to find a leader for instruction TransInst, which is a phi translated
2644 // version of something in our original program. Visited is used to ensure we
2645 // don't infinite loop during translations of cycles. OrigInst is the
2646 // instruction in the original program, and PredBB is the predecessor we
2647 // translated it through.
2648 Value *NewGVN::findLeaderForInst(Instruction *TransInst,
2649  SmallPtrSetImpl<Value *> &Visited,
2650  MemoryAccess *MemAccess, Instruction *OrigInst,
2651  BasicBlock *PredBB) {
2652  unsigned IDFSNum = InstrToDFSNum(OrigInst);
2653  // Make sure it's marked as a temporary instruction.
2654  AllTempInstructions.insert(TransInst);
2655  // and make sure anything that tries to add it's DFS number is
2656  // redirected to the instruction we are making a phi of ops
2657  // for.
2658  TempToBlock.insert({TransInst, PredBB});
2659  InstrDFS.insert({TransInst, IDFSNum});
2660 
2661  auto Res = performSymbolicEvaluation(TransInst, Visited);
2662  const Expression *E = Res.Expr;
2663  addAdditionalUsers(Res, OrigInst);
2664  InstrDFS.erase(TransInst);
2665  AllTempInstructions.erase(TransInst);
2666  TempToBlock.erase(TransInst);
2667  if (MemAccess)
2668  TempToMemory.erase(TransInst);
2669  if (!E)
2670  return nullptr;
2671  auto *FoundVal = findPHIOfOpsLeader(E, OrigInst, PredBB);
2672  if (!FoundVal) {
2673  ExpressionToPhiOfOps[E].insert(OrigInst);
2674  LLVM_DEBUG(dbgs() << "Cannot find phi of ops operand for " << *TransInst
2675  << " in block " << getBlockName(PredBB) << "\n");
2676  return nullptr;
2677  }
2678  if (auto *SI = dyn_cast<StoreInst>(FoundVal))
2679  FoundVal = SI->getValueOperand();
2680  return FoundVal;
2681 }
2682 
2683 // When we see an instruction that is an op of phis, generate the equivalent phi
2684 // of ops form.
2685 const Expression *
2686 NewGVN::makePossiblePHIOfOps(Instruction *I,
2687  SmallPtrSetImpl<Value *> &Visited) {
2688  if (!okayForPHIOfOps(I))
2689  return nullptr;
2690 
2691  if (!Visited.insert(I).second)
2692  return nullptr;
2693  // For now, we require the instruction be cycle free because we don't
2694  // *always* create a phi of ops for instructions that could be done as phi
2695  // of ops, we only do it if we think it is useful. If we did do it all the
2696  // time, we could remove the cycle free check.
2697  if (!isCycleFree(I))
2698  return nullptr;
2699 
2700  SmallPtrSet<const Value *, 8> ProcessedPHIs;
2701  // TODO: We don't do phi translation on memory accesses because it's
2702  // complicated. For a load, we'd need to be able to simulate a new memoryuse,
2703  // which we don't have a good way of doing ATM.
2704  auto *MemAccess = getMemoryAccess(I);
2705  // If the memory operation is defined by a memory operation this block that
2706  // isn't a MemoryPhi, transforming the pointer backwards through a scalar phi
2707  // can't help, as it would still be killed by that memory operation.
2708  if (MemAccess && !isa<MemoryPhi>(MemAccess->getDefiningAccess()) &&
2709  MemAccess->getDefiningAccess()->getBlock() == I->getParent())
2710  return nullptr;
2711 
2712  // Convert op of phis to phi of ops
2713  SmallPtrSet<const Value *, 10> VisitedOps;
2714  SmallVector<Value *, 4> Ops(I->operand_values());
2715  BasicBlock *SamePHIBlock = nullptr;
2716  PHINode *OpPHI = nullptr;
2717  if (!DebugCounter::shouldExecute(PHIOfOpsCounter))
2718  return nullptr;
2719  for (auto *Op : Ops) {
2720  if (!isa<PHINode>(Op)) {
2721  auto *ValuePHI = RealToTemp.lookup(Op);
2722  if (!ValuePHI)
2723  continue;
2724  LLVM_DEBUG(dbgs() << "Found possible dependent phi of ops\n");
2725  Op = ValuePHI;
2726  }
2727  OpPHI = cast<PHINode>(Op);
2728  if (!SamePHIBlock) {
2729  SamePHIBlock = getBlockForValue(OpPHI);
2730  } else if (SamePHIBlock != getBlockForValue(OpPHI)) {
2731  LLVM_DEBUG(
2732  dbgs()
2733  << "PHIs for operands are not all in the same block, aborting\n");
2734  return nullptr;
2735  }
2736  // No point in doing this for one-operand phis.
2737  if (OpPHI->getNumOperands() == 1) {
2738  OpPHI = nullptr;
2739  continue;
2740  }
2741  }
2742 
2743  if (!OpPHI)
2744  return nullptr;
2745 
2746  SmallVector<ValPair, 4> PHIOps;
2748  auto *PHIBlock = getBlockForValue(OpPHI);
2749  RevisitOnReachabilityChange[PHIBlock].reset(InstrToDFSNum(I));
2750  for (unsigned PredNum = 0; PredNum < OpPHI->getNumOperands(); ++PredNum) {
2751  auto *PredBB = OpPHI->getIncomingBlock(PredNum);
2752  Value *FoundVal = nullptr;
2753  SmallPtrSet<Value *, 4> CurrentDeps;
2754  // We could just skip unreachable edges entirely but it's tricky to do
2755  // with rewriting existing phi nodes.
2756  if (ReachableEdges.count({PredBB, PHIBlock})) {
2757  // Clone the instruction, create an expression from it that is
2758  // translated back into the predecessor, and see if we have a leader.
2759  Instruction *ValueOp = I->clone();
2760  if (MemAccess)
2761  TempToMemory.insert({ValueOp, MemAccess});
2762  bool SafeForPHIOfOps = true;
2763  VisitedOps.clear();
2764  for (auto &Op : ValueOp->operands()) {
2765  auto *OrigOp = &*Op;
2766  // When these operand changes, it could change whether there is a
2767  // leader for us or not, so we have to add additional users.
2768  if (isa<PHINode>(Op)) {
2769  Op = Op->DoPHITranslation(PHIBlock, PredBB);
2770  if (Op != OrigOp && Op != I)
2771  CurrentDeps.insert(Op);
2772  } else if (auto *ValuePHI = RealToTemp.lookup(Op)) {
2773  if (getBlockForValue(ValuePHI) == PHIBlock)
2774  Op = ValuePHI->getIncomingValueForBlock(PredBB);
2775  }
2776  // If we phi-translated the op, it must be safe.
2777  SafeForPHIOfOps =
2778  SafeForPHIOfOps &&
2779  (Op != OrigOp || OpIsSafeForPHIOfOps(Op, PHIBlock, VisitedOps));
2780  }
2781  // FIXME: For those things that are not safe we could generate
2782  // expressions all the way down, and see if this comes out to a
2783  // constant. For anything where that is true, and unsafe, we should
2784  // have made a phi-of-ops (or value numbered it equivalent to something)
2785  // for the pieces already.
2786  FoundVal = !SafeForPHIOfOps ? nullptr
2787  : findLeaderForInst(ValueOp, Visited,
2788  MemAccess, I, PredBB);
2789  ValueOp->deleteValue();
2790  if (!FoundVal) {
2791  // We failed to find a leader for the current ValueOp, but this might
2792  // change in case of the translated operands change.
2793  if (SafeForPHIOfOps)
2794  for (auto Dep : CurrentDeps)
2795  addAdditionalUsers(Dep, I);
2796 
2797  return nullptr;
2798  }
2799  Deps.insert(CurrentDeps.begin(), CurrentDeps.end());
2800  } else {
2801  LLVM_DEBUG(dbgs() << "Skipping phi of ops operand for incoming block "
2802  << getBlockName(PredBB)
2803  << " because the block is unreachable\n");
2804  FoundVal = PoisonValue::get(I->getType());
2805  RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2806  }
2807 
2808  PHIOps.push_back({FoundVal, PredBB});
2809  LLVM_DEBUG(dbgs() << "Found phi of ops operand " << *FoundVal << " in "
2810  << getBlockName(PredBB) << "\n");
2811  }
2812  for (auto Dep : Deps)
2813  addAdditionalUsers(Dep, I);
2814  sortPHIOps(PHIOps);
2815  auto *E = performSymbolicPHIEvaluation(PHIOps, I, PHIBlock);
2816  if (isa<ConstantExpression>(E) || isa<VariableExpression>(E)) {
2817  LLVM_DEBUG(
2818  dbgs()
2819  << "Not creating real PHI of ops because it simplified to existing "
2820  "value or constant\n");
2821  // We have leaders for all operands, but do not create a real PHI node with
2822  // those leaders as operands, so the link between the operands and the
2823  // PHI-of-ops is not materialized in the IR. If any of those leaders
2824  // changes, the PHI-of-op may change also, so we need to add the operands as
2825  // additional users.
2826  for (auto &O : PHIOps)
2827  addAdditionalUsers(O.first, I);
2828 
2829  return E;
2830  }
2831  auto *ValuePHI = RealToTemp.lookup(I);
2832  bool NewPHI = false;
2833  if (!ValuePHI) {
2834  ValuePHI =
2835  PHINode::Create(I->getType(), OpPHI->getNumOperands(), "phiofops");
2836  addPhiOfOps(ValuePHI, PHIBlock, I);
2837  NewPHI = true;
2838  NumGVNPHIOfOpsCreated++;
2839  }
2840  if (NewPHI) {
2841  for (auto PHIOp : PHIOps)
2842  ValuePHI->addIncoming(PHIOp.first, PHIOp.second);
2843  } else {
2844  TempToBlock[ValuePHI] = PHIBlock;
2845  unsigned int i = 0;
2846  for (auto PHIOp : PHIOps) {
2847  ValuePHI->setIncomingValue(i, PHIOp.first);
2848  ValuePHI->setIncomingBlock(i, PHIOp.second);
2849  ++i;
2850  }
2851  }
2852  RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(I));
2853  LLVM_DEBUG(dbgs() << "Created phi of ops " << *ValuePHI << " for " << *I
2854  << "\n");
2855 
2856  return E;
2857 }
2858 
2859 // The algorithm initially places the values of the routine in the TOP
2860 // congruence class. The leader of TOP is the undetermined value `poison`.
2861 // When the algorithm has finished, values still in TOP are unreachable.
2862 void NewGVN::initializeCongruenceClasses(Function &F) {
2863  NextCongruenceNum = 0;
2864 
2865  // Note that even though we use the live on entry def as a representative
2866  // MemoryAccess, it is *not* the same as the actual live on entry def. We
2867  // have no real equivalent to poison for MemoryAccesses, and so we really
2868  // should be checking whether the MemoryAccess is top if we want to know if it
2869  // is equivalent to everything. Otherwise, what this really signifies is that
2870  // the access "it reaches all the way back to the beginning of the function"
2871 
2872  // Initialize all other instructions to be in TOP class.
2873  TOPClass = createCongruenceClass(nullptr, nullptr);
2874  TOPClass->setMemoryLeader(MSSA->getLiveOnEntryDef());
2875  // The live on entry def gets put into it's own class
2876  MemoryAccessToClass[MSSA->getLiveOnEntryDef()] =
2877  createMemoryClass(MSSA->getLiveOnEntryDef());
2878 
2879  for (auto DTN : nodes(DT)) {
2880  BasicBlock *BB = DTN->getBlock();
2881  // All MemoryAccesses are equivalent to live on entry to start. They must
2882  // be initialized to something so that initial changes are noticed. For
2883  // the maximal answer, we initialize them all to be the same as
2884  // liveOnEntry.
2885  auto *MemoryBlockDefs = MSSA->getBlockDefs(BB);
2886  if (MemoryBlockDefs)
2887  for (const auto &Def : *MemoryBlockDefs) {
2888  MemoryAccessToClass[&Def] = TOPClass;
2889  auto *MD = dyn_cast<MemoryDef>(&Def);
2890  // Insert the memory phis into the member list.
2891  if (!MD) {
2892  const MemoryPhi *MP = cast<MemoryPhi>(&Def);
2893  TOPClass->memory_insert(MP);
2894  MemoryPhiState.insert({MP, MPS_TOP});
2895  }
2896 
2897  if (MD && isa<StoreInst>(MD->getMemoryInst()))
2898  TOPClass->incStoreCount();
2899  }
2900 
2901  // FIXME: This is trying to discover which instructions are uses of phi
2902  // nodes. We should move this into one of the myriad of places that walk
2903  // all the operands already.
2904  for (auto &I : *BB) {
2905  if (isa<PHINode>(&I))
2906  for (auto *U : I.users())
2907  if (auto *UInst = dyn_cast<Instruction>(U))
2908  if (InstrToDFSNum(UInst) != 0 && okayForPHIOfOps(UInst))
2909  PHINodeUses.insert(UInst);
2910  // Don't insert void terminators into the class. We don't value number
2911  // them, and they just end up sitting in TOP.
2912  if (I.isTerminator() && I.getType()->isVoidTy())
2913  continue;
2914  TOPClass->insert(&I);
2915  ValueToClass[&I] = TOPClass;
2916  }
2917  }
2918 
2919  // Initialize arguments to be in their own unique congruence classes
2920  for (auto &FA : F.args())
2921  createSingletonCongruenceClass(&FA);
2922 }
2923 
2924 void NewGVN::cleanupTables() {
2925  for (unsigned i = 0, e = CongruenceClasses.size(); i != e; ++i) {
2926  LLVM_DEBUG(dbgs() << "Congruence class " << CongruenceClasses[i]->getID()
2927  << " has " << CongruenceClasses[i]->size()
2928  << " members\n");
2929  // Make sure we delete the congruence class (probably worth switching to
2930  // a unique_ptr at some point.
2931  delete CongruenceClasses[i];
2932  CongruenceClasses[i] = nullptr;
2933  }
2934 
2935  // Destroy the value expressions
2936  SmallVector<Instruction *, 8> TempInst(AllTempInstructions.begin(),
2937  AllTempInstructions.end());
2938  AllTempInstructions.clear();
2939 
2940  // We have to drop all references for everything first, so there are no uses
2941  // left as we delete them.
2942  for (auto *I : TempInst) {
2943  I->dropAllReferences();
2944  }
2945 
2946  while (!TempInst.empty()) {
2947  auto *I = TempInst.pop_back_val();
2948  I->deleteValue();
2949  }
2950 
2951  ValueToClass.clear();
2952  ArgRecycler.clear(ExpressionAllocator);
2953  ExpressionAllocator.Reset();
2954  CongruenceClasses.clear();
2955  ExpressionToClass.clear();
2956  ValueToExpression.clear();
2957  RealToTemp.clear();
2958  AdditionalUsers.clear();
2959  ExpressionToPhiOfOps.clear();
2960  TempToBlock.clear();
2961  TempToMemory.clear();
2962  PHINodeUses.clear();
2963  OpSafeForPHIOfOps.clear();
2964  ReachableBlocks.clear();
2965  ReachableEdges.clear();
2966 #ifndef NDEBUG
2967  ProcessedCount.clear();
2968 #endif
2969  InstrDFS.clear();
2970  InstructionsToErase.clear();
2971  DFSToInstr.clear();
2972  BlockInstRange.clear();
2973  TouchedInstructions.clear();
2974  MemoryAccessToClass.clear();
2975  PredicateToUsers.clear();
2976  MemoryToUsers.clear();
2977  RevisitOnReachabilityChange.clear();
2978  IntrinsicInstPred.clear();
2979 }
2980 
2981 // Assign local DFS number mapping to instructions, and leave space for Value
2982 // PHI's.
2983 std::pair<unsigned, unsigned> NewGVN::assignDFSNumbers(BasicBlock *B,
2984  unsigned Start) {
2985  unsigned End = Start;
2986  if (MemoryAccess *MemPhi = getMemoryAccess(B)) {
2987  InstrDFS[MemPhi] = End++;
2988  DFSToInstr.emplace_back(MemPhi);
2989  }
2990 
2991  // Then the real block goes next.
2992  for (auto &I : *B) {
2993  // There's no need to call isInstructionTriviallyDead more than once on
2994  // an instruction. Therefore, once we know that an instruction is dead
2995  // we change its DFS number so that it doesn't get value numbered.
2996  if (isInstructionTriviallyDead(&I, TLI)) {
2997  InstrDFS[&I] = 0;
2998  LLVM_DEBUG(dbgs() << "Skipping trivially dead instruction " << I << "\n");
2999  markInstructionForDeletion(&I);
3000  continue;
3001  }
3002  if (isa<PHINode>(&I))
3003  RevisitOnReachabilityChange[B].set(End);
3004  InstrDFS[&I] = End++;
3005  DFSToInstr.emplace_back(&I);
3006  }
3007 
3008  // All of the range functions taken half-open ranges (open on the end side).
3009  // So we do not subtract one from count, because at this point it is one
3010  // greater than the last instruction.
3011  return std::make_pair(Start, End);
3012 }
3013 
3014 void NewGVN::updateProcessedCount(const Value *V) {
3015 #ifndef NDEBUG
3016  if (ProcessedCount.count(V) == 0) {
3017  ProcessedCount.insert({V, 1});
3018  } else {
3019  ++ProcessedCount[V];
3020  assert(ProcessedCount[V] < 100 &&
3021  "Seem to have processed the same Value a lot");
3022  }
3023 #endif
3024 }
3025 
3026 // Evaluate MemoryPhi nodes symbolically, just like PHI nodes
3027 void NewGVN::valueNumberMemoryPhi(MemoryPhi *MP) {
3028  // If all the arguments are the same, the MemoryPhi has the same value as the
3029  // argument. Filter out unreachable blocks and self phis from our operands.
3030  // TODO: We could do cycle-checking on the memory phis to allow valueizing for
3031  // self-phi checking.
3032  const BasicBlock *PHIBlock = MP->getBlock();
3033  auto Filtered = make_filter_range(MP->operands(), [&](const Use &U) {
3034  return cast<MemoryAccess>(U) != MP &&
3035  !isMemoryAccessTOP(cast<MemoryAccess>(U)) &&
3036  ReachableEdges.count({MP->getIncomingBlock(U), PHIBlock});
3037  });
3038  // If all that is left is nothing, our memoryphi is poison. We keep it as
3039  // InitialClass. Note: The only case this should happen is if we have at
3040  // least one self-argument.
3041  if (Filtered.begin() == Filtered.end()) {
3042  if (setMemoryClass(MP, TOPClass))
3043  markMemoryUsersTouched(MP);
3044  return;
3045  }
3046 
3047  // Transform the remaining operands into operand leaders.
3048  // FIXME: mapped_iterator should have a range version.
3049  auto LookupFunc = [&](const Use &U) {
3050  return lookupMemoryLeader(cast<MemoryAccess>(U));
3051  };
3052  auto MappedBegin = map_iterator(Filtered.begin(), LookupFunc);
3053  auto MappedEnd = map_iterator(Filtered.end(), LookupFunc);
3054 
3055  // and now check if all the elements are equal.
3056  // Sadly, we can't use std::equals since these are random access iterators.
3057  const auto *AllSameValue = *MappedBegin;
3058  ++MappedBegin;
3059  bool AllEqual = std::all_of(
3060  MappedBegin, MappedEnd,
3061  [&AllSameValue](const MemoryAccess *V) { return V == AllSameValue; });
3062 
3063  if (AllEqual)
3064  LLVM_DEBUG(dbgs() << "Memory Phi value numbered to " << *AllSameValue
3065  << "\n");
3066  else
3067  LLVM_DEBUG(dbgs() << "Memory Phi value numbered to itself\n");
3068  // If it's equal to something, it's in that class. Otherwise, it has to be in
3069  // a class where it is the leader (other things may be equivalent to it, but
3070  // it needs to start off in its own class, which means it must have been the
3071  // leader, and it can't have stopped being the leader because it was never
3072  // removed).
3073  CongruenceClass *CC =
3074  AllEqual ? getMemoryClass(AllSameValue) : ensureLeaderOfMemoryClass(MP);
3075  auto OldState = MemoryPhiState.lookup(MP);
3076  assert(OldState != MPS_Invalid && "Invalid memory phi state");
3077  auto NewState = AllEqual ? MPS_Equivalent : MPS_Unique;
3078  MemoryPhiState[MP] = NewState;
3079  if (setMemoryClass(MP, CC) || OldState != NewState)
3080  markMemoryUsersTouched(MP);
3081 }
3082 
3083 // Value number a single instruction, symbolically evaluating, performing
3084 // congruence finding, and updating mappings.
3085 void NewGVN::valueNumberInstruction(Instruction *I) {
3086  LLVM_DEBUG(dbgs() << "Processing instruction " << *I << "\n");
3087  if (!I->isTerminator()) {
3088  const Expression *Symbolized = nullptr;
3089  SmallPtrSet<Value *, 2> Visited;
3090  if (DebugCounter::shouldExecute(VNCounter)) {
3091  auto Res = performSymbolicEvaluation(I, Visited);
3092  Symbolized = Res.Expr;
3093  addAdditionalUsers(Res, I);
3094 
3095  // Make a phi of ops if necessary
3096  if (Symbolized && !isa<ConstantExpression>(Symbolized) &&
3097  !isa<VariableExpression>(Symbolized) && PHINodeUses.count(I)) {
3098  auto *PHIE = makePossiblePHIOfOps(I, Visited);
3099  // If we created a phi of ops, use it.
3100  // If we couldn't create one, make sure we don't leave one lying around
3101  if (PHIE) {
3102  Symbolized = PHIE;
3103  } else if (auto *Op = RealToTemp.lookup(I)) {
3104  removePhiOfOps(I, Op);
3105  }
3106  }
3107  } else {
3108  // Mark the instruction as unused so we don't value number it again.
3109  InstrDFS[I] = 0;
3110  }
3111  // If we couldn't come up with a symbolic expression, use the unknown
3112  // expression
3113  if (Symbolized == nullptr)
3114  Symbolized = createUnknownExpression(I);
3115  performCongruenceFinding(I, Symbolized);
3116  } else {
3117  // Handle terminators that return values. All of them produce values we
3118  // don't currently understand. We don't place non-value producing
3119  // terminators in a class.
3120  if (!I->getType()->isVoidTy()) {
3121  auto *Symbolized = createUnknownExpression(I);
3122  performCongruenceFinding(I, Symbolized);
3123  }
3124  processOutgoingEdges(I, I->getParent());
3125  }
3126 }
3127 
3128 // Check if there is a path, using single or equal argument phi nodes, from
3129 // First to Second.
3130 bool NewGVN::singleReachablePHIPath(
3132  const MemoryAccess *Second) const {
3133  if (First == Second)
3134  return true;
3135  if (MSSA->isLiveOnEntryDef(First))
3136  return false;
3137 
3138  // This is not perfect, but as we're just verifying here, we can live with
3139  // the loss of precision. The real solution would be that of doing strongly
3140  // connected component finding in this routine, and it's probably not worth
3141  // the complexity for the time being. So, we just keep a set of visited
3142  // MemoryAccess and return true when we hit a cycle.
3143  if (Visited.count(First))
3144  return true;
3145  Visited.insert(First);
3146 
3147  const auto *EndDef = First;
3148  for (auto *ChainDef : optimized_def_chain(First)) {
3149  if (ChainDef == Second)
3150  return true;
3151  if (MSSA->isLiveOnEntryDef(ChainDef))
3152  return false;
3153  EndDef = ChainDef;
3154  }
3155  auto *MP = cast<MemoryPhi>(EndDef);
3156  auto ReachableOperandPred = [&](const Use &U) {
3157  return ReachableEdges.count({MP->getIncomingBlock(U), MP->getBlock()});
3158  };
3159  auto FilteredPhiArgs =
3160  make_filter_range(MP->operands(), ReachableOperandPred);
3161  SmallVector<const Value *, 32> OperandList;
3162  llvm::copy(FilteredPhiArgs, std::back_inserter(OperandList));
3163  bool Okay = is_splat(OperandList);
3164  if (Okay)
3165  return singleReachablePHIPath(Visited, cast<MemoryAccess>(OperandList[0]),
3166  Second);
3167  return false;
3168 }
3169 
3170 // Verify the that the memory equivalence table makes sense relative to the
3171 // congruence classes. Note that this checking is not perfect, and is currently
3172 // subject to very rare false negatives. It is only useful for
3173 // testing/debugging.
3174 void NewGVN::verifyMemoryCongruency() const {
3175 #ifndef NDEBUG
3176  // Verify that the memory table equivalence and memory member set match
3177  for (const auto *CC : CongruenceClasses) {
3178  if (CC == TOPClass || CC->isDead())
3179  continue;
3180  if (CC->getStoreCount() != 0) {
3181  assert((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) &&
3182  "Any class with a store as a leader should have a "
3183  "representative stored value");
3184  assert(CC->getMemoryLeader() &&
3185  "Any congruence class with a store should have a "
3186  "representative access");
3187  }
3188 
3189  if (CC->getMemoryLeader())
3190  assert(MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
3191  "Representative MemoryAccess does not appear to be reverse "
3192  "mapped properly");
3193  for (auto M : CC->memory())
3194  assert(MemoryAccessToClass.lookup(M) == CC &&
3195  "Memory member does not appear to be reverse mapped properly");
3196  }
3197 
3198  // Anything equivalent in the MemoryAccess table should be in the same
3199  // congruence class.
3200 
3201  // Filter out the unreachable and trivially dead entries, because they may
3202  // never have been updated if the instructions were not processed.
3203  auto ReachableAccessPred =
3204  [&](const std::pair<const MemoryAccess *, CongruenceClass *> Pair) {
3205  bool Result = ReachableBlocks.count(Pair.first->getBlock());
3206  if (!Result || MSSA->isLiveOnEntryDef(Pair.first) ||
3207  MemoryToDFSNum(Pair.first) == 0)
3208  return false;
3209  if (auto *MemDef = dyn_cast<MemoryDef>(Pair.first))
3210  return !isInstructionTriviallyDead(MemDef->getMemoryInst());
3211 
3212  // We could have phi nodes which operands are all trivially dead,
3213  // so we don't process them.
3214  if (auto *MemPHI = dyn_cast<MemoryPhi>(Pair.first)) {
3215  for (auto &U : MemPHI->incoming_values()) {
3216  if (auto *I = dyn_cast<Instruction>(&*U)) {
3218  return true;
3219  }
3220  }
3221  return false;
3222  }
3223 
3224  return true;
3225  };
3226 
3227  auto Filtered = make_filter_range(MemoryAccessToClass, ReachableAccessPred);
3228  for (auto KV : Filtered) {
3229  if (auto *FirstMUD = dyn_cast<MemoryUseOrDef>(KV.first)) {
3230  auto *SecondMUD = dyn_cast<MemoryUseOrDef>(KV.second->getMemoryLeader());
3231  if (FirstMUD && SecondMUD) {
3233  assert((singleReachablePHIPath(VisitedMAS, FirstMUD, SecondMUD) ||
3234  ValueToClass.lookup(FirstMUD->getMemoryInst()) ==
3235  ValueToClass.lookup(SecondMUD->getMemoryInst())) &&
3236  "The instructions for these memory operations should have "
3237  "been in the same congruence class or reachable through"
3238  "a single argument phi");
3239  }
3240  } else if (auto *FirstMP = dyn_cast<MemoryPhi>(KV.first)) {
3241  // We can only sanely verify that MemoryDefs in the operand list all have
3242  // the same class.
3243  auto ReachableOperandPred = [&](const Use &U) {
3244  return ReachableEdges.count(
3245  {FirstMP->getIncomingBlock(U), FirstMP->getBlock()}) &&
3246  isa<MemoryDef>(U);
3247 
3248  };
3249  // All arguments should in the same class, ignoring unreachable arguments
3250  auto FilteredPhiArgs =
3251  make_filter_range(FirstMP->operands(), ReachableOperandPred);
3253  std::transform(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
3254  std::back_inserter(PhiOpClasses), [&](const Use &U) {
3255  const MemoryDef *MD = cast<MemoryDef>(U);
3256  return ValueToClass.lookup(MD->getMemoryInst());
3257  });
3258  assert(is_splat(PhiOpClasses) &&
3259  "All MemoryPhi arguments should be in the same class");
3260  }
3261  }
3262 #endif
3263 }
3264 
3265 // Verify that the sparse propagation we did actually found the maximal fixpoint
3266 // We do this by storing the value to class mapping, touching all instructions,
3267 // and redoing the iteration to see if anything changed.
3268 void NewGVN::verifyIterationSettled(Function &F) {
3269 #ifndef NDEBUG
3270  LLVM_DEBUG(dbgs() << "Beginning iteration verification\n");
3271  if (DebugCounter::isCounterSet(VNCounter))
3272  DebugCounter::setCounterValue(VNCounter, StartingVNCounter);
3273 
3274  // Note that we have to store the actual classes, as we may change existing
3275  // classes during iteration. This is because our memory iteration propagation
3276  // is not perfect, and so may waste a little work. But it should generate
3277  // exactly the same congruence classes we have now, with different IDs.
3278  std::map<const Value *, CongruenceClass> BeforeIteration;
3279 
3280  for (auto &KV : ValueToClass) {
3281  if (auto *I = dyn_cast<Instruction>(KV.first))
3282  // Skip unused/dead instructions.
3283  if (InstrToDFSNum(I) == 0)
3284  continue;
3285  BeforeIteration.insert({KV.first, *KV.second});
3286  }
3287 
3288  TouchedInstructions.set();
3289  TouchedInstructions.reset(0);
3290  iterateTouchedInstructions();
3292  EqualClasses;
3293  for (const auto &KV : ValueToClass) {
3294  if (auto *I = dyn_cast<Instruction>(KV.first))
3295  // Skip unused/dead instructions.
3296  if (InstrToDFSNum(I) == 0)
3297  continue;
3298  // We could sink these uses, but i think this adds a bit of clarity here as
3299  // to what we are comparing.
3300  auto *BeforeCC = &BeforeIteration.find(KV.first)->second;
3301  auto *AfterCC = KV.second;
3302  // Note that the classes can't change at this point, so we memoize the set
3303  // that are equal.
3304  if (!EqualClasses.count({BeforeCC, AfterCC})) {
3305  assert(BeforeCC->isEquivalentTo(AfterCC) &&
3306  "Value number changed after main loop completed!");
3307  EqualClasses.insert({BeforeCC, AfterCC});
3308  }
3309  }
3310 #endif
3311 }
3312 
3313 // Verify that for each store expression in the expression to class mapping,
3314 // only the latest appears, and multiple ones do not appear.
3315 // Because loads do not use the stored value when doing equality with stores,
3316 // if we don't erase the old store expressions from the table, a load can find
3317 // a no-longer valid StoreExpression.
3318 void NewGVN::verifyStoreExpressions() const {
3319 #ifndef NDEBUG
3320  // This is the only use of this, and it's not worth defining a complicated
3321  // densemapinfo hash/equality function for it.
3322  std::set<
3323  std::pair<const Value *,
3324  std::tuple<const Value *, const CongruenceClass *, Value *>>>
3325  StoreExpressionSet;
3326  for (const auto &KV : ExpressionToClass) {
3327  if (auto *SE = dyn_cast<StoreExpression>(KV.first)) {
3328  // Make sure a version that will conflict with loads is not already there
3329  auto Res = StoreExpressionSet.insert(
3330  {SE->getOperand(0), std::make_tuple(SE->getMemoryLeader(), KV.second,
3331  SE->getStoredValue())});
3332  bool Okay = Res.second;
3333  // It's okay to have the same expression already in there if it is
3334  // identical in nature.
3335  // This can happen when the leader of the stored value changes over time.
3336  if (!Okay)
3337  Okay = (std::get<1>(Res.first->second) == KV.second) &&
3338  (lookupOperandLeader(std::get<2>(Res.first->second)) ==
3339  lookupOperandLeader(SE->getStoredValue()));
3340  assert(Okay && "Stored expression conflict exists in expression table");
3341  auto *ValueExpr = ValueToExpression.lookup(SE->getStoreInst());
3342  assert(ValueExpr && ValueExpr->equals(*SE) &&
3343  "StoreExpression in ExpressionToClass is not latest "
3344  "StoreExpression for value");
3345  }
3346  }
3347 #endif
3348 }
3349 
3350 // This is the main value numbering loop, it iterates over the initial touched
3351 // instruction set, propagating value numbers, marking things touched, etc,
3352 // until the set of touched instructions is completely empty.
3353 void NewGVN::iterateTouchedInstructions() {
3354  unsigned int Iterations = 0;
3355  // Figure out where touchedinstructions starts
3356  int FirstInstr = TouchedInstructions.find_first();
3357  // Nothing set, nothing to iterate, just return.
3358  if (FirstInstr == -1)
3359  return;
3360  const BasicBlock *LastBlock = getBlockForValue(InstrFromDFSNum(FirstInstr));
3361  while (TouchedInstructions.any()) {
3362  ++Iterations;
3363  // Walk through all the instructions in all the blocks in RPO.
3364  // TODO: As we hit a new block, we should push and pop equalities into a
3365  // table lookupOperandLeader can use, to catch things PredicateInfo
3366  // might miss, like edge-only equivalences.
3367  for (unsigned InstrNum : TouchedInstructions.set_bits()) {
3368 
3369  // This instruction was found to be dead. We don't bother looking
3370  // at it again.
3371  if (InstrNum == 0) {
3372  TouchedInstructions.reset(InstrNum);
3373  continue;
3374  }
3375 
3376  Value *V = InstrFromDFSNum(InstrNum);
3377  const BasicBlock *CurrBlock = getBlockForValue(V);
3378 
3379  // If we hit a new block, do reachability processing.
3380  if (CurrBlock != LastBlock) {
3381  LastBlock = CurrBlock;
3382  bool BlockReachable = ReachableBlocks.count(CurrBlock);
3383  const auto &CurrInstRange = BlockInstRange.lookup(CurrBlock);
3384 
3385  // If it's not reachable, erase any touched instructions and move on.
3386  if (!BlockReachable) {
3387  TouchedInstructions.reset(CurrInstRange.first, CurrInstRange.second);
3388  LLVM_DEBUG(dbgs() << "Skipping instructions in block "
3389  << getBlockName(CurrBlock)
3390  << " because it is unreachable\n");
3391  continue;
3392  }
3393  updateProcessedCount(CurrBlock);
3394  }
3395  // Reset after processing (because we may mark ourselves as touched when
3396  // we propagate equalities).
3397  TouchedInstructions.reset(InstrNum);
3398 
3399  if (auto *MP = dyn_cast<MemoryPhi>(V)) {
3400  LLVM_DEBUG(dbgs() << "Processing MemoryPhi " << *MP << "\n");
3401  valueNumberMemoryPhi(MP);
3402  } else if (auto *I = dyn_cast<Instruction>(V)) {
3403  valueNumberInstruction(I);
3404  } else {
3405  llvm_unreachable("Should have been a MemoryPhi or Instruction");
3406  }
3407  updateProcessedCount(V);
3408  }
3409  }
3410  NumGVNMaxIterations = std::max(NumGVNMaxIterations.getValue(), Iterations);
3411 }
3412 
3413 // This is the main transformation entry point.
3414 bool NewGVN::runGVN() {
3415  if (DebugCounter::isCounterSet(VNCounter))
3416  StartingVNCounter = DebugCounter::getCounterValue(VNCounter);
3417  bool Changed = false;
3418  NumFuncArgs = F.arg_size();
3419  MSSAWalker = MSSA->getWalker();
3420  SingletonDeadExpression = new (ExpressionAllocator) DeadExpression();
3421 
3422  // Count number of instructions for sizing of hash tables, and come
3423  // up with a global dfs numbering for instructions.
3424  unsigned ICount = 1;
3425  // Add an empty instruction to account for the fact that we start at 1
3426  DFSToInstr.emplace_back(nullptr);
3427  // Note: We want ideal RPO traversal of the blocks, which is not quite the
3428  // same as dominator tree order, particularly with regard whether backedges
3429  // get visited first or second, given a block with multiple successors.
3430  // If we visit in the wrong order, we will end up performing N times as many
3431  // iterations.
3432  // The dominator tree does guarantee that, for a given dom tree node, it's
3433  // parent must occur before it in the RPO ordering. Thus, we only need to sort
3434  // the siblings.
3436  unsigned Counter = 0;
3437  for (auto &B : RPOT) {
3438  auto *Node = DT->getNode(B);
3439  assert(Node && "RPO and Dominator tree should have same reachability");
3440  RPOOrdering[Node] = ++Counter;
3441  }
3442  // Sort dominator tree children arrays into RPO.
3443  for (auto &B : RPOT) {
3444  auto *Node = DT->getNode(B);
3445  if (Node->getNumChildren() > 1)
3446  llvm::sort(*Node, [&](const DomTreeNode *A, const DomTreeNode *B) {
3447  return RPOOrdering[A] < RPOOrdering[B];
3448  });
3449  }
3450 
3451  // Now a standard depth first ordering of the domtree is equivalent to RPO.
3452  for (auto DTN : depth_first(DT->getRootNode())) {
3453  BasicBlock *B = DTN->getBlock();
3454  const auto &BlockRange = assignDFSNumbers(B, ICount);
3455  BlockInstRange.insert({B, BlockRange});
3456  ICount += BlockRange.second - BlockRange.first;
3457  }
3458  initializeCongruenceClasses(F);
3459 
3460  TouchedInstructions.resize(ICount);
3461  // Ensure we don't end up resizing the expressionToClass map, as
3462  // that can be quite expensive. At most, we have one expression per
3463  // instruction.
3464  ExpressionToClass.reserve(ICount);
3465 
3466  // Initialize the touched instructions to include the entry block.
3467  const auto &InstRange = BlockInstRange.lookup(&F.getEntryBlock());
3468  TouchedInstructions.set(InstRange.first, InstRange.second);
3469  LLVM_DEBUG(dbgs() << "Block " << getBlockName(&F.getEntryBlock())
3470  << " marked reachable\n");
3471  ReachableBlocks.insert(&F.getEntryBlock());
3472 
3473  iterateTouchedInstructions();
3474  verifyMemoryCongruency();
3475  verifyIterationSettled(F);
3476  verifyStoreExpressions();
3477 
3478  Changed |= eliminateInstructions(F);
3479 
3480  // Delete all instructions marked for deletion.
3481  for (Instruction *ToErase : InstructionsToErase) {
3482  if (!ToErase->use_empty())
3483  ToErase->replaceAllUsesWith(PoisonValue::get(ToErase->getType()));
3484 
3485  assert(ToErase->getParent() &&
3486  "BB containing ToErase deleted unexpectedly!");
3487  ToErase->eraseFromParent();
3488  }
3489  Changed |= !InstructionsToErase.empty();
3490 
3491  // Delete all unreachable blocks.
3492  auto UnreachableBlockPred = [&](const BasicBlock &BB) {
3493  return !ReachableBlocks.count(&BB);
3494  };
3495 
3496  for (auto &BB : make_filter_range(F, UnreachableBlockPred)) {
3497  LLVM_DEBUG(dbgs() << "We believe block " << getBlockName(&BB)
3498  << " is unreachable\n");
3499  deleteInstructionsInBlock(&BB);
3500  Changed = true;
3501  }
3502 
3503  cleanupTables();
3504  return Changed;
3505 }
3506 
3508  int DFSIn = 0;
3509  int DFSOut = 0;
3510  int LocalNum = 0;
3511 
3512  // Only one of Def and U will be set.
3513  // The bool in the Def tells us whether the Def is the stored value of a
3514  // store.
3516  Use *U = nullptr;
3517 
3518  bool operator<(const ValueDFS &Other) const {
3519  // It's not enough that any given field be less than - we have sets
3520  // of fields that need to be evaluated together to give a proper ordering.
3521  // For example, if you have;
3522  // DFS (1, 3)
3523  // Val 0
3524  // DFS (1, 2)
3525  // Val 50
3526  // We want the second to be less than the first, but if we just go field
3527  // by field, we will get to Val 0 < Val 50 and say the first is less than
3528  // the second. We only want it to be less than if the DFS orders are equal.
3529  //
3530  // Each LLVM instruction only produces one value, and thus the lowest-level
3531  // differentiator that really matters for the stack (and what we use as as a
3532  // replacement) is the local dfs number.
3533  // Everything else in the structure is instruction level, and only affects
3534  // the order in which we will replace operands of a given instruction.
3535  //
3536  // For a given instruction (IE things with equal dfsin, dfsout, localnum),
3537  // the order of replacement of uses does not matter.
3538  // IE given,
3539  // a = 5
3540  // b = a + a
3541  // When you hit b, you will have two valuedfs with the same dfsin, out, and
3542  // localnum.
3543  // The .val will be the same as well.
3544  // The .u's will be different.
3545  // You will replace both, and it does not matter what order you replace them
3546  // in (IE whether you replace operand 2, then operand 1, or operand 1, then
3547  // operand 2).
3548  // Similarly for the case of same dfsin, dfsout, localnum, but different
3549  // .val's
3550  // a = 5
3551  // b = 6
3552  // c = a + b
3553  // in c, we will a valuedfs for a, and one for b,with everything the same
3554  // but .val and .u.
3555  // It does not matter what order we replace these operands in.
3556  // You will always end up with the same IR, and this is guaranteed.
3557  return std::tie(DFSIn, DFSOut, LocalNum, Def, U) <
3558  std::tie(Other.DFSIn, Other.DFSOut, Other.LocalNum, Other.Def,
3559  Other.U);
3560  }
3561 };
3562 
3563 // This function converts the set of members for a congruence class from values,
3564 // to sets of defs and uses with associated DFS info. The total number of
3565 // reachable uses for each value is stored in UseCount, and instructions that
3566 // seem
3567 // dead (have no non-dead uses) are stored in ProbablyDead.
3568 void NewGVN::convertClassToDFSOrdered(
3569  const CongruenceClass &Dense, SmallVectorImpl<ValueDFS> &DFSOrderedSet,
3571  SmallPtrSetImpl<Instruction *> &ProbablyDead) const {
3572  for (auto D : Dense) {
3573  // First add the value.
3574  BasicBlock *BB = getBlockForValue(D);
3575  // Constants are handled prior to ever calling this function, so
3576  // we should only be left with instructions as members.
3577  assert(BB && "Should have figured out a basic block for value");
3578  ValueDFS VDDef;
3579  DomTreeNode *DomNode = DT->getNode(BB);
3580  VDDef.DFSIn = DomNode->getDFSNumIn();
3581  VDDef.DFSOut = DomNode->getDFSNumOut();
3582  // If it's a store, use the leader of the value operand, if it's always
3583  // available, or the value operand. TODO: We could do dominance checks to
3584  // find a dominating leader, but not worth it ATM.
3585  if (auto *SI = dyn_cast<StoreInst>(D)) {
3586  auto Leader = lookupOperandLeader(SI->getValueOperand());
3587  if (alwaysAvailable(Leader)) {
3588  VDDef.Def.setPointer(Leader);
3589  } else {
3590  VDDef.Def.setPointer(SI->getValueOperand());
3591  VDDef.Def.setInt(true);
3592  }
3593  } else {
3594  VDDef.Def.setPointer(D);
3595  }
3596  assert(isa<Instruction>(D) &&
3597  "The dense set member should always be an instruction");
3598  Instruction *Def = cast<Instruction>(D);
3599  VDDef.LocalNum = InstrToDFSNum(D);
3600  DFSOrderedSet.push_back(VDDef);
3601  // If there is a phi node equivalent, add it
3602  if (auto *PN = RealToTemp.lookup(Def)) {
3603  auto *PHIE =
3604  dyn_cast_or_null<PHIExpression>(ValueToExpression.lookup(Def));
3605  if (PHIE) {
3606  VDDef.Def.setInt(false);
3607  VDDef.Def.setPointer(PN);
3608  VDDef.LocalNum = 0;
3609  DFSOrderedSet.push_back(VDDef);
3610  }
3611  }
3612 
3613  unsigned int UseCount = 0;
3614  // Now add the uses.
3615  for (auto &U : Def->uses()) {
3616  if (auto *I = dyn_cast<Instruction>(U.getUser())) {
3617  // Don't try to replace into dead uses
3618  if (InstructionsToErase.count(I))
3619  continue;
3620  ValueDFS VDUse;
3621  // Put the phi node uses in the incoming block.
3622  BasicBlock *IBlock;
3623  if (auto *P = dyn_cast<PHINode>(I)) {
3624  IBlock = P->getIncomingBlock(U);
3625  // Make phi node users appear last in the incoming block
3626  // they are from.
3627  VDUse.LocalNum = InstrDFS.size() + 1;
3628  } else {
3629  IBlock = getBlockForValue(I);
3630  VDUse.LocalNum = InstrToDFSNum(I);
3631  }
3632 
3633  // Skip uses in unreachable blocks, as we're going
3634  // to delete them.
3635  if (!ReachableBlocks.contains(IBlock))
3636  continue;
3637 
3638  DomTreeNode *DomNode = DT->getNode(IBlock);
3639  VDUse.DFSIn = DomNode->getDFSNumIn();
3640  VDUse.DFSOut = DomNode->getDFSNumOut();
3641  VDUse.U = &U;
3642  ++UseCount;
3643  DFSOrderedSet.emplace_back(VDUse);
3644  }
3645  }
3646 
3647  // If there are no uses, it's probably dead (but it may have side-effects,
3648  // so not definitely dead. Otherwise, store the number of uses so we can
3649  // track if it becomes dead later).
3650  if (UseCount == 0)
3651  ProbablyDead.insert(Def);
3652  else
3653  UseCounts[Def] = UseCount;
3654  }
3655 }
3656 
3657 // This function converts the set of members for a congruence class from values,
3658 // to the set of defs for loads and stores, with associated DFS info.
3659 void NewGVN::convertClassToLoadsAndStores(
3660  const CongruenceClass &Dense,
3661  SmallVectorImpl<ValueDFS> &LoadsAndStores) const {
3662  for (auto D : Dense) {
3663  if (!isa<LoadInst>(D) && !isa<StoreInst>(D))
3664  continue;
3665 
3666  BasicBlock *BB = getBlockForValue(D);
3667  ValueDFS VD;
3668  DomTreeNode *DomNode = DT->getNode(BB);
3669  VD.DFSIn = DomNode->getDFSNumIn();
3670  VD.DFSOut = DomNode->getDFSNumOut();
3671  VD.Def.setPointer(D);
3672 
3673  // If it's an instruction, use the real local dfs number.
3674  if (auto *I = dyn_cast<Instruction>(D))
3675  VD.LocalNum = InstrToDFSNum(I);
3676  else
3677  llvm_unreachable("Should have been an instruction");
3678 
3679  LoadsAndStores.emplace_back(VD);
3680  }
3681 }
3682 
3685  I->replaceAllUsesWith(Repl);
3686 }
3687 
3688 void NewGVN::deleteInstructionsInBlock(BasicBlock *BB) {
3689  LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
3690  ++NumGVNBlocksDeleted;
3691 
3692  // Delete the instructions backwards, as it has a reduced likelihood of having
3693  // to update as many def-use and use-def chains. Start after the terminator.
3694  auto StartPoint = BB->rbegin();
3695  ++StartPoint;
3696  // Note that we explicitly recalculate BB->rend() on each iteration,
3697  // as it may change when we remove the first instruction.
3698  for (BasicBlock::reverse_iterator I(StartPoint); I != BB->rend();) {
3699  Instruction &Inst = *I++;
3700  if (!Inst.use_empty())
3702  if (isa<LandingPadInst>(Inst))
3703  continue;
3704  salvageKnowledge(&Inst, AC);
3705 
3706  Inst.eraseFromParent();
3707  ++NumGVNInstrDeleted;
3708  }
3709  // Now insert something that simplifycfg will turn into an unreachable.
3710  Type *Int8Ty = Type::getInt8Ty(BB->getContext());
3711  new StoreInst(PoisonValue::get(Int8Ty),
3713  BB->getTerminator());
3714 }
3715 
3716 void NewGVN::markInstructionForDeletion(Instruction *I) {
3717  LLVM_DEBUG(dbgs() << "Marking " << *I << " for deletion\n");
3718  InstructionsToErase.insert(I);
3719 }
3720 
3721 void NewGVN::replaceInstruction(Instruction *I, Value *V) {
3722  LLVM_DEBUG(dbgs() << "Replacing " << *I << " with " << *V << "\n");
3724  // We save the actual erasing to avoid invalidating memory
3725  // dependencies until we are done with everything.
3726  markInstructionForDeletion(I);
3727 }
3728 
3729 namespace {
3730 
3731 // This is a stack that contains both the value and dfs info of where
3732 // that value is valid.
3733 class ValueDFSStack {
3734 public:
3735  Value *back() const { return ValueStack.back(); }
3736  std::pair<int, int> dfs_back() const { return DFSStack.back(); }
3737 
3738  void push_back(Value *V, int DFSIn, int DFSOut) {
3739  ValueStack.emplace_back(V);
3740  DFSStack.emplace_back(DFSIn, DFSOut);
3741  }
3742 
3743  bool empty() const { return DFSStack.empty(); }
3744 
3745  bool isInScope(int DFSIn, int DFSOut) const {
3746  if (empty())
3747  return false;
3748  return DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second;
3749  }
3750 
3751  void popUntilDFSScope(int DFSIn, int DFSOut) {
3752 
3753  // These two should always be in sync at this point.
3754  assert(ValueStack.size() == DFSStack.size() &&
3755  "Mismatch between ValueStack and DFSStack");
3756  while (
3757  !DFSStack.empty() &&
3758  !(DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second)) {
3759  DFSStack.pop_back();
3760  ValueStack.pop_back();
3761  }
3762  }
3763 
3764 private:
3765  SmallVector<Value *, 8> ValueStack;
3766  SmallVector<std::pair<int, int>, 8> DFSStack;
3767 };
3768 
3769 } // end anonymous namespace
3770 
3771 // Given an expression, get the congruence class for it.
3772 CongruenceClass *NewGVN::getClassForExpression(const Expression *E) const {
3773  if (auto *VE = dyn_cast<VariableExpression>(E))
3774  return ValueToClass.lookup(VE->getVariableValue());
3775  else if (isa<DeadExpression>(E))
3776  return TOPClass;
3777  return ExpressionToClass.lookup(E);
3778 }
3779 
3780 // Given a value and a basic block we are trying to see if it is available in,
3781 // see if the value has a leader available in that block.
3782 Value *NewGVN::findPHIOfOpsLeader(const Expression *E,
3783  const Instruction *OrigInst,
3784  const BasicBlock *BB) const {
3785  // It would already be constant if we could make it constant
3786  if (auto *CE = dyn_cast<ConstantExpression>(E))
3787  return CE->getConstantValue();
3788  if (auto *VE = dyn_cast<VariableExpression>(E)) {
3789  auto *V = VE->getVariableValue();
3790  if (alwaysAvailable(V) || DT->dominates(getBlockForValue(V), BB))
3791  return VE->getVariableValue();
3792  }
3793 
3794  auto *CC = getClassForExpression(E);
3795  if (!CC)
3796  return nullptr;
3797  if (alwaysAvailable(CC->getLeader()))
3798  return CC->getLeader();
3799 
3800  for (auto Member : *CC) {
3801  auto *MemberInst = dyn_cast<Instruction>(Member);
3802  if (MemberInst == OrigInst)
3803  continue;
3804  // Anything that isn't an instruction is always available.
3805  if (!MemberInst)
3806  return Member;
3807  if (DT->dominates(getBlockForValue(MemberInst), BB))
3808  return Member;
3809  }
3810  return nullptr;
3811 }
3812 
3813 bool NewGVN::eliminateInstructions(Function &F) {
3814  // This is a non-standard eliminator. The normal way to eliminate is
3815  // to walk the dominator tree in order, keeping track of available
3816  // values, and eliminating them. However, this is mildly
3817  // pointless. It requires doing lookups on every instruction,
3818  // regardless of whether we will ever eliminate it. For
3819  // instructions part of most singleton congruence classes, we know we
3820  // will never eliminate them.
3821 
3822  // Instead, this eliminator looks at the congruence classes directly, sorts
3823  // them into a DFS ordering of the dominator tree, and then we just
3824  // perform elimination straight on the sets by walking the congruence
3825  // class member uses in order, and eliminate the ones dominated by the
3826  // last member. This is worst case O(E log E) where E = number of
3827  // instructions in a single congruence class. In theory, this is all
3828  // instructions. In practice, it is much faster, as most instructions are
3829  // either in singleton congruence classes or can't possibly be eliminated
3830  // anyway (if there are no overlapping DFS ranges in class).
3831  // When we find something not dominated, it becomes the new leader
3832  // for elimination purposes.
3833  // TODO: If we wanted to be faster, We could remove any members with no
3834  // overlapping ranges while sorting, as we will never eliminate anything
3835  // with those members, as they don't dominate anything else in our set.
3836 
3837  bool AnythingReplaced = false;
3838 
3839  // Since we are going to walk the domtree anyway, and we can't guarantee the
3840  // DFS numbers are updated, we compute some ourselves.
3841  DT->updateDFSNumbers();
3842 
3843  // Go through all of our phi nodes, and kill the arguments associated with
3844  // unreachable edges.
3845  auto ReplaceUnreachablePHIArgs = [&](PHINode *PHI, BasicBlock *BB) {
3846  for (auto &Operand : PHI->incoming_values())
3847  if (!ReachableEdges.count({PHI->getIncomingBlock(Operand), BB})) {
3848  LLVM_DEBUG(dbgs() << "Replacing incoming value of " << PHI
3849  << " for block "
3850  << getBlockName(PHI->getIncomingBlock(Operand))
3851  << " with poison due to it being unreachable\n");
3852  Operand.set(PoisonValue::get(PHI->getType()));
3853  }
3854  };
3855  // Replace unreachable phi arguments.
3856  // At this point, RevisitOnReachabilityChange only contains:
3857  //
3858  // 1. PHIs
3859  // 2. Temporaries that will convert to PHIs
3860  // 3. Operations that are affected by an unreachable edge but do not fit into
3861  // 1 or 2 (rare).
3862  // So it is a slight overshoot of what we want. We could make it exact by
3863  // using two SparseBitVectors per block.
3864  DenseMap<const BasicBlock *, unsigned> ReachablePredCount;
3865  for (auto &KV : ReachableEdges)
3866  ReachablePredCount[KV.getEnd()]++;
3867  for (auto &BBPair : RevisitOnReachabilityChange) {
3868  for (auto InstNum : BBPair.second) {
3869  auto *Inst = InstrFromDFSNum(InstNum);
3870  auto *PHI = dyn_cast<PHINode>(Inst);
3871  PHI = PHI ? PHI : dyn_cast_or_null<PHINode>(RealToTemp.lookup(Inst));
3872  if (!PHI)
3873  continue;
3874  auto *BB = BBPair.first;
3875  if (ReachablePredCount.lookup(BB) != PHI->getNumIncomingValues())
3876  ReplaceUnreachablePHIArgs(PHI, BB);
3877  }
3878  }
3879 
3880  // Map to store the use counts
3882  for (auto *CC : reverse(CongruenceClasses)) {
3883  LLVM_DEBUG(dbgs() << "Eliminating in congruence class " << CC->getID()
3884  << "\n");
3885  // Track the equivalent store info so we can decide whether to try
3886  // dead store elimination.
3887  SmallVector<ValueDFS, 8> PossibleDeadStores;
3888  SmallPtrSet<Instruction *, 8> ProbablyDead;
3889  if (CC->isDead() || CC->empty())
3890  continue;
3891  // Everything still in the TOP class is unreachable or dead.
3892  if (CC == TOPClass) {
3893  for (auto M : *CC) {
3894  auto *VTE = ValueToExpression.lookup(M);
3895  if (VTE && isa<DeadExpression>(VTE))
3896  markInstructionForDeletion(cast<Instruction>(M));
3897  assert((!ReachableBlocks.count(cast<Instruction>(M)->getParent()) ||
3898  InstructionsToErase.count(cast<Instruction>(M))) &&
3899  "Everything in TOP should be unreachable or dead at this "
3900  "point");
3901  }
3902  continue;
3903  }
3904 
3905  assert(CC->getLeader() && "We should have had a leader");
3906  // If this is a leader that is always available, and it's a
3907  // constant or has no equivalences, just replace everything with
3908  // it. We then update the congruence class with whatever members
3909  // are left.
3910  Value *Leader =
3911  CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
3912  if (alwaysAvailable(Leader)) {
3913  CongruenceClass::MemberSet MembersLeft;
3914  for (auto M : *CC) {
3915  Value *Member = M;
3916  // Void things have no uses we can replace.
3917  if (Member == Leader || !isa<Instruction>(Member) ||
3918  Member->getType()->isVoidTy()) {
3919  MembersLeft.insert(Member);
3920  continue;
3921  }
3922  LLVM_DEBUG(dbgs() << "Found replacement " << *(Leader) << " for "
3923  << *Member << "\n");
3924  auto *I = cast<Instruction>(Member);
3925  assert(Leader != I && "About to accidentally remove our leader");
3926  replaceInstruction(I, Leader);
3927  AnythingReplaced = true;
3928  }
3929  CC->swap(MembersLeft);
3930  } else {
3931  // If this is a singleton, we can skip it.
3932  if (CC->size() != 1 || RealToTemp.count(Leader)) {
3933  // This is a stack because equality replacement/etc may place
3934  // constants in the middle of the member list, and we want to use
3935  // those constant values in preference to the current leader, over
3936  // the scope of those constants.
3937  ValueDFSStack EliminationStack;
3938 
3939  // Convert the members to DFS ordered sets and then merge them.
3940  SmallVector<ValueDFS, 8> DFSOrderedSet;
3941  convertClassToDFSOrdered(*CC, DFSOrderedSet, UseCounts, ProbablyDead);
3942 
3943  // Sort the whole thing.
3944  llvm::sort(DFSOrderedSet);
3945  for (auto &VD : DFSOrderedSet) {
3946  int MemberDFSIn = VD.DFSIn;
3947  int MemberDFSOut = VD.DFSOut;
3948  Value *Def = VD.Def.getPointer();
3949  bool FromStore = VD.Def.getInt();
3950  Use *U = VD.U;
3951  // We ignore void things because we can't get a value from them.
3952  if (Def && Def->getType()->isVoidTy())
3953  continue;
3954  auto *DefInst = dyn_cast_or_null<Instruction>(Def);
3955  if (DefInst && AllTempInstructions.count(DefInst)) {
3956  auto *PN = cast<PHINode>(DefInst);
3957 
3958  // If this is a value phi and that's the expression we used, insert
3959  // it into the program
3960  // remove from temp instruction list.
3961  AllTempInstructions.erase(PN);
3962  auto *DefBlock = getBlockForValue(Def);
3963  LLVM_DEBUG(dbgs() << "Inserting fully real phi of ops" << *Def
3964  << " into block "
3965  << getBlockName(getBlockForValue(Def)) << "\n");
3966  PN->insertBefore(&DefBlock->front());
3967  Def = PN;
3968  NumGVNPHIOfOpsEliminations++;
3969  }
3970 
3971  if (EliminationStack.empty()) {
3972  LLVM_DEBUG(dbgs() << "Elimination Stack is empty\n");
3973  } else {
3974  LLVM_DEBUG(dbgs() << "Elimination Stack Top DFS numbers are ("
3975  << EliminationStack.dfs_back().first << ","
3976  << EliminationStack.dfs_back().second << ")\n");
3977  }
3978 
3979  LLVM_DEBUG(dbgs() << "Current DFS numbers are (" << MemberDFSIn << ","
3980  << MemberDFSOut << ")\n");
3981  // First, we see if we are out of scope or empty. If so,
3982  // and there equivalences, we try to replace the top of
3983  // stack with equivalences (if it's on the stack, it must
3984  // not have been eliminated yet).
3985  // Then we synchronize to our current scope, by
3986  // popping until we are back within a DFS scope that
3987  // dominates the current member.
3988  // Then, what happens depends on a few factors
3989  // If the stack is now empty, we need to push
3990  // If we have a constant or a local equivalence we want to
3991  // start using, we also push.
3992  // Otherwise, we walk along, processing members who are
3993  // dominated by this scope, and eliminate them.
3994  bool ShouldPush = Def && EliminationStack.empty();
3995  bool OutOfScope =
3996  !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut);
3997 
3998  if (OutOfScope || ShouldPush) {
3999  // Sync to our current scope.
4000  EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4001  bool ShouldPush = Def && EliminationStack.empty();
4002  if (ShouldPush) {
4003  EliminationStack.push_back(Def, MemberDFSIn, MemberDFSOut);
4004  }
4005  }
4006 
4007  // Skip the Def's, we only want to eliminate on their uses. But mark
4008  // dominated defs as dead.
4009  if (Def) {
4010  // For anything in this case, what and how we value number
4011  // guarantees that any side-effets that would have occurred (ie
4012  // throwing, etc) can be proven to either still occur (because it's
4013  // dominated by something that has the same side-effects), or never
4014  // occur. Otherwise, we would not have been able to prove it value
4015  // equivalent to something else. For these things, we can just mark
4016  // it all dead. Note that this is different from the "ProbablyDead"
4017  // set, which may not be dominated by anything, and thus, are only
4018  // easy to prove dead if they are also side-effect free. Note that
4019  // because stores are put in terms of the stored value, we skip
4020  // stored values here. If the stored value is really dead, it will
4021  // still be marked for deletion when we process it in its own class.
4022  if (!EliminationStack.empty() && Def != EliminationStack.back() &&
4023  isa<Instruction>(Def) && !FromStore)
4024  markInstructionForDeletion(cast<Instruction>(Def));
4025  continue;
4026  }
4027  // At this point, we know it is a Use we are trying to possibly
4028  // replace.
4029 
4030  assert(isa<Instruction>(U->get()) &&
4031  "Current def should have been an instruction");
4032  assert(isa<Instruction>(U->getUser()) &&
4033  "Current user should have been an instruction");
4034 
4035  // If the thing we are replacing into is already marked to be dead,
4036  // this use is dead. Note that this is true regardless of whether
4037  // we have anything dominating the use or not. We do this here
4038  // because we are already walking all the uses anyway.
4039  Instruction *InstUse = cast<Instruction>(U->getUser());
4040  if (InstructionsToErase.count(InstUse)) {
4041  auto &UseCount = UseCounts[U->get()];
4042  if (--UseCount == 0) {
4043  ProbablyDead.insert(cast<Instruction>(U->get()));
4044  }
4045  }
4046 
4047  // If we get to this point, and the stack is empty we must have a use
4048  // with nothing we can use to eliminate this use, so just skip it.
4049  if (EliminationStack.empty())
4050  continue;
4051 
4052  Value *DominatingLeader = EliminationStack.back();
4053 
4054  auto *II = dyn_cast<IntrinsicInst>(DominatingLeader);
4055  bool isSSACopy = II && II->getIntrinsicID() == Intrinsic::ssa_copy;
4056  if (isSSACopy)
4057  DominatingLeader = II->getOperand(0);
4058 
4059  // Don't replace our existing users with ourselves.
4060  if (U->get() == DominatingLeader)
4061  continue;
4062  LLVM_DEBUG(dbgs()
4063  << "Found replacement " << *DominatingLeader << " for "
4064  << *U->get() << " in " << *(U->getUser()) << "\n");
4065 
4066  // If we replaced something in an instruction, handle the patching of
4067  // metadata. Skip this if we are replacing predicateinfo with its
4068  // original operand, as we already know we can just drop it.
4069  auto *ReplacedInst = cast<Instruction>(U->get());
4070  auto *PI = PredInfo->getPredicateInfoFor(ReplacedInst);
4071  if (!PI || DominatingLeader != PI->OriginalOp)
4072  patchReplacementInstruction(ReplacedInst, DominatingLeader);
4073  U->set(DominatingLeader);
4074  // This is now a use of the dominating leader, which means if the
4075  // dominating leader was dead, it's now live!
4076  auto &LeaderUseCount = UseCounts[DominatingLeader];
4077  // It's about to be alive again.
4078  if (LeaderUseCount == 0 && isa<Instruction>(DominatingLeader))
4079  ProbablyDead.erase(cast<Instruction>(DominatingLeader));
4080  // For copy instructions, we use their operand as a leader,
4081  // which means we remove a user of the copy and it may become dead.
4082  if (isSSACopy) {
4083  unsigned &IIUseCount = UseCounts[II];
4084  if (--IIUseCount == 0)
4085  ProbablyDead.insert(II);
4086  }
4087  ++LeaderUseCount;
4088  AnythingReplaced = true;
4089  }
4090  }
4091  }
4092 
4093  // At this point, anything still in the ProbablyDead set is actually dead if
4094  // would be trivially dead.
4095  for (auto *I : ProbablyDead)
4097  markInstructionForDeletion(I);
4098 
4099  // Cleanup the congruence class.
4100  CongruenceClass::MemberSet MembersLeft;
4101  for (auto *Member : *CC)
4102  if (!isa<Instruction>(Member) ||
4103  !InstructionsToErase.count(cast<Instruction>(Member)))
4104  MembersLeft.insert(Member);
4105  CC->swap(MembersLeft);
4106 
4107  // If we have possible dead stores to look at, try to eliminate them.
4108  if (CC->getStoreCount() > 0) {
4109  convertClassToLoadsAndStores(*CC, PossibleDeadStores);
4110  llvm::sort(PossibleDeadStores);
4111  ValueDFSStack EliminationStack;
4112  for (auto &VD : PossibleDeadStores) {
4113  int MemberDFSIn = VD.DFSIn;
4114  int MemberDFSOut = VD.DFSOut;
4115  Instruction *Member = cast<Instruction>(VD.Def.getPointer());
4116  if (EliminationStack.empty() ||
4117  !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut)) {
4118  // Sync to our current scope.
4119  EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
4120  if (EliminationStack.empty()) {
4121  EliminationStack.push_back(Member, MemberDFSIn, MemberDFSOut);
4122  continue;
4123  }
4124  }
4125  // We already did load elimination, so nothing to do here.
4126  if (isa<LoadInst>(Member))
4127  continue;
4128  assert(!EliminationStack.empty());
4129  Instruction *Leader = cast<Instruction>(EliminationStack.back());
4130  (void)Leader;
4131  assert(DT->dominates(Leader->getParent(), Member->getParent()));
4132  // Member is dominater by Leader, and thus dead
4133  LLVM_DEBUG(dbgs() << "Marking dead store " << *Member
4134  << " that is dominated by " << *Leader << "\n");
4135  markInstructionForDeletion(Member);
4136  CC->erase(Member);
4137  ++NumGVNDeadStores;
4138  }
4139  }
4140  }
4141  return AnythingReplaced;
4142 }
4143 
4144 // This function provides global ranking of operations so that we can place them
4145 // in a canonical order. Note that rank alone is not necessarily enough for a
4146 // complete ordering, as constants all have the same rank. However, generally,
4147 // we will simplify an operation with all constants so that it doesn't matter
4148 // what order they appear in.
4149 unsigned int NewGVN::getRank(const Value *V) const {
4150  // Prefer constants to undef to anything else
4151  // Undef is a constant, have to check it first.
4152  // Prefer poison to undef as it's less defined.
4153  // Prefer smaller constants to constantexprs
4154  // Note that the order here matters because of class inheritance
4155  if (isa<ConstantExpr>(V))
4156  return 3;
4157  if (isa<PoisonValue>(V))
4158  return 1;
4159  if (isa<UndefValue>(V))
4160  return 2;
4161  if (isa<Constant>(V))
4162  return 0;
4163  if (auto *A = dyn_cast<Argument>(V))
4164  return 4 + A->getArgNo();
4165 
4166  // Need to shift the instruction DFS by number of arguments + 5 to account for
4167  // the constant and argument ranking above.
4168  unsigned Result = InstrToDFSNum(V);
4169  if (Result > 0)
4170  return 5 + NumFuncArgs + Result;
4171  // Unreachable or something else, just return a really large number.
4172  return ~0;
4173 }
4174 
4175 // This is a function that says whether two commutative operations should
4176 // have their order swapped when canonicalizing.
4177 bool NewGVN::shouldSwapOperands(const Value *A, const Value *B) const {
4178  // Because we only care about a total ordering, and don't rewrite expressions
4179  // in this order, we order by rank, which will give a strict weak ordering to
4180  // everything but constants, and then we order by pointer address.
4181  return std::make_pair(getRank(A), A) > std::make_pair(getRank(B), B);
4182 }
4183 
4184 bool NewGVN::shouldSwapOperandsForIntrinsic(const Value *A, const Value *B,
4185  const IntrinsicInst *I) const {
4186  auto LookupResult = IntrinsicInstPred.find(I);
4187  if (shouldSwapOperands(A, B)) {
4188  if (LookupResult == IntrinsicInstPred.end())
4189  IntrinsicInstPred.insert({I, B});
4190  else
4191  LookupResult->second = B;
4192  return true;
4193  }
4194 
4195  if (LookupResult != IntrinsicInstPred.end()) {
4196  auto *SeenPredicate = LookupResult->second;
4197  if (SeenPredicate) {
4198  if (SeenPredicate == B)
4199  return true;
4200  else
4201  LookupResult->second = nullptr;
4202  }
4203  }
4204  return false;
4205 }
4206 
4207 namespace {
4208 
4209 class NewGVNLegacyPass : public FunctionPass {
4210 public:
4211  // Pass identification, replacement for typeid.
4212  static char ID;
4213 
4214  NewGVNLegacyPass() : FunctionPass(ID) {
4216  }
4217 
4218  bool runOnFunction(Function &F) override;
4219 
4220 private:
4221  void getAnalysisUsage(AnalysisUsage &AU) const override {
4229  }
4230 };
4231 
4232 } // end anonymous namespace
4233 
4235  if (skipFunction(F))
4236  return false;
4237  return NewGVN(F, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4238  &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
4239  &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F),
4240  &getAnalysis<AAResultsWrapperPass>().getAAResults(),
4241  &getAnalysis<MemorySSAWrapperPass>().getMSSA(),
4242  F.getParent()->getDataLayout())
4243  .runGVN();
4244 }
4245 
4246 char NewGVNLegacyPass::ID = 0;
4247 
4248 INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering",
4249  false, false)
4256 INITIALIZE_PASS_END(NewGVNLegacyPass, "newgvn", "Global Value Numbering", false,
4257  false)
4258 
4259 // createGVNPass - The public interface to this file.
4260 FunctionPass *llvm::createNewGVNPass() { return new NewGVNLegacyPass(); }
4261 
4263  // Apparently the order in which we get these results matter for
4264  // the old GVN (see Chandler's comment in GVN.cpp). I'll keep
4265  // the same order here, just in case.
4266  auto &AC = AM.getResult<AssumptionAnalysis>(F);
4267  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
4268  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
4269  auto &AA = AM.getResult<AAManager>(F);
4270  auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
4271  bool Changed =
4272  NewGVN(F, &DT, &AC, &TLI, &AA, &MSSA, F.getParent()->getDataLayout())
4273  .runGVN();
4274  if (!Changed)
4275  return PreservedAnalyses::all();
4276  PreservedAnalyses PA;
4278  return PA;
4279 }
i
i
Definition: README.txt:29
llvm::PreservedAnalyses
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:152
llvm::VNCoercion::analyzeLoadFromClobberingMemInst
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:356
set
We currently generate a but we really shouldn eax ecx xorl edx divl ecx eax divl ecx movl eax ret A similar code sequence works for division We currently compile i32 v2 eax eax jo LBB1_2 atomic and others It is also currently not done for read modify write instructions It is also current not done if the OF or CF flags are needed The shift operators have the complication that when the shift count is EFLAGS is not set
Definition: README.txt:1277
AssumptionCache.h
patchAndReplaceAllUsesWith
static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl)
Definition: NewGVN.cpp:3683
llvm::AAManager
A manager for alias analyses.
Definition: AliasAnalysis.h:1299
llvm::ValueDFS::LocalNum
unsigned int LocalNum
Definition: PredicateInfo.cpp:96
llvm::CmpInst::getSwappedPredicate
Predicate getSwappedPredicate() const
For example, EQ->EQ, SLE->SGE, ULT->UGT, OEQ->OEQ, ULE->UGE, OLT->OGT, etc.
Definition: InstrTypes.h:849
llvm
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:17
llvm::sys::path::const_iterator::end
friend const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:235
llvm::tgtok::Def
@ Def
Definition: TGLexer.h:50
NewGVN::ValueDFS
Definition: NewGVN.cpp:3507
M
We currently emits eax Perhaps this is what we really should generate is Is imull three or four cycles eax eax The current instruction priority is based on pattern complexity The former is more complex because it folds a load so the latter will not be emitted Perhaps we should use AddedComplexity to give LEA32r a higher priority We should always try to match LEA first since the LEA matching code does some estimate to determine whether the match is profitable if we care more about code then imull is better It s two bytes shorter than movl leal On a Pentium M
Definition: README.txt:252
llvm::CmpInst::ICMP_EQ
@ ICMP_EQ
equal
Definition: InstrTypes.h:740
llvm::make_range
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
Definition: iterator_range.h:53
llvm::set_is_subset
bool set_is_subset(const S1Ty &S1, const S2Ty &S2)
set_is_subset(A, B) - Return true iff A in B
Definition: SetOperations.h:72
llvm::SmallPtrSetImpl::erase
bool erase(PtrType Ptr)
erase - If the set contains the specified pointer, remove it and return true, otherwise return false.
Definition: SmallPtrSet.h:379
getBlockName
static std::string getBlockName(const BasicBlock *B)
Definition: NewGVN.cpp:935
llvm::PHINode::incoming_values
op_range incoming_values()
Definition: Instructions.h:2750
llvm::DataLayout
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:113
llvm::BumpPtrAllocatorImpl::Reset
void Reset()
Deallocate all but the current slab and reset the current pointer to the beginning of it,...
Definition: Allocator.h:121
PredicateInfo.h
llvm::CmpInst::Predicate
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:719
llvm::User::operands
op_range operands()
Definition: User.h:242
IntrinsicInst.h
llvm::SimplifyQuery
Definition: InstructionSimplify.h:93
llvm::AnalysisManager::getResult
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:780
llvm::pdb::PDB_DataKind::Member
@ Member
Scalar.h
SetOperations.h
llvm::Function
Definition: Function.h:60
llvm::optimized_def_chain
iterator_range< def_chain_iterator< T, true > > optimized_def_chain(T MA)
Definition: MemorySSA.h:1371
llvm::DenseMapBase::lookup
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:199
P
This currently compiles esp xmm0 movsd esp eax eax esp ret We should use not the dag combiner This is because dagcombine2 needs to be able to see through the X86ISD::Wrapper which DAGCombine can t really do The code for turning x load into a single vector load is target independent and should be moved to the dag combiner The code for turning x load into a vector load can only handle a direct load from a global or a direct load from the stack It should be generalized to handle any load from P
Definition: README-SSE.txt:411
Pass.h
llvm::BitVector::set
BitVector & set()
Definition: BitVector.h:344
llvm::LocalNum
LocalNum
Definition: PredicateInfo.cpp:81
llvm::BitVector::clear
void clear()
clear - Removes all bits from the bitvector.
Definition: BitVector.h:328
llvm::DenseMapInfo< const Expression * >::getTombstoneKey
static const Expression * getTombstoneKey()
Definition: NewGVN.cpp:451
llvm::SimplifySelectInst
Value * SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, const SimplifyQuery &Q)
Given operands for a SelectInst, fold the result or return null.
Definition: InstructionSimplify.cpp:4517
llvm::SmallVector
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1185
isCopyOfAPHI
static bool isCopyOfAPHI(const Value *V)
Definition: NewGVN.cpp:990
Statistic.h
llvm::detail::DenseSetImpl::find
iterator find(const_arg_type_t< ValueT > V)
Definition: DenseSet.h:179
ErrorHandling.h
Allocator.h
llvm::BitVector::set_bits
iterator_range< const_set_bits_iterator > set_bits() const
Definition: BitVector.h:133
llvm::CmpInst::getInversePredicate
Predicate getInversePredicate() const
For example, EQ -> NE, UGT -> ULE, SLT -> SGE, OEQ -> UNE, UGT -> OLE, OLT -> UGE,...
Definition: InstrTypes.h:833
ValueTracking.h
Local.h
llvm::getInitialValueOfAllocation
Constant * getInitialValueOfAllocation(const CallBase *Alloc, const TargetLibraryInfo *TLI, Type *Ty)
If this allocation function initializes memory to a fixed value, return said value in the requested t...
Definition: MemoryBuiltins.cpp:423
llvm::ValueDFS::DFSIn
int DFSIn
Definition: PredicateInfo.cpp:94
llvm::DominatorTree
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree.
Definition: Dominators.h:166
alwaysAvailable
static bool alwaysAvailable(Value *V)
Definition: NewGVN.cpp:1008
GlobalsModRef.h
llvm::cl::Hidden
@ Hidden
Definition: CommandLine.h:139
llvm::PatternMatch::m_Br
brc_match< Cond_t, bind_ty< BasicBlock >, bind_ty< BasicBlock > > m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F)
Definition: PatternMatch.h:1741
llvm::DenseMapBase::erase
bool erase(const KeyT &Val)
Definition: DenseMap.h:304
llvm::Type
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
NL
#define NL
Definition: DetailedRecordsBackend.cpp:31
DenseMap.h
MemoryBuiltins.h
llvm::BitVector::resize
void resize(unsigned N, bool t=false)
resize - Grow or shrink the bitvector.
Definition: BitVector.h:334
llvm::reverse
auto reverse(ContainerTy &&C, std::enable_if_t< has_rbegin< ContainerTy >::value > *=nullptr)
Definition: STLExtras.h:380
llvm::sys::path::end
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:235
llvm::MemorySSA::getLiveOnEntryDef
MemoryAccess * getLiveOnEntryDef() const
Definition: MemorySSA.h:755
llvm::sys::path::begin
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:226
llvm::BasicBlockEdge
Definition: Dominators.h:98
llvm::copy
OutputIt copy(R &&Range, OutputIt Out)
Definition: STLExtras.h:1656
llvm::DominatorTreeBase::getNode
DomTreeNodeBase< NodeT > * getNode(const NodeT *BB) const
getNode - return the (Post)DominatorTree node for the specified basic block.
Definition: GenericDomTree.h:351
llvm::SPII::Load
@ Load
Definition: SparcInstrInfo.h:32
llvm::Optional
Definition: APInt.h:33
llvm::DenseMapBase::count
size_type count(const_arg_type_t< KeyT > Val) const
Return 1 if the specified key is in the map, 0 otherwise.
Definition: DenseMap.h:147
T
#define T
Definition: Mips16ISelLowering.cpp:341
llvm::SimplifyCmpInst
Value * SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, const SimplifyQuery &Q)
Given operands for a CmpInst, fold the result or return null.
Definition: InstructionSimplify.cpp:5560
Offset
uint64_t Offset
Definition: ELFObjHandler.cpp:79
llvm::SmallPtrSet< const Value *, 8 >
llvm::SimplifyGEPInst
Value * SimplifyGEPInst(Type *SrcTy, Value *Ptr, ArrayRef< Value * > Indices, bool InBounds, const SimplifyQuery &Q)
Given operands for a GetElementPtrInst, fold the result or return null.
Definition: InstructionSimplify.cpp:4662
llvm::GVNExpression::StoreExpression::~StoreExpression
~StoreExpression() override
Hashing.h
llvm::MemoryPhi
Represents phi nodes for memory accesses.
Definition: MemorySSA.h:493
getCopyOf
static Value * getCopyOf(const Value *V)
Definition: NewGVN.cpp:978
equalsLoadStoreHelper
static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS)
Definition: NewGVN.cpp:907
llvm::SmallVectorImpl::pop_back_val
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:654
RHS
Value * RHS
Definition: X86PartialReduction.cpp:76
llvm::LoadInst::getPointerOperand
Value * getPointerOperand()
Definition: Instructions.h:268
llvm::detail::DenseSetImpl< ValueT, DenseMap< ValueT, detail::DenseSetEmpty, DenseMapInfo< ValueT >, detail::DenseSetPair< ValueT > >, DenseMapInfo< ValueT > >::insert
std::pair< iterator, bool > insert(const ValueT &V)
Definition: DenseSet.h:206
llvm::detail::DenseSetImpl< ValueT, DenseMap< ValueT, detail::DenseSetEmpty, DenseMapInfo< ValueT >, detail::DenseSetPair< ValueT > >, DenseMapInfo< ValueT > >::count
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:97
llvm::Type::getInt8Ty
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:237
ConstantFolding.h
Use.h
llvm::DebugCounter::getCounterValue
static int64_t getCounterValue(unsigned ID)
Definition: DebugCounter.h:107
llvm::GVNExpression::CallExpression::~CallExpression
~CallExpression() override
llvm::MemoryPhi::getIncomingBlock
BasicBlock * getIncomingBlock(unsigned I) const
Return incoming basic block number i.
Definition: MemorySSA.h:554
LLVM_DEBUG
#define LLVM_DEBUG(X)
Definition: Debug.h:101
DepthFirstIterator.h
llvm::map_iterator
mapped_iterator< ItTy, FuncTy > map_iterator(ItTy I, FuncTy F)
Definition: STLExtras.h:322
llvm::DebugCounter::isCounterSet
static bool isCounterSet(unsigned ID)
Definition: DebugCounter.h:102
F
#define F(x, y, z)
Definition: MD5.cpp:55
llvm::RISCVFenceField::R
@ R
Definition: RISCVBaseInfo.h:240
llvm::BasicBlock
LLVM Basic Block Representation.
Definition: BasicBlock.h:55
AliasAnalysis.h
result
It looks like we only need to define PPCfmarto for these because according to these instructions perform RTO on fma s result
Definition: README_P9.txt:256
PointerIntPair.h
llvm::dbgs
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
GraphTraits.h
llvm::GVNExpression::UnknownExpression
Definition: GVNExpression.h:625
llvm::DomTreeNodeBase::getDFSNumIn
unsigned getDFSNumIn() const
getDFSNumIn/getDFSNumOut - These return the DFS visitation order for nodes in the dominator tree.
Definition: GenericDomTree.h:143
llvm::SimplifyCastInst
Value * SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty, const SimplifyQuery &Q)
Given operands for a CastInst, fold the result or return null.
Definition: InstructionSimplify.cpp:4890
llvm::isAllocationFn
bool isAllocationFn(const Value *V, const TargetLibraryInfo *TLI)
Tests if a value is a call or invoke to a library function that allocates or reallocates memory (eith...
Definition: MemoryBuiltins.cpp:272
SparseBitVector.h
llvm::DominatorTree::dominates
bool dominates(const BasicBlock *BB, const Use &U) const
Return true if the (end of the) basic block BB dominates the use U.
Definition: Dominators.cpp:122
llvm::DomTreeNodeBase::getDFSNumOut
unsigned getDFSNumOut() const
Definition: GenericDomTree.h:144
Arg
amdgpu Simplify well known AMD library false FunctionCallee Value * Arg
Definition: AMDGPULibCalls.cpp:186
llvm::detail::DenseSetImpl< ValueT, DenseMap< ValueT, detail::DenseSetEmpty, DenseMapInfo< ValueT >, detail::DenseSetPair< ValueT > >, DenseMapInfo< ValueT > >::end
iterator end()
Definition: DenseSet.h:174
llvm::ValueDFS::U
Use * U
Definition: PredicateInfo.cpp:99
Instruction.h
CommandLine.h
llvm::GVNExpression::int_op_inserter
Definition: GVNExpression.h:480
llvm::GVNExpression::ConstantExpression
Definition: GVNExpression.h:588
LHS
Value * LHS
Definition: X86PartialReduction.cpp:75
llvm::ConstantInt
This is the shared class of boolean and integer constants.
Definition: Constants.h:79
llvm::Instruction::getNumSuccessors
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Definition: Instruction.cpp:777
llvm::DenseMapInfo
An information struct used to provide DenseMap with the various necessary components for a given valu...
Definition: APInt.h:34
llvm::all_of
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:1605
llvm::orc::tpctypes::LookupResult
std::vector< JITTargetAddress > LookupResult
Definition: TargetProcessControlTypes.h:119
llvm::MemorySSAWrapperPass
Legacy analysis pass which computes MemorySSA.
Definition: MemorySSA.h:998
llvm::DominatorTreeBase::getRootNode
DomTreeNodeBase< NodeT > * getRootNode()
getRootNode - This returns the entry node for the CFG of the function.
Definition: GenericDomTree.h:370
llvm::PassRegistry::getPassRegistry
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
Definition: PassRegistry.cpp:31
llvm::DenseMapInfo< const Expression * >::isEqual
static bool isEqual(const Expression *LHS, const Expression *RHS)
Definition: NewGVN.cpp:471
NewGVN::ValueDFS::operator<
bool operator<(const ValueDFS &Other) const
Definition: NewGVN.cpp:3518
llvm::MutableArrayRef
MutableArrayRef - Represent a mutable reference to an array (0 or more elements consecutively in memo...
Definition: ArrayRef.h:306
llvm::PredicateConstraint::OtherOp
Value * OtherOp
Definition: PredicateInfo.h:77
llvm::GVNExpression::PHIExpression
Definition: GVNExpression.h:505
Constants.h
llvm::PatternMatch::match
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
llvm::ExactEqualsExpression::ExactEqualsExpression
ExactEqualsExpression(const Expression &E)
Definition: NewGVN.cpp:435
isZero
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:518
llvm::AAResults
Definition: AliasAnalysis.h:511
llvm::getStoredValue
SDValue getStoredValue(SDValue Op)
Definition: VECustomDAG.cpp:332
llvm::MemorySSA::isLiveOnEntryDef
bool isLiveOnEntryDef(const MemoryAccess *MA) const
Return true if MA represents the live on entry value.
Definition: MemorySSA.h:751
P2
This might compile to this xmm1 xorps xmm0 movss xmm0 ret Now consider if the code caused xmm1 to get spilled This might produce this xmm1 movaps xmm0 movaps xmm1 movss xmm0 ret since the reload is only used by these we could fold it into the producing something like xmm1 movaps xmm0 ret saving two instructions The basic idea is that a reload from a spill if only one byte chunk is bring in zeros the one element instead of elements This can be used to simplify a variety of shuffle where the elements are fixed zeros This code generates ugly probably due to costs being off or< 4 x float > * P2
Definition: README-SSE.txt:278
E
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
llvm::User
Definition: User.h:44
llvm::initializeNewGVNLegacyPassPass
void initializeNewGVNLegacyPassPass(PassRegistry &)
C
(vector float) vec_cmpeq(*A, *B) C
Definition: README_ALTIVEC.txt:86
llvm::ARM_PROC::A
@ A
Definition: ARMBaseInfo.h:34
InstrTypes.h
EnableStoreRefinement
static cl::opt< bool > EnableStoreRefinement("enable-store-refinement", cl::init(false), cl::Hidden)
llvm::JumpTable::Simplified
@ Simplified
Definition: TargetOptions.h:47
llvm::createNewGVNPass
FunctionPass * createNewGVNPass()
Definition: NewGVN.cpp:4260
llvm::AnalysisUsage
Represent the analysis usage information of a pass.
Definition: PassAnalysisSupport.h:47
llvm::ms_demangle::QualifierMangleMode::Result
@ Result
NewGVN::ValueDFS::Def
PointerIntPair< Value *, 1, bool > Def
Definition: NewGVN.cpp:3515
llvm::AMDGPU::PALMD::Key
Key
PAL metadata keys.
Definition: AMDGPUMetadata.h:486
TargetLibraryInfo.h
llvm::GVNExpression::LoadExpression::equals
bool equals(const Expression &Other) const override
Definition: NewGVN.cpp:913
AssumeBundleBuilder.h
DenseSet.h
false
Definition: StackSlotColoring.cpp:141
B
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
First
into llvm powi allowing the code generator to produce balanced multiplication trees First
Definition: README.txt:54
llvm::Value::getValueID
unsigned getValueID() const
Return an ID for the concrete type of this object.
Definition: Value.h:532
llvm::VNCoercion
Definition: VNCoercion.h:34
llvm::GVNExpression::Expression
Definition: GVNExpression.h:60
llvm::Instruction
Definition: Instruction.h:42
llvm::SPII::Store
@ Store
Definition: SparcInstrInfo.h:33
llvm::DebugCounter::shouldExecute
static bool shouldExecute(unsigned CounterName)
Definition: DebugCounter.h:74
llvm::DominatorTreeWrapperPass
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:302
GVNExpression.h
llvm::STATISTIC
STATISTIC(NumFunctions, "Total number of functions")
isCopyOfPHI
static bool isCopyOfPHI(const Value *V, const PHINode *PN)
Definition: NewGVN.cpp:986
PointerLikeTypeTraits.h
EnablePhiOfOps
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.
llvm::AArch64CC::LE
@ LE
Definition: AArch64BaseInfo.h:268
llvm::UndefValue::get
static UndefValue * get(Type *T)
Static factory methods - Return an 'undef' object of the specified type.
Definition: Constants.cpp:1769
BitVector.h
llvm::GVNExpression::BasicExpression::~BasicExpression
~BasicExpression() override
llvm::CmpInst::FCMP_OEQ
@ FCMP_OEQ
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:722
SmallPtrSet.h
llvm::Instruction::getSuccessor
BasicBlock * getSuccessor(unsigned Idx) const
Return the specified successor. This instruction must be a terminator.
Definition: Instruction.cpp:789
llvm::BitVector
Definition: BitVector.h:75
llvm::DominatorTreeBase::updateDFSNumbers
void updateDFSNumbers() const
updateDFSNumbers - Assign In and Out numbers to the nodes while walking dominator tree in dfs order.
Definition: GenericDomTree.h:732
PatternMatch.h
llvm::ExactEqualsExpression::getComputedHash
hash_code getComputedHash() const
Definition: NewGVN.cpp:437
llvm::MCID::Call
@ Call
Definition: MCInstrDesc.h:155
place
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM ID Predecessors according to mbb< bb27, 0x8b0a7c0 > Note ADDri is not a two address instruction its result reg1037 is an operand of the PHI node in bb76 and its operand reg1039 is the result of the PHI node We should treat it as a two address code and make sure the ADDri is scheduled after any node that reads reg1039 Use info(i.e. register scavenger) to assign it a free register to allow reuse the collector could move the objects and invalidate the derived pointer This is bad enough in the first place
Definition: README.txt:50
llvm::PHINode::getNumIncomingValues
unsigned getNumIncomingValues() const
Return the number of incoming edges.
Definition: Instructions.h:2756
llvm::GVNExpression::PHIExpression::~PHIExpression
~PHIExpression() override
llvm::MemoryAccess::getBlock
BasicBlock * getBlock() const
Definition: MemorySSA.h:164
llvm::Value::use_empty
bool use_empty() const
Definition: Value.h:344
llvm::CallingConv::ID
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
Definition: CallingConv.h:24
Type.h
X
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
INITIALIZE_PASS_BEGIN
INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering", false, false) INITIALIZE_PASS_END(NewGVNLegacyPass
llvm::PredicateInfo
Encapsulates PredicateInfo, including all data associated with memory accesses.
Definition: PredicateInfo.h:178
INITIALIZE_PASS_END
#define INITIALIZE_PASS_END(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:58
llvm::DenseMapInfo< const Expression * >::isEqual
static bool isEqual(const ExactEqualsExpression &LHS, const Expression *RHS)
Definition: NewGVN.cpp:465
llvm::Instruction::isCommutative
bool isCommutative() const LLVM_READONLY
Return true if the instruction is commutative:
Definition: Instruction.cpp:770
llvm::PredicateConstraint::Predicate
CmpInst::Predicate Predicate
Definition: PredicateInfo.h:76
llvm::DOTGraphTraits
DOTGraphTraits - Template class that can be specialized to customize how graphs are converted to 'dot...
Definition: DOTGraphTraits.h:166
llvm::SPIRV::Decoration::Component
@ Component
llvm::ExactEqualsExpression::operator==
bool operator==(const Expression &Other) const
Definition: NewGVN.cpp:439
llvm::DenseSet
Implements a dense probed hash-table based set.
Definition: DenseSet.h:268
llvm::MemorySSAWalker::getClobberingMemoryAccess
MemoryAccess * getClobberingMemoryAccess(const Instruction *I)
Given a memory Mod/Ref/ModRef'ing instruction, calling this will give you the nearest dominating Memo...
Definition: MemorySSA.h:1058
llvm::wouldInstructionBeTriviallyDead
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:414
llvm::PredicateBase
Definition: PredicateInfo.h:82
DEBUG_COUNTER
DEBUG_COUNTER(VNCounter, "newgvn-vn", "Controls which instructions are value numbered")
BasicBlock.h
llvm::cl::opt< bool >
llvm::ValueDFS
Definition: PredicateInfo.cpp:93
llvm::RISCVFenceField::O
@ O
Definition: RISCVBaseInfo.h:239
llvm::PPC::Predicate
Predicate
Predicate - These are "(BI << 5) | BO" for various predicates.
Definition: PPCPredicates.h:26
llvm::StoreInst
An instruction for storing to memory.
Definition: Instructions.h:305
llvm::DebugCounter::setCounterValue
static void setCounterValue(unsigned ID, int64_t Count)
Definition: DebugCounter.h:115
llvm::Constant
This is an important base class in LLVM.
Definition: Constant.h:41
llvm::Instruction::eraseFromParent
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks 'this' from the containing basic block and deletes it.
Definition: Instruction.cpp:77
CFGPrinter.h
llvm::DenseMapBase< DenseMap< KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >, KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >::clear
void clear()
Definition: DenseMap.h:112
llvm::GVNExpression
Definition: GVNExpression.h:40
llvm::TargetLibraryInfoWrapperPass
Definition: TargetLibraryInfo.h:468
llvm::GVNExpression::AggregateValueExpression::~AggregateValueExpression
~AggregateValueExpression() override
D
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
llvm::BitVector::any
bool any() const
any - Returns true if any bit is set.
Definition: BitVector.h:163
const
aarch64 promote const
Definition: AArch64PromoteConstant.cpp:232
llvm::SmallPtrSetImpl::end
iterator end() const
Definition: SmallPtrSet.h:408
llvm::AssumptionAnalysis
A function analysis which provides an AssumptionCache.
Definition: AssumptionCache.h:173
llvm::PreservedAnalyses::preserve
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:173
INITIALIZE_PASS_DEPENDENCY
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
llvm::SimplifyBinOp
Value * SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, const SimplifyQuery &Q)
Given operands for a BinaryOperator, fold the result or return null.
Definition: InstructionSimplify.cpp:5542
llvm::BumpPtrAllocatorImpl
Allocate memory in an ever growing pool, as if by bump-pointer.
Definition: Allocator.h:63
llvm::numbers::e
constexpr double e
Definition: MathExtras.h:57
llvm::DenseMap
Definition: DenseMap.h:716
llvm::nodes
iterator_range< typename GraphTraits< GraphType >::nodes_iterator > nodes(const GraphType &G)
Definition: GraphTraits.h:110
llvm::MemorySSA
Encapsulates MemorySSA, including all data associated with memory accesses.
Definition: MemorySSA.h:714
I
#define I(x, y, z)
Definition: MD5.cpp:58
llvm::MemorySSA::getMemoryAccess
MemoryUseOrDef * getMemoryAccess(const Instruction *I) const
Given a memory Mod/Ref'ing instruction, get the MemorySSA access associated with it.
Definition: MemorySSA.h:731
VNCoercion.h
llvm::DenseMapInfo< const Expression * >::getHashValue
static unsigned getHashValue(const Expression *E)
Definition: NewGVN.cpp:457
llvm::cl::init
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:432
newgvn
newgvn
Definition: NewGVN.cpp:4256
llvm::SmallPtrSetImpl::begin
iterator begin() const
Definition: SmallPtrSet.h:403
ArrayRef.h
llvm::detail::DenseSetImpl< ValueT, DenseMap< ValueT, detail::DenseSetEmpty, DenseMapInfo< ValueT >, detail::DenseSetPair< ValueT > >, DenseMapInfo< ValueT > >::begin
iterator begin()
Definition: DenseSet.h:173
llvm::pdb::PDB_MemoryType::Stack
@ Stack
llvm::DenseMapBase< DenseMap< KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >, KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >::find
iterator find(const_arg_type_t< KeyT > Val)
Definition: DenseMap.h:152
llvm::sys::path::const_iterator::begin
friend const_iterator begin(StringRef path, Style style)
Get begin iterator over path.
Definition: Path.cpp:226
assert
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
std::swap
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:853
SI
StandardInstrumentations SI(Debug, VerifyEach)
llvm::GVNExpression::StoreExpression
Definition: GVNExpression.h:370
llvm::GVNExpression::VariableExpression
Definition: GVNExpression.h:552
iterator_range.h
okayForPHIOfOps
static bool okayForPHIOfOps(const Instruction *I)
Definition: NewGVN.cpp:2563
llvm::is_splat
bool is_splat(R &&Range)
Wrapper function around std::equal to detect if all elements in a container are same.
Definition: STLExtras.h:1781
llvm::WinEH::EncodingType::CE
@ CE
Windows NT (Windows on ARM)
llvm::salvageKnowledge
void salvageKnowledge(Instruction *I, AssumptionCache *AC=nullptr, DominatorTree *DT=nullptr)
Calls BuildAssumeFromInst and if the resulting llvm.assume is valid insert if before I.
Definition: AssumeBundleBuilder.cpp:293
llvm::MemorySSAAnalysis
An analysis that produces MemorySSA for a function.
Definition: MemorySSA.h:948
llvm::MemorySSA::getWalker
MemorySSAWalker * getWalker()
Definition: MemorySSA.cpp:1604
llvm::SmallPtrSetImpl::count
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:383
llvm::ArrayRecycler::clear
void clear(AllocatorType &Allocator)
Release all the tracked allocations to the allocator.
Definition: ArrayRecycler.h:104
llvm::PatternMatch::m_Value
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:76
llvm::detail::DenseSetImpl< ValueT, DenseMap< ValueT, detail::DenseSetEmpty, DenseMapInfo< ValueT >, detail::DenseSetPair< ValueT > >, DenseMapInfo< ValueT > >::clear
void clear()
Definition: DenseSet.h:92
llvm::size
auto size(R &&Range, std::enable_if_t< std::is_base_of< std::random_access_iterator_tag, typename std::iterator_traits< decltype(Range.begin())>::iterator_category >::value, void > *=nullptr)
Get the size of a range.
Definition: STLExtras.h:1586
llvm::GVNExpression::StoreExpression::equals
bool equals(const Expression &Other) const override
Definition: NewGVN.cpp:917
llvm::AMDGPU::CPol::SCC
@ SCC
Definition: SIDefines.h:296
llvm::ConstantFoldInstOperands
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.
Definition: ConstantFolding.cpp:1165
llvm::Expression
Class representing an expression and its matching format.
Definition: FileCheckImpl.h:237
llvm::SmallPtrSetImplBase::clear
void clear()
Definition: SmallPtrSet.h:95
llvm::AssumptionCacheTracker
An immutable pass that tracks lazily created AssumptionCache objects.
Definition: AssumptionCache.h:202
llvm::ArrayRef
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition: APInt.h:32
llvm::isInstructionTriviallyDead
bool isInstructionTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction is not used, and the instruction will return.
Definition: Local.cpp:395
llvm::VNCoercion::getConstantLoadValueForLoad
Constant * getConstantLoadValueForLoad(Constant *SrcVal, unsigned Offset, Type *LoadTy, const DataLayout &DL)
Definition: VNCoercion.cpp:515
llvm::min
Expected< ExpressionValue > min(const ExpressionValue &Lhs, const ExpressionValue &Rhs)
Definition: FileCheck.cpp:357
Mul
BinaryOperator * Mul
Definition: X86PartialReduction.cpp:70
llvm::any_of
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:1612
Cond
SmallVector< MachineOperand, 4 > Cond
Definition: BasicBlockSections.cpp:178
llvm::LoadInst::isSimple
bool isSimple() const
Definition: Instructions.h:260
llvm::AssumptionCache
A cache of @llvm.assume calls within a function.
Definition: AssumptionCache.h:42
llvm_unreachable
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
Definition: ErrorHandling.h:143
llvm::Value::getType
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
llvm::MemorySSA::getBlockDefs
const DefsList * getBlockDefs(const BasicBlock *BB) const
Return the list of MemoryDef's and MemoryPhi's for a given basic block.
Definition: MemorySSA.h:779
if
if(llvm_vc STREQUAL "") set(fake_version_inc "$
Definition: CMakeLists.txt:14
llvm::ConstantInt::isZero
bool isZero() const
This is just a convenience method to make client code smaller for a common code.
Definition: Constants.h:194
llvm::AnalysisUsage::addPreserved
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
Definition: PassAnalysisSupport.h:98
llvm::Value::replaceAllUsesWith
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:529
A
* A
Definition: README_ALTIVEC.txt:89
DL
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
Definition: AArch64SLSHardening.cpp:76
llvm::ExactEqualsExpression
Definition: NewGVN.cpp:432
S
add sub stmia L5 ldr r0 bl L_printf $stub Instead of a and a wouldn t it be better to do three moves *Return an aggregate type is even return S
Definition: README.txt:210
NewGVN.h
llvm::GVNExpression::CallExpression
Definition: GVNExpression.h:301
llvm::DominatorTreeBase::properlyDominates
bool properlyDominates(const DomTreeNodeBase< NodeT > *A, const DomTreeNodeBase< NodeT > *B) const
properlyDominates - Returns true iff A dominates B and A != B.
Definition: GenericDomTree.h:392
llvm::MemoryAccess
Definition: MemorySSA.h:142
llvm::DomTreeNodeBase< BasicBlock >
llvm::LoadInst
An instruction for reading from memory.
Definition: Instructions.h:176
llvm::DenseMapBase< DenseMap< KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >, KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >::insert
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:209
llvm::make_filter_range
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:524
llvm::DenseMapInfo< const Expression * >::getHashValue
static unsigned getHashValue(const ExactEqualsExpression &E)
Definition: NewGVN.cpp:461
llvm::ValueDFS::Def
Value * Def
Definition: PredicateInfo.cpp:98
llvm::ConstantInt::getFalse
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:874
Argument.h
llvm::ArrayRecycler
Recycle small arrays allocated from a BumpPtrAllocator.
Definition: ArrayRecycler.h:28
llvm::depth_first
iterator_range< df_iterator< T > > depth_first(const T &G)
Definition: DepthFirstIterator.h:230
llvm::MCID::Select
@ Select
Definition: MCInstrDesc.h:164
runOnFunction
static bool runOnFunction(Function &F, bool PostInlining)
Definition: EntryExitInstrumenter.cpp:69
llvm::GVNExpression::Expression::~Expression
virtual ~Expression()
llvm::VNCoercion::getConstantMemInstValueForLoad
Constant * getConstantMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, Type *LoadTy, const DataLayout &DL)
Definition: VNCoercion.cpp:584
llvm::CmpInst::isImpliedFalseByMatchingCmp
static bool isImpliedFalseByMatchingCmp(Predicate Pred1, Predicate Pred2)
Determine if Pred1 implies Pred2 is false when two compares have matching operands.
Definition: Instructions.cpp:4303
llvm::Instruction::isAtomic
bool isAtomic() const
Return true if this instruction has an AtomicOrdering of unordered or higher.
Definition: Instruction.cpp:616
Constant.h
llvm::CmpInst::isImpliedTrueByMatchingCmp
static bool isImpliedTrueByMatchingCmp(Predicate Pred1, Predicate Pred2)
Determine if Pred1 implies Pred2 is true when two compares have matching operands.
Definition: Instructions.cpp:4278
llvm::empty
constexpr bool empty(const T &RangeOrContainer)
Test whether RangeOrContainer is empty. Similar to C++17 std::empty.
Definition: STLExtras.h:268
llvm::ConstantInt::getTrue
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:867
std
Definition: BitVector.h:851
llvm::Constant::getNullValue
static Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition: Constants.cpp:350
llvm::DenseMapBase< DenseMap< KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >, KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >::end
iterator end()
Definition: DenseMap.h:84
llvm::AMDGPU::SendMsg::Op
Op
Definition: SIDefines.h:337
llvm::PreservedAnalyses::all
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:158
llvm::PHINode::Create
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...
Definition: Instructions.h:2706
llvm::ArrayRef::begin
iterator begin() const
Definition: ArrayRef.h:152
llvm::VNCoercion::analyzeLoadFromClobberingStore
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:215
Casting.h
Function.h
llvm::GVNExpression::DeadExpression
Definition: GVNExpression.h:541
DebugCounter.h
llvm::DenseMapBase::size
unsigned size() const
Definition: DenseMap.h:101
llvm::sort
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:1550
llvm::TargetStackID::Value
Value
Definition: TargetFrameLowering.h:27
llvm::Type::getPointerTo
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:774
llvm::TargetLibraryInfo
Provides information about what library functions are available for the current target.
Definition: TargetLibraryInfo.h:222
Numbering
Global Value Numbering
Definition: NewGVN.cpp:4256
llvm::MemoryUseOrDef
Class that has the common methods + fields of memory uses/defs.
Definition: MemorySSA.h:252
transform
instcombine should handle this transform
Definition: README.txt:262
llvm::SmallVectorImpl::clear
void clear()
Definition: SmallVector.h:591
llvm::ExactEqualsExpression::E
const Expression & E
Definition: NewGVN.cpp:433
llvm::MCID::Add
@ Add
Definition: MCInstrDesc.h:185
llvm::GVNExpression::LoadExpression::~LoadExpression
~LoadExpression() override
llvm::ReversePostOrderTraversal
Definition: PostOrderIterator.h:291
llvm::BitVector::reset
BitVector & reset()
Definition: BitVector.h:385
llvm::IntrinsicInst
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:46
llvm::VNCoercion::getConstantStoreValueForLoad
Constant * getConstantStoreValueForLoad(Constant *SrcVal, unsigned Offset, Type *LoadTy, const DataLayout &DL)
Definition: VNCoercion.cpp:456
llvm::makeArrayRef
ArrayRef< T > makeArrayRef(const T &OneElt)
Construct an ArrayRef from a single element.
Definition: ArrayRef.h:475
AA
llvm::NewGVNPass::run
PreservedAnalyses run(Function &F, AnalysisManager< Function > &AM)
Run the pass over the function.
Definition: NewGVN.cpp:4262
llvm::DominatorTreeAnalysis
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:267
llvm::DenseMapInfo< const Expression * >::getEmptyKey
static const Expression * getEmptyKey()
Definition: NewGVN.cpp:445
llvm::BasicBlock::reverse_iterator
InstListType::reverse_iterator reverse_iterator
Definition: BasicBlock.h:89
MemorySSA.h
Instructions.h
PostOrderIterator.h
llvm::GVNExpression::AggregateValueExpression
Definition: GVNExpression.h:411
llvm::User::getNumOperands
unsigned getNumOperands() const
Definition: User.h:191
llvm::PointerIntPair< Value *, 1, bool >
llvm::ValueDFS::DFSOut
int DFSOut
Definition: PredicateInfo.cpp:95
SmallVector.h
lookup
static bool lookup(const GsymReader &GR, DataExtractor &Data, uint64_t &Offset, uint64_t BaseAddr, uint64_t Addr, SourceLocations &SrcLocs, llvm::Error &Err)
A Lookup helper functions.
Definition: InlineInfo.cpp:108
User.h
ArrayRecycler.h
Dominators.h
llvm::BitVector::find_first
int find_first() const
find_first - Returns the index of the first set bit, -1 if none of the bits are set.
Definition: BitVector.h:293
llvm::AAResultsWrapperPass
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object.
Definition: AliasAnalysis.h:1347
llvm::Instruction::getParent
const BasicBlock * getParent() const
Definition: Instruction.h:91
InstructionSimplify.h
llvm::Value::deleteValue
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:106
llvm::VNCoercion::analyzeLoadFromClobberingLoad
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:321
llvm::GlobalsAAWrapperPass
Legacy wrapper pass to provide the GlobalsAAResult object.
Definition: GlobalsModRef.h:148
llvm::PHINode::getIncomingBlock
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
Definition: Instructions.h:2780
llvm::ArrayRef::size
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:164
llvm::max
Align max(MaybeAlign Lhs, Align Rhs)
Definition: Alignment.h:340
llvm::iterator_range
A range adaptor for a pair of iterators.
Definition: iterator_range.h:30
llvm::PHINode
Definition: Instructions.h:2664
llvm::GVNExpression::BasicExpression
Definition: GVNExpression.h:136
llvm::PatternMatch
Definition: PatternMatch.h:47
llvm::detail::DenseSetImpl< ValueT, DenseMap< ValueT, detail::DenseSetEmpty, DenseMapInfo< ValueT >, detail::DenseSetPair< ValueT > >, DenseMapInfo< ValueT > >::erase
bool erase(const ValueT &V)
Definition: DenseSet.h:101
llvm::SmallVectorImpl< Instruction * >
DenseMapInfo.h
llvm::SmallPtrSetImpl< const Value * >
llvm::AnalysisManager
A container for analyses that lazily runs them and caches their results.
Definition: InstructionSimplify.h:42
llvm::FunctionPass
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:308
llvm::CallInst
This class represents a function call, abstracting a target machine's calling convention.
Definition: Instructions.h:1474
llvm::ConstantInt::isOne
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
Definition: Constants.h:200
llvm::isGuaranteedNotToBePoison
bool isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, unsigned Depth=0)
Definition: ValueTracking.cpp:5253
BB
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM BB
Definition: README.txt:39
GEP
Hexagon Common GEP
Definition: HexagonCommonGEP.cpp:172
llvm::GVNExpression::op_inserter
Definition: GVNExpression.h:243
llvm::patchReplacementInstruction
void patchReplacementInstruction(Instruction *I, Value *Repl)
Patch the replacement so that it is not more restrictive than the value being replaced.
Definition: Local.cpp:2667
llvm::AnalysisUsage::addRequired
AnalysisUsage & addRequired()
Definition: PassAnalysisSupport.h:75
From
BlockVerifier::State From
Definition: BlockVerifier.cpp:55
llvm::MemorySSAWalker
This is the generic walker interface for walkers of MemorySSA.
Definition: MemorySSA.h:1029
llvm::User::getOperand
Value * getOperand(unsigned i) const
Definition: User.h:169
raw_ostream.h
llvm::SmallPtrSetImpl::contains
bool contains(ConstPtrType Ptr) const
Definition: SmallPtrSet.h:389
copy
we should consider alternate ways to model stack dependencies Lots of things could be done in WebAssemblyTargetTransformInfo cpp there are numerous optimization related hooks that can be overridden in WebAssemblyTargetLowering Instead of the OptimizeReturned which should consider preserving the returned attribute through to MachineInstrs and extending the MemIntrinsicResults pass to do this optimization on calls too That would also let the WebAssemblyPeephole pass clean up dead defs for such as it does for stores Consider implementing and or getMachineCombinerPatterns Find a clean way to fix the problem which leads to the Shrink Wrapping pass being run after the WebAssembly PEI pass When setting multiple variables to the same we currently get code like const It could be done with a smaller encoding like local tee $pop5 local copy
Definition: README.txt:101
Value.h
llvm::GVNExpression::LoadExpression
Definition: GVNExpression.h:328
InitializePasses.h
llvm::Value
LLVM Value Representation.
Definition: Value.h:74
Debug.h
llvm::TargetLibraryAnalysis
Analysis pass providing the TargetLibraryInfo.
Definition: TargetLibraryInfo.h:443
llvm::Value::users
iterator_range< user_iterator > users()
Definition: Value.h:421
llvm::BumpPtrAllocatorImpl::Deallocate
void Deallocate(const void *Ptr, size_t Size, size_t)
Definition: Allocator.h:216
Other
Optional< std::vector< StOtherPiece > > Other
Definition: ELFYAML.cpp:1225
llvm::Use
A Use represents the edge between a Value definition and its users.
Definition: Use.h:43
llvm::SmallVectorImpl::emplace_back
reference emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:927
llvm::SmallPtrSetImpl::insert
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:365
llvm::Intrinsic::ID
unsigned ID
Definition: TargetTransformInfo.h:37
llvm::hash_code
An opaque object representing a hash code.
Definition: Hashing.h:73
llvm::PoisonValue::get
static PoisonValue * get(Type *T)
Static factory methods - Return an 'poison' object of the specified type.
Definition: Constants.cpp:1788