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