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 "ScalarOptions.h"
56#include "llvm/ADT/ArrayRef.h"
57#include "llvm/ADT/BitVector.h"
58#include "llvm/ADT/DenseMap.h"
59#include "llvm/ADT/DenseMapInfo.h"
60#include "llvm/ADT/DenseSet.h"
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/DebugInfo.h"
86#include "llvm/IR/Dominators.h"
87#include "llvm/IR/Function.h"
88#include "llvm/IR/InstrTypes.h"
89#include "llvm/IR/Instruction.h"
90#include "llvm/IR/Instructions.h"
91#include "llvm/IR/IntrinsicInst.h"
92#include "llvm/IR/PatternMatch.h"
93#include "llvm/IR/Type.h"
94#include "llvm/IR/Use.h"
95#include "llvm/IR/User.h"
96#include "llvm/IR/Value.h"
97#include "llvm/Support/Allocator.h"
98#include "llvm/Support/ArrayRecycler.h"
99#include "llvm/Support/Casting.h"
100#include "llvm/Support/Debug.h"
101#include "llvm/Support/DebugCounter.h"
102#include "llvm/Support/ErrorHandling.h"
103#include "llvm/Support/raw_ostream.h"
104#include "llvm/Transforms/Scalar/GVNExpression.h"
105#include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
106#include "llvm/Transforms/Utils/Local.h"
107#include "llvm/Transforms/Utils/PredicateInfo.h"
108#include "llvm/Transforms/Utils/VNCoercion.h"
109#include <algorithm>
110#include <cassert>
111#include <cstdint>
112#include <iterator>
113#include <map>
114#include <memory>
115#include <set>
116#include <string>
117#include <tuple>
118#include <utility>
119#include <vector>
120
121using namespace llvm;
122using namespace llvm::GVNExpression;
123using namespace llvm::VNCoercion;
124using namespace llvm::PatternMatch;
125
126#define DEBUG_TYPE "newgvn"
127
128STATISTIC(NumGVNInstrDeleted, "Number of instructions deleted");
129STATISTIC(NumGVNBlocksDeleted, "Number of blocks deleted");
130STATISTIC(NumGVNOpsSimplified, "Number of Expressions simplified");
131STATISTIC(NumGVNPhisAllSame, "Number of PHIs whos arguments are all the same");
132STATISTIC(NumGVNMaxIterations,
133 "Maximum Number of iterations it took to converge GVN");
134STATISTIC(NumGVNLeaderChanges, "Number of leader changes");
135STATISTIC(NumGVNSortedLeaderChanges, "Number of sorted leader changes");
136STATISTIC(NumGVNAvoidedSortedLeaderChanges,
137 "Number of avoided sorted leader changes");
138STATISTIC(NumGVNDeadStores, "Number of redundant/dead stores eliminated");
139STATISTIC(NumGVNPHIOfOpsCreated, "Number of PHI of ops created");
140STATISTIC(NumGVNPHIOfOpsEliminations,
141 "Number of things eliminated using PHI of ops");
142DEBUG_COUNTER(VNCounter, "newgvn-vn",
143 "Controls which instructions are value numbered");
144DEBUG_COUNTER(PHIOfOpsCounter, "newgvn-phi",
145 "Controls which instructions we create phi of ops for");
146
147//===----------------------------------------------------------------------===//
148// GVN Pass
149//===----------------------------------------------------------------------===//
150
151// Anchor methods.
152Expression::~Expression() = default;
153BasicExpression::~BasicExpression() = default;
154CallExpression::~CallExpression() = default;
155LoadExpression::~LoadExpression() = default;
156StoreExpression::~StoreExpression() = default;
157AggregateValueExpression::~AggregateValueExpression() = default;
158PHIExpression::~PHIExpression() = default;
159
160namespace {
161
162// Tarjan's SCC finding algorithm with Nuutila's improvements
163// SCCIterator is actually fairly complex for the simple thing we want.
164// It also wants to hand us SCC's that are unrelated to the phi node we ask
165// about, and have us process them there or risk redoing work.
166// Graph traits over a filter iterator also doesn't work that well here.
167// This SCC finder is specialized to walk use-def chains, and only follows
168// instructions,
169// not generic values (arguments, etc).
170struct TarjanSCC {
171 TarjanSCC() : Components(1) {}
172
173 void Start(const Instruction *Start) {
174 if (Root.lookup(Val: Start) == 0)
175 FindSCC(I: Start);
176 }
177
178 const SmallPtrSetImpl<const Value *> &getComponentFor(const Value *V) const {
179 unsigned ComponentID = ValueToComponent.lookup(Val: V);
180
181 assert(ComponentID > 0 &&
182 "Asking for a component for a value we never processed");
183 return Components[ComponentID];
184 }
185
186private:
187 void FindSCC(const Instruction *I) {
188 Root[I] = ++DFSNum;
189 // Store the DFS Number we had before it possibly gets incremented.
190 unsigned int OurDFS = DFSNum;
191 for (const auto &Op : I->operands()) {
192 if (auto *InstOp = dyn_cast<Instruction>(Val: Op)) {
193 if (Root.lookup(Val: Op) == 0)
194 FindSCC(I: InstOp);
195 if (!InComponent.count(Ptr: Op))
196 Root[I] = std::min(a: Root.lookup(Val: I), b: Root.lookup(Val: Op));
197 }
198 }
199 // See if we really were the root of a component, by seeing if we still have
200 // our DFSNumber. If we do, we are the root of the component, and we have
201 // completed a component. If we do not, we are not the root of a component,
202 // and belong on the component stack.
203 if (Root.lookup(Val: I) == OurDFS) {
204 unsigned ComponentID = Components.size();
205 Components.resize(N: Components.size() + 1);
206 auto &Component = Components.back();
207 Component.insert(Ptr: I);
208 LLVM_DEBUG(dbgs() << "Component root is " << *I << "\n");
209 InComponent.insert(Ptr: I);
210 ValueToComponent[I] = ComponentID;
211 // Pop a component off the stack and label it.
212 while (!Stack.empty() && Root.lookup(Val: Stack.back()) >= OurDFS) {
213 auto *Member = Stack.back();
214 LLVM_DEBUG(dbgs() << "Component member is " << *Member << "\n");
215 Component.insert(Ptr: Member);
216 InComponent.insert(Ptr: Member);
217 ValueToComponent[Member] = ComponentID;
218 Stack.pop_back();
219 }
220 } else {
221 // Part of a component, push to stack
222 Stack.push_back(Elt: I);
223 }
224 }
225
226 unsigned int DFSNum = 1;
227 SmallPtrSet<const Value *, 8> InComponent;
228 DenseMap<const Value *, unsigned int> Root;
229 SmallVector<const Value *, 8> Stack;
230
231 // Store the components as vector of ptr sets, because we need the topo order
232 // of SCC's, but not individual member order
233 SmallVector<SmallPtrSet<const Value *, 8>, 8> Components;
234
235 DenseMap<const Value *, unsigned> ValueToComponent;
236};
237
238// Congruence classes represent the set of expressions/instructions
239// that are all the same *during some scope in the function*.
240// That is, because of the way we perform equality propagation, and
241// because of memory value numbering, it is not correct to assume
242// you can willy-nilly replace any member with any other at any
243// point in the function.
244//
245// For any Value in the Member set, it is valid to replace any dominated member
246// with that Value.
247//
248// Every congruence class has a leader, and the leader is used to symbolize
249// instructions in a canonical way (IE every operand of an instruction that is a
250// member of the same congruence class will always be replaced with leader
251// during symbolization). To simplify symbolization, we keep the leader as a
252// constant if class can be proved to be a constant value. Otherwise, the
253// leader is the member of the value set with the smallest DFS number. Each
254// congruence class also has a defining expression, though the expression may be
255// null. If it exists, it can be used for forward propagation and reassociation
256// of values.
257
258// For memory, we also track a representative MemoryAccess, and a set of memory
259// members for MemoryPhis (which have no real instructions). Note that for
260// memory, it seems tempting to try to split the memory members into a
261// MemoryCongruenceClass or something. Unfortunately, this does not work
262// easily. The value numbering of a given memory expression depends on the
263// leader of the memory congruence class, and the leader of memory congruence
264// class depends on the value numbering of a given memory expression. This
265// leads to wasted propagation, and in some cases, missed optimization. For
266// example: If we had value numbered two stores together before, but now do not,
267// we move them to a new value congruence class. This in turn will move at one
268// of the memorydefs to a new memory congruence class. Which in turn, affects
269// the value numbering of the stores we just value numbered (because the memory
270// congruence class is part of the value number). So while theoretically
271// possible to split them up, it turns out to be *incredibly* complicated to get
272// it to work right, because of the interdependency. While structurally
273// slightly messier, it is algorithmically much simpler and faster to do what we
274// do here, and track them both at once in the same class.
275// Note: The default iterators for this class iterate over values
276class CongruenceClass {
277public:
278 using MemberType = Value;
279 using MemberSet = SmallPtrSet<MemberType *, 4>;
280 using MemoryMemberType = MemoryPhi;
281 using MemoryMemberSet = SmallPtrSet<const MemoryMemberType *, 2>;
282
283 explicit CongruenceClass(unsigned ID) : ID(ID) {}
284 CongruenceClass(unsigned ID, std::pair<Value *, unsigned int> Leader,
285 const Expression *E)
286 : ID(ID), RepLeader(Leader), DefiningExpr(E) {}
287
288 unsigned getID() const { return ID; }
289
290 // True if this class has no members left. This is mainly used for assertion
291 // purposes, and for skipping empty classes.
292 bool isDead() const {
293 // If it's both dead from a value perspective, and dead from a memory
294 // perspective, it's really dead.
295 return empty() && memory_empty();
296 }
297
298 // Leader functions
299 Value *getLeader() const { return RepLeader.first; }
300 void setLeader(std::pair<Value *, unsigned int> Leader) {
301 RepLeader = std::move(Leader);
302 }
303 const std::pair<Value *, unsigned int> &getNextLeader() const {
304 return NextLeader;
305 }
306 void resetNextLeader() { NextLeader = {nullptr, ~0}; }
307 bool addPossibleLeader(std::pair<Value *, unsigned int> LeaderPair) {
308 if (LeaderPair.second < RepLeader.second) {
309 NextLeader = RepLeader;
310 RepLeader = std::move(LeaderPair);
311 return true;
312 } else if (LeaderPair.second < NextLeader.second) {
313 NextLeader = std::move(LeaderPair);
314 }
315 return false;
316 }
317
318 Value *getStoredValue() const { return RepStoredValue; }
319 void setStoredValue(Value *Leader) { RepStoredValue = Leader; }
320 const MemoryAccess *getMemoryLeader() const { return RepMemoryAccess; }
321 void setMemoryLeader(const MemoryAccess *Leader) { RepMemoryAccess = Leader; }
322
323 // Forward propagation info
324 const Expression *getDefiningExpr() const { return DefiningExpr; }
325
326 // Value member set
327 bool empty() const { return Members.empty(); }
328 unsigned size() const { return Members.size(); }
329 MemberSet::const_iterator begin() const { return Members.begin(); }
330 MemberSet::const_iterator end() const { return Members.end(); }
331 void insert(MemberType *M) { Members.insert(Ptr: M); }
332 void erase(MemberType *M) { Members.erase(Ptr: M); }
333 void swap(MemberSet &Other) { Members.swap(RHS&: Other); }
334
335 // Memory member set
336 bool memory_empty() const { return MemoryMembers.empty(); }
337 unsigned memory_size() const { return MemoryMembers.size(); }
338 MemoryMemberSet::const_iterator memory_begin() const {
339 return MemoryMembers.begin();
340 }
341 MemoryMemberSet::const_iterator memory_end() const {
342 return MemoryMembers.end();
343 }
344 iterator_range<MemoryMemberSet::const_iterator> memory() const {
345 return make_range(x: memory_begin(), y: memory_end());
346 }
347
348 void memory_insert(const MemoryMemberType *M) { MemoryMembers.insert(Ptr: M); }
349 void memory_erase(const MemoryMemberType *M) { MemoryMembers.erase(Ptr: M); }
350
351 // Store count
352 unsigned getStoreCount() const { return StoreCount; }
353 void incStoreCount() { ++StoreCount; }
354 void decStoreCount() {
355 assert(StoreCount != 0 && "Store count went negative");
356 --StoreCount;
357 }
358
359 // True if this class has no memory members.
360 bool definesNoMemory() const { return StoreCount == 0 && memory_empty(); }
361
362 // Return true if two congruence classes are equivalent to each other. This
363 // means that every field but the ID number and the dead field are equivalent.
364 bool isEquivalentTo(const CongruenceClass *Other) const {
365 if (!Other)
366 return false;
367 if (this == Other)
368 return true;
369
370 if (std::tie(args: StoreCount, args: RepLeader, args: RepStoredValue, args: RepMemoryAccess) !=
371 std::tie(args: Other->StoreCount, args: Other->RepLeader, args: Other->RepStoredValue,
372 args: Other->RepMemoryAccess))
373 return false;
374 if (DefiningExpr != Other->DefiningExpr)
375 if (!DefiningExpr || !Other->DefiningExpr ||
376 *DefiningExpr != *Other->DefiningExpr)
377 return false;
378
379 if (Members.size() != Other->Members.size())
380 return false;
381
382 return llvm::set_is_subset(S1: Members, S2: Other->Members);
383 }
384
385private:
386 unsigned ID;
387
388 // Representative leader and its corresponding RPO number.
389 // The leader must have the lowest RPO number.
390 std::pair<Value *, unsigned int> RepLeader = {nullptr, ~0U};
391
392 // The most dominating leader after our current leader (given by the RPO
393 // number), because the member set is not sorted and is expensive to keep
394 // sorted all the time.
395 std::pair<Value *, unsigned int> NextLeader = {nullptr, ~0U};
396
397 // If this is represented by a store, the value of the store.
398 Value *RepStoredValue = nullptr;
399
400 // If this class contains MemoryDefs or MemoryPhis, this is the leading memory
401 // access.
402 const MemoryAccess *RepMemoryAccess = nullptr;
403
404 // Defining Expression.
405 const Expression *DefiningExpr = nullptr;
406
407 // Actual members of this class.
408 MemberSet Members;
409
410 // This is the set of MemoryPhis that exist in the class. MemoryDefs and
411 // MemoryUses have real instructions representing them, so we only need to
412 // track MemoryPhis here.
413 MemoryMemberSet MemoryMembers;
414
415 // Number of stores in this congruence class.
416 // This is used so we can detect store equivalence changes properly.
417 int StoreCount = 0;
418};
419
420struct ExactEqualsExpression {
421 const Expression &E;
422
423 explicit ExactEqualsExpression(const Expression &E) : E(E) {}
424
425 hash_code getComputedHash() const { return E.getComputedHash(); }
426
427 bool operator==(const Expression &Other) const {
428 return E.exactlyEquals(Other);
429 }
430};
431} // end anonymous namespace
432
433template <> struct llvm::DenseMapInfo<const Expression *> {
434 static unsigned getHashValue(const Expression *E) {
435 return E->getComputedHash();
436 }
437
438 static unsigned getHashValue(const ExactEqualsExpression &E) {
439 return E.getComputedHash();
440 }
441
442 static bool isEqual(const ExactEqualsExpression &LHS, const Expression *RHS) {
443 return LHS == *RHS;
444 }
445
446 static bool isEqual(const Expression *LHS, const Expression *RHS) {
447 if (LHS == RHS)
448 return true;
449 // Compare hashes before equality. This is *not* what the hashtable does,
450 // since it is computing it modulo the number of buckets, whereas we are
451 // using the full hash keyspace. Since the hashes are precomputed, this
452 // check is *much* faster than equality.
453 if (LHS->getComputedHash() != RHS->getComputedHash())
454 return false;
455 return *LHS == *RHS;
456 }
457};
458
459namespace {
460
461class NewGVN {
462 const ScalarOptions &Opts;
463 Function &F;
464 DominatorTree *DT = nullptr;
465 const TargetLibraryInfo *TLI = nullptr;
466 AliasAnalysis *AA = nullptr;
467 MemorySSA *MSSA = nullptr;
468 MemorySSAWalker *MSSAWalker = nullptr;
469 AssumptionCache *AC = nullptr;
470 const DataLayout &DL;
471
472 // These are the only two things the create* functions should have
473 // side-effects on due to allocating memory.
474 mutable BumpPtrAllocator ExpressionAllocator;
475 mutable ArrayRecycler<Value *> ArgRecycler;
476 mutable TarjanSCC SCCFinder;
477
478 std::unique_ptr<PredicateInfo> PredInfo;
479 const SimplifyQuery SQ;
480
481 // Number of function arguments, used by ranking
482 unsigned int NumFuncArgs = 0;
483
484 // RPOOrdering of basic blocks
485 DenseMap<const DomTreeNode *, unsigned> RPOOrdering;
486
487 // Congruence class info.
488
489 // This class is called INITIAL in the paper. It is the class everything
490 // startsout in, and represents any value. Being an optimistic analysis,
491 // anything in the TOP class has the value TOP, which is indeterminate and
492 // equivalent to everything.
493 CongruenceClass *TOPClass = nullptr;
494 std::vector<CongruenceClass *> CongruenceClasses;
495 unsigned NextCongruenceNum = 0;
496
497 // Value Mappings.
498 DenseMap<Value *, CongruenceClass *> ValueToClass;
499 DenseMap<Value *, const Expression *> ValueToExpression;
500
501 // Value PHI handling, used to make equivalence between phi(op, op) and
502 // op(phi, phi).
503 // These mappings just store various data that would normally be part of the
504 // IR.
505 SmallPtrSet<const Instruction *, 8> PHINodeUses;
506
507 // The cached results, in general, are only valid for the specific block where
508 // they were computed. The unsigned part of the key is a unique block
509 // identifier
510 DenseMap<std::pair<const Value *, unsigned>, bool> OpSafeForPHIOfOps;
511 unsigned CacheIdx;
512
513 // Map a temporary instruction we created to a parent block.
514 DenseMap<const Value *, BasicBlock *> TempToBlock;
515
516 // Map between the already in-program instructions and the temporary phis we
517 // created that they are known equivalent to.
518 DenseMap<const Value *, PHINode *> RealToTemp;
519
520 // In order to know when we should re-process instructions that have
521 // phi-of-ops, we track the set of expressions that they needed as
522 // leaders. When we discover new leaders for those expressions, we process the
523 // associated phi-of-op instructions again in case they have changed. The
524 // other way they may change is if they had leaders, and those leaders
525 // disappear. However, at the point they have leaders, there are uses of the
526 // relevant operands in the created phi node, and so they will get reprocessed
527 // through the normal user marking we perform.
528 mutable DenseMap<const Value *, SmallPtrSet<Value *, 2>> AdditionalUsers;
529 DenseMap<const Expression *, SmallPtrSet<Instruction *, 2>>
530 ExpressionToPhiOfOps;
531
532 // Map from temporary operation to MemoryAccess.
533 DenseMap<const Instruction *, MemoryUseOrDef *> TempToMemory;
534
535 // Set of all temporary instructions we created.
536 // Note: This will include instructions that were just created during value
537 // numbering. The way to test if something is using them is to check
538 // RealToTemp.
539 DenseSet<Instruction *> AllTempInstructions;
540
541 // This is the set of instructions to revisit on a reachability change. At
542 // the end of the main iteration loop it will contain at least all the phi of
543 // ops instructions that will be changed to phis, as well as regular phis.
544 // During the iteration loop, it may contain other things, such as phi of ops
545 // instructions that used edge reachability to reach a result, and so need to
546 // be revisited when the edge changes, independent of whether the phi they
547 // depended on changes.
548 DenseMap<BasicBlock *, SparseBitVector<>> RevisitOnReachabilityChange;
549
550 // Mapping from predicate info we used to the instructions we used it with.
551 // In order to correctly ensure propagation, we must keep track of what
552 // comparisons we used, so that when the values of the comparisons change, we
553 // propagate the information to the places we used the comparison.
554 mutable DenseMap<const Value *, SmallPtrSet<Instruction *, 2>>
555 PredicateToUsers;
556
557 // the same reasoning as PredicateToUsers. When we skip MemoryAccesses for
558 // stores, we no longer can rely solely on the def-use chains of MemorySSA.
559 mutable DenseMap<const MemoryAccess *, SmallPtrSet<MemoryAccess *, 2>>
560 MemoryToUsers;
561
562 // A table storing which memorydefs/phis represent a memory state provably
563 // equivalent to another memory state.
564 // We could use the congruence class machinery, but the MemoryAccess's are
565 // abstract memory states, so they can only ever be equivalent to each other,
566 // and not to constants, etc.
567 DenseMap<const MemoryAccess *, CongruenceClass *> MemoryAccessToClass;
568
569 // We could, if we wanted, build MemoryPhiExpressions and
570 // MemoryVariableExpressions, etc, and value number them the same way we value
571 // number phi expressions. For the moment, this seems like overkill. They
572 // can only exist in one of three states: they can be TOP (equal to
573 // everything), Equivalent to something else, or unique. Because we do not
574 // create expressions for them, we need to simulate leader change not just
575 // when they change class, but when they change state. Note: We can do the
576 // same thing for phis, and avoid having phi expressions if we wanted, We
577 // should eventually unify in one direction or the other, so this is a little
578 // bit of an experiment in which turns out easier to maintain.
579 enum MemoryPhiState { MPS_Invalid, MPS_TOP, MPS_Equivalent, MPS_Unique };
580 DenseMap<const MemoryPhi *, MemoryPhiState> MemoryPhiState;
581
582 enum InstCycleState { ICS_Unknown, ICS_CycleFree, ICS_Cycle };
583 mutable DenseMap<const Instruction *, InstCycleState> InstCycleState;
584
585 // Expression to class mapping.
586 using ExpressionClassMap = DenseMap<const Expression *, CongruenceClass *>;
587 ExpressionClassMap ExpressionToClass;
588
589 // We have a single expression that represents currently DeadExpressions.
590 // For dead expressions we can prove will stay dead, we mark them with
591 // DFS number zero. However, it's possible in the case of phi nodes
592 // for us to assume/prove all arguments are dead during fixpointing.
593 // We use DeadExpression for that case.
594 DeadExpression *SingletonDeadExpression = nullptr;
595
596 // Which values have changed as a result of leader changes.
597 SmallPtrSet<Value *, 8> LeaderChanges;
598
599 // Reachability info.
600 using BlockEdge = BasicBlockEdge;
601 DenseSet<BlockEdge> ReachableEdges;
602 SmallPtrSet<const BasicBlock *, 8> ReachableBlocks;
603
604 // This is a bitvector because, on larger functions, we may have
605 // thousands of touched instructions at once (entire blocks,
606 // instructions with hundreds of uses, etc). Even with optimization
607 // for when we mark whole blocks as touched, when this was a
608 // SmallPtrSet or DenseSet, for some functions, we spent >20% of all
609 // the time in GVN just managing this list. The bitvector, on the
610 // other hand, efficiently supports test/set/clear of both
611 // individual and ranges, as well as "find next element" This
612 // enables us to use it as a worklist with essentially 0 cost.
613 BitVector TouchedInstructions;
614
615 DenseMap<const BasicBlock *, std::pair<unsigned, unsigned>> BlockInstRange;
616 mutable DenseMap<const BitCastInst *, const Value *> PredicateSwapChoice;
617
618#ifndef NDEBUG
619 // Debugging for how many times each block and instruction got processed.
620 DenseMap<const Value *, unsigned> ProcessedCount;
621#endif
622
623 // DFS info.
624 // This contains a mapping from Instructions to DFS numbers.
625 // The numbering starts at 1. An instruction with DFS number zero
626 // means that the instruction is dead.
627 DenseMap<const Value *, unsigned> InstrDFS;
628
629 // This contains the mapping DFS numbers to instructions.
630 SmallVector<Value *, 32> DFSToInstr;
631
632 // Deletion info.
633 SmallPtrSet<Instruction *, 8> InstructionsToErase;
634
635public:
636 NewGVN(Function &F, DominatorTree *DT, AssumptionCache *AC,
637 TargetLibraryInfo *TLI, AliasAnalysis *AA, MemorySSA *MSSA,
638 const DataLayout &DL)
639 : Opts(ScalarOptions::Global), F(F), DT(DT), TLI(TLI), AA(AA), MSSA(MSSA),
640 AC(AC), DL(DL),
641 // Reuse ExpressionAllocator for PredicateInfo as well.
642 PredInfo(
643 std::make_unique<PredicateInfo>(args&: F, args&: *DT, args&: *AC, args&: ExpressionAllocator)),
644 SQ(DL, TLI, DT, AC, /*CtxI=*/nullptr, /*UseInstrInfo=*/false,
645 /*CanUseUndef=*/false) {}
646
647 bool runGVN();
648
649private:
650 /// Helper struct return a Expression with an optional extra dependency.
651 struct ExprResult {
652 const Expression *Expr;
653 Value *ExtraDep;
654 const PredicateBase *PredDep;
655
656 ExprResult(const Expression *Expr, Value *ExtraDep = nullptr,
657 const PredicateBase *PredDep = nullptr)
658 : Expr(Expr), ExtraDep(ExtraDep), PredDep(PredDep) {}
659 ExprResult(const ExprResult &) = delete;
660 ExprResult(ExprResult &&Other)
661 : Expr(Other.Expr), ExtraDep(Other.ExtraDep), PredDep(Other.PredDep) {
662 Other.Expr = nullptr;
663 Other.ExtraDep = nullptr;
664 Other.PredDep = nullptr;
665 }
666 ExprResult &operator=(const ExprResult &Other) = delete;
667 ExprResult &operator=(ExprResult &&Other) = delete;
668
669 ~ExprResult() { assert(!ExtraDep && "unhandled ExtraDep"); }
670
671 operator bool() const { return Expr; }
672
673 static ExprResult none() { return {nullptr, nullptr, nullptr}; }
674 static ExprResult some(const Expression *Expr, Value *ExtraDep = nullptr) {
675 return {Expr, ExtraDep, nullptr};
676 }
677 static ExprResult some(const Expression *Expr,
678 const PredicateBase *PredDep) {
679 return {Expr, nullptr, PredDep};
680 }
681 static ExprResult some(const Expression *Expr, Value *ExtraDep,
682 const PredicateBase *PredDep) {
683 return {Expr, ExtraDep, PredDep};
684 }
685 };
686
687 // Expression handling.
688 ExprResult createExpression(Instruction *) const;
689 const Expression *createBinaryExpression(unsigned, Type *, Value *, Value *,
690 Instruction *) const;
691
692 // Our canonical form for phi arguments is a pair of incoming value, incoming
693 // basic block.
694 using ValPair = std::pair<Value *, BasicBlock *>;
695
696 PHIExpression *createPHIExpression(ArrayRef<ValPair>, const Instruction *,
697 BasicBlock *, bool &HasBackEdge,
698 bool &OriginalOpsConstant) const;
699 const DeadExpression *createDeadExpression() const;
700 const VariableExpression *createVariableExpression(Value *) const;
701 const ConstantExpression *createConstantExpression(Constant *) const;
702 const Expression *createVariableOrConstant(Value *V) const;
703 const UnknownExpression *createUnknownExpression(Instruction *) const;
704 const StoreExpression *createStoreExpression(StoreInst *,
705 const MemoryAccess *) const;
706 LoadExpression *createLoadExpression(Type *, Value *, LoadInst *,
707 const MemoryAccess *) const;
708 const CallExpression *createCallExpression(CallInst *,
709 const MemoryAccess *) const;
710 const AggregateValueExpression *
711 createAggregateValueExpression(Instruction *) const;
712 bool setBasicExpressionInfo(Instruction *, BasicExpression *) const;
713
714 // Congruence class handling.
715 CongruenceClass *createCongruenceClass(Value *Leader, const Expression *E) {
716 // Set RPO to 0 for values that are always available (constants and function
717 // args). These should always be made leader.
718 unsigned LeaderDFS = 0;
719
720 // If Leader is not specified, either we have a memory class or the leader
721 // will be set later. Otherwise, if Leader is an Instruction, set LeaderDFS
722 // to its RPO number.
723 if (!Leader)
724 LeaderDFS = ~0;
725 else if (auto *I = dyn_cast<Instruction>(Val: Leader))
726 LeaderDFS = InstrToDFSNum(V: I);
727 auto *result =
728 new CongruenceClass(NextCongruenceNum++, {Leader, LeaderDFS}, E);
729 CongruenceClasses.emplace_back(args&: result);
730 return result;
731 }
732
733 CongruenceClass *createMemoryClass(MemoryAccess *MA) {
734 auto *CC = createCongruenceClass(Leader: nullptr, E: nullptr);
735 CC->setMemoryLeader(MA);
736 return CC;
737 }
738
739 CongruenceClass *ensureLeaderOfMemoryClass(MemoryAccess *MA) {
740 auto *CC = getMemoryClass(MA);
741 if (CC->getMemoryLeader() != MA)
742 CC = createMemoryClass(MA);
743 return CC;
744 }
745
746 CongruenceClass *createSingletonCongruenceClass(Value *Member) {
747 CongruenceClass *CClass = createCongruenceClass(Leader: Member, E: nullptr);
748 CClass->insert(M: Member);
749 ValueToClass[Member] = CClass;
750 return CClass;
751 }
752
753 void initializeCongruenceClasses(Function &F);
754 const Expression *makePossiblePHIOfOps(Instruction *,
755 SmallPtrSetImpl<Value *> &);
756 Value *findLeaderForInst(Instruction *ValueOp,
757 SmallPtrSetImpl<Value *> &Visited,
758 MemoryAccess *MemAccess, Instruction *OrigInst,
759 BasicBlock *PredBB);
760 bool OpIsSafeForPHIOfOps(Value *Op, const BasicBlock *PHIBlock,
761 SmallPtrSetImpl<const Value *> &);
762 void addPhiOfOps(PHINode *Op, BasicBlock *BB, Instruction *ExistingValue);
763 void removePhiOfOps(Instruction *I, PHINode *PHITemp);
764
765 // Value number an Instruction or MemoryPhi.
766 void valueNumberMemoryPhi(MemoryPhi *);
767 void valueNumberInstruction(Instruction *);
768
769 // Symbolic evaluation.
770 ExprResult checkExprResults(Expression *, Instruction *, Value *) const;
771 ExprResult performSymbolicEvaluation(Instruction *,
772 SmallPtrSetImpl<Value *> &) const;
773 const Expression *performSymbolicLoadCoercion(Type *, Value *, LoadInst *,
774 Instruction *,
775 MemoryAccess *) const;
776 const Expression *performSymbolicLoadEvaluation(Instruction *) const;
777 const Expression *performSymbolicStoreEvaluation(Instruction *) const;
778 ExprResult performSymbolicCallEvaluation(Instruction *) const;
779 void sortPHIOps(MutableArrayRef<ValPair> Ops) const;
780 const Expression *performSymbolicPHIEvaluation(ArrayRef<ValPair>,
781 Instruction *I,
782 BasicBlock *PHIBlock) const;
783 const Expression *performSymbolicAggrValueEvaluation(Instruction *) const;
784 ExprResult performSymbolicCmpEvaluation(Instruction *) const;
785 ExprResult performSymbolicPredicateInfoEvaluation(BitCastInst *) const;
786
787 // Congruence finding.
788 bool someEquivalentDominates(const Instruction *, const Instruction *) const;
789 Value *lookupOperandLeader(Value *) const;
790 CongruenceClass *getClassForExpression(const Expression *E) const;
791 void performCongruenceFinding(Instruction *, const Expression *);
792 void moveValueToNewCongruenceClass(Instruction *, const Expression *,
793 CongruenceClass *, CongruenceClass *);
794 void moveMemoryToNewCongruenceClass(Instruction *, MemoryAccess *,
795 CongruenceClass *, CongruenceClass *);
796 Value *getNextValueLeader(CongruenceClass *) const;
797 const MemoryAccess *getNextMemoryLeader(CongruenceClass *) const;
798 bool setMemoryClass(const MemoryAccess *From, CongruenceClass *To);
799 CongruenceClass *getMemoryClass(const MemoryAccess *MA) const;
800 const MemoryAccess *lookupMemoryLeader(const MemoryAccess *) const;
801 bool isMemoryAccessTOP(const MemoryAccess *) const;
802
803 // Ranking
804 unsigned int getRank(const Value *) const;
805 bool shouldSwapOperands(const Value *, const Value *) const;
806 bool shouldSwapOperandsForPredicate(const Value *, const Value *,
807 const BitCastInst *I) const;
808
809 // Reachability handling.
810 void updateReachableEdge(BasicBlock *, BasicBlock *);
811 void processOutgoingEdges(Instruction *, BasicBlock *);
812 Value *findConditionEquivalence(Value *) const;
813
814 // Elimination.
815 struct ValueDFS;
816 void convertClassToDFSOrdered(const CongruenceClass &,
817 SmallVectorImpl<ValueDFS> &,
818 DenseMap<const Value *, unsigned int> &,
819 SmallPtrSetImpl<Instruction *> &) const;
820 void convertClassToLoadsAndStores(const CongruenceClass &,
821 SmallVectorImpl<ValueDFS> &) const;
822
823 bool eliminateInstructions(Function &);
824 void replaceInstruction(Instruction *, Value *);
825 void markInstructionForDeletion(Instruction *);
826 void deleteInstructionsInBlock(BasicBlock *);
827 Value *findPHIOfOpsLeader(const Expression *, const Instruction *,
828 const BasicBlock *) const;
829
830 // Various instruction touch utilities
831 template <typename Map, typename KeyType>
832 void touchAndErase(Map &, const KeyType &);
833 void markUsersTouched(Value *);
834 void markMemoryUsersTouched(const MemoryAccess *);
835 void markMemoryDefTouched(const MemoryAccess *);
836 void markPredicateUsersTouched(Instruction *);
837 void markValueLeaderChangeTouched(CongruenceClass *CC);
838 void markMemoryLeaderChangeTouched(CongruenceClass *CC);
839 void markPhiOfOpsChanged(const Expression *E);
840 void addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const;
841 void addAdditionalUsers(Value *To, Value *User) const;
842 void addAdditionalUsers(ExprResult &Res, Instruction *User) const;
843
844 // Main loop of value numbering
845 void iterateTouchedInstructions();
846
847 // Utilities.
848 void cleanupTables();
849 std::pair<unsigned, unsigned> assignDFSNumbers(BasicBlock *, unsigned);
850 void updateProcessedCount(const Value *V);
851 void verifyMemoryCongruency() const;
852 void verifyIterationSettled(Function &F);
853 void verifyStoreExpressions() const;
854 bool singleReachablePHIPath(SmallPtrSet<const MemoryAccess *, 8> &,
855 const MemoryAccess *, const MemoryAccess *) const;
856 BasicBlock *getBlockForValue(Value *V) const;
857 void deleteExpression(const Expression *E) const;
858 MemoryUseOrDef *getMemoryAccess(const Instruction *) const;
859 MemoryPhi *getMemoryAccess(const BasicBlock *) const;
860 template <class T, class Range> T *getMinDFSOfRange(const Range &) const;
861
862 unsigned InstrToDFSNum(const Value *V) const {
863 assert(isa<Instruction>(V) && "This should not be used for MemoryAccesses");
864 return InstrDFS.lookup(Val: V);
865 }
866
867 unsigned InstrToDFSNum(const MemoryAccess *MA) const {
868 return MemoryToDFSNum(MA);
869 }
870
871 Value *InstrFromDFSNum(unsigned DFSNum) { return DFSToInstr[DFSNum]; }
872
873 // Given a MemoryAccess, return the relevant instruction DFS number. Note:
874 // This deliberately takes a value so it can be used with Use's, which will
875 // auto-convert to Value's but not to MemoryAccess's.
876 unsigned MemoryToDFSNum(const Value *MA) const {
877 assert(isa<MemoryAccess>(MA) &&
878 "This should not be used with instructions");
879 return isa<MemoryUseOrDef>(Val: MA)
880 ? InstrToDFSNum(V: cast<MemoryUseOrDef>(Val: MA)->getMemoryInst())
881 : InstrDFS.lookup(Val: MA);
882 }
883
884 bool isCycleFree(const Instruction *) const;
885 bool isBackedge(BasicBlock *From, BasicBlock *To) const;
886
887 // Debug counter info. When verifying, we have to reset the value numbering
888 // debug counter to the same state it started in to get the same results.
889 DebugCounter::CounterState StartingVNCounter;
890};
891
892} // end anonymous namespace
893
894template <typename T>
895static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS) {
896 if (!isa<LoadExpression>(Val: RHS) && !isa<StoreExpression>(Val: RHS))
897 return false;
898 return LHS.MemoryExpression::equals(RHS);
899}
900
901bool LoadExpression::equals(const Expression &Other) const {
902 return equalsLoadStoreHelper(LHS: *this, RHS: Other);
903}
904
905bool StoreExpression::equals(const Expression &Other) const {
906 if (!equalsLoadStoreHelper(LHS: *this, RHS: Other))
907 return false;
908 // Make sure that store vs store includes the value operand.
909 if (const auto *S = dyn_cast<StoreExpression>(Val: &Other))
910 if (getStoredValue() != S->getStoredValue())
911 return false;
912 return true;
913}
914
915bool CallExpression::equals(const Expression &Other) const {
916 if (!MemoryExpression::equals(Other))
917 return false;
918
919 if (auto *RHS = dyn_cast<CallExpression>(Val: &Other))
920 return Call->getAttributes()
921 .intersectWith(C&: Call->getContext(), Other: RHS->Call->getAttributes())
922 .has_value();
923
924 return false;
925}
926
927// Determine if the edge From->To is a backedge
928bool NewGVN::isBackedge(BasicBlock *From, BasicBlock *To) const {
929 return From == To ||
930 RPOOrdering.lookup(Val: DT->getNode(BB: From)) >=
931 RPOOrdering.lookup(Val: DT->getNode(BB: To));
932}
933
934#ifndef NDEBUG
935static std::string getBlockName(const BasicBlock *B) {
936 return DOTGraphTraits<DOTFuncInfo *>::getSimpleNodeLabel(B, nullptr);
937}
938#endif
939
940// Get a MemoryAccess for an instruction, fake or real.
941MemoryUseOrDef *NewGVN::getMemoryAccess(const Instruction *I) const {
942 auto *Result = MSSA->getMemoryAccess(I);
943 return Result ? Result : TempToMemory.lookup(Val: I);
944}
945
946// Get a MemoryPhi for a basic block. These are all real.
947MemoryPhi *NewGVN::getMemoryAccess(const BasicBlock *BB) const {
948 return MSSA->getMemoryAccess(BB);
949}
950
951// Get the basic block from an instruction/memory value.
952BasicBlock *NewGVN::getBlockForValue(Value *V) const {
953 if (auto *I = dyn_cast<Instruction>(Val: V)) {
954 auto *Parent = I->getParent();
955 if (Parent)
956 return Parent;
957 Parent = TempToBlock.lookup(Val: V);
958 assert(Parent && "Every fake instruction should have a block");
959 return Parent;
960 }
961
962 auto *MP = dyn_cast<MemoryPhi>(Val: V);
963 assert(MP && "Should have been an instruction or a MemoryPhi");
964 return MP->getBlock();
965}
966
967// Delete a definitely dead expression, so it can be reused by the expression
968// allocator. Some of these are not in creation functions, so we have to accept
969// const versions.
970void NewGVN::deleteExpression(const Expression *E) const {
971 assert(isa<BasicExpression>(E));
972 auto *BE = cast<BasicExpression>(Val: E);
973 const_cast<BasicExpression *>(BE)->deallocateOperands(Recycler&: ArgRecycler);
974 ExpressionAllocator.Deallocate(Ptr: E);
975}
976
977// If V is a predicateinfo copy, get the thing it is a copy of.
978static Value *getCopyOf(const Value *V) {
979 if (auto *BC = dyn_cast<BitCastInst>(Val: V))
980 if (BC->getType() == BC->getOperand(i_nocapture: 0)->getType())
981 return BC->getOperand(i_nocapture: 0);
982 return nullptr;
983}
984
985// Return true if V is really PN, even accounting for predicateinfo copies.
986static bool isCopyOfPHI(const Value *V, const PHINode *PN) {
987 return V == PN || getCopyOf(V) == PN;
988}
989
990static bool isCopyOfAPHI(const Value *V) {
991 auto *CO = getCopyOf(V);
992 return CO && isa<PHINode>(Val: CO);
993}
994
995// Sort PHI Operands into a canonical order. What we use here is an RPO
996// order. The BlockInstRange numbers are generated in an RPO walk of the basic
997// blocks.
998void NewGVN::sortPHIOps(MutableArrayRef<ValPair> Ops) const {
999 llvm::sort(C&: Ops, Comp: [&](const ValPair &P1, const ValPair &P2) {
1000 return BlockInstRange.lookup(Val: P1.second).first <
1001 BlockInstRange.lookup(Val: P2.second).first;
1002 });
1003}
1004
1005// Return true if V is a value that will always be available (IE can
1006// be placed anywhere) in the function. We don't do globals here
1007// because they are often worse to put in place.
1008static bool alwaysAvailable(Value *V) {
1009 return isa<Constant>(Val: V) || isa<Argument>(Val: V);
1010}
1011
1012// Create a PHIExpression from an array of {incoming edge, value} pairs. I is
1013// the original instruction we are creating a PHIExpression for (but may not be
1014// a phi node). We require, as an invariant, that all the PHIOperands in the
1015// same block are sorted the same way. sortPHIOps will sort them into a
1016// canonical order.
1017PHIExpression *NewGVN::createPHIExpression(ArrayRef<ValPair> PHIOperands,
1018 const Instruction *I,
1019 BasicBlock *PHIBlock,
1020 bool &HasBackedge,
1021 bool &OriginalOpsConstant) const {
1022 unsigned NumOps = PHIOperands.size();
1023 auto *E = new (ExpressionAllocator) PHIExpression(NumOps, PHIBlock);
1024
1025 E->allocateOperands(Recycler&: ArgRecycler, Allocator&: ExpressionAllocator);
1026 E->setType(PHIOperands.begin()->first->getType());
1027 E->setOpcode(Instruction::PHI);
1028
1029 // Filter out unreachable phi operands.
1030 auto Filtered = make_filter_range(Range&: PHIOperands, Pred: [&](const ValPair &P) {
1031 auto *BB = P.second;
1032 if (auto *PHIOp = dyn_cast<PHINode>(Val: I))
1033 if (isCopyOfPHI(V: P.first, PN: PHIOp))
1034 return false;
1035 if (!ReachableEdges.count(V: {BB, PHIBlock}))
1036 return false;
1037 // Things in TOPClass are equivalent to everything.
1038 if (ValueToClass.lookup(Val: P.first) == TOPClass)
1039 return false;
1040 OriginalOpsConstant = OriginalOpsConstant && isa<Constant>(Val: P.first);
1041 HasBackedge = HasBackedge || isBackedge(From: BB, To: PHIBlock);
1042 return lookupOperandLeader(P.first) != I;
1043 });
1044 llvm::transform(Range&: Filtered, d_first: op_inserter(E), F: [&](const ValPair &P) -> Value * {
1045 return lookupOperandLeader(P.first);
1046 });
1047 return E;
1048}
1049
1050// Set basic expression info (Arguments, type, opcode) for Expression
1051// E from Instruction I in block B.
1052bool NewGVN::setBasicExpressionInfo(Instruction *I, BasicExpression *E) const {
1053 bool AllConstant = true;
1054 if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: I))
1055 E->setType(GEP->getSourceElementType());
1056 else
1057 E->setType(I->getType());
1058 E->setOpcode(I->getOpcode());
1059 E->allocateOperands(Recycler&: ArgRecycler, Allocator&: ExpressionAllocator);
1060
1061 // Transform the operand array into an operand leader array, and keep track of
1062 // whether all members are constant.
1063 std::transform(first: I->op_begin(), last: I->op_end(), result: op_inserter(E), unary_op: [&](Value *O) {
1064 auto Operand = lookupOperandLeader(O);
1065 AllConstant = AllConstant && isa<Constant>(Val: Operand);
1066 return Operand;
1067 });
1068
1069 return AllConstant;
1070}
1071
1072const Expression *NewGVN::createBinaryExpression(unsigned Opcode, Type *T,
1073 Value *Arg1, Value *Arg2,
1074 Instruction *I) const {
1075 auto *E = new (ExpressionAllocator) BasicExpression(2);
1076 // TODO: we need to remove context instruction after Value Tracking
1077 // can run without context instruction
1078 const SimplifyQuery Q = SQ.getWithInstruction(I);
1079
1080 E->setType(T);
1081 E->setOpcode(Opcode);
1082 E->allocateOperands(Recycler&: ArgRecycler, Allocator&: ExpressionAllocator);
1083 if (Instruction::isCommutative(Opcode)) {
1084 // Ensure that commutative instructions that only differ by a permutation
1085 // of their operands get the same value number by sorting the operand value
1086 // numbers. Since all commutative instructions have two operands it is more
1087 // efficient to sort by hand rather than using, say, std::sort.
1088 if (shouldSwapOperands(Arg1, Arg2))
1089 std::swap(a&: Arg1, b&: Arg2);
1090 }
1091 E->op_push_back(Arg: lookupOperandLeader(Arg1));
1092 E->op_push_back(Arg: lookupOperandLeader(Arg2));
1093
1094 Value *V = simplifyBinOp(Opcode, LHS: E->getOperand(N: 0), RHS: E->getOperand(N: 1), Q);
1095 if (auto Simplified = checkExprResults(E, I, V)) {
1096 addAdditionalUsers(Res&: Simplified, User: I);
1097 return Simplified.Expr;
1098 }
1099 return E;
1100}
1101
1102// Take a Value returned by simplification of Expression E/Instruction
1103// I, and see if it resulted in a simpler expression. If so, return
1104// that expression.
1105NewGVN::ExprResult NewGVN::checkExprResults(Expression *E, Instruction *I,
1106 Value *V) const {
1107 if (!V)
1108 return ExprResult::none();
1109
1110 if (auto *C = dyn_cast<Constant>(Val: V)) {
1111 if (I)
1112 LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1113 << " constant " << *C << "\n");
1114 NumGVNOpsSimplified++;
1115 assert(isa<BasicExpression>(E) &&
1116 "We should always have had a basic expression here");
1117 deleteExpression(E);
1118 return ExprResult::some(Expr: createConstantExpression(C));
1119 } else if (isa<Argument>(Val: V) || isa<GlobalVariable>(Val: V)) {
1120 if (I)
1121 LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1122 << " variable " << *V << "\n");
1123 deleteExpression(E);
1124 return ExprResult::some(Expr: createVariableExpression(V));
1125 }
1126
1127 CongruenceClass *CC = ValueToClass.lookup(Val: V);
1128 if (CC) {
1129 if (CC->getLeader() && CC->getLeader() != I) {
1130 return ExprResult::some(Expr: createVariableOrConstant(V: CC->getLeader()), ExtraDep: V);
1131 }
1132 if (CC->getDefiningExpr()) {
1133 if (I)
1134 LLVM_DEBUG(dbgs() << "Simplified " << *I << " to "
1135 << " expression " << *CC->getDefiningExpr() << "\n");
1136 NumGVNOpsSimplified++;
1137 deleteExpression(E);
1138 return ExprResult::some(Expr: CC->getDefiningExpr(), ExtraDep: V);
1139 }
1140 }
1141
1142 return ExprResult::none();
1143}
1144
1145// Create a value expression from the instruction I, replacing operands with
1146// their leaders.
1147
1148NewGVN::ExprResult NewGVN::createExpression(Instruction *I) const {
1149 auto *E = new (ExpressionAllocator) BasicExpression(I->getNumOperands());
1150 // TODO: we need to remove context instruction after Value Tracking
1151 // can run without context instruction
1152 const SimplifyQuery Q = SQ.getWithInstruction(I);
1153
1154 bool AllConstant = setBasicExpressionInfo(I, E);
1155
1156 if (I->isCommutative()) {
1157 // Ensure that commutative instructions that only differ by a permutation
1158 // of their operands get the same value number by sorting the operand value
1159 // numbers. Since all commutative instructions have two operands it is more
1160 // efficient to sort by hand rather than using, say, std::sort.
1161 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
1162 if (shouldSwapOperands(E->getOperand(N: 0), E->getOperand(N: 1)))
1163 E->swapOperands(First: 0, Second: 1);
1164 }
1165 // Perform simplification.
1166 if (auto *CI = dyn_cast<CmpInst>(Val: I)) {
1167 // Sort the operand value numbers so x<y and y>x get the same value
1168 // number.
1169 CmpInst::Predicate Predicate = CI->getPredicate();
1170 if (shouldSwapOperands(E->getOperand(N: 0), E->getOperand(N: 1))) {
1171 E->swapOperands(First: 0, Second: 1);
1172 Predicate = CmpInst::getSwappedPredicate(pred: Predicate);
1173 }
1174 E->setOpcode((CI->getOpcode() << 8) | Predicate);
1175 // TODO: 25% of our time is spent in simplifyCmpInst with pointer operands
1176 assert(I->getOperand(0)->getType() == I->getOperand(1)->getType() &&
1177 "Wrong types on cmp instruction");
1178 assert((E->getOperand(0)->getType() == I->getOperand(0)->getType() &&
1179 E->getOperand(1)->getType() == I->getOperand(1)->getType()));
1180 Value *V =
1181 simplifyCmpInst(Predicate, LHS: E->getOperand(N: 0), RHS: E->getOperand(N: 1), Q);
1182 if (auto Simplified = checkExprResults(E, I, V))
1183 return Simplified;
1184 } else if (isa<SelectInst>(Val: I)) {
1185 if (isa<Constant>(Val: E->getOperand(N: 0)) ||
1186 E->getOperand(N: 1) == E->getOperand(N: 2)) {
1187 assert(E->getOperand(1)->getType() == I->getOperand(1)->getType() &&
1188 E->getOperand(2)->getType() == I->getOperand(2)->getType());
1189 Value *V = simplifySelectInst(Cond: E->getOperand(N: 0), TrueVal: E->getOperand(N: 1),
1190 FalseVal: E->getOperand(N: 2), FMF: FastMathFlags(), Q);
1191 if (auto Simplified = checkExprResults(E, I, V))
1192 return Simplified;
1193 }
1194 } else if (I->isBinaryOp()) {
1195 Value *V =
1196 simplifyBinOp(Opcode: E->getOpcode(), LHS: E->getOperand(N: 0), RHS: E->getOperand(N: 1), Q);
1197 if (auto Simplified = checkExprResults(E, I, V))
1198 return Simplified;
1199 } else if (auto *CI = dyn_cast<CastInst>(Val: I)) {
1200 Value *V =
1201 simplifyCastInst(CastOpc: CI->getOpcode(), Op: E->getOperand(N: 0), Ty: CI->getType(), Q);
1202 if (auto Simplified = checkExprResults(E, I, V))
1203 return Simplified;
1204 } else if (auto *GEPI = dyn_cast<GetElementPtrInst>(Val: I)) {
1205 Value *V = simplifyGEPInst(SrcTy: GEPI->getSourceElementType(), Ptr: *E->op_begin(),
1206 Indices: ArrayRef(std::next(x: E->op_begin()), E->op_end()),
1207 NW: GEPI->getNoWrapFlags(), Q);
1208 if (auto Simplified = checkExprResults(E, I, V))
1209 return Simplified;
1210 } else if (AllConstant) {
1211 // We don't bother trying to simplify unless all of the operands
1212 // were constant.
1213 // TODO: There are a lot of Simplify*'s we could call here, if we
1214 // wanted to. The original motivating case for this code was a
1215 // zext i1 false to i8, which we don't have an interface to
1216 // simplify (IE there is no SimplifyZExt).
1217
1218 SmallVector<Constant *, 8> C;
1219 for (Value *Arg : E->operands())
1220 C.emplace_back(Args: cast<Constant>(Val: Arg));
1221
1222 if (Value *V = ConstantFoldInstOperands(I, Ops: C, DL, TLI))
1223 if (auto Simplified = checkExprResults(E, I, V))
1224 return Simplified;
1225 }
1226 return ExprResult::some(Expr: E);
1227}
1228
1229const AggregateValueExpression *
1230NewGVN::createAggregateValueExpression(Instruction *I) const {
1231 if (auto *II = dyn_cast<InsertValueInst>(Val: I)) {
1232 auto *E = new (ExpressionAllocator)
1233 AggregateValueExpression(I->getNumOperands(), II->getNumIndices());
1234 setBasicExpressionInfo(I, E);
1235 E->allocateIntOperands(Allocator&: ExpressionAllocator);
1236 llvm::copy(Range: II->indices(), Out: int_op_inserter(E));
1237 return E;
1238 } else if (auto *EI = dyn_cast<ExtractValueInst>(Val: I)) {
1239 auto *E = new (ExpressionAllocator)
1240 AggregateValueExpression(I->getNumOperands(), EI->getNumIndices());
1241 setBasicExpressionInfo(I: EI, E);
1242 E->allocateIntOperands(Allocator&: ExpressionAllocator);
1243 llvm::copy(Range: EI->indices(), Out: int_op_inserter(E));
1244 return E;
1245 }
1246 llvm_unreachable("Unhandled type of aggregate value operation");
1247}
1248
1249const DeadExpression *NewGVN::createDeadExpression() const {
1250 // DeadExpression has no arguments and all DeadExpression's are the same,
1251 // so we only need one of them.
1252 return SingletonDeadExpression;
1253}
1254
1255const VariableExpression *NewGVN::createVariableExpression(Value *V) const {
1256 auto *E = new (ExpressionAllocator) VariableExpression(V);
1257 E->setOpcode(V->getValueID());
1258 return E;
1259}
1260
1261const Expression *NewGVN::createVariableOrConstant(Value *V) const {
1262 if (auto *C = dyn_cast<Constant>(Val: V))
1263 return createConstantExpression(C);
1264 return createVariableExpression(V);
1265}
1266
1267const ConstantExpression *NewGVN::createConstantExpression(Constant *C) const {
1268 auto *E = new (ExpressionAllocator) ConstantExpression(C);
1269 E->setOpcode(C->getValueID());
1270 return E;
1271}
1272
1273const UnknownExpression *NewGVN::createUnknownExpression(Instruction *I) const {
1274 auto *E = new (ExpressionAllocator) UnknownExpression(I);
1275 E->setOpcode(I->getOpcode());
1276 return E;
1277}
1278
1279const CallExpression *
1280NewGVN::createCallExpression(CallInst *CI, const MemoryAccess *MA) const {
1281 // FIXME: Add operand bundles for calls.
1282 auto *E =
1283 new (ExpressionAllocator) CallExpression(CI->getNumOperands(), CI, MA);
1284 setBasicExpressionInfo(I: CI, E);
1285 if (CI->isCommutative()) {
1286 // Ensure that commutative intrinsics that only differ by a permutation
1287 // of their operands get the same value number by sorting the operand value
1288 // numbers.
1289 assert(CI->getNumOperands() >= 2 && "Unsupported commutative intrinsic!");
1290 if (shouldSwapOperands(E->getOperand(N: 0), E->getOperand(N: 1)))
1291 E->swapOperands(First: 0, Second: 1);
1292 }
1293 return E;
1294}
1295
1296// Return true if some equivalent of instruction Inst dominates instruction U.
1297bool NewGVN::someEquivalentDominates(const Instruction *Inst,
1298 const Instruction *U) const {
1299 auto *CC = ValueToClass.lookup(Val: Inst);
1300 // This must be an instruction because we are only called from phi nodes
1301 // in the case that the value it needs to check against is an instruction.
1302
1303 // The most likely candidates for dominance are the leader and the next leader.
1304 // The leader or nextleader will dominate in all cases where there is an
1305 // equivalent that is higher up in the dom tree.
1306 // We can't *only* check them, however, because the
1307 // dominator tree could have an infinite number of non-dominating siblings
1308 // with instructions that are in the right congruence class.
1309 // A
1310 // B C D E F G
1311 // |
1312 // H
1313 // Instruction U could be in H, with equivalents in every other sibling.
1314 // Depending on the rpo order picked, the leader could be the equivalent in
1315 // any of these siblings.
1316 if (!CC)
1317 return false;
1318 if (alwaysAvailable(V: CC->getLeader()))
1319 return true;
1320 if (DT->dominates(Def: cast<Instruction>(Val: CC->getLeader()), User: U))
1321 return true;
1322 if (CC->getNextLeader().first &&
1323 DT->dominates(Def: cast<Instruction>(Val: CC->getNextLeader().first), User: U))
1324 return true;
1325 return llvm::any_of(Range&: *CC, P: [&](const Value *Member) {
1326 return Member != CC->getLeader() &&
1327 DT->dominates(Def: cast<Instruction>(Val: Member), User: U);
1328 });
1329}
1330
1331// See if we have a congruence class and leader for this operand, and if so,
1332// return it. Otherwise, return the operand itself.
1333Value *NewGVN::lookupOperandLeader(Value *V) const {
1334 CongruenceClass *CC = ValueToClass.lookup(Val: V);
1335 if (CC) {
1336 // Everything in TOP is represented by poison, as it can be any value.
1337 // We do have to make sure we get the type right though, so we can't set the
1338 // RepLeader to poison.
1339 if (CC == TOPClass)
1340 return PoisonValue::get(T: V->getType());
1341 return CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
1342 }
1343
1344 return V;
1345}
1346
1347const MemoryAccess *NewGVN::lookupMemoryLeader(const MemoryAccess *MA) const {
1348 auto *CC = getMemoryClass(MA);
1349 assert(CC->getMemoryLeader() &&
1350 "Every MemoryAccess should be mapped to a congruence class with a "
1351 "representative memory access");
1352 return CC->getMemoryLeader();
1353}
1354
1355// Return true if the MemoryAccess is really equivalent to everything. This is
1356// equivalent to the lattice value "TOP" in most lattices. This is the initial
1357// state of all MemoryAccesses.
1358bool NewGVN::isMemoryAccessTOP(const MemoryAccess *MA) const {
1359 return getMemoryClass(MA) == TOPClass;
1360}
1361
1362LoadExpression *NewGVN::createLoadExpression(Type *LoadType, Value *PointerOp,
1363 LoadInst *LI,
1364 const MemoryAccess *MA) const {
1365 auto *E =
1366 new (ExpressionAllocator) LoadExpression(1, LI, lookupMemoryLeader(MA));
1367 E->allocateOperands(Recycler&: ArgRecycler, Allocator&: ExpressionAllocator);
1368 E->setType(LoadType);
1369
1370 // Give store and loads same opcode so they value number together.
1371 E->setOpcode(0);
1372 E->op_push_back(Arg: PointerOp);
1373
1374 // TODO: Value number heap versions. We may be able to discover
1375 // things alias analysis can't on it's own (IE that a store and a
1376 // load have the same value, and thus, it isn't clobbering the load).
1377 return E;
1378}
1379
1380const StoreExpression *
1381NewGVN::createStoreExpression(StoreInst *SI, const MemoryAccess *MA) const {
1382 auto *StoredValueLeader = lookupOperandLeader(V: SI->getValueOperand());
1383 auto *E = new (ExpressionAllocator)
1384 StoreExpression(SI->getNumOperands(), SI, StoredValueLeader, MA);
1385 E->allocateOperands(Recycler&: ArgRecycler, Allocator&: ExpressionAllocator);
1386 E->setType(SI->getValueOperand()->getType());
1387
1388 // Give store and loads same opcode so they value number together.
1389 E->setOpcode(0);
1390 E->op_push_back(Arg: lookupOperandLeader(V: SI->getPointerOperand()));
1391
1392 // TODO: Value number heap versions. We may be able to discover
1393 // things alias analysis can't on it's own (IE that a store and a
1394 // load have the same value, and thus, it isn't clobbering the load).
1395 return E;
1396}
1397
1398const Expression *NewGVN::performSymbolicStoreEvaluation(Instruction *I) const {
1399 // Unlike loads, we never try to eliminate stores, so we do not check if they
1400 // are simple and avoid value numbering them.
1401 auto *SI = cast<StoreInst>(Val: I);
1402 auto *StoreAccess = getMemoryAccess(I: SI);
1403 // Get the expression, if any, for the RHS of the MemoryDef.
1404 const MemoryAccess *StoreRHS = StoreAccess->getDefiningAccess();
1405 if (Opts.enable_store_refinement)
1406 StoreRHS = MSSAWalker->getClobberingMemoryAccess(MA: StoreAccess);
1407 // If we bypassed the use-def chains, make sure we add a use.
1408 StoreRHS = lookupMemoryLeader(MA: StoreRHS);
1409 if (StoreRHS != StoreAccess->getDefiningAccess())
1410 addMemoryUsers(To: StoreRHS, U: StoreAccess);
1411 // If we are defined by ourselves, use the live on entry def.
1412 if (StoreRHS == StoreAccess)
1413 StoreRHS = MSSA->getLiveOnEntryDef();
1414
1415 if (SI->isSimple()) {
1416 // See if we are defined by a previous store expression, it already has a
1417 // value, and it's the same value as our current store. FIXME: Right now, we
1418 // only do this for simple stores, we should expand to cover memcpys, etc.
1419 const auto *LastStore = createStoreExpression(SI, MA: StoreRHS);
1420 const auto *LastCC = ExpressionToClass.lookup(Val: LastStore);
1421 // We really want to check whether the expression we matched was a store. No
1422 // easy way to do that. However, we can check that the class we found has a
1423 // store, which, assuming the value numbering state is not corrupt, is
1424 // sufficient, because we must also be equivalent to that store's expression
1425 // for it to be in the same class as the load.
1426 if (LastCC && LastCC->getStoredValue() == LastStore->getStoredValue())
1427 return LastStore;
1428 // Also check if our value operand is defined by a load of the same memory
1429 // location, and the memory state is the same as it was then (otherwise, it
1430 // could have been overwritten later. See test32 in
1431 // transforms/DeadStoreElimination/simple.ll).
1432 if (auto *LI = dyn_cast<LoadInst>(Val: LastStore->getStoredValue()))
1433 if ((lookupOperandLeader(V: LI->getPointerOperand()) ==
1434 LastStore->getOperand(N: 0)) &&
1435 (lookupMemoryLeader(MA: getMemoryAccess(I: LI)->getDefiningAccess()) ==
1436 StoreRHS))
1437 return LastStore;
1438 deleteExpression(E: LastStore);
1439 }
1440
1441 // If the store is not equivalent to anything, value number it as a store that
1442 // produces a unique memory state (instead of using it's MemoryUse, we use
1443 // it's MemoryDef).
1444 return createStoreExpression(SI, MA: StoreAccess);
1445}
1446
1447// See if we can extract the value of a loaded pointer from a load, a store, or
1448// a memory instruction.
1449const Expression *
1450NewGVN::performSymbolicLoadCoercion(Type *LoadType, Value *LoadPtr,
1451 LoadInst *LI, Instruction *DepInst,
1452 MemoryAccess *DefiningAccess) const {
1453 assert((!LI || LI->isSimple()) && "Not a simple load");
1454 if (auto *DepSI = dyn_cast<StoreInst>(Val: DepInst)) {
1455 // Can't forward from non-atomic to atomic without violating memory model.
1456 // Also don't need to coerce if they are the same type, we will just
1457 // propagate.
1458 if (LI->isAtomic() > DepSI->isAtomic() ||
1459 LoadType == DepSI->getValueOperand()->getType())
1460 return nullptr;
1461 int Offset = analyzeLoadFromClobberingStore(LoadTy: LoadType, LoadPtr, DepSI, DL);
1462 if (Offset >= 0) {
1463 if (auto *C = dyn_cast<Constant>(
1464 Val: lookupOperandLeader(V: DepSI->getValueOperand()))) {
1465 if (Constant *Res = getConstantValueForLoad(SrcVal: C, Offset, LoadTy: LoadType, DL)) {
1466 LLVM_DEBUG(dbgs() << "Coercing load from store " << *DepSI
1467 << " to constant " << *Res << "\n");
1468 return createConstantExpression(C: Res);
1469 }
1470 }
1471 }
1472 } else if (auto *DepLI = dyn_cast<LoadInst>(Val: DepInst)) {
1473 // Can't forward from non-atomic to atomic without violating memory model.
1474 if (LI->isAtomic() > DepLI->isAtomic())
1475 return nullptr;
1476 int Offset = analyzeLoadFromClobberingLoad(LoadTy: LoadType, LoadPtr, DepLI, DL);
1477 if (Offset >= 0) {
1478 // We can coerce a constant load into a load.
1479 if (auto *C = dyn_cast<Constant>(Val: lookupOperandLeader(V: DepLI)))
1480 if (auto *PossibleConstant =
1481 getConstantValueForLoad(SrcVal: C, Offset, LoadTy: LoadType, DL)) {
1482 LLVM_DEBUG(dbgs() << "Coercing load from load " << *LI
1483 << " to constant " << *PossibleConstant << "\n");
1484 return createConstantExpression(C: PossibleConstant);
1485 }
1486 }
1487 } else if (auto *DepMI = dyn_cast<MemIntrinsic>(Val: DepInst)) {
1488 int Offset = analyzeLoadFromClobberingMemInst(LoadTy: LoadType, LoadPtr, DepMI, DL);
1489 if (Offset >= 0) {
1490 if (auto *PossibleConstant =
1491 getConstantMemInstValueForLoad(SrcInst: DepMI, Offset, LoadTy: LoadType, DL)) {
1492 LLVM_DEBUG(dbgs() << "Coercing load from meminst " << *DepMI
1493 << " to constant " << *PossibleConstant << "\n");
1494 return createConstantExpression(C: PossibleConstant);
1495 }
1496 }
1497 }
1498
1499 if (auto *II = dyn_cast<IntrinsicInst>(Val: DepInst)) {
1500 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1501 auto *LifetimePtr = II->getOperand(i_nocapture: 0);
1502 if (LoadPtr == lookupOperandLeader(V: LifetimePtr) ||
1503 AA->isMustAlias(V1: LoadPtr, V2: LifetimePtr))
1504 return createConstantExpression(C: UndefValue::get(T: LoadType));
1505 }
1506 }
1507
1508 // All of the below are only true if the loaded pointer is produced
1509 // by the dependent instruction.
1510 if (!DepInst->getType()->isPointerTy() ||
1511 (LoadPtr != lookupOperandLeader(V: DepInst) &&
1512 !AA->isMustAlias(V1: LoadPtr, V2: DepInst)))
1513 return nullptr;
1514 // If this load really doesn't depend on anything, then we must be loading an
1515 // undef value. This can happen when loading for a fresh allocation with no
1516 // intervening stores, for example. Note that this is only true in the case
1517 // that the result of the allocation is pointer equal to the load ptr.
1518 if (isa<AllocaInst>(Val: DepInst)) {
1519 return createConstantExpression(C: UndefValue::get(T: LoadType));
1520 } else if (auto *InitVal =
1521 getInitialValueOfAllocation(V: DepInst, TLI, Ty: LoadType))
1522 return createConstantExpression(C: InitVal);
1523
1524 return nullptr;
1525}
1526
1527const Expression *NewGVN::performSymbolicLoadEvaluation(Instruction *I) const {
1528 auto *LI = cast<LoadInst>(Val: I);
1529
1530 // We can eliminate in favor of non-simple loads, but we won't be able to
1531 // eliminate the loads themselves.
1532 if (!LI->isSimple())
1533 return nullptr;
1534
1535 Value *LoadAddressLeader = lookupOperandLeader(V: LI->getPointerOperand());
1536 // Load of undef is UB.
1537 if (isa<UndefValue>(Val: LoadAddressLeader))
1538 return createConstantExpression(C: PoisonValue::get(T: LI->getType()));
1539 MemoryAccess *OriginalAccess = getMemoryAccess(I);
1540 MemoryAccess *DefiningAccess =
1541 MSSAWalker->getClobberingMemoryAccess(MA: OriginalAccess);
1542
1543 if (!MSSA->isLiveOnEntryDef(MA: DefiningAccess)) {
1544 if (auto *MD = dyn_cast<MemoryDef>(Val: DefiningAccess)) {
1545 Instruction *DefiningInst = MD->getMemoryInst();
1546 // If the defining instruction is not reachable, replace with poison.
1547 if (!ReachableBlocks.count(Ptr: DefiningInst->getParent()))
1548 return createConstantExpression(C: PoisonValue::get(T: LI->getType()));
1549 // This will handle stores and memory insts. We only do if it the
1550 // defining access has a different type, or it is a pointer produced by
1551 // certain memory operations that cause the memory to have a fixed value
1552 // (IE things like calloc).
1553 if (const auto *CoercionResult =
1554 performSymbolicLoadCoercion(LoadType: LI->getType(), LoadPtr: LoadAddressLeader, LI,
1555 DepInst: DefiningInst, DefiningAccess))
1556 return CoercionResult;
1557 }
1558 }
1559
1560 const auto *LE = createLoadExpression(LoadType: LI->getType(), PointerOp: LoadAddressLeader, LI,
1561 MA: DefiningAccess);
1562 // If our MemoryLeader is not our defining access, add a use to the
1563 // MemoryLeader, so that we get reprocessed when it changes.
1564 if (LE->getMemoryLeader() != DefiningAccess)
1565 addMemoryUsers(To: LE->getMemoryLeader(), U: OriginalAccess);
1566 return LE;
1567}
1568
1569NewGVN::ExprResult
1570NewGVN::performSymbolicPredicateInfoEvaluation(BitCastInst *I) const {
1571 auto *PI = PredInfo->getPredicateInfoFor(V: I);
1572 if (!PI)
1573 return ExprResult::none();
1574
1575 LLVM_DEBUG(dbgs() << "Found predicate info from instruction !\n");
1576
1577 const std::optional<PredicateConstraint> &Constraint = PI->getConstraint();
1578 if (!Constraint)
1579 return ExprResult::none();
1580
1581 CmpInst::Predicate Predicate = Constraint->Predicate;
1582 Value *CmpOp0 = I->getOperand(i_nocapture: 0);
1583 Value *CmpOp1 = Constraint->OtherOp;
1584
1585 Value *FirstOp = lookupOperandLeader(V: CmpOp0);
1586 Value *SecondOp = lookupOperandLeader(V: CmpOp1);
1587 Value *AdditionallyUsedValue = CmpOp0;
1588
1589 // Sort the ops.
1590 if (shouldSwapOperandsForPredicate(FirstOp, SecondOp, I)) {
1591 std::swap(a&: FirstOp, b&: SecondOp);
1592 Predicate = CmpInst::getSwappedPredicate(pred: Predicate);
1593 AdditionallyUsedValue = CmpOp1;
1594 }
1595
1596 if (Predicate == CmpInst::ICMP_EQ)
1597 return ExprResult::some(Expr: createVariableOrConstant(V: FirstOp),
1598 ExtraDep: AdditionallyUsedValue, PredDep: PI);
1599
1600 // Handle the special case of floating point.
1601 if (Predicate == CmpInst::FCMP_OEQ && isa<ConstantFP>(Val: FirstOp) &&
1602 !cast<ConstantFP>(Val: FirstOp)->isZero())
1603 return ExprResult::some(Expr: createConstantExpression(C: cast<Constant>(Val: FirstOp)),
1604 ExtraDep: AdditionallyUsedValue, PredDep: PI);
1605
1606 return ExprResult::none();
1607}
1608
1609// Evaluate read only and pure calls, and create an expression result.
1610NewGVN::ExprResult NewGVN::performSymbolicCallEvaluation(Instruction *I) const {
1611 auto *CI = cast<CallInst>(Val: I);
1612
1613 // FIXME: Currently the calls which may access the thread id may
1614 // be considered as not accessing the memory. But this is
1615 // problematic for coroutines, since coroutines may resume in a
1616 // different thread. So we disable the optimization here for the
1617 // correctness. However, it may block many other correct
1618 // optimizations. Revert this one when we detect the memory
1619 // accessing kind more precisely.
1620 if (CI->getFunction()->isPresplitCoroutine())
1621 return ExprResult::none();
1622
1623 // Do not combine convergent calls since they implicitly depend on the set of
1624 // threads that is currently executing, and they might be in different basic
1625 // blocks.
1626 if (CI->isConvergent())
1627 return ExprResult::none();
1628
1629 if (AA->doesNotAccessMemory(Call: CI)) {
1630 return ExprResult::some(
1631 Expr: createCallExpression(CI, MA: TOPClass->getMemoryLeader()));
1632 } else if (AA->onlyReadsMemory(Call: CI)) {
1633 if (auto *MA = MSSA->getMemoryAccess(I: CI)) {
1634 auto *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(MA);
1635 return ExprResult::some(Expr: createCallExpression(CI, MA: DefiningAccess));
1636 } else // MSSA determined that CI does not access memory.
1637 return ExprResult::some(
1638 Expr: createCallExpression(CI, MA: TOPClass->getMemoryLeader()));
1639 }
1640 return ExprResult::none();
1641}
1642
1643// Retrieve the memory class for a given MemoryAccess.
1644CongruenceClass *NewGVN::getMemoryClass(const MemoryAccess *MA) const {
1645 auto *Result = MemoryAccessToClass.lookup(Val: MA);
1646 assert(Result && "Should have found memory class");
1647 return Result;
1648}
1649
1650// Update the MemoryAccess equivalence table to say that From is equal to To,
1651// and return true if this is different from what already existed in the table.
1652bool NewGVN::setMemoryClass(const MemoryAccess *From,
1653 CongruenceClass *NewClass) {
1654 assert(NewClass &&
1655 "Every MemoryAccess should be getting mapped to a non-null class");
1656 LLVM_DEBUG(dbgs() << "Setting " << *From);
1657 LLVM_DEBUG(dbgs() << " equivalent to congruence class ");
1658 LLVM_DEBUG(dbgs() << NewClass->getID()
1659 << " with current MemoryAccess leader ");
1660 LLVM_DEBUG(dbgs() << *NewClass->getMemoryLeader() << "\n");
1661
1662 auto LookupResult = MemoryAccessToClass.find(Val: From);
1663 bool Changed = false;
1664 // If it's already in the table, see if the value changed.
1665 if (LookupResult != MemoryAccessToClass.end()) {
1666 auto *OldClass = LookupResult->second;
1667 if (OldClass != NewClass) {
1668 // If this is a phi, we have to handle memory member updates.
1669 if (auto *MP = dyn_cast<MemoryPhi>(Val: From)) {
1670 OldClass->memory_erase(M: MP);
1671 NewClass->memory_insert(M: MP);
1672 // This may have killed the class if it had no non-memory members
1673 if (OldClass->getMemoryLeader() == From) {
1674 if (OldClass->definesNoMemory()) {
1675 OldClass->setMemoryLeader(nullptr);
1676 } else {
1677 OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));
1678 LLVM_DEBUG(dbgs() << "Memory class leader change for class "
1679 << OldClass->getID() << " to "
1680 << *OldClass->getMemoryLeader()
1681 << " due to removal of a memory member " << *From
1682 << "\n");
1683 markMemoryLeaderChangeTouched(CC: OldClass);
1684 }
1685 }
1686 }
1687 // It wasn't equivalent before, and now it is.
1688 LookupResult->second = NewClass;
1689 Changed = true;
1690 }
1691 }
1692
1693 return Changed;
1694}
1695
1696// Determine if a instruction is cycle-free. That means the values in the
1697// instruction don't depend on any expressions that can change value as a result
1698// of the instruction. For example, a non-cycle free instruction would be v =
1699// phi(0, v+1).
1700bool NewGVN::isCycleFree(const Instruction *I) const {
1701 // In order to compute cycle-freeness, we do SCC finding on the instruction,
1702 // and see what kind of SCC it ends up in. If it is a singleton, it is
1703 // cycle-free. If it is not in a singleton, it is only cycle free if the
1704 // other members are all phi nodes (as they do not compute anything, they are
1705 // copies).
1706 auto ICS = InstCycleState.lookup(Val: I);
1707 if (ICS == ICS_Unknown) {
1708 SCCFinder.Start(Start: I);
1709 auto &SCC = SCCFinder.getComponentFor(V: I);
1710 // It's cycle free if it's size 1 or the SCC is *only* phi nodes.
1711 if (SCC.size() == 1)
1712 InstCycleState.insert(KV: {I, ICS_CycleFree});
1713 else {
1714 bool AllPhis = llvm::all_of(Range: SCC, P: [](const Value *V) {
1715 return isa<PHINode>(Val: V) || isCopyOfAPHI(V);
1716 });
1717 ICS = AllPhis ? ICS_CycleFree : ICS_Cycle;
1718 for (const auto *Member : SCC)
1719 if (auto *MemberPhi = dyn_cast<PHINode>(Val: Member))
1720 InstCycleState.insert(KV: {MemberPhi, ICS});
1721 }
1722 }
1723 if (ICS == ICS_Cycle)
1724 return false;
1725 return true;
1726}
1727
1728// Evaluate PHI nodes symbolically and create an expression result.
1729const Expression *
1730NewGVN::performSymbolicPHIEvaluation(ArrayRef<ValPair> PHIOps,
1731 Instruction *I,
1732 BasicBlock *PHIBlock) const {
1733 // True if one of the incoming phi edges is a backedge.
1734 bool HasBackedge = false;
1735 // All constant tracks the state of whether all the *original* phi operands
1736 // This is really shorthand for "this phi cannot cycle due to forward
1737 // change in value of the phi is guaranteed not to later change the value of
1738 // the phi. IE it can't be v = phi(undef, v+1)
1739 bool OriginalOpsConstant = true;
1740 auto *E = cast<PHIExpression>(Val: createPHIExpression(
1741 PHIOperands: PHIOps, I, PHIBlock, HasBackedge, OriginalOpsConstant));
1742 // We match the semantics of SimplifyPhiNode from InstructionSimplify here.
1743 // See if all arguments are the same.
1744 // We track if any were undef because they need special handling.
1745 bool HasUndef = false, HasPoison = false;
1746 auto Filtered = make_filter_range(Range: E->operands(), Pred: [&](Value *Arg) {
1747 if (isa<PoisonValue>(Val: Arg)) {
1748 HasPoison = true;
1749 return false;
1750 }
1751 if (isa<UndefValue>(Val: Arg)) {
1752 HasUndef = true;
1753 return false;
1754 }
1755 return true;
1756 });
1757 // If we are left with no operands, it's dead.
1758 if (Filtered.empty()) {
1759 // If it has undef or poison at this point, it means there are no-non-undef
1760 // arguments, and thus, the value of the phi node must be undef.
1761 if (HasUndef) {
1762 LLVM_DEBUG(
1763 dbgs() << "PHI Node " << *I
1764 << " has no non-undef arguments, valuing it as undef\n");
1765 return createConstantExpression(C: UndefValue::get(T: I->getType()));
1766 }
1767 if (HasPoison) {
1768 LLVM_DEBUG(
1769 dbgs() << "PHI Node " << *I
1770 << " has no non-poison arguments, valuing it as poison\n");
1771 return createConstantExpression(C: PoisonValue::get(T: I->getType()));
1772 }
1773
1774 LLVM_DEBUG(dbgs() << "No arguments of PHI node " << *I << " are live\n");
1775 deleteExpression(E);
1776 return createDeadExpression();
1777 }
1778 Value *AllSameValue = *(Filtered.begin());
1779 ++Filtered.begin();
1780 // Can't use std::equal here, sadly, because filter.begin moves.
1781 if (llvm::all_of(Range&: Filtered, P: equal_to(Arg&: AllSameValue))) {
1782 // Can't fold phi(undef, X) -> X unless X can't be poison (thus X is undef
1783 // in the worst case).
1784 if (HasUndef && !isGuaranteedNotToBePoison(V: AllSameValue, AC, CtxI: nullptr, DT))
1785 return E;
1786
1787 // In LLVM's non-standard representation of phi nodes, it's possible to have
1788 // phi nodes with cycles (IE dependent on other phis that are .... dependent
1789 // on the original phi node), especially in weird CFG's where some arguments
1790 // are unreachable, or uninitialized along certain paths. This can cause
1791 // infinite loops during evaluation. We work around this by not trying to
1792 // really evaluate them independently, but instead using a variable
1793 // expression to say if one is equivalent to the other.
1794 // We also special case undef/poison, so that if we have an undef, we can't
1795 // use the common value unless it dominates the phi block.
1796 if (HasPoison || HasUndef) {
1797 // If we have undef and at least one other value, this is really a
1798 // multivalued phi, and we need to know if it's cycle free in order to
1799 // evaluate whether we can ignore the undef. The other parts of this are
1800 // just shortcuts. If there is no backedge, or all operands are
1801 // constants, it also must be cycle free.
1802 if (HasBackedge && !OriginalOpsConstant &&
1803 !isa<UndefValue>(Val: AllSameValue) && !isCycleFree(I))
1804 return E;
1805
1806 // Only have to check for instructions
1807 if (auto *AllSameInst = dyn_cast<Instruction>(Val: AllSameValue))
1808 if (!someEquivalentDominates(Inst: AllSameInst, U: I))
1809 return E;
1810 }
1811 // Can't simplify to something that comes later in the iteration.
1812 // Otherwise, when and if it changes congruence class, we will never catch
1813 // up. We will always be a class behind it.
1814 if (isa<Instruction>(Val: AllSameValue) &&
1815 InstrToDFSNum(V: AllSameValue) > InstrToDFSNum(V: I))
1816 return E;
1817 NumGVNPhisAllSame++;
1818 LLVM_DEBUG(dbgs() << "Simplified PHI node " << *I << " to " << *AllSameValue
1819 << "\n");
1820 deleteExpression(E);
1821 return createVariableOrConstant(V: AllSameValue);
1822 }
1823 return E;
1824}
1825
1826const Expression *
1827NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) const {
1828 if (auto *EI = dyn_cast<ExtractValueInst>(Val: I)) {
1829 auto *WO = dyn_cast<WithOverflowInst>(Val: EI->getAggregateOperand());
1830 if (WO && EI->getNumIndices() == 1 && *EI->idx_begin() == 0)
1831 // EI is an extract from one of our with.overflow intrinsics. Synthesize
1832 // a semantically equivalent expression instead of an extract value
1833 // expression.
1834 return createBinaryExpression(Opcode: WO->getBinaryOp(), T: EI->getType(),
1835 Arg1: WO->getLHS(), Arg2: WO->getRHS(), I);
1836 }
1837
1838 return createAggregateValueExpression(I);
1839}
1840
1841NewGVN::ExprResult NewGVN::performSymbolicCmpEvaluation(Instruction *I) const {
1842 assert(isa<CmpInst>(I) && "Expected a cmp instruction.");
1843
1844 auto *CI = cast<CmpInst>(Val: I);
1845 // See if our operands are equal to those of a previous predicate, and if so,
1846 // if it implies true or false.
1847 auto Op0 = lookupOperandLeader(V: CI->getOperand(i_nocapture: 0));
1848 auto Op1 = lookupOperandLeader(V: CI->getOperand(i_nocapture: 1));
1849 auto OurPredicate = CI->getPredicate();
1850 if (shouldSwapOperands(Op0, Op1)) {
1851 std::swap(a&: Op0, b&: Op1);
1852 OurPredicate = CI->getSwappedPredicate();
1853 }
1854
1855 // Avoid processing the same info twice.
1856 const PredicateBase *LastPredInfo = nullptr;
1857 // See if we know something about the comparison itself, like it is the target
1858 // of an assume.
1859 auto *CmpPI = PredInfo->getPredicateInfoFor(V: I);
1860 if (isa_and_nonnull<PredicateAssume>(Val: CmpPI))
1861 return ExprResult::some(
1862 Expr: createConstantExpression(C: ConstantInt::getTrue(Ty: CI->getType())));
1863
1864 if (Op0 == Op1) {
1865 // This condition does not depend on predicates, no need to add users
1866 if (CI->isTrueWhenEqual())
1867 return ExprResult::some(
1868 Expr: createConstantExpression(C: ConstantInt::getTrue(Ty: CI->getType())));
1869 else if (CI->isFalseWhenEqual())
1870 return ExprResult::some(
1871 Expr: createConstantExpression(C: ConstantInt::getFalse(Ty: CI->getType())));
1872 }
1873
1874 // NOTE: Because we are comparing both operands here and below, and using
1875 // previous comparisons, we rely on fact that predicateinfo knows to mark
1876 // comparisons that use renamed operands as users of the earlier comparisons.
1877 // It is *not* enough to just mark predicateinfo renamed operands as users of
1878 // the earlier comparisons, because the *other* operand may have changed in a
1879 // previous iteration.
1880 // Example:
1881 // icmp slt %a, %b
1882 // %b.0 = ssa.copy(%b)
1883 // false branch:
1884 // icmp slt %c, %b.0
1885
1886 // %c and %a may start out equal, and thus, the code below will say the second
1887 // %icmp is false. c may become equal to something else, and in that case the
1888 // %second icmp *must* be reexamined, but would not if only the renamed
1889 // %operands are considered users of the icmp.
1890
1891 // *Currently* we only check one level of comparisons back, and only mark one
1892 // level back as touched when changes happen. If you modify this code to look
1893 // back farther through comparisons, you *must* mark the appropriate
1894 // comparisons as users in PredicateInfo.cpp, or you will cause bugs. See if
1895 // we know something just from the operands themselves
1896
1897 // See if our operands have predicate info, so that we may be able to derive
1898 // something from a previous comparison.
1899 for (const auto &Op : CI->operands()) {
1900 auto *PI = PredInfo->getPredicateInfoFor(V: Op);
1901 if (const auto *PBranch = dyn_cast_or_null<PredicateBranch>(Val: PI)) {
1902 if (PI == LastPredInfo)
1903 continue;
1904 LastPredInfo = PI;
1905 // In phi of ops cases, we may have predicate info that we are evaluating
1906 // in a different context.
1907 if (!DT->dominates(A: PBranch->To, B: I->getParent()))
1908 continue;
1909 // TODO: Along the false edge, we may know more things too, like
1910 // icmp of
1911 // same operands is false.
1912 // TODO: We only handle actual comparison conditions below, not
1913 // and/or.
1914 auto *BranchCond = dyn_cast<CmpInst>(Val: PBranch->Condition);
1915 if (!BranchCond)
1916 continue;
1917 auto *BranchOp0 = lookupOperandLeader(V: BranchCond->getOperand(i_nocapture: 0));
1918 auto *BranchOp1 = lookupOperandLeader(V: BranchCond->getOperand(i_nocapture: 1));
1919 auto BranchPredicate = BranchCond->getPredicate();
1920 if (shouldSwapOperands(BranchOp0, BranchOp1)) {
1921 std::swap(a&: BranchOp0, b&: BranchOp1);
1922 BranchPredicate = BranchCond->getSwappedPredicate();
1923 }
1924 if (BranchOp0 == Op0 && BranchOp1 == Op1) {
1925 if (PBranch->TrueEdge) {
1926 // If we know the previous predicate is true and we are in the true
1927 // edge then we may be implied true or false.
1928 if (auto R = ICmpInst::isImpliedByMatchingCmp(Pred1: BranchPredicate,
1929 Pred2: OurPredicate)) {
1930 auto *C = ConstantInt::getBool(Ty: CI->getType(), V: *R);
1931 return ExprResult::some(Expr: createConstantExpression(C), PredDep: PI);
1932 }
1933 } else {
1934 // Just handle the ne and eq cases, where if we have the same
1935 // operands, we may know something.
1936 if (BranchPredicate == OurPredicate) {
1937 // Same predicate, same ops,we know it was false, so this is false.
1938 return ExprResult::some(
1939 Expr: createConstantExpression(C: ConstantInt::getFalse(Ty: CI->getType())),
1940 PredDep: PI);
1941 } else if (BranchPredicate ==
1942 CmpInst::getInversePredicate(pred: OurPredicate)) {
1943 // Inverse predicate, we know the other was false, so this is true.
1944 return ExprResult::some(
1945 Expr: createConstantExpression(C: ConstantInt::getTrue(Ty: CI->getType())),
1946 PredDep: PI);
1947 }
1948 }
1949 }
1950 }
1951 }
1952 // Create expression will take care of simplifyCmpInst
1953 return createExpression(I);
1954}
1955
1956// Substitute and symbolize the instruction before value numbering.
1957NewGVN::ExprResult
1958NewGVN::performSymbolicEvaluation(Instruction *I,
1959 SmallPtrSetImpl<Value *> &Visited) const {
1960
1961 const Expression *E = nullptr;
1962 // TODO: memory intrinsics.
1963 // TODO: Some day, we should do the forward propagation and reassociation
1964 // parts of the algorithm.
1965 switch (I->getOpcode()) {
1966 case Instruction::ExtractValue:
1967 case Instruction::InsertValue:
1968 E = performSymbolicAggrValueEvaluation(I);
1969 break;
1970 case Instruction::PHI: {
1971 SmallVector<ValPair, 3> Ops;
1972 auto *PN = cast<PHINode>(Val: I);
1973 for (unsigned i = 0; i < PN->getNumOperands(); ++i)
1974 Ops.push_back(Elt: {PN->getIncomingValue(i), PN->getIncomingBlock(i)});
1975 // Sort to ensure the invariant createPHIExpression requires is met.
1976 sortPHIOps(Ops);
1977 E = performSymbolicPHIEvaluation(PHIOps: Ops, I, PHIBlock: getBlockForValue(V: I));
1978 } break;
1979 case Instruction::Call:
1980 return performSymbolicCallEvaluation(I);
1981 break;
1982 case Instruction::Store:
1983 E = performSymbolicStoreEvaluation(I);
1984 break;
1985 case Instruction::Load:
1986 E = performSymbolicLoadEvaluation(I);
1987 break;
1988 case Instruction::BitCast:
1989 // Intrinsics with the returned attribute are copies of arguments.
1990 if (I->getType() == I->getOperand(i: 0)->getType())
1991 if (auto Res =
1992 performSymbolicPredicateInfoEvaluation(I: cast<BitCastInst>(Val: I)))
1993 return Res;
1994 [[fallthrough]];
1995 case Instruction::AddrSpaceCast:
1996 case Instruction::Freeze:
1997 return createExpression(I);
1998 break;
1999 case Instruction::ICmp:
2000 case Instruction::FCmp:
2001 return performSymbolicCmpEvaluation(I);
2002 break;
2003 case Instruction::FNeg:
2004 case Instruction::Add:
2005 case Instruction::FAdd:
2006 case Instruction::Sub:
2007 case Instruction::FSub:
2008 case Instruction::Mul:
2009 case Instruction::FMul:
2010 case Instruction::UDiv:
2011 case Instruction::SDiv:
2012 case Instruction::FDiv:
2013 case Instruction::URem:
2014 case Instruction::SRem:
2015 case Instruction::FRem:
2016 case Instruction::Shl:
2017 case Instruction::LShr:
2018 case Instruction::AShr:
2019 case Instruction::And:
2020 case Instruction::Or:
2021 case Instruction::Xor:
2022 case Instruction::Trunc:
2023 case Instruction::ZExt:
2024 case Instruction::SExt:
2025 case Instruction::FPToUI:
2026 case Instruction::FPToSI:
2027 case Instruction::UIToFP:
2028 case Instruction::SIToFP:
2029 case Instruction::FPTrunc:
2030 case Instruction::FPExt:
2031 case Instruction::PtrToInt:
2032 case Instruction::PtrToAddr:
2033 case Instruction::IntToPtr:
2034 case Instruction::Select:
2035 case Instruction::ExtractElement:
2036 case Instruction::InsertElement:
2037 case Instruction::GetElementPtr:
2038 return createExpression(I);
2039 break;
2040 case Instruction::ShuffleVector:
2041 // FIXME: Add support for shufflevector to createExpression.
2042 return ExprResult::none();
2043 default:
2044 return ExprResult::none();
2045 }
2046 return ExprResult::some(Expr: E);
2047}
2048
2049// Look up a container of values/instructions in a map, and touch all the
2050// instructions in the container. Then erase value from the map.
2051template <typename Map, typename KeyType>
2052void NewGVN::touchAndErase(Map &M, const KeyType &Key) {
2053 const auto Result = M.find_as(Key);
2054 if (Result != M.end()) {
2055 for (const typename Map::mapped_type::value_type Mapped : Result->second)
2056 TouchedInstructions.set(InstrToDFSNum(Mapped));
2057 M.erase(Result);
2058 }
2059}
2060
2061void NewGVN::addAdditionalUsers(Value *To, Value *User) const {
2062 assert(User && To != User);
2063 if (isa<Instruction>(Val: To))
2064 AdditionalUsers[To].insert(Ptr: User);
2065}
2066
2067void NewGVN::addAdditionalUsers(ExprResult &Res, Instruction *User) const {
2068 if (Res.ExtraDep && Res.ExtraDep != User)
2069 addAdditionalUsers(To: Res.ExtraDep, User);
2070 Res.ExtraDep = nullptr;
2071
2072 if (Res.PredDep) {
2073 if (const auto *PBranch = dyn_cast<PredicateBranch>(Val: Res.PredDep))
2074 PredicateToUsers[PBranch->Condition].insert(Ptr: User);
2075 else if (const auto *PAssume =
2076 dyn_cast<PredicateConditionAssume>(Val: Res.PredDep))
2077 PredicateToUsers[PAssume->Condition].insert(Ptr: User);
2078 }
2079 Res.PredDep = nullptr;
2080}
2081
2082void NewGVN::markUsersTouched(Value *V) {
2083 // Now mark the users as touched.
2084 for (auto *User : V->users()) {
2085 assert(isa<Instruction>(User) && "Use of value not within an instruction?");
2086 TouchedInstructions.set(InstrToDFSNum(V: User));
2087 }
2088 touchAndErase(M&: AdditionalUsers, Key: V);
2089}
2090
2091void NewGVN::addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const {
2092 LLVM_DEBUG(dbgs() << "Adding memory user " << *U << " to " << *To << "\n");
2093 MemoryToUsers[To].insert(Ptr: U);
2094}
2095
2096void NewGVN::markMemoryDefTouched(const MemoryAccess *MA) {
2097 TouchedInstructions.set(MemoryToDFSNum(MA));
2098}
2099
2100void NewGVN::markMemoryUsersTouched(const MemoryAccess *MA) {
2101 if (isa<MemoryUse>(Val: MA))
2102 return;
2103 for (const auto *U : MA->users())
2104 TouchedInstructions.set(MemoryToDFSNum(MA: U));
2105 touchAndErase(M&: MemoryToUsers, Key: MA);
2106}
2107
2108// Touch all the predicates that depend on this instruction.
2109void NewGVN::markPredicateUsersTouched(Instruction *I) {
2110 touchAndErase(M&: PredicateToUsers, Key: I);
2111}
2112
2113// Mark users affected by a memory leader change.
2114void NewGVN::markMemoryLeaderChangeTouched(CongruenceClass *CC) {
2115 for (const auto *M : CC->memory())
2116 markMemoryDefTouched(MA: M);
2117}
2118
2119// Touch the instructions that need to be updated after a congruence class has a
2120// leader change, and mark changed values.
2121void NewGVN::markValueLeaderChangeTouched(CongruenceClass *CC) {
2122 for (auto *M : *CC) {
2123 if (auto *I = dyn_cast<Instruction>(Val: M))
2124 TouchedInstructions.set(InstrToDFSNum(V: I));
2125 LeaderChanges.insert(Ptr: M);
2126 }
2127}
2128
2129// Give a range of things that have instruction DFS numbers, this will return
2130// the member of the range with the smallest dfs number.
2131template <class T, class Range>
2132T *NewGVN::getMinDFSOfRange(const Range &R) const {
2133 std::pair<T *, unsigned> MinDFS = {nullptr, ~0U};
2134 for (const auto X : R) {
2135 auto DFSNum = InstrToDFSNum(X);
2136 if (DFSNum < MinDFS.second)
2137 MinDFS = {X, DFSNum};
2138 }
2139 return MinDFS.first;
2140}
2141
2142// This function returns the MemoryAccess that should be the next leader of
2143// congruence class CC, under the assumption that the current leader is going to
2144// disappear.
2145const MemoryAccess *NewGVN::getNextMemoryLeader(CongruenceClass *CC) const {
2146 // TODO: If this ends up to slow, we can maintain a next memory leader like we
2147 // do for regular leaders.
2148 // Make sure there will be a leader to find.
2149 assert(!CC->definesNoMemory() && "Can't get next leader if there is none");
2150 if (CC->getStoreCount() > 0) {
2151 if (auto *NL = dyn_cast_or_null<StoreInst>(Val: CC->getNextLeader().first))
2152 return getMemoryAccess(I: NL);
2153 // Find the store with the minimum DFS number.
2154 auto *V = getMinDFSOfRange<Value>(R: make_filter_range(
2155 Range&: *CC, Pred: [&](const Value *V) { return isa<StoreInst>(Val: V); }));
2156 return getMemoryAccess(I: cast<StoreInst>(Val: V));
2157 }
2158 assert(CC->getStoreCount() == 0);
2159
2160 // Given our assertion, hitting this part must mean
2161 // !OldClass->memory_empty()
2162 if (CC->memory_size() == 1)
2163 return *CC->memory_begin();
2164 return getMinDFSOfRange<const MemoryPhi>(R: CC->memory());
2165}
2166
2167// This function returns the next value leader of a congruence class, under the
2168// assumption that the current leader is going away. This should end up being
2169// the next most dominating member.
2170Value *NewGVN::getNextValueLeader(CongruenceClass *CC) const {
2171 // We don't need to sort members if there is only 1, and we don't care about
2172 // sorting the TOP class because everything either gets out of it or is
2173 // unreachable.
2174
2175 if (CC->size() == 1 || CC == TOPClass) {
2176 return *(CC->begin());
2177 } else if (CC->getNextLeader().first) {
2178 ++NumGVNAvoidedSortedLeaderChanges;
2179 return CC->getNextLeader().first;
2180 } else {
2181 ++NumGVNSortedLeaderChanges;
2182 // NOTE: If this ends up to slow, we can maintain a dual structure for
2183 // member testing/insertion, or keep things mostly sorted, and sort only
2184 // here, or use SparseBitVector or ....
2185 return getMinDFSOfRange<Value>(R: *CC);
2186 }
2187}
2188
2189// Move a MemoryAccess, currently in OldClass, to NewClass, including updates to
2190// the memory members, etc for the move.
2191//
2192// The invariants of this function are:
2193//
2194// - I must be moving to NewClass from OldClass
2195// - The StoreCount of OldClass and NewClass is expected to have been updated
2196// for I already if it is a store.
2197// - The OldClass memory leader has not been updated yet if I was the leader.
2198void NewGVN::moveMemoryToNewCongruenceClass(Instruction *I,
2199 MemoryAccess *InstMA,
2200 CongruenceClass *OldClass,
2201 CongruenceClass *NewClass) {
2202 // If the leader is I, and we had a representative MemoryAccess, it should
2203 // be the MemoryAccess of OldClass.
2204 assert((!InstMA || !OldClass->getMemoryLeader() ||
2205 OldClass->getLeader() != I ||
2206 MemoryAccessToClass.lookup(OldClass->getMemoryLeader()) ==
2207 MemoryAccessToClass.lookup(InstMA)) &&
2208 "Representative MemoryAccess mismatch");
2209 // First, see what happens to the new class
2210 if (!NewClass->getMemoryLeader()) {
2211 // Should be a new class, or a store becoming a leader of a new class.
2212 assert(NewClass->size() == 1 ||
2213 (isa<StoreInst>(I) && NewClass->getStoreCount() == 1));
2214 NewClass->setMemoryLeader(InstMA);
2215 // Mark it touched if we didn't just create a singleton
2216 LLVM_DEBUG(dbgs() << "Memory class leader change for class "
2217 << NewClass->getID()
2218 << " due to new memory instruction becoming leader\n");
2219 markMemoryLeaderChangeTouched(CC: NewClass);
2220 }
2221 setMemoryClass(From: InstMA, NewClass);
2222 // Now, fixup the old class if necessary
2223 if (OldClass->getMemoryLeader() == InstMA) {
2224 if (!OldClass->definesNoMemory()) {
2225 OldClass->setMemoryLeader(getNextMemoryLeader(CC: OldClass));
2226 LLVM_DEBUG(dbgs() << "Memory class leader change for class "
2227 << OldClass->getID() << " to "
2228 << *OldClass->getMemoryLeader()
2229 << " due to removal of old leader " << *InstMA << "\n");
2230 markMemoryLeaderChangeTouched(CC: OldClass);
2231 } else
2232 OldClass->setMemoryLeader(nullptr);
2233 }
2234}
2235
2236// Move a value, currently in OldClass, to be part of NewClass
2237// Update OldClass and NewClass for the move (including changing leaders, etc).
2238void NewGVN::moveValueToNewCongruenceClass(Instruction *I, const Expression *E,
2239 CongruenceClass *OldClass,
2240 CongruenceClass *NewClass) {
2241 if (I == OldClass->getNextLeader().first)
2242 OldClass->resetNextLeader();
2243
2244 OldClass->erase(M: I);
2245 NewClass->insert(M: I);
2246
2247 // Ensure that the leader has the lowest RPO. If the leader changed notify all
2248 // members of the class.
2249 if (NewClass->getLeader() != I &&
2250 NewClass->addPossibleLeader(LeaderPair: {I, InstrToDFSNum(V: I)})) {
2251 markValueLeaderChangeTouched(CC: NewClass);
2252 }
2253
2254 // Handle our special casing of stores.
2255 if (auto *SI = dyn_cast<StoreInst>(Val: I)) {
2256 OldClass->decStoreCount();
2257 // Okay, so when do we want to make a store a leader of a class?
2258 // If we have a store defined by an earlier load, we want the earlier load
2259 // to lead the class.
2260 // If we have a store defined by something else, we want the store to lead
2261 // the class so everything else gets the "something else" as a value.
2262 // If we have a store as the single member of the class, we want the store
2263 // as the leader
2264 if (NewClass->getStoreCount() == 0 && !NewClass->getStoredValue()) {
2265 // If it's a store expression we are using, it means we are not equivalent
2266 // to something earlier.
2267 if (auto *SE = dyn_cast<StoreExpression>(Val: E)) {
2268 NewClass->setStoredValue(SE->getStoredValue());
2269 markValueLeaderChangeTouched(CC: NewClass);
2270 // Shift the new class leader to be the store
2271 LLVM_DEBUG(dbgs() << "Changing leader of congruence class "
2272 << NewClass->getID() << " from "
2273 << *NewClass->getLeader() << " to " << *SI
2274 << " because store joined class\n");
2275 // If we changed the leader, we have to mark it changed because we don't
2276 // know what it will do to symbolic evaluation.
2277 NewClass->setLeader({SI, InstrToDFSNum(V: SI)});
2278 }
2279 // We rely on the code below handling the MemoryAccess change.
2280 }
2281 NewClass->incStoreCount();
2282 }
2283 // True if there is no memory instructions left in a class that had memory
2284 // instructions before.
2285
2286 // If it's not a memory use, set the MemoryAccess equivalence
2287 auto *InstMA = dyn_cast_or_null<MemoryDef>(Val: getMemoryAccess(I));
2288 if (InstMA)
2289 moveMemoryToNewCongruenceClass(I, InstMA, OldClass, NewClass);
2290 ValueToClass[I] = NewClass;
2291 // See if we destroyed the class or need to swap leaders.
2292 if (OldClass->empty() && OldClass != TOPClass) {
2293 if (OldClass->getDefiningExpr()) {
2294 LLVM_DEBUG(dbgs() << "Erasing expression " << *OldClass->getDefiningExpr()
2295 << " from table\n");
2296 // We erase it as an exact expression to make sure we don't just erase an
2297 // equivalent one.
2298 auto Iter = ExpressionToClass.find_as(
2299 Val: ExactEqualsExpression(*OldClass->getDefiningExpr()));
2300 if (Iter != ExpressionToClass.end())
2301 ExpressionToClass.erase(I: Iter);
2302#ifdef EXPENSIVE_CHECKS
2303 assert(
2304 (*OldClass->getDefiningExpr() != *E || ExpressionToClass.lookup(E)) &&
2305 "We erased the expression we just inserted, which should not happen");
2306#endif
2307 }
2308 } else if (OldClass->getLeader() == I) {
2309 // When the leader changes, the value numbering of
2310 // everything may change due to symbolization changes, so we need to
2311 // reprocess.
2312 LLVM_DEBUG(dbgs() << "Value class leader change for class "
2313 << OldClass->getID() << "\n");
2314 ++NumGVNLeaderChanges;
2315 // Destroy the stored value if there are no more stores to represent it.
2316 // Note that this is basically clean up for the expression removal that
2317 // happens below. If we remove stores from a class, we may leave it as a
2318 // class of equivalent memory phis.
2319 if (OldClass->getStoreCount() == 0) {
2320 if (OldClass->getStoredValue())
2321 OldClass->setStoredValue(nullptr);
2322 }
2323 OldClass->setLeader({getNextValueLeader(CC: OldClass),
2324 InstrToDFSNum(V: getNextValueLeader(CC: OldClass))});
2325 OldClass->resetNextLeader();
2326 markValueLeaderChangeTouched(CC: OldClass);
2327 }
2328}
2329
2330// For a given expression, mark the phi of ops instructions that could have
2331// changed as a result.
2332void NewGVN::markPhiOfOpsChanged(const Expression *E) {
2333 touchAndErase(M&: ExpressionToPhiOfOps, Key: E);
2334}
2335
2336// Perform congruence finding on a given value numbering expression.
2337void NewGVN::performCongruenceFinding(Instruction *I, const Expression *E) {
2338 // This is guaranteed to return something, since it will at least find
2339 // TOP.
2340
2341 CongruenceClass *IClass = ValueToClass.lookup(Val: I);
2342 assert(IClass && "Should have found a IClass");
2343 // Dead classes should have been eliminated from the mapping.
2344 assert(!IClass->isDead() && "Found a dead class");
2345
2346 CongruenceClass *EClass = nullptr;
2347 if (const auto *VE = dyn_cast<VariableExpression>(Val: E)) {
2348 EClass = ValueToClass.lookup(Val: VE->getVariableValue());
2349 } else if (isa<DeadExpression>(Val: E)) {
2350 EClass = TOPClass;
2351 }
2352 if (!EClass) {
2353 auto lookupResult = ExpressionToClass.try_emplace(Key: E);
2354
2355 // If it's not in the value table, create a new congruence class.
2356 if (lookupResult.second) {
2357 CongruenceClass *NewClass = createCongruenceClass(Leader: nullptr, E);
2358 auto place = lookupResult.first;
2359 place->second = NewClass;
2360
2361 // Constants and variables should always be made the leader.
2362 if (const auto *CE = dyn_cast<ConstantExpression>(Val: E)) {
2363 NewClass->setLeader({CE->getConstantValue(), 0});
2364 } else if (const auto *SE = dyn_cast<StoreExpression>(Val: E)) {
2365 StoreInst *SI = SE->getStoreInst();
2366 NewClass->setLeader({SI, InstrToDFSNum(V: SI)});
2367 NewClass->setStoredValue(SE->getStoredValue());
2368 // The RepMemoryAccess field will be filled in properly by the
2369 // moveValueToNewCongruenceClass call.
2370 } else {
2371 NewClass->setLeader({I, InstrToDFSNum(V: I)});
2372 }
2373 assert(!isa<VariableExpression>(E) &&
2374 "VariableExpression should have been handled already");
2375
2376 EClass = NewClass;
2377 LLVM_DEBUG(dbgs() << "Created new congruence class for " << *I
2378 << " using expression " << *E << " at "
2379 << NewClass->getID() << " and leader "
2380 << *(NewClass->getLeader()));
2381 if (NewClass->getStoredValue())
2382 LLVM_DEBUG(dbgs() << " and stored value "
2383 << *(NewClass->getStoredValue()));
2384 LLVM_DEBUG(dbgs() << "\n");
2385 } else {
2386 EClass = lookupResult.first->second;
2387 if (isa<ConstantExpression>(Val: E))
2388 assert((isa<Constant>(EClass->getLeader()) ||
2389 (EClass->getStoredValue() &&
2390 isa<Constant>(EClass->getStoredValue()))) &&
2391 "Any class with a constant expression should have a "
2392 "constant leader");
2393
2394 assert(EClass && "Somehow don't have an eclass");
2395
2396 assert(!EClass->isDead() && "We accidentally looked up a dead class");
2397 }
2398 }
2399 bool ClassChanged = IClass != EClass;
2400 bool LeaderChanged = LeaderChanges.erase(Ptr: I);
2401 if (ClassChanged || LeaderChanged) {
2402 LLVM_DEBUG(dbgs() << "New class " << EClass->getID() << " for expression "
2403 << *E << "\n");
2404 if (ClassChanged) {
2405 moveValueToNewCongruenceClass(I, E, OldClass: IClass, NewClass: EClass);
2406 markPhiOfOpsChanged(E);
2407 }
2408
2409 markUsersTouched(V: I);
2410 if (MemoryAccess *MA = getMemoryAccess(I))
2411 markMemoryUsersTouched(MA);
2412 if (auto *CI = dyn_cast<CmpInst>(Val: I))
2413 markPredicateUsersTouched(I: CI);
2414 }
2415 // If we changed the class of the store, we want to ensure nothing finds the
2416 // old store expression. In particular, loads do not compare against stored
2417 // value, so they will find old store expressions (and associated class
2418 // mappings) if we leave them in the table.
2419 if (ClassChanged && isa<StoreInst>(Val: I)) {
2420 auto *OldE = ValueToExpression.lookup(Val: I);
2421 // It could just be that the old class died. We don't want to erase it if we
2422 // just moved classes.
2423 if (OldE && isa<StoreExpression>(Val: OldE) && *E != *OldE) {
2424 // Erase this as an exact expression to ensure we don't erase expressions
2425 // equivalent to it.
2426 auto Iter = ExpressionToClass.find_as(Val: ExactEqualsExpression(*OldE));
2427 if (Iter != ExpressionToClass.end())
2428 ExpressionToClass.erase(I: Iter);
2429 }
2430 }
2431 ValueToExpression[I] = E;
2432}
2433
2434// Process the fact that Edge (from, to) is reachable, including marking
2435// any newly reachable blocks and instructions for processing.
2436void NewGVN::updateReachableEdge(BasicBlock *From, BasicBlock *To) {
2437 // Check if the Edge was reachable before.
2438 if (ReachableEdges.insert(V: {From, To}).second) {
2439 // If this block wasn't reachable before, all instructions are touched.
2440 if (ReachableBlocks.insert(Ptr: To).second) {
2441 LLVM_DEBUG(dbgs() << "Block " << getBlockName(To)
2442 << " marked reachable\n");
2443 const auto &InstRange = BlockInstRange.lookup(Val: To);
2444 TouchedInstructions.set(I: InstRange.first, E: InstRange.second);
2445 } else {
2446 LLVM_DEBUG(dbgs() << "Block " << getBlockName(To)
2447 << " was reachable, but new edge {"
2448 << getBlockName(From) << "," << getBlockName(To)
2449 << "} to it found\n");
2450
2451 // We've made an edge reachable to an existing block, which may
2452 // impact predicates. Otherwise, only mark the phi nodes as touched, as
2453 // they are the only thing that depend on new edges. Anything using their
2454 // values will get propagated to if necessary.
2455 if (MemoryAccess *MemPhi = getMemoryAccess(BB: To))
2456 TouchedInstructions.set(InstrToDFSNum(MA: MemPhi));
2457
2458 // FIXME: We should just add a union op on a Bitvector and
2459 // SparseBitVector. We can do it word by word faster than we are doing it
2460 // here.
2461 for (auto InstNum : RevisitOnReachabilityChange[To])
2462 TouchedInstructions.set(InstNum);
2463 }
2464 }
2465}
2466
2467// Given a predicate condition (from a switch, cmp, or whatever) and a block,
2468// see if we know some constant value for it already.
2469Value *NewGVN::findConditionEquivalence(Value *Cond) const {
2470 auto Result = lookupOperandLeader(V: Cond);
2471 return isa<Constant>(Val: Result) ? Result : nullptr;
2472}
2473
2474// Process the outgoing edges of a block for reachability.
2475void NewGVN::processOutgoingEdges(Instruction *TI, BasicBlock *B) {
2476 // Evaluate reachability of terminator instruction.
2477 Value *Cond;
2478 BasicBlock *TrueSucc, *FalseSucc;
2479 if (match(V: TI, P: m_Br(C: m_Value(V&: Cond), T&: TrueSucc, F&: FalseSucc))) {
2480 Value *CondEvaluated = findConditionEquivalence(Cond);
2481 if (!CondEvaluated) {
2482 if (auto *I = dyn_cast<Instruction>(Val: Cond)) {
2483 SmallPtrSet<Value *, 4> Visited;
2484 auto Res = performSymbolicEvaluation(I, Visited);
2485 if (const auto *CE = dyn_cast_or_null<ConstantExpression>(Val: Res.Expr)) {
2486 CondEvaluated = CE->getConstantValue();
2487 addAdditionalUsers(Res, User: I);
2488 } else {
2489 // Did not use simplification result, no need to add the extra
2490 // dependency.
2491 Res.ExtraDep = nullptr;
2492 }
2493 } else if (isa<ConstantInt>(Val: Cond)) {
2494 CondEvaluated = Cond;
2495 }
2496 }
2497 ConstantInt *CI;
2498 if (CondEvaluated && (CI = dyn_cast<ConstantInt>(Val: CondEvaluated))) {
2499 if (CI->isOne()) {
2500 LLVM_DEBUG(dbgs() << "Condition for Terminator " << *TI
2501 << " evaluated to true\n");
2502 updateReachableEdge(From: B, To: TrueSucc);
2503 } else if (CI->isZero()) {
2504 LLVM_DEBUG(dbgs() << "Condition for Terminator " << *TI
2505 << " evaluated to false\n");
2506 updateReachableEdge(From: B, To: FalseSucc);
2507 }
2508 } else {
2509 updateReachableEdge(From: B, To: TrueSucc);
2510 updateReachableEdge(From: B, To: FalseSucc);
2511 }
2512 } else if (auto *SI = dyn_cast<SwitchInst>(Val: TI)) {
2513 // For switches, propagate the case values into the case
2514 // destinations.
2515
2516 Value *SwitchCond = SI->getCondition();
2517 Value *CondEvaluated = findConditionEquivalence(Cond: SwitchCond);
2518 // See if we were able to turn this switch statement into a constant.
2519 if (CondEvaluated && isa<ConstantInt>(Val: CondEvaluated)) {
2520 auto *CondVal = cast<ConstantInt>(Val: CondEvaluated);
2521 // We should be able to get case value for this.
2522 auto Case = *SI->findCaseValue(C: CondVal);
2523 if (Case.getCaseSuccessor() == SI->getDefaultDest()) {
2524 // We proved the value is outside of the range of the case.
2525 // We can't do anything other than mark the default dest as reachable,
2526 // and go home.
2527 updateReachableEdge(From: B, To: SI->getDefaultDest());
2528 return;
2529 }
2530 // Now get where it goes and mark it reachable.
2531 BasicBlock *TargetBlock = Case.getCaseSuccessor();
2532 updateReachableEdge(From: B, To: TargetBlock);
2533 } else {
2534 for (BasicBlock *TargetBlock : successors(BB: SI->getParent()))
2535 updateReachableEdge(From: B, To: TargetBlock);
2536 }
2537 } else {
2538 // Otherwise this is either unconditional, or a type we have no
2539 // idea about. Just mark successors as reachable.
2540 for (BasicBlock *TargetBlock : successors(BB: TI->getParent()))
2541 updateReachableEdge(From: B, To: TargetBlock);
2542
2543 // This also may be a memory defining terminator, in which case, set it
2544 // equivalent only to itself.
2545 //
2546 auto *MA = getMemoryAccess(I: TI);
2547 if (MA && !isa<MemoryUse>(Val: MA)) {
2548 auto *CC = ensureLeaderOfMemoryClass(MA);
2549 if (setMemoryClass(From: MA, NewClass: CC))
2550 markMemoryUsersTouched(MA);
2551 }
2552 }
2553}
2554
2555// Remove the PHI of Ops PHI for I
2556void NewGVN::removePhiOfOps(Instruction *I, PHINode *PHITemp) {
2557 InstrDFS.erase(Val: PHITemp);
2558 // It's still a temp instruction. We keep it in the array so it gets erased.
2559 // However, it's no longer used by I, or in the block
2560 TempToBlock.erase(Val: PHITemp);
2561 RealToTemp.erase(Val: I);
2562 // We don't remove the users from the phi node uses. This wastes a little
2563 // time, but such is life. We could use two sets to track which were there
2564 // are the start of NewGVN, and which were added, but right nowt he cost of
2565 // tracking is more than the cost of checking for more phi of ops.
2566}
2567
2568// Add PHI Op in BB as a PHI of operations version of ExistingValue.
2569void NewGVN::addPhiOfOps(PHINode *Op, BasicBlock *BB,
2570 Instruction *ExistingValue) {
2571 InstrDFS[Op] = InstrToDFSNum(V: ExistingValue);
2572 AllTempInstructions.insert(V: Op);
2573 TempToBlock[Op] = BB;
2574 RealToTemp[ExistingValue] = Op;
2575 // Add all users to phi node use, as they are now uses of the phi of ops phis
2576 // and may themselves be phi of ops.
2577 for (auto *U : ExistingValue->users())
2578 if (auto *UI = dyn_cast<Instruction>(Val: U))
2579 PHINodeUses.insert(Ptr: UI);
2580}
2581
2582static bool okayForPHIOfOps(const ScalarOptions &Opts, const Instruction *I) {
2583 if (!Opts.enable_phi_of_ops)
2584 return false;
2585 return isa<BinaryOperator>(Val: I) || isa<SelectInst>(Val: I) || isa<CmpInst>(Val: I) ||
2586 isa<LoadInst>(Val: I);
2587}
2588
2589// Return true if this operand will be safe to use for phi of ops.
2590//
2591// The reason some operands are unsafe is that we are not trying to recursively
2592// translate everything back through phi nodes. We actually expect some lookups
2593// of expressions to fail. In particular, a lookup where the expression cannot
2594// exist in the predecessor. This is true even if the expression, as shown, can
2595// be determined to be constant.
2596bool NewGVN::OpIsSafeForPHIOfOps(Value *V, const BasicBlock *PHIBlock,
2597 SmallPtrSetImpl<const Value *> &Visited) {
2598 SmallVector<Value *, 4> Worklist;
2599 Worklist.push_back(Elt: V);
2600 while (!Worklist.empty()) {
2601 auto *I = Worklist.pop_back_val();
2602 if (!isa<Instruction>(Val: I))
2603 continue;
2604
2605 auto OISIt = OpSafeForPHIOfOps.find(Val: {I, CacheIdx});
2606 if (OISIt != OpSafeForPHIOfOps.end())
2607 return OISIt->second;
2608
2609 // Keep walking until we either dominate the phi block, or hit a phi, or run
2610 // out of things to check.
2611 if (DT->properlyDominates(A: getBlockForValue(V: I), B: PHIBlock)) {
2612 OpSafeForPHIOfOps.insert(KV: {{I, CacheIdx}, true});
2613 continue;
2614 }
2615 // PHI in the same block.
2616 if (isa<PHINode>(Val: I) && getBlockForValue(V: I) == PHIBlock) {
2617 OpSafeForPHIOfOps.insert(KV: {{I, CacheIdx}, false});
2618 return false;
2619 }
2620
2621 auto *OrigI = cast<Instruction>(Val: I);
2622 // When we hit an instruction that reads memory (load, call, etc), we must
2623 // consider any store that may happen in the loop. For now, we assume the
2624 // worst: there is a store in the loop that alias with this read.
2625 // The case where the load is outside the loop is already covered by the
2626 // dominator check above.
2627 // TODO: relax this condition
2628 if (OrigI->mayReadFromMemory())
2629 return false;
2630
2631 // Check the operands of the current instruction.
2632 for (auto *Op : OrigI->operand_values()) {
2633 if (!isa<Instruction>(Val: Op))
2634 continue;
2635 // Stop now if we find an unsafe operand.
2636 auto OISIt = OpSafeForPHIOfOps.find(Val: {OrigI, CacheIdx});
2637 if (OISIt != OpSafeForPHIOfOps.end()) {
2638 if (!OISIt->second) {
2639 OpSafeForPHIOfOps.insert(KV: {{I, CacheIdx}, false});
2640 return false;
2641 }
2642 continue;
2643 }
2644 if (!Visited.insert(Ptr: Op).second)
2645 continue;
2646 Worklist.push_back(Elt: cast<Instruction>(Val: Op));
2647 }
2648 }
2649 OpSafeForPHIOfOps.insert(KV: {{V, CacheIdx}, true});
2650 return true;
2651}
2652
2653// Try to find a leader for instruction TransInst, which is a phi translated
2654// version of something in our original program. Visited is used to ensure we
2655// don't infinite loop during translations of cycles. OrigInst is the
2656// instruction in the original program, and PredBB is the predecessor we
2657// translated it through.
2658Value *NewGVN::findLeaderForInst(Instruction *TransInst,
2659 SmallPtrSetImpl<Value *> &Visited,
2660 MemoryAccess *MemAccess, Instruction *OrigInst,
2661 BasicBlock *PredBB) {
2662 unsigned IDFSNum = InstrToDFSNum(V: OrigInst);
2663 // Make sure it's marked as a temporary instruction.
2664 AllTempInstructions.insert(V: TransInst);
2665 // and make sure anything that tries to add it's DFS number is
2666 // redirected to the instruction we are making a phi of ops
2667 // for.
2668 TempToBlock.insert(KV: {TransInst, PredBB});
2669 InstrDFS.insert(KV: {TransInst, IDFSNum});
2670
2671 auto Res = performSymbolicEvaluation(I: TransInst, Visited);
2672 const Expression *E = Res.Expr;
2673 addAdditionalUsers(Res, User: OrigInst);
2674 InstrDFS.erase(Val: TransInst);
2675 AllTempInstructions.erase(V: TransInst);
2676 TempToBlock.erase(Val: TransInst);
2677 if (MemAccess)
2678 TempToMemory.erase(Val: TransInst);
2679 if (!E)
2680 return nullptr;
2681 auto *FoundVal = findPHIOfOpsLeader(E, OrigInst, PredBB);
2682 if (!FoundVal) {
2683 ExpressionToPhiOfOps[E].insert(Ptr: OrigInst);
2684 LLVM_DEBUG(dbgs() << "Cannot find phi of ops operand for " << *TransInst
2685 << " in block " << getBlockName(PredBB) << "\n");
2686 return nullptr;
2687 }
2688 if (auto *SI = dyn_cast<StoreInst>(Val: FoundVal))
2689 FoundVal = SI->getValueOperand();
2690 return FoundVal;
2691}
2692
2693// When we see an instruction that is an op of phis, generate the equivalent phi
2694// of ops form.
2695const Expression *
2696NewGVN::makePossiblePHIOfOps(Instruction *I,
2697 SmallPtrSetImpl<Value *> &Visited) {
2698 if (!okayForPHIOfOps(Opts, I))
2699 return nullptr;
2700
2701 if (!Visited.insert(Ptr: I).second)
2702 return nullptr;
2703 // For now, we require the instruction be cycle free because we don't
2704 // *always* create a phi of ops for instructions that could be done as phi
2705 // of ops, we only do it if we think it is useful. If we did do it all the
2706 // time, we could remove the cycle free check.
2707 if (!isCycleFree(I))
2708 return nullptr;
2709
2710 // TODO: We don't do phi translation on memory accesses because it's
2711 // complicated. For a load, we'd need to be able to simulate a new memoryuse,
2712 // which we don't have a good way of doing ATM.
2713 auto *MemAccess = getMemoryAccess(I);
2714 // If the memory operation is defined by a memory operation this block that
2715 // isn't a MemoryPhi, transforming the pointer backwards through a scalar phi
2716 // can't help, as it would still be killed by that memory operation.
2717 if (MemAccess && !isa<MemoryPhi>(Val: MemAccess->getDefiningAccess()) &&
2718 MemAccess->getDefiningAccess()->getBlock() == I->getParent())
2719 return nullptr;
2720
2721 // Convert op of phis to phi of ops
2722 SmallPtrSet<const Value *, 10> VisitedOps;
2723 SmallVector<Value *, 4> Ops(I->operand_values());
2724 BasicBlock *SamePHIBlock = nullptr;
2725 PHINode *OpPHI = nullptr;
2726 if (!DebugCounter::shouldExecute(Counter&: PHIOfOpsCounter))
2727 return nullptr;
2728 for (auto *Op : Ops) {
2729 if (!isa<PHINode>(Val: Op)) {
2730 auto *ValuePHI = RealToTemp.lookup(Val: Op);
2731 if (!ValuePHI)
2732 continue;
2733 LLVM_DEBUG(dbgs() << "Found possible dependent phi of ops\n");
2734 Op = ValuePHI;
2735 }
2736 OpPHI = cast<PHINode>(Val: Op);
2737 if (!SamePHIBlock) {
2738 SamePHIBlock = getBlockForValue(V: OpPHI);
2739 } else if (SamePHIBlock != getBlockForValue(V: OpPHI)) {
2740 LLVM_DEBUG(
2741 dbgs()
2742 << "PHIs for operands are not all in the same block, aborting\n");
2743 return nullptr;
2744 }
2745 // No point in doing this for one-operand phis.
2746 // Since all PHIs for operands must be in the same block, then they must
2747 // have the same number of operands so we can just abort.
2748 if (OpPHI->getNumOperands() == 1)
2749 return nullptr;
2750 }
2751
2752 if (!OpPHI)
2753 return nullptr;
2754
2755 SmallVector<ValPair, 4> PHIOps;
2756 SmallPtrSet<Value *, 4> Deps;
2757 auto *PHIBlock = getBlockForValue(V: OpPHI);
2758 RevisitOnReachabilityChange[PHIBlock].reset(Idx: InstrToDFSNum(V: I));
2759 for (unsigned PredNum = 0; PredNum < OpPHI->getNumOperands(); ++PredNum) {
2760 auto *PredBB = OpPHI->getIncomingBlock(i: PredNum);
2761 Value *FoundVal = nullptr;
2762 SmallPtrSet<Value *, 4> CurrentDeps;
2763 // We could just skip unreachable edges entirely but it's tricky to do
2764 // with rewriting existing phi nodes.
2765 if (ReachableEdges.count(V: {PredBB, PHIBlock})) {
2766 // Clone the instruction, create an expression from it that is
2767 // translated back into the predecessor, and see if we have a leader.
2768 Instruction *ValueOp = I->clone();
2769 // Emit the temporal instruction in the predecessor basic block where the
2770 // corresponding value is defined.
2771 ValueOp->insertBefore(InsertPos: PredBB->getTerminator()->getIterator());
2772 if (MemAccess)
2773 TempToMemory.insert(KV: {ValueOp, MemAccess});
2774 bool SafeForPHIOfOps = true;
2775 VisitedOps.clear();
2776 for (auto &Op : ValueOp->operands()) {
2777 auto *OrigOp = &*Op;
2778 // When these operand changes, it could change whether there is a
2779 // leader for us or not, so we have to add additional users.
2780 if (isa<PHINode>(Val: Op)) {
2781 Op = Op->DoPHITranslation(CurBB: PHIBlock, PredBB);
2782 if (Op != OrigOp && Op != I)
2783 CurrentDeps.insert(Ptr: Op);
2784 } else if (auto *ValuePHI = RealToTemp.lookup(Val: Op)) {
2785 if (getBlockForValue(V: ValuePHI) == PHIBlock)
2786 Op = ValuePHI->getIncomingValueForBlock(BB: PredBB);
2787 }
2788 // If we phi-translated the op, it must be safe.
2789 SafeForPHIOfOps =
2790 SafeForPHIOfOps &&
2791 (Op != OrigOp || OpIsSafeForPHIOfOps(V: Op, PHIBlock, Visited&: VisitedOps));
2792 }
2793 // FIXME: For those things that are not safe we could generate
2794 // expressions all the way down, and see if this comes out to a
2795 // constant. For anything where that is true, and unsafe, we should
2796 // have made a phi-of-ops (or value numbered it equivalent to something)
2797 // for the pieces already.
2798 FoundVal = !SafeForPHIOfOps ? nullptr
2799 : findLeaderForInst(TransInst: ValueOp, Visited,
2800 MemAccess, OrigInst: I, PredBB);
2801 ValueOp->eraseFromParent();
2802 if (!FoundVal) {
2803 // We failed to find a leader for the current ValueOp, but this might
2804 // change in case of the translated operands change.
2805 if (SafeForPHIOfOps)
2806 for (auto *Dep : CurrentDeps)
2807 addAdditionalUsers(To: Dep, User: I);
2808
2809 return nullptr;
2810 }
2811 Deps.insert_range(R&: CurrentDeps);
2812 } else {
2813 LLVM_DEBUG(dbgs() << "Skipping phi of ops operand for incoming block "
2814 << getBlockName(PredBB)
2815 << " because the block is unreachable\n");
2816 FoundVal = PoisonValue::get(T: I->getType());
2817 RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(V: I));
2818 }
2819
2820 PHIOps.push_back(Elt: {FoundVal, PredBB});
2821 LLVM_DEBUG(dbgs() << "Found phi of ops operand " << *FoundVal << " in "
2822 << getBlockName(PredBB) << "\n");
2823 }
2824 for (auto *Dep : Deps)
2825 addAdditionalUsers(To: Dep, User: I);
2826 sortPHIOps(Ops: PHIOps);
2827 auto *E = performSymbolicPHIEvaluation(PHIOps, I, PHIBlock);
2828 if (isa<ConstantExpression>(Val: E) || isa<VariableExpression>(Val: E)) {
2829 LLVM_DEBUG(
2830 dbgs()
2831 << "Not creating real PHI of ops because it simplified to existing "
2832 "value or constant\n");
2833 // We have leaders for all operands, but do not create a real PHI node with
2834 // those leaders as operands, so the link between the operands and the
2835 // PHI-of-ops is not materialized in the IR. If any of those leaders
2836 // changes, the PHI-of-op may change also, so we need to add the operands as
2837 // additional users.
2838 for (auto &O : PHIOps)
2839 addAdditionalUsers(To: O.first, User: I);
2840
2841 return E;
2842 }
2843 auto *ValuePHI = RealToTemp.lookup(Val: I);
2844 bool NewPHI = false;
2845 if (!ValuePHI) {
2846 ValuePHI =
2847 PHINode::Create(Ty: I->getType(), NumReservedValues: OpPHI->getNumOperands(), NameStr: "phiofops");
2848 addPhiOfOps(Op: ValuePHI, BB: PHIBlock, ExistingValue: I);
2849 NewPHI = true;
2850 NumGVNPHIOfOpsCreated++;
2851 }
2852 if (NewPHI) {
2853 for (auto PHIOp : PHIOps)
2854 ValuePHI->addIncoming(V: PHIOp.first, BB: PHIOp.second);
2855 } else {
2856 TempToBlock[ValuePHI] = PHIBlock;
2857 unsigned int i = 0;
2858 for (auto PHIOp : PHIOps) {
2859 ValuePHI->setIncomingValue(i, V: PHIOp.first);
2860 ValuePHI->setIncomingBlock(i, BB: PHIOp.second);
2861 ++i;
2862 }
2863 }
2864 RevisitOnReachabilityChange[PHIBlock].set(InstrToDFSNum(V: I));
2865 LLVM_DEBUG(dbgs() << "Created phi of ops " << *ValuePHI << " for " << *I
2866 << "\n");
2867
2868 return E;
2869}
2870
2871// The algorithm initially places the values of the routine in the TOP
2872// congruence class. The leader of TOP is the undetermined value `poison`.
2873// When the algorithm has finished, values still in TOP are unreachable.
2874void NewGVN::initializeCongruenceClasses(Function &F) {
2875 NextCongruenceNum = 0;
2876
2877 // Note that even though we use the live on entry def as a representative
2878 // MemoryAccess, it is *not* the same as the actual live on entry def. We
2879 // have no real equivalent to poison for MemoryAccesses, and so we really
2880 // should be checking whether the MemoryAccess is top if we want to know if it
2881 // is equivalent to everything. Otherwise, what this really signifies is that
2882 // the access "it reaches all the way back to the beginning of the function"
2883
2884 // Initialize all other instructions to be in TOP class.
2885 TOPClass = createCongruenceClass(Leader: nullptr, E: nullptr);
2886 TOPClass->setMemoryLeader(MSSA->getLiveOnEntryDef());
2887 // The live on entry def gets put into it's own class
2888 MemoryAccessToClass[MSSA->getLiveOnEntryDef()] =
2889 createMemoryClass(MA: MSSA->getLiveOnEntryDef());
2890
2891 for (auto *DTN : nodes(G: DT)) {
2892 BasicBlock *BB = DTN->getBlock();
2893 // All MemoryAccesses are equivalent to live on entry to start. They must
2894 // be initialized to something so that initial changes are noticed. For
2895 // the maximal answer, we initialize them all to be the same as
2896 // liveOnEntry.
2897 auto *MemoryBlockDefs = MSSA->getBlockDefs(BB);
2898 if (MemoryBlockDefs)
2899 for (const auto &Def : *MemoryBlockDefs) {
2900 MemoryAccessToClass[&Def] = TOPClass;
2901 auto *MD = dyn_cast<MemoryDef>(Val: &Def);
2902 // Insert the memory phis into the member list.
2903 if (!MD) {
2904 const MemoryPhi *MP = cast<MemoryPhi>(Val: &Def);
2905 TOPClass->memory_insert(M: MP);
2906 MemoryPhiState.insert(KV: {MP, MPS_TOP});
2907 }
2908
2909 if (MD && isa<StoreInst>(Val: MD->getMemoryInst()))
2910 TOPClass->incStoreCount();
2911 }
2912
2913 // FIXME: This is trying to discover which instructions are uses of phi
2914 // nodes. We should move this into one of the myriad of places that walk
2915 // all the operands already.
2916 for (auto &I : *BB) {
2917 if (isa<PHINode>(Val: &I))
2918 for (auto *U : I.users())
2919 if (auto *UInst = dyn_cast<Instruction>(Val: U))
2920 if (InstrToDFSNum(V: UInst) != 0 && okayForPHIOfOps(Opts, I: UInst))
2921 PHINodeUses.insert(Ptr: UInst);
2922 // Don't insert void terminators into the class. We don't value number
2923 // them, and they just end up sitting in TOP.
2924 if (I.isTerminator() && I.getType()->isVoidTy())
2925 continue;
2926 TOPClass->insert(M: &I);
2927 ValueToClass[&I] = TOPClass;
2928 }
2929 }
2930
2931 // Initialize arguments to be in their own unique congruence classes
2932 for (auto &FA : F.args())
2933 createSingletonCongruenceClass(Member: &FA);
2934}
2935
2936void NewGVN::cleanupTables() {
2937 for (CongruenceClass *&CC : CongruenceClasses) {
2938 LLVM_DEBUG(dbgs() << "Congruence class " << CC->getID() << " has "
2939 << CC->size() << " members\n");
2940 // Make sure we delete the congruence class (probably worth switching to
2941 // a unique_ptr at some point.
2942 delete CC;
2943 CC = nullptr;
2944 }
2945
2946 // Destroy the value expressions
2947 SmallVector<Instruction *, 8> TempInst(AllTempInstructions.begin(),
2948 AllTempInstructions.end());
2949 AllTempInstructions.clear();
2950
2951 // We have to drop all references for everything first, so there are no uses
2952 // left as we delete them.
2953 for (auto *I : TempInst) {
2954 I->dropAllReferences();
2955 }
2956
2957 while (!TempInst.empty()) {
2958 auto *I = TempInst.pop_back_val();
2959 I->deleteValue();
2960 }
2961
2962 ValueToClass.clear();
2963 ArgRecycler.clear(ExpressionAllocator);
2964 ExpressionAllocator.Reset();
2965 CongruenceClasses.clear();
2966 ExpressionToClass.clear();
2967 ValueToExpression.clear();
2968 RealToTemp.clear();
2969 AdditionalUsers.clear();
2970 ExpressionToPhiOfOps.clear();
2971 TempToBlock.clear();
2972 TempToMemory.clear();
2973 PHINodeUses.clear();
2974 OpSafeForPHIOfOps.clear();
2975 ReachableBlocks.clear();
2976 ReachableEdges.clear();
2977#ifndef NDEBUG
2978 ProcessedCount.clear();
2979#endif
2980 InstrDFS.clear();
2981 InstructionsToErase.clear();
2982 DFSToInstr.clear();
2983 BlockInstRange.clear();
2984 TouchedInstructions.clear();
2985 MemoryAccessToClass.clear();
2986 PredicateToUsers.clear();
2987 MemoryToUsers.clear();
2988 RevisitOnReachabilityChange.clear();
2989 PredicateSwapChoice.clear();
2990}
2991
2992// Assign local DFS number mapping to instructions, and leave space for Value
2993// PHI's.
2994std::pair<unsigned, unsigned> NewGVN::assignDFSNumbers(BasicBlock *B,
2995 unsigned Start) {
2996 unsigned End = Start;
2997 if (MemoryAccess *MemPhi = getMemoryAccess(BB: B)) {
2998 InstrDFS[MemPhi] = End++;
2999 DFSToInstr.emplace_back(Args&: MemPhi);
3000 }
3001
3002 // Then the real block goes next.
3003 for (auto &I : *B) {
3004 // There's no need to call isInstructionTriviallyDead more than once on
3005 // an instruction. Therefore, once we know that an instruction is dead
3006 // we change its DFS number so that it doesn't get value numbered.
3007 if (isInstructionTriviallyDead(I: &I, TLI)) {
3008 InstrDFS[&I] = 0;
3009 LLVM_DEBUG(dbgs() << "Skipping trivially dead instruction " << I << "\n");
3010 salvageDebugInfo(I);
3011 markInstructionForDeletion(&I);
3012 continue;
3013 }
3014 if (isa<PHINode>(Val: &I))
3015 RevisitOnReachabilityChange[B].set(End);
3016 InstrDFS[&I] = End++;
3017 DFSToInstr.emplace_back(Args: &I);
3018 }
3019
3020 // All of the range functions taken half-open ranges (open on the end side).
3021 // So we do not subtract one from count, because at this point it is one
3022 // greater than the last instruction.
3023 return std::make_pair(x&: Start, y&: End);
3024}
3025
3026void NewGVN::updateProcessedCount(const Value *V) {
3027#ifndef NDEBUG
3028 assert(++ProcessedCount[V] < 100 &&
3029 "Seem to have processed the same Value a lot");
3030#endif
3031}
3032
3033// Evaluate MemoryPhi nodes symbolically, just like PHI nodes
3034void NewGVN::valueNumberMemoryPhi(MemoryPhi *MP) {
3035 // If all the arguments are the same, the MemoryPhi has the same value as the
3036 // argument. Filter out unreachable blocks and self phis from our operands.
3037 // TODO: We could do cycle-checking on the memory phis to allow valueizing for
3038 // self-phi checking.
3039 const BasicBlock *PHIBlock = MP->getBlock();
3040 auto Filtered = make_filter_range(Range: MP->operands(), Pred: [&](const Use &U) {
3041 return cast<MemoryAccess>(Val: U) != MP &&
3042 !isMemoryAccessTOP(MA: cast<MemoryAccess>(Val: U)) &&
3043 ReachableEdges.count(V: {MP->getIncomingBlock(U), PHIBlock});
3044 });
3045 // If all that is left is nothing, our memoryphi is poison. We keep it as
3046 // InitialClass. Note: The only case this should happen is if we have at
3047 // least one self-argument.
3048 if (Filtered.begin() == Filtered.end()) {
3049 if (setMemoryClass(From: MP, NewClass: TOPClass))
3050 markMemoryUsersTouched(MA: MP);
3051 return;
3052 }
3053
3054 // Transform the remaining operands into operand leaders.
3055 // FIXME: mapped_iterator should have a range version.
3056 auto LookupFunc = [&](const Use &U) {
3057 return lookupMemoryLeader(MA: cast<MemoryAccess>(Val: U));
3058 };
3059 auto MappedBegin = map_iterator(I: Filtered.begin(), F: LookupFunc);
3060 auto MappedEnd = map_iterator(I: Filtered.end(), F: LookupFunc);
3061
3062 // and now check if all the elements are equal.
3063 // Sadly, we can't use std::equals since these are random access iterators.
3064 const auto *AllSameValue = *MappedBegin;
3065 ++MappedBegin;
3066 bool AllEqual = std::all_of(
3067 first: MappedBegin, last: MappedEnd,
3068 pred: [&AllSameValue](const MemoryAccess *V) { return V == AllSameValue; });
3069
3070 if (AllEqual)
3071 LLVM_DEBUG(dbgs() << "Memory Phi value numbered to " << *AllSameValue
3072 << "\n");
3073 else
3074 LLVM_DEBUG(dbgs() << "Memory Phi value numbered to itself\n");
3075 // If it's equal to something, it's in that class. Otherwise, it has to be in
3076 // a class where it is the leader (other things may be equivalent to it, but
3077 // it needs to start off in its own class, which means it must have been the
3078 // leader, and it can't have stopped being the leader because it was never
3079 // removed).
3080 CongruenceClass *CC =
3081 AllEqual ? getMemoryClass(MA: AllSameValue) : ensureLeaderOfMemoryClass(MA: MP);
3082 auto OldState = MemoryPhiState.lookup(Val: MP);
3083 assert(OldState != MPS_Invalid && "Invalid memory phi state");
3084 auto NewState = AllEqual ? MPS_Equivalent : MPS_Unique;
3085 MemoryPhiState[MP] = NewState;
3086 if (setMemoryClass(From: MP, NewClass: CC) || OldState != NewState)
3087 markMemoryUsersTouched(MA: MP);
3088}
3089
3090// Value number a single instruction, symbolically evaluating, performing
3091// congruence finding, and updating mappings.
3092void NewGVN::valueNumberInstruction(Instruction *I) {
3093 LLVM_DEBUG(dbgs() << "Processing instruction " << *I << "\n");
3094 if (!I->isTerminator()) {
3095 const Expression *Symbolized = nullptr;
3096 SmallPtrSet<Value *, 2> Visited;
3097 if (DebugCounter::shouldExecute(Counter&: VNCounter)) {
3098 auto Res = performSymbolicEvaluation(I, Visited);
3099 Symbolized = Res.Expr;
3100 addAdditionalUsers(Res, User: I);
3101
3102 // Make a phi of ops if necessary
3103 if (Symbolized && !isa<ConstantExpression>(Val: Symbolized) &&
3104 !isa<VariableExpression>(Val: Symbolized) && PHINodeUses.count(Ptr: I)) {
3105 auto *PHIE = makePossiblePHIOfOps(I, Visited);
3106 // If we created a phi of ops, use it.
3107 // If we couldn't create one, make sure we don't leave one lying around
3108 if (PHIE) {
3109 Symbolized = PHIE;
3110 } else if (auto *Op = RealToTemp.lookup(Val: I)) {
3111 removePhiOfOps(I, PHITemp: Op);
3112 }
3113 }
3114 } else {
3115 // Mark the instruction as unused so we don't value number it again.
3116 InstrDFS[I] = 0;
3117 }
3118 // If we couldn't come up with a symbolic expression, use the unknown
3119 // expression
3120 if (Symbolized == nullptr)
3121 Symbolized = createUnknownExpression(I);
3122 performCongruenceFinding(I, E: Symbolized);
3123 } else {
3124 // Handle terminators that return values. All of them produce values we
3125 // don't currently understand. We don't place non-value producing
3126 // terminators in a class.
3127 if (!I->getType()->isVoidTy()) {
3128 auto *Symbolized = createUnknownExpression(I);
3129 performCongruenceFinding(I, E: Symbolized);
3130 }
3131 processOutgoingEdges(TI: I, B: I->getParent());
3132 }
3133}
3134
3135// Check if there is a path, using single or equal argument phi nodes, from
3136// First to Second.
3137bool NewGVN::singleReachablePHIPath(
3138 SmallPtrSet<const MemoryAccess *, 8> &Visited, const MemoryAccess *First,
3139 const MemoryAccess *Second) const {
3140 if (First == Second)
3141 return true;
3142 if (MSSA->isLiveOnEntryDef(MA: First))
3143 return false;
3144
3145 // This is not perfect, but as we're just verifying here, we can live with
3146 // the loss of precision. The real solution would be that of doing strongly
3147 // connected component finding in this routine, and it's probably not worth
3148 // the complexity for the time being. So, we just keep a set of visited
3149 // MemoryAccess and return true when we hit a cycle.
3150 if (!Visited.insert(Ptr: First).second)
3151 return true;
3152
3153 const auto *EndDef = First;
3154 for (const auto *ChainDef : optimized_def_chain(MA: First)) {
3155 if (ChainDef == Second)
3156 return true;
3157 if (MSSA->isLiveOnEntryDef(MA: ChainDef))
3158 return false;
3159 EndDef = ChainDef;
3160 }
3161 auto *MP = cast<MemoryPhi>(Val: EndDef);
3162 auto ReachableOperandPred = [&](const Use &U) {
3163 return ReachableEdges.count(V: {MP->getIncomingBlock(U), MP->getBlock()});
3164 };
3165 auto FilteredPhiArgs =
3166 make_filter_range(Range: MP->operands(), Pred: ReachableOperandPred);
3167 SmallVector<const Value *, 32> OperandList(FilteredPhiArgs);
3168 bool Okay = all_equal(Range&: OperandList);
3169 if (Okay)
3170 return singleReachablePHIPath(Visited, First: cast<MemoryAccess>(Val: OperandList[0]),
3171 Second);
3172 return false;
3173}
3174
3175// Verify the that the memory equivalence table makes sense relative to the
3176// congruence classes. Note that this checking is not perfect, and is currently
3177// subject to very rare false negatives. It is only useful for
3178// testing/debugging.
3179void NewGVN::verifyMemoryCongruency() const {
3180#ifndef NDEBUG
3181 // Verify that the memory table equivalence and memory member set match
3182 for (const auto *CC : CongruenceClasses) {
3183 if (CC == TOPClass || CC->isDead())
3184 continue;
3185 if (CC->getStoreCount() != 0) {
3186 assert((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) &&
3187 "Any class with a store as a leader should have a "
3188 "representative stored value");
3189 assert(CC->getMemoryLeader() &&
3190 "Any congruence class with a store should have a "
3191 "representative access");
3192 }
3193
3194 if (CC->getMemoryLeader())
3195 assert(MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
3196 "Representative MemoryAccess does not appear to be reverse "
3197 "mapped properly");
3198 for (const auto *M : CC->memory())
3199 assert(MemoryAccessToClass.lookup(M) == CC &&
3200 "Memory member does not appear to be reverse mapped properly");
3201 }
3202
3203 // Anything equivalent in the MemoryAccess table should be in the same
3204 // congruence class.
3205
3206 // Filter out the unreachable and trivially dead entries, because they may
3207 // never have been updated if the instructions were not processed.
3208 auto ReachableAccessPred =
3209 [&](const std::pair<const MemoryAccess *, CongruenceClass *> Pair) {
3210 bool Result = ReachableBlocks.count(Pair.first->getBlock());
3211 if (!Result || MSSA->isLiveOnEntryDef(Pair.first) ||
3212 MemoryToDFSNum(Pair.first) == 0)
3213 return false;
3214 if (auto *MemDef = dyn_cast<MemoryDef>(Pair.first))
3215 return !isInstructionTriviallyDead(MemDef->getMemoryInst());
3216
3217 // We could have phi nodes which operands are all trivially dead,
3218 // so we don't process them.
3219 if (auto *MemPHI = dyn_cast<MemoryPhi>(Pair.first)) {
3220 for (const auto &U : MemPHI->incoming_values()) {
3221 if (auto *I = dyn_cast<Instruction>(&*U)) {
3222 if (!isInstructionTriviallyDead(I))
3223 return true;
3224 }
3225 }
3226 return false;
3227 }
3228
3229 return true;
3230 };
3231
3232 auto Filtered = make_filter_range(MemoryAccessToClass, ReachableAccessPred);
3233 for (auto KV : Filtered) {
3234 if (auto *FirstMUD = dyn_cast<MemoryUseOrDef>(KV.first)) {
3235 auto *SecondMUD = dyn_cast<MemoryUseOrDef>(KV.second->getMemoryLeader());
3236 if (FirstMUD && SecondMUD) {
3237 SmallPtrSet<const MemoryAccess *, 8> VisitedMAS;
3238 assert((singleReachablePHIPath(VisitedMAS, FirstMUD, SecondMUD) ||
3239 ValueToClass.lookup(FirstMUD->getMemoryInst()) ==
3240 ValueToClass.lookup(SecondMUD->getMemoryInst())) &&
3241 "The instructions for these memory operations should have "
3242 "been in the same congruence class or reachable through"
3243 "a single argument phi");
3244 }
3245 } else if (auto *FirstMP = dyn_cast<MemoryPhi>(KV.first)) {
3246 // We can only sanely verify that MemoryDefs in the operand list all have
3247 // the same class.
3248 auto ReachableOperandPred = [&](const Use &U) {
3249 return ReachableEdges.count(
3250 {FirstMP->getIncomingBlock(U), FirstMP->getBlock()}) &&
3251 isa<MemoryDef>(U);
3252 };
3253 // All arguments should in the same class, ignoring unreachable arguments
3254 auto FilteredPhiArgs =
3255 make_filter_range(FirstMP->operands(), ReachableOperandPred);
3256 SmallVector<const CongruenceClass *, 16> PhiOpClasses;
3257 std::transform(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
3258 std::back_inserter(PhiOpClasses), [&](const Use &U) {
3259 const MemoryDef *MD = cast<MemoryDef>(U);
3260 return ValueToClass.lookup(MD->getMemoryInst());
3261 });
3262 assert(all_equal(PhiOpClasses) &&
3263 "All MemoryPhi arguments should be in the same class");
3264 }
3265 }
3266#endif
3267}
3268
3269// Verify that the sparse propagation we did actually found the maximal fixpoint
3270// We do this by storing the value to class mapping, touching all instructions,
3271// and redoing the iteration to see if anything changed.
3272void NewGVN::verifyIterationSettled(Function &F) {
3273#ifndef NDEBUG
3274 LLVM_DEBUG(dbgs() << "Beginning iteration verification\n");
3275 if (DebugCounter::isCounterSet(VNCounter))
3276 DebugCounter::setCounterState(VNCounter, StartingVNCounter);
3277
3278 // Note that we have to store the actual classes, as we may change existing
3279 // classes during iteration. This is because our memory iteration propagation
3280 // is not perfect, and so may waste a little work. But it should generate
3281 // exactly the same congruence classes we have now, with different IDs.
3282 std::map<const Value *, CongruenceClass> BeforeIteration;
3283
3284 for (auto &KV : ValueToClass) {
3285 if (auto *I = dyn_cast<Instruction>(KV.first))
3286 // Skip unused/dead instructions.
3287 if (InstrToDFSNum(I) == 0)
3288 continue;
3289 BeforeIteration.insert({KV.first, *KV.second});
3290 }
3291
3292 TouchedInstructions.set();
3293 TouchedInstructions.reset(0);
3294 OpSafeForPHIOfOps.clear();
3295 CacheIdx = 0;
3296 iterateTouchedInstructions();
3297 DenseSet<std::pair<const CongruenceClass *, const CongruenceClass *>>
3298 EqualClasses;
3299 for (const auto &KV : ValueToClass) {
3300 if (auto *I = dyn_cast<Instruction>(KV.first))
3301 // Skip unused/dead instructions.
3302 if (InstrToDFSNum(I) == 0)
3303 continue;
3304 // We could sink these uses, but i think this adds a bit of clarity here as
3305 // to what we are comparing.
3306 auto *BeforeCC = &BeforeIteration.find(KV.first)->second;
3307 auto *AfterCC = KV.second;
3308 // Note that the classes can't change at this point, so we memoize the set
3309 // that are equal.
3310 if (!EqualClasses.count({BeforeCC, AfterCC})) {
3311 assert(BeforeCC->isEquivalentTo(AfterCC) &&
3312 "Value number changed after main loop completed!");
3313 EqualClasses.insert({BeforeCC, AfterCC});
3314 }
3315 }
3316#endif
3317}
3318
3319// Verify that for each store expression in the expression to class mapping,
3320// only the latest appears, and multiple ones do not appear.
3321// Because loads do not use the stored value when doing equality with stores,
3322// if we don't erase the old store expressions from the table, a load can find
3323// a no-longer valid StoreExpression.
3324void NewGVN::verifyStoreExpressions() const {
3325#ifndef NDEBUG
3326 // This is the only use of this, and it's not worth defining a complicated
3327 // densemapinfo hash/equality function for it.
3328 std::set<
3329 std::pair<const Value *,
3330 std::tuple<const Value *, const CongruenceClass *, Value *>>>
3331 StoreExpressionSet;
3332 for (const auto &KV : ExpressionToClass) {
3333 if (auto *SE = dyn_cast<StoreExpression>(KV.first)) {
3334 // Make sure a version that will conflict with loads is not already there
3335 auto Res = StoreExpressionSet.insert(
3336 {SE->getOperand(0), std::make_tuple(SE->getMemoryLeader(), KV.second,
3337 SE->getStoredValue())});
3338 bool Okay = Res.second;
3339 // It's okay to have the same expression already in there if it is
3340 // identical in nature.
3341 // This can happen when the leader of the stored value changes over time.
3342 if (!Okay)
3343 Okay = (std::get<1>(Res.first->second) == KV.second) &&
3344 (lookupOperandLeader(std::get<2>(Res.first->second)) ==
3345 lookupOperandLeader(SE->getStoredValue()));
3346 assert(Okay && "Stored expression conflict exists in expression table");
3347 auto *ValueExpr = ValueToExpression.lookup(SE->getStoreInst());
3348 assert(ValueExpr && ValueExpr->equals(*SE) &&
3349 "StoreExpression in ExpressionToClass is not latest "
3350 "StoreExpression for value");
3351 }
3352 }
3353#endif
3354}
3355
3356// This is the main value numbering loop, it iterates over the initial touched
3357// instruction set, propagating value numbers, marking things touched, etc,
3358// until the set of touched instructions is completely empty.
3359void NewGVN::iterateTouchedInstructions() {
3360 uint64_t Iterations = 0;
3361 // Figure out where touchedinstructions starts
3362 int FirstInstr = TouchedInstructions.find_first();
3363 // Nothing set, nothing to iterate, just return.
3364 if (FirstInstr == -1)
3365 return;
3366 const BasicBlock *LastBlock = getBlockForValue(V: InstrFromDFSNum(DFSNum: FirstInstr));
3367 while (TouchedInstructions.any()) {
3368 ++Iterations;
3369 // Walk through all the instructions in all the blocks in RPO.
3370 // TODO: As we hit a new block, we should push and pop equalities into a
3371 // table lookupOperandLeader can use, to catch things PredicateInfo
3372 // might miss, like edge-only equivalences.
3373 for (unsigned InstrNum : TouchedInstructions.set_bits()) {
3374
3375 // This instruction was found to be dead. We don't bother looking
3376 // at it again.
3377 if (InstrNum == 0) {
3378 TouchedInstructions.reset(Idx: InstrNum);
3379 continue;
3380 }
3381
3382 Value *V = InstrFromDFSNum(DFSNum: InstrNum);
3383 const BasicBlock *CurrBlock = getBlockForValue(V);
3384
3385 // If we hit a new block, do reachability processing.
3386 if (CurrBlock != LastBlock) {
3387 LastBlock = CurrBlock;
3388 bool BlockReachable = ReachableBlocks.count(Ptr: CurrBlock);
3389 const auto &CurrInstRange = BlockInstRange.lookup(Val: CurrBlock);
3390
3391 // If it's not reachable, erase any touched instructions and move on.
3392 if (!BlockReachable) {
3393 TouchedInstructions.reset(I: CurrInstRange.first, E: CurrInstRange.second);
3394 LLVM_DEBUG(dbgs() << "Skipping instructions in block "
3395 << getBlockName(CurrBlock)
3396 << " because it is unreachable\n");
3397 continue;
3398 }
3399 // Use the appropriate cache for "OpIsSafeForPHIOfOps".
3400 CacheIdx = RPOOrdering.lookup(Val: DT->getNode(BB: CurrBlock)) - 1;
3401 updateProcessedCount(V: CurrBlock);
3402 }
3403 // Reset after processing (because we may mark ourselves as touched when
3404 // we propagate equalities).
3405 TouchedInstructions.reset(Idx: InstrNum);
3406
3407 if (auto *MP = dyn_cast<MemoryPhi>(Val: V)) {
3408 LLVM_DEBUG(dbgs() << "Processing MemoryPhi " << *MP << "\n");
3409 valueNumberMemoryPhi(MP);
3410 } else if (auto *I = dyn_cast<Instruction>(Val: V)) {
3411 valueNumberInstruction(I);
3412 } else {
3413 llvm_unreachable("Should have been a MemoryPhi or Instruction");
3414 }
3415 updateProcessedCount(V);
3416 }
3417 }
3418 NumGVNMaxIterations = std::max(a: NumGVNMaxIterations.getValue(), b: Iterations);
3419}
3420
3421// This is the main transformation entry point.
3422bool NewGVN::runGVN() {
3423 if (DebugCounter::isCounterSet(Info&: VNCounter))
3424 StartingVNCounter = DebugCounter::getCounterState(Info&: VNCounter);
3425 bool Changed = false;
3426 NumFuncArgs = F.arg_size();
3427 MSSAWalker = MSSA->getWalker();
3428 SingletonDeadExpression = new (ExpressionAllocator) DeadExpression();
3429
3430 // Count number of instructions for sizing of hash tables, and come
3431 // up with a global dfs numbering for instructions.
3432 unsigned ICount = 1;
3433 // Add an empty instruction to account for the fact that we start at 1
3434 DFSToInstr.emplace_back(Args: nullptr);
3435 // Note: Number the blocks in RPO to put every definition before its uses,
3436 // except for a PHI operand arriving along a back edge. A wrong order costs
3437 // iterations.
3438 ReversePostOrderTraversal<Function *> RPOT(&F);
3439 unsigned Counter = 0;
3440 for (BasicBlock *B : RPOT) {
3441 auto *Node = DT->getNode(BB: B);
3442 assert(Node && "RPO and Dominator tree should have same reachability");
3443 RPOOrdering[Node] = ++Counter;
3444 const auto &BlockRange = assignDFSNumbers(B, Start: ICount);
3445 BlockInstRange.insert(KV: {B, BlockRange});
3446 ICount += BlockRange.second - BlockRange.first;
3447 }
3448 initializeCongruenceClasses(F);
3449
3450 TouchedInstructions.resize(N: ICount);
3451 // Ensure we don't end up resizing the expressionToClass map, as
3452 // that can be quite expensive. At most, we have one expression per
3453 // instruction.
3454 ExpressionToClass.reserve(NumEntries: ICount);
3455
3456 // Initialize the touched instructions to include the entry block.
3457 const auto &InstRange = BlockInstRange.lookup(Val: &F.getEntryBlock());
3458 TouchedInstructions.set(I: InstRange.first, E: InstRange.second);
3459 LLVM_DEBUG(dbgs() << "Block " << getBlockName(&F.getEntryBlock())
3460 << " marked reachable\n");
3461 ReachableBlocks.insert(Ptr: &F.getEntryBlock());
3462 // Use index corresponding to entry block.
3463 CacheIdx = 0;
3464
3465 iterateTouchedInstructions();
3466 verifyMemoryCongruency();
3467 verifyIterationSettled(F);
3468 verifyStoreExpressions();
3469
3470 Changed |= eliminateInstructions(F);
3471
3472 // Delete all instructions marked for deletion.
3473 for (Instruction *ToErase : InstructionsToErase) {
3474 if (!ToErase->use_empty())
3475 ToErase->replaceAllUsesWith(V: PoisonValue::get(T: ToErase->getType()));
3476
3477 assert(ToErase->getParent() &&
3478 "BB containing ToErase deleted unexpectedly!");
3479 ToErase->eraseFromParent();
3480 }
3481 Changed |= !InstructionsToErase.empty();
3482
3483 // Delete all unreachable blocks.
3484 auto UnreachableBlockPred = [&](const BasicBlock &BB) {
3485 return !ReachableBlocks.count(Ptr: &BB);
3486 };
3487
3488 for (auto &BB : make_filter_range(Range&: F, Pred: UnreachableBlockPred)) {
3489 LLVM_DEBUG(dbgs() << "We believe block " << getBlockName(&BB)
3490 << " is unreachable\n");
3491 deleteInstructionsInBlock(&BB);
3492 Changed = true;
3493 }
3494
3495 cleanupTables();
3496 return Changed;
3497}
3498
3499struct NewGVN::ValueDFS {
3500 int DFSIn = 0;
3501 int DFSOut = 0;
3502 int LocalNum = 0;
3503
3504 // Only one of Def and U will be set.
3505 // The bool in the Def tells us whether the Def is the stored value of a
3506 // store.
3507 PointerIntPair<Value *, 1, bool> Def;
3508 Use *U = nullptr;
3509
3510 bool operator<(const ValueDFS &Other) const {
3511 // It's not enough that any given field be less than - we have sets
3512 // of fields that need to be evaluated together to give a proper ordering.
3513 // For example, if you have;
3514 // DFS (1, 3)
3515 // Val 0
3516 // DFS (1, 2)
3517 // Val 50
3518 // We want the second to be less than the first, but if we just go field
3519 // by field, we will get to Val 0 < Val 50 and say the first is less than
3520 // the second. We only want it to be less than if the DFS orders are equal.
3521 //
3522 // Each LLVM instruction only produces one value, and thus the lowest-level
3523 // differentiator that really matters for the stack (and what we use as a
3524 // replacement) is the local dfs number.
3525 // Everything else in the structure is instruction level, and only affects
3526 // the order in which we will replace operands of a given instruction.
3527 //
3528 // For a given instruction (IE things with equal dfsin, dfsout, localnum),
3529 // the order of replacement of uses does not matter.
3530 // IE given,
3531 // a = 5
3532 // b = a + a
3533 // When you hit b, you will have two valuedfs with the same dfsin, out, and
3534 // localnum.
3535 // The .val will be the same as well.
3536 // The .u's will be different.
3537 // You will replace both, and it does not matter what order you replace them
3538 // in (IE whether you replace operand 2, then operand 1, or operand 1, then
3539 // operand 2).
3540 // Similarly for the case of same dfsin, dfsout, localnum, but different
3541 // .val's
3542 // a = 5
3543 // b = 6
3544 // c = a + b
3545 // in c, we will a valuedfs for a, and one for b,with everything the same
3546 // but .val and .u.
3547 // It does not matter what order we replace these operands in.
3548 // You will always end up with the same IR, and this is guaranteed.
3549 return std::tie(args: DFSIn, args: DFSOut, args: LocalNum, args: Def, args: U) <
3550 std::tie(args: Other.DFSIn, args: Other.DFSOut, args: Other.LocalNum, args: Other.Def,
3551 args: Other.U);
3552 }
3553};
3554
3555// This function converts the set of members for a congruence class from values,
3556// to sets of defs and uses with associated DFS info. The total number of
3557// reachable uses for each value is stored in UseCount, and instructions that
3558// seem
3559// dead (have no non-dead uses) are stored in ProbablyDead.
3560void NewGVN::convertClassToDFSOrdered(
3561 const CongruenceClass &Dense, SmallVectorImpl<ValueDFS> &DFSOrderedSet,
3562 DenseMap<const Value *, unsigned int> &UseCounts,
3563 SmallPtrSetImpl<Instruction *> &ProbablyDead) const {
3564 for (auto *D : Dense) {
3565 // First add the value.
3566 BasicBlock *BB = getBlockForValue(V: D);
3567 // Constants are handled prior to ever calling this function, so
3568 // we should only be left with instructions as members.
3569 assert(BB && "Should have figured out a basic block for value");
3570 ValueDFS VDDef;
3571 DomTreeNode *DomNode = DT->getNode(BB);
3572 VDDef.DFSIn = DomNode->getDFSNumIn();
3573 VDDef.DFSOut = DomNode->getDFSNumOut();
3574 // If it's a store, use the leader of the value operand, if it's always
3575 // available, or the value operand. TODO: We could do dominance checks to
3576 // find a dominating leader, but not worth it ATM.
3577 if (auto *SI = dyn_cast<StoreInst>(Val: D)) {
3578 auto Leader = lookupOperandLeader(V: SI->getValueOperand());
3579 if (alwaysAvailable(V: Leader)) {
3580 VDDef.Def.setPointer(Leader);
3581 } else {
3582 VDDef.Def.setPointer(SI->getValueOperand());
3583 VDDef.Def.setInt(true);
3584 }
3585 } else {
3586 VDDef.Def.setPointer(D);
3587 }
3588 assert(isa<Instruction>(D) &&
3589 "The dense set member should always be an instruction");
3590 Instruction *Def = cast<Instruction>(Val: D);
3591 VDDef.LocalNum = InstrToDFSNum(V: D);
3592 DFSOrderedSet.push_back(Elt: VDDef);
3593 // If there is a phi node equivalent, add it
3594 if (auto *PN = RealToTemp.lookup(Val: Def)) {
3595 auto *PHIE =
3596 dyn_cast_or_null<PHIExpression>(Val: ValueToExpression.lookup(Val: Def));
3597 if (PHIE) {
3598 VDDef.Def.setInt(false);
3599 VDDef.Def.setPointer(PN);
3600 VDDef.LocalNum = 0;
3601 DFSOrderedSet.push_back(Elt: VDDef);
3602 }
3603 }
3604
3605 unsigned int UseCount = 0;
3606 // Now add the uses.
3607 for (auto &U : Def->uses()) {
3608 if (auto *I = dyn_cast<Instruction>(Val: U.getUser())) {
3609 // Don't try to replace into dead uses
3610 if (InstructionsToErase.count(Ptr: I))
3611 continue;
3612 ValueDFS VDUse;
3613 // Put the phi node uses in the incoming block.
3614 BasicBlock *IBlock;
3615 if (auto *P = dyn_cast<PHINode>(Val: I)) {
3616 IBlock = P->getIncomingBlock(U);
3617 // Make phi node users appear last in the incoming block
3618 // they are from.
3619 VDUse.LocalNum = InstrDFS.size() + 1;
3620 } else {
3621 IBlock = getBlockForValue(V: I);
3622 VDUse.LocalNum = InstrToDFSNum(V: I);
3623 }
3624
3625 // Skip uses in unreachable blocks, as we're going
3626 // to delete them.
3627 if (!ReachableBlocks.contains(Ptr: IBlock))
3628 continue;
3629
3630 DomTreeNode *DomNode = DT->getNode(BB: IBlock);
3631 VDUse.DFSIn = DomNode->getDFSNumIn();
3632 VDUse.DFSOut = DomNode->getDFSNumOut();
3633 VDUse.U = &U;
3634 ++UseCount;
3635 DFSOrderedSet.emplace_back(Args&: VDUse);
3636 }
3637 }
3638
3639 // If there are no uses, it's probably dead (but it may have side-effects,
3640 // so not definitely dead. Otherwise, store the number of uses so we can
3641 // track if it becomes dead later).
3642 if (UseCount == 0)
3643 ProbablyDead.insert(Ptr: Def);
3644 else
3645 UseCounts[Def] = UseCount;
3646 }
3647}
3648
3649// This function converts the set of members for a congruence class from values,
3650// to the set of defs for loads and stores, with associated DFS info.
3651void NewGVN::convertClassToLoadsAndStores(
3652 const CongruenceClass &Dense,
3653 SmallVectorImpl<ValueDFS> &LoadsAndStores) const {
3654 for (auto *D : Dense) {
3655 if (!isa<LoadInst>(Val: D) && !isa<StoreInst>(Val: D))
3656 continue;
3657
3658 BasicBlock *BB = getBlockForValue(V: D);
3659 ValueDFS VD;
3660 DomTreeNode *DomNode = DT->getNode(BB);
3661 VD.DFSIn = DomNode->getDFSNumIn();
3662 VD.DFSOut = DomNode->getDFSNumOut();
3663 VD.Def.setPointer(D);
3664
3665 // If it's an instruction, use the real local dfs number.
3666 if (auto *I = dyn_cast<Instruction>(Val: D))
3667 VD.LocalNum = InstrToDFSNum(V: I);
3668 else
3669 llvm_unreachable("Should have been an instruction");
3670
3671 LoadsAndStores.emplace_back(Args&: VD);
3672 }
3673}
3674
3675static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
3676 patchReplacementInstruction(I, Repl);
3677 I->replaceAllUsesWith(V: Repl);
3678}
3679
3680void NewGVN::deleteInstructionsInBlock(BasicBlock *BB) {
3681 LLVM_DEBUG(dbgs() << " BasicBlock Dead:" << *BB);
3682 ++NumGVNBlocksDeleted;
3683
3684 // Delete the instructions backwards, as it has a reduced likelihood of having
3685 // to update as many def-use and use-def chains. Start after the terminator.
3686 auto StartPoint = BB->rbegin();
3687 ++StartPoint;
3688 // Note that we explicitly recalculate BB->rend() on each iteration,
3689 // as it may change when we remove the first instruction.
3690 for (BasicBlock::reverse_iterator I(StartPoint); I != BB->rend();) {
3691 Instruction &Inst = *I++;
3692 if (!Inst.use_empty())
3693 Inst.replaceAllUsesWith(V: PoisonValue::get(T: Inst.getType()));
3694 if (isa<LandingPadInst>(Val: Inst))
3695 continue;
3696 salvageKnowledge(I: &Inst, AC);
3697
3698 Inst.eraseFromParent();
3699 ++NumGVNInstrDeleted;
3700 }
3701 // Now insert something that simplifycfg will turn into an unreachable.
3702 Type *Int8Ty = Type::getInt8Ty(C&: BB->getContext());
3703 new StoreInst(
3704 PoisonValue::get(T: Int8Ty),
3705 Constant::getNullValue(Ty: PointerType::getUnqual(C&: BB->getContext())),
3706 BB->getTerminator()->getIterator());
3707}
3708
3709void NewGVN::markInstructionForDeletion(Instruction *I) {
3710 LLVM_DEBUG(dbgs() << "Marking " << *I << " for deletion\n");
3711 InstructionsToErase.insert(Ptr: I);
3712}
3713
3714void NewGVN::replaceInstruction(Instruction *I, Value *V) {
3715 LLVM_DEBUG(dbgs() << "Replacing " << *I << " with " << *V << "\n");
3716 patchAndReplaceAllUsesWith(I, Repl: V);
3717 // We save the actual erasing to avoid invalidating memory
3718 // dependencies until we are done with everything.
3719 markInstructionForDeletion(I);
3720}
3721
3722namespace {
3723
3724// This is a stack that contains both the value and dfs info of where
3725// that value is valid.
3726class ValueDFSStack {
3727public:
3728 Value *back() const { return ValueStack.back(); }
3729 std::pair<int, int> dfs_back() const { return DFSStack.back(); }
3730
3731 void push_back(Value *V, int DFSIn, int DFSOut) {
3732 ValueStack.emplace_back(Args&: V);
3733 DFSStack.emplace_back(Args&: DFSIn, Args&: DFSOut);
3734 }
3735
3736 bool empty() const { return DFSStack.empty(); }
3737
3738 bool isInScope(int DFSIn, int DFSOut) const {
3739 if (empty())
3740 return false;
3741 return DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second;
3742 }
3743
3744 void popUntilDFSScope(int DFSIn, int DFSOut) {
3745
3746 // These two should always be in sync at this point.
3747 assert(ValueStack.size() == DFSStack.size() &&
3748 "Mismatch between ValueStack and DFSStack");
3749 while (
3750 !DFSStack.empty() &&
3751 !(DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second)) {
3752 DFSStack.pop_back();
3753 ValueStack.pop_back();
3754 }
3755 }
3756
3757private:
3758 SmallVector<Value *, 8> ValueStack;
3759 SmallVector<std::pair<int, int>, 8> DFSStack;
3760};
3761
3762} // end anonymous namespace
3763
3764// Given an expression, get the congruence class for it.
3765CongruenceClass *NewGVN::getClassForExpression(const Expression *E) const {
3766 if (auto *VE = dyn_cast<VariableExpression>(Val: E))
3767 return ValueToClass.lookup(Val: VE->getVariableValue());
3768 else if (isa<DeadExpression>(Val: E))
3769 return TOPClass;
3770 return ExpressionToClass.lookup(Val: E);
3771}
3772
3773// Given a value and a basic block we are trying to see if it is available in,
3774// see if the value has a leader available in that block.
3775Value *NewGVN::findPHIOfOpsLeader(const Expression *E,
3776 const Instruction *OrigInst,
3777 const BasicBlock *BB) const {
3778 // It would already be constant if we could make it constant
3779 if (auto *CE = dyn_cast<ConstantExpression>(Val: E))
3780 return CE->getConstantValue();
3781 if (auto *VE = dyn_cast<VariableExpression>(Val: E)) {
3782 auto *V = VE->getVariableValue();
3783 if (alwaysAvailable(V) || DT->dominates(A: getBlockForValue(V), B: BB))
3784 return VE->getVariableValue();
3785 }
3786
3787 auto *CC = getClassForExpression(E);
3788 if (!CC)
3789 return nullptr;
3790 if (alwaysAvailable(V: CC->getLeader()))
3791 return CC->getLeader();
3792
3793 for (auto *Member : *CC) {
3794 auto *MemberInst = dyn_cast<Instruction>(Val: Member);
3795 if (MemberInst == OrigInst)
3796 continue;
3797 // Anything that isn't an instruction is always available.
3798 if (!MemberInst)
3799 return Member;
3800 if (DT->dominates(A: getBlockForValue(V: MemberInst), B: BB))
3801 return Member;
3802 }
3803 return nullptr;
3804}
3805
3806bool NewGVN::eliminateInstructions(Function &F) {
3807 // This is a non-standard eliminator. The normal way to eliminate is
3808 // to walk the dominator tree in order, keeping track of available
3809 // values, and eliminating them. However, this is mildly
3810 // pointless. It requires doing lookups on every instruction,
3811 // regardless of whether we will ever eliminate it. For
3812 // instructions part of most singleton congruence classes, we know we
3813 // will never eliminate them.
3814
3815 // Instead, this eliminator looks at the congruence classes directly, sorts
3816 // them into a DFS ordering of the dominator tree, and then we just
3817 // perform elimination straight on the sets by walking the congruence
3818 // class member uses in order, and eliminate the ones dominated by the
3819 // last member. This is worst case O(E log E) where E = number of
3820 // instructions in a single congruence class. In theory, this is all
3821 // instructions. In practice, it is much faster, as most instructions are
3822 // either in singleton congruence classes or can't possibly be eliminated
3823 // anyway (if there are no overlapping DFS ranges in class).
3824 // When we find something not dominated, it becomes the new leader
3825 // for elimination purposes.
3826 // TODO: If we wanted to be faster, We could remove any members with no
3827 // overlapping ranges while sorting, as we will never eliminate anything
3828 // with those members, as they don't dominate anything else in our set.
3829
3830 bool AnythingReplaced = false;
3831
3832 // Since we are going to walk the domtree anyway, and we can't guarantee the
3833 // DFS numbers are updated, we compute some ourselves.
3834 DT->updateDFSNumbers();
3835
3836 // Go through all of our phi nodes, and kill the arguments associated with
3837 // unreachable edges.
3838 auto ReplaceUnreachablePHIArgs = [&](PHINode *PHI, BasicBlock *BB) {
3839 for (auto &Operand : PHI->incoming_values())
3840 if (!ReachableEdges.count(V: {PHI->getIncomingBlock(U: Operand), BB})) {
3841 LLVM_DEBUG(dbgs() << "Replacing incoming value of " << PHI
3842 << " for block "
3843 << getBlockName(PHI->getIncomingBlock(Operand))
3844 << " with poison due to it being unreachable\n");
3845 Operand.set(PoisonValue::get(T: PHI->getType()));
3846 }
3847 };
3848 // Replace unreachable phi arguments.
3849 // At this point, RevisitOnReachabilityChange only contains:
3850 //
3851 // 1. PHIs
3852 // 2. Temporaries that will convert to PHIs
3853 // 3. Operations that are affected by an unreachable edge but do not fit into
3854 // 1 or 2 (rare).
3855 // So it is a slight overshoot of what we want. We could make it exact by
3856 // using two SparseBitVectors per block.
3857 DenseMap<const BasicBlock *, unsigned> ReachablePredCount;
3858 for (auto &KV : ReachableEdges)
3859 ReachablePredCount[KV.getEnd()]++;
3860 for (auto &BBPair : RevisitOnReachabilityChange) {
3861 for (auto InstNum : BBPair.second) {
3862 auto *Inst = InstrFromDFSNum(DFSNum: InstNum);
3863 auto *PHI = dyn_cast<PHINode>(Val: Inst);
3864 PHI = PHI ? PHI : dyn_cast_or_null<PHINode>(Val: RealToTemp.lookup(Val: Inst));
3865 if (!PHI)
3866 continue;
3867 auto *BB = BBPair.first;
3868 if (ReachablePredCount.lookup(Val: BB) != PHI->getNumIncomingValues())
3869 ReplaceUnreachablePHIArgs(PHI, BB);
3870 }
3871 }
3872
3873 // Map to store the use counts
3874 DenseMap<const Value *, unsigned int> UseCounts;
3875 for (auto *CC : reverse(C&: CongruenceClasses)) {
3876 LLVM_DEBUG(dbgs() << "Eliminating in congruence class " << CC->getID()
3877 << "\n");
3878 // Track the equivalent store info so we can decide whether to try
3879 // dead store elimination.
3880 SmallVector<ValueDFS, 8> PossibleDeadStores;
3881 SmallPtrSet<Instruction *, 8> ProbablyDead;
3882 if (CC->isDead() || CC->empty())
3883 continue;
3884 // Everything still in the TOP class is unreachable or dead.
3885 if (CC == TOPClass) {
3886 for (auto *M : *CC) {
3887 auto *VTE = ValueToExpression.lookup(Val: M);
3888 if (VTE && isa<DeadExpression>(Val: VTE))
3889 markInstructionForDeletion(I: cast<Instruction>(Val: M));
3890 assert((!ReachableBlocks.count(cast<Instruction>(M)->getParent()) ||
3891 InstructionsToErase.count(cast<Instruction>(M))) &&
3892 "Everything in TOP should be unreachable or dead at this "
3893 "point");
3894 }
3895 continue;
3896 }
3897
3898 assert(CC->getLeader() && "We should have had a leader");
3899 // If this is a leader that is always available, and it's a
3900 // constant or has no equivalences, just replace everything with
3901 // it. We then update the congruence class with whatever members
3902 // are left.
3903 Value *Leader =
3904 CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();
3905 if (alwaysAvailable(V: Leader)) {
3906 CongruenceClass::MemberSet MembersLeft;
3907 for (auto *M : *CC) {
3908 Value *Member = M;
3909 // Void things have no uses we can replace.
3910 if (Member == Leader || !isa<Instruction>(Val: Member) ||
3911 Member->getType()->isVoidTy()) {
3912 MembersLeft.insert(Ptr: Member);
3913 continue;
3914 }
3915
3916 LLVM_DEBUG(dbgs() << "Found replacement " << *(Leader) << " for "
3917 << *Member << "\n");
3918 auto *I = cast<Instruction>(Val: Member);
3919 assert(Leader != I && "About to accidentally remove our leader");
3920 replaceInstruction(I, V: Leader);
3921 AnythingReplaced = true;
3922 }
3923 CC->swap(Other&: MembersLeft);
3924 } else {
3925 // If this is a singleton, we can skip it.
3926 if (CC->size() != 1 || RealToTemp.count(Val: Leader)) {
3927 // This is a stack because equality replacement/etc may place
3928 // constants in the middle of the member list, and we want to use
3929 // those constant values in preference to the current leader, over
3930 // the scope of those constants.
3931 ValueDFSStack EliminationStack;
3932
3933 // Convert the members to DFS ordered sets and then merge them.
3934 SmallVector<ValueDFS, 8> DFSOrderedSet;
3935 convertClassToDFSOrdered(Dense: *CC, DFSOrderedSet, UseCounts, ProbablyDead);
3936
3937 // Sort the whole thing.
3938 llvm::sort(C&: DFSOrderedSet);
3939 for (auto &VD : DFSOrderedSet) {
3940 int MemberDFSIn = VD.DFSIn;
3941 int MemberDFSOut = VD.DFSOut;
3942 Value *Def = VD.Def.getPointer();
3943 bool FromStore = VD.Def.getInt();
3944 Use *U = VD.U;
3945 // We ignore void things because we can't get a value from them.
3946 if (Def && Def->getType()->isVoidTy())
3947 continue;
3948 auto *DefInst = dyn_cast_or_null<Instruction>(Val: Def);
3949 if (DefInst && AllTempInstructions.count(V: DefInst)) {
3950 auto *PN = cast<PHINode>(Val: DefInst);
3951
3952 // If this is a value phi and that's the expression we used, insert
3953 // it into the program
3954 // remove from temp instruction list.
3955 AllTempInstructions.erase(V: PN);
3956 auto *DefBlock = getBlockForValue(V: Def);
3957 LLVM_DEBUG(dbgs() << "Inserting fully real phi of ops" << *Def
3958 << " into block "
3959 << getBlockName(getBlockForValue(Def)) << "\n");
3960 PN->insertBefore(InsertPos: DefBlock->begin());
3961 Def = PN;
3962 NumGVNPHIOfOpsEliminations++;
3963 }
3964
3965 if (EliminationStack.empty()) {
3966 LLVM_DEBUG(dbgs() << "Elimination Stack is empty\n");
3967 } else {
3968 LLVM_DEBUG(dbgs() << "Elimination Stack Top DFS numbers are ("
3969 << EliminationStack.dfs_back().first << ","
3970 << EliminationStack.dfs_back().second << ")\n");
3971 }
3972
3973 LLVM_DEBUG(dbgs() << "Current DFS numbers are (" << MemberDFSIn << ","
3974 << MemberDFSOut << ")\n");
3975 // First, we see if we are out of scope or empty. If so,
3976 // and there equivalences, we try to replace the top of
3977 // stack with equivalences (if it's on the stack, it must
3978 // not have been eliminated yet).
3979 // Then we synchronize to our current scope, by
3980 // popping until we are back within a DFS scope that
3981 // dominates the current member.
3982 // Then, what happens depends on a few factors
3983 // If the stack is now empty, we need to push
3984 // If we have a constant or a local equivalence we want to
3985 // start using, we also push.
3986 // Otherwise, we walk along, processing members who are
3987 // dominated by this scope, and eliminate them.
3988 bool ShouldPush = Def && EliminationStack.empty();
3989 bool OutOfScope =
3990 !EliminationStack.isInScope(DFSIn: MemberDFSIn, DFSOut: MemberDFSOut);
3991
3992 if (OutOfScope || ShouldPush) {
3993 // Sync to our current scope.
3994 EliminationStack.popUntilDFSScope(DFSIn: MemberDFSIn, DFSOut: MemberDFSOut);
3995 bool ShouldPush = Def && EliminationStack.empty();
3996 if (ShouldPush) {
3997 EliminationStack.push_back(V: Def, DFSIn: MemberDFSIn, DFSOut: MemberDFSOut);
3998 }
3999 }
4000
4001 // Skip the Def's, we only want to eliminate on their uses. But mark
4002 // dominated defs as dead.
4003 if (Def) {
4004 // For anything in this case, what and how we value number
4005 // guarantees that any side-effects that would have occurred (ie
4006 // throwing, etc) can be proven to either still occur (because it's
4007 // dominated by something that has the same side-effects), or never
4008 // occur. Otherwise, we would not have been able to prove it value
4009 // equivalent to something else. For these things, we can just mark
4010 // it all dead. Note that this is different from the "ProbablyDead"
4011 // set, which may not be dominated by anything, and thus, are only
4012 // easy to prove dead if they are also side-effect free. Note that
4013 // because stores are put in terms of the stored value, we skip
4014 // stored values here. If the stored value is really dead, it will
4015 // still be marked for deletion when we process it in its own class.
4016 auto *DefI = dyn_cast<Instruction>(Val: Def);
4017 if (!EliminationStack.empty() && DefI && !FromStore) {
4018 Value *DominatingLeader = EliminationStack.back();
4019 if (DominatingLeader != Def) {
4020 // Even if the instruction is removed, we still need to update
4021 // flags/metadata due to downstreams users of the leader.
4022 patchReplacementInstruction(I: DefI, Repl: DominatingLeader);
4023
4024 SmallVector<DbgVariableRecord *> DVRUsers;
4025 findDbgUsers(V: DefI, DbgVariableRecords&: DVRUsers);
4026
4027 for (auto *DVR : DVRUsers)
4028 DVR->replaceVariableLocationOp(OldValue: DefI, NewValue: DominatingLeader);
4029
4030 markInstructionForDeletion(I: DefI);
4031 }
4032 }
4033 continue;
4034 }
4035 // At this point, we know it is a Use we are trying to possibly
4036 // replace.
4037
4038 assert(isa<Instruction>(U->get()) &&
4039 "Current def should have been an instruction");
4040 assert(isa<Instruction>(U->getUser()) &&
4041 "Current user should have been an instruction");
4042
4043 // If the thing we are replacing into is already marked to be dead,
4044 // this use is dead. Note that this is true regardless of whether
4045 // we have anything dominating the use or not. We do this here
4046 // because we are already walking all the uses anyway.
4047 Instruction *InstUse = cast<Instruction>(Val: U->getUser());
4048 if (InstructionsToErase.count(Ptr: InstUse)) {
4049 auto &UseCount = UseCounts[U->get()];
4050 if (--UseCount == 0) {
4051 ProbablyDead.insert(Ptr: cast<Instruction>(Val: U->get()));
4052 }
4053 }
4054
4055 // If we get to this point, and the stack is empty we must have a use
4056 // with nothing we can use to eliminate this use, so just skip it.
4057 if (EliminationStack.empty())
4058 continue;
4059
4060 Value *DominatingLeader = EliminationStack.back();
4061
4062 Instruction *SSACopy = nullptr;
4063 if (auto *BC = dyn_cast<BitCastInst>(Val: DominatingLeader)) {
4064 if (BC->getType() == BC->getOperand(i_nocapture: 0)->getType() &&
4065 PredInfo->getPredicateInfoFor(V: DominatingLeader)) {
4066 SSACopy = BC;
4067 DominatingLeader = BC->getOperand(i_nocapture: 0);
4068 }
4069 }
4070
4071 // Don't replace our existing users with ourselves.
4072 if (U->get() == DominatingLeader)
4073 continue;
4074
4075 // If we replaced something in an instruction, handle the patching of
4076 // metadata. Skip this if we are replacing predicateinfo with its
4077 // original operand, as we already know we can just drop it.
4078 auto *ReplacedInst = cast<Instruction>(Val: U->get());
4079 auto *PI = PredInfo->getPredicateInfoFor(V: ReplacedInst);
4080 if (!PI || DominatingLeader != PI->OriginalOp)
4081 patchReplacementInstruction(I: ReplacedInst, Repl: DominatingLeader);
4082
4083 LLVM_DEBUG(dbgs()
4084 << "Found replacement " << *DominatingLeader << " for "
4085 << *U->get() << " in " << *(U->getUser()) << "\n");
4086 U->set(DominatingLeader);
4087 // This is now a use of the dominating leader, which means if the
4088 // dominating leader was dead, it's now live!
4089 auto &LeaderUseCount = UseCounts[DominatingLeader];
4090 // It's about to be alive again.
4091 if (LeaderUseCount == 0 && isa<Instruction>(Val: DominatingLeader))
4092 ProbablyDead.erase(Ptr: cast<Instruction>(Val: DominatingLeader));
4093 // For copy instructions, we use their operand as a leader,
4094 // which means we remove a user of the copy and it may become dead.
4095 if (SSACopy) {
4096 auto It = UseCounts.find(Val: SSACopy);
4097 if (It != UseCounts.end()) {
4098 unsigned &IIUseCount = It->second;
4099 if (--IIUseCount == 0)
4100 ProbablyDead.insert(Ptr: SSACopy);
4101 }
4102 }
4103 ++LeaderUseCount;
4104 AnythingReplaced = true;
4105 }
4106 }
4107 }
4108
4109 // At this point, anything still in the ProbablyDead set is actually dead if
4110 // would be trivially dead.
4111 for (auto *I : ProbablyDead)
4112 if (wouldInstructionBeTriviallyDead(I))
4113 markInstructionForDeletion(I);
4114
4115 // Cleanup the congruence class.
4116 CongruenceClass::MemberSet MembersLeft;
4117 for (auto *Member : *CC)
4118 if (!isa<Instruction>(Val: Member) ||
4119 !InstructionsToErase.count(Ptr: cast<Instruction>(Val: Member)))
4120 MembersLeft.insert(Ptr: Member);
4121 CC->swap(Other&: MembersLeft);
4122
4123 // If we have possible dead stores to look at, try to eliminate them.
4124 if (CC->getStoreCount() > 0) {
4125 convertClassToLoadsAndStores(Dense: *CC, LoadsAndStores&: PossibleDeadStores);
4126 llvm::sort(C&: PossibleDeadStores);
4127 ValueDFSStack EliminationStack;
4128 for (auto &VD : PossibleDeadStores) {
4129 int MemberDFSIn = VD.DFSIn;
4130 int MemberDFSOut = VD.DFSOut;
4131 Instruction *Member = cast<Instruction>(Val: VD.Def.getPointer());
4132 if (EliminationStack.empty() ||
4133 !EliminationStack.isInScope(DFSIn: MemberDFSIn, DFSOut: MemberDFSOut)) {
4134 // Sync to our current scope.
4135 EliminationStack.popUntilDFSScope(DFSIn: MemberDFSIn, DFSOut: MemberDFSOut);
4136 if (EliminationStack.empty()) {
4137 EliminationStack.push_back(V: Member, DFSIn: MemberDFSIn, DFSOut: MemberDFSOut);
4138 continue;
4139 }
4140 }
4141 // We already did load elimination, so nothing to do here.
4142 if (isa<LoadInst>(Val: Member))
4143 continue;
4144 assert(!EliminationStack.empty());
4145 Instruction *Leader = cast<Instruction>(Val: EliminationStack.back());
4146 (void)Leader;
4147 assert(DT->dominates(Leader->getParent(), Member->getParent()));
4148 // Member is dominater by Leader, and thus dead
4149 LLVM_DEBUG(dbgs() << "Marking dead store " << *Member
4150 << " that is dominated by " << *Leader << "\n");
4151 markInstructionForDeletion(I: Member);
4152 CC->erase(M: Member);
4153 ++NumGVNDeadStores;
4154 }
4155 }
4156 }
4157 return AnythingReplaced;
4158}
4159
4160// This function provides global ranking of operations so that we can place them
4161// in a canonical order. Note that rank alone is not necessarily enough for a
4162// complete ordering, as constants all have the same rank. However, generally,
4163// we will simplify an operation with all constants so that it doesn't matter
4164// what order they appear in.
4165unsigned int NewGVN::getRank(const Value *V) const {
4166 // Prefer constants to undef to anything else
4167 // Undef is a constant, have to check it first.
4168 // Prefer poison to undef as it's less defined.
4169 // Prefer smaller constants to constantexprs
4170 // Note that the order here matters because of class inheritance
4171 if (isa<ConstantExpr>(Val: V))
4172 return 3;
4173 if (isa<PoisonValue>(Val: V))
4174 return 1;
4175 if (isa<UndefValue>(Val: V))
4176 return 2;
4177 if (isa<Constant>(Val: V))
4178 return 0;
4179 if (auto *A = dyn_cast<Argument>(Val: V))
4180 return 4 + A->getArgNo();
4181
4182 // Need to shift the instruction DFS by number of arguments + 5 to account for
4183 // the constant and argument ranking above.
4184 unsigned Result = InstrToDFSNum(V);
4185 if (Result > 0)
4186 return 5 + NumFuncArgs + Result;
4187 // Unreachable or something else, just return a really large number.
4188 return ~0;
4189}
4190
4191// This is a function that says whether two commutative operations should
4192// have their order swapped when canonicalizing.
4193bool NewGVN::shouldSwapOperands(const Value *A, const Value *B) const {
4194 // Because we only care about a total ordering, and don't rewrite expressions
4195 // in this order, we order by rank, which will give a strict weak ordering to
4196 // everything but constants, and then we order by pointer address.
4197 return std::make_pair(x: getRank(V: A), y&: A) > std::make_pair(x: getRank(V: B), y&: B);
4198}
4199
4200bool NewGVN::shouldSwapOperandsForPredicate(const Value *A, const Value *B,
4201 const BitCastInst *I) const {
4202 if (shouldSwapOperands(A, B)) {
4203 PredicateSwapChoice[I] = B;
4204 return true;
4205 }
4206
4207 auto LookupResult = PredicateSwapChoice.find(Val: I);
4208 if (LookupResult != PredicateSwapChoice.end()) {
4209 auto *SeenPredicate = LookupResult->second;
4210 if (SeenPredicate) {
4211 // We previously decided to swap B to the left. Keep that choice.
4212 if (SeenPredicate == B)
4213 return true;
4214 else
4215 LookupResult->second = nullptr;
4216 }
4217 }
4218 return false;
4219}
4220
4221PreservedAnalyses NewGVNPass::run(Function &F, AnalysisManager<Function> &AM) {
4222 // Apparently the order in which we get these results matter for
4223 // the old GVN (see Chandler's comment in GVN.cpp). I'll keep
4224 // the same order here, just in case.
4225 auto &AC = AM.getResult<AssumptionAnalysis>(IR&: F);
4226 auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F);
4227 auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F);
4228 auto &AA = AM.getResult<AAManager>(IR&: F);
4229 auto &MSSA = AM.getResult<MemorySSAAnalysis>(IR&: F).getMSSA();
4230 bool Changed =
4231 NewGVN(F, &DT, &AC, &TLI, &AA, &MSSA, F.getDataLayout())
4232 .runGVN();
4233 if (!Changed)
4234 return PreservedAnalyses::all();
4235 PreservedAnalyses PA;
4236 PA.preserve<DominatorTreeAnalysis>();
4237 return PA;
4238}
4239