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