/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp

Bug Summary

File:	lib/Transforms/Scalar/NewGVN.cpp
Warning:	line 1179, column 9 Called C++ object pointer is null

Annotated Source Code

//===---- NewGVN.cpp - Global Value Numbering Pass --------------*- C++ -*-===//

// The LLVM Compiler Infrastructure

// This file is distributed under the University of Illinois Open Source

// License. See LICENSE.TXT for details.

//===----------------------------------------------------------------------===//

/// \file

/// This file implements the new LLVM's Global Value Numbering pass.

/// GVN partitions values computed by a function into congruence classes.

/// Values ending up in the same congruence class are guaranteed to be the same

/// for every execution of the program. In that respect, congruency is a

/// compile-time approximation of equivalence of values at runtime.

/// The algorithm implemented here uses a sparse formulation and it's based

/// on the ideas described in the paper:

/// "A Sparse Algorithm for Predicated Global Value Numbering" from

/// Karthik Gargi.

///

/// A brief overview of the algorithm: The algorithm is essentially the same as

/// the standard RPO value numbering algorithm (a good reference is the paper

/// "SCC based value numbering" by L. Taylor Simpson) with one major difference:

/// The RPO algorithm proceeds, on every iteration, to process every reachable

/// block and every instruction in that block. This is because the standard RPO

/// algorithm does not track what things have the same value number, it only

/// tracks what the value number of a given operation is (the mapping is

/// operation -> value number). Thus, when a value number of an operation

/// changes, it must reprocess everything to ensure all uses of a value number

/// get updated properly. In constrast, the sparse algorithm we use *also*

/// tracks what operations have a given value number (IE it also tracks the

/// reverse mapping from value number -> operations with that value number), so

/// that it only needs to reprocess the instructions that are affected when

/// something's value number changes. The rest of the algorithm is devoted to

/// performing symbolic evaluation, forward propagation, and simplification of

/// operations based on the value numbers deduced so far.

///

/// We also do not perform elimination by using any published algorithm. All

/// published algorithms are O(Instructions). Instead, we use a technique that

/// is O(number of operations with the same value number), enabling us to skip

/// trying to eliminate things that have unique value numbers.

//===----------------------------------------------------------------------===//

#include "llvm/Transforms/Scalar/NewGVN.h"

#include "llvm/ADT/BitVector.h"

#include "llvm/ADT/DenseMap.h"

#include "llvm/ADT/DenseSet.h"

#include "llvm/ADT/DepthFirstIterator.h"

#include "llvm/ADT/Hashing.h"

#include "llvm/ADT/MapVector.h"

#include "llvm/ADT/PostOrderIterator.h"

#include "llvm/ADT/STLExtras.h"

#include "llvm/ADT/SmallPtrSet.h"

#include "llvm/ADT/SmallSet.h"

#include "llvm/ADT/SparseBitVector.h"

#include "llvm/ADT/Statistic.h"

#include "llvm/ADT/TinyPtrVector.h"

#include "llvm/Analysis/AliasAnalysis.h"

#include "llvm/Analysis/AssumptionCache.h"

#include "llvm/Analysis/CFG.h"

#include "llvm/Analysis/CFGPrinter.h"

#include "llvm/Analysis/ConstantFolding.h"

#include "llvm/Analysis/GlobalsModRef.h"

#include "llvm/Analysis/InstructionSimplify.h"

#include "llvm/Analysis/MemoryBuiltins.h"

#include "llvm/Analysis/MemoryLocation.h"

#include "llvm/Analysis/MemorySSA.h"

#include "llvm/Analysis/TargetLibraryInfo.h"

#include "llvm/IR/DataLayout.h"

#include "llvm/IR/Dominators.h"

#include "llvm/IR/GlobalVariable.h"

#include "llvm/IR/IRBuilder.h"

#include "llvm/IR/IntrinsicInst.h"

#include "llvm/IR/LLVMContext.h"

#include "llvm/IR/Metadata.h"

#include "llvm/IR/PatternMatch.h"

#include "llvm/IR/Type.h"

#include "llvm/Support/Allocator.h"

#include "llvm/Support/CommandLine.h"

#include "llvm/Support/Debug.h"

#include "llvm/Support/DebugCounter.h"

#include "llvm/Transforms/Scalar.h"

#include "llvm/Transforms/Scalar/GVNExpression.h"

#include "llvm/Transforms/Utils/BasicBlockUtils.h"

#include "llvm/Transforms/Utils/Local.h"

#include "llvm/Transforms/Utils/PredicateInfo.h"

#include "llvm/Transforms/Utils/VNCoercion.h"

#include <numeric>

#include <unordered_map>

#include <utility>

#include <vector>

using namespace llvm;

using namespace PatternMatch;

using namespace llvm::GVNExpression;

using namespace llvm::VNCoercion;

#define DEBUG_TYPE"newgvn" "newgvn"

STATISTIC(NumGVNInstrDeleted, "Number of instructions deleted")static llvm::Statistic NumGVNInstrDeleted = {"newgvn", "NumGVNInstrDeleted"
, "Number of instructions deleted", {0}, false};

STATISTIC(NumGVNBlocksDeleted, "Number of blocks deleted")static llvm::Statistic NumGVNBlocksDeleted = {"newgvn", "NumGVNBlocksDeleted"
, "Number of blocks deleted", {0}, false};

STATISTIC(NumGVNOpsSimplified, "Number of Expressions simplified")static llvm::Statistic NumGVNOpsSimplified = {"newgvn", "NumGVNOpsSimplified"
, "Number of Expressions simplified", {0}, false};

100

STATISTIC(NumGVNPhisAllSame, "Number of PHIs whos arguments are all the same")static llvm::Statistic NumGVNPhisAllSame = {"newgvn", "NumGVNPhisAllSame"
, "Number of PHIs whos arguments are all the same", {0}, false
};

101

STATISTIC(NumGVNMaxIterations,static llvm::Statistic NumGVNMaxIterations = {"newgvn", "NumGVNMaxIterations"
, "Maximum Number of iterations it took to converge GVN", {0}
, false}

102

"Maximum Number of iterations it took to converge GVN")static llvm::Statistic NumGVNMaxIterations = {"newgvn", "NumGVNMaxIterations"
, "Maximum Number of iterations it took to converge GVN", {0}
, false};

103

STATISTIC(NumGVNLeaderChanges, "Number of leader changes")static llvm::Statistic NumGVNLeaderChanges = {"newgvn", "NumGVNLeaderChanges"
, "Number of leader changes", {0}, false};

104

STATISTIC(NumGVNSortedLeaderChanges, "Number of sorted leader changes")static llvm::Statistic NumGVNSortedLeaderChanges = {"newgvn",
"NumGVNSortedLeaderChanges", "Number of sorted leader changes"
, {0}, false};

105

STATISTIC(NumGVNAvoidedSortedLeaderChanges,static llvm::Statistic NumGVNAvoidedSortedLeaderChanges = {"newgvn"
, "NumGVNAvoidedSortedLeaderChanges", "Number of avoided sorted leader changes"
, {0}, false}

106

"Number of avoided sorted leader changes")static llvm::Statistic NumGVNAvoidedSortedLeaderChanges = {"newgvn"
, "NumGVNAvoidedSortedLeaderChanges", "Number of avoided sorted leader changes"
, {0}, false};

107

STATISTIC(NumGVNNotMostDominatingLeader,static llvm::Statistic NumGVNNotMostDominatingLeader = {"newgvn"
, "NumGVNNotMostDominatingLeader", "Number of times a member dominated it's new classes' leader"
, {0}, false}

108

"Number of times a member dominated it's new classes' leader")static llvm::Statistic NumGVNNotMostDominatingLeader = {"newgvn"
, "NumGVNNotMostDominatingLeader", "Number of times a member dominated it's new classes' leader"
, {0}, false};

109

STATISTIC(NumGVNDeadStores, "Number of redundant/dead stores eliminated")static llvm::Statistic NumGVNDeadStores = {"newgvn", "NumGVNDeadStores"
, "Number of redundant/dead stores eliminated", {0}, false};

110

DEBUG_COUNTER(VNCounter, "newgvn-vn",static const unsigned VNCounter = DebugCounter::registerCounter
("newgvn-vn", "Controls which instructions are value numbered"
);

111

"Controls which instructions are value numbered")static const unsigned VNCounter = DebugCounter::registerCounter
("newgvn-vn", "Controls which instructions are value numbered"
);

112

113

// Currently store defining access refinement is too slow due to basicaa being

114

// egregiously slow. This flag lets us keep it working while we work on this

115

// issue.

116

static cl::opt<bool> EnableStoreRefinement("enable-store-refinement",

117

cl::init(false), cl::Hidden);

118

119

//===----------------------------------------------------------------------===//

120

// GVN Pass

121

//===----------------------------------------------------------------------===//

122

123

// Anchor methods.

124

namespace llvm {

125

namespace GVNExpression {

126

Expression::~Expression() = default;

127

BasicExpression::~BasicExpression() = default;

128

CallExpression::~CallExpression() = default;

129

LoadExpression::~LoadExpression() = default;

130

StoreExpression::~StoreExpression() = default;

131

AggregateValueExpression::~AggregateValueExpression() = default;

132

PHIExpression::~PHIExpression() = default;

133

}

134

}

135

136

// Tarjan's SCC finding algorithm with Nuutila's improvements

137

// SCCIterator is actually fairly complex for the simple thing we want.

138

// It also wants to hand us SCC's that are unrelated to the phi node we ask

139

// about, and have us process them there or risk redoing work.

140

// Graph traits over a filter iterator also doesn't work that well here.

141

// This SCC finder is specialized to walk use-def chains, and only follows

142

// instructions,

143

// not generic values (arguments, etc).

144

struct TarjanSCC {

145

146

TarjanSCC() : Components(1) {}

147

148

void Start(const Instruction *Start) {

149

if (Root.lookup(Start) == 0)

150

FindSCC(Start);

151

}

152

153

const SmallPtrSetImpl<const Value *> &getComponentFor(const Value *V) const {

154

unsigned ComponentID = ValueToComponent.lookup(V);

155

156

assert(ComponentID > 0 &&((ComponentID > 0 && "Asking for a component for a value we never processed"
) ? static_cast<void> (0) : __assert_fail ("ComponentID > 0 && \"Asking for a component for a value we never processed\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 157, __PRETTY_FUNCTION__))

157

"Asking for a component for a value we never processed")((ComponentID > 0 && "Asking for a component for a value we never processed"
) ? static_cast<void> (0) : __assert_fail ("ComponentID > 0 && \"Asking for a component for a value we never processed\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 157, __PRETTY_FUNCTION__));

158

return Components[ComponentID];

159

}

160

161

private:

162

void FindSCC(const Instruction *I) {

163

Root[I] = ++DFSNum;

164

// Store the DFS Number we had before it possibly gets incremented.

165

unsigned int OurDFS = DFSNum;

166

for (auto &Op : I->operands()) {

167

if (auto *InstOp = dyn_cast<Instruction>(Op)) {

168

if (Root.lookup(Op) == 0)

169

FindSCC(InstOp);

170

if (!InComponent.count(Op))

171

Root[I] = std::min(Root.lookup(I), Root.lookup(Op));

172

}

173

}

174

// See if we really were the root of a component, by seeing if we still have

175

// our DFSNumber.

176

// If we do, we are the root of the component, and we have completed a

177

// component. If we do not,

178

// we are not the root of a component, and belong on the component stack.

179

if (Root.lookup(I) == OurDFS) {

180

unsigned ComponentID = Components.size();

181

Components.resize(Components.size() + 1);

182

auto &Component = Components.back();

183

Component.insert(I);

184

DEBUG(dbgs() << "Component root is " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Component root is " << *
I << "\n"; } } while (false);

185

InComponent.insert(I);

186

ValueToComponent[I] = ComponentID;

187

// Pop a component off the stack and label it.

188

while (!Stack.empty() && Root.lookup(Stack.back()) >= OurDFS) {

189

auto *Member = Stack.back();

190

DEBUG(dbgs() << "Component member is " << *Member << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Component member is " <<
*Member << "\n"; } } while (false);

191

Component.insert(Member);

192

InComponent.insert(Member);

193

ValueToComponent[Member] = ComponentID;

194

Stack.pop_back();

195

}

196

} else {

197

// Part of a component, push to stack

198

Stack.push_back(I);

199

}

200

}

201

unsigned int DFSNum = 1;

202

SmallPtrSet<const Value *, 8> InComponent;

203

DenseMap<const Value *, unsigned int> Root;

204

SmallVector<const Value *, 8> Stack;

205

// Store the components as vector of ptr sets, because we need the topo order

206

// of SCC's, but not individual member order

207

SmallVector<SmallPtrSet<const Value *, 8>, 8> Components;

208

DenseMap<const Value *, unsigned> ValueToComponent;

209

};

210

// Congruence classes represent the set of expressions/instructions

211

// that are all the same *during some scope in the function*.

212

// That is, because of the way we perform equality propagation, and

213

// because of memory value numbering, it is not correct to assume

214

// you can willy-nilly replace any member with any other at any

215

// point in the function.

216

217

// For any Value in the Member set, it is valid to replace any dominated member

218

// with that Value.

219

220

// Every congruence class has a leader, and the leader is used to symbolize

221

// instructions in a canonical way (IE every operand of an instruction that is a

222

// member of the same congruence class will always be replaced with leader

223

// during symbolization). To simplify symbolization, we keep the leader as a

224

// constant if class can be proved to be a constant value. Otherwise, the

225

// leader is the member of the value set with the smallest DFS number. Each

226

// congruence class also has a defining expression, though the expression may be

227

// null. If it exists, it can be used for forward propagation and reassociation

228

// of values.

229

230

// For memory, we also track a representative MemoryAccess, and a set of memory

231

// members for MemoryPhis (which have no real instructions). Note that for

232

// memory, it seems tempting to try to split the memory members into a

233

// MemoryCongruenceClass or something. Unfortunately, this does not work

234

// easily. The value numbering of a given memory expression depends on the

235

// leader of the memory congruence class, and the leader of memory congruence

236

// class depends on the value numbering of a given memory expression. This

237

// leads to wasted propagation, and in some cases, missed optimization. For

238

// example: If we had value numbered two stores together before, but now do not,

239

// we move them to a new value congruence class. This in turn will move at one

240

// of the memorydefs to a new memory congruence class. Which in turn, affects

241

// the value numbering of the stores we just value numbered (because the memory

242

// congruence class is part of the value number). So while theoretically

243

// possible to split them up, it turns out to be *incredibly* complicated to get

244

// it to work right, because of the interdependency. While structurally

245

// slightly messier, it is algorithmically much simpler and faster to do what we

246

// do here, and track them both at once in the same class.

247

// Note: The default iterators for this class iterate over values

248

class CongruenceClass {

249

public:

250

using MemberType = Value;

251

using MemberSet = SmallPtrSet<MemberType *, 4>;

252

using MemoryMemberType = MemoryPhi;

253

using MemoryMemberSet = SmallPtrSet<const MemoryMemberType *, 2>;

254

255

explicit CongruenceClass(unsigned ID) : ID(ID) {}

256

CongruenceClass(unsigned ID, Value *Leader, const Expression *E)

257

: ID(ID), RepLeader(Leader), DefiningExpr(E) {}

258

unsigned getID() const { return ID; }

259

// True if this class has no members left. This is mainly used for assertion

260

// purposes, and for skipping empty classes.

261

bool isDead() const {

262

// If it's both dead from a value perspective, and dead from a memory

263

// perspective, it's really dead.

264

return empty() && memory_empty();

265

}

266

// Leader functions

267

Value *getLeader() const { return RepLeader; }

268

void setLeader(Value *Leader) { RepLeader = Leader; }

269

const std::pair<Value *, unsigned int> &getNextLeader() const {

270

return NextLeader;

271

}

272

void resetNextLeader() { NextLeader = {nullptr, ~0}; }

273

274

void addPossibleNextLeader(std::pair<Value *, unsigned int> LeaderPair) {

275

if (LeaderPair.second < NextLeader.second)

276

NextLeader = LeaderPair;

277

}

278

279

Value *getStoredValue() const { return RepStoredValue; }

280

void setStoredValue(Value *Leader) { RepStoredValue = Leader; }

281

const MemoryAccess *getMemoryLeader() const { return RepMemoryAccess; }

282

void setMemoryLeader(const MemoryAccess *Leader) { RepMemoryAccess = Leader; }

283

284

// Forward propagation info

285

const Expression *getDefiningExpr() const { return DefiningExpr; }

286

287

// Value member set

288

bool empty() const { return Members.empty(); }

289

unsigned size() const { return Members.size(); }

290

MemberSet::const_iterator begin() const { return Members.begin(); }

291

MemberSet::const_iterator end() const { return Members.end(); }

292

void insert(MemberType *M) { Members.insert(M); }

293

void erase(MemberType *M) { Members.erase(M); }

294

void swap(MemberSet &Other) { Members.swap(Other); }

295

296

// Memory member set

297

bool memory_empty() const { return MemoryMembers.empty(); }

298

unsigned memory_size() const { return MemoryMembers.size(); }

299

MemoryMemberSet::const_iterator memory_begin() const {

300

return MemoryMembers.begin();

301

}

302

MemoryMemberSet::const_iterator memory_end() const {

303

return MemoryMembers.end();

304

}

305

iterator_range<MemoryMemberSet::const_iterator> memory() const {

306

return make_range(memory_begin(), memory_end());

307

}

308

void memory_insert(const MemoryMemberType *M) { MemoryMembers.insert(M); }

309

void memory_erase(const MemoryMemberType *M) { MemoryMembers.erase(M); }

310

311

// Store count

312

unsigned getStoreCount() const { return StoreCount; }

313

void incStoreCount() { ++StoreCount; }

314

void decStoreCount() {

315

assert(StoreCount != 0 && "Store count went negative")((StoreCount != 0 && "Store count went negative") ? static_cast
<void> (0) : __assert_fail ("StoreCount != 0 && \"Store count went negative\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 315, __PRETTY_FUNCTION__));

316

--StoreCount;

317

}

318

319

// True if this class has no memory members.

320

bool definesNoMemory() const { return StoreCount == 0 && memory_empty(); }

321

322

// Return true if two congruence classes are equivalent to each other. This

323

// means

324

// that every field but the ID number and the dead field are equivalent.

325

bool isEquivalentTo(const CongruenceClass *Other) const {

326

if (!Other)

327

return false;

328

if (this == Other)

329

return true;

330

331

if (std::tie(StoreCount, RepLeader, RepStoredValue, RepMemoryAccess) !=

332

std::tie(Other->StoreCount, Other->RepLeader, Other->RepStoredValue,

333

Other->RepMemoryAccess))

334

return false;

335

if (DefiningExpr != Other->DefiningExpr)

336

if (!DefiningExpr || !Other->DefiningExpr ||

337

*DefiningExpr != *Other->DefiningExpr)

338

return false;

339

// We need some ordered set

340

std::set<Value *> AMembers(Members.begin(), Members.end());

341

std::set<Value *> BMembers(Members.begin(), Members.end());

342

return AMembers == BMembers;

343

}

344

345

private:

346

unsigned ID;

347

// Representative leader.

348

Value *RepLeader = nullptr;

349

// The most dominating leader after our current leader, because the member set

350

// is not sorted and is expensive to keep sorted all the time.

351

std::pair<Value *, unsigned int> NextLeader = {nullptr, ~0U};

352

// If this is represented by a store, the value of the store.

353

Value *RepStoredValue = nullptr;

354

// If this class contains MemoryDefs or MemoryPhis, this is the leading memory

355

// access.

356

const MemoryAccess *RepMemoryAccess = nullptr;

357

// Defining Expression.

358

const Expression *DefiningExpr = nullptr;

359

// Actual members of this class.

360

MemberSet Members;

361

// This is the set of MemoryPhis that exist in the class. MemoryDefs and

362

// MemoryUses have real instructions representing them, so we only need to

363

// track MemoryPhis here.

364

MemoryMemberSet MemoryMembers;

365

// Number of stores in this congruence class.

366

// This is used so we can detect store equivalence changes properly.

367

int StoreCount = 0;

368

};

369

370

namespace llvm {

371

template <> struct DenseMapInfo<const Expression *> {

372

static const Expression *getEmptyKey() {

373

auto Val = static_cast<uintptr_t>(-1);

374

Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;

375

return reinterpret_cast<const Expression *>(Val);

376

}

377

static const Expression *getTombstoneKey() {

378

auto Val = static_cast<uintptr_t>(~1U);

379

Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;

380

return reinterpret_cast<const Expression *>(Val);

381

}

382

static unsigned getHashValue(const Expression *V) {

383

return static_cast<unsigned>(V->getHashValue());

384

}

385

static bool isEqual(const Expression *LHS, const Expression *RHS) {

386

if (LHS == RHS)

387

return true;

388

if (LHS == getTombstoneKey() || RHS == getTombstoneKey() ||

389

LHS == getEmptyKey() || RHS == getEmptyKey())

390

return false;

391

return *LHS == *RHS;

392

}

393

};

394

} // end namespace llvm

395

396

namespace {

397

class NewGVN {

398

Function &F;

399

DominatorTree *DT;

400

const TargetLibraryInfo *TLI;

401

AliasAnalysis *AA;

402

MemorySSA *MSSA;

403

MemorySSAWalker *MSSAWalker;

404

const DataLayout &DL;

405

std::unique_ptr<PredicateInfo> PredInfo;

406

407

// These are the only two things the create* functions should have

408

// side-effects on due to allocating memory.

409

mutable BumpPtrAllocator ExpressionAllocator;

410

mutable ArrayRecycler<Value *> ArgRecycler;

411

mutable TarjanSCC SCCFinder;

412

const SimplifyQuery SQ;

413

414

// Number of function arguments, used by ranking

415

unsigned int NumFuncArgs;

416

417

// RPOOrdering of basic blocks

418

DenseMap<const DomTreeNode *, unsigned> RPOOrdering;

419

420

// Congruence class info.

421

422

// This class is called INITIAL in the paper. It is the class everything

423

// startsout in, and represents any value. Being an optimistic analysis,

424

// anything in the TOP class has the value TOP, which is indeterminate and

425

// equivalent to everything.

426

CongruenceClass *TOPClass;

427

std::vector<CongruenceClass *> CongruenceClasses;

428

unsigned NextCongruenceNum;

429

430

// Value Mappings.

431

DenseMap<Value *, CongruenceClass *> ValueToClass;

432

DenseMap<Value *, const Expression *> ValueToExpression;

433

434

// Mapping from predicate info we used to the instructions we used it with.

435

// In order to correctly ensure propagation, we must keep track of what

436

// comparisons we used, so that when the values of the comparisons change, we

437

// propagate the information to the places we used the comparison.

438

mutable DenseMap<const Value *, SmallPtrSet<Instruction *, 2>>

439

PredicateToUsers;

440

// the same reasoning as PredicateToUsers. When we skip MemoryAccesses for

441

// stores, we no longer can rely solely on the def-use chains of MemorySSA.

442

mutable DenseMap<const MemoryAccess *, SmallPtrSet<MemoryAccess *, 2>>

443

MemoryToUsers;

444

445

// A table storing which memorydefs/phis represent a memory state provably

446

// equivalent to another memory state.

447

// We could use the congruence class machinery, but the MemoryAccess's are

448

// abstract memory states, so they can only ever be equivalent to each other,

449

// and not to constants, etc.

450

DenseMap<const MemoryAccess *, CongruenceClass *> MemoryAccessToClass;

451

452

// We could, if we wanted, build MemoryPhiExpressions and

453

// MemoryVariableExpressions, etc, and value number them the same way we value

454

// number phi expressions. For the moment, this seems like overkill. They

455

// can only exist in one of three states: they can be TOP (equal to

456

// everything), Equivalent to something else, or unique. Because we do not

457

// create expressions for them, we need to simulate leader change not just

458

// when they change class, but when they change state. Note: We can do the

459

// same thing for phis, and avoid having phi expressions if we wanted, We

460

// should eventually unify in one direction or the other, so this is a little

461

// bit of an experiment in which turns out easier to maintain.

462

enum MemoryPhiState { MPS_Invalid, MPS_TOP, MPS_Equivalent, MPS_Unique };

463

DenseMap<const MemoryPhi *, MemoryPhiState> MemoryPhiState;

464

465

enum PhiCycleState { PCS_Unknown, PCS_CycleFree, PCS_Cycle };

466

mutable DenseMap<const PHINode *, PhiCycleState> PhiCycleState;

467

// Expression to class mapping.

468

using ExpressionClassMap = DenseMap<const Expression *, CongruenceClass *>;

469

ExpressionClassMap ExpressionToClass;

470

471

// Which values have changed as a result of leader changes.

472

SmallPtrSet<Value *, 8> LeaderChanges;

473

474

// Reachability info.

475

using BlockEdge = BasicBlockEdge;

476

DenseSet<BlockEdge> ReachableEdges;

477

SmallPtrSet<const BasicBlock *, 8> ReachableBlocks;

478

479

// This is a bitvector because, on larger functions, we may have

480

// thousands of touched instructions at once (entire blocks,

481

// instructions with hundreds of uses, etc). Even with optimization

482

// for when we mark whole blocks as touched, when this was a

483

// SmallPtrSet or DenseSet, for some functions, we spent >20% of all

484

// the time in GVN just managing this list. The bitvector, on the

485

// other hand, efficiently supports test/set/clear of both

486

// individual and ranges, as well as "find next element" This

487

// enables us to use it as a worklist with essentially 0 cost.

488

BitVector TouchedInstructions;

489

490

DenseMap<const BasicBlock *, std::pair<unsigned, unsigned>> BlockInstRange;

491

492

#ifndef NDEBUG

493

// Debugging for how many times each block and instruction got processed.

494

DenseMap<const Value *, unsigned> ProcessedCount;

495

#endif

496

497

// DFS info.

498

// This contains a mapping from Instructions to DFS numbers.

499

// The numbering starts at 1. An instruction with DFS number zero

500

// means that the instruction is dead.

501

DenseMap<const Value *, unsigned> InstrDFS;

502

503

// This contains the mapping DFS numbers to instructions.

504

SmallVector<Value *, 32> DFSToInstr;

505

506

// Deletion info.

507

SmallPtrSet<Instruction *, 8> InstructionsToErase;

508

509

public:

510

NewGVN(Function &F, DominatorTree *DT, AssumptionCache *AC,

511

TargetLibraryInfo *TLI, AliasAnalysis *AA, MemorySSA *MSSA,

512

const DataLayout &DL)

513

: F(F), DT(DT), TLI(TLI), AA(AA), MSSA(MSSA), DL(DL),

514

PredInfo(make_unique<PredicateInfo>(F, *DT, *AC)), SQ(DL, TLI, DT, AC) {

515

}

516

bool runGVN();

517

518

private:

519

// Expression handling.

520

const Expression *createExpression(Instruction *) const;

521

const Expression *createBinaryExpression(unsigned, Type *, Value *,

522

Value *) const;

523

PHIExpression *createPHIExpression(Instruction *, bool &HasBackEdge,

524

bool &AllConstant) const;

525

const VariableExpression *createVariableExpression(Value *) const;

526

const ConstantExpression *createConstantExpression(Constant *) const;

527

const Expression *createVariableOrConstant(Value *V) const;

528

const UnknownExpression *createUnknownExpression(Instruction *) const;

529

const StoreExpression *createStoreExpression(StoreInst *,

530

const MemoryAccess *) const;

531

LoadExpression *createLoadExpression(Type *, Value *, LoadInst *,

532

const MemoryAccess *) const;

533

const CallExpression *createCallExpression(CallInst *,

534

const MemoryAccess *) const;

535

const AggregateValueExpression *

536

createAggregateValueExpression(Instruction *) const;

537

bool setBasicExpressionInfo(Instruction *, BasicExpression *) const;

538

539

// Congruence class handling.

540

CongruenceClass *createCongruenceClass(Value *Leader, const Expression *E) {

541

auto *result = new CongruenceClass(NextCongruenceNum++, Leader, E);

542

CongruenceClasses.emplace_back(result);

543

return result;

544

}

545

546

CongruenceClass *createMemoryClass(MemoryAccess *MA) {

547

auto *CC = createCongruenceClass(nullptr, nullptr);

548

CC->setMemoryLeader(MA);

549

return CC;

550

}

551

CongruenceClass *ensureLeaderOfMemoryClass(MemoryAccess *MA) {

552

auto *CC = getMemoryClass(MA);

553

if (CC->getMemoryLeader() != MA)

554

CC = createMemoryClass(MA);

555

return CC;

556

}

557

558

CongruenceClass *createSingletonCongruenceClass(Value *Member) {

559

CongruenceClass *CClass = createCongruenceClass(Member, nullptr);

560

CClass->insert(Member);

561

ValueToClass[Member] = CClass;

562

return CClass;

563

}

564

void initializeCongruenceClasses(Function &F);

565

566

// Value number an Instruction or MemoryPhi.

567

void valueNumberMemoryPhi(MemoryPhi *);

568

void valueNumberInstruction(Instruction *);

569

570

// Symbolic evaluation.

571

const Expression *checkSimplificationResults(Expression *, Instruction *,

572

Value *) const;

573

const Expression *performSymbolicEvaluation(Value *) const;

574

const Expression *performSymbolicLoadCoercion(Type *, Value *, LoadInst *,

575

Instruction *,

576

MemoryAccess *) const;

577

const Expression *performSymbolicLoadEvaluation(Instruction *) const;

578

const Expression *performSymbolicStoreEvaluation(Instruction *) const;

579

const Expression *performSymbolicCallEvaluation(Instruction *) const;

580

const Expression *performSymbolicPHIEvaluation(Instruction *) const;

581

const Expression *performSymbolicAggrValueEvaluation(Instruction *) const;

582

const Expression *performSymbolicCmpEvaluation(Instruction *) const;

583

const Expression *performSymbolicPredicateInfoEvaluation(Instruction *) const;

584

585

// Congruence finding.

586

bool someEquivalentDominates(const Instruction *, const Instruction *) const;

587

Value *lookupOperandLeader(Value *) const;

588

void performCongruenceFinding(Instruction *, const Expression *);

589

void moveValueToNewCongruenceClass(Instruction *, const Expression *,

590

CongruenceClass *, CongruenceClass *);

591

void moveMemoryToNewCongruenceClass(Instruction *, MemoryAccess *,

592

CongruenceClass *, CongruenceClass *);

593

Value *getNextValueLeader(CongruenceClass *) const;

594

const MemoryAccess *getNextMemoryLeader(CongruenceClass *) const;

595

bool setMemoryClass(const MemoryAccess *From, CongruenceClass *To);

596

CongruenceClass *getMemoryClass(const MemoryAccess *MA) const;

597

const MemoryAccess *lookupMemoryLeader(const MemoryAccess *) const;

598

bool isMemoryAccessTop(const MemoryAccess *) const;

599

600

// Ranking

601

unsigned int getRank(const Value *) const;

602

bool shouldSwapOperands(const Value *, const Value *) const;

603

604

// Reachability handling.

605

void updateReachableEdge(BasicBlock *, BasicBlock *);

606

void processOutgoingEdges(TerminatorInst *, BasicBlock *);

607

Value *findConditionEquivalence(Value *) const;

608

609

// Elimination.

610

struct ValueDFS;

611

void convertClassToDFSOrdered(const CongruenceClass &,

612

SmallVectorImpl<ValueDFS> &,

613

DenseMap<const Value *, unsigned int> &,

614

SmallPtrSetImpl<Instruction *> &) const;

615

void convertClassToLoadsAndStores(const CongruenceClass &,

616

SmallVectorImpl<ValueDFS> &) const;

617

618

bool eliminateInstructions(Function &);

619

void replaceInstruction(Instruction *, Value *);

620

void markInstructionForDeletion(Instruction *);

621

void deleteInstructionsInBlock(BasicBlock *);

622

623

// New instruction creation.

624

void handleNewInstruction(Instruction *){};

625

626

// Various instruction touch utilities

627

void markUsersTouched(Value *);

628

void markMemoryUsersTouched(const MemoryAccess *);

629

void markMemoryDefTouched(const MemoryAccess *);

630

void markPredicateUsersTouched(Instruction *);

631

void markValueLeaderChangeTouched(CongruenceClass *CC);

632

void markMemoryLeaderChangeTouched(CongruenceClass *CC);

633

void addPredicateUsers(const PredicateBase *, Instruction *) const;

634

void addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const;

635

636

// Main loop of value numbering

637

void iterateTouchedInstructions();

638

639

// Utilities.

640

void cleanupTables();

641

std::pair<unsigned, unsigned> assignDFSNumbers(BasicBlock *, unsigned);

642

void updateProcessedCount(Value *V);

643

void verifyMemoryCongruency() const;

644

void verifyIterationSettled(Function &F);

645

void verifyStoreExpressions() const;

646

bool singleReachablePHIPath(const MemoryAccess *, const MemoryAccess *) const;

647

BasicBlock *getBlockForValue(Value *V) const;

648

void deleteExpression(const Expression *E) const;

649

unsigned InstrToDFSNum(const Value *V) const {

650

assert(isa<Instruction>(V) && "This should not be used for MemoryAccesses")((isa<Instruction>(V) && "This should not be used for MemoryAccesses"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(V) && \"This should not be used for MemoryAccesses\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 650, __PRETTY_FUNCTION__));

651

return InstrDFS.lookup(V);

652

}

653

654

unsigned InstrToDFSNum(const MemoryAccess *MA) const {

655

return MemoryToDFSNum(MA);

656

}

657

Value *InstrFromDFSNum(unsigned DFSNum) { return DFSToInstr[DFSNum]; }

658

// Given a MemoryAccess, return the relevant instruction DFS number. Note:

659

// This deliberately takes a value so it can be used with Use's, which will

660

// auto-convert to Value's but not to MemoryAccess's.

661

unsigned MemoryToDFSNum(const Value *MA) const {

662

assert(isa<MemoryAccess>(MA) &&((isa<MemoryAccess>(MA) && "This should not be used with instructions"
) ? static_cast<void> (0) : __assert_fail ("isa<MemoryAccess>(MA) && \"This should not be used with instructions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 663, __PRETTY_FUNCTION__))

663

"This should not be used with instructions")((isa<MemoryAccess>(MA) && "This should not be used with instructions"
) ? static_cast<void> (0) : __assert_fail ("isa<MemoryAccess>(MA) && \"This should not be used with instructions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 663, __PRETTY_FUNCTION__));

664

return isa<MemoryUseOrDef>(MA)

665

? InstrToDFSNum(cast<MemoryUseOrDef>(MA)->getMemoryInst())

666

: InstrDFS.lookup(MA);

667

}

668

bool isCycleFree(const PHINode *PN) const;

669

template <class T, class Range> T *getMinDFSOfRange(const Range &) const;

670

// Debug counter info. When verifying, we have to reset the value numbering

671

// debug counter to the same state it started in to get the same results.

672

std::pair<int, int> StartingVNCounter;

673

};

674

} // end anonymous namespace

675

676

template <typename T>

677

static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS) {

678

if (!isa<LoadExpression>(RHS) && !isa<StoreExpression>(RHS))

679

return false;

680

return LHS.MemoryExpression::equals(RHS);

681

}

682

683

bool LoadExpression::equals(const Expression &Other) const {

684

return equalsLoadStoreHelper(*this, Other);

685

}

686

687

bool StoreExpression::equals(const Expression &Other) const {

688

if (!equalsLoadStoreHelper(*this, Other))

689

return false;

690

// Make sure that store vs store includes the value operand.

691

if (const auto *S = dyn_cast<StoreExpression>(&Other))

692

if (getStoredValue() != S->getStoredValue())

693

return false;

694

return true;

695

}

696

697

#ifndef NDEBUG

698

static std::string getBlockName(const BasicBlock *B) {

699

return DOTGraphTraits<const Function *>::getSimpleNodeLabel(B, nullptr);

700

}

701

#endif

702

703

// Get the basic block from an instruction/memory value.

704

BasicBlock *NewGVN::getBlockForValue(Value *V) const {

705

if (auto *I = dyn_cast<Instruction>(V))

706

return I->getParent();

707

else if (auto *MP = dyn_cast<MemoryPhi>(V))

708

return MP->getBlock();

709

llvm_unreachable("Should have been able to figure out a block for our value")::llvm::llvm_unreachable_internal("Should have been able to figure out a block for our value"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 709);

710

return nullptr;

711

}

712

713

// Delete a definitely dead expression, so it can be reused by the expression

714

// allocator. Some of these are not in creation functions, so we have to accept

715

// const versions.

716

void NewGVN::deleteExpression(const Expression *E) const {

717

assert(isa<BasicExpression>(E))((isa<BasicExpression>(E)) ? static_cast<void> (0
) : __assert_fail ("isa<BasicExpression>(E)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 717, __PRETTY_FUNCTION__));

718

auto *BE = cast<BasicExpression>(E);

719

const_cast<BasicExpression *>(BE)->deallocateOperands(ArgRecycler);

720

ExpressionAllocator.Deallocate(E);

721

}

722

723

PHIExpression *NewGVN::createPHIExpression(Instruction *I, bool &HasBackedge,

724

bool &AllConstant) const {

725

BasicBlock *PHIBlock = I->getParent();

726

auto *PN = cast<PHINode>(I);

727

auto *E =

728

new (ExpressionAllocator) PHIExpression(PN->getNumOperands(), PHIBlock);

729

730

E->allocateOperands(ArgRecycler, ExpressionAllocator);

731

E->setType(I->getType());

732

E->setOpcode(I->getOpcode());

733

734

unsigned PHIRPO = RPOOrdering.lookup(DT->getNode(PHIBlock));

735

736

// NewGVN assumes the operands of a PHI node are in a consistent order across

737

// PHIs. LLVM doesn't seem to always guarantee this. While we need to fix

738

// this in LLVM at some point we don't want GVN to find wrong congruences.

739

// Therefore, here we sort uses in predecessor order.

740

// We're sorting the values by pointer. In theory this might be cause of

741

// non-determinism, but here we don't rely on the ordering for anything

742

// significant, e.g. we don't create new instructions based on it so we're

743

// fine.

744

SmallVector<const Use *, 4> PHIOperands;

745

for (const Use &U : PN->operands())

746

PHIOperands.push_back(&U);

747

std::sort(PHIOperands.begin(), PHIOperands.end(),

748

[&](const Use *U1, const Use *U2) {

749

return PN->getIncomingBlock(*U1) < PN->getIncomingBlock(*U2);

750

});

751

752

// Filter out unreachable phi operands.

753

auto Filtered = make_filter_range(PHIOperands, [&](const Use *U) {

754

return ReachableEdges.count({PN->getIncomingBlock(*U), PHIBlock});

755

});

756

757

std::transform(Filtered.begin(), Filtered.end(), op_inserter(E),

758

[&](const Use *U) -> Value * {

759

auto *BB = PN->getIncomingBlock(*U);

760

auto *DTN = DT->getNode(BB);

761

if (RPOOrdering.lookup(DTN) >= PHIRPO)

762

HasBackedge = true;

763

AllConstant &= isa<UndefValue>(*U) || isa<Constant>(*U);

764

765

// Don't try to transform self-defined phis.

766

if (*U == PN)

767

return PN;

768

return lookupOperandLeader(*U);

769

});

770

return E;

771

}

772

773

// Set basic expression info (Arguments, type, opcode) for Expression

774

// E from Instruction I in block B.

775

bool NewGVN::setBasicExpressionInfo(Instruction *I, BasicExpression *E) const {

776

bool AllConstant = true;

777

if (auto *GEP = dyn_cast<GetElementPtrInst>(I))

778

E->setType(GEP->getSourceElementType());

779

else

780

E->setType(I->getType());

781

E->setOpcode(I->getOpcode());

782

E->allocateOperands(ArgRecycler, ExpressionAllocator);

783

784

// Transform the operand array into an operand leader array, and keep track of

785

// whether all members are constant.

786

std::transform(I->op_begin(), I->op_end(), op_inserter(E), [&](Value *O) {

787

auto Operand = lookupOperandLeader(O);

788

AllConstant &= isa<Constant>(Operand);

789

return Operand;

790

});

791

792

return AllConstant;

793

}

794

795

const Expression *NewGVN::createBinaryExpression(unsigned Opcode, Type *T,

796

Value *Arg1,

797

Value *Arg2) const {

798

auto *E = new (ExpressionAllocator) BasicExpression(2);

799

800

E->setType(T);

801

E->setOpcode(Opcode);

802

E->allocateOperands(ArgRecycler, ExpressionAllocator);

803

if (Instruction::isCommutative(Opcode)) {

804

// Ensure that commutative instructions that only differ by a permutation

805

// of their operands get the same value number by sorting the operand value

806

// numbers. Since all commutative instructions have two operands it is more

807

// efficient to sort by hand rather than using, say, std::sort.

808

if (shouldSwapOperands(Arg1, Arg2))

809

std::swap(Arg1, Arg2);

810

}

811

E->op_push_back(lookupOperandLeader(Arg1));

812

E->op_push_back(lookupOperandLeader(Arg2));

813

814

Value *V = SimplifyBinOp(Opcode, E->getOperand(0), E->getOperand(1), SQ);

815

if (const Expression *SimplifiedE = checkSimplificationResults(E, nullptr, V))

816

return SimplifiedE;

817

return E;

818

}

819

820

// Take a Value returned by simplification of Expression E/Instruction

821

// I, and see if it resulted in a simpler expression. If so, return

822

// that expression.

823

// TODO: Once finished, this should not take an Instruction, we only

824

// use it for printing.

825

const Expression *NewGVN::checkSimplificationResults(Expression *E,

826

Instruction *I,

827

Value *V) const {

828

if (!V)

829

return nullptr;

830

if (auto *C = dyn_cast<Constant>(V)) {

831

if (I)

832

DEBUG(dbgs() << "Simplified " << *I << " to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " constant " << *C << "\n"; } } while
(false)

833

<< " constant " << *C << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " constant " << *C << "\n"; } } while
(false);

834

NumGVNOpsSimplified++;

835

assert(isa<BasicExpression>(E) &&((isa<BasicExpression>(E) && "We should always have had a basic expression here"
) ? static_cast<void> (0) : __assert_fail ("isa<BasicExpression>(E) && \"We should always have had a basic expression here\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 836, __PRETTY_FUNCTION__))

836

"We should always have had a basic expression here")((isa<BasicExpression>(E) && "We should always have had a basic expression here"
) ? static_cast<void> (0) : __assert_fail ("isa<BasicExpression>(E) && \"We should always have had a basic expression here\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 836, __PRETTY_FUNCTION__));

837

deleteExpression(E);

838

return createConstantExpression(C);

839

} else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {

840

if (I)

841

DEBUG(dbgs() << "Simplified " << *I << " to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " variable " << *V << "\n"; } } while
(false)

842

<< " variable " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " variable " << *V << "\n"; } } while
(false);

843

deleteExpression(E);

844

return createVariableExpression(V);

845

}

846

847

CongruenceClass *CC = ValueToClass.lookup(V);

848

if (CC && CC->getDefiningExpr()) {

849

if (I)

850

DEBUG(dbgs() << "Simplified " << *I << " to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " expression " << *V << "\n"; } }
while (false)

851

<< " expression " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " expression " << *V << "\n"; } }
while (false);

852

NumGVNOpsSimplified++;

853

deleteExpression(E);

854

return CC->getDefiningExpr();

855

}

856

return nullptr;

857

}

858

859

const Expression *NewGVN::createExpression(Instruction *I) const {

860

auto *E = new (ExpressionAllocator) BasicExpression(I->getNumOperands());

861

862

bool AllConstant = setBasicExpressionInfo(I, E);

863

864

if (I->isCommutative()) {

865

// Ensure that commutative instructions that only differ by a permutation

866

// of their operands get the same value number by sorting the operand value

867

// numbers. Since all commutative instructions have two operands it is more

868

// efficient to sort by hand rather than using, say, std::sort.

869

assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!")((I->getNumOperands() == 2 && "Unsupported commutative instruction!"
) ? static_cast<void> (0) : __assert_fail ("I->getNumOperands() == 2 && \"Unsupported commutative instruction!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 869, __PRETTY_FUNCTION__));

870

if (shouldSwapOperands(E->getOperand(0), E->getOperand(1)))

871

E->swapOperands(0, 1);

872

}

873

874

// Perform simplificaiton

875

// TODO: Right now we only check to see if we get a constant result.

876

// We may get a less than constant, but still better, result for

877

// some operations.

878

// IE

879

// add 0, x -> x

880

// and x, x -> x

881

// We should handle this by simply rewriting the expression.

882

if (auto *CI = dyn_cast<CmpInst>(I)) {

883

// Sort the operand value numbers so x<y and y>x get the same value

884

// number.

885

CmpInst::Predicate Predicate = CI->getPredicate();

886

if (shouldSwapOperands(E->getOperand(0), E->getOperand(1))) {

887

E->swapOperands(0, 1);

888

Predicate = CmpInst::getSwappedPredicate(Predicate);

889

}

890

E->setOpcode((CI->getOpcode() << 8) | Predicate);

891

// TODO: 25% of our time is spent in SimplifyCmpInst with pointer operands

892

assert(I->getOperand(0)->getType() == I->getOperand(1)->getType() &&((I->getOperand(0)->getType() == I->getOperand(1)->
getType() && "Wrong types on cmp instruction") ? static_cast
<void> (0) : __assert_fail ("I->getOperand(0)->getType() == I->getOperand(1)->getType() && \"Wrong types on cmp instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 893, __PRETTY_FUNCTION__))

893

"Wrong types on cmp instruction")((I->getOperand(0)->getType() == I->getOperand(1)->
getType() && "Wrong types on cmp instruction") ? static_cast
<void> (0) : __assert_fail ("I->getOperand(0)->getType() == I->getOperand(1)->getType() && \"Wrong types on cmp instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 893, __PRETTY_FUNCTION__));

894

assert((E->getOperand(0)->getType() == I->getOperand(0)->getType() &&(((E->getOperand(0)->getType() == I->getOperand(0)->
getType() && E->getOperand(1)->getType() == I->
getOperand(1)->getType())) ? static_cast<void> (0) :
__assert_fail ("(E->getOperand(0)->getType() == I->getOperand(0)->getType() && E->getOperand(1)->getType() == I->getOperand(1)->getType())"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 895, __PRETTY_FUNCTION__))

895

E->getOperand(1)->getType() == I->getOperand(1)->getType()))(((E->getOperand(0)->getType() == I->getOperand(0)->
getType() && E->getOperand(1)->getType() == I->
getOperand(1)->getType())) ? static_cast<void> (0) :
__assert_fail ("(E->getOperand(0)->getType() == I->getOperand(0)->getType() && E->getOperand(1)->getType() == I->getOperand(1)->getType())"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 895, __PRETTY_FUNCTION__));

896

Value *V =

897

SimplifyCmpInst(Predicate, E->getOperand(0), E->getOperand(1), SQ);

898

if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))

899

return SimplifiedE;

900

} else if (isa<SelectInst>(I)) {

901

if (isa<Constant>(E->getOperand(0)) ||

902

E->getOperand(0) == E->getOperand(1)) {

903

assert(E->getOperand(1)->getType() == I->getOperand(1)->getType() &&((E->getOperand(1)->getType() == I->getOperand(1)->
getType() && E->getOperand(2)->getType() == I->
getOperand(2)->getType()) ? static_cast<void> (0) : __assert_fail
("E->getOperand(1)->getType() == I->getOperand(1)->getType() && E->getOperand(2)->getType() == I->getOperand(2)->getType()"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 904, __PRETTY_FUNCTION__))

904

E->getOperand(2)->getType() == I->getOperand(2)->getType())((E->getOperand(1)->getType() == I->getOperand(1)->
getType() && E->getOperand(2)->getType() == I->
getOperand(2)->getType()) ? static_cast<void> (0) : __assert_fail
("E->getOperand(1)->getType() == I->getOperand(1)->getType() && E->getOperand(2)->getType() == I->getOperand(2)->getType()"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 904, __PRETTY_FUNCTION__));

905

Value *V = SimplifySelectInst(E->getOperand(0), E->getOperand(1),

906

E->getOperand(2), SQ);

907

if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))

908

return SimplifiedE;

909

}

910

} else if (I->isBinaryOp()) {

911

Value *V =

912

SimplifyBinOp(E->getOpcode(), E->getOperand(0), E->getOperand(1), SQ);

913

if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))

914

return SimplifiedE;

915

} else if (auto *BI = dyn_cast<BitCastInst>(I)) {

916

Value *V =

917

SimplifyCastInst(BI->getOpcode(), BI->getOperand(0), BI->getType(), SQ);

918

if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))

919

return SimplifiedE;

920

} else if (isa<GetElementPtrInst>(I)) {

921

Value *V = SimplifyGEPInst(

922

E->getType(), ArrayRef<Value *>(E->op_begin(), E->op_end()), SQ);

923

if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))

924

return SimplifiedE;

925

} else if (AllConstant) {

926

// We don't bother trying to simplify unless all of the operands

927

// were constant.

928

// TODO: There are a lot of Simplify*'s we could call here, if we

929

// wanted to. The original motivating case for this code was a

930

// zext i1 false to i8, which we don't have an interface to

931

// simplify (IE there is no SimplifyZExt).

932

933

SmallVector<Constant *, 8> C;

934

for (Value *Arg : E->operands())

935

C.emplace_back(cast<Constant>(Arg));

936

937

if (Value *V = ConstantFoldInstOperands(I, C, DL, TLI))

938

if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))

939

return SimplifiedE;

940

}

941

return E;

942

}

943

944

const AggregateValueExpression *

945

NewGVN::createAggregateValueExpression(Instruction *I) const {

946

if (auto *II = dyn_cast<InsertValueInst>(I)) {

947

auto *E = new (ExpressionAllocator)

948

AggregateValueExpression(I->getNumOperands(), II->getNumIndices());

949

setBasicExpressionInfo(I, E);

950

E->allocateIntOperands(ExpressionAllocator);

951

std::copy(II->idx_begin(), II->idx_end(), int_op_inserter(E));

952

return E;

953

} else if (auto *EI = dyn_cast<ExtractValueInst>(I)) {

954

auto *E = new (ExpressionAllocator)

955

AggregateValueExpression(I->getNumOperands(), EI->getNumIndices());

956

setBasicExpressionInfo(EI, E);

957

E->allocateIntOperands(ExpressionAllocator);

958

std::copy(EI->idx_begin(), EI->idx_end(), int_op_inserter(E));

959

return E;

960

}

961

llvm_unreachable("Unhandled type of aggregate value operation")::llvm::llvm_unreachable_internal("Unhandled type of aggregate value operation"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 961);

962

}

963

964

const VariableExpression *NewGVN::createVariableExpression(Value *V) const {

965

auto *E = new (ExpressionAllocator) VariableExpression(V);

966

E->setOpcode(V->getValueID());

967

return E;

968

}

969

970

const Expression *NewGVN::createVariableOrConstant(Value *V) const {

971

if (auto *C = dyn_cast<Constant>(V))

972

return createConstantExpression(C);

973

return createVariableExpression(V);

974

}

975

976

const ConstantExpression *NewGVN::createConstantExpression(Constant *C) const {

977

auto *E = new (ExpressionAllocator) ConstantExpression(C);

978

E->setOpcode(C->getValueID());

979

return E;

980

}

981

982

const UnknownExpression *NewGVN::createUnknownExpression(Instruction *I) const {

983

auto *E = new (ExpressionAllocator) UnknownExpression(I);

984

E->setOpcode(I->getOpcode());

985

return E;

986

}

987

988

const CallExpression *

989

NewGVN::createCallExpression(CallInst *CI, const MemoryAccess *MA) const {

990

// FIXME: Add operand bundles for calls.

991

auto *E =

992

new (ExpressionAllocator) CallExpression(CI->getNumOperands(), CI, MA);

993

setBasicExpressionInfo(CI, E);

994

return E;

995

}

996

997

// Return true if some equivalent of instruction Inst dominates instruction U.

998

bool NewGVN::someEquivalentDominates(const Instruction *Inst,

999

const Instruction *U) const {

1000

auto *CC = ValueToClass.lookup(Inst);

1001

// This must be an instruction because we are only called from phi nodes

1002

// in the case that the value it needs to check against is an instruction.

1003

1004

// The most likely candiates for dominance are the leader and the next leader.

1005

// The leader or nextleader will dominate in all cases where there is an

1006

// equivalent that is higher up in the dom tree.

1007

// We can't *only* check them, however, because the

1008

// dominator tree could have an infinite number of non-dominating siblings

1009

// with instructions that are in the right congruence class.

1010

// A

1011

// B C D E F G

1012

// |

1013

// H

1014

// Instruction U could be in H, with equivalents in every other sibling.

1015

// Depending on the rpo order picked, the leader could be the equivalent in

1016

// any of these siblings.

1017

if (!CC)

1018

return false;

1019

if (DT->dominates(cast<Instruction>(CC->getLeader()), U))

1020

return true;

1021

if (CC->getNextLeader().first &&

1022

DT->dominates(cast<Instruction>(CC->getNextLeader().first), U))

1023

return true;

1024

return llvm::any_of(*CC, [&](const Value *Member) {

1025

return Member != CC->getLeader() &&

1026

DT->dominates(cast<Instruction>(Member), U);

1027

});

1028

}

1029

1030

// See if we have a congruence class and leader for this operand, and if so,

1031

// return it. Otherwise, return the operand itself.

1032

Value *NewGVN::lookupOperandLeader(Value *V) const {

1033

CongruenceClass *CC = ValueToClass.lookup(V);

1034

if (CC) {

1035

// Everything in TOP is represneted by undef, as it can be any value.

1036

// We do have to make sure we get the type right though, so we can't set the

1037

// RepLeader to undef.

1038

if (CC == TOPClass)

1039

return UndefValue::get(V->getType());

1040

return CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();

1041

}

1042

1043

return V;

1044

}

1045

1046

const MemoryAccess *NewGVN::lookupMemoryLeader(const MemoryAccess *MA) const {

1047

auto *CC = getMemoryClass(MA);

1048

assert(CC->getMemoryLeader() &&((CC->getMemoryLeader() && "Every MemoryAccess should be mapped to a congruence class with a "
"representative memory access") ? static_cast<void> (0
) : __assert_fail ("CC->getMemoryLeader() && \"Every MemoryAccess should be mapped to a congruence class with a \" \"representative memory access\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1050, __PRETTY_FUNCTION__))

1049

"Every MemoryAccess should be mapped to a congruence class with a "((CC->getMemoryLeader() && "Every MemoryAccess should be mapped to a congruence class with a "
"representative memory access") ? static_cast<void> (0
) : __assert_fail ("CC->getMemoryLeader() && \"Every MemoryAccess should be mapped to a congruence class with a \" \"representative memory access\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1050, __PRETTY_FUNCTION__))

1050

"representative memory access")((CC->getMemoryLeader() && "Every MemoryAccess should be mapped to a congruence class with a "
"representative memory access") ? static_cast<void> (0
) : __assert_fail ("CC->getMemoryLeader() && \"Every MemoryAccess should be mapped to a congruence class with a \" \"representative memory access\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1050, __PRETTY_FUNCTION__));

1051

return CC->getMemoryLeader();

1052

}

1053

1054

// Return true if the MemoryAccess is really equivalent to everything. This is

1055

// equivalent to the lattice value "TOP" in most lattices. This is the initial

1056

// state of all MemoryAccesses.

1057

bool NewGVN::isMemoryAccessTop(const MemoryAccess *MA) const {

1058

return getMemoryClass(MA) == TOPClass;

1059

}

1060

1061

LoadExpression *NewGVN::createLoadExpression(Type *LoadType, Value *PointerOp,

1062

LoadInst *LI,

1063

const MemoryAccess *MA) const {

1064

auto *E =

1065

new (ExpressionAllocator) LoadExpression(1, LI, lookupMemoryLeader(MA));

1066

E->allocateOperands(ArgRecycler, ExpressionAllocator);

1067

E->setType(LoadType);

1068

1069

// Give store and loads same opcode so they value number together.

1070

E->setOpcode(0);

1071

E->op_push_back(PointerOp);

1072

if (LI)

1073

E->setAlignment(LI->getAlignment());

1074

1075

// TODO: Value number heap versions. We may be able to discover

1076

// things alias analysis can't on it's own (IE that a store and a

1077

// load have the same value, and thus, it isn't clobbering the load).

1078

return E;

1079

}

1080

1081

const StoreExpression *

1082

NewGVN::createStoreExpression(StoreInst *SI, const MemoryAccess *MA) const {

1083

auto *StoredValueLeader = lookupOperandLeader(SI->getValueOperand());

1084

auto *E = new (ExpressionAllocator)

1085

StoreExpression(SI->getNumOperands(), SI, StoredValueLeader, MA);

1086

E->allocateOperands(ArgRecycler, ExpressionAllocator);

1087

E->setType(SI->getValueOperand()->getType());

1088

1089

// Give store and loads same opcode so they value number together.

1090

E->setOpcode(0);

1091

E->op_push_back(lookupOperandLeader(SI->getPointerOperand()));

1092

1093

// TODO: Value number heap versions. We may be able to discover

1094

// things alias analysis can't on it's own (IE that a store and a

1095

// load have the same value, and thus, it isn't clobbering the load).

1096

return E;

1097

}

1098

1099

const Expression *NewGVN::performSymbolicStoreEvaluation(Instruction *I) const {

1100

// Unlike loads, we never try to eliminate stores, so we do not check if they

1101

// are simple and avoid value numbering them.

1102

auto *SI = cast<StoreInst>(I);

1103

auto *StoreAccess = MSSA->getMemoryAccess(SI);

1104

// Get the expression, if any, for the RHS of the MemoryDef.

1105

const MemoryAccess *StoreRHS = StoreAccess->getDefiningAccess();

1106

if (EnableStoreRefinement)

1107

StoreRHS = MSSAWalker->getClobberingMemoryAccess(StoreAccess);

1108

// If we bypassed the use-def chains, make sure we add a use.

1109

if (StoreRHS != StoreAccess->getDefiningAccess())

1110

addMemoryUsers(StoreRHS, StoreAccess);

1111

1112

StoreRHS = lookupMemoryLeader(StoreRHS);

1113

// If we are defined by ourselves, use the live on entry def.

1114

if (StoreRHS == StoreAccess)

1115

StoreRHS = MSSA->getLiveOnEntryDef();

1116

1117

if (SI->isSimple()) {

1118

// See if we are defined by a previous store expression, it already has a

1119

// value, and it's the same value as our current store. FIXME: Right now, we

1120

// only do this for simple stores, we should expand to cover memcpys, etc.

1121

const auto *LastStore = createStoreExpression(SI, StoreRHS);

1122

const auto *LastCC = ExpressionToClass.lookup(LastStore);

1123

// Basically, check if the congruence class the store is in is defined by a

1124

// store that isn't us, and has the same value. MemorySSA takes care of

1125

// ensuring the store has the same memory state as us already.

1126

// The RepStoredValue gets nulled if all the stores disappear in a class, so

1127

// we don't need to check if the class contains a store besides us.

1128

if (LastCC &&

1129

LastCC->getStoredValue() == lookupOperandLeader(SI->getValueOperand()))

1130

return LastStore;

1131

deleteExpression(LastStore);

1132

// Also check if our value operand is defined by a load of the same memory

1133

// location, and the memory state is the same as it was then (otherwise, it

1134

// could have been overwritten later. See test32 in

1135

// transforms/DeadStoreElimination/simple.ll).

1136

if (auto *LI =

1137

dyn_cast<LoadInst>(lookupOperandLeader(SI->getValueOperand()))) {

1138

if ((lookupOperandLeader(LI->getPointerOperand()) ==

1139

lookupOperandLeader(SI->getPointerOperand())) &&

1140

(lookupMemoryLeader(MSSA->getMemoryAccess(LI)->getDefiningAccess()) ==

1141

StoreRHS))

1142

return createVariableExpression(LI);

1143

}

1144

}

1145

1146

// If the store is not equivalent to anything, value number it as a store that

1147

// produces a unique memory state (instead of using it's MemoryUse, we use

1148

// it's MemoryDef).

1149

return createStoreExpression(SI, StoreAccess);

1150

}

1151

1152

// See if we can extract the value of a loaded pointer from a load, a store, or

1153

// a memory instruction.

1154

const Expression *

1155

NewGVN::performSymbolicLoadCoercion(Type *LoadType, Value *LoadPtr,

1156

LoadInst *LI, Instruction *DepInst,

1157

MemoryAccess *DefiningAccess) const {

1158

assert((!LI || LI->isSimple()) && "Not a simple load")(((!LI || LI->isSimple()) && "Not a simple load") ?
static_cast<void> (0) : __assert_fail ("(!LI || LI->isSimple()) && \"Not a simple load\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1158, __PRETTY_FUNCTION__));

Within the expansion of the macro 'assert':

→

a	Assuming 'LI' is null

1159

if (auto *DepSI = dyn_cast<StoreInst>(DepInst)) {

←

Taking false branch

→

1160

// Can't forward from non-atomic to atomic without violating memory model.

1161

// Also don't need to coerce if they are the same type, we will just

1162

// propogate..

1163

if (LI->isAtomic() > DepSI->isAtomic() ||

1164

LoadType == DepSI->getValueOperand()->getType())

1165

return nullptr;

1166

int Offset = analyzeLoadFromClobberingStore(LoadType, LoadPtr, DepSI, DL);

1167

if (Offset >= 0) {

1168

if (auto *C = dyn_cast<Constant>(

1169

lookupOperandLeader(DepSI->getValueOperand()))) {

1170

DEBUG(dbgs() << "Coercing load from store " << *DepSI << " to constant "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Coercing load from store " <<
*DepSI << " to constant " << *C << "\n"; }
} while (false)

1171

<< *C << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Coercing load from store " <<
*DepSI << " to constant " << *C << "\n"; }
} while (false);

1172

return createConstantExpression(

1173

getConstantStoreValueForLoad(C, Offset, LoadType, DL));

1174

}

1175

}

1176

1177

} else if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {

←

Assuming 'DepLI' is non-null

→

←

Taking true branch

→

1178

// Can't forward from non-atomic to atomic without violating memory model.

1179

if (LI->isAtomic() > DepLI->isAtomic())

←

Called C++ object pointer is null

1180

return nullptr;

1181

int Offset = analyzeLoadFromClobberingLoad(LoadType, LoadPtr, DepLI, DL);

1182

if (Offset >= 0) {

1183

// We can coerce a constant load into a load

1184

if (auto *C = dyn_cast<Constant>(lookupOperandLeader(DepLI)))

1185

if (auto *PossibleConstant =

1186

getConstantLoadValueForLoad(C, Offset, LoadType, DL)) {

1187

DEBUG(dbgs() << "Coercing load from load " << *LI << " to constant "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Coercing load from load " <<
*LI << " to constant " << *PossibleConstant <<
"\n"; } } while (false)

1188

<< *PossibleConstant << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Coercing load from load " <<
*LI << " to constant " << *PossibleConstant <<
"\n"; } } while (false);

1189

return createConstantExpression(PossibleConstant);

1190

}

1191

}

1192

1193

} else if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {

1194

int Offset = analyzeLoadFromClobberingMemInst(LoadType, LoadPtr, DepMI, DL);

1195

if (Offset >= 0) {

1196

if (auto *PossibleConstant =

1197

getConstantMemInstValueForLoad(DepMI, Offset, LoadType, DL)) {

1198

DEBUG(dbgs() << "Coercing load from meminst " << *DepMIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Coercing load from meminst " <<
*DepMI << " to constant " << *PossibleConstant <<
"\n"; } } while (false)

1199

<< " to constant " << *PossibleConstant << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Coercing load from meminst " <<
*DepMI << " to constant " << *PossibleConstant <<
"\n"; } } while (false);

1200

return createConstantExpression(PossibleConstant);

1201

}

1202

}

1203

}

1204

1205

// All of the below are only true if the loaded pointer is produced

1206

// by the dependent instruction.

1207

if (LoadPtr != lookupOperandLeader(DepInst) &&

1208

!AA->isMustAlias(LoadPtr, DepInst))

1209

return nullptr;

1210

// If this load really doesn't depend on anything, then we must be loading an

1211

// undef value. This can happen when loading for a fresh allocation with no

1212

// intervening stores, for example. Note that this is only true in the case

1213

// that the result of the allocation is pointer equal to the load ptr.

1214

if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {

1215

return createConstantExpression(UndefValue::get(LoadType));

1216

}

1217

// If this load occurs either right after a lifetime begin,

1218

// then the loaded value is undefined.

1219

else if (auto *II = dyn_cast<IntrinsicInst>(DepInst)) {

1220

if (II->getIntrinsicID() == Intrinsic::lifetime_start)

1221

return createConstantExpression(UndefValue::get(LoadType));

1222

}

1223

// If this load follows a calloc (which zero initializes memory),

1224

// then the loaded value is zero

1225

else if (isCallocLikeFn(DepInst, TLI)) {

1226

return createConstantExpression(Constant::getNullValue(LoadType));

1227

}

1228

1229

return nullptr;

1230

}

1231

1232

const Expression *NewGVN::performSymbolicLoadEvaluation(Instruction *I) const {

1233

auto *LI = cast<LoadInst>(I);

1234

1235

// We can eliminate in favor of non-simple loads, but we won't be able to

1236

// eliminate the loads themselves.

1237

if (!LI->isSimple())

1238

return nullptr;

1239

1240

Value *LoadAddressLeader = lookupOperandLeader(LI->getPointerOperand());

1241

// Load of undef is undef.

1242

if (isa<UndefValue>(LoadAddressLeader))

1243

return createConstantExpression(UndefValue::get(LI->getType()));

1244

1245

MemoryAccess *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(I);

1246

1247

if (!MSSA->isLiveOnEntryDef(DefiningAccess)) {

1248

if (auto *MD = dyn_cast<MemoryDef>(DefiningAccess)) {

1249

Instruction *DefiningInst = MD->getMemoryInst();

1250

// If the defining instruction is not reachable, replace with undef.

1251

if (!ReachableBlocks.count(DefiningInst->getParent()))

1252

return createConstantExpression(UndefValue::get(LI->getType()));

1253

// This will handle stores and memory insts. We only do if it the

1254

// defining access has a different type, or it is a pointer produced by

1255

// certain memory operations that cause the memory to have a fixed value

1256

// (IE things like calloc).

1257

if (const auto *CoercionResult =

1258

performSymbolicLoadCoercion(LI->getType(), LoadAddressLeader, LI,

1259

DefiningInst, DefiningAccess))

1260

return CoercionResult;

1261

}

1262

}

1263

1264

const Expression *E = createLoadExpression(LI->getType(), LoadAddressLeader,

1265

LI, DefiningAccess);

1266

return E;

1267

}

1268

1269

const Expression *

1270

NewGVN::performSymbolicPredicateInfoEvaluation(Instruction *I) const {

1271

auto *PI = PredInfo->getPredicateInfoFor(I);

1272

if (!PI)

1273

return nullptr;

1274

1275

DEBUG(dbgs() << "Found predicate info from instruction !\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found predicate info from instruction !\n"
; } } while (false);

1276

1277

auto *PWC = dyn_cast<PredicateWithCondition>(PI);

1278

if (!PWC)

1279

return nullptr;

1280

1281

auto *CopyOf = I->getOperand(0);

1282

auto *Cond = PWC->Condition;

1283

1284

// If this a copy of the condition, it must be either true or false depending

1285

// on the predicate info type and edge

1286

if (CopyOf == Cond) {

1287

// We should not need to add predicate users because the predicate info is

1288

// already a use of this operand.

1289

if (isa<PredicateAssume>(PI))

1290

return createConstantExpression(ConstantInt::getTrue(Cond->getType()));

1291

if (auto *PBranch = dyn_cast<PredicateBranch>(PI)) {

1292

if (PBranch->TrueEdge)

1293

return createConstantExpression(ConstantInt::getTrue(Cond->getType()));

1294

return createConstantExpression(ConstantInt::getFalse(Cond->getType()));

1295

}

1296

if (auto *PSwitch = dyn_cast<PredicateSwitch>(PI))

1297

return createConstantExpression(cast<Constant>(PSwitch->CaseValue));

1298

}

1299

1300

// Not a copy of the condition, so see what the predicates tell us about this

1301

// value. First, though, we check to make sure the value is actually a copy

1302

// of one of the condition operands. It's possible, in certain cases, for it

1303

// to be a copy of a predicateinfo copy. In particular, if two branch

1304

// operations use the same condition, and one branch dominates the other, we

1305

// will end up with a copy of a copy. This is currently a small deficiency in

1306

// predicateinfo. What will end up happening here is that we will value

1307

// number both copies the same anyway.

1308

1309

// Everything below relies on the condition being a comparison.

1310

auto *Cmp = dyn_cast<CmpInst>(Cond);

1311

if (!Cmp)

1312

return nullptr;

1313

1314

if (CopyOf != Cmp->getOperand(0) && CopyOf != Cmp->getOperand(1)) {

1315

DEBUG(dbgs() << "Copy is not of any condition operands!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Copy is not of any condition operands!\n"
; } } while (false);

1316

return nullptr;

1317

}

1318

Value *FirstOp = lookupOperandLeader(Cmp->getOperand(0));

1319

Value *SecondOp = lookupOperandLeader(Cmp->getOperand(1));

1320

bool SwappedOps = false;

1321

// Sort the ops

1322

if (shouldSwapOperands(FirstOp, SecondOp)) {

1323

std::swap(FirstOp, SecondOp);

1324

SwappedOps = true;

1325

}

1326

CmpInst::Predicate Predicate =

1327

SwappedOps ? Cmp->getSwappedPredicate() : Cmp->getPredicate();

1328

1329

if (isa<PredicateAssume>(PI)) {

1330

// If the comparison is true when the operands are equal, then we know the

1331

// operands are equal, because assumes must always be true.

1332

if (CmpInst::isTrueWhenEqual(Predicate)) {

1333

addPredicateUsers(PI, I);

1334

return createVariableOrConstant(FirstOp);

1335

}

1336

}

1337

if (const auto *PBranch = dyn_cast<PredicateBranch>(PI)) {

1338

// If we are *not* a copy of the comparison, we may equal to the other

1339

// operand when the predicate implies something about equality of

1340

// operations. In particular, if the comparison is true/false when the

1341

// operands are equal, and we are on the right edge, we know this operation

1342

// is equal to something.

1343

if ((PBranch->TrueEdge && Predicate == CmpInst::ICMP_EQ) ||

1344

(!PBranch->TrueEdge && Predicate == CmpInst::ICMP_NE)) {

1345

addPredicateUsers(PI, I);

1346

return createVariableOrConstant(FirstOp);

1347

}

1348

// Handle the special case of floating point.

1349

if (((PBranch->TrueEdge && Predicate == CmpInst::FCMP_OEQ) ||

1350

(!PBranch->TrueEdge && Predicate == CmpInst::FCMP_UNE)) &&

1351

isa<ConstantFP>(FirstOp) && !cast<ConstantFP>(FirstOp)->isZero()) {

1352

addPredicateUsers(PI, I);

1353

return createConstantExpression(cast<Constant>(FirstOp));

1354

}

1355

}

1356

return nullptr;

1357

}

1358

1359

// Evaluate read only and pure calls, and create an expression result.

1360

const Expression *NewGVN::performSymbolicCallEvaluation(Instruction *I) const {

1361

auto *CI = cast<CallInst>(I);

1362

if (auto *II = dyn_cast<IntrinsicInst>(I)) {

1363

// Instrinsics with the returned attribute are copies of arguments.

1364

if (auto *ReturnedValue = II->getReturnedArgOperand()) {

1365

if (II->getIntrinsicID() == Intrinsic::ssa_copy)

1366

if (const auto *Result = performSymbolicPredicateInfoEvaluation(I))

1367

return Result;

1368

return createVariableOrConstant(ReturnedValue);

1369

}

1370

}

1371

if (AA->doesNotAccessMemory(CI)) {

1372

return createCallExpression(CI, TOPClass->getMemoryLeader());

1373

} else if (AA->onlyReadsMemory(CI)) {

1374

MemoryAccess *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(CI);

1375

return createCallExpression(CI, DefiningAccess);

1376

}

1377

return nullptr;

1378

}

1379

1380

// Retrieve the memory class for a given MemoryAccess.

1381

CongruenceClass *NewGVN::getMemoryClass(const MemoryAccess *MA) const {

1382

1383

auto *Result = MemoryAccessToClass.lookup(MA);

1384

assert(Result && "Should have found memory class")((Result && "Should have found memory class") ? static_cast
<void> (0) : __assert_fail ("Result && \"Should have found memory class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1384, __PRETTY_FUNCTION__));

1385

return Result;

1386

}

1387

1388

// Update the MemoryAccess equivalence table to say that From is equal to To,

1389

// and return true if this is different from what already existed in the table.

1390

bool NewGVN::setMemoryClass(const MemoryAccess *From,

1391

CongruenceClass *NewClass) {

1392

assert(NewClass &&((NewClass && "Every MemoryAccess should be getting mapped to a non-null class"
) ? static_cast<void> (0) : __assert_fail ("NewClass && \"Every MemoryAccess should be getting mapped to a non-null class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1393, __PRETTY_FUNCTION__))

1393

"Every MemoryAccess should be getting mapped to a non-null class")((NewClass && "Every MemoryAccess should be getting mapped to a non-null class"
) ? static_cast<void> (0) : __assert_fail ("NewClass && \"Every MemoryAccess should be getting mapped to a non-null class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1393, __PRETTY_FUNCTION__));

1394

DEBUG(dbgs() << "Setting " << *From)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Setting " << *From; } } while
(false);

1395

DEBUG(dbgs() << " equivalent to congruence class ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << " equivalent to congruence class "
; } } while (false);

1396

DEBUG(dbgs() << NewClass->getID() << " with current MemoryAccess leader ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << NewClass->getID() << " with current MemoryAccess leader "
; } } while (false);

1397

DEBUG(dbgs() << *NewClass->getMemoryLeader() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << *NewClass->getMemoryLeader()
<< "\n"; } } while (false);

1398

1399

auto LookupResult = MemoryAccessToClass.find(From);

1400

bool Changed = false;

1401

// If it's already in the table, see if the value changed.

1402

if (LookupResult != MemoryAccessToClass.end()) {

1403

auto *OldClass = LookupResult->second;

1404

if (OldClass != NewClass) {

1405

// If this is a phi, we have to handle memory member updates.

1406

if (auto *MP = dyn_cast<MemoryPhi>(From)) {

1407

OldClass->memory_erase(MP);

1408

NewClass->memory_insert(MP);

1409

// This may have killed the class if it had no non-memory members

1410

if (OldClass->getMemoryLeader() == From) {

1411

if (OldClass->definesNoMemory()) {

1412

OldClass->setMemoryLeader(nullptr);

1413

} else {

1414

OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));

1415

DEBUG(dbgs() << "Memory class leader change for class "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory class leader change for class "
<< OldClass->getID() << " to " << *OldClass
->getMemoryLeader() << " due to removal of a memory member "
<< *From << "\n"; } } while (false)

1416

<< OldClass->getID() << " to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory class leader change for class "
<< OldClass->getID() << " to " << *OldClass
->getMemoryLeader() << " due to removal of a memory member "
<< *From << "\n"; } } while (false)

1417

<< *OldClass->getMemoryLeader()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory class leader change for class "
<< OldClass->getID() << " to " << *OldClass
->getMemoryLeader() << " due to removal of a memory member "
<< *From << "\n"; } } while (false)

1418

<< " due to removal of a memory member " << *Fromdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory class leader change for class "
<< OldClass->getID() << " to " << *OldClass
->getMemoryLeader() << " due to removal of a memory member "
<< *From << "\n"; } } while (false)

1419

<< "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory class leader change for class "
<< OldClass->getID() << " to " << *OldClass
->getMemoryLeader() << " due to removal of a memory member "
<< *From << "\n"; } } while (false);

1420

markMemoryLeaderChangeTouched(OldClass);

1421

}

1422

}

1423

}

1424

// It wasn't equivalent before, and now it is.

1425

LookupResult->second = NewClass;

1426

Changed = true;

1427

}

1428

}

1429

1430

return Changed;

1431

}

1432

1433

// Determine if a phi is cycle-free. That means the values in the phi don't

1434

// depend on any expressions that can change value as a result of the phi.

1435

// For example, a non-cycle free phi would be v = phi(0, v+1).

1436

bool NewGVN::isCycleFree(const PHINode *PN) const {

1437

// In order to compute cycle-freeness, we do SCC finding on the phi, and see

1438

// what kind of SCC it ends up in. If it is a singleton, it is cycle-free.

1439

// If it is not in a singleton, it is only cycle free if the other members are

1440

// all phi nodes (as they do not compute anything, they are copies). TODO:

1441

// There are likely a few other intrinsics or expressions that could be

1442

// included here, but this happens so infrequently already that it is not

1443

// likely to be worth it.

1444

auto PCS = PhiCycleState.lookup(PN);

1445

if (PCS == PCS_Unknown) {

1446

SCCFinder.Start(PN);

1447

auto &SCC = SCCFinder.getComponentFor(PN);

1448

// It's cycle free if it's size 1 or or the SCC is *only* phi nodes.

1449

if (SCC.size() == 1)

1450

PhiCycleState.insert({PN, PCS_CycleFree});

1451

else {

1452

bool AllPhis =

1453

llvm::all_of(SCC, [](const Value *V) { return isa<PHINode>(V); });

1454

PCS = AllPhis ? PCS_CycleFree : PCS_Cycle;

1455

for (auto *Member : SCC)

1456

if (auto *MemberPhi = dyn_cast<PHINode>(Member))

1457

PhiCycleState.insert({MemberPhi, PCS});

1458

}

1459

}

1460

if (PCS == PCS_Cycle)

1461

return false;

1462

return true;

1463

}

1464

1465

// Evaluate PHI nodes symbolically, and create an expression result.

1466

const Expression *NewGVN::performSymbolicPHIEvaluation(Instruction *I) const {

1467

// True if one of the incoming phi edges is a backedge.

1468

bool HasBackedge = false;

1469

// All constant tracks the state of whether all the *original* phi operands

1470

// were constant. This is really shorthand for "this phi cannot cycle due

1471

// to forward propagation", as any change in value of the phi is guaranteed

1472

// not to later change the value of the phi.

1473

// IE it can't be v = phi(undef, v+1)

1474

bool AllConstant = true;

1475

auto *E =

1476

cast<PHIExpression>(createPHIExpression(I, HasBackedge, AllConstant));

1477

// We match the semantics of SimplifyPhiNode from InstructionSimplify here.

1478

// See if all arguments are the same.

1479

// We track if any were undef because they need special handling.

1480

bool HasUndef = false;

1481

auto Filtered = make_filter_range(E->operands(), [&](const Value *Arg) {

1482

if (Arg == I)

1483

return false;

1484

if (isa<UndefValue>(Arg)) {

1485

HasUndef = true;

1486

return false;

1487

}

1488

return true;

1489

});

1490

// If we are left with no operands, it's undef

1491

if (Filtered.begin() == Filtered.end()) {

1492

DEBUG(dbgs() << "Simplified PHI node " << *I << " to undef"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified PHI node " <<
*I << " to undef" << "\n"; } } while (false)

1493

<< "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified PHI node " <<
*I << " to undef" << "\n"; } } while (false);

1494

deleteExpression(E);

1495

return createConstantExpression(UndefValue::get(I->getType()));

1496

}

1497

unsigned NumOps = 0;

1498

Value *AllSameValue = *(Filtered.begin());

1499

++Filtered.begin();

1500

// Can't use std::equal here, sadly, because filter.begin moves.

1501

if (llvm::all_of(Filtered, [AllSameValue, &NumOps](const Value *V) {

1502

++NumOps;

1503

return V == AllSameValue;

1504

})) {

1505

// In LLVM's non-standard representation of phi nodes, it's possible to have

1506

// phi nodes with cycles (IE dependent on other phis that are .... dependent

1507

// on the original phi node), especially in weird CFG's where some arguments

1508

// are unreachable, or uninitialized along certain paths. This can cause

1509

// infinite loops during evaluation. We work around this by not trying to

1510

// really evaluate them independently, but instead using a variable

1511

// expression to say if one is equivalent to the other.

1512

// We also special case undef, so that if we have an undef, we can't use the

1513

// common value unless it dominates the phi block.

1514

if (HasUndef) {

1515

// If we have undef and at least one other value, this is really a

1516

// multivalued phi, and we need to know if it's cycle free in order to

1517

// evaluate whether we can ignore the undef. The other parts of this are

1518

// just shortcuts. If there is no backedge, or all operands are

1519

// constants, or all operands are ignored but the undef, it also must be

1520

// cycle free.

1521

if (!AllConstant && HasBackedge && NumOps > 0 &&

1522

!isa<UndefValue>(AllSameValue) && !isCycleFree(cast<PHINode>(I)))

1523

return E;

1524

1525

// Only have to check for instructions

1526

if (auto *AllSameInst = dyn_cast<Instruction>(AllSameValue))

1527

if (!someEquivalentDominates(AllSameInst, I))

1528

return E;

1529

}

1530

1531

NumGVNPhisAllSame++;

1532

DEBUG(dbgs() << "Simplified PHI node " << *I << " to " << *AllSameValuedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified PHI node " <<
*I << " to " << *AllSameValue << "\n"; } }
while (false)

1533

<< "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified PHI node " <<
*I << " to " << *AllSameValue << "\n"; } }
while (false);

1534

deleteExpression(E);

1535

return createVariableOrConstant(AllSameValue);

1536

}

1537

return E;

1538

}

1539

1540

const Expression *

1541

NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) const {

1542

if (auto *EI = dyn_cast<ExtractValueInst>(I)) {

1543

auto *II = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());

1544

if (II && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {

1545

unsigned Opcode = 0;

1546

// EI might be an extract from one of our recognised intrinsics. If it

1547

// is we'll synthesize a semantically equivalent expression instead on

1548

// an extract value expression.

1549

switch (II->getIntrinsicID()) {

1550

case Intrinsic::sadd_with_overflow:

1551

case Intrinsic::uadd_with_overflow:

1552

Opcode = Instruction::Add;

1553

break;

1554

case Intrinsic::ssub_with_overflow:

1555

case Intrinsic::usub_with_overflow:

1556

Opcode = Instruction::Sub;

1557

break;

1558

case Intrinsic::smul_with_overflow:

1559

case Intrinsic::umul_with_overflow:

1560

Opcode = Instruction::Mul;

1561

break;

1562

default:

1563

break;

1564

}

1565

1566

if (Opcode != 0) {

1567

// Intrinsic recognized. Grab its args to finish building the

1568

// expression.

1569

assert(II->getNumArgOperands() == 2 &&((II->getNumArgOperands() == 2 && "Expect two args for recognised intrinsics."
) ? static_cast<void> (0) : __assert_fail ("II->getNumArgOperands() == 2 && \"Expect two args for recognised intrinsics.\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1570, __PRETTY_FUNCTION__))

1570

"Expect two args for recognised intrinsics.")((II->getNumArgOperands() == 2 && "Expect two args for recognised intrinsics."
) ? static_cast<void> (0) : __assert_fail ("II->getNumArgOperands() == 2 && \"Expect two args for recognised intrinsics.\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1570, __PRETTY_FUNCTION__));

1571

return createBinaryExpression(

1572

Opcode, EI->getType(), II->getArgOperand(0), II->getArgOperand(1));

1573

}

1574

}

1575

}

1576

1577

return createAggregateValueExpression(I);

1578

}

1579

const Expression *NewGVN::performSymbolicCmpEvaluation(Instruction *I) const {

1580

auto *CI = dyn_cast<CmpInst>(I);

1581

// See if our operands are equal to those of a previous predicate, and if so,

1582

// if it implies true or false.

1583

auto Op0 = lookupOperandLeader(CI->getOperand(0));

1584

auto Op1 = lookupOperandLeader(CI->getOperand(1));

1585

auto OurPredicate = CI->getPredicate();

1586

if (shouldSwapOperands(Op0, Op1)) {

1587

std::swap(Op0, Op1);

1588

OurPredicate = CI->getSwappedPredicate();

1589

}

1590

1591

// Avoid processing the same info twice

1592

const PredicateBase *LastPredInfo = nullptr;

1593

// See if we know something about the comparison itself, like it is the target

1594

// of an assume.

1595

auto *CmpPI = PredInfo->getPredicateInfoFor(I);

1596

if (dyn_cast_or_null<PredicateAssume>(CmpPI))

1597

return createConstantExpression(ConstantInt::getTrue(CI->getType()));

1598

1599

if (Op0 == Op1) {

1600

// This condition does not depend on predicates, no need to add users

1601

if (CI->isTrueWhenEqual())

1602

return createConstantExpression(ConstantInt::getTrue(CI->getType()));

1603

else if (CI->isFalseWhenEqual())

1604

return createConstantExpression(ConstantInt::getFalse(CI->getType()));

1605

}

1606

1607

// NOTE: Because we are comparing both operands here and below, and using

1608

// previous comparisons, we rely on fact that predicateinfo knows to mark

1609

// comparisons that use renamed operands as users of the earlier comparisons.

1610

// It is *not* enough to just mark predicateinfo renamed operands as users of

1611

// the earlier comparisons, because the *other* operand may have changed in a

1612

// previous iteration.

1613

// Example:

1614

// icmp slt %a, %b

1615

// %b.0 = ssa.copy(%b)

1616

// false branch:

1617

// icmp slt %c, %b.0

1618

1619

// %c and %a may start out equal, and thus, the code below will say the second

1620

// %icmp is false. c may become equal to something else, and in that case the

1621

// %second icmp *must* be reexamined, but would not if only the renamed

1622

// %operands are considered users of the icmp.

1623

1624

// *Currently* we only check one level of comparisons back, and only mark one

1625

// level back as touched when changes appen . If you modify this code to look

1626

// back farther through comparisons, you *must* mark the appropriate

1627

// comparisons as users in PredicateInfo.cpp, or you will cause bugs. See if

1628

// we know something just from the operands themselves

1629

1630

// See if our operands have predicate info, so that we may be able to derive

1631

// something from a previous comparison.

1632

for (const auto &Op : CI->operands()) {

1633

auto *PI = PredInfo->getPredicateInfoFor(Op);

1634

if (const auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI)) {

1635

if (PI == LastPredInfo)

1636

continue;

1637

LastPredInfo = PI;

1638

1639

// TODO: Along the false edge, we may know more things too, like icmp of

1640

// same operands is false.

1641

// TODO: We only handle actual comparison conditions below, not and/or.

1642

auto *BranchCond = dyn_cast<CmpInst>(PBranch->Condition);

1643

if (!BranchCond)

1644

continue;

1645

auto *BranchOp0 = lookupOperandLeader(BranchCond->getOperand(0));

1646

auto *BranchOp1 = lookupOperandLeader(BranchCond->getOperand(1));

1647

auto BranchPredicate = BranchCond->getPredicate();

1648

if (shouldSwapOperands(BranchOp0, BranchOp1)) {

1649

std::swap(BranchOp0, BranchOp1);

1650

BranchPredicate = BranchCond->getSwappedPredicate();

1651

}

1652

if (BranchOp0 == Op0 && BranchOp1 == Op1) {

1653

if (PBranch->TrueEdge) {

1654

// If we know the previous predicate is true and we are in the true

1655

// edge then we may be implied true or false.

1656

if (CmpInst::isImpliedTrueByMatchingCmp(BranchPredicate,

1657

OurPredicate)) {

1658

addPredicateUsers(PI, I);

1659

return createConstantExpression(

1660

ConstantInt::getTrue(CI->getType()));

1661

}

1662

1663

if (CmpInst::isImpliedFalseByMatchingCmp(BranchPredicate,

1664

OurPredicate)) {

1665

addPredicateUsers(PI, I);

1666

return createConstantExpression(

1667

ConstantInt::getFalse(CI->getType()));

1668

}

1669

1670

} else {

1671

// Just handle the ne and eq cases, where if we have the same

1672

// operands, we may know something.

1673

if (BranchPredicate == OurPredicate) {

1674

addPredicateUsers(PI, I);

1675

// Same predicate, same ops,we know it was false, so this is false.

1676

return createConstantExpression(

1677

ConstantInt::getFalse(CI->getType()));

1678

} else if (BranchPredicate ==

1679

CmpInst::getInversePredicate(OurPredicate)) {

1680

addPredicateUsers(PI, I);

1681

// Inverse predicate, we know the other was false, so this is true.

1682

return createConstantExpression(

1683

ConstantInt::getTrue(CI->getType()));

1684

}

1685

}

1686

}

1687

}

1688

}

1689

// Create expression will take care of simplifyCmpInst

1690

return createExpression(I);

1691

}

1692

1693

// Substitute and symbolize the value before value numbering.

1694

const Expression *NewGVN::performSymbolicEvaluation(Value *V) const {

1695

const Expression *E = nullptr;

1696

if (auto *C = dyn_cast<Constant>(V))

1697

E = createConstantExpression(C);

1698

else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {

1699

E = createVariableExpression(V);

1700

} else {

1701

// TODO: memory intrinsics.

1702

// TODO: Some day, we should do the forward propagation and reassociation

1703

// parts of the algorithm.

1704

auto *I = cast<Instruction>(V);

1705

switch (I->getOpcode()) {

1706

case Instruction::ExtractValue:

1707

case Instruction::InsertValue:

1708

E = performSymbolicAggrValueEvaluation(I);

1709

break;

1710

case Instruction::PHI:

1711

E = performSymbolicPHIEvaluation(I);

1712

break;

1713

case Instruction::Call:

1714

E = performSymbolicCallEvaluation(I);

1715

break;

1716

case Instruction::Store:

1717

E = performSymbolicStoreEvaluation(I);

1718

break;

1719

case Instruction::Load:

1720

E = performSymbolicLoadEvaluation(I);

1721

break;

1722

case Instruction::BitCast: {

1723

E = createExpression(I);

1724

} break;

1725

case Instruction::ICmp:

1726

case Instruction::FCmp: {

1727

E = performSymbolicCmpEvaluation(I);

1728

} break;

1729

case Instruction::Add:

1730

case Instruction::FAdd:

1731

case Instruction::Sub:

1732

case Instruction::FSub:

1733

case Instruction::Mul:

1734

case Instruction::FMul:

1735

case Instruction::UDiv:

1736

case Instruction::SDiv:

1737

case Instruction::FDiv:

1738

case Instruction::URem:

1739

case Instruction::SRem:

1740

case Instruction::FRem:

1741

case Instruction::Shl:

1742

case Instruction::LShr:

1743

case Instruction::AShr:

1744

case Instruction::And:

1745

case Instruction::Or:

1746

case Instruction::Xor:

1747

case Instruction::Trunc:

1748

case Instruction::ZExt:

1749

case Instruction::SExt:

1750

case Instruction::FPToUI:

1751

case Instruction::FPToSI:

1752

case Instruction::UIToFP:

1753

case Instruction::SIToFP:

1754

case Instruction::FPTrunc:

1755

case Instruction::FPExt:

1756

case Instruction::PtrToInt:

1757

case Instruction::IntToPtr:

1758

case Instruction::Select:

1759

case Instruction::ExtractElement:

1760

case Instruction::InsertElement:

1761

case Instruction::ShuffleVector:

1762

case Instruction::GetElementPtr:

1763

E = createExpression(I);

1764

break;

1765

default:

1766

return nullptr;

1767

}

1768

}

1769

return E;

1770

}

1771

1772

void NewGVN::markUsersTouched(Value *V) {

1773

// Now mark the users as touched.

1774

for (auto *User : V->users()) {

1775

assert(isa<Instruction>(User) && "Use of value not within an instruction?")((isa<Instruction>(User) && "Use of value not within an instruction?"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(User) && \"Use of value not within an instruction?\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1775, __PRETTY_FUNCTION__));

1776

TouchedInstructions.set(InstrToDFSNum(User));

1777

}

1778

}

1779

1780

void NewGVN::addMemoryUsers(const MemoryAccess *To, MemoryAccess *U) const {

1781

DEBUG(dbgs() << "Adding memory user " << *U << " to " << *To << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Adding memory user " << *
U << " to " << *To << "\n"; } } while (false
);

1782

MemoryToUsers[To].insert(U);

1783

}

1784

1785

void NewGVN::markMemoryDefTouched(const MemoryAccess *MA) {

1786

TouchedInstructions.set(MemoryToDFSNum(MA));

1787

}

1788

1789

void NewGVN::markMemoryUsersTouched(const MemoryAccess *MA) {

1790

if (isa<MemoryUse>(MA))

1791

return;

1792

for (auto U : MA->users())

1793

TouchedInstructions.set(MemoryToDFSNum(U));

1794

const auto Result = MemoryToUsers.find(MA);

1795

if (Result != MemoryToUsers.end()) {

1796

for (auto *User : Result->second)

1797

TouchedInstructions.set(MemoryToDFSNum(User));

1798

MemoryToUsers.erase(Result);

1799

}

1800

}

1801

1802

// Add I to the set of users of a given predicate.

1803

void NewGVN::addPredicateUsers(const PredicateBase *PB, Instruction *I) const {

1804

if (auto *PBranch = dyn_cast<PredicateBranch>(PB))

1805

PredicateToUsers[PBranch->Condition].insert(I);

1806

else if (auto *PAssume = dyn_cast<PredicateBranch>(PB))

1807

PredicateToUsers[PAssume->Condition].insert(I);

1808

}

1809

1810

// Touch all the predicates that depend on this instruction.

1811

void NewGVN::markPredicateUsersTouched(Instruction *I) {

1812

const auto Result = PredicateToUsers.find(I);

1813

if (Result != PredicateToUsers.end()) {

1814

for (auto *User : Result->second)

1815

TouchedInstructions.set(InstrToDFSNum(User));

1816

PredicateToUsers.erase(Result);

1817

}

1818

}

1819

1820

// Mark users affected by a memory leader change.

1821

void NewGVN::markMemoryLeaderChangeTouched(CongruenceClass *CC) {

1822

for (auto M : CC->memory())

1823

markMemoryDefTouched(M);

1824

}

1825

1826

// Touch the instructions that need to be updated after a congruence class has a

1827

// leader change, and mark changed values.

1828

void NewGVN::markValueLeaderChangeTouched(CongruenceClass *CC) {

1829

for (auto M : *CC) {

1830

if (auto *I = dyn_cast<Instruction>(M))

1831

TouchedInstructions.set(InstrToDFSNum(I));

1832

LeaderChanges.insert(M);

1833

}

1834

}

1835

1836

// Give a range of things that have instruction DFS numbers, this will return

1837

// the member of the range with the smallest dfs number.

1838

template <class T, class Range>

1839

T *NewGVN::getMinDFSOfRange(const Range &R) const {

1840

std::pair<T *, unsigned> MinDFS = {nullptr, ~0U};

1841

for (const auto X : R) {

1842

auto DFSNum = InstrToDFSNum(X);

1843

if (DFSNum < MinDFS.second)

1844

MinDFS = {X, DFSNum};

1845

}

1846

return MinDFS.first;

1847

}

1848

1849

// This function returns the MemoryAccess that should be the next leader of

1850

// congruence class CC, under the assumption that the current leader is going to

1851

// disappear.

1852

const MemoryAccess *NewGVN::getNextMemoryLeader(CongruenceClass *CC) const {

1853

// TODO: If this ends up to slow, we can maintain a next memory leader like we

1854

// do for regular leaders.

1855

// Make sure there will be a leader to find

1856

assert(!CC->definesNoMemory() && "Can't get next leader if there is none")((!CC->definesNoMemory() && "Can't get next leader if there is none"
) ? static_cast<void> (0) : __assert_fail ("!CC->definesNoMemory() && \"Can't get next leader if there is none\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1856, __PRETTY_FUNCTION__));

1857

if (CC->getStoreCount() > 0) {

1858

if (auto *NL = dyn_cast_or_null<StoreInst>(CC->getNextLeader().first))

1859

return MSSA->getMemoryAccess(NL);

1860

// Find the store with the minimum DFS number.

1861

auto *V = getMinDFSOfRange<Value>(make_filter_range(

1862

*CC, [&](const Value *V) { return isa<StoreInst>(V); }));

1863

return MSSA->getMemoryAccess(cast<StoreInst>(V));

1864

}

1865

assert(CC->getStoreCount() == 0)((CC->getStoreCount() == 0) ? static_cast<void> (0) :
__assert_fail ("CC->getStoreCount() == 0", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1865, __PRETTY_FUNCTION__));

1866

1867

// Given our assertion, hitting this part must mean

1868

// !OldClass->memory_empty()

1869

if (CC->memory_size() == 1)

1870

return *CC->memory_begin();

1871

return getMinDFSOfRange<const MemoryPhi>(CC->memory());

1872

}

1873

1874

// This function returns the next value leader of a congruence class, under the

1875

// assumption that the current leader is going away. This should end up being

1876

// the next most dominating member.

1877

Value *NewGVN::getNextValueLeader(CongruenceClass *CC) const {

1878

// We don't need to sort members if there is only 1, and we don't care about

1879

// sorting the TOP class because everything either gets out of it or is

1880

// unreachable.

1881

1882

if (CC->size() == 1 || CC == TOPClass) {

1883

return *(CC->begin());

1884

} else if (CC->getNextLeader().first) {

1885

++NumGVNAvoidedSortedLeaderChanges;

1886

return CC->getNextLeader().first;

1887

} else {

1888

++NumGVNSortedLeaderChanges;

1889

// NOTE: If this ends up to slow, we can maintain a dual structure for

1890

// member testing/insertion, or keep things mostly sorted, and sort only

1891

// here, or use SparseBitVector or ....

1892

return getMinDFSOfRange<Value>(*CC);

1893

}

1894

}

1895

1896

// Move a MemoryAccess, currently in OldClass, to NewClass, including updates to

1897

// the memory members, etc for the move.

1898

1899

// The invariants of this function are:

1900

1901

// I must be moving to NewClass from OldClass The StoreCount of OldClass and

1902

// NewClass is expected to have been updated for I already if it is is a store.

1903

// The OldClass memory leader has not been updated yet if I was the leader.

1904

void NewGVN::moveMemoryToNewCongruenceClass(Instruction *I,

1905

MemoryAccess *InstMA,

1906

CongruenceClass *OldClass,

1907

CongruenceClass *NewClass) {

1908

// If the leader is I, and we had a represenative MemoryAccess, it should

1909

// be the MemoryAccess of OldClass.

1910

assert((!InstMA || !OldClass->getMemoryLeader() ||(((!InstMA || !OldClass->getMemoryLeader() || OldClass->
getLeader() != I || OldClass->getMemoryLeader() == InstMA)
&& "Representative MemoryAccess mismatch") ? static_cast
<void> (0) : __assert_fail ("(!InstMA || !OldClass->getMemoryLeader() || OldClass->getLeader() != I || OldClass->getMemoryLeader() == InstMA) && \"Representative MemoryAccess mismatch\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1913, __PRETTY_FUNCTION__))

1911

OldClass->getLeader() != I ||(((!InstMA || !OldClass->getMemoryLeader() || OldClass->
getLeader() != I || OldClass->getMemoryLeader() == InstMA)
&& "Representative MemoryAccess mismatch") ? static_cast
<void> (0) : __assert_fail ("(!InstMA || !OldClass->getMemoryLeader() || OldClass->getLeader() != I || OldClass->getMemoryLeader() == InstMA) && \"Representative MemoryAccess mismatch\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1913, __PRETTY_FUNCTION__))

1912

OldClass->getMemoryLeader() == InstMA) &&(((!InstMA || !OldClass->getMemoryLeader() || OldClass->
getLeader() != I || OldClass->getMemoryLeader() == InstMA)
&& "Representative MemoryAccess mismatch") ? static_cast
<void> (0) : __assert_fail ("(!InstMA || !OldClass->getMemoryLeader() || OldClass->getLeader() != I || OldClass->getMemoryLeader() == InstMA) && \"Representative MemoryAccess mismatch\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1913, __PRETTY_FUNCTION__))

1913

"Representative MemoryAccess mismatch")(((!InstMA || !OldClass->getMemoryLeader() || OldClass->
getLeader() != I || OldClass->getMemoryLeader() == InstMA)
&& "Representative MemoryAccess mismatch") ? static_cast
<void> (0) : __assert_fail ("(!InstMA || !OldClass->getMemoryLeader() || OldClass->getLeader() != I || OldClass->getMemoryLeader() == InstMA) && \"Representative MemoryAccess mismatch\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1913, __PRETTY_FUNCTION__));

1914

// First, see what happens to the new class

1915

if (!NewClass->getMemoryLeader()) {

1916

// Should be a new class, or a store becoming a leader of a new class.

1917

assert(NewClass->size() == 1 ||((NewClass->size() == 1 || (isa<StoreInst>(I) &&
NewClass->getStoreCount() == 1)) ? static_cast<void>
(0) : __assert_fail ("NewClass->size() == 1 || (isa<StoreInst>(I) && NewClass->getStoreCount() == 1)"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1918, __PRETTY_FUNCTION__))

1918

(isa<StoreInst>(I) && NewClass->getStoreCount() == 1))((NewClass->size() == 1 || (isa<StoreInst>(I) &&
NewClass->getStoreCount() == 1)) ? static_cast<void>
(0) : __assert_fail ("NewClass->size() == 1 || (isa<StoreInst>(I) && NewClass->getStoreCount() == 1)"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1918, __PRETTY_FUNCTION__));

1919

NewClass->setMemoryLeader(InstMA);

1920

// Mark it touched if we didn't just create a singleton

1921

DEBUG(dbgs() << "Memory class leader change for class " << NewClass->getID()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory class leader change for class "
<< NewClass->getID() << " due to new memory instruction becoming leader\n"
; } } while (false)

1922

<< " due to new memory instruction becoming leader\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory class leader change for class "
<< NewClass->getID() << " due to new memory instruction becoming leader\n"
; } } while (false);

1923

markMemoryLeaderChangeTouched(NewClass);

1924

}

1925

setMemoryClass(InstMA, NewClass);

1926

// Now, fixup the old class if necessary

1927

if (OldClass->getMemoryLeader() == InstMA) {

1928

if (!OldClass->definesNoMemory()) {

1929

OldClass->setMemoryLeader(getNextMemoryLeader(OldClass));

1930

1931

1932

1933

<< " due to removal of old leader " << *InstMA << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory class leader change for class "
<< OldClass->getID() << " to " << *OldClass
->getMemoryLeader() << " due to removal of old leader "
<< *InstMA << "\n"; } } while (false);

1934

markMemoryLeaderChangeTouched(OldClass);

1935

} else

1936

OldClass->setMemoryLeader(nullptr);

1937

}

1938

}

1939

1940

// Move a value, currently in OldClass, to be part of NewClass

1941

// Update OldClass and NewClass for the move (including changing leaders, etc).

1942

void NewGVN::moveValueToNewCongruenceClass(Instruction *I, const Expression *E,

1943

CongruenceClass *OldClass,

1944

CongruenceClass *NewClass) {

1945

if (I == OldClass->getNextLeader().first)

1946

OldClass->resetNextLeader();

1947

1948

// It's possible, though unlikely, for us to discover equivalences such

1949

// that the current leader does not dominate the old one.

1950

// This statistic tracks how often this happens.

1951

// We assert on phi nodes when this happens, currently, for debugging, because

1952

// we want to make sure we name phi node cycles properly.

1953

if (isa<Instruction>(NewClass->getLeader()) && NewClass->getLeader() &&

1954

I != NewClass->getLeader()) {

1955

auto *IBB = I->getParent();

1956

auto *NCBB = cast<Instruction>(NewClass->getLeader())->getParent();

1957

bool Dominated =

1958

IBB == NCBB && InstrToDFSNum(I) < InstrToDFSNum(NewClass->getLeader());

1959

Dominated = Dominated || DT->properlyDominates(IBB, NCBB);

1960

if (Dominated) {

1961

++NumGVNNotMostDominatingLeader;

1962

assert(((!isa<PHINode>(I) && "New class for instruction should not be dominated by instruction"
) ? static_cast<void> (0) : __assert_fail ("!isa<PHINode>(I) && \"New class for instruction should not be dominated by instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1964, __PRETTY_FUNCTION__))

1963

!isa<PHINode>(I) &&((!isa<PHINode>(I) && "New class for instruction should not be dominated by instruction"
) ? static_cast<void> (0) : __assert_fail ("!isa<PHINode>(I) && \"New class for instruction should not be dominated by instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1964, __PRETTY_FUNCTION__))

1964

"New class for instruction should not be dominated by instruction")((!isa<PHINode>(I) && "New class for instruction should not be dominated by instruction"
) ? static_cast<void> (0) : __assert_fail ("!isa<PHINode>(I) && \"New class for instruction should not be dominated by instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1964, __PRETTY_FUNCTION__));

1965

}

1966

}

1967

1968

if (NewClass->getLeader() != I)

1969

NewClass->addPossibleNextLeader({I, InstrToDFSNum(I)});

1970

1971

OldClass->erase(I);

1972

NewClass->insert(I);

1973

// Handle our special casing of stores.

1974

if (auto *SI = dyn_cast<StoreInst>(I)) {

1975

OldClass->decStoreCount();

1976

// Okay, so when do we want to make a store a leader of a class?

1977

// If we have a store defined by an earlier load, we want the earlier load

1978

// to lead the class.

1979

// If we have a store defined by something else, we want the store to lead

1980

// the class so everything else gets the "something else" as a value.

1981

// If we have a store as the single member of the class, we want the store

1982

// as the leader

1983

if (NewClass->getStoreCount() == 0 && !NewClass->getStoredValue()) {

1984

// If it's a store expression we are using, it means we are not equivalent

1985

// to something earlier.

1986

if (auto *SE = dyn_cast<StoreExpression>(E)) {

1987

assert(SE->getStoredValue() != NewClass->getLeader())((SE->getStoredValue() != NewClass->getLeader()) ? static_cast
<void> (0) : __assert_fail ("SE->getStoredValue() != NewClass->getLeader()"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 1987, __PRETTY_FUNCTION__));

1988

NewClass->setStoredValue(SE->getStoredValue());

1989

markValueLeaderChangeTouched(NewClass);

1990

// Shift the new class leader to be the store

1991

DEBUG(dbgs() << "Changing leader of congruence class "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Changing leader of congruence class "
<< NewClass->getID() << " from " << *NewClass
->getLeader() << " to " << *SI << " because store joined class\n"
; } } while (false)

1992

<< NewClass->getID() << " from " << *NewClass->getLeader()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Changing leader of congruence class "
<< NewClass->getID() << " from " << *NewClass
->getLeader() << " to " << *SI << " because store joined class\n"
; } } while (false)

1993

<< " to " << *SI << " because store joined class\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Changing leader of congruence class "
<< NewClass->getID() << " from " << *NewClass
->getLeader() << " to " << *SI << " because store joined class\n"
; } } while (false);

1994

// If we changed the leader, we have to mark it changed because we don't

1995

// know what it will do to symbolic evlauation.

1996

NewClass->setLeader(SI);

1997

}

1998

// We rely on the code below handling the MemoryAccess change.

1999

}

2000

NewClass->incStoreCount();

2001

}

2002

// True if there is no memory instructions left in a class that had memory

2003

// instructions before.

2004

2005

// If it's not a memory use, set the MemoryAccess equivalence

2006

auto *InstMA = dyn_cast_or_null<MemoryDef>(MSSA->getMemoryAccess(I));

2007

if (InstMA)

2008

moveMemoryToNewCongruenceClass(I, InstMA, OldClass, NewClass);

2009

ValueToClass[I] = NewClass;

2010

// See if we destroyed the class or need to swap leaders.

2011

if (OldClass->empty() && OldClass != TOPClass) {

2012

if (OldClass->getDefiningExpr()) {

2013

DEBUG(dbgs() << "Erasing expression " << *OldClass->getDefiningExpr()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Erasing expression " << *
OldClass->getDefiningExpr() << " from table\n"; } } while
(false)

2014

<< " from table\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Erasing expression " << *
OldClass->getDefiningExpr() << " from table\n"; } } while
(false);

2015

ExpressionToClass.erase(OldClass->getDefiningExpr());

2016

}

2017

} else if (OldClass->getLeader() == I) {

2018

// When the leader changes, the value numbering of

2019

// everything may change due to symbolization changes, so we need to

2020

// reprocess.

2021

DEBUG(dbgs() << "Value class leader change for class " << OldClass->getID()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Value class leader change for class "
<< OldClass->getID() << "\n"; } } while (false
)

2022

<< "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Value class leader change for class "
<< OldClass->getID() << "\n"; } } while (false
);

2023

++NumGVNLeaderChanges;

2024

// Destroy the stored value if there are no more stores to represent it.

2025

// Note that this is basically clean up for the expression removal that

2026

// happens below. If we remove stores from a class, we may leave it as a

2027

// class of equivalent memory phis.

2028

if (OldClass->getStoreCount() == 0) {

2029

if (OldClass->getStoredValue())

2030

OldClass->setStoredValue(nullptr);

2031

}

2032

OldClass->setLeader(getNextValueLeader(OldClass));

2033

OldClass->resetNextLeader();

2034

markValueLeaderChangeTouched(OldClass);

2035

}

2036

}

2037

2038

// Perform congruence finding on a given value numbering expression.

2039

void NewGVN::performCongruenceFinding(Instruction *I, const Expression *E) {

2040

// This is guaranteed to return something, since it will at least find

2041

// TOP.

2042

2043

CongruenceClass *IClass = ValueToClass[I];

2044

assert(IClass && "Should have found a IClass")((IClass && "Should have found a IClass") ? static_cast
<void> (0) : __assert_fail ("IClass && \"Should have found a IClass\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2044, __PRETTY_FUNCTION__));

2045

// Dead classes should have been eliminated from the mapping.

2046

assert(!IClass->isDead() && "Found a dead class")((!IClass->isDead() && "Found a dead class") ? static_cast
<void> (0) : __assert_fail ("!IClass->isDead() && \"Found a dead class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2046, __PRETTY_FUNCTION__));

2047

2048

CongruenceClass *EClass;

2049

if (const auto *VE = dyn_cast<VariableExpression>(E)) {

2050

EClass = ValueToClass[VE->getVariableValue()];

2051

} else {

2052

auto lookupResult = ExpressionToClass.insert({E, nullptr});

2053

2054

// If it's not in the value table, create a new congruence class.

2055

if (lookupResult.second) {

2056

CongruenceClass *NewClass = createCongruenceClass(nullptr, E);

2057

auto place = lookupResult.first;

2058

place->second = NewClass;

2059

2060

// Constants and variables should always be made the leader.

2061

if (const auto *CE = dyn_cast<ConstantExpression>(E)) {

2062

NewClass->setLeader(CE->getConstantValue());

2063

} else if (const auto *SE = dyn_cast<StoreExpression>(E)) {

2064

StoreInst *SI = SE->getStoreInst();

2065

NewClass->setLeader(SI);

2066

NewClass->setStoredValue(SE->getStoredValue());

2067

// The RepMemoryAccess field will be filled in properly by the

2068

// moveValueToNewCongruenceClass call.

2069

} else {

2070

NewClass->setLeader(I);

2071

}

2072

assert(!isa<VariableExpression>(E) &&((!isa<VariableExpression>(E) && "VariableExpression should have been handled already"
) ? static_cast<void> (0) : __assert_fail ("!isa<VariableExpression>(E) && \"VariableExpression should have been handled already\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2073, __PRETTY_FUNCTION__))

2073

"VariableExpression should have been handled already")((!isa<VariableExpression>(E) && "VariableExpression should have been handled already"
) ? static_cast<void> (0) : __assert_fail ("!isa<VariableExpression>(E) && \"VariableExpression should have been handled already\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2073, __PRETTY_FUNCTION__));

2074

2075

EClass = NewClass;

2076

DEBUG(dbgs() << "Created new congruence class for " << *Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Created new congruence class for "
<< *I << " using expression " << *E <<
" at " << NewClass->getID() << " and leader "
<< *(NewClass->getLeader()); } } while (false)

2077

<< " using expression " << *E << " at " << NewClass->getID()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Created new congruence class for "
<< *I << " using expression " << *E <<
" at " << NewClass->getID() << " and leader "
<< *(NewClass->getLeader()); } } while (false)

2078

<< " and leader " << *(NewClass->getLeader()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Created new congruence class for "
<< *I << " using expression " << *E <<
" at " << NewClass->getID() << " and leader "
<< *(NewClass->getLeader()); } } while (false);

2079

if (NewClass->getStoredValue())

2080

DEBUG(dbgs() << " and stored value " << *(NewClass->getStoredValue()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << " and stored value " << *
(NewClass->getStoredValue()); } } while (false);

2081

DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "\n"; } } while (false);

2082

} else {

2083

EClass = lookupResult.first->second;

2084

if (isa<ConstantExpression>(E))

2085

assert((isa<Constant>(EClass->getLeader()) ||(((isa<Constant>(EClass->getLeader()) || (EClass->
getStoredValue() && isa<Constant>(EClass->getStoredValue
()))) && "Any class with a constant expression should have a "
"constant leader") ? static_cast<void> (0) : __assert_fail
("(isa<Constant>(EClass->getLeader()) || (EClass->getStoredValue() && isa<Constant>(EClass->getStoredValue()))) && \"Any class with a constant expression should have a \" \"constant leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2089, __PRETTY_FUNCTION__))

2086

(EClass->getStoredValue() &&(((isa<Constant>(EClass->getLeader()) || (EClass->
getStoredValue() && isa<Constant>(EClass->getStoredValue
()))) && "Any class with a constant expression should have a "
"constant leader") ? static_cast<void> (0) : __assert_fail
("(isa<Constant>(EClass->getLeader()) || (EClass->getStoredValue() && isa<Constant>(EClass->getStoredValue()))) && \"Any class with a constant expression should have a \" \"constant leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2089, __PRETTY_FUNCTION__))

2087

isa<Constant>(EClass->getStoredValue()))) &&(((isa<Constant>(EClass->getLeader()) || (EClass->
getStoredValue() && isa<Constant>(EClass->getStoredValue
()))) && "Any class with a constant expression should have a "
"constant leader") ? static_cast<void> (0) : __assert_fail
("(isa<Constant>(EClass->getLeader()) || (EClass->getStoredValue() && isa<Constant>(EClass->getStoredValue()))) && \"Any class with a constant expression should have a \" \"constant leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2089, __PRETTY_FUNCTION__))

2088

"Any class with a constant expression should have a "(((isa<Constant>(EClass->getLeader()) || (EClass->
getStoredValue() && isa<Constant>(EClass->getStoredValue
()))) && "Any class with a constant expression should have a "
"constant leader") ? static_cast<void> (0) : __assert_fail
("(isa<Constant>(EClass->getLeader()) || (EClass->getStoredValue() && isa<Constant>(EClass->getStoredValue()))) && \"Any class with a constant expression should have a \" \"constant leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2089, __PRETTY_FUNCTION__))

2089

"constant leader")(((isa<Constant>(EClass->getLeader()) || (EClass->
getStoredValue() && isa<Constant>(EClass->getStoredValue
()))) && "Any class with a constant expression should have a "
"constant leader") ? static_cast<void> (0) : __assert_fail
("(isa<Constant>(EClass->getLeader()) || (EClass->getStoredValue() && isa<Constant>(EClass->getStoredValue()))) && \"Any class with a constant expression should have a \" \"constant leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2089, __PRETTY_FUNCTION__));

2090

2091

assert(EClass && "Somehow don't have an eclass")((EClass && "Somehow don't have an eclass") ? static_cast
<void> (0) : __assert_fail ("EClass && \"Somehow don't have an eclass\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2091, __PRETTY_FUNCTION__));

2092

2093

assert(!EClass->isDead() && "We accidentally looked up a dead class")((!EClass->isDead() && "We accidentally looked up a dead class"
) ? static_cast<void> (0) : __assert_fail ("!EClass->isDead() && \"We accidentally looked up a dead class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2093, __PRETTY_FUNCTION__));

2094

}

2095

}

2096

bool ClassChanged = IClass != EClass;

2097

bool LeaderChanged = LeaderChanges.erase(I);

2098

if (ClassChanged || LeaderChanged) {

2099

DEBUG(dbgs() << "New class " << EClass->getID() << " for expression " << *Edo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "New class " << EClass->
getID() << " for expression " << *E << "\n"
; } } while (false)

2100

<< "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "New class " << EClass->
getID() << " for expression " << *E << "\n"
; } } while (false);

2101

if (ClassChanged)

2102

moveValueToNewCongruenceClass(I, E, IClass, EClass);

2103

markUsersTouched(I);

2104

if (MemoryAccess *MA = MSSA->getMemoryAccess(I))

2105

markMemoryUsersTouched(MA);

2106

if (auto *CI = dyn_cast<CmpInst>(I))

2107

markPredicateUsersTouched(CI);

2108

}

2109

// If we changed the class of the store, we want to ensure nothing finds the

2110

// old store expression. In particular, loads do not compare against stored

2111

// value, so they will find old store expressions (and associated class

2112

// mappings) if we leave them in the table.

2113

if (ClassChanged && isa<StoreExpression>(E)) {

2114

auto *OldE = ValueToExpression.lookup(I);

2115

// It could just be that the old class died. We don't want to erase it if we

2116

// just moved classes.

2117

if (OldE && isa<StoreExpression>(OldE) && !OldE->equals(*E))

2118

ExpressionToClass.erase(OldE);

2119

}

2120

ValueToExpression[I] = E;

2121

}

2122

2123

// Process the fact that Edge (from, to) is reachable, including marking

2124

// any newly reachable blocks and instructions for processing.

2125

void NewGVN::updateReachableEdge(BasicBlock *From, BasicBlock *To) {

2126

// Check if the Edge was reachable before.

2127

if (ReachableEdges.insert({From, To}).second) {

2128

// If this block wasn't reachable before, all instructions are touched.

2129

if (ReachableBlocks.insert(To).second) {

2130

DEBUG(dbgs() << "Block " << getBlockName(To) << " marked reachable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Block " << getBlockName(
To) << " marked reachable\n"; } } while (false);

2131

const auto &InstRange = BlockInstRange.lookup(To);

2132

TouchedInstructions.set(InstRange.first, InstRange.second);

2133

} else {

2134

DEBUG(dbgs() << "Block " << getBlockName(To)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Block " << getBlockName(
To) << " was reachable, but new edge {" << getBlockName
(From) << "," << getBlockName(To) << "} to it found\n"
; } } while (false)

2135

<< " was reachable, but new edge {" << getBlockName(From)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Block " << getBlockName(
To) << " was reachable, but new edge {" << getBlockName
(From) << "," << getBlockName(To) << "} to it found\n"
; } } while (false)

2136

<< "," << getBlockName(To) << "} to it found\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Block " << getBlockName(
To) << " was reachable, but new edge {" << getBlockName
(From) << "," << getBlockName(To) << "} to it found\n"
; } } while (false);

2137

2138

// We've made an edge reachable to an existing block, which may

2139

// impact predicates. Otherwise, only mark the phi nodes as touched, as

2140

// they are the only thing that depend on new edges. Anything using their

2141

// values will get propagated to if necessary.

2142

if (MemoryAccess *MemPhi = MSSA->getMemoryAccess(To))

2143

TouchedInstructions.set(InstrToDFSNum(MemPhi));

2144

2145

auto BI = To->begin();

2146

while (isa<PHINode>(BI)) {

2147

TouchedInstructions.set(InstrToDFSNum(&*BI));

2148

++BI;

2149

}

2150

}

2151

}

2152

}

2153

2154

// Given a predicate condition (from a switch, cmp, or whatever) and a block,

2155

// see if we know some constant value for it already.

2156

Value *NewGVN::findConditionEquivalence(Value *Cond) const {

2157

auto Result = lookupOperandLeader(Cond);

2158

if (isa<Constant>(Result))

2159

return Result;

2160

return nullptr;

2161

}

2162

2163

// Process the outgoing edges of a block for reachability.

2164

void NewGVN::processOutgoingEdges(TerminatorInst *TI, BasicBlock *B) {

2165

// Evaluate reachability of terminator instruction.

2166

BranchInst *BR;

2167

if ((BR = dyn_cast<BranchInst>(TI)) && BR->isConditional()) {

2168

Value *Cond = BR->getCondition();

2169

Value *CondEvaluated = findConditionEquivalence(Cond);

2170

if (!CondEvaluated) {

2171

if (auto *I = dyn_cast<Instruction>(Cond)) {

2172

const Expression *E = createExpression(I);

2173

if (const auto *CE = dyn_cast<ConstantExpression>(E)) {

2174

CondEvaluated = CE->getConstantValue();

2175

}

2176

} else if (isa<ConstantInt>(Cond)) {

2177

CondEvaluated = Cond;

2178

}

2179

}

2180

ConstantInt *CI;

2181

BasicBlock *TrueSucc = BR->getSuccessor(0);

2182

BasicBlock *FalseSucc = BR->getSuccessor(1);

2183

if (CondEvaluated && (CI = dyn_cast<ConstantInt>(CondEvaluated))) {

2184

if (CI->isOne()) {

2185

DEBUG(dbgs() << "Condition for Terminator " << *TIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Condition for Terminator " <<
*TI << " evaluated to true\n"; } } while (false)

2186

<< " evaluated to true\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Condition for Terminator " <<
*TI << " evaluated to true\n"; } } while (false);

2187

updateReachableEdge(B, TrueSucc);

2188

} else if (CI->isZero()) {

2189

2190

<< " evaluated to false\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Condition for Terminator " <<
*TI << " evaluated to false\n"; } } while (false);

2191

updateReachableEdge(B, FalseSucc);

2192

}

2193

} else {

2194

updateReachableEdge(B, TrueSucc);

2195

updateReachableEdge(B, FalseSucc);

2196

}

2197

} else if (auto *SI = dyn_cast<SwitchInst>(TI)) {

2198

// For switches, propagate the case values into the case

2199

// destinations.

2200

2201

// Remember how many outgoing edges there are to every successor.

2202

SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;

2203

2204

Value *SwitchCond = SI->getCondition();

2205

Value *CondEvaluated = findConditionEquivalence(SwitchCond);

2206

// See if we were able to turn this switch statement into a constant.

2207

if (CondEvaluated && isa<ConstantInt>(CondEvaluated)) {

2208

auto *CondVal = cast<ConstantInt>(CondEvaluated);

2209

// We should be able to get case value for this.

2210

auto Case = *SI->findCaseValue(CondVal);

2211

if (Case.getCaseSuccessor() == SI->getDefaultDest()) {

2212

// We proved the value is outside of the range of the case.

2213

// We can't do anything other than mark the default dest as reachable,

2214

// and go home.

2215

updateReachableEdge(B, SI->getDefaultDest());

2216

return;

2217

}

2218

// Now get where it goes and mark it reachable.

2219

BasicBlock *TargetBlock = Case.getCaseSuccessor();

2220

updateReachableEdge(B, TargetBlock);

2221

} else {

2222

for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) {

2223

BasicBlock *TargetBlock = SI->getSuccessor(i);

2224

++SwitchEdges[TargetBlock];

2225

updateReachableEdge(B, TargetBlock);

2226

}

2227

}

2228

} else {

2229

// Otherwise this is either unconditional, or a type we have no

2230

// idea about. Just mark successors as reachable.

2231

for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {

2232

BasicBlock *TargetBlock = TI->getSuccessor(i);

2233

updateReachableEdge(B, TargetBlock);

2234

}

2235

2236

// This also may be a memory defining terminator, in which case, set it

2237

// equivalent only to itself.

2238

2239

auto *MA = MSSA->getMemoryAccess(TI);

2240

if (MA && !isa<MemoryUse>(MA)) {

2241

auto *CC = ensureLeaderOfMemoryClass(MA);

2242

if (setMemoryClass(MA, CC))

2243

markMemoryUsersTouched(MA);

2244

}

2245

}

2246

}

2247

2248

// The algorithm initially places the values of the routine in the TOP

2249

// congruence class. The leader of TOP is the undetermined value `undef`.

2250

// When the algorithm has finished, values still in TOP are unreachable.

2251

void NewGVN::initializeCongruenceClasses(Function &F) {

2252

NextCongruenceNum = 0;

2253

2254

// Note that even though we use the live on entry def as a representative

2255

// MemoryAccess, it is *not* the same as the actual live on entry def. We

2256

// have no real equivalemnt to undef for MemoryAccesses, and so we really

2257

// should be checking whether the MemoryAccess is top if we want to know if it

2258

// is equivalent to everything. Otherwise, what this really signifies is that

2259

// the access "it reaches all the way back to the beginning of the function"

2260

2261

// Initialize all other instructions to be in TOP class.

2262

TOPClass = createCongruenceClass(nullptr, nullptr);

2263

TOPClass->setMemoryLeader(MSSA->getLiveOnEntryDef());

2264

// The live on entry def gets put into it's own class

2265

MemoryAccessToClass[MSSA->getLiveOnEntryDef()] =

2266

createMemoryClass(MSSA->getLiveOnEntryDef());

2267

2268

for (auto DTN : nodes(DT)) {

2269

BasicBlock *BB = DTN->getBlock();

2270

// All MemoryAccesses are equivalent to live on entry to start. They must

2271

// be initialized to something so that initial changes are noticed. For

2272

// the maximal answer, we initialize them all to be the same as

2273

// liveOnEntry.

2274

auto *MemoryBlockDefs = MSSA->getBlockDefs(BB);

2275

if (MemoryBlockDefs)

2276

for (const auto &Def : *MemoryBlockDefs) {

2277

MemoryAccessToClass[&Def] = TOPClass;

2278

auto *MD = dyn_cast<MemoryDef>(&Def);

2279

// Insert the memory phis into the member list.

2280

if (!MD) {

2281

const MemoryPhi *MP = cast<MemoryPhi>(&Def);

2282

TOPClass->memory_insert(MP);

2283

MemoryPhiState.insert({MP, MPS_TOP});

2284

}

2285

2286

if (MD && isa<StoreInst>(MD->getMemoryInst()))

2287

TOPClass->incStoreCount();

2288

}

2289

for (auto &I : *BB) {

2290

// Don't insert void terminators into the class. We don't value number

2291

// them, and they just end up sitting in TOP.

2292

if (isa<TerminatorInst>(I) && I.getType()->isVoidTy())

2293

continue;

2294

TOPClass->insert(&I);

2295

ValueToClass[&I] = TOPClass;

2296

}

2297

}

2298

2299

// Initialize arguments to be in their own unique congruence classes

2300

for (auto &FA : F.args())

2301

createSingletonCongruenceClass(&FA);

2302

}

2303

2304

void NewGVN::cleanupTables() {

2305

for (unsigned i = 0, e = CongruenceClasses.size(); i != e; ++i) {

2306

DEBUG(dbgs() << "Congruence class " << CongruenceClasses[i]->getID()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Congruence class " << CongruenceClasses
[i]->getID() << " has " << CongruenceClasses[i
]->size() << " members\n"; } } while (false)

2307

<< " has " << CongruenceClasses[i]->size() << " members\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Congruence class " << CongruenceClasses
[i]->getID() << " has " << CongruenceClasses[i
]->size() << " members\n"; } } while (false);

2308

// Make sure we delete the congruence class (probably worth switching to

2309

// a unique_ptr at some point.

2310

delete CongruenceClasses[i];

2311

CongruenceClasses[i] = nullptr;

2312

}

2313

2314

ValueToClass.clear();

2315

ArgRecycler.clear(ExpressionAllocator);

2316

ExpressionAllocator.Reset();

2317

CongruenceClasses.clear();

2318

ExpressionToClass.clear();

2319

ValueToExpression.clear();

2320

ReachableBlocks.clear();

2321

ReachableEdges.clear();

2322

#ifndef NDEBUG

2323

ProcessedCount.clear();

2324

#endif

2325

InstrDFS.clear();

2326

InstructionsToErase.clear();

2327

DFSToInstr.clear();

2328

BlockInstRange.clear();

2329

TouchedInstructions.clear();

2330

MemoryAccessToClass.clear();

2331

PredicateToUsers.clear();

2332

MemoryToUsers.clear();

2333

}

2334

2335

std::pair<unsigned, unsigned> NewGVN::assignDFSNumbers(BasicBlock *B,

2336

unsigned Start) {

2337

unsigned End = Start;

2338

if (MemoryAccess *MemPhi = MSSA->getMemoryAccess(B)) {

2339

InstrDFS[MemPhi] = End++;

2340

DFSToInstr.emplace_back(MemPhi);

2341

}

2342

2343

for (auto &I : *B) {

2344

// There's no need to call isInstructionTriviallyDead more than once on

2345

// an instruction. Therefore, once we know that an instruction is dead

2346

// we change its DFS number so that it doesn't get value numbered.

2347

if (isInstructionTriviallyDead(&I, TLI)) {

2348

InstrDFS[&I] = 0;

2349

DEBUG(dbgs() << "Skipping trivially dead instruction " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Skipping trivially dead instruction "
<< I << "\n"; } } while (false);

2350

markInstructionForDeletion(&I);

2351

continue;

2352

}

2353

2354

InstrDFS[&I] = End++;

2355

DFSToInstr.emplace_back(&I);

2356

}

2357

2358

// All of the range functions taken half-open ranges (open on the end side).

2359

// So we do not subtract one from count, because at this point it is one

2360

// greater than the last instruction.

2361

return std::make_pair(Start, End);

2362

}

2363

2364

void NewGVN::updateProcessedCount(Value *V) {

2365

#ifndef NDEBUG

2366

if (ProcessedCount.count(V) == 0) {

2367

ProcessedCount.insert({V, 1});

2368

} else {

2369

++ProcessedCount[V];

2370

assert(ProcessedCount[V] < 100 &&((ProcessedCount[V] < 100 && "Seem to have processed the same Value a lot"
) ? static_cast<void> (0) : __assert_fail ("ProcessedCount[V] < 100 && \"Seem to have processed the same Value a lot\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2371, __PRETTY_FUNCTION__))

2371

"Seem to have processed the same Value a lot")((ProcessedCount[V] < 100 && "Seem to have processed the same Value a lot"
) ? static_cast<void> (0) : __assert_fail ("ProcessedCount[V] < 100 && \"Seem to have processed the same Value a lot\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2371, __PRETTY_FUNCTION__));

2372

}

2373

#endif

2374

}

2375

// Evaluate MemoryPhi nodes symbolically, just like PHI nodes

2376

void NewGVN::valueNumberMemoryPhi(MemoryPhi *MP) {

2377

// If all the arguments are the same, the MemoryPhi has the same value as the

2378

// argument.

2379

// Filter out unreachable blocks and self phis from our operands.

2380

const BasicBlock *PHIBlock = MP->getBlock();

2381

auto Filtered = make_filter_range(MP->operands(), [&](const Use &U) {

2382

return lookupMemoryLeader(cast<MemoryAccess>(U)) != MP &&

2383

!isMemoryAccessTop(cast<MemoryAccess>(U)) &&

2384

ReachableEdges.count({MP->getIncomingBlock(U), PHIBlock});

2385

});

2386

// If all that is left is nothing, our memoryphi is undef. We keep it as

2387

// InitialClass. Note: The only case this should happen is if we have at

2388

// least one self-argument.

2389

if (Filtered.begin() == Filtered.end()) {

2390

if (setMemoryClass(MP, TOPClass))

2391

markMemoryUsersTouched(MP);

2392

return;

2393

}

2394

2395

// Transform the remaining operands into operand leaders.

2396

// FIXME: mapped_iterator should have a range version.

2397

auto LookupFunc = [&](const Use &U) {

2398

return lookupMemoryLeader(cast<MemoryAccess>(U));

2399

};

2400

auto MappedBegin = map_iterator(Filtered.begin(), LookupFunc);

2401

auto MappedEnd = map_iterator(Filtered.end(), LookupFunc);

2402

2403

// and now check if all the elements are equal.

2404

// Sadly, we can't use std::equals since these are random access iterators.

2405

const auto *AllSameValue = *MappedBegin;

2406

++MappedBegin;

2407

bool AllEqual = std::all_of(

2408

MappedBegin, MappedEnd,

2409

[&AllSameValue](const MemoryAccess *V) { return V == AllSameValue; });

2410

2411

if (AllEqual)

2412

DEBUG(dbgs() << "Memory Phi value numbered to " << *AllSameValue << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory Phi value numbered to "
<< *AllSameValue << "\n"; } } while (false);

2413

else

2414

DEBUG(dbgs() << "Memory Phi value numbered to itself\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Memory Phi value numbered to itself\n"
; } } while (false);

2415

// If it's equal to something, it's in that class. Otherwise, it has to be in

2416

// a class where it is the leader (other things may be equivalent to it, but

2417

// it needs to start off in its own class, which means it must have been the

2418

// leader, and it can't have stopped being the leader because it was never

2419

// removed).

2420

CongruenceClass *CC =

2421

AllEqual ? getMemoryClass(AllSameValue) : ensureLeaderOfMemoryClass(MP);

2422

auto OldState = MemoryPhiState.lookup(MP);

2423

assert(OldState != MPS_Invalid && "Invalid memory phi state")((OldState != MPS_Invalid && "Invalid memory phi state"
) ? static_cast<void> (0) : __assert_fail ("OldState != MPS_Invalid && \"Invalid memory phi state\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2423, __PRETTY_FUNCTION__));

2424

auto NewState = AllEqual ? MPS_Equivalent : MPS_Unique;

2425

MemoryPhiState[MP] = NewState;

2426

if (setMemoryClass(MP, CC) || OldState != NewState)

2427

markMemoryUsersTouched(MP);

2428

}

2429

2430

// Value number a single instruction, symbolically evaluating, performing

2431

// congruence finding, and updating mappings.

2432

void NewGVN::valueNumberInstruction(Instruction *I) {

2433

DEBUG(dbgs() << "Processing instruction " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Processing instruction " <<
*I << "\n"; } } while (false);

2434

if (!I->isTerminator()) {

2435

const Expression *Symbolized = nullptr;

2436

if (DebugCounter::shouldExecute(VNCounter)) {

2437

Symbolized = performSymbolicEvaluation(I);

2438

} else {

2439

// Mark the instruction as unused so we don't value number it again.

2440

InstrDFS[I] = 0;

2441

}

2442

// If we couldn't come up with a symbolic expression, use the unknown

2443

// expression

2444

if (Symbolized == nullptr) {

2445

Symbolized = createUnknownExpression(I);

2446

}

2447

2448

performCongruenceFinding(I, Symbolized);

2449

} else {

2450

// Handle terminators that return values. All of them produce values we

2451

// don't currently understand. We don't place non-value producing

2452

// terminators in a class.

2453

if (!I->getType()->isVoidTy()) {

2454

auto *Symbolized = createUnknownExpression(I);

2455

performCongruenceFinding(I, Symbolized);

2456

}

2457

processOutgoingEdges(dyn_cast<TerminatorInst>(I), I->getParent());

2458

}

2459

}

2460

2461

// Check if there is a path, using single or equal argument phi nodes, from

2462

// First to Second.

2463

bool NewGVN::singleReachablePHIPath(const MemoryAccess *First,

2464

const MemoryAccess *Second) const {

2465

if (First == Second)

2466

return true;

2467

if (MSSA->isLiveOnEntryDef(First))

2468

return false;

2469

2470

const auto *EndDef = First;

2471

for (auto *ChainDef : optimized_def_chain(First)) {

2472

if (ChainDef == Second)

2473

return true;

2474

if (MSSA->isLiveOnEntryDef(ChainDef))

2475

return false;

2476

EndDef = ChainDef;

2477

}

2478

auto *MP = cast<MemoryPhi>(EndDef);

2479

auto ReachableOperandPred = [&](const Use &U) {

2480

return ReachableEdges.count({MP->getIncomingBlock(U), MP->getBlock()});

2481

};

2482

auto FilteredPhiArgs =

2483

make_filter_range(MP->operands(), ReachableOperandPred);

2484

SmallVector<const Value *, 32> OperandList;

2485

std::copy(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),

2486

std::back_inserter(OperandList));

2487

bool Okay = OperandList.size() == 1;

2488

if (!Okay)

2489

Okay =

2490

std::equal(OperandList.begin(), OperandList.end(), OperandList.begin());

2491

if (Okay)

2492

return singleReachablePHIPath(cast<MemoryAccess>(OperandList[0]), Second);

2493

return false;

2494

}

2495

2496

// Verify the that the memory equivalence table makes sense relative to the

2497

// congruence classes. Note that this checking is not perfect, and is currently

2498

// subject to very rare false negatives. It is only useful for

2499

// testing/debugging.

2500

void NewGVN::verifyMemoryCongruency() const {

2501

#ifndef NDEBUG

2502

// Verify that the memory table equivalence and memory member set match

2503

for (const auto *CC : CongruenceClasses) {

2504

if (CC == TOPClass || CC->isDead())

2505

continue;

2506

if (CC->getStoreCount() != 0) {

2507

assert((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) &&(((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader
())) && "Any class with a store as a leader should have a "
"representative stored value") ? static_cast<void> (0)
: __assert_fail ("(CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) && \"Any class with a store as a leader should have a \" \"representative stored value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2509, __PRETTY_FUNCTION__))

2508

"Any class with a store as a leader should have a "(((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader
())) && "Any class with a store as a leader should have a "
"representative stored value") ? static_cast<void> (0)
: __assert_fail ("(CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) && \"Any class with a store as a leader should have a \" \"representative stored value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2509, __PRETTY_FUNCTION__))

2509

"representative stored value")(((CC->getStoredValue() || !isa<StoreInst>(CC->getLeader
())) && "Any class with a store as a leader should have a "
"representative stored value") ? static_cast<void> (0)
: __assert_fail ("(CC->getStoredValue() || !isa<StoreInst>(CC->getLeader())) && \"Any class with a store as a leader should have a \" \"representative stored value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2509, __PRETTY_FUNCTION__));

2510

assert(CC->getMemoryLeader() &&((CC->getMemoryLeader() && "Any congruence class with a store should have a "
"representative access") ? static_cast<void> (0) : __assert_fail
("CC->getMemoryLeader() && \"Any congruence class with a store should have a \" \"representative access\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2512, __PRETTY_FUNCTION__))

2511

"Any congruence class with a store should have a "((CC->getMemoryLeader() && "Any congruence class with a store should have a "
"representative access") ? static_cast<void> (0) : __assert_fail
("CC->getMemoryLeader() && \"Any congruence class with a store should have a \" \"representative access\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2512, __PRETTY_FUNCTION__))

2512

"representative access")((CC->getMemoryLeader() && "Any congruence class with a store should have a "
"representative access") ? static_cast<void> (0) : __assert_fail
("CC->getMemoryLeader() && \"Any congruence class with a store should have a \" \"representative access\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2512, __PRETTY_FUNCTION__));

2513

}

2514

2515

if (CC->getMemoryLeader())

2516

assert(MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&((MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
"Representative MemoryAccess does not appear to be reverse "
"mapped properly") ? static_cast<void> (0) : __assert_fail
("MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC && \"Representative MemoryAccess does not appear to be reverse \" \"mapped properly\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2518, __PRETTY_FUNCTION__))

2517

"Representative MemoryAccess does not appear to be reverse "((MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
"Representative MemoryAccess does not appear to be reverse "
"mapped properly") ? static_cast<void> (0) : __assert_fail
("MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC && \"Representative MemoryAccess does not appear to be reverse \" \"mapped properly\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2518, __PRETTY_FUNCTION__))

2518

"mapped properly")((MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC &&
"Representative MemoryAccess does not appear to be reverse "
"mapped properly") ? static_cast<void> (0) : __assert_fail
("MemoryAccessToClass.lookup(CC->getMemoryLeader()) == CC && \"Representative MemoryAccess does not appear to be reverse \" \"mapped properly\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2518, __PRETTY_FUNCTION__));

2519

for (auto M : CC->memory())

2520

assert(MemoryAccessToClass.lookup(M) == CC &&((MemoryAccessToClass.lookup(M) == CC && "Memory member does not appear to be reverse mapped properly"
) ? static_cast<void> (0) : __assert_fail ("MemoryAccessToClass.lookup(M) == CC && \"Memory member does not appear to be reverse mapped properly\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2521, __PRETTY_FUNCTION__))

2521

"Memory member does not appear to be reverse mapped properly")((MemoryAccessToClass.lookup(M) == CC && "Memory member does not appear to be reverse mapped properly"
) ? static_cast<void> (0) : __assert_fail ("MemoryAccessToClass.lookup(M) == CC && \"Memory member does not appear to be reverse mapped properly\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2521, __PRETTY_FUNCTION__));

2522

}

2523

2524

// Anything equivalent in the MemoryAccess table should be in the same

2525

// congruence class.

2526

2527

// Filter out the unreachable and trivially dead entries, because they may

2528

// never have been updated if the instructions were not processed.

2529

auto ReachableAccessPred =

2530

[&](const std::pair<const MemoryAccess *, CongruenceClass *> Pair) {

2531

bool Result = ReachableBlocks.count(Pair.first->getBlock());

2532

if (!Result || MSSA->isLiveOnEntryDef(Pair.first) ||

2533

MemoryToDFSNum(Pair.first) == 0)

2534

return false;

2535

if (auto *MemDef = dyn_cast<MemoryDef>(Pair.first))

2536

return !isInstructionTriviallyDead(MemDef->getMemoryInst());

2537

2538

// We could have phi nodes which operands are all trivially dead,

2539

// so we don't process them.

2540

if (auto *MemPHI = dyn_cast<MemoryPhi>(Pair.first)) {

2541

for (auto &U : MemPHI->incoming_values()) {

2542

if (Instruction *I = dyn_cast<Instruction>(U.get())) {

2543

if (!isInstructionTriviallyDead(I))

2544

return true;

2545

}

2546

}

2547

return false;

2548

}

2549

2550

return true;

2551

};

2552

2553

auto Filtered = make_filter_range(MemoryAccessToClass, ReachableAccessPred);

2554

for (auto KV : Filtered) {

2555

assert(KV.second != TOPClass &&((KV.second != TOPClass && "Memory not unreachable but ended up in TOP"
) ? static_cast<void> (0) : __assert_fail ("KV.second != TOPClass && \"Memory not unreachable but ended up in TOP\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2556, __PRETTY_FUNCTION__))

2556

"Memory not unreachable but ended up in TOP")((KV.second != TOPClass && "Memory not unreachable but ended up in TOP"
) ? static_cast<void> (0) : __assert_fail ("KV.second != TOPClass && \"Memory not unreachable but ended up in TOP\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2556, __PRETTY_FUNCTION__));

2557

if (auto *FirstMUD = dyn_cast<MemoryUseOrDef>(KV.first)) {

2558

auto *SecondMUD = dyn_cast<MemoryUseOrDef>(KV.second->getMemoryLeader());

2559

if (FirstMUD && SecondMUD)

2560

assert((singleReachablePHIPath(FirstMUD, SecondMUD) ||(((singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass
.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(
SecondMUD->getMemoryInst())) && "The instructions for these memory operations should have "
"been in the same congruence class or reachable through" "a single argument phi"
) ? static_cast<void> (0) : __assert_fail ("(singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(SecondMUD->getMemoryInst())) && \"The instructions for these memory operations should have \" \"been in the same congruence class or reachable through\" \"a single argument phi\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2565, __PRETTY_FUNCTION__))

2561

ValueToClass.lookup(FirstMUD->getMemoryInst()) ==(((singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass
.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(
SecondMUD->getMemoryInst())) && "The instructions for these memory operations should have "
"been in the same congruence class or reachable through" "a single argument phi"
) ? static_cast<void> (0) : __assert_fail ("(singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(SecondMUD->getMemoryInst())) && \"The instructions for these memory operations should have \" \"been in the same congruence class or reachable through\" \"a single argument phi\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2565, __PRETTY_FUNCTION__))

2562

ValueToClass.lookup(SecondMUD->getMemoryInst())) &&(((singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass
.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(
SecondMUD->getMemoryInst())) && "The instructions for these memory operations should have "
"been in the same congruence class or reachable through" "a single argument phi"
) ? static_cast<void> (0) : __assert_fail ("(singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(SecondMUD->getMemoryInst())) && \"The instructions for these memory operations should have \" \"been in the same congruence class or reachable through\" \"a single argument phi\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2565, __PRETTY_FUNCTION__))

2563

"The instructions for these memory operations should have "(((singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass
.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(
SecondMUD->getMemoryInst())) && "The instructions for these memory operations should have "
"been in the same congruence class or reachable through" "a single argument phi"
) ? static_cast<void> (0) : __assert_fail ("(singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(SecondMUD->getMemoryInst())) && \"The instructions for these memory operations should have \" \"been in the same congruence class or reachable through\" \"a single argument phi\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2565, __PRETTY_FUNCTION__))

2564

"been in the same congruence class or reachable through"(((singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass
.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(
SecondMUD->getMemoryInst())) && "The instructions for these memory operations should have "
"been in the same congruence class or reachable through" "a single argument phi"
) ? static_cast<void> (0) : __assert_fail ("(singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(SecondMUD->getMemoryInst())) && \"The instructions for these memory operations should have \" \"been in the same congruence class or reachable through\" \"a single argument phi\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2565, __PRETTY_FUNCTION__))

2565

"a single argument phi")(((singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass
.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(
SecondMUD->getMemoryInst())) && "The instructions for these memory operations should have "
"been in the same congruence class or reachable through" "a single argument phi"
) ? static_cast<void> (0) : __assert_fail ("(singleReachablePHIPath(FirstMUD, SecondMUD) || ValueToClass.lookup(FirstMUD->getMemoryInst()) == ValueToClass.lookup(SecondMUD->getMemoryInst())) && \"The instructions for these memory operations should have \" \"been in the same congruence class or reachable through\" \"a single argument phi\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2565, __PRETTY_FUNCTION__));

2566

} else if (auto *FirstMP = dyn_cast<MemoryPhi>(KV.first)) {

2567

// We can only sanely verify that MemoryDefs in the operand list all have

2568

// the same class.

2569

auto ReachableOperandPred = [&](const Use &U) {

2570

return ReachableEdges.count(

2571

{FirstMP->getIncomingBlock(U), FirstMP->getBlock()}) &&

2572

isa<MemoryDef>(U);

2573

2574

};

2575

// All arguments should in the same class, ignoring unreachable arguments

2576

auto FilteredPhiArgs =

2577

make_filter_range(FirstMP->operands(), ReachableOperandPred);

2578

SmallVector<const CongruenceClass *, 16> PhiOpClasses;

2579

std::transform(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),

2580

std::back_inserter(PhiOpClasses), [&](const Use &U) {

2581

const MemoryDef *MD = cast<MemoryDef>(U);

2582

return ValueToClass.lookup(MD->getMemoryInst());

2583

});

2584

assert(std::equal(PhiOpClasses.begin(), PhiOpClasses.end(),((std::equal(PhiOpClasses.begin(), PhiOpClasses.end(), PhiOpClasses
.begin()) && "All MemoryPhi arguments should be in the same class"
) ? static_cast<void> (0) : __assert_fail ("std::equal(PhiOpClasses.begin(), PhiOpClasses.end(), PhiOpClasses.begin()) && \"All MemoryPhi arguments should be in the same class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2586, __PRETTY_FUNCTION__))

2585

PhiOpClasses.begin()) &&((std::equal(PhiOpClasses.begin(), PhiOpClasses.end(), PhiOpClasses
.begin()) && "All MemoryPhi arguments should be in the same class"
) ? static_cast<void> (0) : __assert_fail ("std::equal(PhiOpClasses.begin(), PhiOpClasses.end(), PhiOpClasses.begin()) && \"All MemoryPhi arguments should be in the same class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2586, __PRETTY_FUNCTION__))

2586

"All MemoryPhi arguments should be in the same class")((std::equal(PhiOpClasses.begin(), PhiOpClasses.end(), PhiOpClasses
.begin()) && "All MemoryPhi arguments should be in the same class"
) ? static_cast<void> (0) : __assert_fail ("std::equal(PhiOpClasses.begin(), PhiOpClasses.end(), PhiOpClasses.begin()) && \"All MemoryPhi arguments should be in the same class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2586, __PRETTY_FUNCTION__));

2587

}

2588

}

2589

#endif

2590

}

2591

2592

// Verify that the sparse propagation we did actually found the maximal fixpoint

2593

// We do this by storing the value to class mapping, touching all instructions,

2594

// and redoing the iteration to see if anything changed.

2595

void NewGVN::verifyIterationSettled(Function &F) {

2596

#ifndef NDEBUG

2597

DEBUG(dbgs() << "Beginning iteration verification\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Beginning iteration verification\n"
; } } while (false);

2598

if (DebugCounter::isCounterSet(VNCounter))

2599

DebugCounter::setCounterValue(VNCounter, StartingVNCounter);

2600

2601

// Note that we have to store the actual classes, as we may change existing

2602

// classes during iteration. This is because our memory iteration propagation

2603

// is not perfect, and so may waste a little work. But it should generate

2604

// exactly the same congruence classes we have now, with different IDs.

2605

std::map<const Value *, CongruenceClass> BeforeIteration;

2606

2607

for (auto &KV : ValueToClass) {

2608

if (auto *I = dyn_cast<Instruction>(KV.first))

2609

// Skip unused/dead instructions.

2610

if (InstrToDFSNum(I) == 0)

2611

continue;

2612

BeforeIteration.insert({KV.first, *KV.second});

2613

}

2614

2615

TouchedInstructions.set();

2616

TouchedInstructions.reset(0);

2617

iterateTouchedInstructions();

2618

DenseSet<std::pair<const CongruenceClass *, const CongruenceClass *>>

2619

EqualClasses;

2620

for (const auto &KV : ValueToClass) {

2621

if (auto *I = dyn_cast<Instruction>(KV.first))

2622

// Skip unused/dead instructions.

2623

if (InstrToDFSNum(I) == 0)

2624

continue;

2625

// We could sink these uses, but i think this adds a bit of clarity here as

2626

// to what we are comparing.

2627

auto *BeforeCC = &BeforeIteration.find(KV.first)->second;

2628

auto *AfterCC = KV.second;

2629

// Note that the classes can't change at this point, so we memoize the set

2630

// that are equal.

2631

if (!EqualClasses.count({BeforeCC, AfterCC})) {

2632

assert(BeforeCC->isEquivalentTo(AfterCC) &&((BeforeCC->isEquivalentTo(AfterCC) && "Value number changed after main loop completed!"
) ? static_cast<void> (0) : __assert_fail ("BeforeCC->isEquivalentTo(AfterCC) && \"Value number changed after main loop completed!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2633, __PRETTY_FUNCTION__))

2633

"Value number changed after main loop completed!")((BeforeCC->isEquivalentTo(AfterCC) && "Value number changed after main loop completed!"
) ? static_cast<void> (0) : __assert_fail ("BeforeCC->isEquivalentTo(AfterCC) && \"Value number changed after main loop completed!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2633, __PRETTY_FUNCTION__));

2634

EqualClasses.insert({BeforeCC, AfterCC});

2635

}

2636

}

2637

#endif

2638

}

2639

2640

// Verify that for each store expression in the expression to class mapping,

2641

// only the latest appears, and multiple ones do not appear.

2642

// Because loads do not use the stored value when doing equality with stores,

2643

// if we don't erase the old store expressions from the table, a load can find

2644

// a no-longer valid StoreExpression.

2645

void NewGVN::verifyStoreExpressions() const {

2646

#ifndef NDEBUG

2647

DenseSet<std::pair<const Value *, const Value *>> StoreExpressionSet;

2648

for (const auto &KV : ExpressionToClass) {

2649

if (auto *SE = dyn_cast<StoreExpression>(KV.first)) {

2650

// Make sure a version that will conflict with loads is not already there

2651

auto Res =

2652

StoreExpressionSet.insert({SE->getOperand(0), SE->getMemoryLeader()});

2653

assert(Res.second &&((Res.second && "Stored expression conflict exists in expression table"
) ? static_cast<void> (0) : __assert_fail ("Res.second && \"Stored expression conflict exists in expression table\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2654, __PRETTY_FUNCTION__))

2654

"Stored expression conflict exists in expression table")((Res.second && "Stored expression conflict exists in expression table"
) ? static_cast<void> (0) : __assert_fail ("Res.second && \"Stored expression conflict exists in expression table\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2654, __PRETTY_FUNCTION__));

2655

auto *ValueExpr = ValueToExpression.lookup(SE->getStoreInst());

2656

assert(ValueExpr && ValueExpr->equals(*SE) &&((ValueExpr && ValueExpr->equals(*SE) && "StoreExpression in ExpressionToClass is not latest "
"StoreExpression for value") ? static_cast<void> (0) :
__assert_fail ("ValueExpr && ValueExpr->equals(*SE) && \"StoreExpression in ExpressionToClass is not latest \" \"StoreExpression for value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2658, __PRETTY_FUNCTION__))

2657

"StoreExpression in ExpressionToClass is not latest "((ValueExpr && ValueExpr->equals(*SE) && "StoreExpression in ExpressionToClass is not latest "
"StoreExpression for value") ? static_cast<void> (0) :
__assert_fail ("ValueExpr && ValueExpr->equals(*SE) && \"StoreExpression in ExpressionToClass is not latest \" \"StoreExpression for value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2658, __PRETTY_FUNCTION__))

2658

"StoreExpression for value")((ValueExpr && ValueExpr->equals(*SE) && "StoreExpression in ExpressionToClass is not latest "
"StoreExpression for value") ? static_cast<void> (0) :
__assert_fail ("ValueExpr && ValueExpr->equals(*SE) && \"StoreExpression in ExpressionToClass is not latest \" \"StoreExpression for value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2658, __PRETTY_FUNCTION__));

2659

}

2660

}

2661

#endif

2662

}

2663

2664

// This is the main value numbering loop, it iterates over the initial touched

2665

// instruction set, propagating value numbers, marking things touched, etc,

2666

// until the set of touched instructions is completely empty.

2667

void NewGVN::iterateTouchedInstructions() {

2668

unsigned int Iterations = 0;

2669

// Figure out where touchedinstructions starts

2670

int FirstInstr = TouchedInstructions.find_first();

2671

// Nothing set, nothing to iterate, just return.

2672

if (FirstInstr == -1)

2673

return;

2674

BasicBlock *LastBlock = getBlockForValue(InstrFromDFSNum(FirstInstr));

2675

while (TouchedInstructions.any()) {

2676

++Iterations;

2677

// Walk through all the instructions in all the blocks in RPO.

2678

// TODO: As we hit a new block, we should push and pop equalities into a

2679

// table lookupOperandLeader can use, to catch things PredicateInfo

2680

// might miss, like edge-only equivalences.

2681

for (unsigned InstrNum : TouchedInstructions.set_bits()) {

2682

2683

// This instruction was found to be dead. We don't bother looking

2684

// at it again.

2685

if (InstrNum == 0) {

2686

TouchedInstructions.reset(InstrNum);

2687

continue;

2688

}

2689

2690

Value *V = InstrFromDFSNum(InstrNum);

2691

BasicBlock *CurrBlock = getBlockForValue(V);

2692

2693

// If we hit a new block, do reachability processing.

2694

if (CurrBlock != LastBlock) {

2695

LastBlock = CurrBlock;

2696

bool BlockReachable = ReachableBlocks.count(CurrBlock);

2697

const auto &CurrInstRange = BlockInstRange.lookup(CurrBlock);

2698

2699

// If it's not reachable, erase any touched instructions and move on.

2700

if (!BlockReachable) {

2701

TouchedInstructions.reset(CurrInstRange.first, CurrInstRange.second);

2702

DEBUG(dbgs() << "Skipping instructions in block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Skipping instructions in block "
<< getBlockName(CurrBlock) << " because it is unreachable\n"
; } } while (false)

2703

<< getBlockName(CurrBlock)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Skipping instructions in block "
<< getBlockName(CurrBlock) << " because it is unreachable\n"
; } } while (false)

2704

<< " because it is unreachable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Skipping instructions in block "
<< getBlockName(CurrBlock) << " because it is unreachable\n"
; } } while (false);

2705

continue;

2706

}

2707

updateProcessedCount(CurrBlock);

2708

}

2709

2710

if (auto *MP = dyn_cast<MemoryPhi>(V)) {

2711

DEBUG(dbgs() << "Processing MemoryPhi " << *MP << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Processing MemoryPhi " <<
*MP << "\n"; } } while (false);

2712

valueNumberMemoryPhi(MP);

2713

} else if (auto *I = dyn_cast<Instruction>(V)) {

2714

valueNumberInstruction(I);

2715

} else {

2716

llvm_unreachable("Should have been a MemoryPhi or Instruction")::llvm::llvm_unreachable_internal("Should have been a MemoryPhi or Instruction"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2716);

2717

}

2718

updateProcessedCount(V);

2719

// Reset after processing (because we may mark ourselves as touched when

2720

// we propagate equalities).

2721

TouchedInstructions.reset(InstrNum);

2722

}

2723

}

2724

NumGVNMaxIterations = std::max(NumGVNMaxIterations.getValue(), Iterations);

2725

}

2726

2727

// This is the main transformation entry point.

2728

bool NewGVN::runGVN() {

2729

if (DebugCounter::isCounterSet(VNCounter))

2730

StartingVNCounter = DebugCounter::getCounterValue(VNCounter);

2731

bool Changed = false;

2732

NumFuncArgs = F.arg_size();

2733

MSSAWalker = MSSA->getWalker();

2734

2735

// Count number of instructions for sizing of hash tables, and come

2736

// up with a global dfs numbering for instructions.

2737

unsigned ICount = 1;

2738

// Add an empty instruction to account for the fact that we start at 1

2739

DFSToInstr.emplace_back(nullptr);

2740

// Note: We want ideal RPO traversal of the blocks, which is not quite the

2741

// same as dominator tree order, particularly with regard whether backedges

2742

// get visited first or second, given a block with multiple successors.

2743

// If we visit in the wrong order, we will end up performing N times as many

2744

// iterations.

2745

// The dominator tree does guarantee that, for a given dom tree node, it's

2746

// parent must occur before it in the RPO ordering. Thus, we only need to sort

2747

// the siblings.

2748

ReversePostOrderTraversal<Function *> RPOT(&F);

2749

unsigned Counter = 0;

2750

for (auto &B : RPOT) {

2751

auto *Node = DT->getNode(B);

2752

assert(Node && "RPO and Dominator tree should have same reachability")((Node && "RPO and Dominator tree should have same reachability"
) ? static_cast<void> (0) : __assert_fail ("Node && \"RPO and Dominator tree should have same reachability\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2752, __PRETTY_FUNCTION__));

2753

RPOOrdering[Node] = ++Counter;

2754

}

2755

// Sort dominator tree children arrays into RPO.

2756

for (auto &B : RPOT) {

2757

auto *Node = DT->getNode(B);

2758

if (Node->getChildren().size() > 1)

2759

std::sort(Node->begin(), Node->end(),

2760

[&](const DomTreeNode *A, const DomTreeNode *B) {

2761

return RPOOrdering[A] < RPOOrdering[B];

2762

});

2763

}

2764

2765

// Now a standard depth first ordering of the domtree is equivalent to RPO.

2766

for (auto DTN : depth_first(DT->getRootNode())) {

2767

BasicBlock *B = DTN->getBlock();

2768

const auto &BlockRange = assignDFSNumbers(B, ICount);

2769

BlockInstRange.insert({B, BlockRange});

2770

ICount += BlockRange.second - BlockRange.first;

2771

}

2772

2773

TouchedInstructions.resize(ICount);

2774

// Ensure we don't end up resizing the expressionToClass map, as

2775

// that can be quite expensive. At most, we have one expression per

2776

// instruction.

2777

ExpressionToClass.reserve(ICount);

2778

2779

// Initialize the touched instructions to include the entry block.

2780

const auto &InstRange = BlockInstRange.lookup(&F.getEntryBlock());

2781

TouchedInstructions.set(InstRange.first, InstRange.second);

2782

ReachableBlocks.insert(&F.getEntryBlock());

2783

2784

initializeCongruenceClasses(F);

2785

iterateTouchedInstructions();

2786

verifyMemoryCongruency();

2787

verifyIterationSettled(F);

2788

verifyStoreExpressions();

2789

2790

Changed |= eliminateInstructions(F);

2791

2792

// Delete all instructions marked for deletion.

2793

for (Instruction *ToErase : InstructionsToErase) {

2794

if (!ToErase->use_empty())

2795

ToErase->replaceAllUsesWith(UndefValue::get(ToErase->getType()));

2796

2797

ToErase->eraseFromParent();

2798

}

2799

2800

// Delete all unreachable blocks.

2801

auto UnreachableBlockPred = [&](const BasicBlock &BB) {

2802

return !ReachableBlocks.count(&BB);

2803

};

2804

2805

for (auto &BB : make_filter_range(F, UnreachableBlockPred)) {

2806

DEBUG(dbgs() << "We believe block " << getBlockName(&BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "We believe block " << getBlockName
(&BB) << " is unreachable\n"; } } while (false)

2807

<< " is unreachable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "We believe block " << getBlockName
(&BB) << " is unreachable\n"; } } while (false);

2808

deleteInstructionsInBlock(&BB);

2809

Changed = true;

2810

}

2811

2812

cleanupTables();

2813

return Changed;

2814

}

2815

2816

// Return true if V is a value that will always be available (IE can

2817

// be placed anywhere) in the function. We don't do globals here

2818

// because they are often worse to put in place.

2819

// TODO: Separate cost from availability

2820

static bool alwaysAvailable(Value *V) {

2821

return isa<Constant>(V) || isa<Argument>(V);

2822

}

2823

2824

struct NewGVN::ValueDFS {

2825

int DFSIn = 0;

2826

int DFSOut = 0;

2827

int LocalNum = 0;

2828

// Only one of Def and U will be set.

2829

// The bool in the Def tells us whether the Def is the stored value of a

2830

// store.

2831

PointerIntPair<Value *, 1, bool> Def;

2832

Use *U = nullptr;

2833

bool operator<(const ValueDFS &Other) const {

2834

// It's not enough that any given field be less than - we have sets

2835

// of fields that need to be evaluated together to give a proper ordering.

2836

// For example, if you have;

2837

// DFS (1, 3)

2838

// Val 0

2839

// DFS (1, 2)

2840

// Val 50

2841

// We want the second to be less than the first, but if we just go field

2842

// by field, we will get to Val 0 < Val 50 and say the first is less than

2843

// the second. We only want it to be less than if the DFS orders are equal.

2844

2845

// Each LLVM instruction only produces one value, and thus the lowest-level

2846

// differentiator that really matters for the stack (and what we use as as a

2847

// replacement) is the local dfs number.

2848

// Everything else in the structure is instruction level, and only affects

2849

// the order in which we will replace operands of a given instruction.

2850

2851

// For a given instruction (IE things with equal dfsin, dfsout, localnum),

2852

// the order of replacement of uses does not matter.

2853

// IE given,

2854

// a = 5

2855

// b = a + a

2856

// When you hit b, you will have two valuedfs with the same dfsin, out, and

2857

// localnum.

2858

// The .val will be the same as well.

2859

// The .u's will be different.

2860

// You will replace both, and it does not matter what order you replace them

2861

// in (IE whether you replace operand 2, then operand 1, or operand 1, then

2862

// operand 2).

2863

// Similarly for the case of same dfsin, dfsout, localnum, but different

2864

// .val's

2865

// a = 5

2866

// b = 6

2867

// c = a + b

2868

// in c, we will a valuedfs for a, and one for b,with everything the same

2869

// but .val and .u.

2870

// It does not matter what order we replace these operands in.

2871

// You will always end up with the same IR, and this is guaranteed.

2872

return std::tie(DFSIn, DFSOut, LocalNum, Def, U) <

2873

std::tie(Other.DFSIn, Other.DFSOut, Other.LocalNum, Other.Def,

2874

Other.U);

2875

}

2876

};

2877

2878

// This function converts the set of members for a congruence class from values,

2879

// to sets of defs and uses with associated DFS info. The total number of

2880

// reachable uses for each value is stored in UseCount, and instructions that

2881

// seem

2882

// dead (have no non-dead uses) are stored in ProbablyDead.

2883

void NewGVN::convertClassToDFSOrdered(

2884

const CongruenceClass &Dense, SmallVectorImpl<ValueDFS> &DFSOrderedSet,

2885

DenseMap<const Value *, unsigned int> &UseCounts,

2886

SmallPtrSetImpl<Instruction *> &ProbablyDead) const {

2887

for (auto D : Dense) {

2888

// First add the value.

2889

BasicBlock *BB = getBlockForValue(D);

2890

// Constants are handled prior to ever calling this function, so

2891

// we should only be left with instructions as members.

2892

assert(BB && "Should have figured out a basic block for value")((BB && "Should have figured out a basic block for value"
) ? static_cast<void> (0) : __assert_fail ("BB && \"Should have figured out a basic block for value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2892, __PRETTY_FUNCTION__));

2893

ValueDFS VDDef;

2894

DomTreeNode *DomNode = DT->getNode(BB);

2895

VDDef.DFSIn = DomNode->getDFSNumIn();

2896

VDDef.DFSOut = DomNode->getDFSNumOut();

2897

// If it's a store, use the leader of the value operand, if it's always

2898

// available, or the value operand. TODO: We could do dominance checks to

2899

// find a dominating leader, but not worth it ATM.

2900

if (auto *SI = dyn_cast<StoreInst>(D)) {

2901

auto Leader = lookupOperandLeader(SI->getValueOperand());

2902

if (alwaysAvailable(Leader)) {

2903

VDDef.Def.setPointer(Leader);

2904

} else {

2905

VDDef.Def.setPointer(SI->getValueOperand());

2906

VDDef.Def.setInt(true);

2907

}

2908

} else {

2909

VDDef.Def.setPointer(D);

2910

}

2911

assert(isa<Instruction>(D) &&((isa<Instruction>(D) && "The dense set member should always be an instruction"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(D) && \"The dense set member should always be an instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2912, __PRETTY_FUNCTION__))

2912

"The dense set member should always be an instruction")((isa<Instruction>(D) && "The dense set member should always be an instruction"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(D) && \"The dense set member should always be an instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2912, __PRETTY_FUNCTION__));

2913

VDDef.LocalNum = InstrToDFSNum(D);

2914

DFSOrderedSet.emplace_back(VDDef);

2915

Instruction *Def = cast<Instruction>(D);

2916

unsigned int UseCount = 0;

2917

// Now add the uses.

2918

for (auto &U : Def->uses()) {

2919

if (auto *I = dyn_cast<Instruction>(U.getUser())) {

2920

// Don't try to replace into dead uses

2921

if (InstructionsToErase.count(I))

2922

continue;

2923

ValueDFS VDUse;

2924

// Put the phi node uses in the incoming block.

2925

BasicBlock *IBlock;

2926

if (auto *P = dyn_cast<PHINode>(I)) {

2927

IBlock = P->getIncomingBlock(U);

2928

// Make phi node users appear last in the incoming block

2929

// they are from.

2930

VDUse.LocalNum = InstrDFS.size() + 1;

2931

} else {

2932

IBlock = I->getParent();

2933

VDUse.LocalNum = InstrToDFSNum(I);

2934

}

2935

2936

// Skip uses in unreachable blocks, as we're going

2937

// to delete them.

2938

if (ReachableBlocks.count(IBlock) == 0)

2939

continue;

2940

2941

DomTreeNode *DomNode = DT->getNode(IBlock);

2942

VDUse.DFSIn = DomNode->getDFSNumIn();

2943

VDUse.DFSOut = DomNode->getDFSNumOut();

2944

VDUse.U = &U;

2945

++UseCount;

2946

DFSOrderedSet.emplace_back(VDUse);

2947

}

2948

}

2949

2950

// If there are no uses, it's probably dead (but it may have side-effects,

2951

// so not definitely dead. Otherwise, store the number of uses so we can

2952

// track if it becomes dead later).

2953

if (UseCount == 0)

2954

ProbablyDead.insert(Def);

2955

else

2956

UseCounts[Def] = UseCount;

2957

}

2958

}

2959

2960

// This function converts the set of members for a congruence class from values,

2961

// to the set of defs for loads and stores, with associated DFS info.

2962

void NewGVN::convertClassToLoadsAndStores(

2963

const CongruenceClass &Dense,

2964

SmallVectorImpl<ValueDFS> &LoadsAndStores) const {

2965

for (auto D : Dense) {

2966

if (!isa<LoadInst>(D) && !isa<StoreInst>(D))

2967

continue;

2968

2969

BasicBlock *BB = getBlockForValue(D);

2970

ValueDFS VD;

2971

DomTreeNode *DomNode = DT->getNode(BB);

2972

VD.DFSIn = DomNode->getDFSNumIn();

2973

VD.DFSOut = DomNode->getDFSNumOut();

2974

VD.Def.setPointer(D);

2975

2976

// If it's an instruction, use the real local dfs number.

2977

if (auto *I = dyn_cast<Instruction>(D))

2978

VD.LocalNum = InstrToDFSNum(I);

2979

else

2980

llvm_unreachable("Should have been an instruction")::llvm::llvm_unreachable_internal("Should have been an instruction"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 2980);

2981

2982

LoadsAndStores.emplace_back(VD);

2983

}

2984

}

2985

2986

static void patchReplacementInstruction(Instruction *I, Value *Repl) {

2987

auto *ReplInst = dyn_cast<Instruction>(Repl);

2988

if (!ReplInst)

2989

return;

2990

2991

// Patch the replacement so that it is not more restrictive than the value

2992

// being replaced.

2993

// Note that if 'I' is a load being replaced by some operation,

2994

// for example, by an arithmetic operation, then andIRFlags()

2995

// would just erase all math flags from the original arithmetic

2996

// operation, which is clearly not wanted and not needed.

2997

if (!isa<LoadInst>(I))

2998

ReplInst->andIRFlags(I);

2999

3000

// FIXME: If both the original and replacement value are part of the

3001

// same control-flow region (meaning that the execution of one

3002

// guarantees the execution of the other), then we can combine the

3003

// noalias scopes here and do better than the general conservative

3004

// answer used in combineMetadata().

3005

3006

// In general, GVN unifies expressions over different control-flow

3007

// regions, and so we need a conservative combination of the noalias

3008

// scopes.

3009

static const unsigned KnownIDs[] = {

3010

LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,

3011

LLVMContext::MD_noalias, LLVMContext::MD_range,

3012

LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,

3013

LLVMContext::MD_invariant_group};

3014

combineMetadata(ReplInst, I, KnownIDs);

3015

}

3016

3017

static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {

3018

patchReplacementInstruction(I, Repl);

3019

I->replaceAllUsesWith(Repl);

3020

}

3021

3022

void NewGVN::deleteInstructionsInBlock(BasicBlock *BB) {

3023

DEBUG(dbgs() << " BasicBlock Dead:" << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << " BasicBlock Dead:" << *
BB; } } while (false);

3024

++NumGVNBlocksDeleted;

3025

3026

// Delete the instructions backwards, as it has a reduced likelihood of having

3027

// to update as many def-use and use-def chains. Start after the terminator.

3028

auto StartPoint = BB->rbegin();

3029

++StartPoint;

3030

// Note that we explicitly recalculate BB->rend() on each iteration,

3031

// as it may change when we remove the first instruction.

3032

for (BasicBlock::reverse_iterator I(StartPoint); I != BB->rend();) {

3033

Instruction &Inst = *I++;

3034

if (!Inst.use_empty())

3035

Inst.replaceAllUsesWith(UndefValue::get(Inst.getType()));

3036

if (isa<LandingPadInst>(Inst))

3037

continue;

3038

3039

Inst.eraseFromParent();

3040

++NumGVNInstrDeleted;

3041

}

3042

// Now insert something that simplifycfg will turn into an unreachable.

3043

Type *Int8Ty = Type::getInt8Ty(BB->getContext());

3044

new StoreInst(UndefValue::get(Int8Ty),

3045

Constant::getNullValue(Int8Ty->getPointerTo()),

3046

BB->getTerminator());

3047

}

3048

3049

void NewGVN::markInstructionForDeletion(Instruction *I) {

3050

DEBUG(dbgs() << "Marking " << *I << " for deletion\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Marking " << *I <<
" for deletion\n"; } } while (false);

3051

InstructionsToErase.insert(I);

3052

}

3053

3054

void NewGVN::replaceInstruction(Instruction *I, Value *V) {

3055

3056

DEBUG(dbgs() << "Replacing " << *I << " with " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Replacing " << *I <<
" with " << *V << "\n"; } } while (false);

3057

patchAndReplaceAllUsesWith(I, V);

3058

// We save the actual erasing to avoid invalidating memory

3059

// dependencies until we are done with everything.

3060

markInstructionForDeletion(I);

3061

}

3062

3063

namespace {

3064

3065

// This is a stack that contains both the value and dfs info of where

3066

// that value is valid.

3067

class ValueDFSStack {

3068

public:

3069

Value *back() const { return ValueStack.back(); }

3070

std::pair<int, int> dfs_back() const { return DFSStack.back(); }

3071

3072

void push_back(Value *V, int DFSIn, int DFSOut) {

3073

ValueStack.emplace_back(V);

3074

DFSStack.emplace_back(DFSIn, DFSOut);

3075

}

3076

bool empty() const { return DFSStack.empty(); }

3077

bool isInScope(int DFSIn, int DFSOut) const {

3078

if (empty())

3079

return false;

3080

return DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second;

3081

}

3082

3083

void popUntilDFSScope(int DFSIn, int DFSOut) {

3084

3085

// These two should always be in sync at this point.

3086

assert(ValueStack.size() == DFSStack.size() &&((ValueStack.size() == DFSStack.size() && "Mismatch between ValueStack and DFSStack"
) ? static_cast<void> (0) : __assert_fail ("ValueStack.size() == DFSStack.size() && \"Mismatch between ValueStack and DFSStack\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3087, __PRETTY_FUNCTION__))

3087

"Mismatch between ValueStack and DFSStack")((ValueStack.size() == DFSStack.size() && "Mismatch between ValueStack and DFSStack"
) ? static_cast<void> (0) : __assert_fail ("ValueStack.size() == DFSStack.size() && \"Mismatch between ValueStack and DFSStack\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3087, __PRETTY_FUNCTION__));

3088

while (

3089

!DFSStack.empty() &&

3090

!(DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second)) {

3091

DFSStack.pop_back();

3092

ValueStack.pop_back();

3093

}

3094

}

3095

3096

private:

3097

SmallVector<Value *, 8> ValueStack;

3098

SmallVector<std::pair<int, int>, 8> DFSStack;

3099

};

3100

}

3101

3102

bool NewGVN::eliminateInstructions(Function &F) {

3103

// This is a non-standard eliminator. The normal way to eliminate is

3104

// to walk the dominator tree in order, keeping track of available

3105

// values, and eliminating them. However, this is mildly

3106

// pointless. It requires doing lookups on every instruction,

3107

// regardless of whether we will ever eliminate it. For

3108

// instructions part of most singleton congruence classes, we know we

3109

// will never eliminate them.

3110

3111

// Instead, this eliminator looks at the congruence classes directly, sorts

3112

// them into a DFS ordering of the dominator tree, and then we just

3113

// perform elimination straight on the sets by walking the congruence

3114

// class member uses in order, and eliminate the ones dominated by the

3115

// last member. This is worst case O(E log E) where E = number of

3116

// instructions in a single congruence class. In theory, this is all

3117

// instructions. In practice, it is much faster, as most instructions are

3118

// either in singleton congruence classes or can't possibly be eliminated

3119

// anyway (if there are no overlapping DFS ranges in class).

3120

// When we find something not dominated, it becomes the new leader

3121

// for elimination purposes.

3122

// TODO: If we wanted to be faster, We could remove any members with no

3123

// overlapping ranges while sorting, as we will never eliminate anything

3124

// with those members, as they don't dominate anything else in our set.

3125

3126

bool AnythingReplaced = false;

3127

3128

// Since we are going to walk the domtree anyway, and we can't guarantee the

3129

// DFS numbers are updated, we compute some ourselves.

3130

DT->updateDFSNumbers();

3131

3132

for (auto &B : F) {

3133

if (!ReachableBlocks.count(&B)) {

3134

for (const auto S : successors(&B)) {

3135

for (auto II = S->begin(); isa<PHINode>(II); ++II) {

3136

auto &Phi = cast<PHINode>(*II);

3137

DEBUG(dbgs() << "Replacing incoming value of " << *II << " for block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Replacing incoming value of " <<
*II << " for block " << getBlockName(&B) <<
" with undef due to it being unreachable\n"; } } while (false
)

3138

<< getBlockName(&B)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Replacing incoming value of " <<
*II << " for block " << getBlockName(&B) <<
" with undef due to it being unreachable\n"; } } while (false
)

3139

<< " with undef due to it being unreachable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Replacing incoming value of " <<
*II << " for block " << getBlockName(&B) <<
" with undef due to it being unreachable\n"; } } while (false
);

3140

for (auto &Operand : Phi.incoming_values())

3141

if (Phi.getIncomingBlock(Operand) == &B)

3142

Operand.set(UndefValue::get(Phi.getType()));

3143

}

3144

}

3145

}

3146

}

3147

3148

// Map to store the use counts

3149

DenseMap<const Value *, unsigned int> UseCounts;

3150

for (CongruenceClass *CC : reverse(CongruenceClasses)) {

3151

// Track the equivalent store info so we can decide whether to try

3152

// dead store elimination.

3153

SmallVector<ValueDFS, 8> PossibleDeadStores;

3154

SmallPtrSet<Instruction *, 8> ProbablyDead;

3155

if (CC->isDead() || CC->empty())

3156

continue;

3157

// Everything still in the TOP class is unreachable or dead.

3158

if (CC == TOPClass) {

3159

#ifndef NDEBUG

3160

for (auto M : *CC)

3161

assert((!ReachableBlocks.count(cast<Instruction>(M)->getParent()) ||(((!ReachableBlocks.count(cast<Instruction>(M)->getParent
()) || InstructionsToErase.count(cast<Instruction>(M)))
&& "Everything in TOP should be unreachable or dead at this "
"point") ? static_cast<void> (0) : __assert_fail ("(!ReachableBlocks.count(cast<Instruction>(M)->getParent()) || InstructionsToErase.count(cast<Instruction>(M))) && \"Everything in TOP should be unreachable or dead at this \" \"point\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3164, __PRETTY_FUNCTION__))

3162

InstructionsToErase.count(cast<Instruction>(M))) &&(((!ReachableBlocks.count(cast<Instruction>(M)->getParent
()) || InstructionsToErase.count(cast<Instruction>(M)))
&& "Everything in TOP should be unreachable or dead at this "
"point") ? static_cast<void> (0) : __assert_fail ("(!ReachableBlocks.count(cast<Instruction>(M)->getParent()) || InstructionsToErase.count(cast<Instruction>(M))) && \"Everything in TOP should be unreachable or dead at this \" \"point\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3164, __PRETTY_FUNCTION__))

3163

"Everything in TOP should be unreachable or dead at this "(((!ReachableBlocks.count(cast<Instruction>(M)->getParent
()) || InstructionsToErase.count(cast<Instruction>(M)))
&& "Everything in TOP should be unreachable or dead at this "
"point") ? static_cast<void> (0) : __assert_fail ("(!ReachableBlocks.count(cast<Instruction>(M)->getParent()) || InstructionsToErase.count(cast<Instruction>(M))) && \"Everything in TOP should be unreachable or dead at this \" \"point\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3164, __PRETTY_FUNCTION__))

3164

"point")(((!ReachableBlocks.count(cast<Instruction>(M)->getParent
()) || InstructionsToErase.count(cast<Instruction>(M)))
&& "Everything in TOP should be unreachable or dead at this "
"point") ? static_cast<void> (0) : __assert_fail ("(!ReachableBlocks.count(cast<Instruction>(M)->getParent()) || InstructionsToErase.count(cast<Instruction>(M))) && \"Everything in TOP should be unreachable or dead at this \" \"point\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3164, __PRETTY_FUNCTION__));

3165

#endif

3166

continue;

3167

}

3168

3169

assert(CC->getLeader() && "We should have had a leader")((CC->getLeader() && "We should have had a leader"
) ? static_cast<void> (0) : __assert_fail ("CC->getLeader() && \"We should have had a leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3169, __PRETTY_FUNCTION__));

3170

// If this is a leader that is always available, and it's a

3171

// constant or has no equivalences, just replace everything with

3172

// it. We then update the congruence class with whatever members

3173

// are left.

3174

Value *Leader =

3175

CC->getStoredValue() ? CC->getStoredValue() : CC->getLeader();

3176

if (alwaysAvailable(Leader)) {

3177

CongruenceClass::MemberSet MembersLeft;

3178

for (auto M : *CC) {

3179

Value *Member = M;

3180

// Void things have no uses we can replace.

3181

if (Member == Leader || !isa<Instruction>(Member) ||

3182

Member->getType()->isVoidTy()) {

3183

MembersLeft.insert(Member);

3184

continue;

3185

}

3186

DEBUG(dbgs() << "Found replacement " << *(Leader) << " for " << *Memberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found replacement " << *
(Leader) << " for " << *Member << "\n"; } }
while (false)

3187

<< "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found replacement " << *
(Leader) << " for " << *Member << "\n"; } }
while (false);

3188

auto *I = cast<Instruction>(Member);

3189

assert(Leader != I && "About to accidentally remove our leader")((Leader != I && "About to accidentally remove our leader"
) ? static_cast<void> (0) : __assert_fail ("Leader != I && \"About to accidentally remove our leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3189, __PRETTY_FUNCTION__));

3190

replaceInstruction(I, Leader);

3191

AnythingReplaced = true;

3192

}

3193

CC->swap(MembersLeft);

3194

} else {

3195

DEBUG(dbgs() << "Eliminating in congruence class " << CC->getID()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Eliminating in congruence class "
<< CC->getID() << "\n"; } } while (false)

3196

<< "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Eliminating in congruence class "
<< CC->getID() << "\n"; } } while (false);

3197

// If this is a singleton, we can skip it.

3198

if (CC->size() != 1) {

3199

// This is a stack because equality replacement/etc may place

3200

// constants in the middle of the member list, and we want to use

3201

// those constant values in preference to the current leader, over

3202

// the scope of those constants.

3203

ValueDFSStack EliminationStack;

3204

3205

// Convert the members to DFS ordered sets and then merge them.

3206

SmallVector<ValueDFS, 8> DFSOrderedSet;

3207

convertClassToDFSOrdered(*CC, DFSOrderedSet, UseCounts, ProbablyDead);

3208

3209

// Sort the whole thing.

3210

std::sort(DFSOrderedSet.begin(), DFSOrderedSet.end());

3211

for (auto &VD : DFSOrderedSet) {

3212

int MemberDFSIn = VD.DFSIn;

3213

int MemberDFSOut = VD.DFSOut;

3214

Value *Def = VD.Def.getPointer();

3215

bool FromStore = VD.Def.getInt();

3216

Use *U = VD.U;

3217

// We ignore void things because we can't get a value from them.

3218

if (Def && Def->getType()->isVoidTy())

3219

continue;

3220

3221

if (EliminationStack.empty()) {

3222

DEBUG(dbgs() << "Elimination Stack is empty\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Elimination Stack is empty\n";
} } while (false);

3223

} else {

3224

DEBUG(dbgs() << "Elimination Stack Top DFS numbers are ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Elimination Stack Top DFS numbers are ("
<< EliminationStack.dfs_back().first << "," <<
EliminationStack.dfs_back().second << ")\n"; } } while
(false)

3225

<< EliminationStack.dfs_back().first << ","do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Elimination Stack Top DFS numbers are ("
<< EliminationStack.dfs_back().first << "," <<
EliminationStack.dfs_back().second << ")\n"; } } while
(false)

3226

<< EliminationStack.dfs_back().second << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Elimination Stack Top DFS numbers are ("
<< EliminationStack.dfs_back().first << "," <<
EliminationStack.dfs_back().second << ")\n"; } } while
(false);

3227

}

3228

3229

DEBUG(dbgs() << "Current DFS numbers are (" << MemberDFSIn << ","do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Current DFS numbers are (" <<
MemberDFSIn << "," << MemberDFSOut << ")\n"
; } } while (false)

3230

<< MemberDFSOut << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Current DFS numbers are (" <<
MemberDFSIn << "," << MemberDFSOut << ")\n"
; } } while (false);

3231

// First, we see if we are out of scope or empty. If so,

3232

// and there equivalences, we try to replace the top of

3233

// stack with equivalences (if it's on the stack, it must

3234

// not have been eliminated yet).

3235

// Then we synchronize to our current scope, by

3236

// popping until we are back within a DFS scope that

3237

// dominates the current member.

3238

// Then, what happens depends on a few factors

3239

// If the stack is now empty, we need to push

3240

// If we have a constant or a local equivalence we want to

3241

// start using, we also push.

3242

// Otherwise, we walk along, processing members who are

3243

// dominated by this scope, and eliminate them.

3244

bool ShouldPush = Def && EliminationStack.empty();

3245

bool OutOfScope =

3246

!EliminationStack.isInScope(MemberDFSIn, MemberDFSOut);

3247

3248

if (OutOfScope || ShouldPush) {

3249

// Sync to our current scope.

3250

EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);

3251

bool ShouldPush = Def && EliminationStack.empty();

3252

if (ShouldPush) {

3253

EliminationStack.push_back(Def, MemberDFSIn, MemberDFSOut);

3254

}

3255

}

3256

3257

// Skip the Def's, we only want to eliminate on their uses. But mark

3258

// dominated defs as dead.

3259

if (Def) {

3260

// For anything in this case, what and how we value number

3261

// guarantees that any side-effets that would have occurred (ie

3262

// throwing, etc) can be proven to either still occur (because it's

3263

// dominated by something that has the same side-effects), or never

3264

// occur. Otherwise, we would not have been able to prove it value

3265

// equivalent to something else. For these things, we can just mark

3266

// it all dead. Note that this is different from the "ProbablyDead"

3267

// set, which may not be dominated by anything, and thus, are only

3268

// easy to prove dead if they are also side-effect free. Note that

3269

// because stores are put in terms of the stored value, we skip

3270

// stored values here. If the stored value is really dead, it will

3271

// still be marked for deletion when we process it in its own class.

3272

if (!EliminationStack.empty() && Def != EliminationStack.back() &&

3273

isa<Instruction>(Def) && !FromStore)

3274

markInstructionForDeletion(cast<Instruction>(Def));

3275

continue;

3276

}

3277

// At this point, we know it is a Use we are trying to possibly

3278

// replace.

3279

3280

assert(isa<Instruction>(U->get()) &&((isa<Instruction>(U->get()) && "Current def should have been an instruction"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(U->get()) && \"Current def should have been an instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3281, __PRETTY_FUNCTION__))

3281

"Current def should have been an instruction")((isa<Instruction>(U->get()) && "Current def should have been an instruction"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(U->get()) && \"Current def should have been an instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3281, __PRETTY_FUNCTION__));

3282

assert(isa<Instruction>(U->getUser()) &&((isa<Instruction>(U->getUser()) && "Current user should have been an instruction"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(U->getUser()) && \"Current user should have been an instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3283, __PRETTY_FUNCTION__))

3283

"Current user should have been an instruction")((isa<Instruction>(U->getUser()) && "Current user should have been an instruction"
) ? static_cast<void> (0) : __assert_fail ("isa<Instruction>(U->getUser()) && \"Current user should have been an instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3283, __PRETTY_FUNCTION__));

3284

3285

// If the thing we are replacing into is already marked to be dead,

3286

// this use is dead. Note that this is true regardless of whether

3287

// we have anything dominating the use or not. We do this here

3288

// because we are already walking all the uses anyway.

3289

Instruction *InstUse = cast<Instruction>(U->getUser());

3290

if (InstructionsToErase.count(InstUse)) {

3291

auto &UseCount = UseCounts[U->get()];

3292

if (--UseCount == 0) {

3293

ProbablyDead.insert(cast<Instruction>(U->get()));

3294

}

3295

}

3296

3297

// If we get to this point, and the stack is empty we must have a use

3298

// with nothing we can use to eliminate this use, so just skip it.

3299

if (EliminationStack.empty())

3300

continue;

3301

3302

Value *DominatingLeader = EliminationStack.back();

3303

3304

// Don't replace our existing users with ourselves.

3305

if (U->get() == DominatingLeader)

3306

continue;

3307

DEBUG(dbgs() << "Found replacement " << *DominatingLeader << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found replacement " << *
DominatingLeader << " for " << *U->get() <<
" in " << *(U->getUser()) << "\n"; } } while (
false)

3308

<< *U->get() << " in " << *(U->getUser()) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found replacement " << *
DominatingLeader << " for " << *U->get() <<
" in " << *(U->getUser()) << "\n"; } } while (
false);

3309

3310

// If we replaced something in an instruction, handle the patching of

3311

// metadata. Skip this if we are replacing predicateinfo with its

3312

// original operand, as we already know we can just drop it.

3313

auto *ReplacedInst = cast<Instruction>(U->get());

3314

auto *PI = PredInfo->getPredicateInfoFor(ReplacedInst);

3315

if (!PI || DominatingLeader != PI->OriginalOp)

3316

patchReplacementInstruction(ReplacedInst, DominatingLeader);

3317

U->set(DominatingLeader);

3318

// This is now a use of the dominating leader, which means if the

3319

// dominating leader was dead, it's now live!

3320

auto &LeaderUseCount = UseCounts[DominatingLeader];

3321

// It's about to be alive again.

3322

if (LeaderUseCount == 0 && isa<Instruction>(DominatingLeader))

3323

ProbablyDead.erase(cast<Instruction>(DominatingLeader));

3324

++LeaderUseCount;

3325

AnythingReplaced = true;

3326

}

3327

}

3328

}

3329

3330

// At this point, anything still in the ProbablyDead set is actually dead if

3331

// would be trivially dead.

3332

for (auto *I : ProbablyDead)

3333

if (wouldInstructionBeTriviallyDead(I))

3334

markInstructionForDeletion(I);

3335

3336

// Cleanup the congruence class.

3337

CongruenceClass::MemberSet MembersLeft;

3338

for (auto *Member : *CC)

3339

if (!isa<Instruction>(Member) ||

3340

!InstructionsToErase.count(cast<Instruction>(Member)))

3341

MembersLeft.insert(Member);

3342

CC->swap(MembersLeft);

3343

3344

// If we have possible dead stores to look at, try to eliminate them.

3345

if (CC->getStoreCount() > 0) {

3346

convertClassToLoadsAndStores(*CC, PossibleDeadStores);

3347

std::sort(PossibleDeadStores.begin(), PossibleDeadStores.end());

3348

ValueDFSStack EliminationStack;

3349

for (auto &VD : PossibleDeadStores) {

3350

int MemberDFSIn = VD.DFSIn;

3351

int MemberDFSOut = VD.DFSOut;

3352

Instruction *Member = cast<Instruction>(VD.Def.getPointer());

3353

if (EliminationStack.empty() ||

3354

!EliminationStack.isInScope(MemberDFSIn, MemberDFSOut)) {

3355

// Sync to our current scope.

3356

EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);

3357

if (EliminationStack.empty()) {

3358

EliminationStack.push_back(Member, MemberDFSIn, MemberDFSOut);

3359

continue;

3360

}

3361

}

3362

// We already did load elimination, so nothing to do here.

3363

if (isa<LoadInst>(Member))

3364

continue;

3365

assert(!EliminationStack.empty())((!EliminationStack.empty()) ? static_cast<void> (0) : __assert_fail
("!EliminationStack.empty()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3365, __PRETTY_FUNCTION__));

3366

Instruction *Leader = cast<Instruction>(EliminationStack.back());

3367

(void)Leader;

3368

assert(DT->dominates(Leader->getParent(), Member->getParent()))((DT->dominates(Leader->getParent(), Member->getParent
())) ? static_cast<void> (0) : __assert_fail ("DT->dominates(Leader->getParent(), Member->getParent())"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn303373/lib/Transforms/Scalar/NewGVN.cpp"
, 3368, __PRETTY_FUNCTION__));

3369

// Member is dominater by Leader, and thus dead

3370

DEBUG(dbgs() << "Marking dead store " << *Memberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Marking dead store " << *
Member << " that is dominated by " << *Leader <<
"\n"; } } while (false)

3371

<< " that is dominated by " << *Leader << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Marking dead store " << *
Member << " that is dominated by " << *Leader <<
"\n"; } } while (false);

3372

markInstructionForDeletion(Member);

3373

CC->erase(Member);

3374

++NumGVNDeadStores;

3375

}

3376

}

3377

}

3378

3379

return AnythingReplaced;

3380

}

3381

3382

// This function provides global ranking of operations so that we can place them

3383

// in a canonical order. Note that rank alone is not necessarily enough for a

3384

// complete ordering, as constants all have the same rank. However, generally,

3385

// we will simplify an operation with all constants so that it doesn't matter

3386

// what order they appear in.

3387

unsigned int NewGVN::getRank(const Value *V) const {

3388

// Prefer undef to anything else

3389

if (isa<UndefValue>(V))

3390

return 0;

3391

if (isa<Constant>(V))

3392

return 1;

3393

else if (auto *A = dyn_cast<Argument>(V))

3394

return 2 + A->getArgNo();

3395

3396

// Need to shift the instruction DFS by number of arguments + 3 to account for

3397

// the constant and argument ranking above.

3398

unsigned Result = InstrToDFSNum(V);

3399

if (Result > 0)

3400

return 3 + NumFuncArgs + Result;

3401

// Unreachable or something else, just return a really large number.

3402

return ~0;

3403

}

3404

3405

// This is a function that says whether two commutative operations should

3406

// have their order swapped when canonicalizing.

3407

bool NewGVN::shouldSwapOperands(const Value *A, const Value *B) const {

3408

// Because we only care about a total ordering, and don't rewrite expressions

3409

// in this order, we order by rank, which will give a strict weak ordering to

3410

// everything but constants, and then we order by pointer address.

3411

return std::make_pair(getRank(A), A) > std::make_pair(getRank(B), B);

3412

}

3413

3414

class NewGVNLegacyPass : public FunctionPass {

3415

public:

3416

static char ID; // Pass identification, replacement for typeid.

3417

NewGVNLegacyPass() : FunctionPass(ID) {

3418

initializeNewGVNLegacyPassPass(*PassRegistry::getPassRegistry());

3419

}

3420

bool runOnFunction(Function &F) override;

3421

3422

private:

3423

void getAnalysisUsage(AnalysisUsage &AU) const override {

3424

AU.addRequired<AssumptionCacheTracker>();

3425

AU.addRequired<DominatorTreeWrapperPass>();

3426

AU.addRequired<TargetLibraryInfoWrapperPass>();

3427

AU.addRequired<MemorySSAWrapperPass>();

3428

AU.addRequired<AAResultsWrapperPass>();

3429

AU.addPreserved<DominatorTreeWrapperPass>();

3430

AU.addPreserved<GlobalsAAWrapperPass>();

3431

}

3432

};

3433

3434

bool NewGVNLegacyPass::runOnFunction(Function &F) {

3435

if (skipFunction(F))

3436

return false;

3437

return NewGVN(F, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),

3438

&getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),

3439

&getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),

3440

&getAnalysis<AAResultsWrapperPass>().getAAResults(),

3441

&getAnalysis<MemorySSAWrapperPass>().getMSSA(),

3442

F.getParent()->getDataLayout())

3443

.runGVN();

3444

}

3445

3446

INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering",static void *initializeNewGVNLegacyPassPassOnce(PassRegistry &
Registry) {

3447

false, false)static void *initializeNewGVNLegacyPassPassOnce(PassRegistry &
Registry) {

3448

INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);

3449

INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry);

3450

INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);

3451

INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);

3452

INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);

3453

INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry);

3454

INITIALIZE_PASS_END(NewGVNLegacyPass, "newgvn", "Global Value Numbering", false,PassInfo *PI = new PassInfo( "Global Value Numbering", "newgvn"
, &NewGVNLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<NewGVNLegacyPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeNewGVNLegacyPassPassFlag
; void llvm::initializeNewGVNLegacyPassPass(PassRegistry &
Registry) { llvm::call_once(InitializeNewGVNLegacyPassPassFlag
, initializeNewGVNLegacyPassPassOnce, std::ref(Registry)); }

3455

false)PassInfo *PI = new PassInfo( "Global Value Numbering", "newgvn"
, &NewGVNLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<NewGVNLegacyPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeNewGVNLegacyPassPassFlag
; void llvm::initializeNewGVNLegacyPassPass(PassRegistry &
Registry) { llvm::call_once(InitializeNewGVNLegacyPassPassFlag
, initializeNewGVNLegacyPassPassOnce, std::ref(Registry)); }

3456

3457

char NewGVNLegacyPass::ID = 0;

3458

3459

// createGVNPass - The public interface to this file.

3460

FunctionPass *llvm::createNewGVNPass() { return new NewGVNLegacyPass(); }

3461

3462

PreservedAnalyses NewGVNPass::run(Function &F, AnalysisManager<Function> &AM) {

3463

// Apparently the order in which we get these results matter for

3464

// the old GVN (see Chandler's comment in GVN.cpp). I'll keep

3465

// the same order here, just in case.

3466

auto &AC = AM.getResult<AssumptionAnalysis>(F);

3467

auto &DT = AM.getResult<DominatorTreeAnalysis>(F);

3468

auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);

3469

auto &AA = AM.getResult<AAManager>(F);

3470

auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();

3471

bool Changed =

3472

NewGVN(F, &DT, &AC, &TLI, &AA, &MSSA, F.getParent()->getDataLayout())

3473

.runGVN();

3474

if (!Changed)

3475

return PreservedAnalyses::all();

3476

PreservedAnalyses PA;

3477

PA.preserve<DominatorTreeAnalysis>();

3478

PA.preserve<GlobalsAA>();

3479

return PA;

3480

}