Bug Summary

File:lib/Transforms/Scalar/NewGVN.cpp
Warning:line 190, column 7
Forming reference to null pointer

Annotated Source Code

1//===---- NewGVN.cpp - Global Value Numbering Pass --------------*- C++ -*-===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9/// \file
10/// This file implements the new LLVM's Global Value Numbering pass.
11/// GVN partitions values computed by a function into congruence classes.
12/// Values ending up in the same congruence class are guaranteed to be the same
13/// for every execution of the program. In that respect, congruency is a
14/// compile-time approximation of equivalence of values at runtime.
15/// The algorithm implemented here uses a sparse formulation and it's based
16/// on the ideas described in the paper:
17/// "A Sparse Algorithm for Predicated Global Value Numbering" from
18/// Karthik Gargi.
19///
20/// A brief overview of the algorithm: The algorithm is essentially the same as
21/// the standard RPO value numbering algorithm (a good reference is the paper
22/// "SCC based value numbering" by L. Taylor Simpson) with one major difference:
23/// The RPO algorithm proceeds, on every iteration, to process every reachable
24/// block and every instruction in that block. This is because the standard RPO
25/// algorithm does not track what things have the same value number, it only
26/// tracks what the value number of a given operation is (the mapping is
27/// operation -> value number). Thus, when a value number of an operation
28/// changes, it must reprocess everything to ensure all uses of a value number
29/// get updated properly. In constrast, the sparse algorithm we use *also*
30/// tracks what operations have a given value number (IE it also tracks the
31/// reverse mapping from value number -> operations with that value number), so
32/// that it only needs to reprocess the instructions that are affected when
33/// something's value number changes. The rest of the algorithm is devoted to
34/// performing symbolic evaluation, forward propagation, and simplification of
35/// operations based on the value numbers deduced so far.
36///
37/// We also do not perform elimination by using any published algorithm. All
38/// published algorithms are O(Instructions). Instead, we use a technique that
39/// is O(number of operations with the same value number), enabling us to skip
40/// trying to eliminate things that have unique value numbers.
41//===----------------------------------------------------------------------===//
42
43#include "llvm/Transforms/Scalar/NewGVN.h"
44#include "llvm/ADT/BitVector.h"
45#include "llvm/ADT/DenseMap.h"
46#include "llvm/ADT/DenseSet.h"
47#include "llvm/ADT/DepthFirstIterator.h"
48#include "llvm/ADT/Hashing.h"
49#include "llvm/ADT/MapVector.h"
50#include "llvm/ADT/PostOrderIterator.h"
51#include "llvm/ADT/STLExtras.h"
52#include "llvm/ADT/SmallPtrSet.h"
53#include "llvm/ADT/SmallSet.h"
54#include "llvm/ADT/SparseBitVector.h"
55#include "llvm/ADT/Statistic.h"
56#include "llvm/ADT/TinyPtrVector.h"
57#include "llvm/Analysis/AliasAnalysis.h"
58#include "llvm/Analysis/AssumptionCache.h"
59#include "llvm/Analysis/CFG.h"
60#include "llvm/Analysis/CFGPrinter.h"
61#include "llvm/Analysis/ConstantFolding.h"
62#include "llvm/Analysis/GlobalsModRef.h"
63#include "llvm/Analysis/InstructionSimplify.h"
64#include "llvm/Analysis/MemoryBuiltins.h"
65#include "llvm/Analysis/MemoryLocation.h"
66#include "llvm/Analysis/TargetLibraryInfo.h"
67#include "llvm/IR/DataLayout.h"
68#include "llvm/IR/Dominators.h"
69#include "llvm/IR/GlobalVariable.h"
70#include "llvm/IR/IRBuilder.h"
71#include "llvm/IR/IntrinsicInst.h"
72#include "llvm/IR/LLVMContext.h"
73#include "llvm/IR/Metadata.h"
74#include "llvm/IR/PatternMatch.h"
75#include "llvm/IR/Type.h"
76#include "llvm/Support/Allocator.h"
77#include "llvm/Support/CommandLine.h"
78#include "llvm/Support/Debug.h"
79#include "llvm/Support/DebugCounter.h"
80#include "llvm/Transforms/Scalar.h"
81#include "llvm/Transforms/Scalar/GVNExpression.h"
82#include "llvm/Transforms/Utils/BasicBlockUtils.h"
83#include "llvm/Transforms/Utils/Local.h"
84#include "llvm/Transforms/Utils/MemorySSA.h"
85#include "llvm/Transforms/Utils/PredicateInfo.h"
86#include "llvm/Transforms/Utils/VNCoercion.h"
87#include <unordered_map>
88#include <utility>
89#include <vector>
90using namespace llvm;
91using namespace PatternMatch;
92using namespace llvm::GVNExpression;
93using namespace llvm::VNCoercion;
94#define DEBUG_TYPE"newgvn" "newgvn"
95
96STATISTIC(NumGVNInstrDeleted, "Number of instructions deleted")static llvm::Statistic NumGVNInstrDeleted = {"newgvn", "NumGVNInstrDeleted"
, "Number of instructions deleted", {0}, false};
97STATISTIC(NumGVNBlocksDeleted, "Number of blocks deleted")static llvm::Statistic NumGVNBlocksDeleted = {"newgvn", "NumGVNBlocksDeleted"
, "Number of blocks deleted", {0}, false};
98STATISTIC(NumGVNOpsSimplified, "Number of Expressions simplified")static llvm::Statistic NumGVNOpsSimplified = {"newgvn", "NumGVNOpsSimplified"
, "Number of Expressions simplified", {0}, false};
99STATISTIC(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
};
100STATISTIC(NumGVNMaxIterations,static llvm::Statistic NumGVNMaxIterations = {"newgvn", "NumGVNMaxIterations"
, "Maximum Number of iterations it took to converge GVN", {0}
, false}
101 "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};
102STATISTIC(NumGVNLeaderChanges, "Number of leader changes")static llvm::Statistic NumGVNLeaderChanges = {"newgvn", "NumGVNLeaderChanges"
, "Number of leader changes", {0}, false};
103STATISTIC(NumGVNSortedLeaderChanges, "Number of sorted leader changes")static llvm::Statistic NumGVNSortedLeaderChanges = {"newgvn",
"NumGVNSortedLeaderChanges", "Number of sorted leader changes"
, {0}, false};
104STATISTIC(NumGVNAvoidedSortedLeaderChanges,static llvm::Statistic NumGVNAvoidedSortedLeaderChanges = {"newgvn"
, "NumGVNAvoidedSortedLeaderChanges", "Number of avoided sorted leader changes"
, {0}, false}
105 "Number of avoided sorted leader changes")static llvm::Statistic NumGVNAvoidedSortedLeaderChanges = {"newgvn"
, "NumGVNAvoidedSortedLeaderChanges", "Number of avoided sorted leader changes"
, {0}, false};
106STATISTIC(NumGVNNotMostDominatingLeader,static llvm::Statistic NumGVNNotMostDominatingLeader = {"newgvn"
, "NumGVNNotMostDominatingLeader", "Number of times a member dominated it's new classes' leader"
, {0}, false}
107 "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};
108STATISTIC(NumGVNDeadStores, "Number of redundant/dead stores eliminated")static llvm::Statistic NumGVNDeadStores = {"newgvn", "NumGVNDeadStores"
, "Number of redundant/dead stores eliminated", {0}, false};
109DEBUG_COUNTER(VNCounter, "newgvn-vn",static const unsigned VNCounter = DebugCounter::registerCounter
("newgvn-vn", "Controls which instructions are value numbered"
);
110 "Controls which instructions are value numbered")static const unsigned VNCounter = DebugCounter::registerCounter
("newgvn-vn", "Controls which instructions are value numbered"
);
111//===----------------------------------------------------------------------===//
112// GVN Pass
113//===----------------------------------------------------------------------===//
114
115// Anchor methods.
116namespace llvm {
117namespace GVNExpression {
118Expression::~Expression() = default;
119BasicExpression::~BasicExpression() = default;
120CallExpression::~CallExpression() = default;
121LoadExpression::~LoadExpression() = default;
122StoreExpression::~StoreExpression() = default;
123AggregateValueExpression::~AggregateValueExpression() = default;
124PHIExpression::~PHIExpression() = default;
125}
126}
127
128// Congruence classes represent the set of expressions/instructions
129// that are all the same *during some scope in the function*.
130// That is, because of the way we perform equality propagation, and
131// because of memory value numbering, it is not correct to assume
132// you can willy-nilly replace any member with any other at any
133// point in the function.
134//
135// For any Value in the Member set, it is valid to replace any dominated member
136// with that Value.
137//
138// Every congruence class has a leader, and the leader is used to
139// symbolize instructions in a canonical way (IE every operand of an
140// instruction that is a member of the same congruence class will
141// always be replaced with leader during symbolization).
142// To simplify symbolization, we keep the leader as a constant if class can be
143// proved to be a constant value.
144// Otherwise, the leader is a randomly chosen member of the value set, it does
145// not matter which one is chosen.
146// Each congruence class also has a defining expression,
147// though the expression may be null. If it exists, it can be used for forward
148// propagation and reassociation of values.
149//
150struct CongruenceClass {
151 using MemberSet = SmallPtrSet<Value *, 4>;
152 unsigned ID;
153 // Representative leader.
154 Value *RepLeader = nullptr;
155 // If this is represented by a store, the value.
156 Value *RepStoredValue = nullptr;
157 // If this class contains MemoryDefs, what is the represented memory state.
158 MemoryAccess *RepMemoryAccess = nullptr;
159 // Defining Expression.
160 const Expression *DefiningExpr = nullptr;
161 // Actual members of this class.
162 MemberSet Members;
163
164 // True if this class has no members left. This is mainly used for assertion
165 // purposes, and for skipping empty classes.
166 bool Dead = false;
167
168 // Number of stores in this congruence class.
169 // This is used so we can detect store equivalence changes properly.
170 int StoreCount = 0;
171
172 // The most dominating leader after our current leader, because the member set
173 // is not sorted and is expensive to keep sorted all the time.
174 std::pair<Value *, unsigned int> NextLeader = {nullptr, ~0U};
175
176 explicit CongruenceClass(unsigned ID) : ID(ID) {}
177 CongruenceClass(unsigned ID, Value *Leader, const Expression *E)
178 : ID(ID), RepLeader(Leader), DefiningExpr(E) {}
179};
180
181// Return true if two congruence classes are equivalent to each other. This
182// means
183// that every field but the ID number and the dead field are equivalent.
184bool areClassesEquivalent(const CongruenceClass *A, const CongruenceClass *B) {
185 if (A == B)
7
←
Taking false branch
→
186 return true;
187 if ((A && !B) || (B && !A))
8
←
Assuming 'A' is null
→
9
←
Assuming 'B' is null
→
188 return false;
189
190 if (std::tie(A->StoreCount, A->RepLeader, A->RepStoredValue,
10
←
Forming reference to null pointer
191 A->RepMemoryAccess) != std::tie(B->StoreCount, B->RepLeader,
192 B->RepStoredValue,
193 B->RepMemoryAccess))
194 return false;
195 if (A->DefiningExpr != B->DefiningExpr)
196 if (!A->DefiningExpr || !B->DefiningExpr ||
197 *A->DefiningExpr != *B->DefiningExpr)
198 return false;
199 // We need some ordered set
200 std::set<Value *> AMembers(A->Members.begin(), A->Members.end());
201 std::set<Value *> BMembers(B->Members.begin(), B->Members.end());
202 return AMembers == BMembers;
203}
204
205namespace llvm {
206template <> struct DenseMapInfo<const Expression *> {
207 static const Expression *getEmptyKey() {
208 auto Val = static_cast<uintptr_t>(-1);
209 Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
210 return reinterpret_cast<const Expression *>(Val);
211 }
212 static const Expression *getTombstoneKey() {
213 auto Val = static_cast<uintptr_t>(~1U);
214 Val <<= PointerLikeTypeTraits<const Expression *>::NumLowBitsAvailable;
215 return reinterpret_cast<const Expression *>(Val);
216 }
217 static unsigned getHashValue(const Expression *V) {
218 return static_cast<unsigned>(V->getHashValue());
219 }
220 static bool isEqual(const Expression *LHS, const Expression *RHS) {
221 if (LHS == RHS)
222 return true;
223 if (LHS == getTombstoneKey() || RHS == getTombstoneKey() ||
224 LHS == getEmptyKey() || RHS == getEmptyKey())
225 return false;
226 return *LHS == *RHS;
227 }
228};
229} // end namespace llvm
230
231namespace {
232class NewGVN {
233 Function &F;
234 DominatorTree *DT;
235 AssumptionCache *AC;
236 const TargetLibraryInfo *TLI;
237 AliasAnalysis *AA;
238 MemorySSA *MSSA;
239 MemorySSAWalker *MSSAWalker;
240 const DataLayout &DL;
241 std::unique_ptr<PredicateInfo> PredInfo;
242 BumpPtrAllocator ExpressionAllocator;
243 ArrayRecycler<Value *> ArgRecycler;
244
245 // Number of function arguments, used by ranking
246 unsigned int NumFuncArgs;
247
248 // Congruence class info.
249
250 // This class is called INITIAL in the paper. It is the class everything
251 // startsout in, and represents any value. Being an optimistic analysis,
252 // anything in the TOP class has the value TOP, which is indeterminate and
253 // equivalent to everything.
254 CongruenceClass *TOPClass;
255 std::vector<CongruenceClass *> CongruenceClasses;
256 unsigned NextCongruenceNum;
257
258 // Value Mappings.
259 DenseMap<Value *, CongruenceClass *> ValueToClass;
260 DenseMap<Value *, const Expression *> ValueToExpression;
261
262 // Mapping from predicate info we used to the instructions we used it with.
263 // In order to correctly ensure propagation, we must keep track of what
264 // comparisons we used, so that when the values of the comparisons change, we
265 // propagate the information to the places we used the comparison.
266 DenseMap<const Value *, SmallPtrSet<Instruction *, 2>> PredicateToUsers;
267
268 // A table storing which memorydefs/phis represent a memory state provably
269 // equivalent to another memory state.
270 // We could use the congruence class machinery, but the MemoryAccess's are
271 // abstract memory states, so they can only ever be equivalent to each other,
272 // and not to constants, etc.
273 DenseMap<const MemoryAccess *, CongruenceClass *> MemoryAccessToClass;
274
275 // Expression to class mapping.
276 using ExpressionClassMap = DenseMap<const Expression *, CongruenceClass *>;
277 ExpressionClassMap ExpressionToClass;
278
279 // Which values have changed as a result of leader changes.
280 SmallPtrSet<Value *, 8> LeaderChanges;
281
282 // Reachability info.
283 using BlockEdge = BasicBlockEdge;
284 DenseSet<BlockEdge> ReachableEdges;
285 SmallPtrSet<const BasicBlock *, 8> ReachableBlocks;
286
287 // This is a bitvector because, on larger functions, we may have
288 // thousands of touched instructions at once (entire blocks,
289 // instructions with hundreds of uses, etc). Even with optimization
290 // for when we mark whole blocks as touched, when this was a
291 // SmallPtrSet or DenseSet, for some functions, we spent >20% of all
292 // the time in GVN just managing this list. The bitvector, on the
293 // other hand, efficiently supports test/set/clear of both
294 // individual and ranges, as well as "find next element" This
295 // enables us to use it as a worklist with essentially 0 cost.
296 BitVector TouchedInstructions;
297
298 DenseMap<const BasicBlock *, std::pair<unsigned, unsigned>> BlockInstRange;
299
300#ifndef NDEBUG
301 // Debugging for how many times each block and instruction got processed.
302 DenseMap<const Value *, unsigned> ProcessedCount;
303#endif
304
305 // DFS info.
306 // This contains a mapping from Instructions to DFS numbers.
307 // The numbering starts at 1. An instruction with DFS number zero
308 // means that the instruction is dead.
309 DenseMap<const Value *, unsigned> InstrDFS;
310
311 // This contains the mapping DFS numbers to instructions.
312 SmallVector<Value *, 32> DFSToInstr;
313
314 // Deletion info.
315 SmallPtrSet<Instruction *, 8> InstructionsToErase;
316
317public:
318 NewGVN(Function &F, DominatorTree *DT, AssumptionCache *AC,
319 TargetLibraryInfo *TLI, AliasAnalysis *AA, MemorySSA *MSSA,
320 const DataLayout &DL)
321 : F(F), DT(DT), AC(AC), TLI(TLI), AA(AA), MSSA(MSSA), DL(DL),
322 PredInfo(make_unique<PredicateInfo>(F, *DT, *AC)) {}
323 bool runGVN();
324
325private:
326 // Expression handling.
327 const Expression *createExpression(Instruction *);
328 const Expression *createBinaryExpression(unsigned, Type *, Value *, Value *);
329 PHIExpression *createPHIExpression(Instruction *);
330 const VariableExpression *createVariableExpression(Value *);
331 const ConstantExpression *createConstantExpression(Constant *);
332 const Expression *createVariableOrConstant(Value *V);
333 const UnknownExpression *createUnknownExpression(Instruction *);
334 const StoreExpression *createStoreExpression(StoreInst *, MemoryAccess *);
335 LoadExpression *createLoadExpression(Type *, Value *, LoadInst *,
336 MemoryAccess *);
337 const CallExpression *createCallExpression(CallInst *, MemoryAccess *);
338 const AggregateValueExpression *createAggregateValueExpression(Instruction *);
339 bool setBasicExpressionInfo(Instruction *, BasicExpression *);
340
341 // Congruence class handling.
342 CongruenceClass *createCongruenceClass(Value *Leader, const Expression *E) {
343 auto *result = new CongruenceClass(NextCongruenceNum++, Leader, E);
344 CongruenceClasses.emplace_back(result);
345 return result;
346 }
347
348 CongruenceClass *createSingletonCongruenceClass(Value *Member) {
349 CongruenceClass *CClass = createCongruenceClass(Member, nullptr);
350 CClass->Members.insert(Member);
351 ValueToClass[Member] = CClass;
352 return CClass;
353 }
354 void initializeCongruenceClasses(Function &F);
355
356 // Value number an Instruction or MemoryPhi.
357 void valueNumberMemoryPhi(MemoryPhi *);
358 void valueNumberInstruction(Instruction *);
359
360 // Symbolic evaluation.
361 const Expression *checkSimplificationResults(Expression *, Instruction *,
362 Value *);
363 const Expression *performSymbolicEvaluation(Value *);
364 const Expression *performSymbolicLoadCoercion(Type *, Value *, LoadInst *,
365 Instruction *, MemoryAccess *);
366 const Expression *performSymbolicLoadEvaluation(Instruction *);
367 const Expression *performSymbolicStoreEvaluation(Instruction *);
368 const Expression *performSymbolicCallEvaluation(Instruction *);
369 const Expression *performSymbolicPHIEvaluation(Instruction *);
370 const Expression *performSymbolicAggrValueEvaluation(Instruction *);
371 const Expression *performSymbolicCmpEvaluation(Instruction *);
372 const Expression *performSymbolicPredicateInfoEvaluation(Instruction *);
373
374 // Congruence finding.
375 bool someEquivalentDominates(const Instruction *, const Instruction *) const;
376 Value *lookupOperandLeader(Value *) const;
377 void performCongruenceFinding(Instruction *, const Expression *);
378 void moveValueToNewCongruenceClass(Instruction *, CongruenceClass *,
379 CongruenceClass *);
380 bool setMemoryAccessEquivTo(MemoryAccess *From, CongruenceClass *To);
381 MemoryAccess *lookupMemoryAccessEquiv(MemoryAccess *) const;
382 bool isMemoryAccessTop(const MemoryAccess *) const;
383 // Ranking
384 unsigned int getRank(const Value *) const;
385 bool shouldSwapOperands(const Value *, const Value *) const;
386
387 // Reachability handling.
388 void updateReachableEdge(BasicBlock *, BasicBlock *);
389 void processOutgoingEdges(TerminatorInst *, BasicBlock *);
390 Value *findConditionEquivalence(Value *) const;
391
392 // Elimination.
393 struct ValueDFS;
394 void convertClassToDFSOrdered(const CongruenceClass::MemberSet &,
395 SmallVectorImpl<ValueDFS> &,
396 DenseMap<const Value *, unsigned int> &,
397 SmallPtrSetImpl<Instruction *> &);
398 void convertClassToLoadsAndStores(const CongruenceClass::MemberSet &,
399 SmallVectorImpl<ValueDFS> &);
400
401 bool eliminateInstructions(Function &);
402 void replaceInstruction(Instruction *, Value *);
403 void markInstructionForDeletion(Instruction *);
404 void deleteInstructionsInBlock(BasicBlock *);
405
406 // New instruction creation.
407 void handleNewInstruction(Instruction *){};
408
409 // Various instruction touch utilities
410 void markUsersTouched(Value *);
411 void markMemoryUsersTouched(MemoryAccess *);
412 void markPredicateUsersTouched(Instruction *);
413 void markLeaderChangeTouched(CongruenceClass *CC);
414 void addPredicateUsers(const PredicateBase *, Instruction *);
415
416 // Main loop of value numbering
417 void iterateTouchedInstructions();
418
419 // Utilities.
420 void cleanupTables();
421 std::pair<unsigned, unsigned> assignDFSNumbers(BasicBlock *, unsigned);
422 void updateProcessedCount(Value *V);
423 void verifyMemoryCongruency() const;
424 void verifyIterationSettled(Function &F);
425 bool singleReachablePHIPath(const MemoryAccess *, const MemoryAccess *) const;
426 BasicBlock *getBlockForValue(Value *V) const;
427 void deleteExpression(const Expression *E);
428 // Debug counter info. When verifying, we have to reset the value numbering
429 // debug counter to the same state it started in to get the same results.
430 std::pair<int, int> StartingVNCounter;
431};
432} // end anonymous namespace
433
434template <typename T>
435static bool equalsLoadStoreHelper(const T &LHS, const Expression &RHS) {
436 if (!isa<LoadExpression>(RHS) && !isa<StoreExpression>(RHS))
437 return false;
438 return LHS.MemoryExpression::equals(RHS);
439}
440
441bool LoadExpression::equals(const Expression &Other) const {
442 return equalsLoadStoreHelper(*this, Other);
443}
444
445bool StoreExpression::equals(const Expression &Other) const {
446 if (!equalsLoadStoreHelper(*this, Other))
447 return false;
448 // Make sure that store vs store includes the value operand.
449 if (const auto *S = dyn_cast<StoreExpression>(&Other))
450 if (getStoredValue() != S->getStoredValue())
451 return false;
452 return true;
453}
454
455#ifndef NDEBUG
456static std::string getBlockName(const BasicBlock *B) {
457 return DOTGraphTraits<const Function *>::getSimpleNodeLabel(B, nullptr);
458}
459#endif
460
461// Get the basic block from an instruction/memory value.
462BasicBlock *NewGVN::getBlockForValue(Value *V) const {
463 if (auto *I = dyn_cast<Instruction>(V))
464 return I->getParent();
465 else if (auto *MP = dyn_cast<MemoryPhi>(V))
466 return MP->getBlock();
467 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 467);
468 return nullptr;
469}
470
471// Delete a definitely dead expression, so it can be reused by the expression
472// allocator. Some of these are not in creation functions, so we have to accept
473// const versions.
474void NewGVN::deleteExpression(const Expression *E) {
475 assert(isa<BasicExpression>(E))((isa<BasicExpression>(E)) ? static_cast<void> (0
) : __assert_fail ("isa<BasicExpression>(E)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 475, __PRETTY_FUNCTION__));
476 auto *BE = cast<BasicExpression>(E);
477 const_cast<BasicExpression *>(BE)->deallocateOperands(ArgRecycler);
478 ExpressionAllocator.Deallocate(E);
479}
480
481PHIExpression *NewGVN::createPHIExpression(Instruction *I) {
482 BasicBlock *PHIBlock = I->getParent();
483 auto *PN = cast<PHINode>(I);
484 auto *E =
485 new (ExpressionAllocator) PHIExpression(PN->getNumOperands(), PHIBlock);
486
487 E->allocateOperands(ArgRecycler, ExpressionAllocator);
488 E->setType(I->getType());
489 E->setOpcode(I->getOpcode());
490
491 // Filter out unreachable phi operands.
492 auto Filtered = make_filter_range(PN->operands(), [&](const Use &U) {
493 return ReachableEdges.count({PN->getIncomingBlock(U), PHIBlock});
494 });
495
496 std::transform(Filtered.begin(), Filtered.end(), op_inserter(E),
497 [&](const Use &U) -> Value * {
498 // Don't try to transform self-defined phis.
499 if (U == PN)
500 return PN;
501 return lookupOperandLeader(U);
502 });
503 return E;
504}
505
506// Set basic expression info (Arguments, type, opcode) for Expression
507// E from Instruction I in block B.
508bool NewGVN::setBasicExpressionInfo(Instruction *I, BasicExpression *E) {
509 bool AllConstant = true;
510 if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
511 E->setType(GEP->getSourceElementType());
512 else
513 E->setType(I->getType());
514 E->setOpcode(I->getOpcode());
515 E->allocateOperands(ArgRecycler, ExpressionAllocator);
516
517 // Transform the operand array into an operand leader array, and keep track of
518 // whether all members are constant.
519 std::transform(I->op_begin(), I->op_end(), op_inserter(E), [&](Value *O) {
520 auto Operand = lookupOperandLeader(O);
521 AllConstant &= isa<Constant>(Operand);
522 return Operand;
523 });
524
525 return AllConstant;
526}
527
528const Expression *NewGVN::createBinaryExpression(unsigned Opcode, Type *T,
529 Value *Arg1, Value *Arg2) {
530 auto *E = new (ExpressionAllocator) BasicExpression(2);
531
532 E->setType(T);
533 E->setOpcode(Opcode);
534 E->allocateOperands(ArgRecycler, ExpressionAllocator);
535 if (Instruction::isCommutative(Opcode)) {
536 // Ensure that commutative instructions that only differ by a permutation
537 // of their operands get the same value number by sorting the operand value
538 // numbers. Since all commutative instructions have two operands it is more
539 // efficient to sort by hand rather than using, say, std::sort.
540 if (shouldSwapOperands(Arg1, Arg2))
541 std::swap(Arg1, Arg2);
542 }
543 E->op_push_back(lookupOperandLeader(Arg1));
544 E->op_push_back(lookupOperandLeader(Arg2));
545
546 Value *V = SimplifyBinOp(Opcode, E->getOperand(0), E->getOperand(1), DL, TLI,
547 DT, AC);
548 if (const Expression *SimplifiedE = checkSimplificationResults(E, nullptr, V))
549 return SimplifiedE;
550 return E;
551}
552
553// Take a Value returned by simplification of Expression E/Instruction
554// I, and see if it resulted in a simpler expression. If so, return
555// that expression.
556// TODO: Once finished, this should not take an Instruction, we only
557// use it for printing.
558const Expression *NewGVN::checkSimplificationResults(Expression *E,
559 Instruction *I, Value *V) {
560 if (!V)
561 return nullptr;
562 if (auto *C = dyn_cast<Constant>(V)) {
563 if (I)
564 DEBUG(dbgs() << "Simplified " << *I << " to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " constant " << *C << "\n"; } } while
(false)
565 << " constant " << *C << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " constant " << *C << "\n"; } } while
(false);
566 NumGVNOpsSimplified++;
567 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 568, __PRETTY_FUNCTION__))
568 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 568, __PRETTY_FUNCTION__));
569 deleteExpression(E);
570 return createConstantExpression(C);
571 } else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
572 if (I)
573 DEBUG(dbgs() << "Simplified " << *I << " to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " variable " << *V << "\n"; } } while
(false)
574 << " variable " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " variable " << *V << "\n"; } } while
(false);
575 deleteExpression(E);
576 return createVariableExpression(V);
577 }
578
579 CongruenceClass *CC = ValueToClass.lookup(V);
580 if (CC && CC->DefiningExpr) {
581 if (I)
582 DEBUG(dbgs() << "Simplified " << *I << " to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " expression " << *V << "\n"; } }
while (false)
583 << " expression " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified " << *I <<
" to " << " expression " << *V << "\n"; } }
while (false);
584 NumGVNOpsSimplified++;
585 deleteExpression(E);
586 return CC->DefiningExpr;
587 }
588 return nullptr;
589}
590
591const Expression *NewGVN::createExpression(Instruction *I) {
592 auto *E = new (ExpressionAllocator) BasicExpression(I->getNumOperands());
593
594 bool AllConstant = setBasicExpressionInfo(I, E);
595
596 if (I->isCommutative()) {
597 // Ensure that commutative instructions that only differ by a permutation
598 // of their operands get the same value number by sorting the operand value
599 // numbers. Since all commutative instructions have two operands it is more
600 // efficient to sort by hand rather than using, say, std::sort.
601 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 601, __PRETTY_FUNCTION__));
602 if (shouldSwapOperands(E->getOperand(0), E->getOperand(1)))
603 E->swapOperands(0, 1);
604 }
605
606 // Perform simplificaiton
607 // TODO: Right now we only check to see if we get a constant result.
608 // We may get a less than constant, but still better, result for
609 // some operations.
610 // IE
611 // add 0, x -> x
612 // and x, x -> x
613 // We should handle this by simply rewriting the expression.
614 if (auto *CI = dyn_cast<CmpInst>(I)) {
615 // Sort the operand value numbers so x<y and y>x get the same value
616 // number.
617 CmpInst::Predicate Predicate = CI->getPredicate();
618 if (shouldSwapOperands(E->getOperand(0), E->getOperand(1))) {
619 E->swapOperands(0, 1);
620 Predicate = CmpInst::getSwappedPredicate(Predicate);
621 }
622 E->setOpcode((CI->getOpcode() << 8) | Predicate);
623 // TODO: 25% of our time is spent in SimplifyCmpInst with pointer operands
624 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 625, __PRETTY_FUNCTION__))
625 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 625, __PRETTY_FUNCTION__));
626 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 627, __PRETTY_FUNCTION__))
627 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 627, __PRETTY_FUNCTION__));
628 Value *V = SimplifyCmpInst(Predicate, E->getOperand(0), E->getOperand(1),
629 DL, TLI, DT, AC);
630 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
631 return SimplifiedE;
632 } else if (isa<SelectInst>(I)) {
633 if (isa<Constant>(E->getOperand(0)) ||
634 E->getOperand(0) == E->getOperand(1)) {
635 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 636, __PRETTY_FUNCTION__))
636 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 636, __PRETTY_FUNCTION__));
637 Value *V = SimplifySelectInst(E->getOperand(0), E->getOperand(1),
638 E->getOperand(2), DL, TLI, DT, AC);
639 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
640 return SimplifiedE;
641 }
642 } else if (I->isBinaryOp()) {
643 Value *V = SimplifyBinOp(E->getOpcode(), E->getOperand(0), E->getOperand(1),
644 DL, TLI, DT, AC);
645 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
646 return SimplifiedE;
647 } else if (auto *BI = dyn_cast<BitCastInst>(I)) {
648 Value *V = SimplifyInstruction(BI, DL, TLI, DT, AC);
649 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
650 return SimplifiedE;
651 } else if (isa<GetElementPtrInst>(I)) {
652 Value *V = SimplifyGEPInst(E->getType(),
653 ArrayRef<Value *>(E->op_begin(), E->op_end()),
654 DL, TLI, DT, AC);
655 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
656 return SimplifiedE;
657 } else if (AllConstant) {
658 // We don't bother trying to simplify unless all of the operands
659 // were constant.
660 // TODO: There are a lot of Simplify*'s we could call here, if we
661 // wanted to. The original motivating case for this code was a
662 // zext i1 false to i8, which we don't have an interface to
663 // simplify (IE there is no SimplifyZExt).
664
665 SmallVector<Constant *, 8> C;
666 for (Value *Arg : E->operands())
667 C.emplace_back(cast<Constant>(Arg));
668
669 if (Value *V = ConstantFoldInstOperands(I, C, DL, TLI))
670 if (const Expression *SimplifiedE = checkSimplificationResults(E, I, V))
671 return SimplifiedE;
672 }
673 return E;
674}
675
676const AggregateValueExpression *
677NewGVN::createAggregateValueExpression(Instruction *I) {
678 if (auto *II = dyn_cast<InsertValueInst>(I)) {
679 auto *E = new (ExpressionAllocator)
680 AggregateValueExpression(I->getNumOperands(), II->getNumIndices());
681 setBasicExpressionInfo(I, E);
682 E->allocateIntOperands(ExpressionAllocator);
683 std::copy(II->idx_begin(), II->idx_end(), int_op_inserter(E));
684 return E;
685 } else if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
686 auto *E = new (ExpressionAllocator)
687 AggregateValueExpression(I->getNumOperands(), EI->getNumIndices());
688 setBasicExpressionInfo(EI, E);
689 E->allocateIntOperands(ExpressionAllocator);
690 std::copy(EI->idx_begin(), EI->idx_end(), int_op_inserter(E));
691 return E;
692 }
693 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 693);
694}
695
696const VariableExpression *NewGVN::createVariableExpression(Value *V) {
697 auto *E = new (ExpressionAllocator) VariableExpression(V);
698 E->setOpcode(V->getValueID());
699 return E;
700}
701
702const Expression *NewGVN::createVariableOrConstant(Value *V) {
703 if (auto *C = dyn_cast<Constant>(V))
704 return createConstantExpression(C);
705 return createVariableExpression(V);
706}
707
708const ConstantExpression *NewGVN::createConstantExpression(Constant *C) {
709 auto *E = new (ExpressionAllocator) ConstantExpression(C);
710 E->setOpcode(C->getValueID());
711 return E;
712}
713
714const UnknownExpression *NewGVN::createUnknownExpression(Instruction *I) {
715 auto *E = new (ExpressionAllocator) UnknownExpression(I);
716 E->setOpcode(I->getOpcode());
717 return E;
718}
719
720const CallExpression *NewGVN::createCallExpression(CallInst *CI,
721 MemoryAccess *HV) {
722 // FIXME: Add operand bundles for calls.
723 auto *E =
724 new (ExpressionAllocator) CallExpression(CI->getNumOperands(), CI, HV);
725 setBasicExpressionInfo(CI, E);
726 return E;
727}
728
729// Return true if some equivalent of instruction Inst dominates instruction U.
730bool NewGVN::someEquivalentDominates(const Instruction *Inst,
731 const Instruction *U) const {
732 auto *CC = ValueToClass.lookup(Inst);
733 // This must be an instruction because we are only called from phi nodes
734 // in the case that the value it needs to check against is an instruction.
735
736 // The most likely candiates for dominance are the leader and the next leader.
737 // The leader or nextleader will dominate in all cases where there is an
738 // equivalent that is higher up in the dom tree.
739 // We can't *only* check them, however, because the
740 // dominator tree could have an infinite number of non-dominating siblings
741 // with instructions that are in the right congruence class.
742 // A
743 // B C D E F G
744 // |
745 // H
746 // Instruction U could be in H, with equivalents in every other sibling.
747 // Depending on the rpo order picked, the leader could be the equivalent in
748 // any of these siblings.
749 if (!CC)
750 return false;
751 if (DT->dominates(cast<Instruction>(CC->RepLeader), U))
752 return true;
753 if (CC->NextLeader.first &&
754 DT->dominates(cast<Instruction>(CC->NextLeader.first), U))
755 return true;
756 return llvm::any_of(CC->Members, [&](const Value *Member) {
757 return Member != CC->RepLeader &&
758 DT->dominates(cast<Instruction>(Member), U);
759 });
760}
761
762// See if we have a congruence class and leader for this operand, and if so,
763// return it. Otherwise, return the operand itself.
764Value *NewGVN::lookupOperandLeader(Value *V) const {
765 CongruenceClass *CC = ValueToClass.lookup(V);
766 if (CC) {
767 // Everything in TOP is represneted by undef, as it can be any value.
768 // We do have to make sure we get the type right though, so we can't set the
769 // RepLeader to undef.
770 if (CC == TOPClass)
771 return UndefValue::get(V->getType());
772 return CC->RepStoredValue ? CC->RepStoredValue : CC->RepLeader;
773 }
774
775 return V;
776}
777
778MemoryAccess *NewGVN::lookupMemoryAccessEquiv(MemoryAccess *MA) const {
779 auto *CC = MemoryAccessToClass.lookup(MA);
780 if (CC && CC->RepMemoryAccess)
781 return CC->RepMemoryAccess;
782 // FIXME: We need to audit all the places that current set a nullptr To, and
783 // fix them. There should always be *some* congruence class, even if it is
784 // singular. Right now, we don't bother setting congruence classes for
785 // anything but stores, which means we have to return the original access
786 // here. Otherwise, this should be unreachable.
787 return MA;
788}
789
790// Return true if the MemoryAccess is really equivalent to everything. This is
791// equivalent to the lattice value "TOP" in most lattices. This is the initial
792// state of all memory accesses.
793bool NewGVN::isMemoryAccessTop(const MemoryAccess *MA) const {
794 return MemoryAccessToClass.lookup(MA) == TOPClass;
795}
796
797LoadExpression *NewGVN::createLoadExpression(Type *LoadType, Value *PointerOp,
798 LoadInst *LI, MemoryAccess *DA) {
799 auto *E = new (ExpressionAllocator) LoadExpression(1, LI, DA);
800 E->allocateOperands(ArgRecycler, ExpressionAllocator);
801 E->setType(LoadType);
802
803 // Give store and loads same opcode so they value number together.
804 E->setOpcode(0);
805 E->op_push_back(lookupOperandLeader(PointerOp));
806 if (LI)
807 E->setAlignment(LI->getAlignment());
808
809 // TODO: Value number heap versions. We may be able to discover
810 // things alias analysis can't on it's own (IE that a store and a
811 // load have the same value, and thus, it isn't clobbering the load).
812 return E;
813}
814
815const StoreExpression *NewGVN::createStoreExpression(StoreInst *SI,
816 MemoryAccess *DA) {
817 auto *StoredValueLeader = lookupOperandLeader(SI->getValueOperand());
818 auto *E = new (ExpressionAllocator)
819 StoreExpression(SI->getNumOperands(), SI, StoredValueLeader, DA);
820 E->allocateOperands(ArgRecycler, ExpressionAllocator);
821 E->setType(SI->getValueOperand()->getType());
822
823 // Give store and loads same opcode so they value number together.
824 E->setOpcode(0);
825 E->op_push_back(lookupOperandLeader(SI->getPointerOperand()));
826
827 // TODO: Value number heap versions. We may be able to discover
828 // things alias analysis can't on it's own (IE that a store and a
829 // load have the same value, and thus, it isn't clobbering the load).
830 return E;
831}
832
833const Expression *NewGVN::performSymbolicStoreEvaluation(Instruction *I) {
834 // Unlike loads, we never try to eliminate stores, so we do not check if they
835 // are simple and avoid value numbering them.
836 auto *SI = cast<StoreInst>(I);
837 MemoryAccess *StoreAccess = MSSA->getMemoryAccess(SI);
838 // Get the expression, if any, for the RHS of the MemoryDef.
839 MemoryAccess *StoreRHS = lookupMemoryAccessEquiv(
840 cast<MemoryDef>(StoreAccess)->getDefiningAccess());
841 // If we are defined by ourselves, use the live on entry def.
842 if (StoreRHS == StoreAccess)
843 StoreRHS = MSSA->getLiveOnEntryDef();
844
845 if (SI->isSimple()) {
846 // See if we are defined by a previous store expression, it already has a
847 // value, and it's the same value as our current store. FIXME: Right now, we
848 // only do this for simple stores, we should expand to cover memcpys, etc.
849 const Expression *OldStore = createStoreExpression(SI, StoreRHS);
850 CongruenceClass *CC = ExpressionToClass.lookup(OldStore);
851 // Basically, check if the congruence class the store is in is defined by a
852 // store that isn't us, and has the same value. MemorySSA takes care of
853 // ensuring the store has the same memory state as us already.
854 // The RepStoredValue gets nulled if all the stores disappear in a class, so
855 // we don't need to check if the class contains a store besides us.
856 if (CC && CC->RepStoredValue == lookupOperandLeader(SI->getValueOperand()))
857 return OldStore;
858 deleteExpression(OldStore);
859 // Also check if our value operand is defined by a load of the same memory
860 // location, and the memory state is the same as it was then
861 // (otherwise, it could have been overwritten later. See test32 in
862 // transforms/DeadStoreElimination/simple.ll)
863 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand())) {
864 if ((lookupOperandLeader(LI->getPointerOperand()) ==
865 lookupOperandLeader(SI->getPointerOperand())) &&
866 (lookupMemoryAccessEquiv(
867 MSSA->getMemoryAccess(LI)->getDefiningAccess()) == StoreRHS))
868 return createVariableExpression(LI);
869 }
870 }
871 return createStoreExpression(SI, StoreAccess);
872}
873
874// See if we can extract the value of a loaded pointer from a load, a store, or
875// a memory instruction.
876const Expression *
877NewGVN::performSymbolicLoadCoercion(Type *LoadType, Value *LoadPtr,
878 LoadInst *LI, Instruction *DepInst,
879 MemoryAccess *DefiningAccess) {
880 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 880, __PRETTY_FUNCTION__));
881 if (auto *DepSI = dyn_cast<StoreInst>(DepInst)) {
882 // Can't forward from non-atomic to atomic without violating memory model.
883 // Also don't need to coerce if they are the same type, we will just
884 // propogate..
885 if (LI->isAtomic() > DepSI->isAtomic() ||
886 LoadType == DepSI->getValueOperand()->getType())
887 return nullptr;
888 int Offset = analyzeLoadFromClobberingStore(LoadType, LoadPtr, DepSI, DL);
889 if (Offset >= 0) {
890 if (auto *C = dyn_cast<Constant>(
891 lookupOperandLeader(DepSI->getValueOperand()))) {
892 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)
893 << *C << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Coercing load from store " <<
*DepSI << " to constant " << *C << "\n"; }
} while (false);
894 return createConstantExpression(
895 getConstantStoreValueForLoad(C, Offset, LoadType, DL));
896 }
897 }
898
899 } else if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
900 // Can't forward from non-atomic to atomic without violating memory model.
901 if (LI->isAtomic() > DepLI->isAtomic())
902 return nullptr;
903 int Offset = analyzeLoadFromClobberingLoad(LoadType, LoadPtr, DepLI, DL);
904 if (Offset >= 0) {
905 // We can coerce a constant load into a load
906 if (auto *C = dyn_cast<Constant>(lookupOperandLeader(DepLI)))
907 if (auto *PossibleConstant =
908 getConstantLoadValueForLoad(C, Offset, LoadType, DL)) {
909 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)
910 << *PossibleConstant << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Coercing load from load " <<
*LI << " to constant " << *PossibleConstant <<
"\n"; } } while (false);
911 return createConstantExpression(PossibleConstant);
912 }
913 }
914
915 } else if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
916 int Offset = analyzeLoadFromClobberingMemInst(LoadType, LoadPtr, DepMI, DL);
917 if (Offset >= 0) {
918 if (auto *PossibleConstant =
919 getConstantMemInstValueForLoad(DepMI, Offset, LoadType, DL)) {
920 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)
921 << " to constant " << *PossibleConstant << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Coercing load from meminst " <<
*DepMI << " to constant " << *PossibleConstant <<
"\n"; } } while (false);
922 return createConstantExpression(PossibleConstant);
923 }
924 }
925 }
926
927 // All of the below are only true if the loaded pointer is produced
928 // by the dependent instruction.
929 if (LoadPtr != lookupOperandLeader(DepInst) &&
930 !AA->isMustAlias(LoadPtr, DepInst))
931 return nullptr;
932 // If this load really doesn't depend on anything, then we must be loading an
933 // undef value. This can happen when loading for a fresh allocation with no
934 // intervening stores, for example. Note that this is only true in the case
935 // that the result of the allocation is pointer equal to the load ptr.
936 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
937 return createConstantExpression(UndefValue::get(LoadType));
938 }
939 // If this load occurs either right after a lifetime begin,
940 // then the loaded value is undefined.
941 else if (auto *II = dyn_cast<IntrinsicInst>(DepInst)) {
942 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
943 return createConstantExpression(UndefValue::get(LoadType));
944 }
945 // If this load follows a calloc (which zero initializes memory),
946 // then the loaded value is zero
947 else if (isCallocLikeFn(DepInst, TLI)) {
948 return createConstantExpression(Constant::getNullValue(LoadType));
949 }
950
951 return nullptr;
952}
953
954const Expression *NewGVN::performSymbolicLoadEvaluation(Instruction *I) {
955 auto *LI = cast<LoadInst>(I);
956
957 // We can eliminate in favor of non-simple loads, but we won't be able to
958 // eliminate the loads themselves.
959 if (!LI->isSimple())
960 return nullptr;
961
962 Value *LoadAddressLeader = lookupOperandLeader(LI->getPointerOperand());
963 // Load of undef is undef.
964 if (isa<UndefValue>(LoadAddressLeader))
965 return createConstantExpression(UndefValue::get(LI->getType()));
966
967 MemoryAccess *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(I);
968
969 if (!MSSA->isLiveOnEntryDef(DefiningAccess)) {
970 if (auto *MD = dyn_cast<MemoryDef>(DefiningAccess)) {
971 Instruction *DefiningInst = MD->getMemoryInst();
972 // If the defining instruction is not reachable, replace with undef.
973 if (!ReachableBlocks.count(DefiningInst->getParent()))
974 return createConstantExpression(UndefValue::get(LI->getType()));
975 // This will handle stores and memory insts. We only do if it the
976 // defining access has a different type, or it is a pointer produced by
977 // certain memory operations that cause the memory to have a fixed value
978 // (IE things like calloc).
979 const Expression *CoercionResult = performSymbolicLoadCoercion(
980 LI->getType(), LoadAddressLeader, LI, DefiningInst, DefiningAccess);
981 if (CoercionResult)
982 return CoercionResult;
983 }
984 }
985
986 const Expression *E =
987 createLoadExpression(LI->getType(), LoadAddressLeader, LI,
988 lookupMemoryAccessEquiv(DefiningAccess));
989 return E;
990}
991
992const Expression *
993NewGVN::performSymbolicPredicateInfoEvaluation(Instruction *I) {
994 auto *PI = PredInfo->getPredicateInfoFor(I);
995 if (!PI)
996 return nullptr;
997
998 DEBUG(dbgs() << "Found predicate info from instruction !\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found predicate info from instruction !\n"
; } } while (false);
999
1000 auto *PWC = dyn_cast<PredicateWithCondition>(PI);
1001 if (!PWC)
1002 return nullptr;
1003
1004 auto *CopyOf = I->getOperand(0);
1005 auto *Cond = PWC->Condition;
1006
1007 // If this a copy of the condition, it must be either true or false depending
1008 // on the predicate info type and edge
1009 if (CopyOf == Cond) {
1010 // We should not need to add predicate users because the predicate info is
1011 // already a use of this operand.
1012 if (isa<PredicateAssume>(PI))
1013 return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1014 if (auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1015 if (PBranch->TrueEdge)
1016 return createConstantExpression(ConstantInt::getTrue(Cond->getType()));
1017 return createConstantExpression(ConstantInt::getFalse(Cond->getType()));
1018 }
1019 if (auto *PSwitch = dyn_cast<PredicateSwitch>(PI))
1020 return createConstantExpression(cast<Constant>(PSwitch->CaseValue));
1021 }
1022
1023 // Not a copy of the condition, so see what the predicates tell us about this
1024 // value. First, though, we check to make sure the value is actually a copy
1025 // of one of the condition operands. It's possible, in certain cases, for it
1026 // to be a copy of a predicateinfo copy. In particular, if two branch
1027 // operations use the same condition, and one branch dominates the other, we
1028 // will end up with a copy of a copy. This is currently a small deficiency in
1029 // predicateinfo. What will end up happening here is that we will value
1030 // number both copies the same anyway.
1031
1032 // Everything below relies on the condition being a comparison.
1033 auto *Cmp = dyn_cast<CmpInst>(Cond);
1034 if (!Cmp)
1035 return nullptr;
1036
1037 if (CopyOf != Cmp->getOperand(0) && CopyOf != Cmp->getOperand(1)) {
1038 DEBUG(dbgs() << "Copy is not of any condition operands!")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Copy is not of any condition operands!"
; } } while (false);
1039 return nullptr;
1040 }
1041 Value *FirstOp = lookupOperandLeader(Cmp->getOperand(0));
1042 Value *SecondOp = lookupOperandLeader(Cmp->getOperand(1));
1043 bool SwappedOps = false;
1044 // Sort the ops
1045 if (shouldSwapOperands(FirstOp, SecondOp)) {
1046 std::swap(FirstOp, SecondOp);
1047 SwappedOps = true;
1048 }
1049 CmpInst::Predicate Predicate =
1050 SwappedOps ? Cmp->getSwappedPredicate() : Cmp->getPredicate();
1051
1052 if (isa<PredicateAssume>(PI)) {
1053 // If the comparison is true when the operands are equal, then we know the
1054 // operands are equal, because assumes must always be true.
1055 if (CmpInst::isTrueWhenEqual(Predicate)) {
1056 addPredicateUsers(PI, I);
1057 return createVariableOrConstant(FirstOp);
1058 }
1059 }
1060 if (const auto *PBranch = dyn_cast<PredicateBranch>(PI)) {
1061 // If we are *not* a copy of the comparison, we may equal to the other
1062 // operand when the predicate implies something about equality of
1063 // operations. In particular, if the comparison is true/false when the
1064 // operands are equal, and we are on the right edge, we know this operation
1065 // is equal to something.
1066 if ((PBranch->TrueEdge && Predicate == CmpInst::ICMP_EQ) ||
1067 (!PBranch->TrueEdge && Predicate == CmpInst::ICMP_NE)) {
1068 addPredicateUsers(PI, I);
1069 return createVariableOrConstant(FirstOp);
1070 }
1071 // Handle the special case of floating point.
1072 if (((PBranch->TrueEdge && Predicate == CmpInst::FCMP_OEQ) ||
1073 (!PBranch->TrueEdge && Predicate == CmpInst::FCMP_UNE)) &&
1074 isa<ConstantFP>(FirstOp) && !cast<ConstantFP>(FirstOp)->isZero()) {
1075 addPredicateUsers(PI, I);
1076 return createConstantExpression(cast<Constant>(FirstOp));
1077 }
1078 }
1079 return nullptr;
1080}
1081
1082// Evaluate read only and pure calls, and create an expression result.
1083const Expression *NewGVN::performSymbolicCallEvaluation(Instruction *I) {
1084 auto *CI = cast<CallInst>(I);
1085 if (auto *II = dyn_cast<IntrinsicInst>(I)) {
1086 // Instrinsics with the returned attribute are copies of arguments.
1087 if (auto *ReturnedValue = II->getReturnedArgOperand()) {
1088 if (II->getIntrinsicID() == Intrinsic::ssa_copy)
1089 if (const auto *Result = performSymbolicPredicateInfoEvaluation(I))
1090 return Result;
1091 return createVariableOrConstant(ReturnedValue);
1092 }
1093 }
1094 if (AA->doesNotAccessMemory(CI)) {
1095 return createCallExpression(CI, nullptr);
1096 } else if (AA->onlyReadsMemory(CI)) {
1097 MemoryAccess *DefiningAccess = MSSAWalker->getClobberingMemoryAccess(CI);
1098 return createCallExpression(CI, lookupMemoryAccessEquiv(DefiningAccess));
1099 }
1100 return nullptr;
1101}
1102
1103// Update the memory access equivalence table to say that From is equal to To,
1104// and return true if this is different from what already existed in the table.
1105// FIXME: We need to audit all the places that current set a nullptr To, and fix
1106// them. There should always be *some* congruence class, even if it is singular.
1107bool NewGVN::setMemoryAccessEquivTo(MemoryAccess *From, CongruenceClass *To) {
1108 DEBUG(dbgs() << "Setting " << *From)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Setting " << *From; } } while
(false);
1109 if (To) {
1110 DEBUG(dbgs() << " equivalent to congruence class ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << " equivalent to congruence class "
; } } while (false);
1111 DEBUG(dbgs() << To->ID << " with current memory access leader ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << To->ID << " with current memory access leader "
; } } while (false);
1112 DEBUG(dbgs() << *To->RepMemoryAccess)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << *To->RepMemoryAccess; } } while
(false);
1113 } else {
1114 DEBUG(dbgs() << " equivalent to itself")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << " equivalent to itself"; } } while
(false);
1115 }
1116 DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "\n"; } } while (false);
1117
1118 auto LookupResult = MemoryAccessToClass.find(From);
1119 bool Changed = false;
1120 // If it's already in the table, see if the value changed.
1121 if (LookupResult != MemoryAccessToClass.end()) {
1122 if (To && LookupResult->second != To) {
1123 // It wasn't equivalent before, and now it is.
1124 LookupResult->second = To;
1125 Changed = true;
1126 } else if (!To) {
1127 // It used to be equivalent to something, and now it's not.
1128 MemoryAccessToClass.erase(LookupResult);
1129 Changed = true;
1130 }
1131 } else {
1132 assert(!To &&((!To && "Memory equivalence should never change from nothing to something"
) ? static_cast<void> (0) : __assert_fail ("!To && \"Memory equivalence should never change from nothing to something\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1133, __PRETTY_FUNCTION__))
1133 "Memory equivalence should never change from nothing to something")((!To && "Memory equivalence should never change from nothing to something"
) ? static_cast<void> (0) : __assert_fail ("!To && \"Memory equivalence should never change from nothing to something\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1133, __PRETTY_FUNCTION__));
1134 }
1135
1136 return Changed;
1137}
1138
1139// Evaluate PHI nodes symbolically, and create an expression result.
1140const Expression *NewGVN::performSymbolicPHIEvaluation(Instruction *I) {
1141 auto *E = cast<PHIExpression>(createPHIExpression(I));
1142 // We match the semantics of SimplifyPhiNode from InstructionSimplify here.
1143
1144 // See if all arguaments are the same.
1145 // We track if any were undef because they need special handling.
1146 bool HasUndef = false;
1147 auto Filtered = make_filter_range(E->operands(), [&](const Value *Arg) {
1148 if (Arg == I)
1149 return false;
1150 if (isa<UndefValue>(Arg)) {
1151 HasUndef = true;
1152 return false;
1153 }
1154 return true;
1155 });
1156 // If we are left with no operands, it's undef
1157 if (Filtered.begin() == Filtered.end()) {
1158 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)
1159 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified PHI node " <<
*I << " to undef" << "\n"; } } while (false);
1160 deleteExpression(E);
1161 return createConstantExpression(UndefValue::get(I->getType()));
1162 }
1163 Value *AllSameValue = *(Filtered.begin());
1164 ++Filtered.begin();
1165 // Can't use std::equal here, sadly, because filter.begin moves.
1166 if (llvm::all_of(Filtered, [AllSameValue](const Value *V) {
1167 return V == AllSameValue;
1168 })) {
1169 // In LLVM's non-standard representation of phi nodes, it's possible to have
1170 // phi nodes with cycles (IE dependent on other phis that are .... dependent
1171 // on the original phi node), especially in weird CFG's where some arguments
1172 // are unreachable, or uninitialized along certain paths. This can cause
1173 // infinite loops during evaluation. We work around this by not trying to
1174 // really evaluate them independently, but instead using a variable
1175 // expression to say if one is equivalent to the other.
1176 // We also special case undef, so that if we have an undef, we can't use the
1177 // common value unless it dominates the phi block.
1178 if (HasUndef) {
1179 // Only have to check for instructions
1180 if (auto *AllSameInst = dyn_cast<Instruction>(AllSameValue))
1181 if (!someEquivalentDominates(AllSameInst, I))
1182 return E;
1183 }
1184
1185 NumGVNPhisAllSame++;
1186 DEBUG(dbgs() << "Simplified PHI node " << *I << " to " << *AllSameValuedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified PHI node " <<
*I << " to " << *AllSameValue << "\n"; } }
while (false)
1187 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Simplified PHI node " <<
*I << " to " << *AllSameValue << "\n"; } }
while (false);
1188 deleteExpression(E);
1189 return createVariableOrConstant(AllSameValue);
1190 }
1191 return E;
1192}
1193
1194const Expression *NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) {
1195 if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
1196 auto *II = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
1197 if (II && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
1198 unsigned Opcode = 0;
1199 // EI might be an extract from one of our recognised intrinsics. If it
1200 // is we'll synthesize a semantically equivalent expression instead on
1201 // an extract value expression.
1202 switch (II->getIntrinsicID()) {
1203 case Intrinsic::sadd_with_overflow:
1204 case Intrinsic::uadd_with_overflow:
1205 Opcode = Instruction::Add;
1206 break;
1207 case Intrinsic::ssub_with_overflow:
1208 case Intrinsic::usub_with_overflow:
1209 Opcode = Instruction::Sub;
1210 break;
1211 case Intrinsic::smul_with_overflow:
1212 case Intrinsic::umul_with_overflow:
1213 Opcode = Instruction::Mul;
1214 break;
1215 default:
1216 break;
1217 }
1218
1219 if (Opcode != 0) {
1220 // Intrinsic recognized. Grab its args to finish building the
1221 // expression.
1222 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1223, __PRETTY_FUNCTION__))
1223 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1223, __PRETTY_FUNCTION__));
1224 return createBinaryExpression(
1225 Opcode, EI->getType(), II->getArgOperand(0), II->getArgOperand(1));
1226 }
1227 }
1228 }
1229
1230 return createAggregateValueExpression(I);
1231}
1232const Expression *NewGVN::performSymbolicCmpEvaluation(Instruction *I) {
1233 auto *CI = dyn_cast<CmpInst>(I);
1234 // See if our operands are equal to those of a previous predicate, and if so,
1235 // if it implies true or false.
1236 auto Op0 = lookupOperandLeader(CI->getOperand(0));
1237 auto Op1 = lookupOperandLeader(CI->getOperand(1));
1238 auto OurPredicate = CI->getPredicate();
1239 if (shouldSwapOperands(Op0, Op1)) {
1240 std::swap(Op0, Op1);
1241 OurPredicate = CI->getSwappedPredicate();
1242 }
1243
1244 // Avoid processing the same info twice
1245 const PredicateBase *LastPredInfo = nullptr;
1246 // See if we know something about the comparison itself, like it is the target
1247 // of an assume.
1248 auto *CmpPI = PredInfo->getPredicateInfoFor(I);
1249 if (dyn_cast_or_null<PredicateAssume>(CmpPI))
1250 return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1251
1252 if (Op0 == Op1) {
1253 // This condition does not depend on predicates, no need to add users
1254 if (CI->isTrueWhenEqual())
1255 return createConstantExpression(ConstantInt::getTrue(CI->getType()));
1256 else if (CI->isFalseWhenEqual())
1257 return createConstantExpression(ConstantInt::getFalse(CI->getType()));
1258 }
1259
1260 // NOTE: Because we are comparing both operands here and below, and using
1261 // previous comparisons, we rely on fact that predicateinfo knows to mark
1262 // comparisons that use renamed operands as users of the earlier comparisons.
1263 // It is *not* enough to just mark predicateinfo renamed operands as users of
1264 // the earlier comparisons, because the *other* operand may have changed in a
1265 // previous iteration.
1266 // Example:
1267 // icmp slt %a, %b
1268 // %b.0 = ssa.copy(%b)
1269 // false branch:
1270 // icmp slt %c, %b.0
1271
1272 // %c and %a may start out equal, and thus, the code below will say the second
1273 // %icmp is false. c may become equal to something else, and in that case the
1274 // %second icmp *must* be reexamined, but would not if only the renamed
1275 // %operands are considered users of the icmp.
1276
1277 // *Currently* we only check one level of comparisons back, and only mark one
1278 // level back as touched when changes appen . If you modify this code to look
1279 // back farther through comparisons, you *must* mark the appropriate
1280 // comparisons as users in PredicateInfo.cpp, or you will cause bugs. See if
1281 // we know something just from the operands themselves
1282
1283 // See if our operands have predicate info, so that we may be able to derive
1284 // something from a previous comparison.
1285 for (const auto &Op : CI->operands()) {
1286 auto *PI = PredInfo->getPredicateInfoFor(Op);
1287 if (const auto *PBranch = dyn_cast_or_null<PredicateBranch>(PI)) {
1288 if (PI == LastPredInfo)
1289 continue;
1290 LastPredInfo = PI;
1291
1292 // TODO: Along the false edge, we may know more things too, like icmp of
1293 // same operands is false.
1294 // TODO: We only handle actual comparison conditions below, not and/or.
1295 auto *BranchCond = dyn_cast<CmpInst>(PBranch->Condition);
1296 if (!BranchCond)
1297 continue;
1298 auto *BranchOp0 = lookupOperandLeader(BranchCond->getOperand(0));
1299 auto *BranchOp1 = lookupOperandLeader(BranchCond->getOperand(1));
1300 auto BranchPredicate = BranchCond->getPredicate();
1301 if (shouldSwapOperands(BranchOp0, BranchOp1)) {
1302 std::swap(BranchOp0, BranchOp1);
1303 BranchPredicate = BranchCond->getSwappedPredicate();
1304 }
1305 if (BranchOp0 == Op0 && BranchOp1 == Op1) {
1306 if (PBranch->TrueEdge) {
1307 // If we know the previous predicate is true and we are in the true
1308 // edge then we may be implied true or false.
1309 if (CmpInst::isImpliedTrueByMatchingCmp(OurPredicate,
1310 BranchPredicate)) {
1311 addPredicateUsers(PI, I);
1312 return createConstantExpression(
1313 ConstantInt::getTrue(CI->getType()));
1314 }
1315
1316 if (CmpInst::isImpliedFalseByMatchingCmp(OurPredicate,
1317 BranchPredicate)) {
1318 addPredicateUsers(PI, I);
1319 return createConstantExpression(
1320 ConstantInt::getFalse(CI->getType()));
1321 }
1322
1323 } else {
1324 // Just handle the ne and eq cases, where if we have the same
1325 // operands, we may know something.
1326 if (BranchPredicate == OurPredicate) {
1327 addPredicateUsers(PI, I);
1328 // Same predicate, same ops,we know it was false, so this is false.
1329 return createConstantExpression(
1330 ConstantInt::getFalse(CI->getType()));
1331 } else if (BranchPredicate ==
1332 CmpInst::getInversePredicate(OurPredicate)) {
1333 addPredicateUsers(PI, I);
1334 // Inverse predicate, we know the other was false, so this is true.
1335 // FIXME: Double check this
1336 return createConstantExpression(
1337 ConstantInt::getTrue(CI->getType()));
1338 }
1339 }
1340 }
1341 }
1342 }
1343 // Create expression will take care of simplifyCmpInst
1344 return createExpression(I);
1345}
1346
1347// Substitute and symbolize the value before value numbering.
1348const Expression *NewGVN::performSymbolicEvaluation(Value *V) {
1349 const Expression *E = nullptr;
1350 if (auto *C = dyn_cast<Constant>(V))
1351 E = createConstantExpression(C);
1352 else if (isa<Argument>(V) || isa<GlobalVariable>(V)) {
1353 E = createVariableExpression(V);
1354 } else {
1355 // TODO: memory intrinsics.
1356 // TODO: Some day, we should do the forward propagation and reassociation
1357 // parts of the algorithm.
1358 auto *I = cast<Instruction>(V);
1359 switch (I->getOpcode()) {
1360 case Instruction::ExtractValue:
1361 case Instruction::InsertValue:
1362 E = performSymbolicAggrValueEvaluation(I);
1363 break;
1364 case Instruction::PHI:
1365 E = performSymbolicPHIEvaluation(I);
1366 break;
1367 case Instruction::Call:
1368 E = performSymbolicCallEvaluation(I);
1369 break;
1370 case Instruction::Store:
1371 E = performSymbolicStoreEvaluation(I);
1372 break;
1373 case Instruction::Load:
1374 E = performSymbolicLoadEvaluation(I);
1375 break;
1376 case Instruction::BitCast: {
1377 E = createExpression(I);
1378 } break;
1379 case Instruction::ICmp:
1380 case Instruction::FCmp: {
1381 E = performSymbolicCmpEvaluation(I);
1382 } break;
1383 case Instruction::Add:
1384 case Instruction::FAdd:
1385 case Instruction::Sub:
1386 case Instruction::FSub:
1387 case Instruction::Mul:
1388 case Instruction::FMul:
1389 case Instruction::UDiv:
1390 case Instruction::SDiv:
1391 case Instruction::FDiv:
1392 case Instruction::URem:
1393 case Instruction::SRem:
1394 case Instruction::FRem:
1395 case Instruction::Shl:
1396 case Instruction::LShr:
1397 case Instruction::AShr:
1398 case Instruction::And:
1399 case Instruction::Or:
1400 case Instruction::Xor:
1401 case Instruction::Trunc:
1402 case Instruction::ZExt:
1403 case Instruction::SExt:
1404 case Instruction::FPToUI:
1405 case Instruction::FPToSI:
1406 case Instruction::UIToFP:
1407 case Instruction::SIToFP:
1408 case Instruction::FPTrunc:
1409 case Instruction::FPExt:
1410 case Instruction::PtrToInt:
1411 case Instruction::IntToPtr:
1412 case Instruction::Select:
1413 case Instruction::ExtractElement:
1414 case Instruction::InsertElement:
1415 case Instruction::ShuffleVector:
1416 case Instruction::GetElementPtr:
1417 E = createExpression(I);
1418 break;
1419 default:
1420 return nullptr;
1421 }
1422 }
1423 return E;
1424}
1425
1426void NewGVN::markUsersTouched(Value *V) {
1427 // Now mark the users as touched.
1428 for (auto *User : V->users()) {
1429 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1429, __PRETTY_FUNCTION__));
1430 TouchedInstructions.set(InstrDFS.lookup(User));
1431 }
1432}
1433
1434void NewGVN::markMemoryUsersTouched(MemoryAccess *MA) {
1435 for (auto U : MA->users()) {
1436 if (auto *MUD = dyn_cast<MemoryUseOrDef>(U))
1437 TouchedInstructions.set(InstrDFS.lookup(MUD->getMemoryInst()));
1438 else
1439 TouchedInstructions.set(InstrDFS.lookup(U));
1440 }
1441}
1442
1443// Add I to the set of users of a given predicate.
1444void NewGVN::addPredicateUsers(const PredicateBase *PB, Instruction *I) {
1445 if (auto *PBranch = dyn_cast<PredicateBranch>(PB))
1446 PredicateToUsers[PBranch->Condition].insert(I);
1447 else if (auto *PAssume = dyn_cast<PredicateBranch>(PB))
1448 PredicateToUsers[PAssume->Condition].insert(I);
1449}
1450
1451// Touch all the predicates that depend on this instruction.
1452void NewGVN::markPredicateUsersTouched(Instruction *I) {
1453 const auto Result = PredicateToUsers.find(I);
1454 if (Result != PredicateToUsers.end()) {
1455 for (auto *User : Result->second)
1456 TouchedInstructions.set(InstrDFS.lookup(User));
1457 PredicateToUsers.erase(Result);
1458 }
1459}
1460
1461// Touch the instructions that need to be updated after a congruence class has a
1462// leader change, and mark changed values.
1463void NewGVN::markLeaderChangeTouched(CongruenceClass *CC) {
1464 for (auto M : CC->Members) {
1465 if (auto *I = dyn_cast<Instruction>(M))
1466 TouchedInstructions.set(InstrDFS.lookup(I));
1467 LeaderChanges.insert(M);
1468 }
1469}
1470
1471// Move a value, currently in OldClass, to be part of NewClass
1472// Update OldClass for the move (including changing leaders, etc)
1473void NewGVN::moveValueToNewCongruenceClass(Instruction *I,
1474 CongruenceClass *OldClass,
1475 CongruenceClass *NewClass) {
1476 DEBUG(dbgs() << "New congruence class for " << I << " is " << NewClass->IDdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "New congruence class for " <<
I << " is " << NewClass->ID << "\n"; } }
while (false)
1477 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "New congruence class for " <<
I << " is " << NewClass->ID << "\n"; } }
while (false);
1478
1479 if (I == OldClass->NextLeader.first)
1480 OldClass->NextLeader = {nullptr, ~0U};
1481
1482 // It's possible, though unlikely, for us to discover equivalences such
1483 // that the current leader does not dominate the old one.
1484 // This statistic tracks how often this happens.
1485 // We assert on phi nodes when this happens, currently, for debugging, because
1486 // we want to make sure we name phi node cycles properly.
1487 if (isa<Instruction>(NewClass->RepLeader) && NewClass->RepLeader &&
1488 I != NewClass->RepLeader) {
1489 auto *IBB = I->getParent();
1490 auto *NCBB = cast<Instruction>(NewClass->RepLeader)->getParent();
1491 bool Dominated = IBB == NCBB &&
1492 InstrDFS.lookup(I) < InstrDFS.lookup(NewClass->RepLeader);
1493 Dominated = Dominated || DT->properlyDominates(IBB, NCBB);
1494 if (Dominated) {
1495 ++NumGVNNotMostDominatingLeader;
1496 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1498, __PRETTY_FUNCTION__))
1497 !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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1498, __PRETTY_FUNCTION__))
1498 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1498, __PRETTY_FUNCTION__));
1499 }
1500 }
1501
1502 if (NewClass->RepLeader != I) {
1503 auto DFSNum = InstrDFS.lookup(I);
1504 if (DFSNum < NewClass->NextLeader.second)
1505 NewClass->NextLeader = {I, DFSNum};
1506 }
1507
1508 OldClass->Members.erase(I);
1509 NewClass->Members.insert(I);
1510 MemoryAccess *StoreAccess = nullptr;
1511 if (auto *SI = dyn_cast<StoreInst>(I)) {
1512 StoreAccess = MSSA->getMemoryAccess(SI);
1513 --OldClass->StoreCount;
1514 assert(OldClass->StoreCount >= 0)((OldClass->StoreCount >= 0) ? static_cast<void> (
0) : __assert_fail ("OldClass->StoreCount >= 0", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1514, __PRETTY_FUNCTION__));
1515 ++NewClass->StoreCount;
1516 assert(NewClass->StoreCount > 0)((NewClass->StoreCount > 0) ? static_cast<void> (
0) : __assert_fail ("NewClass->StoreCount > 0", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1516, __PRETTY_FUNCTION__));
1517 if (!NewClass->RepMemoryAccess) {
1518 // If we don't have a representative memory access, it better be the only
1519 // store in there.
1520 assert(NewClass->StoreCount == 1)((NewClass->StoreCount == 1) ? static_cast<void> (0)
: __assert_fail ("NewClass->StoreCount == 1", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1520, __PRETTY_FUNCTION__));
1521 NewClass->RepMemoryAccess = StoreAccess;
1522 }
1523 setMemoryAccessEquivTo(StoreAccess, NewClass);
1524 }
1525
1526 ValueToClass[I] = NewClass;
1527 // See if we destroyed the class or need to swap leaders.
1528 if (OldClass->Members.empty() && OldClass != TOPClass) {
1529 if (OldClass->DefiningExpr) {
1530 OldClass->Dead = true;
1531 DEBUG(dbgs() << "Erasing expression " << OldClass->DefiningExprdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Erasing expression " << OldClass
->DefiningExpr << " from table\n"; } } while (false)
1532 << " from table\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Erasing expression " << OldClass
->DefiningExpr << " from table\n"; } } while (false);
1533 ExpressionToClass.erase(OldClass->DefiningExpr);
1534 }
1535 } else if (OldClass->RepLeader == I) {
1536 // When the leader changes, the value numbering of
1537 // everything may change due to symbolization changes, so we need to
1538 // reprocess.
1539 DEBUG(dbgs() << "Leader change!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Leader change!\n"; } } while (
false);
1540 ++NumGVNLeaderChanges;
1541 // Destroy the stored value if there are no more stores to represent it.
1542 if (OldClass->StoreCount == 0) {
1543 if (OldClass->RepStoredValue != nullptr)
1544 OldClass->RepStoredValue = nullptr;
1545 if (OldClass->RepMemoryAccess != nullptr)
1546 OldClass->RepMemoryAccess = nullptr;
1547 }
1548
1549 // If we destroy the old access leader, we have to effectively destroy the
1550 // congruence class. When it comes to scalars, anything with the same value
1551 // is as good as any other. That means that one leader is as good as
1552 // another, and as long as you have some leader for the value, you are
1553 // good.. When it comes to *memory states*, only one particular thing really
1554 // represents the definition of a given memory state. Once it goes away, we
1555 // need to re-evaluate which pieces of memory are really still
1556 // equivalent. The best way to do this is to re-value number things. The
1557 // only way to really make that happen is to destroy the rest of the class.
1558 // In order to effectively destroy the class, we reset ExpressionToClass for
1559 // each by using the ValueToExpression mapping. The members later get
1560 // marked as touched due to the leader change. We will create new
1561 // congruence classes, and the pieces that are still equivalent will end
1562 // back together in a new class. If this becomes too expensive, it is
1563 // possible to use a versioning scheme for the congruence classes to avoid
1564 // the expressions finding this old class.
1565 if (OldClass->StoreCount > 0 && OldClass->RepMemoryAccess == StoreAccess) {
1566 DEBUG(dbgs() << "Kicking everything out of class " << OldClass->IDdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Kicking everything out of class "
<< OldClass->ID << " because memory access leader changed"
; } } while (false)
1567 << " because memory access leader changed")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Kicking everything out of class "
<< OldClass->ID << " because memory access leader changed"
; } } while (false);
1568 for (auto Member : OldClass->Members)
1569 ExpressionToClass.erase(ValueToExpression.lookup(Member));
1570 }
1571
1572 // We don't need to sort members if there is only 1, and we don't care about
1573 // sorting the TOP class because everything either gets out of it or is
1574 // unreachable.
1575 if (OldClass->Members.size() == 1 || OldClass == TOPClass) {
1576 OldClass->RepLeader = *(OldClass->Members.begin());
1577 } else if (OldClass->NextLeader.first) {
1578 ++NumGVNAvoidedSortedLeaderChanges;
1579 OldClass->RepLeader = OldClass->NextLeader.first;
1580 OldClass->NextLeader = {nullptr, ~0U};
1581 } else {
1582 ++NumGVNSortedLeaderChanges;
1583 // TODO: If this ends up to slow, we can maintain a dual structure for
1584 // member testing/insertion, or keep things mostly sorted, and sort only
1585 // here, or ....
1586 std::pair<Value *, unsigned> MinDFS = {nullptr, ~0U};
1587 for (const auto X : OldClass->Members) {
1588 auto DFSNum = InstrDFS.lookup(X);
1589 if (DFSNum < MinDFS.second)
1590 MinDFS = {X, DFSNum};
1591 }
1592 OldClass->RepLeader = MinDFS.first;
1593 }
1594 markLeaderChangeTouched(OldClass);
1595 }
1596}
1597
1598// Perform congruence finding on a given value numbering expression.
1599void NewGVN::performCongruenceFinding(Instruction *I, const Expression *E) {
1600 ValueToExpression[I] = E;
1601 // This is guaranteed to return something, since it will at least find
1602 // TOP.
1603
1604 CongruenceClass *IClass = ValueToClass[I];
1605 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1605, __PRETTY_FUNCTION__));
1606 // Dead classes should have been eliminated from the mapping.
1607 assert(!IClass->Dead && "Found a dead class")((!IClass->Dead && "Found a dead class") ? static_cast
<void> (0) : __assert_fail ("!IClass->Dead && \"Found a dead class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1607, __PRETTY_FUNCTION__));
1608
1609 CongruenceClass *EClass;
1610 if (const auto *VE = dyn_cast<VariableExpression>(E)) {
1611 EClass = ValueToClass[VE->getVariableValue()];
1612 } else {
1613 auto lookupResult = ExpressionToClass.insert({E, nullptr});
1614
1615 // If it's not in the value table, create a new congruence class.
1616 if (lookupResult.second) {
1617 CongruenceClass *NewClass = createCongruenceClass(nullptr, E);
1618 auto place = lookupResult.first;
1619 place->second = NewClass;
1620
1621 // Constants and variables should always be made the leader.
1622 if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
1623 NewClass->RepLeader = CE->getConstantValue();
1624 } else if (const auto *SE = dyn_cast<StoreExpression>(E)) {
1625 StoreInst *SI = SE->getStoreInst();
1626 NewClass->RepLeader = SI;
1627 NewClass->RepStoredValue = lookupOperandLeader(SI->getValueOperand());
1628 // The RepMemoryAccess field will be filled in properly by the
1629 // moveValueToNewCongruenceClass call.
1630 } else {
1631 NewClass->RepLeader = I;
1632 }
1633 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1634, __PRETTY_FUNCTION__))
1634 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1634, __PRETTY_FUNCTION__));
1635
1636 EClass = NewClass;
1637 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->ID << " and leader " <<
*(NewClass->RepLeader); } } while (false)
1638 << " using expression " << *E << " at " << NewClass->IDdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Created new congruence class for "
<< *I << " using expression " << *E <<
" at " << NewClass->ID << " and leader " <<
*(NewClass->RepLeader); } } while (false)
1639 << " and leader " << *(NewClass->RepLeader))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Created new congruence class for "
<< *I << " using expression " << *E <<
" at " << NewClass->ID << " and leader " <<
*(NewClass->RepLeader); } } while (false);
1640 if (NewClass->RepStoredValue)
1641 DEBUG(dbgs() << " and stored value " << *(NewClass->RepStoredValue))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << " and stored value " << *
(NewClass->RepStoredValue); } } while (false);
1642 DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "\n"; } } while (false);
1643 DEBUG(dbgs() << "Hash value was " << E->getHashValue() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Hash value was " << E->
getHashValue() << "\n"; } } while (false);
1644 } else {
1645 EClass = lookupResult.first->second;
1646 if (isa<ConstantExpression>(E))
1647 assert(isa<Constant>(EClass->RepLeader) &&((isa<Constant>(EClass->RepLeader) && "Any class with a constant expression should have a "
"constant leader") ? static_cast<void> (0) : __assert_fail
("isa<Constant>(EClass->RepLeader) && \"Any class with a constant expression should have a \" \"constant leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1649, __PRETTY_FUNCTION__))
1648 "Any class with a constant expression should have a "((isa<Constant>(EClass->RepLeader) && "Any class with a constant expression should have a "
"constant leader") ? static_cast<void> (0) : __assert_fail
("isa<Constant>(EClass->RepLeader) && \"Any class with a constant expression should have a \" \"constant leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1649, __PRETTY_FUNCTION__))
1649 "constant leader")((isa<Constant>(EClass->RepLeader) && "Any class with a constant expression should have a "
"constant leader") ? static_cast<void> (0) : __assert_fail
("isa<Constant>(EClass->RepLeader) && \"Any class with a constant expression should have a \" \"constant leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1649, __PRETTY_FUNCTION__));
1650
1651 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1651, __PRETTY_FUNCTION__));
1652
1653 assert(!EClass->Dead && "We accidentally looked up a dead class")((!EClass->Dead && "We accidentally looked up a dead class"
) ? static_cast<void> (0) : __assert_fail ("!EClass->Dead && \"We accidentally looked up a dead class\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1653, __PRETTY_FUNCTION__));
1654 }
1655 }
1656 bool ClassChanged = IClass != EClass;
1657 bool LeaderChanged = LeaderChanges.erase(I);
1658 if (ClassChanged || LeaderChanged) {
1659 DEBUG(dbgs() << "Found class " << EClass->ID << " for expression " << Edo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found class " << EClass->
ID << " for expression " << E << "\n"; } } while
(false)
1660 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found class " << EClass->
ID << " for expression " << E << "\n"; } } while
(false);
1661
1662 if (ClassChanged)
1663 moveValueToNewCongruenceClass(I, IClass, EClass);
1664 markUsersTouched(I);
1665 if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
1666 markMemoryUsersTouched(MA);
1667 if (auto *CI = dyn_cast<CmpInst>(I))
1668 markPredicateUsersTouched(CI);
1669 }
1670}
1671
1672// Process the fact that Edge (from, to) is reachable, including marking
1673// any newly reachable blocks and instructions for processing.
1674void NewGVN::updateReachableEdge(BasicBlock *From, BasicBlock *To) {
1675 // Check if the Edge was reachable before.
1676 if (ReachableEdges.insert({From, To}).second) {
1677 // If this block wasn't reachable before, all instructions are touched.
1678 if (ReachableBlocks.insert(To).second) {
1679 DEBUG(dbgs() << "Block " << getBlockName(To) << " marked reachable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Block " << getBlockName(
To) << " marked reachable\n"; } } while (false);
1680 const auto &InstRange = BlockInstRange.lookup(To);
1681 TouchedInstructions.set(InstRange.first, InstRange.second);
1682 } else {
1683 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)
1684 << " 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)
1685 << "," << 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);
1686
1687 // We've made an edge reachable to an existing block, which may
1688 // impact predicates. Otherwise, only mark the phi nodes as touched, as
1689 // they are the only thing that depend on new edges. Anything using their
1690 // values will get propagated to if necessary.
1691 if (MemoryAccess *MemPhi = MSSA->getMemoryAccess(To))
1692 TouchedInstructions.set(InstrDFS.lookup(MemPhi));
1693
1694 auto BI = To->begin();
1695 while (isa<PHINode>(BI)) {
1696 TouchedInstructions.set(InstrDFS.lookup(&*BI));
1697 ++BI;
1698 }
1699 }
1700 }
1701}
1702
1703// Given a predicate condition (from a switch, cmp, or whatever) and a block,
1704// see if we know some constant value for it already.
1705Value *NewGVN::findConditionEquivalence(Value *Cond) const {
1706 auto Result = lookupOperandLeader(Cond);
1707 if (isa<Constant>(Result))
1708 return Result;
1709 return nullptr;
1710}
1711
1712// Process the outgoing edges of a block for reachability.
1713void NewGVN::processOutgoingEdges(TerminatorInst *TI, BasicBlock *B) {
1714 // Evaluate reachability of terminator instruction.
1715 BranchInst *BR;
1716 if ((BR = dyn_cast<BranchInst>(TI)) && BR->isConditional()) {
1717 Value *Cond = BR->getCondition();
1718 Value *CondEvaluated = findConditionEquivalence(Cond);
1719 if (!CondEvaluated) {
1720 if (auto *I = dyn_cast<Instruction>(Cond)) {
1721 const Expression *E = createExpression(I);
1722 if (const auto *CE = dyn_cast<ConstantExpression>(E)) {
1723 CondEvaluated = CE->getConstantValue();
1724 }
1725 } else if (isa<ConstantInt>(Cond)) {
1726 CondEvaluated = Cond;
1727 }
1728 }
1729 ConstantInt *CI;
1730 BasicBlock *TrueSucc = BR->getSuccessor(0);
1731 BasicBlock *FalseSucc = BR->getSuccessor(1);
1732 if (CondEvaluated && (CI = dyn_cast<ConstantInt>(CondEvaluated))) {
1733 if (CI->isOne()) {
1734 DEBUG(dbgs() << "Condition for Terminator " << *TIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Condition for Terminator " <<
*TI << " evaluated to true\n"; } } while (false)
1735 << " evaluated to true\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Condition for Terminator " <<
*TI << " evaluated to true\n"; } } while (false);
1736 updateReachableEdge(B, TrueSucc);
1737 } else if (CI->isZero()) {
1738 DEBUG(dbgs() << "Condition for Terminator " << *TIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Condition for Terminator " <<
*TI << " evaluated to false\n"; } } while (false)
1739 << " evaluated to false\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Condition for Terminator " <<
*TI << " evaluated to false\n"; } } while (false);
1740 updateReachableEdge(B, FalseSucc);
1741 }
1742 } else {
1743 updateReachableEdge(B, TrueSucc);
1744 updateReachableEdge(B, FalseSucc);
1745 }
1746 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
1747 // For switches, propagate the case values into the case
1748 // destinations.
1749
1750 // Remember how many outgoing edges there are to every successor.
1751 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
1752
1753 Value *SwitchCond = SI->getCondition();
1754 Value *CondEvaluated = findConditionEquivalence(SwitchCond);
1755 // See if we were able to turn this switch statement into a constant.
1756 if (CondEvaluated && isa<ConstantInt>(CondEvaluated)) {
1757 auto *CondVal = cast<ConstantInt>(CondEvaluated);
1758 // We should be able to get case value for this.
1759 auto CaseVal = SI->findCaseValue(CondVal);
1760 if (CaseVal.getCaseSuccessor() == SI->getDefaultDest()) {
1761 // We proved the value is outside of the range of the case.
1762 // We can't do anything other than mark the default dest as reachable,
1763 // and go home.
1764 updateReachableEdge(B, SI->getDefaultDest());
1765 return;
1766 }
1767 // Now get where it goes and mark it reachable.
1768 BasicBlock *TargetBlock = CaseVal.getCaseSuccessor();
1769 updateReachableEdge(B, TargetBlock);
1770 } else {
1771 for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) {
1772 BasicBlock *TargetBlock = SI->getSuccessor(i);
1773 ++SwitchEdges[TargetBlock];
1774 updateReachableEdge(B, TargetBlock);
1775 }
1776 }
1777 } else {
1778 // Otherwise this is either unconditional, or a type we have no
1779 // idea about. Just mark successors as reachable.
1780 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
1781 BasicBlock *TargetBlock = TI->getSuccessor(i);
1782 updateReachableEdge(B, TargetBlock);
1783 }
1784
1785 // This also may be a memory defining terminator, in which case, set it
1786 // equivalent to nothing.
1787 if (MemoryAccess *MA = MSSA->getMemoryAccess(TI))
1788 setMemoryAccessEquivTo(MA, nullptr);
1789 }
1790}
1791
1792// The algorithm initially places the values of the routine in the TOP
1793// congruence class. The leader of TOP is the undetermined value `undef`.
1794// When the algorithm has finished, values still in TOP are unreachable.
1795void NewGVN::initializeCongruenceClasses(Function &F) {
1796 // FIXME now i can't remember why this is 2
1797 NextCongruenceNum = 2;
1798 // Initialize all other instructions to be in TOP class.
1799 CongruenceClass::MemberSet InitialValues;
1800 TOPClass = createCongruenceClass(nullptr, nullptr);
1801 TOPClass->RepMemoryAccess = MSSA->getLiveOnEntryDef();
1802 for (auto &B : F) {
1803 if (auto *MP = MSSA->getMemoryAccess(&B))
1804 MemoryAccessToClass[MP] = TOPClass;
1805
1806 for (auto &I : B) {
1807 // Don't insert void terminators into the class. We don't value number
1808 // them, and they just end up sitting in TOP.
1809 if (isa<TerminatorInst>(I) && I.getType()->isVoidTy())
1810 continue;
1811 InitialValues.insert(&I);
1812 ValueToClass[&I] = TOPClass;
1813
1814 // All memory accesses are equivalent to live on entry to start. They must
1815 // be initialized to something so that initial changes are noticed. For
1816 // the maximal answer, we initialize them all to be the same as
1817 // liveOnEntry. Note that to save time, we only initialize the
1818 // MemoryDef's for stores and all MemoryPhis to be equal. Right now, no
1819 // other expression can generate a memory equivalence. If we start
1820 // handling memcpy/etc, we can expand this.
1821 if (isa<StoreInst>(&I)) {
1822 MemoryAccessToClass[MSSA->getMemoryAccess(&I)] = TOPClass;
1823 ++TOPClass->StoreCount;
1824 assert(TOPClass->StoreCount > 0)((TOPClass->StoreCount > 0) ? static_cast<void> (
0) : __assert_fail ("TOPClass->StoreCount > 0", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1824, __PRETTY_FUNCTION__));
1825 }
1826 }
1827 }
1828 TOPClass->Members.swap(InitialValues);
1829
1830 // Initialize arguments to be in their own unique congruence classes
1831 for (auto &FA : F.args())
1832 createSingletonCongruenceClass(&FA);
1833}
1834
1835void NewGVN::cleanupTables() {
1836 for (unsigned i = 0, e = CongruenceClasses.size(); i != e; ++i) {
1837 DEBUG(dbgs() << "Congruence class " << CongruenceClasses[i]->ID << " has "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Congruence class " << CongruenceClasses
[i]->ID << " has " << CongruenceClasses[i]->
Members.size() << " members\n"; } } while (false)
1838 << CongruenceClasses[i]->Members.size() << " members\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Congruence class " << CongruenceClasses
[i]->ID << " has " << CongruenceClasses[i]->
Members.size() << " members\n"; } } while (false);
1839 // Make sure we delete the congruence class (probably worth switching to
1840 // a unique_ptr at some point.
1841 delete CongruenceClasses[i];
1842 CongruenceClasses[i] = nullptr;
1843 }
1844
1845 ValueToClass.clear();
1846 ArgRecycler.clear(ExpressionAllocator);
1847 ExpressionAllocator.Reset();
1848 CongruenceClasses.clear();
1849 ExpressionToClass.clear();
1850 ValueToExpression.clear();
1851 ReachableBlocks.clear();
1852 ReachableEdges.clear();
1853#ifndef NDEBUG
1854 ProcessedCount.clear();
1855#endif
1856 InstrDFS.clear();
1857 InstructionsToErase.clear();
1858 DFSToInstr.clear();
1859 BlockInstRange.clear();
1860 TouchedInstructions.clear();
1861 MemoryAccessToClass.clear();
1862 PredicateToUsers.clear();
1863}
1864
1865std::pair<unsigned, unsigned> NewGVN::assignDFSNumbers(BasicBlock *B,
1866 unsigned Start) {
1867 unsigned End = Start;
1868 if (MemoryAccess *MemPhi = MSSA->getMemoryAccess(B)) {
1869 InstrDFS[MemPhi] = End++;
1870 DFSToInstr.emplace_back(MemPhi);
1871 }
1872
1873 for (auto &I : *B) {
1874 // There's no need to call isInstructionTriviallyDead more than once on
1875 // an instruction. Therefore, once we know that an instruction is dead
1876 // we change its DFS number so that it doesn't get value numbered.
1877 if (isInstructionTriviallyDead(&I, TLI)) {
1878 InstrDFS[&I] = 0;
1879 DEBUG(dbgs() << "Skipping trivially dead instruction " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Skipping trivially dead instruction "
<< I << "\n"; } } while (false);
1880 markInstructionForDeletion(&I);
1881 continue;
1882 }
1883
1884 InstrDFS[&I] = End++;
1885 DFSToInstr.emplace_back(&I);
1886 }
1887
1888 // All of the range functions taken half-open ranges (open on the end side).
1889 // So we do not subtract one from count, because at this point it is one
1890 // greater than the last instruction.
1891 return std::make_pair(Start, End);
1892}
1893
1894void NewGVN::updateProcessedCount(Value *V) {
1895#ifndef NDEBUG
1896 if (ProcessedCount.count(V) == 0) {
1897 ProcessedCount.insert({V, 1});
1898 } else {
1899 ++ProcessedCount[V];
1900 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1901, __PRETTY_FUNCTION__))
1901 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 1901, __PRETTY_FUNCTION__));
1902 }
1903#endif
1904}
1905// Evaluate MemoryPhi nodes symbolically, just like PHI nodes
1906void NewGVN::valueNumberMemoryPhi(MemoryPhi *MP) {
1907 // If all the arguments are the same, the MemoryPhi has the same value as the
1908 // argument.
1909 // Filter out unreachable blocks and self phis from our operands.
1910 const BasicBlock *PHIBlock = MP->getBlock();
1911 auto Filtered = make_filter_range(MP->operands(), [&](const Use &U) {
1912 return lookupMemoryAccessEquiv(cast<MemoryAccess>(U)) != MP &&
1913 !isMemoryAccessTop(cast<MemoryAccess>(U)) &&
1914 ReachableEdges.count({MP->getIncomingBlock(U), PHIBlock});
1915 });
1916 // If all that is left is nothing, our memoryphi is undef. We keep it as
1917 // InitialClass. Note: The only case this should happen is if we have at
1918 // least one self-argument.
1919 if (Filtered.begin() == Filtered.end()) {
1920 if (setMemoryAccessEquivTo(MP, TOPClass))
1921 markMemoryUsersTouched(MP);
1922 return;
1923 }
1924
1925 // Transform the remaining operands into operand leaders.
1926 // FIXME: mapped_iterator should have a range version.
1927 auto LookupFunc = [&](const Use &U) {
1928 return lookupMemoryAccessEquiv(cast<MemoryAccess>(U));
1929 };
1930 auto MappedBegin = map_iterator(Filtered.begin(), LookupFunc);
1931 auto MappedEnd = map_iterator(Filtered.end(), LookupFunc);
1932
1933 // and now check if all the elements are equal.
1934 // Sadly, we can't use std::equals since these are random access iterators.
1935 MemoryAccess *AllSameValue = *MappedBegin;
1936 ++MappedBegin;
1937 bool AllEqual = std::all_of(
1938 MappedBegin, MappedEnd,
1939 [&AllSameValue](const MemoryAccess *V) { return V == AllSameValue; });
1940
1941 if (AllEqual)
1942 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);
1943 else
1944 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);
1945
1946 if (setMemoryAccessEquivTo(
1947 MP, AllEqual ? MemoryAccessToClass.lookup(AllSameValue) : nullptr))
1948 markMemoryUsersTouched(MP);
1949}
1950
1951// Value number a single instruction, symbolically evaluating, performing
1952// congruence finding, and updating mappings.
1953void NewGVN::valueNumberInstruction(Instruction *I) {
1954 DEBUG(dbgs() << "Processing instruction " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Processing instruction " <<
*I << "\n"; } } while (false);
1955 if (!I->isTerminator()) {
1956 const Expression *Symbolized = nullptr;
1957 if (DebugCounter::shouldExecute(VNCounter)) {
1958 Symbolized = performSymbolicEvaluation(I);
1959 } else {
1960 // Mark the instruction as unused so we don't value number it again.
1961 InstrDFS[I] = 0;
1962 }
1963 // If we couldn't come up with a symbolic expression, use the unknown
1964 // expression
1965 if (Symbolized == nullptr)
1966 Symbolized = createUnknownExpression(I);
1967 performCongruenceFinding(I, Symbolized);
1968 } else {
1969 // Handle terminators that return values. All of them produce values we
1970 // don't currently understand. We don't place non-value producing
1971 // terminators in a class.
1972 if (!I->getType()->isVoidTy()) {
1973 auto *Symbolized = createUnknownExpression(I);
1974 performCongruenceFinding(I, Symbolized);
1975 }
1976 processOutgoingEdges(dyn_cast<TerminatorInst>(I), I->getParent());
1977 }
1978}
1979
1980// Check if there is a path, using single or equal argument phi nodes, from
1981// First to Second.
1982bool NewGVN::singleReachablePHIPath(const MemoryAccess *First,
1983 const MemoryAccess *Second) const {
1984 if (First == Second)
1985 return true;
1986 if (MSSA->isLiveOnEntryDef(First))
1987 return false;
1988 const auto *EndDef = First;
1989 for (auto *ChainDef : optimized_def_chain(First)) {
1990 if (ChainDef == Second)
1991 return true;
1992 if (MSSA->isLiveOnEntryDef(ChainDef))
1993 return false;
1994 EndDef = ChainDef;
1995 }
1996 auto *MP = cast<MemoryPhi>(EndDef);
1997 auto ReachableOperandPred = [&](const Use &U) {
1998 return ReachableEdges.count({MP->getIncomingBlock(U), MP->getBlock()});
1999 };
2000 auto FilteredPhiArgs =
2001 make_filter_range(MP->operands(), ReachableOperandPred);
2002 SmallVector<const Value *, 32> OperandList;
2003 std::copy(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
2004 std::back_inserter(OperandList));
2005 bool Okay = OperandList.size() == 1;
2006 if (!Okay)
2007 Okay =
2008 std::equal(OperandList.begin(), OperandList.end(), OperandList.begin());
2009 if (Okay)
2010 return singleReachablePHIPath(cast<MemoryAccess>(OperandList[0]), Second);
2011 return false;
2012}
2013
2014// Verify the that the memory equivalence table makes sense relative to the
2015// congruence classes. Note that this checking is not perfect, and is currently
2016// subject to very rare false negatives. It is only useful for
2017// testing/debugging.
2018void NewGVN::verifyMemoryCongruency() const {
2019#ifndef NDEBUG
2020 // Anything equivalent in the memory access table should be in the same
2021 // congruence class.
2022
2023 // Filter out the unreachable and trivially dead entries, because they may
2024 // never have been updated if the instructions were not processed.
2025 auto ReachableAccessPred =
2026 [&](const std::pair<const MemoryAccess *, CongruenceClass *> Pair) {
2027 bool Result = ReachableBlocks.count(Pair.first->getBlock());
2028 if (!Result)
2029 return false;
2030 if (auto *MemDef = dyn_cast<MemoryDef>(Pair.first))
2031 return !isInstructionTriviallyDead(MemDef->getMemoryInst());
2032 return true;
2033 };
2034
2035 auto Filtered = make_filter_range(MemoryAccessToClass, ReachableAccessPred);
2036 for (auto KV : Filtered) {
2037 // Unreachable instructions may not have changed because we never process
2038 // them.
2039 if (!ReachableBlocks.count(KV.first->getBlock()))
2040 continue;
2041 if (auto *FirstMUD = dyn_cast<MemoryUseOrDef>(KV.first)) {
2042 auto *SecondMUD = dyn_cast<MemoryUseOrDef>(KV.second->RepMemoryAccess);
2043 if (FirstMUD && SecondMUD)
2044 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2049, __PRETTY_FUNCTION__))
2045 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2049, __PRETTY_FUNCTION__))
2046 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2049, __PRETTY_FUNCTION__))
2047 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2049, __PRETTY_FUNCTION__))
2048 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2049, __PRETTY_FUNCTION__))
2049 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2049, __PRETTY_FUNCTION__));
2050 } else if (auto *FirstMP = dyn_cast<MemoryPhi>(KV.first)) {
2051
2052 // We can only sanely verify that MemoryDefs in the operand list all have
2053 // the same class.
2054 auto ReachableOperandPred = [&](const Use &U) {
2055 return ReachableEdges.count(
2056 {FirstMP->getIncomingBlock(U), FirstMP->getBlock()}) &&
2057 isa<MemoryDef>(U);
2058
2059 };
2060 // All arguments should in the same class, ignoring unreachable arguments
2061 auto FilteredPhiArgs =
2062 make_filter_range(FirstMP->operands(), ReachableOperandPred);
2063 SmallVector<const CongruenceClass *, 16> PhiOpClasses;
2064 std::transform(FilteredPhiArgs.begin(), FilteredPhiArgs.end(),
2065 std::back_inserter(PhiOpClasses), [&](const Use &U) {
2066 const MemoryDef *MD = cast<MemoryDef>(U);
2067 return ValueToClass.lookup(MD->getMemoryInst());
2068 });
2069 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2071, __PRETTY_FUNCTION__))
2070 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2071, __PRETTY_FUNCTION__))
2071 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2071, __PRETTY_FUNCTION__));
2072 }
2073 }
2074#endif
2075}
2076
2077// Verify that the sparse propagation we did actually found the maximal fixpoint
2078// We do this by storing the value to class mapping, touching all instructions,
2079// and redoing the iteration to see if anything changed.
2080void NewGVN::verifyIterationSettled(Function &F) {
2081#ifndef NDEBUG
2082 if (DebugCounter::isCounterSet(VNCounter))
1
Assuming the condition is false
→
2
←
Taking false branch
→
2083 DebugCounter::setCounterValue(VNCounter, StartingVNCounter);
2084
2085 // Note that we have to store the actual classes, as we may change existing
2086 // classes during iteration. This is because our memory iteration propagation
2087 // is not perfect, and so may waste a little work. But it should generate
2088 // exactly the same congruence classes we have now, with different IDs.
2089 std::map<const Value *, CongruenceClass> BeforeIteration;
2090
2091 for (auto &KV : ValueToClass) {
2092 if (auto *I = dyn_cast<Instruction>(KV.first))
2093 // Skip unused/dead instructions.
2094 if (InstrDFS.lookup(I) == 0)
2095 continue;
2096 BeforeIteration.insert({KV.first, *KV.second});
2097 }
2098
2099 TouchedInstructions.set();
2100 TouchedInstructions.reset(0);
2101 iterateTouchedInstructions();
2102 DenseSet<std::pair<const CongruenceClass *, const CongruenceClass *>>
2103 EqualClasses;
2104 for (const auto &KV : ValueToClass) {
2105 if (auto *I = dyn_cast<Instruction>(KV.first))
3
←
Taking false branch
→
2106 // Skip unused/dead instructions.
2107 if (InstrDFS.lookup(I) == 0)
2108 continue;
2109 // We could sink these uses, but i think this adds a bit of clarity here as
2110 // to what we are comparing.
2111 auto *BeforeCC = &BeforeIteration.find(KV.first)->second;
2112 auto *AfterCC = KV.second;
2113 // Note that the classes can't change at this point, so we memoize the set
2114 // that are equal.
2115 if (!EqualClasses.count({BeforeCC, AfterCC})) {
4
←
Assuming the condition is true
→
5
←
Taking true branch
→
2116 assert(areClassesEquivalent(BeforeCC, AfterCC) &&((areClassesEquivalent(BeforeCC, AfterCC) && "Value number changed after main loop completed!"
) ? static_cast<void> (0) : __assert_fail ("areClassesEquivalent(BeforeCC, AfterCC) && \"Value number changed after main loop completed!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2117, __PRETTY_FUNCTION__))
6
←
Within the expansion of the macro 'assert':
→
a
Passing value via 1st parameter 'A'
b
Calling 'areClassesEquivalent'
2117 "Value number changed after main loop completed!")((areClassesEquivalent(BeforeCC, AfterCC) && "Value number changed after main loop completed!"
) ? static_cast<void> (0) : __assert_fail ("areClassesEquivalent(BeforeCC, AfterCC) && \"Value number changed after main loop completed!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2117, __PRETTY_FUNCTION__));
2118 EqualClasses.insert({BeforeCC, AfterCC});
2119 }
2120 }
2121#endif
2122}
2123
2124// This is the main value numbering loop, it iterates over the initial touched
2125// instruction set, propagating value numbers, marking things touched, etc,
2126// until the set of touched instructions is completely empty.
2127void NewGVN::iterateTouchedInstructions() {
2128 unsigned int Iterations = 0;
2129 // Figure out where touchedinstructions starts
2130 int FirstInstr = TouchedInstructions.find_first();
2131 // Nothing set, nothing to iterate, just return.
2132 if (FirstInstr == -1)
2133 return;
2134 BasicBlock *LastBlock = getBlockForValue(DFSToInstr[FirstInstr]);
2135 while (TouchedInstructions.any()) {
2136 ++Iterations;
2137 // Walk through all the instructions in all the blocks in RPO.
2138 // TODO: As we hit a new block, we should push and pop equalities into a
2139 // table lookupOperandLeader can use, to catch things PredicateInfo
2140 // might miss, like edge-only equivalences.
2141 for (int InstrNum = TouchedInstructions.find_first(); InstrNum != -1;
2142 InstrNum = TouchedInstructions.find_next(InstrNum)) {
2143
2144 // This instruction was found to be dead. We don't bother looking
2145 // at it again.
2146 if (InstrNum == 0) {
2147 TouchedInstructions.reset(InstrNum);
2148 continue;
2149 }
2150
2151 Value *V = DFSToInstr[InstrNum];
2152 BasicBlock *CurrBlock = getBlockForValue(V);
2153
2154 // If we hit a new block, do reachability processing.
2155 if (CurrBlock != LastBlock) {
2156 LastBlock = CurrBlock;
2157 bool BlockReachable = ReachableBlocks.count(CurrBlock);
2158 const auto &CurrInstRange = BlockInstRange.lookup(CurrBlock);
2159
2160 // If it's not reachable, erase any touched instructions and move on.
2161 if (!BlockReachable) {
2162 TouchedInstructions.reset(CurrInstRange.first, CurrInstRange.second);
2163 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)
2164 << getBlockName(CurrBlock)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Skipping instructions in block "
<< getBlockName(CurrBlock) << " because it is unreachable\n"
; } } while (false)
2165 << " 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);
2166 continue;
2167 }
2168 updateProcessedCount(CurrBlock);
2169 }
2170
2171 if (auto *MP = dyn_cast<MemoryPhi>(V)) {
2172 DEBUG(dbgs() << "Processing MemoryPhi " << *MP << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Processing MemoryPhi " <<
*MP << "\n"; } } while (false);
2173 valueNumberMemoryPhi(MP);
2174 } else if (auto *I = dyn_cast<Instruction>(V)) {
2175 valueNumberInstruction(I);
2176 } else {
2177 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2177);
2178 }
2179 updateProcessedCount(V);
2180 // Reset after processing (because we may mark ourselves as touched when
2181 // we propagate equalities).
2182 TouchedInstructions.reset(InstrNum);
2183 }
2184 }
2185 NumGVNMaxIterations = std::max(NumGVNMaxIterations.getValue(), Iterations);
2186}
2187
2188// This is the main transformation entry point.
2189bool NewGVN::runGVN() {
2190 if (DebugCounter::isCounterSet(VNCounter))
2191 StartingVNCounter = DebugCounter::getCounterValue(VNCounter);
2192 bool Changed = false;
2193 NumFuncArgs = F.arg_size();
2194 MSSAWalker = MSSA->getWalker();
2195
2196 // Count number of instructions for sizing of hash tables, and come
2197 // up with a global dfs numbering for instructions.
2198 unsigned ICount = 1;
2199 // Add an empty instruction to account for the fact that we start at 1
2200 DFSToInstr.emplace_back(nullptr);
2201 // Note: We want ideal RPO traversal of the blocks, which is not quite the
2202 // same as dominator tree order, particularly with regard whether backedges
2203 // get visited first or second, given a block with multiple successors.
2204 // If we visit in the wrong order, we will end up performing N times as many
2205 // iterations.
2206 // The dominator tree does guarantee that, for a given dom tree node, it's
2207 // parent must occur before it in the RPO ordering. Thus, we only need to sort
2208 // the siblings.
2209 DenseMap<const DomTreeNode *, unsigned> RPOOrdering;
2210 ReversePostOrderTraversal<Function *> RPOT(&F);
2211 unsigned Counter = 0;
2212 for (auto &B : RPOT) {
2213 auto *Node = DT->getNode(B);
2214 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2214, __PRETTY_FUNCTION__));
2215 RPOOrdering[Node] = ++Counter;
2216 }
2217 // Sort dominator tree children arrays into RPO.
2218 for (auto &B : RPOT) {
2219 auto *Node = DT->getNode(B);
2220 if (Node->getChildren().size() > 1)
2221 std::sort(Node->begin(), Node->end(),
2222 [&RPOOrdering](const DomTreeNode *A, const DomTreeNode *B) {
2223 return RPOOrdering[A] < RPOOrdering[B];
2224 });
2225 }
2226
2227 // Now a standard depth first ordering of the domtree is equivalent to RPO.
2228 auto DFI = df_begin(DT->getRootNode());
2229 for (auto DFE = df_end(DT->getRootNode()); DFI != DFE; ++DFI) {
2230 BasicBlock *B = DFI->getBlock();
2231 const auto &BlockRange = assignDFSNumbers(B, ICount);
2232 BlockInstRange.insert({B, BlockRange});
2233 ICount += BlockRange.second - BlockRange.first;
2234 }
2235
2236 // Handle forward unreachable blocks and figure out which blocks
2237 // have single preds.
2238 for (auto &B : F) {
2239 // Assign numbers to unreachable blocks.
2240 if (!DFI.nodeVisited(DT->getNode(&B))) {
2241 const auto &BlockRange = assignDFSNumbers(&B, ICount);
2242 BlockInstRange.insert({&B, BlockRange});
2243 ICount += BlockRange.second - BlockRange.first;
2244 }
2245 }
2246
2247 TouchedInstructions.resize(ICount);
2248 // Ensure we don't end up resizing the expressionToClass map, as
2249 // that can be quite expensive. At most, we have one expression per
2250 // instruction.
2251 ExpressionToClass.reserve(ICount);
2252
2253 // Initialize the touched instructions to include the entry block.
2254 const auto &InstRange = BlockInstRange.lookup(&F.getEntryBlock());
2255 TouchedInstructions.set(InstRange.first, InstRange.second);
2256 ReachableBlocks.insert(&F.getEntryBlock());
2257
2258 initializeCongruenceClasses(F);
2259 iterateTouchedInstructions();
2260 verifyMemoryCongruency();
2261 verifyIterationSettled(F);
2262
2263 Changed |= eliminateInstructions(F);
2264
2265 // Delete all instructions marked for deletion.
2266 for (Instruction *ToErase : InstructionsToErase) {
2267 if (!ToErase->use_empty())
2268 ToErase->replaceAllUsesWith(UndefValue::get(ToErase->getType()));
2269
2270 ToErase->eraseFromParent();
2271 }
2272
2273 // Delete all unreachable blocks.
2274 auto UnreachableBlockPred = [&](const BasicBlock &BB) {
2275 return !ReachableBlocks.count(&BB);
2276 };
2277
2278 for (auto &BB : make_filter_range(F, UnreachableBlockPred)) {
2279 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)
2280 << " is unreachable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "We believe block " << getBlockName
(&BB) << " is unreachable\n"; } } while (false);
2281 deleteInstructionsInBlock(&BB);
2282 Changed = true;
2283 }
2284
2285 cleanupTables();
2286 return Changed;
2287}
2288
2289// Return true if V is a value that will always be available (IE can
2290// be placed anywhere) in the function. We don't do globals here
2291// because they are often worse to put in place.
2292// TODO: Separate cost from availability
2293static bool alwaysAvailable(Value *V) {
2294 return isa<Constant>(V) || isa<Argument>(V);
2295}
2296
2297struct NewGVN::ValueDFS {
2298 int DFSIn = 0;
2299 int DFSOut = 0;
2300 int LocalNum = 0;
2301 // Only one of Def and U will be set.
2302 // The bool in the Def tells us whether the Def is the stored value of a
2303 // store.
2304 PointerIntPair<Value *, 1, bool> Def;
2305 Use *U = nullptr;
2306 bool operator<(const ValueDFS &Other) const {
2307 // It's not enough that any given field be less than - we have sets
2308 // of fields that need to be evaluated together to give a proper ordering.
2309 // For example, if you have;
2310 // DFS (1, 3)
2311 // Val 0
2312 // DFS (1, 2)
2313 // Val 50
2314 // We want the second to be less than the first, but if we just go field
2315 // by field, we will get to Val 0 < Val 50 and say the first is less than
2316 // the second. We only want it to be less than if the DFS orders are equal.
2317 //
2318 // Each LLVM instruction only produces one value, and thus the lowest-level
2319 // differentiator that really matters for the stack (and what we use as as a
2320 // replacement) is the local dfs number.
2321 // Everything else in the structure is instruction level, and only affects
2322 // the order in which we will replace operands of a given instruction.
2323 //
2324 // For a given instruction (IE things with equal dfsin, dfsout, localnum),
2325 // the order of replacement of uses does not matter.
2326 // IE given,
2327 // a = 5
2328 // b = a + a
2329 // When you hit b, you will have two valuedfs with the same dfsin, out, and
2330 // localnum.
2331 // The .val will be the same as well.
2332 // The .u's will be different.
2333 // You will replace both, and it does not matter what order you replace them
2334 // in (IE whether you replace operand 2, then operand 1, or operand 1, then
2335 // operand 2).
2336 // Similarly for the case of same dfsin, dfsout, localnum, but different
2337 // .val's
2338 // a = 5
2339 // b = 6
2340 // c = a + b
2341 // in c, we will a valuedfs for a, and one for b,with everything the same
2342 // but .val and .u.
2343 // It does not matter what order we replace these operands in.
2344 // You will always end up with the same IR, and this is guaranteed.
2345 return std::tie(DFSIn, DFSOut, LocalNum, Def, U) <
2346 std::tie(Other.DFSIn, Other.DFSOut, Other.LocalNum, Other.Def,
2347 Other.U);
2348 }
2349};
2350
2351// This function converts the set of members for a congruence class from values,
2352// to sets of defs and uses with associated DFS info. The total number of
2353// reachable uses for each value is stored in UseCount, and instructions that
2354// seem
2355// dead (have no non-dead uses) are stored in ProbablyDead.
2356void NewGVN::convertClassToDFSOrdered(
2357 const CongruenceClass::MemberSet &Dense,
2358 SmallVectorImpl<ValueDFS> &DFSOrderedSet,
2359 DenseMap<const Value *, unsigned int> &UseCounts,
2360 SmallPtrSetImpl<Instruction *> &ProbablyDead) {
2361 for (auto D : Dense) {
2362 // First add the value.
2363 BasicBlock *BB = getBlockForValue(D);
2364 // Constants are handled prior to ever calling this function, so
2365 // we should only be left with instructions as members.
2366 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2366, __PRETTY_FUNCTION__));
2367 ValueDFS VDDef;
2368 DomTreeNode *DomNode = DT->getNode(BB);
2369 VDDef.DFSIn = DomNode->getDFSNumIn();
2370 VDDef.DFSOut = DomNode->getDFSNumOut();
2371 // If it's a store, use the leader of the value operand, if it's always
2372 // available, or the value operand. TODO: We could do dominance checks to
2373 // find a dominating leader, but not worth it ATM.
2374 if (auto *SI = dyn_cast<StoreInst>(D)) {
2375 auto Leader = lookupOperandLeader(SI->getValueOperand());
2376 if (alwaysAvailable(Leader)) {
2377 VDDef.Def.setPointer(Leader);
2378 } else {
2379 VDDef.Def.setPointer(SI->getValueOperand());
2380 VDDef.Def.setInt(true);
2381 }
2382 } else {
2383 VDDef.Def.setPointer(D);
2384 }
2385 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2386, __PRETTY_FUNCTION__))
2386 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2386, __PRETTY_FUNCTION__));
2387 VDDef.LocalNum = InstrDFS.lookup(D);
2388 DFSOrderedSet.emplace_back(VDDef);
2389 Instruction *Def = cast<Instruction>(D);
2390 unsigned int UseCount = 0;
2391 // Now add the uses.
2392 for (auto &U : Def->uses()) {
2393 if (auto *I = dyn_cast<Instruction>(U.getUser())) {
2394 // Don't try to replace into dead uses
2395 if (InstructionsToErase.count(I))
2396 continue;
2397 ValueDFS VDUse;
2398 // Put the phi node uses in the incoming block.
2399 BasicBlock *IBlock;
2400 if (auto *P = dyn_cast<PHINode>(I)) {
2401 IBlock = P->getIncomingBlock(U);
2402 // Make phi node users appear last in the incoming block
2403 // they are from.
2404 VDUse.LocalNum = InstrDFS.size() + 1;
2405 } else {
2406 IBlock = I->getParent();
2407 VDUse.LocalNum = InstrDFS.lookup(I);
2408 }
2409
2410 // Skip uses in unreachable blocks, as we're going
2411 // to delete them.
2412 if (ReachableBlocks.count(IBlock) == 0)
2413 continue;
2414
2415 DomTreeNode *DomNode = DT->getNode(IBlock);
2416 VDUse.DFSIn = DomNode->getDFSNumIn();
2417 VDUse.DFSOut = DomNode->getDFSNumOut();
2418 VDUse.U = &U;
2419 ++UseCount;
2420 DFSOrderedSet.emplace_back(VDUse);
2421 }
2422 }
2423
2424 // If there are no uses, it's probably dead (but it may have side-effects,
2425 // so not definitely dead. Otherwise, store the number of uses so we can
2426 // track if it becomes dead later).
2427 if (UseCount == 0)
2428 ProbablyDead.insert(Def);
2429 else
2430 UseCounts[Def] = UseCount;
2431 }
2432}
2433
2434// This function converts the set of members for a congruence class from values,
2435// to the set of defs for loads and stores, with associated DFS info.
2436void NewGVN::convertClassToLoadsAndStores(
2437 const CongruenceClass::MemberSet &Dense,
2438 SmallVectorImpl<ValueDFS> &LoadsAndStores) {
2439 for (auto D : Dense) {
2440 if (!isa<LoadInst>(D) && !isa<StoreInst>(D))
2441 continue;
2442
2443 BasicBlock *BB = getBlockForValue(D);
2444 ValueDFS VD;
2445 DomTreeNode *DomNode = DT->getNode(BB);
2446 VD.DFSIn = DomNode->getDFSNumIn();
2447 VD.DFSOut = DomNode->getDFSNumOut();
2448 VD.Def.setPointer(D);
2449
2450 // If it's an instruction, use the real local dfs number.
2451 if (auto *I = dyn_cast<Instruction>(D))
2452 VD.LocalNum = InstrDFS.lookup(I);
2453 else
2454 llvm_unreachable("Should have been an instruction")::llvm::llvm_unreachable_internal("Should have been an instruction"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2454);
2455
2456 LoadsAndStores.emplace_back(VD);
2457 }
2458}
2459
2460static void patchReplacementInstruction(Instruction *I, Value *Repl) {
2461 auto *ReplInst = dyn_cast<Instruction>(Repl);
2462 if (!ReplInst)
2463 return;
2464
2465 // Patch the replacement so that it is not more restrictive than the value
2466 // being replaced.
2467 // Note that if 'I' is a load being replaced by some operation,
2468 // for example, by an arithmetic operation, then andIRFlags()
2469 // would just erase all math flags from the original arithmetic
2470 // operation, which is clearly not wanted and not needed.
2471 if (!isa<LoadInst>(I))
2472 ReplInst->andIRFlags(I);
2473
2474 // FIXME: If both the original and replacement value are part of the
2475 // same control-flow region (meaning that the execution of one
2476 // guarantees the execution of the other), then we can combine the
2477 // noalias scopes here and do better than the general conservative
2478 // answer used in combineMetadata().
2479
2480 // In general, GVN unifies expressions over different control-flow
2481 // regions, and so we need a conservative combination of the noalias
2482 // scopes.
2483 static const unsigned KnownIDs[] = {
2484 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
2485 LLVMContext::MD_noalias, LLVMContext::MD_range,
2486 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
2487 LLVMContext::MD_invariant_group};
2488 combineMetadata(ReplInst, I, KnownIDs);
2489}
2490
2491static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
2492 patchReplacementInstruction(I, Repl);
2493 I->replaceAllUsesWith(Repl);
2494}
2495
2496void NewGVN::deleteInstructionsInBlock(BasicBlock *BB) {
2497 DEBUG(dbgs() << " BasicBlock Dead:" << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << " BasicBlock Dead:" << *
BB; } } while (false);
2498 ++NumGVNBlocksDeleted;
2499
2500 // Delete the instructions backwards, as it has a reduced likelihood of having
2501 // to update as many def-use and use-def chains. Start after the terminator.
2502 auto StartPoint = BB->rbegin();
2503 ++StartPoint;
2504 // Note that we explicitly recalculate BB->rend() on each iteration,
2505 // as it may change when we remove the first instruction.
2506 for (BasicBlock::reverse_iterator I(StartPoint); I != BB->rend();) {
2507 Instruction &Inst = *I++;
2508 if (!Inst.use_empty())
2509 Inst.replaceAllUsesWith(UndefValue::get(Inst.getType()));
2510 if (isa<LandingPadInst>(Inst))
2511 continue;
2512
2513 Inst.eraseFromParent();
2514 ++NumGVNInstrDeleted;
2515 }
2516 // Now insert something that simplifycfg will turn into an unreachable.
2517 Type *Int8Ty = Type::getInt8Ty(BB->getContext());
2518 new StoreInst(UndefValue::get(Int8Ty),
2519 Constant::getNullValue(Int8Ty->getPointerTo()),
2520 BB->getTerminator());
2521}
2522
2523void NewGVN::markInstructionForDeletion(Instruction *I) {
2524 DEBUG(dbgs() << "Marking " << *I << " for deletion\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Marking " << *I <<
" for deletion\n"; } } while (false);
2525 InstructionsToErase.insert(I);
2526}
2527
2528void NewGVN::replaceInstruction(Instruction *I, Value *V) {
2529
2530 DEBUG(dbgs() << "Replacing " << *I << " with " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Replacing " << *I <<
" with " << *V << "\n"; } } while (false);
2531 patchAndReplaceAllUsesWith(I, V);
2532 // We save the actual erasing to avoid invalidating memory
2533 // dependencies until we are done with everything.
2534 markInstructionForDeletion(I);
2535}
2536
2537namespace {
2538
2539// This is a stack that contains both the value and dfs info of where
2540// that value is valid.
2541class ValueDFSStack {
2542public:
2543 Value *back() const { return ValueStack.back(); }
2544 std::pair<int, int> dfs_back() const { return DFSStack.back(); }
2545
2546 void push_back(Value *V, int DFSIn, int DFSOut) {
2547 ValueStack.emplace_back(V);
2548 DFSStack.emplace_back(DFSIn, DFSOut);
2549 }
2550 bool empty() const { return DFSStack.empty(); }
2551 bool isInScope(int DFSIn, int DFSOut) const {
2552 if (empty())
2553 return false;
2554 return DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second;
2555 }
2556
2557 void popUntilDFSScope(int DFSIn, int DFSOut) {
2558
2559 // These two should always be in sync at this point.
2560 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2561, __PRETTY_FUNCTION__))
2561 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2561, __PRETTY_FUNCTION__));
2562 while (
2563 !DFSStack.empty() &&
2564 !(DFSIn >= DFSStack.back().first && DFSOut <= DFSStack.back().second)) {
2565 DFSStack.pop_back();
2566 ValueStack.pop_back();
2567 }
2568 }
2569
2570private:
2571 SmallVector<Value *, 8> ValueStack;
2572 SmallVector<std::pair<int, int>, 8> DFSStack;
2573};
2574}
2575
2576bool NewGVN::eliminateInstructions(Function &F) {
2577 // This is a non-standard eliminator. The normal way to eliminate is
2578 // to walk the dominator tree in order, keeping track of available
2579 // values, and eliminating them. However, this is mildly
2580 // pointless. It requires doing lookups on every instruction,
2581 // regardless of whether we will ever eliminate it. For
2582 // instructions part of most singleton congruence classes, we know we
2583 // will never eliminate them.
2584
2585 // Instead, this eliminator looks at the congruence classes directly, sorts
2586 // them into a DFS ordering of the dominator tree, and then we just
2587 // perform elimination straight on the sets by walking the congruence
2588 // class member uses in order, and eliminate the ones dominated by the
2589 // last member. This is worst case O(E log E) where E = number of
2590 // instructions in a single congruence class. In theory, this is all
2591 // instructions. In practice, it is much faster, as most instructions are
2592 // either in singleton congruence classes or can't possibly be eliminated
2593 // anyway (if there are no overlapping DFS ranges in class).
2594 // When we find something not dominated, it becomes the new leader
2595 // for elimination purposes.
2596 // TODO: If we wanted to be faster, We could remove any members with no
2597 // overlapping ranges while sorting, as we will never eliminate anything
2598 // with those members, as they don't dominate anything else in our set.
2599
2600 bool AnythingReplaced = false;
2601
2602 // Since we are going to walk the domtree anyway, and we can't guarantee the
2603 // DFS numbers are updated, we compute some ourselves.
2604 DT->updateDFSNumbers();
2605
2606 for (auto &B : F) {
2607 if (!ReachableBlocks.count(&B)) {
2608 for (const auto S : successors(&B)) {
2609 for (auto II = S->begin(); isa<PHINode>(II); ++II) {
2610 auto &Phi = cast<PHINode>(*II);
2611 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
)
2612 << 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
)
2613 << " 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
);
2614 for (auto &Operand : Phi.incoming_values())
2615 if (Phi.getIncomingBlock(Operand) == &B)
2616 Operand.set(UndefValue::get(Phi.getType()));
2617 }
2618 }
2619 }
2620 }
2621
2622 // Map to store the use counts
2623 DenseMap<const Value *, unsigned int> UseCounts;
2624 for (CongruenceClass *CC : reverse(CongruenceClasses)) {
2625 // Track the equivalent store info so we can decide whether to try
2626 // dead store elimination.
2627 SmallVector<ValueDFS, 8> PossibleDeadStores;
2628 SmallPtrSet<Instruction *, 8> ProbablyDead;
2629 if (CC->Dead)
2630 continue;
2631 // Everything still in the TOP class is unreachable or dead.
2632 if (CC == TOPClass) {
2633#ifndef NDEBUG
2634 for (auto M : CC->Members)
2635 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2638, __PRETTY_FUNCTION__))
2636 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2638, __PRETTY_FUNCTION__))
2637 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2638, __PRETTY_FUNCTION__))
2638 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2638, __PRETTY_FUNCTION__));
2639#endif
2640 continue;
2641 }
2642
2643 assert(CC->RepLeader && "We should have had a leader")((CC->RepLeader && "We should have had a leader") ?
static_cast<void> (0) : __assert_fail ("CC->RepLeader && \"We should have had a leader\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2643, __PRETTY_FUNCTION__));
2644
2645 // If this is a leader that is always available, and it's a
2646 // constant or has no equivalences, just replace everything with
2647 // it. We then update the congruence class with whatever members
2648 // are left.
2649 Value *Leader = CC->RepStoredValue ? CC->RepStoredValue : CC->RepLeader;
2650 if (alwaysAvailable(Leader)) {
2651 SmallPtrSet<Value *, 4> MembersLeft;
2652 for (auto M : CC->Members) {
2653 Value *Member = M;
2654 // Void things have no uses we can replace.
2655 if (Member == Leader || Member->getType()->isVoidTy()) {
2656 MembersLeft.insert(Member);
2657 continue;
2658 }
2659 DEBUG(dbgs() << "Found replacement " << *(Leader) << " for " << *Memberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found replacement " << *
(Leader) << " for " << *Member << "\n"; } }
while (false)
2660 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Found replacement " << *
(Leader) << " for " << *Member << "\n"; } }
while (false);
2661 // Due to equality propagation, these may not always be
2662 // instructions, they may be real values. We don't really
2663 // care about trying to replace the non-instructions.
2664 if (auto *I = dyn_cast<Instruction>(Member)) {
2665 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2665, __PRETTY_FUNCTION__));
2666 replaceInstruction(I, Leader);
2667 AnythingReplaced = true;
2668 continue;
2669 } else {
2670 MembersLeft.insert(I);
2671 }
2672 }
2673 CC->Members.swap(MembersLeft);
2674 } else {
2675 DEBUG(dbgs() << "Eliminating in congruence class " << CC->ID << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Eliminating in congruence class "
<< CC->ID << "\n"; } } while (false);
2676 // If this is a singleton, we can skip it.
2677 if (CC->Members.size() != 1) {
2678
2679 // This is a stack because equality replacement/etc may place
2680 // constants in the middle of the member list, and we want to use
2681 // those constant values in preference to the current leader, over
2682 // the scope of those constants.
2683 ValueDFSStack EliminationStack;
2684
2685 // Convert the members to DFS ordered sets and then merge them.
2686 SmallVector<ValueDFS, 8> DFSOrderedSet;
2687 convertClassToDFSOrdered(CC->Members, DFSOrderedSet, UseCounts,
2688 ProbablyDead);
2689
2690 // Sort the whole thing.
2691 std::sort(DFSOrderedSet.begin(), DFSOrderedSet.end());
2692 for (auto &VD : DFSOrderedSet) {
2693 int MemberDFSIn = VD.DFSIn;
2694 int MemberDFSOut = VD.DFSOut;
2695 Value *Def = VD.Def.getPointer();
2696 bool FromStore = VD.Def.getInt();
2697 Use *U = VD.U;
2698 // We ignore void things because we can't get a value from them.
2699 if (Def && Def->getType()->isVoidTy())
2700 continue;
2701
2702 if (EliminationStack.empty()) {
2703 DEBUG(dbgs() << "Elimination Stack is empty\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Elimination Stack is empty\n";
} } while (false);
2704 } else {
2705 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)
2706 << 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)
2707 << 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);
2708 }
2709
2710 DEBUG(dbgs() << "Current DFS numbers are (" << MemberDFSIn << ","do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Current DFS numbers are (" <<
MemberDFSIn << "," << MemberDFSOut << ")\n"
; } } while (false)
2711 << MemberDFSOut << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("newgvn")) { dbgs() << "Current DFS numbers are (" <<
MemberDFSIn << "," << MemberDFSOut << ")\n"
; } } while (false);
2712 // First, we see if we are out of scope or empty. If so,
2713 // and there equivalences, we try to replace the top of
2714 // stack with equivalences (if it's on the stack, it must
2715 // not have been eliminated yet).
2716 // Then we synchronize to our current scope, by
2717 // popping until we are back within a DFS scope that
2718 // dominates the current member.
2719 // Then, what happens depends on a few factors
2720 // If the stack is now empty, we need to push
2721 // If we have a constant or a local equivalence we want to
2722 // start using, we also push.
2723 // Otherwise, we walk along, processing members who are
2724 // dominated by this scope, and eliminate them.
2725 bool ShouldPush = Def && EliminationStack.empty();
2726 bool OutOfScope =
2727 !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut);
2728
2729 if (OutOfScope || ShouldPush) {
2730 // Sync to our current scope.
2731 EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
2732 bool ShouldPush = Def && EliminationStack.empty();
2733 if (ShouldPush) {
2734 EliminationStack.push_back(Def, MemberDFSIn, MemberDFSOut);
2735 }
2736 }
2737
2738 // Skip the Def's, we only want to eliminate on their uses. But mark
2739 // dominated defs as dead.
2740 if (Def) {
2741 // For anything in this case, what and how we value number
2742 // guarantees that any side-effets that would have occurred (ie
2743 // throwing, etc) can be proven to either still occur (because it's
2744 // dominated by something that has the same side-effects), or never
2745 // occur. Otherwise, we would not have been able to prove it value
2746 // equivalent to something else. For these things, we can just mark
2747 // it all dead. Note that this is different from the "ProbablyDead"
2748 // set, which may not be dominated by anything, and thus, are only
2749 // easy to prove dead if they are also side-effect free. Note that
2750 // because stores are put in terms of the stored value, we skip
2751 // stored values here. If the stored value is really dead, it will
2752 // still be marked for deletion when we process it in its own class.
2753 if (!EliminationStack.empty() && Def != EliminationStack.back() &&
2754 isa<Instruction>(Def) && !FromStore)
2755 markInstructionForDeletion(cast<Instruction>(Def));
2756 continue;
2757 }
2758 // At this point, we know it is a Use we are trying to possibly
2759 // replace.
2760
2761 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2762, __PRETTY_FUNCTION__))
2762 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2762, __PRETTY_FUNCTION__));
2763 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2764, __PRETTY_FUNCTION__))
2764 "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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2764, __PRETTY_FUNCTION__));
2765
2766 // If the thing we are replacing into is already marked to be dead,
2767 // this use is dead. Note that this is true regardless of whether
2768 // we have anything dominating the use or not. We do this here
2769 // because we are already walking all the uses anyway.
2770 Instruction *InstUse = cast<Instruction>(U->getUser());
2771 if (InstructionsToErase.count(InstUse)) {
2772 auto &UseCount = UseCounts[U->get()];
2773 if (--UseCount == 0) {
2774 ProbablyDead.insert(cast<Instruction>(U->get()));
2775 }
2776 }
2777
2778 // If we get to this point, and the stack is empty we must have a use
2779 // with nothing we can use to eliminate this use, so just skip it.
2780 if (EliminationStack.empty())
2781 continue;
2782
2783 Value *DominatingLeader = EliminationStack.back();
2784
2785 // Don't replace our existing users with ourselves.
2786 if (U->get() == DominatingLeader)
2787 continue;
2788 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)
2789 << *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);
2790
2791 // If we replaced something in an instruction, handle the patching of
2792 // metadata. Skip this if we are replacing predicateinfo with its
2793 // original operand, as we already know we can just drop it.
2794 auto *ReplacedInst = cast<Instruction>(U->get());
2795 auto *PI = PredInfo->getPredicateInfoFor(ReplacedInst);
2796 if (!PI || DominatingLeader != PI->OriginalOp)
2797 patchReplacementInstruction(ReplacedInst, DominatingLeader);
2798 U->set(DominatingLeader);
2799 // This is now a use of the dominating leader, which means if the
2800 // dominating leader was dead, it's now live!
2801 auto &LeaderUseCount = UseCounts[DominatingLeader];
2802 // It's about to be alive again.
2803 if (LeaderUseCount == 0 && isa<Instruction>(DominatingLeader))
2804 ProbablyDead.erase(cast<Instruction>(DominatingLeader));
2805 ++LeaderUseCount;
2806 AnythingReplaced = true;
2807 }
2808 }
2809 }
2810
2811 // At this point, anything still in the ProbablyDead set is actually dead if
2812 // would be trivially dead.
2813 for (auto *I : ProbablyDead)
2814 if (wouldInstructionBeTriviallyDead(I))
2815 markInstructionForDeletion(I);
2816
2817 // Cleanup the congruence class.
2818 SmallPtrSet<Value *, 4> MembersLeft;
2819 for (Value *Member : CC->Members) {
2820 if (Member->getType()->isVoidTy()) {
2821 MembersLeft.insert(Member);
2822 continue;
2823 }
2824
2825 MembersLeft.insert(Member);
2826 }
2827 CC->Members.swap(MembersLeft);
2828
2829 // If we have possible dead stores to look at, try to eliminate them.
2830 if (CC->StoreCount > 0) {
2831 convertClassToLoadsAndStores(CC->Members, PossibleDeadStores);
2832 std::sort(PossibleDeadStores.begin(), PossibleDeadStores.end());
2833 ValueDFSStack EliminationStack;
2834 for (auto &VD : PossibleDeadStores) {
2835 int MemberDFSIn = VD.DFSIn;
2836 int MemberDFSOut = VD.DFSOut;
2837 Instruction *Member = cast<Instruction>(VD.Def.getPointer());
2838 if (EliminationStack.empty() ||
2839 !EliminationStack.isInScope(MemberDFSIn, MemberDFSOut)) {
2840 // Sync to our current scope.
2841 EliminationStack.popUntilDFSScope(MemberDFSIn, MemberDFSOut);
2842 if (EliminationStack.empty()) {
2843 EliminationStack.push_back(Member, MemberDFSIn, MemberDFSOut);
2844 continue;
2845 }
2846 }
2847 // We already did load elimination, so nothing to do here.
2848 if (isa<LoadInst>(Member))
2849 continue;
2850 assert(!EliminationStack.empty())((!EliminationStack.empty()) ? static_cast<void> (0) : __assert_fail
("!EliminationStack.empty()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2850, __PRETTY_FUNCTION__));
2851 Instruction *Leader = cast<Instruction>(EliminationStack.back());
2852 (void)Leader;
2853 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~svn299582/lib/Transforms/Scalar/NewGVN.cpp"
, 2853, __PRETTY_FUNCTION__));
2854 // Member is dominater by Leader, and thus dead
2855 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)
2856 << " 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);
2857 markInstructionForDeletion(Member);
2858 CC->Members.erase(Member);
2859 ++NumGVNDeadStores;
2860 }
2861 }
2862 }
2863
2864 return AnythingReplaced;
2865}
2866
2867// This function provides global ranking of operations so that we can place them
2868// in a canonical order. Note that rank alone is not necessarily enough for a
2869// complete ordering, as constants all have the same rank. However, generally,
2870// we will simplify an operation with all constants so that it doesn't matter
2871// what order they appear in.
2872unsigned int NewGVN::getRank(const Value *V) const {
2873 // Prefer undef to anything else
2874 if (isa<UndefValue>(V))
2875 return 0;
2876 if (isa<Constant>(V))
2877 return 1;
2878 else if (auto *A = dyn_cast<Argument>(V))
2879 return 2 + A->getArgNo();
2880
2881 // Need to shift the instruction DFS by number of arguments + 3 to account for
2882 // the constant and argument ranking above.
2883 unsigned Result = InstrDFS.lookup(V);
2884 if (Result > 0)
2885 return 3 + NumFuncArgs + Result;
2886 // Unreachable or something else, just return a really large number.
2887 return ~0;
2888}
2889
2890// This is a function that says whether two commutative operations should
2891// have their order swapped when canonicalizing.
2892bool NewGVN::shouldSwapOperands(const Value *A, const Value *B) const {
2893 // Because we only care about a total ordering, and don't rewrite expressions
2894 // in this order, we order by rank, which will give a strict weak ordering to
2895 // everything but constants, and then we order by pointer address.
2896 return std::make_pair(getRank(A), A) > std::make_pair(getRank(B), B);
2897}
2898
2899class NewGVNLegacyPass : public FunctionPass {
2900public:
2901 static char ID; // Pass identification, replacement for typeid.
2902 NewGVNLegacyPass() : FunctionPass(ID) {
2903 initializeNewGVNLegacyPassPass(*PassRegistry::getPassRegistry());
2904 }
2905 bool runOnFunction(Function &F) override;
2906
2907private:
2908 void getAnalysisUsage(AnalysisUsage &AU) const override {
2909 AU.addRequired<AssumptionCacheTracker>();
2910 AU.addRequired<DominatorTreeWrapperPass>();
2911 AU.addRequired<TargetLibraryInfoWrapperPass>();
2912 AU.addRequired<MemorySSAWrapperPass>();
2913 AU.addRequired<AAResultsWrapperPass>();
2914 AU.addPreserved<DominatorTreeWrapperPass>();
2915 AU.addPreserved<GlobalsAAWrapperPass>();
2916 }
2917};
2918
2919bool NewGVNLegacyPass::runOnFunction(Function &F) {
2920 if (skipFunction(F))
2921 return false;
2922 return NewGVN(F, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2923 &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2924 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2925 &getAnalysis<AAResultsWrapperPass>().getAAResults(),
2926 &getAnalysis<MemorySSAWrapperPass>().getMSSA(),
2927 F.getParent()->getDataLayout())
2928 .runGVN();
2929}
2930
2931INITIALIZE_PASS_BEGIN(NewGVNLegacyPass, "newgvn", "Global Value Numbering",static void *initializeNewGVNLegacyPassPassOnce(PassRegistry &
Registry) {
2932 false, false)static void *initializeNewGVNLegacyPassPassOnce(PassRegistry &
Registry) {
2933INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
2934INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry);
2935INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
2936INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
2937INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
2938INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry);
2939INITIALIZE_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)); }
2940 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)); }
2941
2942char NewGVNLegacyPass::ID = 0;
2943
2944// createGVNPass - The public interface to this file.
2945FunctionPass *llvm::createNewGVNPass() { return new NewGVNLegacyPass(); }
2946
2947PreservedAnalyses NewGVNPass::run(Function &F, AnalysisManager<Function> &AM) {
2948 // Apparently the order in which we get these results matter for
2949 // the old GVN (see Chandler's comment in GVN.cpp). I'll keep
2950 // the same order here, just in case.
2951 auto &AC = AM.getResult<AssumptionAnalysis>(F);
2952 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
2953 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
2954 auto &AA = AM.getResult<AAManager>(F);
2955 auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
2956 bool Changed =
2957 NewGVN(F, &DT, &AC, &TLI, &AA, &MSSA, F.getParent()->getDataLayout())
2958 .runGVN();
2959 if (!Changed)
2960 return PreservedAnalyses::all();
2961 PreservedAnalyses PA;
2962 PA.preserve<DominatorTreeAnalysis>();
2963 PA.preserve<GlobalsAA>();
2964 return PA;
2965}