1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
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 //
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
12 //
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
15 //
16 //===----------------------------------------------------------------------===//
17
18 #include "llvm/Transforms/Scalar/GVN.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/Hashing.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AssumptionCache.h"
29 #include "llvm/Analysis/CFG.h"
30 #include "llvm/Analysis/ConstantFolding.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/InstructionSimplify.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/MemoryBuiltins.h"
35 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
36 #include "llvm/Analysis/PHITransAddr.h"
37 #include "llvm/Analysis/TargetLibraryInfo.h"
38 #include "llvm/Analysis/ValueTracking.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/IntrinsicInst.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/Metadata.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 #include "llvm/Transforms/Utils/Local.h"
52 #include "llvm/Transforms/Utils/SSAUpdater.h"
53 #include <vector>
54 using namespace llvm;
55 using namespace llvm::gvn;
56 using namespace PatternMatch;
57
58 #define DEBUG_TYPE "gvn"
59
60 STATISTIC(NumGVNInstr, "Number of instructions deleted");
61 STATISTIC(NumGVNLoad, "Number of loads deleted");
62 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
63 STATISTIC(NumGVNBlocks, "Number of blocks merged");
64 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
65 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
66 STATISTIC(NumPRELoad, "Number of loads PRE'd");
67
68 static cl::opt<bool> EnablePRE("enable-pre",
69 cl::init(true), cl::Hidden);
70 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
71
72 // Maximum allowed recursion depth.
73 static cl::opt<uint32_t>
74 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
75 cl::desc("Max recurse depth (default = 1000)"));
76
77 struct llvm::GVN::Expression {
78 uint32_t opcode;
79 Type *type;
80 SmallVector<uint32_t, 4> varargs;
81
Expressionllvm::GVN::Expression82 Expression(uint32_t o = ~2U) : opcode(o) {}
83
operator ==llvm::GVN::Expression84 bool operator==(const Expression &other) const {
85 if (opcode != other.opcode)
86 return false;
87 if (opcode == ~0U || opcode == ~1U)
88 return true;
89 if (type != other.type)
90 return false;
91 if (varargs != other.varargs)
92 return false;
93 return true;
94 }
95
hash_value(const Expression & Value)96 friend hash_code hash_value(const Expression &Value) {
97 return hash_combine(
98 Value.opcode, Value.type,
99 hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
100 }
101 };
102
103 namespace llvm {
104 template <> struct DenseMapInfo<GVN::Expression> {
getEmptyKeyllvm::DenseMapInfo105 static inline GVN::Expression getEmptyKey() { return ~0U; }
106
getTombstoneKeyllvm::DenseMapInfo107 static inline GVN::Expression getTombstoneKey() { return ~1U; }
108
getHashValuellvm::DenseMapInfo109 static unsigned getHashValue(const GVN::Expression &e) {
110 using llvm::hash_value;
111 return static_cast<unsigned>(hash_value(e));
112 }
isEqualllvm::DenseMapInfo113 static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
114 return LHS == RHS;
115 }
116 };
117 } // End llvm namespace.
118
119 /// Represents a particular available value that we know how to materialize.
120 /// Materialization of an AvailableValue never fails. An AvailableValue is
121 /// implicitly associated with a rematerialization point which is the
122 /// location of the instruction from which it was formed.
123 struct llvm::gvn::AvailableValue {
124 enum ValType {
125 SimpleVal, // A simple offsetted value that is accessed.
126 LoadVal, // A value produced by a load.
127 MemIntrin, // A memory intrinsic which is loaded from.
128 UndefVal // A UndefValue representing a value from dead block (which
129 // is not yet physically removed from the CFG).
130 };
131
132 /// V - The value that is live out of the block.
133 PointerIntPair<Value *, 2, ValType> Val;
134
135 /// Offset - The byte offset in Val that is interesting for the load query.
136 unsigned Offset;
137
getllvm::gvn::AvailableValue138 static AvailableValue get(Value *V, unsigned Offset = 0) {
139 AvailableValue Res;
140 Res.Val.setPointer(V);
141 Res.Val.setInt(SimpleVal);
142 Res.Offset = Offset;
143 return Res;
144 }
145
getMIllvm::gvn::AvailableValue146 static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
147 AvailableValue Res;
148 Res.Val.setPointer(MI);
149 Res.Val.setInt(MemIntrin);
150 Res.Offset = Offset;
151 return Res;
152 }
153
getLoadllvm::gvn::AvailableValue154 static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
155 AvailableValue Res;
156 Res.Val.setPointer(LI);
157 Res.Val.setInt(LoadVal);
158 Res.Offset = Offset;
159 return Res;
160 }
161
getUndefllvm::gvn::AvailableValue162 static AvailableValue getUndef() {
163 AvailableValue Res;
164 Res.Val.setPointer(nullptr);
165 Res.Val.setInt(UndefVal);
166 Res.Offset = 0;
167 return Res;
168 }
169
isSimpleValuellvm::gvn::AvailableValue170 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
isCoercedLoadValuellvm::gvn::AvailableValue171 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
isMemIntrinValuellvm::gvn::AvailableValue172 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
isUndefValuellvm::gvn::AvailableValue173 bool isUndefValue() const { return Val.getInt() == UndefVal; }
174
getSimpleValuellvm::gvn::AvailableValue175 Value *getSimpleValue() const {
176 assert(isSimpleValue() && "Wrong accessor");
177 return Val.getPointer();
178 }
179
getCoercedLoadValuellvm::gvn::AvailableValue180 LoadInst *getCoercedLoadValue() const {
181 assert(isCoercedLoadValue() && "Wrong accessor");
182 return cast<LoadInst>(Val.getPointer());
183 }
184
getMemIntrinValuellvm::gvn::AvailableValue185 MemIntrinsic *getMemIntrinValue() const {
186 assert(isMemIntrinValue() && "Wrong accessor");
187 return cast<MemIntrinsic>(Val.getPointer());
188 }
189
190 /// Emit code at the specified insertion point to adjust the value defined
191 /// here to the specified type. This handles various coercion cases.
192 Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
193 GVN &gvn) const;
194 };
195
196 /// Represents an AvailableValue which can be rematerialized at the end of
197 /// the associated BasicBlock.
198 struct llvm::gvn::AvailableValueInBlock {
199 /// BB - The basic block in question.
200 BasicBlock *BB;
201
202 /// AV - The actual available value
203 AvailableValue AV;
204
getllvm::gvn::AvailableValueInBlock205 static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) {
206 AvailableValueInBlock Res;
207 Res.BB = BB;
208 Res.AV = std::move(AV);
209 return Res;
210 }
211
getllvm::gvn::AvailableValueInBlock212 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
213 unsigned Offset = 0) {
214 return get(BB, AvailableValue::get(V, Offset));
215 }
getUndefllvm::gvn::AvailableValueInBlock216 static AvailableValueInBlock getUndef(BasicBlock *BB) {
217 return get(BB, AvailableValue::getUndef());
218 }
219
220 /// Emit code at the end of this block to adjust the value defined here to
221 /// the specified type. This handles various coercion cases.
MaterializeAdjustedValuellvm::gvn::AvailableValueInBlock222 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const {
223 return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
224 }
225 };
226
227 //===----------------------------------------------------------------------===//
228 // ValueTable Internal Functions
229 //===----------------------------------------------------------------------===//
230
createExpr(Instruction * I)231 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
232 Expression e;
233 e.type = I->getType();
234 e.opcode = I->getOpcode();
235 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
236 OI != OE; ++OI)
237 e.varargs.push_back(lookupOrAdd(*OI));
238 if (I->isCommutative()) {
239 // Ensure that commutative instructions that only differ by a permutation
240 // of their operands get the same value number by sorting the operand value
241 // numbers. Since all commutative instructions have two operands it is more
242 // efficient to sort by hand rather than using, say, std::sort.
243 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
244 if (e.varargs[0] > e.varargs[1])
245 std::swap(e.varargs[0], e.varargs[1]);
246 }
247
248 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
249 // Sort the operand value numbers so x<y and y>x get the same value number.
250 CmpInst::Predicate Predicate = C->getPredicate();
251 if (e.varargs[0] > e.varargs[1]) {
252 std::swap(e.varargs[0], e.varargs[1]);
253 Predicate = CmpInst::getSwappedPredicate(Predicate);
254 }
255 e.opcode = (C->getOpcode() << 8) | Predicate;
256 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
257 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
258 II != IE; ++II)
259 e.varargs.push_back(*II);
260 }
261
262 return e;
263 }
264
createCmpExpr(unsigned Opcode,CmpInst::Predicate Predicate,Value * LHS,Value * RHS)265 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
266 CmpInst::Predicate Predicate,
267 Value *LHS, Value *RHS) {
268 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
269 "Not a comparison!");
270 Expression e;
271 e.type = CmpInst::makeCmpResultType(LHS->getType());
272 e.varargs.push_back(lookupOrAdd(LHS));
273 e.varargs.push_back(lookupOrAdd(RHS));
274
275 // Sort the operand value numbers so x<y and y>x get the same value number.
276 if (e.varargs[0] > e.varargs[1]) {
277 std::swap(e.varargs[0], e.varargs[1]);
278 Predicate = CmpInst::getSwappedPredicate(Predicate);
279 }
280 e.opcode = (Opcode << 8) | Predicate;
281 return e;
282 }
283
createExtractvalueExpr(ExtractValueInst * EI)284 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
285 assert(EI && "Not an ExtractValueInst?");
286 Expression e;
287 e.type = EI->getType();
288 e.opcode = 0;
289
290 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
291 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
292 // EI might be an extract from one of our recognised intrinsics. If it
293 // is we'll synthesize a semantically equivalent expression instead on
294 // an extract value expression.
295 switch (I->getIntrinsicID()) {
296 case Intrinsic::sadd_with_overflow:
297 case Intrinsic::uadd_with_overflow:
298 e.opcode = Instruction::Add;
299 break;
300 case Intrinsic::ssub_with_overflow:
301 case Intrinsic::usub_with_overflow:
302 e.opcode = Instruction::Sub;
303 break;
304 case Intrinsic::smul_with_overflow:
305 case Intrinsic::umul_with_overflow:
306 e.opcode = Instruction::Mul;
307 break;
308 default:
309 break;
310 }
311
312 if (e.opcode != 0) {
313 // Intrinsic recognized. Grab its args to finish building the expression.
314 assert(I->getNumArgOperands() == 2 &&
315 "Expect two args for recognised intrinsics.");
316 e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
317 e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
318 return e;
319 }
320 }
321
322 // Not a recognised intrinsic. Fall back to producing an extract value
323 // expression.
324 e.opcode = EI->getOpcode();
325 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
326 OI != OE; ++OI)
327 e.varargs.push_back(lookupOrAdd(*OI));
328
329 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
330 II != IE; ++II)
331 e.varargs.push_back(*II);
332
333 return e;
334 }
335
336 //===----------------------------------------------------------------------===//
337 // ValueTable External Functions
338 //===----------------------------------------------------------------------===//
339
ValueTable()340 GVN::ValueTable::ValueTable() : nextValueNumber(1) {}
ValueTable(const ValueTable & Arg)341 GVN::ValueTable::ValueTable(const ValueTable &Arg)
342 : valueNumbering(Arg.valueNumbering),
343 expressionNumbering(Arg.expressionNumbering), AA(Arg.AA), MD(Arg.MD),
344 DT(Arg.DT), nextValueNumber(Arg.nextValueNumber) {}
ValueTable(ValueTable && Arg)345 GVN::ValueTable::ValueTable(ValueTable &&Arg)
346 : valueNumbering(std::move(Arg.valueNumbering)),
347 expressionNumbering(std::move(Arg.expressionNumbering)),
348 AA(std::move(Arg.AA)), MD(std::move(Arg.MD)), DT(std::move(Arg.DT)),
349 nextValueNumber(std::move(Arg.nextValueNumber)) {}
~ValueTable()350 GVN::ValueTable::~ValueTable() {}
351
352 /// add - Insert a value into the table with a specified value number.
add(Value * V,uint32_t num)353 void GVN::ValueTable::add(Value *V, uint32_t num) {
354 valueNumbering.insert(std::make_pair(V, num));
355 }
356
lookupOrAddCall(CallInst * C)357 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
358 if (AA->doesNotAccessMemory(C)) {
359 Expression exp = createExpr(C);
360 uint32_t &e = expressionNumbering[exp];
361 if (!e) e = nextValueNumber++;
362 valueNumbering[C] = e;
363 return e;
364 } else if (AA->onlyReadsMemory(C)) {
365 Expression exp = createExpr(C);
366 uint32_t &e = expressionNumbering[exp];
367 if (!e) {
368 e = nextValueNumber++;
369 valueNumbering[C] = e;
370 return e;
371 }
372 if (!MD) {
373 e = nextValueNumber++;
374 valueNumbering[C] = e;
375 return e;
376 }
377
378 MemDepResult local_dep = MD->getDependency(C);
379
380 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
381 valueNumbering[C] = nextValueNumber;
382 return nextValueNumber++;
383 }
384
385 if (local_dep.isDef()) {
386 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
387
388 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
389 valueNumbering[C] = nextValueNumber;
390 return nextValueNumber++;
391 }
392
393 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
394 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
395 uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
396 if (c_vn != cd_vn) {
397 valueNumbering[C] = nextValueNumber;
398 return nextValueNumber++;
399 }
400 }
401
402 uint32_t v = lookupOrAdd(local_cdep);
403 valueNumbering[C] = v;
404 return v;
405 }
406
407 // Non-local case.
408 const MemoryDependenceResults::NonLocalDepInfo &deps =
409 MD->getNonLocalCallDependency(CallSite(C));
410 // FIXME: Move the checking logic to MemDep!
411 CallInst* cdep = nullptr;
412
413 // Check to see if we have a single dominating call instruction that is
414 // identical to C.
415 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
416 const NonLocalDepEntry *I = &deps[i];
417 if (I->getResult().isNonLocal())
418 continue;
419
420 // We don't handle non-definitions. If we already have a call, reject
421 // instruction dependencies.
422 if (!I->getResult().isDef() || cdep != nullptr) {
423 cdep = nullptr;
424 break;
425 }
426
427 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
428 // FIXME: All duplicated with non-local case.
429 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
430 cdep = NonLocalDepCall;
431 continue;
432 }
433
434 cdep = nullptr;
435 break;
436 }
437
438 if (!cdep) {
439 valueNumbering[C] = nextValueNumber;
440 return nextValueNumber++;
441 }
442
443 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
444 valueNumbering[C] = nextValueNumber;
445 return nextValueNumber++;
446 }
447 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
448 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
449 uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
450 if (c_vn != cd_vn) {
451 valueNumbering[C] = nextValueNumber;
452 return nextValueNumber++;
453 }
454 }
455
456 uint32_t v = lookupOrAdd(cdep);
457 valueNumbering[C] = v;
458 return v;
459
460 } else {
461 valueNumbering[C] = nextValueNumber;
462 return nextValueNumber++;
463 }
464 }
465
466 /// Returns true if a value number exists for the specified value.
exists(Value * V) const467 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
468
469 /// lookup_or_add - Returns the value number for the specified value, assigning
470 /// it a new number if it did not have one before.
lookupOrAdd(Value * V)471 uint32_t GVN::ValueTable::lookupOrAdd(Value *V) {
472 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
473 if (VI != valueNumbering.end())
474 return VI->second;
475
476 if (!isa<Instruction>(V)) {
477 valueNumbering[V] = nextValueNumber;
478 return nextValueNumber++;
479 }
480
481 Instruction* I = cast<Instruction>(V);
482 Expression exp;
483 switch (I->getOpcode()) {
484 case Instruction::Call:
485 return lookupOrAddCall(cast<CallInst>(I));
486 case Instruction::Add:
487 case Instruction::FAdd:
488 case Instruction::Sub:
489 case Instruction::FSub:
490 case Instruction::Mul:
491 case Instruction::FMul:
492 case Instruction::UDiv:
493 case Instruction::SDiv:
494 case Instruction::FDiv:
495 case Instruction::URem:
496 case Instruction::SRem:
497 case Instruction::FRem:
498 case Instruction::Shl:
499 case Instruction::LShr:
500 case Instruction::AShr:
501 case Instruction::And:
502 case Instruction::Or:
503 case Instruction::Xor:
504 case Instruction::ICmp:
505 case Instruction::FCmp:
506 case Instruction::Trunc:
507 case Instruction::ZExt:
508 case Instruction::SExt:
509 case Instruction::FPToUI:
510 case Instruction::FPToSI:
511 case Instruction::UIToFP:
512 case Instruction::SIToFP:
513 case Instruction::FPTrunc:
514 case Instruction::FPExt:
515 case Instruction::PtrToInt:
516 case Instruction::IntToPtr:
517 case Instruction::BitCast:
518 case Instruction::Select:
519 case Instruction::ExtractElement:
520 case Instruction::InsertElement:
521 case Instruction::ShuffleVector:
522 case Instruction::InsertValue:
523 case Instruction::GetElementPtr:
524 exp = createExpr(I);
525 break;
526 case Instruction::ExtractValue:
527 exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
528 break;
529 default:
530 valueNumbering[V] = nextValueNumber;
531 return nextValueNumber++;
532 }
533
534 uint32_t& e = expressionNumbering[exp];
535 if (!e) e = nextValueNumber++;
536 valueNumbering[V] = e;
537 return e;
538 }
539
540 /// Returns the value number of the specified value. Fails if
541 /// the value has not yet been numbered.
lookup(Value * V) const542 uint32_t GVN::ValueTable::lookup(Value *V) const {
543 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
544 assert(VI != valueNumbering.end() && "Value not numbered?");
545 return VI->second;
546 }
547
548 /// Returns the value number of the given comparison,
549 /// assigning it a new number if it did not have one before. Useful when
550 /// we deduced the result of a comparison, but don't immediately have an
551 /// instruction realizing that comparison to hand.
lookupOrAddCmp(unsigned Opcode,CmpInst::Predicate Predicate,Value * LHS,Value * RHS)552 uint32_t GVN::ValueTable::lookupOrAddCmp(unsigned Opcode,
553 CmpInst::Predicate Predicate,
554 Value *LHS, Value *RHS) {
555 Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
556 uint32_t& e = expressionNumbering[exp];
557 if (!e) e = nextValueNumber++;
558 return e;
559 }
560
561 /// Remove all entries from the ValueTable.
clear()562 void GVN::ValueTable::clear() {
563 valueNumbering.clear();
564 expressionNumbering.clear();
565 nextValueNumber = 1;
566 }
567
568 /// Remove a value from the value numbering.
erase(Value * V)569 void GVN::ValueTable::erase(Value *V) {
570 valueNumbering.erase(V);
571 }
572
573 /// verifyRemoved - Verify that the value is removed from all internal data
574 /// structures.
verifyRemoved(const Value * V) const575 void GVN::ValueTable::verifyRemoved(const Value *V) const {
576 for (DenseMap<Value*, uint32_t>::const_iterator
577 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
578 assert(I->first != V && "Inst still occurs in value numbering map!");
579 }
580 }
581
582 //===----------------------------------------------------------------------===//
583 // GVN Pass
584 //===----------------------------------------------------------------------===//
585
run(Function & F,AnalysisManager<Function> & AM)586 PreservedAnalyses GVN::run(Function &F, AnalysisManager<Function> &AM) {
587 // FIXME: The order of evaluation of these 'getResult' calls is very
588 // significant! Re-ordering these variables will cause GVN when run alone to
589 // be less effective! We should fix memdep and basic-aa to not exhibit this
590 // behavior, but until then don't change the order here.
591 auto &AC = AM.getResult<AssumptionAnalysis>(F);
592 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
593 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
594 auto &AA = AM.getResult<AAManager>(F);
595 auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
596 bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep);
597 if (!Changed)
598 return PreservedAnalyses::all();
599 PreservedAnalyses PA;
600 PA.preserve<DominatorTreeAnalysis>();
601 PA.preserve<GlobalsAA>();
602 return PA;
603 }
604
605 LLVM_DUMP_METHOD
dump(DenseMap<uint32_t,Value * > & d)606 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
607 errs() << "{\n";
608 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
609 E = d.end(); I != E; ++I) {
610 errs() << I->first << "\n";
611 I->second->dump();
612 }
613 errs() << "}\n";
614 }
615
616 /// Return true if we can prove that the value
617 /// we're analyzing is fully available in the specified block. As we go, keep
618 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
619 /// map is actually a tri-state map with the following values:
620 /// 0) we know the block *is not* fully available.
621 /// 1) we know the block *is* fully available.
622 /// 2) we do not know whether the block is fully available or not, but we are
623 /// currently speculating that it will be.
624 /// 3) we are speculating for this block and have used that to speculate for
625 /// other blocks.
IsValueFullyAvailableInBlock(BasicBlock * BB,DenseMap<BasicBlock *,char> & FullyAvailableBlocks,uint32_t RecurseDepth)626 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
627 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
628 uint32_t RecurseDepth) {
629 if (RecurseDepth > MaxRecurseDepth)
630 return false;
631
632 // Optimistically assume that the block is fully available and check to see
633 // if we already know about this block in one lookup.
634 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
635 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
636
637 // If the entry already existed for this block, return the precomputed value.
638 if (!IV.second) {
639 // If this is a speculative "available" value, mark it as being used for
640 // speculation of other blocks.
641 if (IV.first->second == 2)
642 IV.first->second = 3;
643 return IV.first->second != 0;
644 }
645
646 // Otherwise, see if it is fully available in all predecessors.
647 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
648
649 // If this block has no predecessors, it isn't live-in here.
650 if (PI == PE)
651 goto SpeculationFailure;
652
653 for (; PI != PE; ++PI)
654 // If the value isn't fully available in one of our predecessors, then it
655 // isn't fully available in this block either. Undo our previous
656 // optimistic assumption and bail out.
657 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
658 goto SpeculationFailure;
659
660 return true;
661
662 // If we get here, we found out that this is not, after
663 // all, a fully-available block. We have a problem if we speculated on this and
664 // used the speculation to mark other blocks as available.
665 SpeculationFailure:
666 char &BBVal = FullyAvailableBlocks[BB];
667
668 // If we didn't speculate on this, just return with it set to false.
669 if (BBVal == 2) {
670 BBVal = 0;
671 return false;
672 }
673
674 // If we did speculate on this value, we could have blocks set to 1 that are
675 // incorrect. Walk the (transitive) successors of this block and mark them as
676 // 0 if set to one.
677 SmallVector<BasicBlock*, 32> BBWorklist;
678 BBWorklist.push_back(BB);
679
680 do {
681 BasicBlock *Entry = BBWorklist.pop_back_val();
682 // Note that this sets blocks to 0 (unavailable) if they happen to not
683 // already be in FullyAvailableBlocks. This is safe.
684 char &EntryVal = FullyAvailableBlocks[Entry];
685 if (EntryVal == 0) continue; // Already unavailable.
686
687 // Mark as unavailable.
688 EntryVal = 0;
689
690 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
691 } while (!BBWorklist.empty());
692
693 return false;
694 }
695
696
697 /// Return true if CoerceAvailableValueToLoadType will succeed.
CanCoerceMustAliasedValueToLoad(Value * StoredVal,Type * LoadTy,const DataLayout & DL)698 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
699 Type *LoadTy,
700 const DataLayout &DL) {
701 // If the loaded or stored value is an first class array or struct, don't try
702 // to transform them. We need to be able to bitcast to integer.
703 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
704 StoredVal->getType()->isStructTy() ||
705 StoredVal->getType()->isArrayTy())
706 return false;
707
708 // The store has to be at least as big as the load.
709 if (DL.getTypeSizeInBits(StoredVal->getType()) <
710 DL.getTypeSizeInBits(LoadTy))
711 return false;
712
713 return true;
714 }
715
716 /// If we saw a store of a value to memory, and
717 /// then a load from a must-aliased pointer of a different type, try to coerce
718 /// the stored value. LoadedTy is the type of the load we want to replace.
719 /// IRB is IRBuilder used to insert new instructions.
720 ///
721 /// If we can't do it, return null.
CoerceAvailableValueToLoadType(Value * StoredVal,Type * LoadedTy,IRBuilder<> & IRB,const DataLayout & DL)722 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
723 IRBuilder<> &IRB,
724 const DataLayout &DL) {
725 assert(CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL) &&
726 "precondition violation - materialization can't fail");
727
728 // If this is already the right type, just return it.
729 Type *StoredValTy = StoredVal->getType();
730
731 uint64_t StoredValSize = DL.getTypeSizeInBits(StoredValTy);
732 uint64_t LoadedValSize = DL.getTypeSizeInBits(LoadedTy);
733
734 // If the store and reload are the same size, we can always reuse it.
735 if (StoredValSize == LoadedValSize) {
736 // Pointer to Pointer -> use bitcast.
737 if (StoredValTy->getScalarType()->isPointerTy() &&
738 LoadedTy->getScalarType()->isPointerTy())
739 return IRB.CreateBitCast(StoredVal, LoadedTy);
740
741 // Convert source pointers to integers, which can be bitcast.
742 if (StoredValTy->getScalarType()->isPointerTy()) {
743 StoredValTy = DL.getIntPtrType(StoredValTy);
744 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
745 }
746
747 Type *TypeToCastTo = LoadedTy;
748 if (TypeToCastTo->getScalarType()->isPointerTy())
749 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
750
751 if (StoredValTy != TypeToCastTo)
752 StoredVal = IRB.CreateBitCast(StoredVal, TypeToCastTo);
753
754 // Cast to pointer if the load needs a pointer type.
755 if (LoadedTy->getScalarType()->isPointerTy())
756 StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy);
757
758 return StoredVal;
759 }
760
761 // If the loaded value is smaller than the available value, then we can
762 // extract out a piece from it. If the available value is too small, then we
763 // can't do anything.
764 assert(StoredValSize >= LoadedValSize &&
765 "CanCoerceMustAliasedValueToLoad fail");
766
767 // Convert source pointers to integers, which can be manipulated.
768 if (StoredValTy->getScalarType()->isPointerTy()) {
769 StoredValTy = DL.getIntPtrType(StoredValTy);
770 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
771 }
772
773 // Convert vectors and fp to integer, which can be manipulated.
774 if (!StoredValTy->isIntegerTy()) {
775 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoredValSize);
776 StoredVal = IRB.CreateBitCast(StoredVal, StoredValTy);
777 }
778
779 // If this is a big-endian system, we need to shift the value down to the low
780 // bits so that a truncate will work.
781 if (DL.isBigEndian()) {
782 uint64_t ShiftAmt = DL.getTypeStoreSizeInBits(StoredValTy) -
783 DL.getTypeStoreSizeInBits(LoadedTy);
784 StoredVal = IRB.CreateLShr(StoredVal, ShiftAmt, "tmp");
785 }
786
787 // Truncate the integer to the right size now.
788 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadedValSize);
789 StoredVal = IRB.CreateTrunc(StoredVal, NewIntTy, "trunc");
790
791 if (LoadedTy == NewIntTy)
792 return StoredVal;
793
794 // If the result is a pointer, inttoptr.
795 if (LoadedTy->getScalarType()->isPointerTy())
796 return IRB.CreateIntToPtr(StoredVal, LoadedTy, "inttoptr");
797
798 // Otherwise, bitcast.
799 return IRB.CreateBitCast(StoredVal, LoadedTy, "bitcast");
800 }
801
802 /// This function is called when we have a
803 /// memdep query of a load that ends up being a clobbering memory write (store,
804 /// memset, memcpy, memmove). This means that the write *may* provide bits used
805 /// by the load but we can't be sure because the pointers don't mustalias.
806 ///
807 /// Check this case to see if there is anything more we can do before we give
808 /// up. This returns -1 if we have to give up, or a byte number in the stored
809 /// value of the piece that feeds the load.
AnalyzeLoadFromClobberingWrite(Type * LoadTy,Value * LoadPtr,Value * WritePtr,uint64_t WriteSizeInBits,const DataLayout & DL)810 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
811 Value *WritePtr,
812 uint64_t WriteSizeInBits,
813 const DataLayout &DL) {
814 // If the loaded or stored value is a first class array or struct, don't try
815 // to transform them. We need to be able to bitcast to integer.
816 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
817 return -1;
818
819 int64_t StoreOffset = 0, LoadOffset = 0;
820 Value *StoreBase =
821 GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
822 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
823 if (StoreBase != LoadBase)
824 return -1;
825
826 // If the load and store are to the exact same address, they should have been
827 // a must alias. AA must have gotten confused.
828 // FIXME: Study to see if/when this happens. One case is forwarding a memset
829 // to a load from the base of the memset.
830 #if 0
831 if (LoadOffset == StoreOffset) {
832 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
833 << "Base = " << *StoreBase << "\n"
834 << "Store Ptr = " << *WritePtr << "\n"
835 << "Store Offs = " << StoreOffset << "\n"
836 << "Load Ptr = " << *LoadPtr << "\n";
837 abort();
838 }
839 #endif
840
841 // If the load and store don't overlap at all, the store doesn't provide
842 // anything to the load. In this case, they really don't alias at all, AA
843 // must have gotten confused.
844 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
845
846 if ((WriteSizeInBits & 7) | (LoadSize & 7))
847 return -1;
848 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
849 LoadSize >>= 3;
850
851
852 bool isAAFailure = false;
853 if (StoreOffset < LoadOffset)
854 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
855 else
856 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
857
858 if (isAAFailure) {
859 #if 0
860 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
861 << "Base = " << *StoreBase << "\n"
862 << "Store Ptr = " << *WritePtr << "\n"
863 << "Store Offs = " << StoreOffset << "\n"
864 << "Load Ptr = " << *LoadPtr << "\n";
865 abort();
866 #endif
867 return -1;
868 }
869
870 // If the Load isn't completely contained within the stored bits, we don't
871 // have all the bits to feed it. We could do something crazy in the future
872 // (issue a smaller load then merge the bits in) but this seems unlikely to be
873 // valuable.
874 if (StoreOffset > LoadOffset ||
875 StoreOffset+StoreSize < LoadOffset+LoadSize)
876 return -1;
877
878 // Okay, we can do this transformation. Return the number of bytes into the
879 // store that the load is.
880 return LoadOffset-StoreOffset;
881 }
882
883 /// This function is called when we have a
884 /// memdep query of a load that ends up being a clobbering store.
AnalyzeLoadFromClobberingStore(Type * LoadTy,Value * LoadPtr,StoreInst * DepSI)885 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
886 StoreInst *DepSI) {
887 // Cannot handle reading from store of first-class aggregate yet.
888 if (DepSI->getValueOperand()->getType()->isStructTy() ||
889 DepSI->getValueOperand()->getType()->isArrayTy())
890 return -1;
891
892 const DataLayout &DL = DepSI->getModule()->getDataLayout();
893 Value *StorePtr = DepSI->getPointerOperand();
894 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
895 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
896 StorePtr, StoreSize, DL);
897 }
898
899 /// This function is called when we have a
900 /// memdep query of a load that ends up being clobbered by another load. See if
901 /// the other load can feed into the second load.
AnalyzeLoadFromClobberingLoad(Type * LoadTy,Value * LoadPtr,LoadInst * DepLI,const DataLayout & DL)902 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
903 LoadInst *DepLI, const DataLayout &DL){
904 // Cannot handle reading from store of first-class aggregate yet.
905 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
906 return -1;
907
908 Value *DepPtr = DepLI->getPointerOperand();
909 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
910 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
911 if (R != -1) return R;
912
913 // If we have a load/load clobber an DepLI can be widened to cover this load,
914 // then we should widen it!
915 int64_t LoadOffs = 0;
916 const Value *LoadBase =
917 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, DL);
918 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
919
920 unsigned Size = MemoryDependenceResults::getLoadLoadClobberFullWidthSize(
921 LoadBase, LoadOffs, LoadSize, DepLI);
922 if (Size == 0) return -1;
923
924 // Check non-obvious conditions enforced by MDA which we rely on for being
925 // able to materialize this potentially available value
926 assert(DepLI->isSimple() && "Cannot widen volatile/atomic load!");
927 assert(DepLI->getType()->isIntegerTy() && "Can't widen non-integer load");
928
929 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
930 }
931
932
933
AnalyzeLoadFromClobberingMemInst(Type * LoadTy,Value * LoadPtr,MemIntrinsic * MI,const DataLayout & DL)934 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
935 MemIntrinsic *MI,
936 const DataLayout &DL) {
937 // If the mem operation is a non-constant size, we can't handle it.
938 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
939 if (!SizeCst) return -1;
940 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
941
942 // If this is memset, we just need to see if the offset is valid in the size
943 // of the memset..
944 if (MI->getIntrinsicID() == Intrinsic::memset)
945 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
946 MemSizeInBits, DL);
947
948 // If we have a memcpy/memmove, the only case we can handle is if this is a
949 // copy from constant memory. In that case, we can read directly from the
950 // constant memory.
951 MemTransferInst *MTI = cast<MemTransferInst>(MI);
952
953 Constant *Src = dyn_cast<Constant>(MTI->getSource());
954 if (!Src) return -1;
955
956 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, DL));
957 if (!GV || !GV->isConstant()) return -1;
958
959 // See if the access is within the bounds of the transfer.
960 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
961 MI->getDest(), MemSizeInBits, DL);
962 if (Offset == -1)
963 return Offset;
964
965 unsigned AS = Src->getType()->getPointerAddressSpace();
966 // Otherwise, see if we can constant fold a load from the constant with the
967 // offset applied as appropriate.
968 Src = ConstantExpr::getBitCast(Src,
969 Type::getInt8PtrTy(Src->getContext(), AS));
970 Constant *OffsetCst =
971 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
972 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
973 OffsetCst);
974 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
975 if (ConstantFoldLoadFromConstPtr(Src, LoadTy, DL))
976 return Offset;
977 return -1;
978 }
979
980
981 /// This function is called when we have a
982 /// memdep query of a load that ends up being a clobbering store. This means
983 /// that the store provides bits used by the load but we the pointers don't
984 /// mustalias. Check this case to see if there is anything more we can do
985 /// before we give up.
GetStoreValueForLoad(Value * SrcVal,unsigned Offset,Type * LoadTy,Instruction * InsertPt,const DataLayout & DL)986 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
987 Type *LoadTy,
988 Instruction *InsertPt, const DataLayout &DL){
989 LLVMContext &Ctx = SrcVal->getType()->getContext();
990
991 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
992 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
993
994 IRBuilder<> Builder(InsertPt);
995
996 // Compute which bits of the stored value are being used by the load. Convert
997 // to an integer type to start with.
998 if (SrcVal->getType()->getScalarType()->isPointerTy())
999 SrcVal = Builder.CreatePtrToInt(SrcVal,
1000 DL.getIntPtrType(SrcVal->getType()));
1001 if (!SrcVal->getType()->isIntegerTy())
1002 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1003
1004 // Shift the bits to the least significant depending on endianness.
1005 unsigned ShiftAmt;
1006 if (DL.isLittleEndian())
1007 ShiftAmt = Offset*8;
1008 else
1009 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1010
1011 if (ShiftAmt)
1012 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1013
1014 if (LoadSize != StoreSize)
1015 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1016
1017 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
1018 }
1019
1020 /// This function is called when we have a
1021 /// memdep query of a load that ends up being a clobbering load. This means
1022 /// that the load *may* provide bits used by the load but we can't be sure
1023 /// because the pointers don't mustalias. Check this case to see if there is
1024 /// anything more we can do before we give up.
GetLoadValueForLoad(LoadInst * SrcVal,unsigned Offset,Type * LoadTy,Instruction * InsertPt,GVN & gvn)1025 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1026 Type *LoadTy, Instruction *InsertPt,
1027 GVN &gvn) {
1028 const DataLayout &DL = SrcVal->getModule()->getDataLayout();
1029 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1030 // widen SrcVal out to a larger load.
1031 unsigned SrcValStoreSize = DL.getTypeStoreSize(SrcVal->getType());
1032 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1033 if (Offset+LoadSize > SrcValStoreSize) {
1034 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1035 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1036 // If we have a load/load clobber an DepLI can be widened to cover this
1037 // load, then we should widen it to the next power of 2 size big enough!
1038 unsigned NewLoadSize = Offset+LoadSize;
1039 if (!isPowerOf2_32(NewLoadSize))
1040 NewLoadSize = NextPowerOf2(NewLoadSize);
1041
1042 Value *PtrVal = SrcVal->getPointerOperand();
1043
1044 // Insert the new load after the old load. This ensures that subsequent
1045 // memdep queries will find the new load. We can't easily remove the old
1046 // load completely because it is already in the value numbering table.
1047 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1048 Type *DestPTy =
1049 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1050 DestPTy = PointerType::get(DestPTy,
1051 PtrVal->getType()->getPointerAddressSpace());
1052 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1053 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1054 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1055 NewLoad->takeName(SrcVal);
1056 NewLoad->setAlignment(SrcVal->getAlignment());
1057
1058 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1059 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1060
1061 // Replace uses of the original load with the wider load. On a big endian
1062 // system, we need to shift down to get the relevant bits.
1063 Value *RV = NewLoad;
1064 if (DL.isBigEndian())
1065 RV = Builder.CreateLShr(RV, (NewLoadSize - SrcValStoreSize) * 8);
1066 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1067 SrcVal->replaceAllUsesWith(RV);
1068
1069 // We would like to use gvn.markInstructionForDeletion here, but we can't
1070 // because the load is already memoized into the leader map table that GVN
1071 // tracks. It is potentially possible to remove the load from the table,
1072 // but then there all of the operations based on it would need to be
1073 // rehashed. Just leave the dead load around.
1074 gvn.getMemDep().removeInstruction(SrcVal);
1075 SrcVal = NewLoad;
1076 }
1077
1078 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1079 }
1080
1081
1082 /// This function is called when we have a
1083 /// memdep query of a load that ends up being a clobbering mem intrinsic.
GetMemInstValueForLoad(MemIntrinsic * SrcInst,unsigned Offset,Type * LoadTy,Instruction * InsertPt,const DataLayout & DL)1084 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1085 Type *LoadTy, Instruction *InsertPt,
1086 const DataLayout &DL){
1087 LLVMContext &Ctx = LoadTy->getContext();
1088 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1089
1090 IRBuilder<> Builder(InsertPt);
1091
1092 // We know that this method is only called when the mem transfer fully
1093 // provides the bits for the load.
1094 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1095 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1096 // independently of what the offset is.
1097 Value *Val = MSI->getValue();
1098 if (LoadSize != 1)
1099 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1100
1101 Value *OneElt = Val;
1102
1103 // Splat the value out to the right number of bits.
1104 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1105 // If we can double the number of bytes set, do it.
1106 if (NumBytesSet*2 <= LoadSize) {
1107 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1108 Val = Builder.CreateOr(Val, ShVal);
1109 NumBytesSet <<= 1;
1110 continue;
1111 }
1112
1113 // Otherwise insert one byte at a time.
1114 Value *ShVal = Builder.CreateShl(Val, 1*8);
1115 Val = Builder.CreateOr(OneElt, ShVal);
1116 ++NumBytesSet;
1117 }
1118
1119 return CoerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
1120 }
1121
1122 // Otherwise, this is a memcpy/memmove from a constant global.
1123 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1124 Constant *Src = cast<Constant>(MTI->getSource());
1125 unsigned AS = Src->getType()->getPointerAddressSpace();
1126
1127 // Otherwise, see if we can constant fold a load from the constant with the
1128 // offset applied as appropriate.
1129 Src = ConstantExpr::getBitCast(Src,
1130 Type::getInt8PtrTy(Src->getContext(), AS));
1131 Constant *OffsetCst =
1132 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1133 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1134 OffsetCst);
1135 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1136 return ConstantFoldLoadFromConstPtr(Src, LoadTy, DL);
1137 }
1138
1139
1140 /// Given a set of loads specified by ValuesPerBlock,
1141 /// construct SSA form, allowing us to eliminate LI. This returns the value
1142 /// that should be used at LI's definition site.
ConstructSSAForLoadSet(LoadInst * LI,SmallVectorImpl<AvailableValueInBlock> & ValuesPerBlock,GVN & gvn)1143 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1144 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1145 GVN &gvn) {
1146 // Check for the fully redundant, dominating load case. In this case, we can
1147 // just use the dominating value directly.
1148 if (ValuesPerBlock.size() == 1 &&
1149 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1150 LI->getParent())) {
1151 assert(!ValuesPerBlock[0].AV.isUndefValue() &&
1152 "Dead BB dominate this block");
1153 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
1154 }
1155
1156 // Otherwise, we have to construct SSA form.
1157 SmallVector<PHINode*, 8> NewPHIs;
1158 SSAUpdater SSAUpdate(&NewPHIs);
1159 SSAUpdate.Initialize(LI->getType(), LI->getName());
1160
1161 for (const AvailableValueInBlock &AV : ValuesPerBlock) {
1162 BasicBlock *BB = AV.BB;
1163
1164 if (SSAUpdate.HasValueForBlock(BB))
1165 continue;
1166
1167 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
1168 }
1169
1170 // Perform PHI construction.
1171 return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1172 }
1173
MaterializeAdjustedValue(LoadInst * LI,Instruction * InsertPt,GVN & gvn) const1174 Value *AvailableValue::MaterializeAdjustedValue(LoadInst *LI,
1175 Instruction *InsertPt,
1176 GVN &gvn) const {
1177 Value *Res;
1178 Type *LoadTy = LI->getType();
1179 const DataLayout &DL = LI->getModule()->getDataLayout();
1180 if (isSimpleValue()) {
1181 Res = getSimpleValue();
1182 if (Res->getType() != LoadTy) {
1183 Res = GetStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
1184
1185 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1186 << *getSimpleValue() << '\n'
1187 << *Res << '\n' << "\n\n\n");
1188 }
1189 } else if (isCoercedLoadValue()) {
1190 LoadInst *Load = getCoercedLoadValue();
1191 if (Load->getType() == LoadTy && Offset == 0) {
1192 Res = Load;
1193 } else {
1194 Res = GetLoadValueForLoad(Load, Offset, LoadTy, InsertPt, gvn);
1195
1196 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1197 << *getCoercedLoadValue() << '\n'
1198 << *Res << '\n' << "\n\n\n");
1199 }
1200 } else if (isMemIntrinValue()) {
1201 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
1202 InsertPt, DL);
1203 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1204 << " " << *getMemIntrinValue() << '\n'
1205 << *Res << '\n' << "\n\n\n");
1206 } else {
1207 assert(isUndefValue() && "Should be UndefVal");
1208 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1209 return UndefValue::get(LoadTy);
1210 }
1211 assert(Res && "failed to materialize?");
1212 return Res;
1213 }
1214
isLifetimeStart(const Instruction * Inst)1215 static bool isLifetimeStart(const Instruction *Inst) {
1216 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1217 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1218 return false;
1219 }
1220
AnalyzeLoadAvailability(LoadInst * LI,MemDepResult DepInfo,Value * Address,AvailableValue & Res)1221 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
1222 Value *Address, AvailableValue &Res) {
1223
1224 assert((DepInfo.isDef() || DepInfo.isClobber()) &&
1225 "expected a local dependence");
1226 assert(LI->isUnordered() && "rules below are incorrect for ordered access");
1227
1228 const DataLayout &DL = LI->getModule()->getDataLayout();
1229
1230 if (DepInfo.isClobber()) {
1231 // If the dependence is to a store that writes to a superset of the bits
1232 // read by the load, we can extract the bits we need for the load from the
1233 // stored value.
1234 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1235 // Can't forward from non-atomic to atomic without violating memory model.
1236 if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
1237 int Offset =
1238 AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
1239 if (Offset != -1) {
1240 Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
1241 return true;
1242 }
1243 }
1244 }
1245
1246 // Check to see if we have something like this:
1247 // load i32* P
1248 // load i8* (P+1)
1249 // if we have this, replace the later with an extraction from the former.
1250 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1251 // If this is a clobber and L is the first instruction in its block, then
1252 // we have the first instruction in the entry block.
1253 // Can't forward from non-atomic to atomic without violating memory model.
1254 if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
1255 int Offset =
1256 AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
1257
1258 if (Offset != -1) {
1259 Res = AvailableValue::getLoad(DepLI, Offset);
1260 return true;
1261 }
1262 }
1263 }
1264
1265 // If the clobbering value is a memset/memcpy/memmove, see if we can
1266 // forward a value on from it.
1267 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1268 if (Address && !LI->isAtomic()) {
1269 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1270 DepMI, DL);
1271 if (Offset != -1) {
1272 Res = AvailableValue::getMI(DepMI, Offset);
1273 return true;
1274 }
1275 }
1276 }
1277 // Nothing known about this clobber, have to be conservative
1278 DEBUG(
1279 // fast print dep, using operator<< on instruction is too slow.
1280 dbgs() << "GVN: load ";
1281 LI->printAsOperand(dbgs());
1282 Instruction *I = DepInfo.getInst();
1283 dbgs() << " is clobbered by " << *I << '\n';
1284 );
1285 return false;
1286 }
1287 assert(DepInfo.isDef() && "follows from above");
1288
1289 Instruction *DepInst = DepInfo.getInst();
1290
1291 // Loading the allocation -> undef.
1292 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1293 // Loading immediately after lifetime begin -> undef.
1294 isLifetimeStart(DepInst)) {
1295 Res = AvailableValue::get(UndefValue::get(LI->getType()));
1296 return true;
1297 }
1298
1299 // Loading from calloc (which zero initializes memory) -> zero
1300 if (isCallocLikeFn(DepInst, TLI)) {
1301 Res = AvailableValue::get(Constant::getNullValue(LI->getType()));
1302 return true;
1303 }
1304
1305 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1306 // Reject loads and stores that are to the same address but are of
1307 // different types if we have to. If the stored value is larger or equal to
1308 // the loaded value, we can reuse it.
1309 if (S->getValueOperand()->getType() != LI->getType() &&
1310 !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1311 LI->getType(), DL))
1312 return false;
1313
1314 // Can't forward from non-atomic to atomic without violating memory model.
1315 if (S->isAtomic() < LI->isAtomic())
1316 return false;
1317
1318 Res = AvailableValue::get(S->getValueOperand());
1319 return true;
1320 }
1321
1322 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1323 // If the types mismatch and we can't handle it, reject reuse of the load.
1324 // If the stored value is larger or equal to the loaded value, we can reuse
1325 // it.
1326 if (LD->getType() != LI->getType() &&
1327 !CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
1328 return false;
1329
1330 // Can't forward from non-atomic to atomic without violating memory model.
1331 if (LD->isAtomic() < LI->isAtomic())
1332 return false;
1333
1334 Res = AvailableValue::getLoad(LD);
1335 return true;
1336 }
1337
1338 // Unknown def - must be conservative
1339 DEBUG(
1340 // fast print dep, using operator<< on instruction is too slow.
1341 dbgs() << "GVN: load ";
1342 LI->printAsOperand(dbgs());
1343 dbgs() << " has unknown def " << *DepInst << '\n';
1344 );
1345 return false;
1346 }
1347
AnalyzeLoadAvailability(LoadInst * LI,LoadDepVect & Deps,AvailValInBlkVect & ValuesPerBlock,UnavailBlkVect & UnavailableBlocks)1348 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1349 AvailValInBlkVect &ValuesPerBlock,
1350 UnavailBlkVect &UnavailableBlocks) {
1351
1352 // Filter out useless results (non-locals, etc). Keep track of the blocks
1353 // where we have a value available in repl, also keep track of whether we see
1354 // dependencies that produce an unknown value for the load (such as a call
1355 // that could potentially clobber the load).
1356 unsigned NumDeps = Deps.size();
1357 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1358 BasicBlock *DepBB = Deps[i].getBB();
1359 MemDepResult DepInfo = Deps[i].getResult();
1360
1361 if (DeadBlocks.count(DepBB)) {
1362 // Dead dependent mem-op disguise as a load evaluating the same value
1363 // as the load in question.
1364 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1365 continue;
1366 }
1367
1368 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1369 UnavailableBlocks.push_back(DepBB);
1370 continue;
1371 }
1372
1373 // The address being loaded in this non-local block may not be the same as
1374 // the pointer operand of the load if PHI translation occurs. Make sure
1375 // to consider the right address.
1376 Value *Address = Deps[i].getAddress();
1377
1378 AvailableValue AV;
1379 if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1380 // subtlety: because we know this was a non-local dependency, we know
1381 // it's safe to materialize anywhere between the instruction within
1382 // DepInfo and the end of it's block.
1383 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1384 std::move(AV)));
1385 } else {
1386 UnavailableBlocks.push_back(DepBB);
1387 }
1388 }
1389
1390 assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1391 "post condition violation");
1392 }
1393
PerformLoadPRE(LoadInst * LI,AvailValInBlkVect & ValuesPerBlock,UnavailBlkVect & UnavailableBlocks)1394 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1395 UnavailBlkVect &UnavailableBlocks) {
1396 // Okay, we have *some* definitions of the value. This means that the value
1397 // is available in some of our (transitive) predecessors. Lets think about
1398 // doing PRE of this load. This will involve inserting a new load into the
1399 // predecessor when it's not available. We could do this in general, but
1400 // prefer to not increase code size. As such, we only do this when we know
1401 // that we only have to insert *one* load (which means we're basically moving
1402 // the load, not inserting a new one).
1403
1404 SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1405 UnavailableBlocks.end());
1406
1407 // Let's find the first basic block with more than one predecessor. Walk
1408 // backwards through predecessors if needed.
1409 BasicBlock *LoadBB = LI->getParent();
1410 BasicBlock *TmpBB = LoadBB;
1411
1412 while (TmpBB->getSinglePredecessor()) {
1413 TmpBB = TmpBB->getSinglePredecessor();
1414 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1415 return false;
1416 if (Blockers.count(TmpBB))
1417 return false;
1418
1419 // If any of these blocks has more than one successor (i.e. if the edge we
1420 // just traversed was critical), then there are other paths through this
1421 // block along which the load may not be anticipated. Hoisting the load
1422 // above this block would be adding the load to execution paths along
1423 // which it was not previously executed.
1424 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1425 return false;
1426 }
1427
1428 assert(TmpBB);
1429 LoadBB = TmpBB;
1430
1431 // Check to see how many predecessors have the loaded value fully
1432 // available.
1433 MapVector<BasicBlock *, Value *> PredLoads;
1434 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1435 for (const AvailableValueInBlock &AV : ValuesPerBlock)
1436 FullyAvailableBlocks[AV.BB] = true;
1437 for (BasicBlock *UnavailableBB : UnavailableBlocks)
1438 FullyAvailableBlocks[UnavailableBB] = false;
1439
1440 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1441 for (BasicBlock *Pred : predecessors(LoadBB)) {
1442 // If any predecessor block is an EH pad that does not allow non-PHI
1443 // instructions before the terminator, we can't PRE the load.
1444 if (Pred->getTerminator()->isEHPad()) {
1445 DEBUG(dbgs()
1446 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1447 << Pred->getName() << "': " << *LI << '\n');
1448 return false;
1449 }
1450
1451 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1452 continue;
1453 }
1454
1455 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1456 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1457 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1458 << Pred->getName() << "': " << *LI << '\n');
1459 return false;
1460 }
1461
1462 if (LoadBB->isEHPad()) {
1463 DEBUG(dbgs()
1464 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1465 << Pred->getName() << "': " << *LI << '\n');
1466 return false;
1467 }
1468
1469 CriticalEdgePred.push_back(Pred);
1470 } else {
1471 // Only add the predecessors that will not be split for now.
1472 PredLoads[Pred] = nullptr;
1473 }
1474 }
1475
1476 // Decide whether PRE is profitable for this load.
1477 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1478 assert(NumUnavailablePreds != 0 &&
1479 "Fully available value should already be eliminated!");
1480
1481 // If this load is unavailable in multiple predecessors, reject it.
1482 // FIXME: If we could restructure the CFG, we could make a common pred with
1483 // all the preds that don't have an available LI and insert a new load into
1484 // that one block.
1485 if (NumUnavailablePreds != 1)
1486 return false;
1487
1488 // Split critical edges, and update the unavailable predecessors accordingly.
1489 for (BasicBlock *OrigPred : CriticalEdgePred) {
1490 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1491 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1492 PredLoads[NewPred] = nullptr;
1493 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1494 << LoadBB->getName() << '\n');
1495 }
1496
1497 // Check if the load can safely be moved to all the unavailable predecessors.
1498 bool CanDoPRE = true;
1499 const DataLayout &DL = LI->getModule()->getDataLayout();
1500 SmallVector<Instruction*, 8> NewInsts;
1501 for (auto &PredLoad : PredLoads) {
1502 BasicBlock *UnavailablePred = PredLoad.first;
1503
1504 // Do PHI translation to get its value in the predecessor if necessary. The
1505 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1506
1507 // If all preds have a single successor, then we know it is safe to insert
1508 // the load on the pred (?!?), so we can insert code to materialize the
1509 // pointer if it is not available.
1510 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1511 Value *LoadPtr = nullptr;
1512 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1513 *DT, NewInsts);
1514
1515 // If we couldn't find or insert a computation of this phi translated value,
1516 // we fail PRE.
1517 if (!LoadPtr) {
1518 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1519 << *LI->getPointerOperand() << "\n");
1520 CanDoPRE = false;
1521 break;
1522 }
1523
1524 PredLoad.second = LoadPtr;
1525 }
1526
1527 if (!CanDoPRE) {
1528 while (!NewInsts.empty()) {
1529 Instruction *I = NewInsts.pop_back_val();
1530 if (MD) MD->removeInstruction(I);
1531 I->eraseFromParent();
1532 }
1533 // HINT: Don't revert the edge-splitting as following transformation may
1534 // also need to split these critical edges.
1535 return !CriticalEdgePred.empty();
1536 }
1537
1538 // Okay, we can eliminate this load by inserting a reload in the predecessor
1539 // and using PHI construction to get the value in the other predecessors, do
1540 // it.
1541 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1542 DEBUG(if (!NewInsts.empty())
1543 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1544 << *NewInsts.back() << '\n');
1545
1546 // Assign value numbers to the new instructions.
1547 for (Instruction *I : NewInsts) {
1548 // FIXME: We really _ought_ to insert these value numbers into their
1549 // parent's availability map. However, in doing so, we risk getting into
1550 // ordering issues. If a block hasn't been processed yet, we would be
1551 // marking a value as AVAIL-IN, which isn't what we intend.
1552 VN.lookupOrAdd(I);
1553 }
1554
1555 for (const auto &PredLoad : PredLoads) {
1556 BasicBlock *UnavailablePred = PredLoad.first;
1557 Value *LoadPtr = PredLoad.second;
1558
1559 auto *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre",
1560 LI->isVolatile(), LI->getAlignment(),
1561 LI->getOrdering(), LI->getSynchScope(),
1562 UnavailablePred->getTerminator());
1563
1564 // Transfer the old load's AA tags to the new load.
1565 AAMDNodes Tags;
1566 LI->getAAMetadata(Tags);
1567 if (Tags)
1568 NewLoad->setAAMetadata(Tags);
1569
1570 if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1571 NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1572 if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1573 NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1574 if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1575 NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1576
1577 // Transfer DebugLoc.
1578 NewLoad->setDebugLoc(LI->getDebugLoc());
1579
1580 // Add the newly created load.
1581 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1582 NewLoad));
1583 MD->invalidateCachedPointerInfo(LoadPtr);
1584 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1585 }
1586
1587 // Perform PHI construction.
1588 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1589 LI->replaceAllUsesWith(V);
1590 if (isa<PHINode>(V))
1591 V->takeName(LI);
1592 if (Instruction *I = dyn_cast<Instruction>(V))
1593 I->setDebugLoc(LI->getDebugLoc());
1594 if (V->getType()->getScalarType()->isPointerTy())
1595 MD->invalidateCachedPointerInfo(V);
1596 markInstructionForDeletion(LI);
1597 ++NumPRELoad;
1598 return true;
1599 }
1600
1601 /// Attempt to eliminate a load whose dependencies are
1602 /// non-local by performing PHI construction.
processNonLocalLoad(LoadInst * LI)1603 bool GVN::processNonLocalLoad(LoadInst *LI) {
1604 // non-local speculations are not allowed under asan.
1605 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeAddress))
1606 return false;
1607
1608 // Step 1: Find the non-local dependencies of the load.
1609 LoadDepVect Deps;
1610 MD->getNonLocalPointerDependency(LI, Deps);
1611
1612 // If we had to process more than one hundred blocks to find the
1613 // dependencies, this load isn't worth worrying about. Optimizing
1614 // it will be too expensive.
1615 unsigned NumDeps = Deps.size();
1616 if (NumDeps > 100)
1617 return false;
1618
1619 // If we had a phi translation failure, we'll have a single entry which is a
1620 // clobber in the current block. Reject this early.
1621 if (NumDeps == 1 &&
1622 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1623 DEBUG(
1624 dbgs() << "GVN: non-local load ";
1625 LI->printAsOperand(dbgs());
1626 dbgs() << " has unknown dependencies\n";
1627 );
1628 return false;
1629 }
1630
1631 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1632 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1633 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1634 OE = GEP->idx_end();
1635 OI != OE; ++OI)
1636 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1637 performScalarPRE(I);
1638 }
1639
1640 // Step 2: Analyze the availability of the load
1641 AvailValInBlkVect ValuesPerBlock;
1642 UnavailBlkVect UnavailableBlocks;
1643 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1644
1645 // If we have no predecessors that produce a known value for this load, exit
1646 // early.
1647 if (ValuesPerBlock.empty())
1648 return false;
1649
1650 // Step 3: Eliminate fully redundancy.
1651 //
1652 // If all of the instructions we depend on produce a known value for this
1653 // load, then it is fully redundant and we can use PHI insertion to compute
1654 // its value. Insert PHIs and remove the fully redundant value now.
1655 if (UnavailableBlocks.empty()) {
1656 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1657
1658 // Perform PHI construction.
1659 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1660 LI->replaceAllUsesWith(V);
1661
1662 if (isa<PHINode>(V))
1663 V->takeName(LI);
1664 if (Instruction *I = dyn_cast<Instruction>(V))
1665 if (LI->getDebugLoc())
1666 I->setDebugLoc(LI->getDebugLoc());
1667 if (V->getType()->getScalarType()->isPointerTy())
1668 MD->invalidateCachedPointerInfo(V);
1669 markInstructionForDeletion(LI);
1670 ++NumGVNLoad;
1671 return true;
1672 }
1673
1674 // Step 4: Eliminate partial redundancy.
1675 if (!EnablePRE || !EnableLoadPRE)
1676 return false;
1677
1678 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1679 }
1680
processAssumeIntrinsic(IntrinsicInst * IntrinsicI)1681 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1682 assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1683 "This function can only be called with llvm.assume intrinsic");
1684 Value *V = IntrinsicI->getArgOperand(0);
1685
1686 if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1687 if (Cond->isZero()) {
1688 Type *Int8Ty = Type::getInt8Ty(V->getContext());
1689 // Insert a new store to null instruction before the load to indicate that
1690 // this code is not reachable. FIXME: We could insert unreachable
1691 // instruction directly because we can modify the CFG.
1692 new StoreInst(UndefValue::get(Int8Ty),
1693 Constant::getNullValue(Int8Ty->getPointerTo()),
1694 IntrinsicI);
1695 }
1696 markInstructionForDeletion(IntrinsicI);
1697 return false;
1698 }
1699
1700 Constant *True = ConstantInt::getTrue(V->getContext());
1701 bool Changed = false;
1702
1703 for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1704 BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1705
1706 // This property is only true in dominated successors, propagateEquality
1707 // will check dominance for us.
1708 Changed |= propagateEquality(V, True, Edge, false);
1709 }
1710
1711 // We can replace assume value with true, which covers cases like this:
1712 // call void @llvm.assume(i1 %cmp)
1713 // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1714 ReplaceWithConstMap[V] = True;
1715
1716 // If one of *cmp *eq operand is const, adding it to map will cover this:
1717 // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1718 // call void @llvm.assume(i1 %cmp)
1719 // ret float %0 ; will change it to ret float 3.000000e+00
1720 if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1721 if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1722 CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1723 (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1724 CmpI->getFastMathFlags().noNaNs())) {
1725 Value *CmpLHS = CmpI->getOperand(0);
1726 Value *CmpRHS = CmpI->getOperand(1);
1727 if (isa<Constant>(CmpLHS))
1728 std::swap(CmpLHS, CmpRHS);
1729 auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1730
1731 // If only one operand is constant.
1732 if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1733 ReplaceWithConstMap[CmpLHS] = RHSConst;
1734 }
1735 }
1736 return Changed;
1737 }
1738
patchReplacementInstruction(Instruction * I,Value * Repl)1739 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1740 auto *ReplInst = dyn_cast<Instruction>(Repl);
1741 if (!ReplInst)
1742 return;
1743
1744 // Patch the replacement so that it is not more restrictive than the value
1745 // being replaced.
1746 ReplInst->andIRFlags(I);
1747
1748 // FIXME: If both the original and replacement value are part of the
1749 // same control-flow region (meaning that the execution of one
1750 // guarantees the execution of the other), then we can combine the
1751 // noalias scopes here and do better than the general conservative
1752 // answer used in combineMetadata().
1753
1754 // In general, GVN unifies expressions over different control-flow
1755 // regions, and so we need a conservative combination of the noalias
1756 // scopes.
1757 static const unsigned KnownIDs[] = {
1758 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
1759 LLVMContext::MD_noalias, LLVMContext::MD_range,
1760 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1761 LLVMContext::MD_invariant_group};
1762 combineMetadata(ReplInst, I, KnownIDs);
1763 }
1764
patchAndReplaceAllUsesWith(Instruction * I,Value * Repl)1765 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1766 patchReplacementInstruction(I, Repl);
1767 I->replaceAllUsesWith(Repl);
1768 }
1769
1770 /// Attempt to eliminate a load, first by eliminating it
1771 /// locally, and then attempting non-local elimination if that fails.
processLoad(LoadInst * L)1772 bool GVN::processLoad(LoadInst *L) {
1773 if (!MD)
1774 return false;
1775
1776 // This code hasn't been audited for ordered or volatile memory access
1777 if (!L->isUnordered())
1778 return false;
1779
1780 if (L->use_empty()) {
1781 markInstructionForDeletion(L);
1782 return true;
1783 }
1784
1785 // ... to a pointer that has been loaded from before...
1786 MemDepResult Dep = MD->getDependency(L);
1787
1788 // If it is defined in another block, try harder.
1789 if (Dep.isNonLocal())
1790 return processNonLocalLoad(L);
1791
1792 // Only handle the local case below
1793 if (!Dep.isDef() && !Dep.isClobber()) {
1794 // This might be a NonFuncLocal or an Unknown
1795 DEBUG(
1796 // fast print dep, using operator<< on instruction is too slow.
1797 dbgs() << "GVN: load ";
1798 L->printAsOperand(dbgs());
1799 dbgs() << " has unknown dependence\n";
1800 );
1801 return false;
1802 }
1803
1804 AvailableValue AV;
1805 if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1806 Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1807
1808 // Replace the load!
1809 patchAndReplaceAllUsesWith(L, AvailableValue);
1810 markInstructionForDeletion(L);
1811 ++NumGVNLoad;
1812 // Tell MDA to rexamine the reused pointer since we might have more
1813 // information after forwarding it.
1814 if (MD && AvailableValue->getType()->getScalarType()->isPointerTy())
1815 MD->invalidateCachedPointerInfo(AvailableValue);
1816 return true;
1817 }
1818
1819 return false;
1820 }
1821
1822 // In order to find a leader for a given value number at a
1823 // specific basic block, we first obtain the list of all Values for that number,
1824 // and then scan the list to find one whose block dominates the block in
1825 // question. This is fast because dominator tree queries consist of only
1826 // a few comparisons of DFS numbers.
findLeader(const BasicBlock * BB,uint32_t num)1827 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1828 LeaderTableEntry Vals = LeaderTable[num];
1829 if (!Vals.Val) return nullptr;
1830
1831 Value *Val = nullptr;
1832 if (DT->dominates(Vals.BB, BB)) {
1833 Val = Vals.Val;
1834 if (isa<Constant>(Val)) return Val;
1835 }
1836
1837 LeaderTableEntry* Next = Vals.Next;
1838 while (Next) {
1839 if (DT->dominates(Next->BB, BB)) {
1840 if (isa<Constant>(Next->Val)) return Next->Val;
1841 if (!Val) Val = Next->Val;
1842 }
1843
1844 Next = Next->Next;
1845 }
1846
1847 return Val;
1848 }
1849
1850 /// There is an edge from 'Src' to 'Dst'. Return
1851 /// true if every path from the entry block to 'Dst' passes via this edge. In
1852 /// particular 'Dst' must not be reachable via another edge from 'Src'.
isOnlyReachableViaThisEdge(const BasicBlockEdge & E,DominatorTree * DT)1853 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
1854 DominatorTree *DT) {
1855 // While in theory it is interesting to consider the case in which Dst has
1856 // more than one predecessor, because Dst might be part of a loop which is
1857 // only reachable from Src, in practice it is pointless since at the time
1858 // GVN runs all such loops have preheaders, which means that Dst will have
1859 // been changed to have only one predecessor, namely Src.
1860 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1861 assert((!Pred || Pred == E.getStart()) &&
1862 "No edge between these basic blocks!");
1863 return Pred != nullptr;
1864 }
1865
1866 // Tries to replace instruction with const, using information from
1867 // ReplaceWithConstMap.
replaceOperandsWithConsts(Instruction * Instr) const1868 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1869 bool Changed = false;
1870 for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1871 Value *Operand = Instr->getOperand(OpNum);
1872 auto it = ReplaceWithConstMap.find(Operand);
1873 if (it != ReplaceWithConstMap.end()) {
1874 assert(!isa<Constant>(Operand) &&
1875 "Replacing constants with constants is invalid");
1876 DEBUG(dbgs() << "GVN replacing: " << *Operand << " with " << *it->second
1877 << " in instruction " << *Instr << '\n');
1878 Instr->setOperand(OpNum, it->second);
1879 Changed = true;
1880 }
1881 }
1882 return Changed;
1883 }
1884
1885 /// The given values are known to be equal in every block
1886 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1887 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1888 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1889 /// value starting from the end of Root.Start.
propagateEquality(Value * LHS,Value * RHS,const BasicBlockEdge & Root,bool DominatesByEdge)1890 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1891 bool DominatesByEdge) {
1892 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
1893 Worklist.push_back(std::make_pair(LHS, RHS));
1894 bool Changed = false;
1895 // For speed, compute a conservative fast approximation to
1896 // DT->dominates(Root, Root.getEnd());
1897 const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1898
1899 while (!Worklist.empty()) {
1900 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1901 LHS = Item.first; RHS = Item.second;
1902
1903 if (LHS == RHS)
1904 continue;
1905 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1906
1907 // Don't try to propagate equalities between constants.
1908 if (isa<Constant>(LHS) && isa<Constant>(RHS))
1909 continue;
1910
1911 // Prefer a constant on the right-hand side, or an Argument if no constants.
1912 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1913 std::swap(LHS, RHS);
1914 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1915
1916 // If there is no obvious reason to prefer the left-hand side over the
1917 // right-hand side, ensure the longest lived term is on the right-hand side,
1918 // so the shortest lived term will be replaced by the longest lived.
1919 // This tends to expose more simplifications.
1920 uint32_t LVN = VN.lookupOrAdd(LHS);
1921 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1922 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1923 // Move the 'oldest' value to the right-hand side, using the value number
1924 // as a proxy for age.
1925 uint32_t RVN = VN.lookupOrAdd(RHS);
1926 if (LVN < RVN) {
1927 std::swap(LHS, RHS);
1928 LVN = RVN;
1929 }
1930 }
1931
1932 // If value numbering later sees that an instruction in the scope is equal
1933 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1934 // the invariant that instructions only occur in the leader table for their
1935 // own value number (this is used by removeFromLeaderTable), do not do this
1936 // if RHS is an instruction (if an instruction in the scope is morphed into
1937 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1938 // using the leader table is about compiling faster, not optimizing better).
1939 // The leader table only tracks basic blocks, not edges. Only add to if we
1940 // have the simple case where the edge dominates the end.
1941 if (RootDominatesEnd && !isa<Instruction>(RHS))
1942 addToLeaderTable(LVN, RHS, Root.getEnd());
1943
1944 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1945 // LHS always has at least one use that is not dominated by Root, this will
1946 // never do anything if LHS has only one use.
1947 if (!LHS->hasOneUse()) {
1948 unsigned NumReplacements =
1949 DominatesByEdge
1950 ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1951 : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1952
1953 Changed |= NumReplacements > 0;
1954 NumGVNEqProp += NumReplacements;
1955 }
1956
1957 // Now try to deduce additional equalities from this one. For example, if
1958 // the known equality was "(A != B)" == "false" then it follows that A and B
1959 // are equal in the scope. Only boolean equalities with an explicit true or
1960 // false RHS are currently supported.
1961 if (!RHS->getType()->isIntegerTy(1))
1962 // Not a boolean equality - bail out.
1963 continue;
1964 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1965 if (!CI)
1966 // RHS neither 'true' nor 'false' - bail out.
1967 continue;
1968 // Whether RHS equals 'true'. Otherwise it equals 'false'.
1969 bool isKnownTrue = CI->isAllOnesValue();
1970 bool isKnownFalse = !isKnownTrue;
1971
1972 // If "A && B" is known true then both A and B are known true. If "A || B"
1973 // is known false then both A and B are known false.
1974 Value *A, *B;
1975 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1976 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1977 Worklist.push_back(std::make_pair(A, RHS));
1978 Worklist.push_back(std::make_pair(B, RHS));
1979 continue;
1980 }
1981
1982 // If we are propagating an equality like "(A == B)" == "true" then also
1983 // propagate the equality A == B. When propagating a comparison such as
1984 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1985 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1986 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1987
1988 // If "A == B" is known true, or "A != B" is known false, then replace
1989 // A with B everywhere in the scope.
1990 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1991 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1992 Worklist.push_back(std::make_pair(Op0, Op1));
1993
1994 // Handle the floating point versions of equality comparisons too.
1995 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1996 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1997
1998 // Floating point -0.0 and 0.0 compare equal, so we can only
1999 // propagate values if we know that we have a constant and that
2000 // its value is non-zero.
2001
2002 // FIXME: We should do this optimization if 'no signed zeros' is
2003 // applicable via an instruction-level fast-math-flag or some other
2004 // indicator that relaxed FP semantics are being used.
2005
2006 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
2007 Worklist.push_back(std::make_pair(Op0, Op1));
2008 }
2009
2010 // If "A >= B" is known true, replace "A < B" with false everywhere.
2011 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2012 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2013 // Since we don't have the instruction "A < B" immediately to hand, work
2014 // out the value number that it would have and use that to find an
2015 // appropriate instruction (if any).
2016 uint32_t NextNum = VN.getNextUnusedValueNumber();
2017 uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2018 // If the number we were assigned was brand new then there is no point in
2019 // looking for an instruction realizing it: there cannot be one!
2020 if (Num < NextNum) {
2021 Value *NotCmp = findLeader(Root.getEnd(), Num);
2022 if (NotCmp && isa<Instruction>(NotCmp)) {
2023 unsigned NumReplacements =
2024 DominatesByEdge
2025 ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
2026 : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
2027 Root.getStart());
2028 Changed |= NumReplacements > 0;
2029 NumGVNEqProp += NumReplacements;
2030 }
2031 }
2032 // Ensure that any instruction in scope that gets the "A < B" value number
2033 // is replaced with false.
2034 // The leader table only tracks basic blocks, not edges. Only add to if we
2035 // have the simple case where the edge dominates the end.
2036 if (RootDominatesEnd)
2037 addToLeaderTable(Num, NotVal, Root.getEnd());
2038
2039 continue;
2040 }
2041 }
2042
2043 return Changed;
2044 }
2045
2046 /// When calculating availability, handle an instruction
2047 /// by inserting it into the appropriate sets
processInstruction(Instruction * I)2048 bool GVN::processInstruction(Instruction *I) {
2049 // Ignore dbg info intrinsics.
2050 if (isa<DbgInfoIntrinsic>(I))
2051 return false;
2052
2053 // If the instruction can be easily simplified then do so now in preference
2054 // to value numbering it. Value numbering often exposes redundancies, for
2055 // example if it determines that %y is equal to %x then the instruction
2056 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2057 const DataLayout &DL = I->getModule()->getDataLayout();
2058 if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2059 bool Changed = false;
2060 if (!I->use_empty()) {
2061 I->replaceAllUsesWith(V);
2062 Changed = true;
2063 }
2064 if (isInstructionTriviallyDead(I, TLI)) {
2065 markInstructionForDeletion(I);
2066 Changed = true;
2067 }
2068 if (Changed) {
2069 if (MD && V->getType()->getScalarType()->isPointerTy())
2070 MD->invalidateCachedPointerInfo(V);
2071 ++NumGVNSimpl;
2072 return true;
2073 }
2074 }
2075
2076 if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
2077 if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
2078 return processAssumeIntrinsic(IntrinsicI);
2079
2080 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2081 if (processLoad(LI))
2082 return true;
2083
2084 unsigned Num = VN.lookupOrAdd(LI);
2085 addToLeaderTable(Num, LI, LI->getParent());
2086 return false;
2087 }
2088
2089 // For conditional branches, we can perform simple conditional propagation on
2090 // the condition value itself.
2091 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2092 if (!BI->isConditional())
2093 return false;
2094
2095 if (isa<Constant>(BI->getCondition()))
2096 return processFoldableCondBr(BI);
2097
2098 Value *BranchCond = BI->getCondition();
2099 BasicBlock *TrueSucc = BI->getSuccessor(0);
2100 BasicBlock *FalseSucc = BI->getSuccessor(1);
2101 // Avoid multiple edges early.
2102 if (TrueSucc == FalseSucc)
2103 return false;
2104
2105 BasicBlock *Parent = BI->getParent();
2106 bool Changed = false;
2107
2108 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2109 BasicBlockEdge TrueE(Parent, TrueSucc);
2110 Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
2111
2112 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2113 BasicBlockEdge FalseE(Parent, FalseSucc);
2114 Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
2115
2116 return Changed;
2117 }
2118
2119 // For switches, propagate the case values into the case destinations.
2120 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2121 Value *SwitchCond = SI->getCondition();
2122 BasicBlock *Parent = SI->getParent();
2123 bool Changed = false;
2124
2125 // Remember how many outgoing edges there are to every successor.
2126 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2127 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2128 ++SwitchEdges[SI->getSuccessor(i)];
2129
2130 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2131 i != e; ++i) {
2132 BasicBlock *Dst = i.getCaseSuccessor();
2133 // If there is only a single edge, propagate the case value into it.
2134 if (SwitchEdges.lookup(Dst) == 1) {
2135 BasicBlockEdge E(Parent, Dst);
2136 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E, true);
2137 }
2138 }
2139 return Changed;
2140 }
2141
2142 // Instructions with void type don't return a value, so there's
2143 // no point in trying to find redundancies in them.
2144 if (I->getType()->isVoidTy())
2145 return false;
2146
2147 uint32_t NextNum = VN.getNextUnusedValueNumber();
2148 unsigned Num = VN.lookupOrAdd(I);
2149
2150 // Allocations are always uniquely numbered, so we can save time and memory
2151 // by fast failing them.
2152 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2153 addToLeaderTable(Num, I, I->getParent());
2154 return false;
2155 }
2156
2157 // If the number we were assigned was a brand new VN, then we don't
2158 // need to do a lookup to see if the number already exists
2159 // somewhere in the domtree: it can't!
2160 if (Num >= NextNum) {
2161 addToLeaderTable(Num, I, I->getParent());
2162 return false;
2163 }
2164
2165 // Perform fast-path value-number based elimination of values inherited from
2166 // dominators.
2167 Value *Repl = findLeader(I->getParent(), Num);
2168 if (!Repl) {
2169 // Failure, just remember this instance for future use.
2170 addToLeaderTable(Num, I, I->getParent());
2171 return false;
2172 } else if (Repl == I) {
2173 // If I was the result of a shortcut PRE, it might already be in the table
2174 // and the best replacement for itself. Nothing to do.
2175 return false;
2176 }
2177
2178 // Remove it!
2179 patchAndReplaceAllUsesWith(I, Repl);
2180 if (MD && Repl->getType()->getScalarType()->isPointerTy())
2181 MD->invalidateCachedPointerInfo(Repl);
2182 markInstructionForDeletion(I);
2183 return true;
2184 }
2185
2186 /// runOnFunction - This is the main transformation entry point for a function.
runImpl(Function & F,AssumptionCache & RunAC,DominatorTree & RunDT,const TargetLibraryInfo & RunTLI,AAResults & RunAA,MemoryDependenceResults * RunMD)2187 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
2188 const TargetLibraryInfo &RunTLI, AAResults &RunAA,
2189 MemoryDependenceResults *RunMD) {
2190 AC = &RunAC;
2191 DT = &RunDT;
2192 VN.setDomTree(DT);
2193 TLI = &RunTLI;
2194 VN.setAliasAnalysis(&RunAA);
2195 MD = RunMD;
2196 VN.setMemDep(MD);
2197
2198 bool Changed = false;
2199 bool ShouldContinue = true;
2200
2201 // Merge unconditional branches, allowing PRE to catch more
2202 // optimization opportunities.
2203 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2204 BasicBlock *BB = &*FI++;
2205
2206 bool removedBlock =
2207 MergeBlockIntoPredecessor(BB, DT, /* LoopInfo */ nullptr, MD);
2208 if (removedBlock) ++NumGVNBlocks;
2209
2210 Changed |= removedBlock;
2211 }
2212
2213 unsigned Iteration = 0;
2214 while (ShouldContinue) {
2215 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2216 ShouldContinue = iterateOnFunction(F);
2217 Changed |= ShouldContinue;
2218 ++Iteration;
2219 }
2220
2221 if (EnablePRE) {
2222 // Fabricate val-num for dead-code in order to suppress assertion in
2223 // performPRE().
2224 assignValNumForDeadCode();
2225 bool PREChanged = true;
2226 while (PREChanged) {
2227 PREChanged = performPRE(F);
2228 Changed |= PREChanged;
2229 }
2230 }
2231
2232 // FIXME: Should perform GVN again after PRE does something. PRE can move
2233 // computations into blocks where they become fully redundant. Note that
2234 // we can't do this until PRE's critical edge splitting updates memdep.
2235 // Actually, when this happens, we should just fully integrate PRE into GVN.
2236
2237 cleanupGlobalSets();
2238 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2239 // iteration.
2240 DeadBlocks.clear();
2241
2242 return Changed;
2243 }
2244
processBlock(BasicBlock * BB)2245 bool GVN::processBlock(BasicBlock *BB) {
2246 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2247 // (and incrementing BI before processing an instruction).
2248 assert(InstrsToErase.empty() &&
2249 "We expect InstrsToErase to be empty across iterations");
2250 if (DeadBlocks.count(BB))
2251 return false;
2252
2253 // Clearing map before every BB because it can be used only for single BB.
2254 ReplaceWithConstMap.clear();
2255 bool ChangedFunction = false;
2256
2257 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2258 BI != BE;) {
2259 if (!ReplaceWithConstMap.empty())
2260 ChangedFunction |= replaceOperandsWithConsts(&*BI);
2261 ChangedFunction |= processInstruction(&*BI);
2262
2263 if (InstrsToErase.empty()) {
2264 ++BI;
2265 continue;
2266 }
2267
2268 // If we need some instructions deleted, do it now.
2269 NumGVNInstr += InstrsToErase.size();
2270
2271 // Avoid iterator invalidation.
2272 bool AtStart = BI == BB->begin();
2273 if (!AtStart)
2274 --BI;
2275
2276 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2277 E = InstrsToErase.end(); I != E; ++I) {
2278 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2279 if (MD) MD->removeInstruction(*I);
2280 DEBUG(verifyRemoved(*I));
2281 (*I)->eraseFromParent();
2282 }
2283 InstrsToErase.clear();
2284
2285 if (AtStart)
2286 BI = BB->begin();
2287 else
2288 ++BI;
2289 }
2290
2291 return ChangedFunction;
2292 }
2293
2294 // Instantiate an expression in a predecessor that lacked it.
performScalarPREInsertion(Instruction * Instr,BasicBlock * Pred,unsigned int ValNo)2295 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2296 unsigned int ValNo) {
2297 // Because we are going top-down through the block, all value numbers
2298 // will be available in the predecessor by the time we need them. Any
2299 // that weren't originally present will have been instantiated earlier
2300 // in this loop.
2301 bool success = true;
2302 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2303 Value *Op = Instr->getOperand(i);
2304 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2305 continue;
2306 // This could be a newly inserted instruction, in which case, we won't
2307 // find a value number, and should give up before we hurt ourselves.
2308 // FIXME: Rewrite the infrastructure to let it easier to value number
2309 // and process newly inserted instructions.
2310 if (!VN.exists(Op)) {
2311 success = false;
2312 break;
2313 }
2314 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2315 Instr->setOperand(i, V);
2316 } else {
2317 success = false;
2318 break;
2319 }
2320 }
2321
2322 // Fail out if we encounter an operand that is not available in
2323 // the PRE predecessor. This is typically because of loads which
2324 // are not value numbered precisely.
2325 if (!success)
2326 return false;
2327
2328 Instr->insertBefore(Pred->getTerminator());
2329 Instr->setName(Instr->getName() + ".pre");
2330 Instr->setDebugLoc(Instr->getDebugLoc());
2331 VN.add(Instr, ValNo);
2332
2333 // Update the availability map to include the new instruction.
2334 addToLeaderTable(ValNo, Instr, Pred);
2335 return true;
2336 }
2337
performScalarPRE(Instruction * CurInst)2338 bool GVN::performScalarPRE(Instruction *CurInst) {
2339 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2340 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2341 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2342 isa<DbgInfoIntrinsic>(CurInst))
2343 return false;
2344
2345 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2346 // sinking the compare again, and it would force the code generator to
2347 // move the i1 from processor flags or predicate registers into a general
2348 // purpose register.
2349 if (isa<CmpInst>(CurInst))
2350 return false;
2351
2352 // We don't currently value number ANY inline asm calls.
2353 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2354 if (CallI->isInlineAsm())
2355 return false;
2356
2357 uint32_t ValNo = VN.lookup(CurInst);
2358
2359 // Look for the predecessors for PRE opportunities. We're
2360 // only trying to solve the basic diamond case, where
2361 // a value is computed in the successor and one predecessor,
2362 // but not the other. We also explicitly disallow cases
2363 // where the successor is its own predecessor, because they're
2364 // more complicated to get right.
2365 unsigned NumWith = 0;
2366 unsigned NumWithout = 0;
2367 BasicBlock *PREPred = nullptr;
2368 BasicBlock *CurrentBlock = CurInst->getParent();
2369
2370 SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap;
2371 for (BasicBlock *P : predecessors(CurrentBlock)) {
2372 // We're not interested in PRE where the block is its
2373 // own predecessor, or in blocks with predecessors
2374 // that are not reachable.
2375 if (P == CurrentBlock) {
2376 NumWithout = 2;
2377 break;
2378 } else if (!DT->isReachableFromEntry(P)) {
2379 NumWithout = 2;
2380 break;
2381 }
2382
2383 Value *predV = findLeader(P, ValNo);
2384 if (!predV) {
2385 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2386 PREPred = P;
2387 ++NumWithout;
2388 } else if (predV == CurInst) {
2389 /* CurInst dominates this predecessor. */
2390 NumWithout = 2;
2391 break;
2392 } else {
2393 predMap.push_back(std::make_pair(predV, P));
2394 ++NumWith;
2395 }
2396 }
2397
2398 // Don't do PRE when it might increase code size, i.e. when
2399 // we would need to insert instructions in more than one pred.
2400 if (NumWithout > 1 || NumWith == 0)
2401 return false;
2402
2403 // We may have a case where all predecessors have the instruction,
2404 // and we just need to insert a phi node. Otherwise, perform
2405 // insertion.
2406 Instruction *PREInstr = nullptr;
2407
2408 if (NumWithout != 0) {
2409 // Don't do PRE across indirect branch.
2410 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2411 return false;
2412
2413 // We can't do PRE safely on a critical edge, so instead we schedule
2414 // the edge to be split and perform the PRE the next time we iterate
2415 // on the function.
2416 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2417 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2418 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2419 return false;
2420 }
2421 // We need to insert somewhere, so let's give it a shot
2422 PREInstr = CurInst->clone();
2423 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2424 // If we failed insertion, make sure we remove the instruction.
2425 DEBUG(verifyRemoved(PREInstr));
2426 delete PREInstr;
2427 return false;
2428 }
2429 }
2430
2431 // Either we should have filled in the PRE instruction, or we should
2432 // not have needed insertions.
2433 assert (PREInstr != nullptr || NumWithout == 0);
2434
2435 ++NumGVNPRE;
2436
2437 // Create a PHI to make the value available in this block.
2438 PHINode *Phi =
2439 PHINode::Create(CurInst->getType(), predMap.size(),
2440 CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2441 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2442 if (Value *V = predMap[i].first)
2443 Phi->addIncoming(V, predMap[i].second);
2444 else
2445 Phi->addIncoming(PREInstr, PREPred);
2446 }
2447
2448 VN.add(Phi, ValNo);
2449 addToLeaderTable(ValNo, Phi, CurrentBlock);
2450 Phi->setDebugLoc(CurInst->getDebugLoc());
2451 CurInst->replaceAllUsesWith(Phi);
2452 if (MD && Phi->getType()->getScalarType()->isPointerTy())
2453 MD->invalidateCachedPointerInfo(Phi);
2454 VN.erase(CurInst);
2455 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2456
2457 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2458 if (MD)
2459 MD->removeInstruction(CurInst);
2460 DEBUG(verifyRemoved(CurInst));
2461 CurInst->eraseFromParent();
2462 ++NumGVNInstr;
2463
2464 return true;
2465 }
2466
2467 /// Perform a purely local form of PRE that looks for diamond
2468 /// control flow patterns and attempts to perform simple PRE at the join point.
performPRE(Function & F)2469 bool GVN::performPRE(Function &F) {
2470 bool Changed = false;
2471 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2472 // Nothing to PRE in the entry block.
2473 if (CurrentBlock == &F.getEntryBlock())
2474 continue;
2475
2476 // Don't perform PRE on an EH pad.
2477 if (CurrentBlock->isEHPad())
2478 continue;
2479
2480 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2481 BE = CurrentBlock->end();
2482 BI != BE;) {
2483 Instruction *CurInst = &*BI++;
2484 Changed |= performScalarPRE(CurInst);
2485 }
2486 }
2487
2488 if (splitCriticalEdges())
2489 Changed = true;
2490
2491 return Changed;
2492 }
2493
2494 /// Split the critical edge connecting the given two blocks, and return
2495 /// the block inserted to the critical edge.
splitCriticalEdges(BasicBlock * Pred,BasicBlock * Succ)2496 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2497 BasicBlock *BB =
2498 SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2499 if (MD)
2500 MD->invalidateCachedPredecessors();
2501 return BB;
2502 }
2503
2504 /// Split critical edges found during the previous
2505 /// iteration that may enable further optimization.
splitCriticalEdges()2506 bool GVN::splitCriticalEdges() {
2507 if (toSplit.empty())
2508 return false;
2509 do {
2510 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2511 SplitCriticalEdge(Edge.first, Edge.second,
2512 CriticalEdgeSplittingOptions(DT));
2513 } while (!toSplit.empty());
2514 if (MD) MD->invalidateCachedPredecessors();
2515 return true;
2516 }
2517
2518 /// Executes one iteration of GVN
iterateOnFunction(Function & F)2519 bool GVN::iterateOnFunction(Function &F) {
2520 cleanupGlobalSets();
2521
2522 // Top-down walk of the dominator tree
2523 bool Changed = false;
2524 // Save the blocks this function have before transformation begins. GVN may
2525 // split critical edge, and hence may invalidate the RPO/DT iterator.
2526 //
2527 std::vector<BasicBlock *> BBVect;
2528 BBVect.reserve(256);
2529 // Needed for value numbering with phi construction to work.
2530 ReversePostOrderTraversal<Function *> RPOT(&F);
2531 for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2532 RE = RPOT.end();
2533 RI != RE; ++RI)
2534 BBVect.push_back(*RI);
2535
2536 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2537 I != E; I++)
2538 Changed |= processBlock(*I);
2539
2540 return Changed;
2541 }
2542
cleanupGlobalSets()2543 void GVN::cleanupGlobalSets() {
2544 VN.clear();
2545 LeaderTable.clear();
2546 TableAllocator.Reset();
2547 }
2548
2549 /// Verify that the specified instruction does not occur in our
2550 /// internal data structures.
verifyRemoved(const Instruction * Inst) const2551 void GVN::verifyRemoved(const Instruction *Inst) const {
2552 VN.verifyRemoved(Inst);
2553
2554 // Walk through the value number scope to make sure the instruction isn't
2555 // ferreted away in it.
2556 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2557 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2558 const LeaderTableEntry *Node = &I->second;
2559 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2560
2561 while (Node->Next) {
2562 Node = Node->Next;
2563 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2564 }
2565 }
2566 }
2567
2568 /// BB is declared dead, which implied other blocks become dead as well. This
2569 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2570 /// live successors, update their phi nodes by replacing the operands
2571 /// corresponding to dead blocks with UndefVal.
addDeadBlock(BasicBlock * BB)2572 void GVN::addDeadBlock(BasicBlock *BB) {
2573 SmallVector<BasicBlock *, 4> NewDead;
2574 SmallSetVector<BasicBlock *, 4> DF;
2575
2576 NewDead.push_back(BB);
2577 while (!NewDead.empty()) {
2578 BasicBlock *D = NewDead.pop_back_val();
2579 if (DeadBlocks.count(D))
2580 continue;
2581
2582 // All blocks dominated by D are dead.
2583 SmallVector<BasicBlock *, 8> Dom;
2584 DT->getDescendants(D, Dom);
2585 DeadBlocks.insert(Dom.begin(), Dom.end());
2586
2587 // Figure out the dominance-frontier(D).
2588 for (BasicBlock *B : Dom) {
2589 for (BasicBlock *S : successors(B)) {
2590 if (DeadBlocks.count(S))
2591 continue;
2592
2593 bool AllPredDead = true;
2594 for (BasicBlock *P : predecessors(S))
2595 if (!DeadBlocks.count(P)) {
2596 AllPredDead = false;
2597 break;
2598 }
2599
2600 if (!AllPredDead) {
2601 // S could be proved dead later on. That is why we don't update phi
2602 // operands at this moment.
2603 DF.insert(S);
2604 } else {
2605 // While S is not dominated by D, it is dead by now. This could take
2606 // place if S already have a dead predecessor before D is declared
2607 // dead.
2608 NewDead.push_back(S);
2609 }
2610 }
2611 }
2612 }
2613
2614 // For the dead blocks' live successors, update their phi nodes by replacing
2615 // the operands corresponding to dead blocks with UndefVal.
2616 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2617 I != E; I++) {
2618 BasicBlock *B = *I;
2619 if (DeadBlocks.count(B))
2620 continue;
2621
2622 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2623 for (BasicBlock *P : Preds) {
2624 if (!DeadBlocks.count(P))
2625 continue;
2626
2627 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2628 if (BasicBlock *S = splitCriticalEdges(P, B))
2629 DeadBlocks.insert(P = S);
2630 }
2631
2632 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2633 PHINode &Phi = cast<PHINode>(*II);
2634 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2635 UndefValue::get(Phi.getType()));
2636 }
2637 }
2638 }
2639 }
2640
2641 // If the given branch is recognized as a foldable branch (i.e. conditional
2642 // branch with constant condition), it will perform following analyses and
2643 // transformation.
2644 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2645 // R be the target of the dead out-coming edge.
2646 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2647 // edge. The result of this step will be {X| X is dominated by R}
2648 // 2) Identify those blocks which haves at least one dead predecessor. The
2649 // result of this step will be dominance-frontier(R).
2650 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2651 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2652 //
2653 // Return true iff *NEW* dead code are found.
processFoldableCondBr(BranchInst * BI)2654 bool GVN::processFoldableCondBr(BranchInst *BI) {
2655 if (!BI || BI->isUnconditional())
2656 return false;
2657
2658 // If a branch has two identical successors, we cannot declare either dead.
2659 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2660 return false;
2661
2662 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2663 if (!Cond)
2664 return false;
2665
2666 BasicBlock *DeadRoot =
2667 Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2668 if (DeadBlocks.count(DeadRoot))
2669 return false;
2670
2671 if (!DeadRoot->getSinglePredecessor())
2672 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2673
2674 addDeadBlock(DeadRoot);
2675 return true;
2676 }
2677
2678 // performPRE() will trigger assert if it comes across an instruction without
2679 // associated val-num. As it normally has far more live instructions than dead
2680 // instructions, it makes more sense just to "fabricate" a val-number for the
2681 // dead code than checking if instruction involved is dead or not.
assignValNumForDeadCode()2682 void GVN::assignValNumForDeadCode() {
2683 for (BasicBlock *BB : DeadBlocks) {
2684 for (Instruction &Inst : *BB) {
2685 unsigned ValNum = VN.lookupOrAdd(&Inst);
2686 addToLeaderTable(ValNum, &Inst, BB);
2687 }
2688 }
2689 }
2690
2691 class llvm::gvn::GVNLegacyPass : public FunctionPass {
2692 public:
2693 static char ID; // Pass identification, replacement for typeid
GVNLegacyPass(bool NoLoads=false)2694 explicit GVNLegacyPass(bool NoLoads = false)
2695 : FunctionPass(ID), NoLoads(NoLoads) {
2696 initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry());
2697 }
2698
runOnFunction(Function & F)2699 bool runOnFunction(Function &F) override {
2700 if (skipFunction(F))
2701 return false;
2702
2703 return Impl.runImpl(
2704 F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2705 getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2706 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2707 getAnalysis<AAResultsWrapperPass>().getAAResults(),
2708 NoLoads ? nullptr
2709 : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep());
2710 }
2711
getAnalysisUsage(AnalysisUsage & AU) const2712 void getAnalysisUsage(AnalysisUsage &AU) const override {
2713 AU.addRequired<AssumptionCacheTracker>();
2714 AU.addRequired<DominatorTreeWrapperPass>();
2715 AU.addRequired<TargetLibraryInfoWrapperPass>();
2716 if (!NoLoads)
2717 AU.addRequired<MemoryDependenceWrapperPass>();
2718 AU.addRequired<AAResultsWrapperPass>();
2719
2720 AU.addPreserved<DominatorTreeWrapperPass>();
2721 AU.addPreserved<GlobalsAAWrapperPass>();
2722 }
2723
2724 private:
2725 bool NoLoads;
2726 GVN Impl;
2727 };
2728
2729 char GVNLegacyPass::ID = 0;
2730
2731 // The public interface to this file...
createGVNPass(bool NoLoads)2732 FunctionPass *llvm::createGVNPass(bool NoLoads) {
2733 return new GVNLegacyPass(NoLoads);
2734 }
2735
2736 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2737 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2738 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
2739 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2740 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
2741 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2742 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
2743 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2744