LLVM  9.0.0svn
GVN.cpp
Go to the documentation of this file.
1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass performs global value numbering to eliminate fully redundant
10 // instructions. It also performs simple dead load elimination.
11 //
12 // Note that this pass does the value numbering itself; it does not use the
13 // ValueNumbering analysis passes.
14 //
15 //===----------------------------------------------------------------------===//
16 
18 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/Hashing.h"
21 #include "llvm/ADT/MapVector.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SetVector.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/CFG.h"
35 #include "llvm/Analysis/LoopInfo.h"
42 #include "llvm/Config/llvm-config.h"
43 #include "llvm/IR/Attributes.h"
44 #include "llvm/IR/BasicBlock.h"
45 #include "llvm/IR/CallSite.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
50 #include "llvm/IR/DebugLoc.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
53 #include "llvm/IR/InstrTypes.h"
54 #include "llvm/IR/Instruction.h"
55 #include "llvm/IR/Instructions.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/Intrinsics.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/Metadata.h"
60 #include "llvm/IR/Module.h"
61 #include "llvm/IR/Operator.h"
62 #include "llvm/IR/PassManager.h"
63 #include "llvm/IR/PatternMatch.h"
64 #include "llvm/IR/Type.h"
65 #include "llvm/IR/Use.h"
66 #include "llvm/IR/Value.h"
67 #include "llvm/Pass.h"
68 #include "llvm/Support/Casting.h"
70 #include "llvm/Support/Compiler.h"
71 #include "llvm/Support/Debug.h"
77 #include <algorithm>
78 #include <cassert>
79 #include <cstdint>
80 #include <utility>
81 #include <vector>
82 
83 using namespace llvm;
84 using namespace llvm::gvn;
85 using namespace llvm::VNCoercion;
86 using namespace PatternMatch;
87 
88 #define DEBUG_TYPE "gvn"
89 
90 STATISTIC(NumGVNInstr, "Number of instructions deleted");
91 STATISTIC(NumGVNLoad, "Number of loads deleted");
92 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
93 STATISTIC(NumGVNBlocks, "Number of blocks merged");
94 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
95 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
96 STATISTIC(NumPRELoad, "Number of loads PRE'd");
97 
98 static cl::opt<bool> EnablePRE("enable-pre",
99  cl::init(true), cl::Hidden);
100 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
101 static cl::opt<bool> EnableMemDep("enable-gvn-memdep", cl::init(true));
102 
103 // Maximum allowed recursion depth.
104 static cl::opt<uint32_t>
105 MaxRecurseDepth("gvn-max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
106  cl::desc("Max recurse depth in GVN (default = 1000)"));
107 
109  "gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore,
110  cl::desc("Max number of dependences to attempt Load PRE (default = 100)"));
111 
115  bool commutative = false;
117 
118  Expression(uint32_t o = ~2U) : opcode(o) {}
119 
120  bool operator==(const Expression &other) const {
121  if (opcode != other.opcode)
122  return false;
123  if (opcode == ~0U || opcode == ~1U)
124  return true;
125  if (type != other.type)
126  return false;
127  if (varargs != other.varargs)
128  return false;
129  return true;
130  }
131 
133  return hash_combine(
134  Value.opcode, Value.type,
135  hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
136  }
137 };
138 
139 namespace llvm {
140 
141 template <> struct DenseMapInfo<GVN::Expression> {
142  static inline GVN::Expression getEmptyKey() { return ~0U; }
143  static inline GVN::Expression getTombstoneKey() { return ~1U; }
144 
145  static unsigned getHashValue(const GVN::Expression &e) {
146  using llvm::hash_value;
147 
148  return static_cast<unsigned>(hash_value(e));
149  }
150 
151  static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
152  return LHS == RHS;
153  }
154 };
155 
156 } // end namespace llvm
157 
158 /// Represents a particular available value that we know how to materialize.
159 /// Materialization of an AvailableValue never fails. An AvailableValue is
160 /// implicitly associated with a rematerialization point which is the
161 /// location of the instruction from which it was formed.
163  enum ValType {
164  SimpleVal, // A simple offsetted value that is accessed.
165  LoadVal, // A value produced by a load.
166  MemIntrin, // A memory intrinsic which is loaded from.
167  UndefVal // A UndefValue representing a value from dead block (which
168  // is not yet physically removed from the CFG).
169  };
170 
171  /// V - The value that is live out of the block.
173 
174  /// Offset - The byte offset in Val that is interesting for the load query.
175  unsigned Offset;
176 
177  static AvailableValue get(Value *V, unsigned Offset = 0) {
178  AvailableValue Res;
179  Res.Val.setPointer(V);
180  Res.Val.setInt(SimpleVal);
181  Res.Offset = Offset;
182  return Res;
183  }
184 
185  static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
186  AvailableValue Res;
187  Res.Val.setPointer(MI);
188  Res.Val.setInt(MemIntrin);
189  Res.Offset = Offset;
190  return Res;
191  }
192 
193  static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
194  AvailableValue Res;
195  Res.Val.setPointer(LI);
196  Res.Val.setInt(LoadVal);
197  Res.Offset = Offset;
198  return Res;
199  }
200 
202  AvailableValue Res;
203  Res.Val.setPointer(nullptr);
204  Res.Val.setInt(UndefVal);
205  Res.Offset = 0;
206  return Res;
207  }
208 
209  bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
210  bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
211  bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
212  bool isUndefValue() const { return Val.getInt() == UndefVal; }
213 
215  assert(isSimpleValue() && "Wrong accessor");
216  return Val.getPointer();
217  }
218 
220  assert(isCoercedLoadValue() && "Wrong accessor");
221  return cast<LoadInst>(Val.getPointer());
222  }
223 
225  assert(isMemIntrinValue() && "Wrong accessor");
226  return cast<MemIntrinsic>(Val.getPointer());
227  }
228 
229  /// Emit code at the specified insertion point to adjust the value defined
230  /// here to the specified type. This handles various coercion cases.
231  Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
232  GVN &gvn) const;
233 };
234 
235 /// Represents an AvailableValue which can be rematerialized at the end of
236 /// the associated BasicBlock.
238  /// BB - The basic block in question.
240 
241  /// AV - The actual available value
243 
246  Res.BB = BB;
247  Res.AV = std::move(AV);
248  return Res;
249  }
250 
252  unsigned Offset = 0) {
253  return get(BB, AvailableValue::get(V, Offset));
254  }
255 
257  return get(BB, AvailableValue::getUndef());
258  }
259 
260  /// Emit code at the end of this block to adjust the value defined here to
261  /// the specified type. This handles various coercion cases.
263  return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
264  }
265 };
266 
267 //===----------------------------------------------------------------------===//
268 // ValueTable Internal Functions
269 //===----------------------------------------------------------------------===//
270 
271 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
272  Expression e;
273  e.type = I->getType();
274  e.opcode = I->getOpcode();
275  for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
276  OI != OE; ++OI)
277  e.varargs.push_back(lookupOrAdd(*OI));
278  if (I->isCommutative()) {
279  // Ensure that commutative instructions that only differ by a permutation
280  // of their operands get the same value number by sorting the operand value
281  // numbers. Since all commutative instructions have two operands it is more
282  // efficient to sort by hand rather than using, say, std::sort.
283  assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
284  if (e.varargs[0] > e.varargs[1])
285  std::swap(e.varargs[0], e.varargs[1]);
286  e.commutative = true;
287  }
288 
289  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
290  // Sort the operand value numbers so x<y and y>x get the same value number.
291  CmpInst::Predicate Predicate = C->getPredicate();
292  if (e.varargs[0] > e.varargs[1]) {
293  std::swap(e.varargs[0], e.varargs[1]);
294  Predicate = CmpInst::getSwappedPredicate(Predicate);
295  }
296  e.opcode = (C->getOpcode() << 8) | Predicate;
297  e.commutative = true;
298  } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
299  for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
300  II != IE; ++II)
301  e.varargs.push_back(*II);
302  }
303 
304  return e;
305 }
306 
307 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
309  Value *LHS, Value *RHS) {
310  assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
311  "Not a comparison!");
312  Expression e;
313  e.type = CmpInst::makeCmpResultType(LHS->getType());
314  e.varargs.push_back(lookupOrAdd(LHS));
315  e.varargs.push_back(lookupOrAdd(RHS));
316 
317  // Sort the operand value numbers so x<y and y>x get the same value number.
318  if (e.varargs[0] > e.varargs[1]) {
319  std::swap(e.varargs[0], e.varargs[1]);
320  Predicate = CmpInst::getSwappedPredicate(Predicate);
321  }
322  e.opcode = (Opcode << 8) | Predicate;
323  e.commutative = true;
324  return e;
325 }
326 
327 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
328  assert(EI && "Not an ExtractValueInst?");
329  Expression e;
330  e.type = EI->getType();
331  e.opcode = 0;
332 
334  if (WO != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
335  // EI is an extract from one of our with.overflow intrinsics. Synthesize
336  // a semantically equivalent expression instead of an extract value
337  // expression.
338  e.opcode = WO->getBinaryOp();
339  e.varargs.push_back(lookupOrAdd(WO->getLHS()));
340  e.varargs.push_back(lookupOrAdd(WO->getRHS()));
341  return e;
342  }
343 
344  // Not a recognised intrinsic. Fall back to producing an extract value
345  // expression.
346  e.opcode = EI->getOpcode();
347  for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
348  OI != OE; ++OI)
349  e.varargs.push_back(lookupOrAdd(*OI));
350 
351  for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
352  II != IE; ++II)
353  e.varargs.push_back(*II);
354 
355  return e;
356 }
357 
358 //===----------------------------------------------------------------------===//
359 // ValueTable External Functions
360 //===----------------------------------------------------------------------===//
361 
362 GVN::ValueTable::ValueTable() = default;
363 GVN::ValueTable::ValueTable(const ValueTable &) = default;
364 GVN::ValueTable::ValueTable(ValueTable &&) = default;
365 GVN::ValueTable::~ValueTable() = default;
366 
367 /// add - Insert a value into the table with a specified value number.
369  valueNumbering.insert(std::make_pair(V, num));
370  if (PHINode *PN = dyn_cast<PHINode>(V))
371  NumberingPhi[num] = PN;
372 }
373 
374 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
375  if (AA->doesNotAccessMemory(C)) {
376  Expression exp = createExpr(C);
377  uint32_t e = assignExpNewValueNum(exp).first;
378  valueNumbering[C] = e;
379  return e;
380  } else if (MD && AA->onlyReadsMemory(C)) {
381  Expression exp = createExpr(C);
382  auto ValNum = assignExpNewValueNum(exp);
383  if (ValNum.second) {
384  valueNumbering[C] = ValNum.first;
385  return ValNum.first;
386  }
387 
388  MemDepResult local_dep = MD->getDependency(C);
389 
390  if (!local_dep.isDef() && !local_dep.isNonLocal()) {
391  valueNumbering[C] = nextValueNumber;
392  return nextValueNumber++;
393  }
394 
395  if (local_dep.isDef()) {
396  CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
397 
398  if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
399  valueNumbering[C] = nextValueNumber;
400  return nextValueNumber++;
401  }
402 
403  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
404  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
405  uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
406  if (c_vn != cd_vn) {
407  valueNumbering[C] = nextValueNumber;
408  return nextValueNumber++;
409  }
410  }
411 
412  uint32_t v = lookupOrAdd(local_cdep);
413  valueNumbering[C] = v;
414  return v;
415  }
416 
417  // Non-local case.
419  MD->getNonLocalCallDependency(C);
420  // FIXME: Move the checking logic to MemDep!
421  CallInst* cdep = nullptr;
422 
423  // Check to see if we have a single dominating call instruction that is
424  // identical to C.
425  for (unsigned i = 0, e = deps.size(); i != e; ++i) {
426  const NonLocalDepEntry *I = &deps[i];
427  if (I->getResult().isNonLocal())
428  continue;
429 
430  // We don't handle non-definitions. If we already have a call, reject
431  // instruction dependencies.
432  if (!I->getResult().isDef() || cdep != nullptr) {
433  cdep = nullptr;
434  break;
435  }
436 
437  CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
438  // FIXME: All duplicated with non-local case.
439  if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
440  cdep = NonLocalDepCall;
441  continue;
442  }
443 
444  cdep = nullptr;
445  break;
446  }
447 
448  if (!cdep) {
449  valueNumbering[C] = nextValueNumber;
450  return nextValueNumber++;
451  }
452 
453  if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
454  valueNumbering[C] = nextValueNumber;
455  return nextValueNumber++;
456  }
457  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
458  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
459  uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
460  if (c_vn != cd_vn) {
461  valueNumbering[C] = nextValueNumber;
462  return nextValueNumber++;
463  }
464  }
465 
466  uint32_t v = lookupOrAdd(cdep);
467  valueNumbering[C] = v;
468  return v;
469  } else {
470  valueNumbering[C] = nextValueNumber;
471  return nextValueNumber++;
472  }
473 }
474 
475 /// Returns true if a value number exists for the specified value.
476 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
477 
478 /// lookup_or_add - Returns the value number for the specified value, assigning
479 /// it a new number if it did not have one before.
481  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
482  if (VI != valueNumbering.end())
483  return VI->second;
484 
485  if (!isa<Instruction>(V)) {
486  valueNumbering[V] = nextValueNumber;
487  return nextValueNumber++;
488  }
489 
490  Instruction* I = cast<Instruction>(V);
491  Expression exp;
492  switch (I->getOpcode()) {
493  case Instruction::Call:
494  return lookupOrAddCall(cast<CallInst>(I));
495  case Instruction::Add:
496  case Instruction::FAdd:
497  case Instruction::Sub:
498  case Instruction::FSub:
499  case Instruction::Mul:
500  case Instruction::FMul:
501  case Instruction::UDiv:
502  case Instruction::SDiv:
503  case Instruction::FDiv:
504  case Instruction::URem:
505  case Instruction::SRem:
506  case Instruction::FRem:
507  case Instruction::Shl:
508  case Instruction::LShr:
509  case Instruction::AShr:
510  case Instruction::And:
511  case Instruction::Or:
512  case Instruction::Xor:
513  case Instruction::ICmp:
514  case Instruction::FCmp:
515  case Instruction::Trunc:
516  case Instruction::ZExt:
517  case Instruction::SExt:
518  case Instruction::FPToUI:
519  case Instruction::FPToSI:
520  case Instruction::UIToFP:
521  case Instruction::SIToFP:
522  case Instruction::FPTrunc:
523  case Instruction::FPExt:
524  case Instruction::PtrToInt:
525  case Instruction::IntToPtr:
526  case Instruction::BitCast:
527  case Instruction::Select:
528  case Instruction::ExtractElement:
529  case Instruction::InsertElement:
530  case Instruction::ShuffleVector:
531  case Instruction::InsertValue:
532  case Instruction::GetElementPtr:
533  exp = createExpr(I);
534  break;
535  case Instruction::ExtractValue:
536  exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
537  break;
538  case Instruction::PHI:
539  valueNumbering[V] = nextValueNumber;
540  NumberingPhi[nextValueNumber] = cast<PHINode>(V);
541  return nextValueNumber++;
542  default:
543  valueNumbering[V] = nextValueNumber;
544  return nextValueNumber++;
545  }
546 
547  uint32_t e = assignExpNewValueNum(exp).first;
548  valueNumbering[V] = e;
549  return e;
550 }
551 
552 /// Returns the value number of the specified value. Fails if
553 /// the value has not yet been numbered.
555  DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
556  if (Verify) {
557  assert(VI != valueNumbering.end() && "Value not numbered?");
558  return VI->second;
559  }
560  return (VI != valueNumbering.end()) ? VI->second : 0;
561 }
562 
563 /// Returns the value number of the given comparison,
564 /// assigning it a new number if it did not have one before. Useful when
565 /// we deduced the result of a comparison, but don't immediately have an
566 /// instruction realizing that comparison to hand.
568  CmpInst::Predicate Predicate,
569  Value *LHS, Value *RHS) {
570  Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
571  return assignExpNewValueNum(exp).first;
572 }
573 
574 /// Remove all entries from the ValueTable.
576  valueNumbering.clear();
577  expressionNumbering.clear();
578  NumberingPhi.clear();
579  PhiTranslateTable.clear();
580  nextValueNumber = 1;
581  Expressions.clear();
582  ExprIdx.clear();
583  nextExprNumber = 0;
584 }
585 
586 /// Remove a value from the value numbering.
588  uint32_t Num = valueNumbering.lookup(V);
589  valueNumbering.erase(V);
590  // If V is PHINode, V <--> value number is an one-to-one mapping.
591  if (isa<PHINode>(V))
592  NumberingPhi.erase(Num);
593 }
594 
595 /// verifyRemoved - Verify that the value is removed from all internal data
596 /// structures.
597 void GVN::ValueTable::verifyRemoved(const Value *V) const {
599  I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
600  assert(I->first != V && "Inst still occurs in value numbering map!");
601  }
602 }
603 
604 //===----------------------------------------------------------------------===//
605 // GVN Pass
606 //===----------------------------------------------------------------------===//
607 
609  // FIXME: The order of evaluation of these 'getResult' calls is very
610  // significant! Re-ordering these variables will cause GVN when run alone to
611  // be less effective! We should fix memdep and basic-aa to not exhibit this
612  // behavior, but until then don't change the order here.
613  auto &AC = AM.getResult<AssumptionAnalysis>(F);
614  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
615  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
616  auto &AA = AM.getResult<AAManager>(F);
617  auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
618  auto *LI = AM.getCachedResult<LoopAnalysis>(F);
620  bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
621  if (!Changed)
622  return PreservedAnalyses::all();
625  PA.preserve<GlobalsAA>();
627  return PA;
628 }
629 
630 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
631 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
632  errs() << "{\n";
634  E = d.end(); I != E; ++I) {
635  errs() << I->first << "\n";
636  I->second->dump();
637  }
638  errs() << "}\n";
639 }
640 #endif
641 
642 /// Return true if we can prove that the value
643 /// we're analyzing is fully available in the specified block. As we go, keep
644 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
645 /// map is actually a tri-state map with the following values:
646 /// 0) we know the block *is not* fully available.
647 /// 1) we know the block *is* fully available.
648 /// 2) we do not know whether the block is fully available or not, but we are
649 /// currently speculating that it will be.
650 /// 3) we are speculating for this block and have used that to speculate for
651 /// other blocks.
653  DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
654  uint32_t RecurseDepth) {
655  if (RecurseDepth > MaxRecurseDepth)
656  return false;
657 
658  // Optimistically assume that the block is fully available and check to see
659  // if we already know about this block in one lookup.
660  std::pair<DenseMap<BasicBlock*, char>::iterator, bool> IV =
661  FullyAvailableBlocks.insert(std::make_pair(BB, 2));
662 
663  // If the entry already existed for this block, return the precomputed value.
664  if (!IV.second) {
665  // If this is a speculative "available" value, mark it as being used for
666  // speculation of other blocks.
667  if (IV.first->second == 2)
668  IV.first->second = 3;
669  return IV.first->second != 0;
670  }
671 
672  // Otherwise, see if it is fully available in all predecessors.
673  pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
674 
675  // If this block has no predecessors, it isn't live-in here.
676  if (PI == PE)
677  goto SpeculationFailure;
678 
679  for (; PI != PE; ++PI)
680  // If the value isn't fully available in one of our predecessors, then it
681  // isn't fully available in this block either. Undo our previous
682  // optimistic assumption and bail out.
683  if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
684  goto SpeculationFailure;
685 
686  return true;
687 
688 // If we get here, we found out that this is not, after
689 // all, a fully-available block. We have a problem if we speculated on this and
690 // used the speculation to mark other blocks as available.
691 SpeculationFailure:
692  char &BBVal = FullyAvailableBlocks[BB];
693 
694  // If we didn't speculate on this, just return with it set to false.
695  if (BBVal == 2) {
696  BBVal = 0;
697  return false;
698  }
699 
700  // If we did speculate on this value, we could have blocks set to 1 that are
701  // incorrect. Walk the (transitive) successors of this block and mark them as
702  // 0 if set to one.
703  SmallVector<BasicBlock*, 32> BBWorklist;
704  BBWorklist.push_back(BB);
705 
706  do {
707  BasicBlock *Entry = BBWorklist.pop_back_val();
708  // Note that this sets blocks to 0 (unavailable) if they happen to not
709  // already be in FullyAvailableBlocks. This is safe.
710  char &EntryVal = FullyAvailableBlocks[Entry];
711  if (EntryVal == 0) continue; // Already unavailable.
712 
713  // Mark as unavailable.
714  EntryVal = 0;
715 
716  BBWorklist.append(succ_begin(Entry), succ_end(Entry));
717  } while (!BBWorklist.empty());
718 
719  return false;
720 }
721 
722 /// Given a set of loads specified by ValuesPerBlock,
723 /// construct SSA form, allowing us to eliminate LI. This returns the value
724 /// that should be used at LI's definition site.
727  GVN &gvn) {
728  // Check for the fully redundant, dominating load case. In this case, we can
729  // just use the dominating value directly.
730  if (ValuesPerBlock.size() == 1 &&
731  gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
732  LI->getParent())) {
733  assert(!ValuesPerBlock[0].AV.isUndefValue() &&
734  "Dead BB dominate this block");
735  return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
736  }
737 
738  // Otherwise, we have to construct SSA form.
739  SmallVector<PHINode*, 8> NewPHIs;
740  SSAUpdater SSAUpdate(&NewPHIs);
741  SSAUpdate.Initialize(LI->getType(), LI->getName());
742 
743  for (const AvailableValueInBlock &AV : ValuesPerBlock) {
744  BasicBlock *BB = AV.BB;
745 
746  if (SSAUpdate.HasValueForBlock(BB))
747  continue;
748 
749  // If the value is the load that we will be eliminating, and the block it's
750  // available in is the block that the load is in, then don't add it as
751  // SSAUpdater will resolve the value to the relevant phi which may let it
752  // avoid phi construction entirely if there's actually only one value.
753  if (BB == LI->getParent() &&
754  ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == LI) ||
755  (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == LI)))
756  continue;
757 
758  SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
759  }
760 
761  // Perform PHI construction.
762  return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
763 }
764 
766  Instruction *InsertPt,
767  GVN &gvn) const {
768  Value *Res;
769  Type *LoadTy = LI->getType();
770  const DataLayout &DL = LI->getModule()->getDataLayout();
771  if (isSimpleValue()) {
772  Res = getSimpleValue();
773  if (Res->getType() != LoadTy) {
774  Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
775 
776  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset
777  << " " << *getSimpleValue() << '\n'
778  << *Res << '\n'
779  << "\n\n\n");
780  }
781  } else if (isCoercedLoadValue()) {
782  LoadInst *Load = getCoercedLoadValue();
783  if (Load->getType() == LoadTy && Offset == 0) {
784  Res = Load;
785  } else {
786  Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
787  // We would like to use gvn.markInstructionForDeletion here, but we can't
788  // because the load is already memoized into the leader map table that GVN
789  // tracks. It is potentially possible to remove the load from the table,
790  // but then there all of the operations based on it would need to be
791  // rehashed. Just leave the dead load around.
792  gvn.getMemDep().removeInstruction(Load);
793  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset
794  << " " << *getCoercedLoadValue() << '\n'
795  << *Res << '\n'
796  << "\n\n\n");
797  }
798  } else if (isMemIntrinValue()) {
799  Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
800  InsertPt, DL);
801  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
802  << " " << *getMemIntrinValue() << '\n'
803  << *Res << '\n'
804  << "\n\n\n");
805  } else {
806  assert(isUndefValue() && "Should be UndefVal");
807  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
808  return UndefValue::get(LoadTy);
809  }
810  assert(Res && "failed to materialize?");
811  return Res;
812 }
813 
814 static bool isLifetimeStart(const Instruction *Inst) {
815  if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
816  return II->getIntrinsicID() == Intrinsic::lifetime_start;
817  return false;
818 }
819 
820 /// Try to locate the three instruction involved in a missed
821 /// load-elimination case that is due to an intervening store.
823  DominatorTree *DT,
825  using namespace ore;
826 
827  User *OtherAccess = nullptr;
828 
829  OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
830  R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
831  << setExtraArgs();
832 
833  for (auto *U : LI->getPointerOperand()->users())
834  if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
835  DT->dominates(cast<Instruction>(U), LI)) {
836  // FIXME: for now give up if there are multiple memory accesses that
837  // dominate the load. We need further analysis to decide which one is
838  // that we're forwarding from.
839  if (OtherAccess)
840  OtherAccess = nullptr;
841  else
842  OtherAccess = U;
843  }
844 
845  if (OtherAccess)
846  R << " in favor of " << NV("OtherAccess", OtherAccess);
847 
848  R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
849 
850  ORE->emit(R);
851 }
852 
853 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
854  Value *Address, AvailableValue &Res) {
855  assert((DepInfo.isDef() || DepInfo.isClobber()) &&
856  "expected a local dependence");
857  assert(LI->isUnordered() && "rules below are incorrect for ordered access");
858 
859  const DataLayout &DL = LI->getModule()->getDataLayout();
860 
861  if (DepInfo.isClobber()) {
862  // If the dependence is to a store that writes to a superset of the bits
863  // read by the load, we can extract the bits we need for the load from the
864  // stored value.
865  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
866  // Can't forward from non-atomic to atomic without violating memory model.
867  if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
868  int Offset =
870  if (Offset != -1) {
871  Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
872  return true;
873  }
874  }
875  }
876 
877  // Check to see if we have something like this:
878  // load i32* P
879  // load i8* (P+1)
880  // if we have this, replace the later with an extraction from the former.
881  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
882  // If this is a clobber and L is the first instruction in its block, then
883  // we have the first instruction in the entry block.
884  // Can't forward from non-atomic to atomic without violating memory model.
885  if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
886  int Offset =
887  analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
888 
889  if (Offset != -1) {
890  Res = AvailableValue::getLoad(DepLI, Offset);
891  return true;
892  }
893  }
894  }
895 
896  // If the clobbering value is a memset/memcpy/memmove, see if we can
897  // forward a value on from it.
898  if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
899  if (Address && !LI->isAtomic()) {
901  DepMI, DL);
902  if (Offset != -1) {
903  Res = AvailableValue::getMI(DepMI, Offset);
904  return true;
905  }
906  }
907  }
908  // Nothing known about this clobber, have to be conservative
909  LLVM_DEBUG(
910  // fast print dep, using operator<< on instruction is too slow.
911  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
912  Instruction *I = DepInfo.getInst();
913  dbgs() << " is clobbered by " << *I << '\n';);
914  if (ORE->allowExtraAnalysis(DEBUG_TYPE))
915  reportMayClobberedLoad(LI, DepInfo, DT, ORE);
916 
917  return false;
918  }
919  assert(DepInfo.isDef() && "follows from above");
920 
921  Instruction *DepInst = DepInfo.getInst();
922 
923  // Loading the allocation -> undef.
924  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
925  // Loading immediately after lifetime begin -> undef.
926  isLifetimeStart(DepInst)) {
928  return true;
929  }
930 
931  // Loading from calloc (which zero initializes memory) -> zero
932  if (isCallocLikeFn(DepInst, TLI)) {
934  return true;
935  }
936 
937  if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
938  // Reject loads and stores that are to the same address but are of
939  // different types if we have to. If the stored value is larger or equal to
940  // the loaded value, we can reuse it.
941  if (S->getValueOperand()->getType() != LI->getType() &&
942  !canCoerceMustAliasedValueToLoad(S->getValueOperand(),
943  LI->getType(), DL))
944  return false;
945 
946  // Can't forward from non-atomic to atomic without violating memory model.
947  if (S->isAtomic() < LI->isAtomic())
948  return false;
949 
950  Res = AvailableValue::get(S->getValueOperand());
951  return true;
952  }
953 
954  if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
955  // If the types mismatch and we can't handle it, reject reuse of the load.
956  // If the stored value is larger or equal to the loaded value, we can reuse
957  // it.
958  if (LD->getType() != LI->getType() &&
960  return false;
961 
962  // Can't forward from non-atomic to atomic without violating memory model.
963  if (LD->isAtomic() < LI->isAtomic())
964  return false;
965 
967  return true;
968  }
969 
970  // Unknown def - must be conservative
971  LLVM_DEBUG(
972  // fast print dep, using operator<< on instruction is too slow.
973  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
974  dbgs() << " has unknown def " << *DepInst << '\n';);
975  return false;
976 }
977 
978 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
979  AvailValInBlkVect &ValuesPerBlock,
980  UnavailBlkVect &UnavailableBlocks) {
981  // Filter out useless results (non-locals, etc). Keep track of the blocks
982  // where we have a value available in repl, also keep track of whether we see
983  // dependencies that produce an unknown value for the load (such as a call
984  // that could potentially clobber the load).
985  unsigned NumDeps = Deps.size();
986  for (unsigned i = 0, e = NumDeps; i != e; ++i) {
987  BasicBlock *DepBB = Deps[i].getBB();
988  MemDepResult DepInfo = Deps[i].getResult();
989 
990  if (DeadBlocks.count(DepBB)) {
991  // Dead dependent mem-op disguise as a load evaluating the same value
992  // as the load in question.
993  ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
994  continue;
995  }
996 
997  if (!DepInfo.isDef() && !DepInfo.isClobber()) {
998  UnavailableBlocks.push_back(DepBB);
999  continue;
1000  }
1001 
1002  // The address being loaded in this non-local block may not be the same as
1003  // the pointer operand of the load if PHI translation occurs. Make sure
1004  // to consider the right address.
1005  Value *Address = Deps[i].getAddress();
1006 
1007  AvailableValue AV;
1008  if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1009  // subtlety: because we know this was a non-local dependency, we know
1010  // it's safe to materialize anywhere between the instruction within
1011  // DepInfo and the end of it's block.
1012  ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1013  std::move(AV)));
1014  } else {
1015  UnavailableBlocks.push_back(DepBB);
1016  }
1017  }
1018 
1019  assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1020  "post condition violation");
1021 }
1022 
1023 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1024  UnavailBlkVect &UnavailableBlocks) {
1025  // Okay, we have *some* definitions of the value. This means that the value
1026  // is available in some of our (transitive) predecessors. Lets think about
1027  // doing PRE of this load. This will involve inserting a new load into the
1028  // predecessor when it's not available. We could do this in general, but
1029  // prefer to not increase code size. As such, we only do this when we know
1030  // that we only have to insert *one* load (which means we're basically moving
1031  // the load, not inserting a new one).
1032 
1033  SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1034  UnavailableBlocks.end());
1035 
1036  // Let's find the first basic block with more than one predecessor. Walk
1037  // backwards through predecessors if needed.
1038  BasicBlock *LoadBB = LI->getParent();
1039  BasicBlock *TmpBB = LoadBB;
1040  bool IsSafeToSpeculativelyExecute = isSafeToSpeculativelyExecute(LI);
1041 
1042  // Check that there is no implicit control flow instructions above our load in
1043  // its block. If there is an instruction that doesn't always pass the
1044  // execution to the following instruction, then moving through it may become
1045  // invalid. For example:
1046  //
1047  // int arr[LEN];
1048  // int index = ???;
1049  // ...
1050  // guard(0 <= index && index < LEN);
1051  // use(arr[index]);
1052  //
1053  // It is illegal to move the array access to any point above the guard,
1054  // because if the index is out of bounds we should deoptimize rather than
1055  // access the array.
1056  // Check that there is no guard in this block above our instruction.
1057  if (!IsSafeToSpeculativelyExecute && ICF->isDominatedByICFIFromSameBlock(LI))
1058  return false;
1059  while (TmpBB->getSinglePredecessor()) {
1060  TmpBB = TmpBB->getSinglePredecessor();
1061  if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1062  return false;
1063  if (Blockers.count(TmpBB))
1064  return false;
1065 
1066  // If any of these blocks has more than one successor (i.e. if the edge we
1067  // just traversed was critical), then there are other paths through this
1068  // block along which the load may not be anticipated. Hoisting the load
1069  // above this block would be adding the load to execution paths along
1070  // which it was not previously executed.
1071  if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1072  return false;
1073 
1074  // Check that there is no implicit control flow in a block above.
1075  if (!IsSafeToSpeculativelyExecute && ICF->hasICF(TmpBB))
1076  return false;
1077  }
1078 
1079  assert(TmpBB);
1080  LoadBB = TmpBB;
1081 
1082  // Check to see how many predecessors have the loaded value fully
1083  // available.
1085  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1086  for (const AvailableValueInBlock &AV : ValuesPerBlock)
1087  FullyAvailableBlocks[AV.BB] = true;
1088  for (BasicBlock *UnavailableBB : UnavailableBlocks)
1089  FullyAvailableBlocks[UnavailableBB] = false;
1090 
1091  SmallVector<BasicBlock *, 4> CriticalEdgePred;
1092  for (BasicBlock *Pred : predecessors(LoadBB)) {
1093  // If any predecessor block is an EH pad that does not allow non-PHI
1094  // instructions before the terminator, we can't PRE the load.
1095  if (Pred->getTerminator()->isEHPad()) {
1096  LLVM_DEBUG(
1097  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1098  << Pred->getName() << "': " << *LI << '\n');
1099  return false;
1100  }
1101 
1102  if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1103  continue;
1104  }
1105 
1106  if (Pred->getTerminator()->getNumSuccessors() != 1) {
1107  if (isa<IndirectBrInst>(Pred->getTerminator())) {
1108  LLVM_DEBUG(
1109  dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1110  << Pred->getName() << "': " << *LI << '\n');
1111  return false;
1112  }
1113 
1114  // FIXME: Can we support the fallthrough edge?
1115  if (isa<CallBrInst>(Pred->getTerminator())) {
1116  LLVM_DEBUG(
1117  dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '"
1118  << Pred->getName() << "': " << *LI << '\n');
1119  return false;
1120  }
1121 
1122  if (LoadBB->isEHPad()) {
1123  LLVM_DEBUG(
1124  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1125  << Pred->getName() << "': " << *LI << '\n');
1126  return false;
1127  }
1128 
1129  CriticalEdgePred.push_back(Pred);
1130  } else {
1131  // Only add the predecessors that will not be split for now.
1132  PredLoads[Pred] = nullptr;
1133  }
1134  }
1135 
1136  // Decide whether PRE is profitable for this load.
1137  unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1138  assert(NumUnavailablePreds != 0 &&
1139  "Fully available value should already be eliminated!");
1140 
1141  // If this load is unavailable in multiple predecessors, reject it.
1142  // FIXME: If we could restructure the CFG, we could make a common pred with
1143  // all the preds that don't have an available LI and insert a new load into
1144  // that one block.
1145  if (NumUnavailablePreds != 1)
1146  return false;
1147 
1148  // Split critical edges, and update the unavailable predecessors accordingly.
1149  for (BasicBlock *OrigPred : CriticalEdgePred) {
1150  BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1151  assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1152  PredLoads[NewPred] = nullptr;
1153  LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1154  << LoadBB->getName() << '\n');
1155  }
1156 
1157  // Check if the load can safely be moved to all the unavailable predecessors.
1158  bool CanDoPRE = true;
1159  const DataLayout &DL = LI->getModule()->getDataLayout();
1161  for (auto &PredLoad : PredLoads) {
1162  BasicBlock *UnavailablePred = PredLoad.first;
1163 
1164  // Do PHI translation to get its value in the predecessor if necessary. The
1165  // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1166 
1167  // If all preds have a single successor, then we know it is safe to insert
1168  // the load on the pred (?!?), so we can insert code to materialize the
1169  // pointer if it is not available.
1170  PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1171  Value *LoadPtr = nullptr;
1172  LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1173  *DT, NewInsts);
1174 
1175  // If we couldn't find or insert a computation of this phi translated value,
1176  // we fail PRE.
1177  if (!LoadPtr) {
1178  LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1179  << *LI->getPointerOperand() << "\n");
1180  CanDoPRE = false;
1181  break;
1182  }
1183 
1184  PredLoad.second = LoadPtr;
1185  }
1186 
1187  if (!CanDoPRE) {
1188  while (!NewInsts.empty()) {
1189  Instruction *I = NewInsts.pop_back_val();
1190  markInstructionForDeletion(I);
1191  }
1192  // HINT: Don't revert the edge-splitting as following transformation may
1193  // also need to split these critical edges.
1194  return !CriticalEdgePred.empty();
1195  }
1196 
1197  // Okay, we can eliminate this load by inserting a reload in the predecessor
1198  // and using PHI construction to get the value in the other predecessors, do
1199  // it.
1200  LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1201  LLVM_DEBUG(if (!NewInsts.empty()) dbgs()
1202  << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back()
1203  << '\n');
1204 
1205  // Assign value numbers to the new instructions.
1206  for (Instruction *I : NewInsts) {
1207  // Instructions that have been inserted in predecessor(s) to materialize
1208  // the load address do not retain their original debug locations. Doing
1209  // so could lead to confusing (but correct) source attributions.
1210  if (const DebugLoc &DL = I->getDebugLoc())
1211  I->setDebugLoc(DebugLoc::get(0, 0, DL.getScope(), DL.getInlinedAt()));
1212 
1213  // FIXME: We really _ought_ to insert these value numbers into their
1214  // parent's availability map. However, in doing so, we risk getting into
1215  // ordering issues. If a block hasn't been processed yet, we would be
1216  // marking a value as AVAIL-IN, which isn't what we intend.
1217  VN.lookupOrAdd(I);
1218  }
1219 
1220  for (const auto &PredLoad : PredLoads) {
1221  BasicBlock *UnavailablePred = PredLoad.first;
1222  Value *LoadPtr = PredLoad.second;
1223 
1224  auto *NewLoad =
1225  new LoadInst(LI->getType(), LoadPtr, LI->getName() + ".pre",
1226  LI->isVolatile(), LI->getAlignment(), LI->getOrdering(),
1227  LI->getSyncScopeID(), UnavailablePred->getTerminator());
1228  NewLoad->setDebugLoc(LI->getDebugLoc());
1229 
1230  // Transfer the old load's AA tags to the new load.
1231  AAMDNodes Tags;
1232  LI->getAAMetadata(Tags);
1233  if (Tags)
1234  NewLoad->setAAMetadata(Tags);
1235 
1236  if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1237  NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1238  if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1239  NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1240  if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1241  NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1242 
1243  // We do not propagate the old load's debug location, because the new
1244  // load now lives in a different BB, and we want to avoid a jumpy line
1245  // table.
1246  // FIXME: How do we retain source locations without causing poor debugging
1247  // behavior?
1248 
1249  // Add the newly created load.
1250  ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1251  NewLoad));
1252  MD->invalidateCachedPointerInfo(LoadPtr);
1253  LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1254  }
1255 
1256  // Perform PHI construction.
1257  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1258  LI->replaceAllUsesWith(V);
1259  if (isa<PHINode>(V))
1260  V->takeName(LI);
1261  if (Instruction *I = dyn_cast<Instruction>(V))
1262  I->setDebugLoc(LI->getDebugLoc());
1263  if (V->getType()->isPtrOrPtrVectorTy())
1264  MD->invalidateCachedPointerInfo(V);
1265  markInstructionForDeletion(LI);
1266  ORE->emit([&]() {
1267  return OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1268  << "load eliminated by PRE";
1269  });
1270  ++NumPRELoad;
1271  return true;
1272 }
1273 
1276  using namespace ore;
1277 
1278  ORE->emit([&]() {
1279  return OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1280  << "load of type " << NV("Type", LI->getType()) << " eliminated"
1281  << setExtraArgs() << " in favor of "
1282  << NV("InfavorOfValue", AvailableValue);
1283  });
1284 }
1285 
1286 /// Attempt to eliminate a load whose dependencies are
1287 /// non-local by performing PHI construction.
1288 bool GVN::processNonLocalLoad(LoadInst *LI) {
1289  // non-local speculations are not allowed under asan.
1290  if (LI->getParent()->getParent()->hasFnAttribute(
1291  Attribute::SanitizeAddress) ||
1293  Attribute::SanitizeHWAddress))
1294  return false;
1295 
1296  // Step 1: Find the non-local dependencies of the load.
1297  LoadDepVect Deps;
1298  MD->getNonLocalPointerDependency(LI, Deps);
1299 
1300  // If we had to process more than one hundred blocks to find the
1301  // dependencies, this load isn't worth worrying about. Optimizing
1302  // it will be too expensive.
1303  unsigned NumDeps = Deps.size();
1304  if (NumDeps > MaxNumDeps)
1305  return false;
1306 
1307  // If we had a phi translation failure, we'll have a single entry which is a
1308  // clobber in the current block. Reject this early.
1309  if (NumDeps == 1 &&
1310  !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1311  LLVM_DEBUG(dbgs() << "GVN: non-local load "; LI->printAsOperand(dbgs());
1312  dbgs() << " has unknown dependencies\n";);
1313  return false;
1314  }
1315 
1316  // If this load follows a GEP, see if we can PRE the indices before analyzing.
1317  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1318  for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1319  OE = GEP->idx_end();
1320  OI != OE; ++OI)
1321  if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1322  performScalarPRE(I);
1323  }
1324 
1325  // Step 2: Analyze the availability of the load
1326  AvailValInBlkVect ValuesPerBlock;
1327  UnavailBlkVect UnavailableBlocks;
1328  AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1329 
1330  // If we have no predecessors that produce a known value for this load, exit
1331  // early.
1332  if (ValuesPerBlock.empty())
1333  return false;
1334 
1335  // Step 3: Eliminate fully redundancy.
1336  //
1337  // If all of the instructions we depend on produce a known value for this
1338  // load, then it is fully redundant and we can use PHI insertion to compute
1339  // its value. Insert PHIs and remove the fully redundant value now.
1340  if (UnavailableBlocks.empty()) {
1341  LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1342 
1343  // Perform PHI construction.
1344  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1345  LI->replaceAllUsesWith(V);
1346 
1347  if (isa<PHINode>(V))
1348  V->takeName(LI);
1349  if (Instruction *I = dyn_cast<Instruction>(V))
1350  // If instruction I has debug info, then we should not update it.
1351  // Also, if I has a null DebugLoc, then it is still potentially incorrect
1352  // to propagate LI's DebugLoc because LI may not post-dominate I.
1353  if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1354  I->setDebugLoc(LI->getDebugLoc());
1355  if (V->getType()->isPtrOrPtrVectorTy())
1356  MD->invalidateCachedPointerInfo(V);
1357  markInstructionForDeletion(LI);
1358  ++NumGVNLoad;
1359  reportLoadElim(LI, V, ORE);
1360  return true;
1361  }
1362 
1363  // Step 4: Eliminate partial redundancy.
1364  if (!EnablePRE || !EnableLoadPRE)
1365  return false;
1366 
1367  return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1368 }
1369 
1370 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1371  assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1372  "This function can only be called with llvm.assume intrinsic");
1373  Value *V = IntrinsicI->getArgOperand(0);
1374 
1375  if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1376  if (Cond->isZero()) {
1377  Type *Int8Ty = Type::getInt8Ty(V->getContext());
1378  // Insert a new store to null instruction before the load to indicate that
1379  // this code is not reachable. FIXME: We could insert unreachable
1380  // instruction directly because we can modify the CFG.
1381  new StoreInst(UndefValue::get(Int8Ty),
1383  IntrinsicI);
1384  }
1385  markInstructionForDeletion(IntrinsicI);
1386  return false;
1387  } else if (isa<Constant>(V)) {
1388  // If it's not false, and constant, it must evaluate to true. This means our
1389  // assume is assume(true), and thus, pointless, and we don't want to do
1390  // anything more here.
1391  return false;
1392  }
1393 
1394  Constant *True = ConstantInt::getTrue(V->getContext());
1395  bool Changed = false;
1396 
1397  for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1398  BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1399 
1400  // This property is only true in dominated successors, propagateEquality
1401  // will check dominance for us.
1402  Changed |= propagateEquality(V, True, Edge, false);
1403  }
1404 
1405  // We can replace assume value with true, which covers cases like this:
1406  // call void @llvm.assume(i1 %cmp)
1407  // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1408  ReplaceWithConstMap[V] = True;
1409 
1410  // If one of *cmp *eq operand is const, adding it to map will cover this:
1411  // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1412  // call void @llvm.assume(i1 %cmp)
1413  // ret float %0 ; will change it to ret float 3.000000e+00
1414  if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1415  if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1416  CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1417  (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1418  CmpI->getFastMathFlags().noNaNs())) {
1419  Value *CmpLHS = CmpI->getOperand(0);
1420  Value *CmpRHS = CmpI->getOperand(1);
1421  if (isa<Constant>(CmpLHS))
1422  std::swap(CmpLHS, CmpRHS);
1423  auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1424 
1425  // If only one operand is constant.
1426  if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1427  ReplaceWithConstMap[CmpLHS] = RHSConst;
1428  }
1429  }
1430  return Changed;
1431 }
1432 
1434  patchReplacementInstruction(I, Repl);
1435  I->replaceAllUsesWith(Repl);
1436 }
1437 
1438 /// Attempt to eliminate a load, first by eliminating it
1439 /// locally, and then attempting non-local elimination if that fails.
1440 bool GVN::processLoad(LoadInst *L) {
1441  if (!MD)
1442  return false;
1443 
1444  // This code hasn't been audited for ordered or volatile memory access
1445  if (!L->isUnordered())
1446  return false;
1447 
1448  if (L->use_empty()) {
1449  markInstructionForDeletion(L);
1450  return true;
1451  }
1452 
1453  // ... to a pointer that has been loaded from before...
1454  MemDepResult Dep = MD->getDependency(L);
1455 
1456  // If it is defined in another block, try harder.
1457  if (Dep.isNonLocal())
1458  return processNonLocalLoad(L);
1459 
1460  // Only handle the local case below
1461  if (!Dep.isDef() && !Dep.isClobber()) {
1462  // This might be a NonFuncLocal or an Unknown
1463  LLVM_DEBUG(
1464  // fast print dep, using operator<< on instruction is too slow.
1465  dbgs() << "GVN: load "; L->printAsOperand(dbgs());
1466  dbgs() << " has unknown dependence\n";);
1467  return false;
1468  }
1469 
1470  AvailableValue AV;
1471  if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1472  Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1473 
1474  // Replace the load!
1475  patchAndReplaceAllUsesWith(L, AvailableValue);
1476  markInstructionForDeletion(L);
1477  ++NumGVNLoad;
1478  reportLoadElim(L, AvailableValue, ORE);
1479  // Tell MDA to rexamine the reused pointer since we might have more
1480  // information after forwarding it.
1481  if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1482  MD->invalidateCachedPointerInfo(AvailableValue);
1483  return true;
1484  }
1485 
1486  return false;
1487 }
1488 
1489 /// Return a pair the first field showing the value number of \p Exp and the
1490 /// second field showing whether it is a value number newly created.
1491 std::pair<uint32_t, bool>
1492 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1493  uint32_t &e = expressionNumbering[Exp];
1494  bool CreateNewValNum = !e;
1495  if (CreateNewValNum) {
1496  Expressions.push_back(Exp);
1497  if (ExprIdx.size() < nextValueNumber + 1)
1498  ExprIdx.resize(nextValueNumber * 2);
1499  e = nextValueNumber;
1500  ExprIdx[nextValueNumber++] = nextExprNumber++;
1501  }
1502  return {e, CreateNewValNum};
1503 }
1504 
1505 /// Return whether all the values related with the same \p num are
1506 /// defined in \p BB.
1507 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1508  GVN &Gvn) {
1509  LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1510  while (Vals && Vals->BB == BB)
1511  Vals = Vals->Next;
1512  return !Vals;
1513 }
1514 
1515 /// Wrap phiTranslateImpl to provide caching functionality.
1517  const BasicBlock *PhiBlock, uint32_t Num,
1518  GVN &Gvn) {
1519  auto FindRes = PhiTranslateTable.find({Num, Pred});
1520  if (FindRes != PhiTranslateTable.end())
1521  return FindRes->second;
1522  uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1523  PhiTranslateTable.insert({{Num, Pred}, NewNum});
1524  return NewNum;
1525 }
1526 
1527 /// Translate value number \p Num using phis, so that it has the values of
1528 /// the phis in BB.
1529 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1530  const BasicBlock *PhiBlock,
1531  uint32_t Num, GVN &Gvn) {
1532  if (PHINode *PN = NumberingPhi[Num]) {
1533  for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1534  if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1535  if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1536  return TransVal;
1537  }
1538  return Num;
1539  }
1540 
1541  // If there is any value related with Num is defined in a BB other than
1542  // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1543  // a backedge. We can do an early exit in that case to save compile time.
1544  if (!areAllValsInBB(Num, PhiBlock, Gvn))
1545  return Num;
1546 
1547  if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1548  return Num;
1549  Expression Exp = Expressions[ExprIdx[Num]];
1550 
1551  for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1552  // For InsertValue and ExtractValue, some varargs are index numbers
1553  // instead of value numbers. Those index numbers should not be
1554  // translated.
1555  if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1556  (i > 0 && Exp.opcode == Instruction::ExtractValue))
1557  continue;
1558  Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1559  }
1560 
1561  if (Exp.commutative) {
1562  assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1563  if (Exp.varargs[0] > Exp.varargs[1]) {
1564  std::swap(Exp.varargs[0], Exp.varargs[1]);
1565  uint32_t Opcode = Exp.opcode >> 8;
1566  if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1567  Exp.opcode = (Opcode << 8) |
1569  static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1570  }
1571  }
1572 
1573  if (uint32_t NewNum = expressionNumbering[Exp])
1574  return NewNum;
1575  return Num;
1576 }
1577 
1578 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1579 /// again.
1581  const BasicBlock &CurrBlock) {
1582  for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1583  auto FindRes = PhiTranslateTable.find({Num, Pred});
1584  if (FindRes != PhiTranslateTable.end())
1585  PhiTranslateTable.erase(FindRes);
1586  }
1587 }
1588 
1589 // In order to find a leader for a given value number at a
1590 // specific basic block, we first obtain the list of all Values for that number,
1591 // and then scan the list to find one whose block dominates the block in
1592 // question. This is fast because dominator tree queries consist of only
1593 // a few comparisons of DFS numbers.
1594 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1595  LeaderTableEntry Vals = LeaderTable[num];
1596  if (!Vals.Val) return nullptr;
1597 
1598  Value *Val = nullptr;
1599  if (DT->dominates(Vals.BB, BB)) {
1600  Val = Vals.Val;
1601  if (isa<Constant>(Val)) return Val;
1602  }
1603 
1604  LeaderTableEntry* Next = Vals.Next;
1605  while (Next) {
1606  if (DT->dominates(Next->BB, BB)) {
1607  if (isa<Constant>(Next->Val)) return Next->Val;
1608  if (!Val) Val = Next->Val;
1609  }
1610 
1611  Next = Next->Next;
1612  }
1613 
1614  return Val;
1615 }
1616 
1617 /// There is an edge from 'Src' to 'Dst'. Return
1618 /// true if every path from the entry block to 'Dst' passes via this edge. In
1619 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1621  DominatorTree *DT) {
1622  // While in theory it is interesting to consider the case in which Dst has
1623  // more than one predecessor, because Dst might be part of a loop which is
1624  // only reachable from Src, in practice it is pointless since at the time
1625  // GVN runs all such loops have preheaders, which means that Dst will have
1626  // been changed to have only one predecessor, namely Src.
1627  const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1628  assert((!Pred || Pred == E.getStart()) &&
1629  "No edge between these basic blocks!");
1630  return Pred != nullptr;
1631 }
1632 
1633 void GVN::assignBlockRPONumber(Function &F) {
1634  BlockRPONumber.clear();
1635  uint32_t NextBlockNumber = 1;
1637  for (BasicBlock *BB : RPOT)
1638  BlockRPONumber[BB] = NextBlockNumber++;
1639  InvalidBlockRPONumbers = false;
1640 }
1641 
1642 // Tries to replace instruction with const, using information from
1643 // ReplaceWithConstMap.
1644 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1645  bool Changed = false;
1646  for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1647  Value *Operand = Instr->getOperand(OpNum);
1648  auto it = ReplaceWithConstMap.find(Operand);
1649  if (it != ReplaceWithConstMap.end()) {
1650  assert(!isa<Constant>(Operand) &&
1651  "Replacing constants with constants is invalid");
1652  LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "
1653  << *it->second << " in instruction " << *Instr << '\n');
1654  Instr->setOperand(OpNum, it->second);
1655  Changed = true;
1656  }
1657  }
1658  return Changed;
1659 }
1660 
1661 /// The given values are known to be equal in every block
1662 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1663 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1664 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1665 /// value starting from the end of Root.Start.
1666 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1667  bool DominatesByEdge) {
1669  Worklist.push_back(std::make_pair(LHS, RHS));
1670  bool Changed = false;
1671  // For speed, compute a conservative fast approximation to
1672  // DT->dominates(Root, Root.getEnd());
1673  const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1674 
1675  while (!Worklist.empty()) {
1676  std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1677  LHS = Item.first; RHS = Item.second;
1678 
1679  if (LHS == RHS)
1680  continue;
1681  assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1682 
1683  // Don't try to propagate equalities between constants.
1684  if (isa<Constant>(LHS) && isa<Constant>(RHS))
1685  continue;
1686 
1687  // Prefer a constant on the right-hand side, or an Argument if no constants.
1688  if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1689  std::swap(LHS, RHS);
1690  assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1691 
1692  // If there is no obvious reason to prefer the left-hand side over the
1693  // right-hand side, ensure the longest lived term is on the right-hand side,
1694  // so the shortest lived term will be replaced by the longest lived.
1695  // This tends to expose more simplifications.
1696  uint32_t LVN = VN.lookupOrAdd(LHS);
1697  if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1698  (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1699  // Move the 'oldest' value to the right-hand side, using the value number
1700  // as a proxy for age.
1701  uint32_t RVN = VN.lookupOrAdd(RHS);
1702  if (LVN < RVN) {
1703  std::swap(LHS, RHS);
1704  LVN = RVN;
1705  }
1706  }
1707 
1708  // If value numbering later sees that an instruction in the scope is equal
1709  // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1710  // the invariant that instructions only occur in the leader table for their
1711  // own value number (this is used by removeFromLeaderTable), do not do this
1712  // if RHS is an instruction (if an instruction in the scope is morphed into
1713  // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1714  // using the leader table is about compiling faster, not optimizing better).
1715  // The leader table only tracks basic blocks, not edges. Only add to if we
1716  // have the simple case where the edge dominates the end.
1717  if (RootDominatesEnd && !isa<Instruction>(RHS))
1718  addToLeaderTable(LVN, RHS, Root.getEnd());
1719 
1720  // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1721  // LHS always has at least one use that is not dominated by Root, this will
1722  // never do anything if LHS has only one use.
1723  if (!LHS->hasOneUse()) {
1724  unsigned NumReplacements =
1725  DominatesByEdge
1726  ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1727  : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1728 
1729  Changed |= NumReplacements > 0;
1730  NumGVNEqProp += NumReplacements;
1731  // Cached information for anything that uses LHS will be invalid.
1732  if (MD)
1733  MD->invalidateCachedPointerInfo(LHS);
1734  }
1735 
1736  // Now try to deduce additional equalities from this one. For example, if
1737  // the known equality was "(A != B)" == "false" then it follows that A and B
1738  // are equal in the scope. Only boolean equalities with an explicit true or
1739  // false RHS are currently supported.
1740  if (!RHS->getType()->isIntegerTy(1))
1741  // Not a boolean equality - bail out.
1742  continue;
1743  ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1744  if (!CI)
1745  // RHS neither 'true' nor 'false' - bail out.
1746  continue;
1747  // Whether RHS equals 'true'. Otherwise it equals 'false'.
1748  bool isKnownTrue = CI->isMinusOne();
1749  bool isKnownFalse = !isKnownTrue;
1750 
1751  // If "A && B" is known true then both A and B are known true. If "A || B"
1752  // is known false then both A and B are known false.
1753  Value *A, *B;
1754  if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1755  (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1756  Worklist.push_back(std::make_pair(A, RHS));
1757  Worklist.push_back(std::make_pair(B, RHS));
1758  continue;
1759  }
1760 
1761  // If we are propagating an equality like "(A == B)" == "true" then also
1762  // propagate the equality A == B. When propagating a comparison such as
1763  // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1764  if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1765  Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1766 
1767  // If "A == B" is known true, or "A != B" is known false, then replace
1768  // A with B everywhere in the scope.
1769  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1770  (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1771  Worklist.push_back(std::make_pair(Op0, Op1));
1772 
1773  // Handle the floating point versions of equality comparisons too.
1774  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1775  (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1776 
1777  // Floating point -0.0 and 0.0 compare equal, so we can only
1778  // propagate values if we know that we have a constant and that
1779  // its value is non-zero.
1780 
1781  // FIXME: We should do this optimization if 'no signed zeros' is
1782  // applicable via an instruction-level fast-math-flag or some other
1783  // indicator that relaxed FP semantics are being used.
1784 
1785  if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1786  Worklist.push_back(std::make_pair(Op0, Op1));
1787  }
1788 
1789  // If "A >= B" is known true, replace "A < B" with false everywhere.
1790  CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1791  Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1792  // Since we don't have the instruction "A < B" immediately to hand, work
1793  // out the value number that it would have and use that to find an
1794  // appropriate instruction (if any).
1795  uint32_t NextNum = VN.getNextUnusedValueNumber();
1796  uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1797  // If the number we were assigned was brand new then there is no point in
1798  // looking for an instruction realizing it: there cannot be one!
1799  if (Num < NextNum) {
1800  Value *NotCmp = findLeader(Root.getEnd(), Num);
1801  if (NotCmp && isa<Instruction>(NotCmp)) {
1802  unsigned NumReplacements =
1803  DominatesByEdge
1804  ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1805  : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1806  Root.getStart());
1807  Changed |= NumReplacements > 0;
1808  NumGVNEqProp += NumReplacements;
1809  // Cached information for anything that uses NotCmp will be invalid.
1810  if (MD)
1811  MD->invalidateCachedPointerInfo(NotCmp);
1812  }
1813  }
1814  // Ensure that any instruction in scope that gets the "A < B" value number
1815  // is replaced with false.
1816  // The leader table only tracks basic blocks, not edges. Only add to if we
1817  // have the simple case where the edge dominates the end.
1818  if (RootDominatesEnd)
1819  addToLeaderTable(Num, NotVal, Root.getEnd());
1820 
1821  continue;
1822  }
1823  }
1824 
1825  return Changed;
1826 }
1827 
1828 /// When calculating availability, handle an instruction
1829 /// by inserting it into the appropriate sets
1830 bool GVN::processInstruction(Instruction *I) {
1831  // Ignore dbg info intrinsics.
1832  if (isa<DbgInfoIntrinsic>(I))
1833  return false;
1834 
1835  // If the instruction can be easily simplified then do so now in preference
1836  // to value numbering it. Value numbering often exposes redundancies, for
1837  // example if it determines that %y is equal to %x then the instruction
1838  // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1839  const DataLayout &DL = I->getModule()->getDataLayout();
1840  if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1841  bool Changed = false;
1842  if (!I->use_empty()) {
1843  I->replaceAllUsesWith(V);
1844  Changed = true;
1845  }
1846  if (isInstructionTriviallyDead(I, TLI)) {
1847  markInstructionForDeletion(I);
1848  Changed = true;
1849  }
1850  if (Changed) {
1851  if (MD && V->getType()->isPtrOrPtrVectorTy())
1852  MD->invalidateCachedPointerInfo(V);
1853  ++NumGVNSimpl;
1854  return true;
1855  }
1856  }
1857 
1858  if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1859  if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1860  return processAssumeIntrinsic(IntrinsicI);
1861 
1862  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1863  if (processLoad(LI))
1864  return true;
1865 
1866  unsigned Num = VN.lookupOrAdd(LI);
1867  addToLeaderTable(Num, LI, LI->getParent());
1868  return false;
1869  }
1870 
1871  // For conditional branches, we can perform simple conditional propagation on
1872  // the condition value itself.
1873  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1874  if (!BI->isConditional())
1875  return false;
1876 
1877  if (isa<Constant>(BI->getCondition()))
1878  return processFoldableCondBr(BI);
1879 
1880  Value *BranchCond = BI->getCondition();
1881  BasicBlock *TrueSucc = BI->getSuccessor(0);
1882  BasicBlock *FalseSucc = BI->getSuccessor(1);
1883  // Avoid multiple edges early.
1884  if (TrueSucc == FalseSucc)
1885  return false;
1886 
1887  BasicBlock *Parent = BI->getParent();
1888  bool Changed = false;
1889 
1890  Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1891  BasicBlockEdge TrueE(Parent, TrueSucc);
1892  Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1893 
1894  Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1895  BasicBlockEdge FalseE(Parent, FalseSucc);
1896  Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1897 
1898  return Changed;
1899  }
1900 
1901  // For switches, propagate the case values into the case destinations.
1902  if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1903  Value *SwitchCond = SI->getCondition();
1904  BasicBlock *Parent = SI->getParent();
1905  bool Changed = false;
1906 
1907  // Remember how many outgoing edges there are to every successor.
1909  for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1910  ++SwitchEdges[SI->getSuccessor(i)];
1911 
1912  for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1913  i != e; ++i) {
1914  BasicBlock *Dst = i->getCaseSuccessor();
1915  // If there is only a single edge, propagate the case value into it.
1916  if (SwitchEdges.lookup(Dst) == 1) {
1917  BasicBlockEdge E(Parent, Dst);
1918  Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1919  }
1920  }
1921  return Changed;
1922  }
1923 
1924  // Instructions with void type don't return a value, so there's
1925  // no point in trying to find redundancies in them.
1926  if (I->getType()->isVoidTy())
1927  return false;
1928 
1929  uint32_t NextNum = VN.getNextUnusedValueNumber();
1930  unsigned Num = VN.lookupOrAdd(I);
1931 
1932  // Allocations are always uniquely numbered, so we can save time and memory
1933  // by fast failing them.
1934  if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) {
1935  addToLeaderTable(Num, I, I->getParent());
1936  return false;
1937  }
1938 
1939  // If the number we were assigned was a brand new VN, then we don't
1940  // need to do a lookup to see if the number already exists
1941  // somewhere in the domtree: it can't!
1942  if (Num >= NextNum) {
1943  addToLeaderTable(Num, I, I->getParent());
1944  return false;
1945  }
1946 
1947  // Perform fast-path value-number based elimination of values inherited from
1948  // dominators.
1949  Value *Repl = findLeader(I->getParent(), Num);
1950  if (!Repl) {
1951  // Failure, just remember this instance for future use.
1952  addToLeaderTable(Num, I, I->getParent());
1953  return false;
1954  } else if (Repl == I) {
1955  // If I was the result of a shortcut PRE, it might already be in the table
1956  // and the best replacement for itself. Nothing to do.
1957  return false;
1958  }
1959 
1960  // Remove it!
1961  patchAndReplaceAllUsesWith(I, Repl);
1962  if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1963  MD->invalidateCachedPointerInfo(Repl);
1964  markInstructionForDeletion(I);
1965  return true;
1966 }
1967 
1968 /// runOnFunction - This is the main transformation entry point for a function.
1969 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1970  const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1971  MemoryDependenceResults *RunMD, LoopInfo *LI,
1972  OptimizationRemarkEmitter *RunORE) {
1973  AC = &RunAC;
1974  DT = &RunDT;
1975  VN.setDomTree(DT);
1976  TLI = &RunTLI;
1977  VN.setAliasAnalysis(&RunAA);
1978  MD = RunMD;
1979  ImplicitControlFlowTracking ImplicitCFT(DT);
1980  ICF = &ImplicitCFT;
1981  VN.setMemDep(MD);
1982  ORE = RunORE;
1983  InvalidBlockRPONumbers = true;
1984 
1985  bool Changed = false;
1986  bool ShouldContinue = true;
1987 
1989  // Merge unconditional branches, allowing PRE to catch more
1990  // optimization opportunities.
1991  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1992  BasicBlock *BB = &*FI++;
1993 
1994  bool removedBlock = MergeBlockIntoPredecessor(BB, &DTU, LI, nullptr, MD);
1995  if (removedBlock)
1996  ++NumGVNBlocks;
1997 
1998  Changed |= removedBlock;
1999  }
2000 
2001  unsigned Iteration = 0;
2002  while (ShouldContinue) {
2003  LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2004  ShouldContinue = iterateOnFunction(F);
2005  Changed |= ShouldContinue;
2006  ++Iteration;
2007  }
2008 
2009  if (EnablePRE) {
2010  // Fabricate val-num for dead-code in order to suppress assertion in
2011  // performPRE().
2012  assignValNumForDeadCode();
2013  bool PREChanged = true;
2014  while (PREChanged) {
2015  PREChanged = performPRE(F);
2016  Changed |= PREChanged;
2017  }
2018  }
2019 
2020  // FIXME: Should perform GVN again after PRE does something. PRE can move
2021  // computations into blocks where they become fully redundant. Note that
2022  // we can't do this until PRE's critical edge splitting updates memdep.
2023  // Actually, when this happens, we should just fully integrate PRE into GVN.
2024 
2025  cleanupGlobalSets();
2026  // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2027  // iteration.
2028  DeadBlocks.clear();
2029 
2030  return Changed;
2031 }
2032 
2033 bool GVN::processBlock(BasicBlock *BB) {
2034  // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2035  // (and incrementing BI before processing an instruction).
2036  assert(InstrsToErase.empty() &&
2037  "We expect InstrsToErase to be empty across iterations");
2038  if (DeadBlocks.count(BB))
2039  return false;
2040 
2041  // Clearing map before every BB because it can be used only for single BB.
2042  ReplaceWithConstMap.clear();
2043  bool ChangedFunction = false;
2044 
2045  for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2046  BI != BE;) {
2047  if (!ReplaceWithConstMap.empty())
2048  ChangedFunction |= replaceOperandsWithConsts(&*BI);
2049  ChangedFunction |= processInstruction(&*BI);
2050 
2051  if (InstrsToErase.empty()) {
2052  ++BI;
2053  continue;
2054  }
2055 
2056  // If we need some instructions deleted, do it now.
2057  NumGVNInstr += InstrsToErase.size();
2058 
2059  // Avoid iterator invalidation.
2060  bool AtStart = BI == BB->begin();
2061  if (!AtStart)
2062  --BI;
2063 
2064  for (auto *I : InstrsToErase) {
2065  assert(I->getParent() == BB && "Removing instruction from wrong block?");
2066  LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n');
2067  salvageDebugInfo(*I);
2068  if (MD) MD->removeInstruction(I);
2069  LLVM_DEBUG(verifyRemoved(I));
2070  ICF->removeInstruction(I);
2071  I->eraseFromParent();
2072  }
2073  InstrsToErase.clear();
2074 
2075  if (AtStart)
2076  BI = BB->begin();
2077  else
2078  ++BI;
2079  }
2080 
2081  return ChangedFunction;
2082 }
2083 
2084 // Instantiate an expression in a predecessor that lacked it.
2085 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2086  BasicBlock *Curr, unsigned int ValNo) {
2087  // Because we are going top-down through the block, all value numbers
2088  // will be available in the predecessor by the time we need them. Any
2089  // that weren't originally present will have been instantiated earlier
2090  // in this loop.
2091  bool success = true;
2092  for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2093  Value *Op = Instr->getOperand(i);
2094  if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2095  continue;
2096  // This could be a newly inserted instruction, in which case, we won't
2097  // find a value number, and should give up before we hurt ourselves.
2098  // FIXME: Rewrite the infrastructure to let it easier to value number
2099  // and process newly inserted instructions.
2100  if (!VN.exists(Op)) {
2101  success = false;
2102  break;
2103  }
2104  uint32_t TValNo =
2105  VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2106  if (Value *V = findLeader(Pred, TValNo)) {
2107  Instr->setOperand(i, V);
2108  } else {
2109  success = false;
2110  break;
2111  }
2112  }
2113 
2114  // Fail out if we encounter an operand that is not available in
2115  // the PRE predecessor. This is typically because of loads which
2116  // are not value numbered precisely.
2117  if (!success)
2118  return false;
2119 
2120  Instr->insertBefore(Pred->getTerminator());
2121  Instr->setName(Instr->getName() + ".pre");
2122  Instr->setDebugLoc(Instr->getDebugLoc());
2123 
2124  unsigned Num = VN.lookupOrAdd(Instr);
2125  VN.add(Instr, Num);
2126 
2127  // Update the availability map to include the new instruction.
2128  addToLeaderTable(Num, Instr, Pred);
2129  return true;
2130 }
2131 
2132 bool GVN::performScalarPRE(Instruction *CurInst) {
2133  if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() ||
2134  isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2135  CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2136  isa<DbgInfoIntrinsic>(CurInst))
2137  return false;
2138 
2139  // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2140  // sinking the compare again, and it would force the code generator to
2141  // move the i1 from processor flags or predicate registers into a general
2142  // purpose register.
2143  if (isa<CmpInst>(CurInst))
2144  return false;
2145 
2146  // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from
2147  // sinking the addressing mode computation back to its uses. Extending the
2148  // GEP's live range increases the register pressure, and therefore it can
2149  // introduce unnecessary spills.
2150  //
2151  // This doesn't prevent Load PRE. PHI translation will make the GEP available
2152  // to the load by moving it to the predecessor block if necessary.
2153  if (isa<GetElementPtrInst>(CurInst))
2154  return false;
2155 
2156  // We don't currently value number ANY inline asm calls.
2157  if (auto *CallB = dyn_cast<CallBase>(CurInst))
2158  if (CallB->isInlineAsm())
2159  return false;
2160 
2161  uint32_t ValNo = VN.lookup(CurInst);
2162 
2163  // Look for the predecessors for PRE opportunities. We're
2164  // only trying to solve the basic diamond case, where
2165  // a value is computed in the successor and one predecessor,
2166  // but not the other. We also explicitly disallow cases
2167  // where the successor is its own predecessor, because they're
2168  // more complicated to get right.
2169  unsigned NumWith = 0;
2170  unsigned NumWithout = 0;
2171  BasicBlock *PREPred = nullptr;
2172  BasicBlock *CurrentBlock = CurInst->getParent();
2173 
2174  // Update the RPO numbers for this function.
2175  if (InvalidBlockRPONumbers)
2176  assignBlockRPONumber(*CurrentBlock->getParent());
2177 
2179  for (BasicBlock *P : predecessors(CurrentBlock)) {
2180  // We're not interested in PRE where blocks with predecessors that are
2181  // not reachable.
2182  if (!DT->isReachableFromEntry(P)) {
2183  NumWithout = 2;
2184  break;
2185  }
2186  // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2187  // when CurInst has operand defined in CurrentBlock (so it may be defined
2188  // by phi in the loop header).
2189  assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) &&
2190  "Invalid BlockRPONumber map.");
2191  if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2192  llvm::any_of(CurInst->operands(), [&](const Use &U) {
2193  if (auto *Inst = dyn_cast<Instruction>(U.get()))
2194  return Inst->getParent() == CurrentBlock;
2195  return false;
2196  })) {
2197  NumWithout = 2;
2198  break;
2199  }
2200 
2201  uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2202  Value *predV = findLeader(P, TValNo);
2203  if (!predV) {
2204  predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2205  PREPred = P;
2206  ++NumWithout;
2207  } else if (predV == CurInst) {
2208  /* CurInst dominates this predecessor. */
2209  NumWithout = 2;
2210  break;
2211  } else {
2212  predMap.push_back(std::make_pair(predV, P));
2213  ++NumWith;
2214  }
2215  }
2216 
2217  // Don't do PRE when it might increase code size, i.e. when
2218  // we would need to insert instructions in more than one pred.
2219  if (NumWithout > 1 || NumWith == 0)
2220  return false;
2221 
2222  // We may have a case where all predecessors have the instruction,
2223  // and we just need to insert a phi node. Otherwise, perform
2224  // insertion.
2225  Instruction *PREInstr = nullptr;
2226 
2227  if (NumWithout != 0) {
2228  if (!isSafeToSpeculativelyExecute(CurInst)) {
2229  // It is only valid to insert a new instruction if the current instruction
2230  // is always executed. An instruction with implicit control flow could
2231  // prevent us from doing it. If we cannot speculate the execution, then
2232  // PRE should be prohibited.
2233  if (ICF->isDominatedByICFIFromSameBlock(CurInst))
2234  return false;
2235  }
2236 
2237  // Don't do PRE across indirect branch.
2238  if (isa<IndirectBrInst>(PREPred->getTerminator()))
2239  return false;
2240 
2241  // Don't do PRE across callbr.
2242  // FIXME: Can we do this across the fallthrough edge?
2243  if (isa<CallBrInst>(PREPred->getTerminator()))
2244  return false;
2245 
2246  // We can't do PRE safely on a critical edge, so instead we schedule
2247  // the edge to be split and perform the PRE the next time we iterate
2248  // on the function.
2249  unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2250  if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2251  toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2252  return false;
2253  }
2254  // We need to insert somewhere, so let's give it a shot
2255  PREInstr = CurInst->clone();
2256  if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2257  // If we failed insertion, make sure we remove the instruction.
2258  LLVM_DEBUG(verifyRemoved(PREInstr));
2259  PREInstr->deleteValue();
2260  return false;
2261  }
2262  }
2263 
2264  // Either we should have filled in the PRE instruction, or we should
2265  // not have needed insertions.
2266  assert(PREInstr != nullptr || NumWithout == 0);
2267 
2268  ++NumGVNPRE;
2269 
2270  // Create a PHI to make the value available in this block.
2271  PHINode *Phi =
2272  PHINode::Create(CurInst->getType(), predMap.size(),
2273  CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2274  for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2275  if (Value *V = predMap[i].first) {
2276  // If we use an existing value in this phi, we have to patch the original
2277  // value because the phi will be used to replace a later value.
2278  patchReplacementInstruction(CurInst, V);
2279  Phi->addIncoming(V, predMap[i].second);
2280  } else
2281  Phi->addIncoming(PREInstr, PREPred);
2282  }
2283 
2284  VN.add(Phi, ValNo);
2285  // After creating a new PHI for ValNo, the phi translate result for ValNo will
2286  // be changed, so erase the related stale entries in phi translate cache.
2287  VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2288  addToLeaderTable(ValNo, Phi, CurrentBlock);
2289  Phi->setDebugLoc(CurInst->getDebugLoc());
2290  CurInst->replaceAllUsesWith(Phi);
2291  if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2292  MD->invalidateCachedPointerInfo(Phi);
2293  VN.erase(CurInst);
2294  removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2295 
2296  LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2297  if (MD)
2298  MD->removeInstruction(CurInst);
2299  LLVM_DEBUG(verifyRemoved(CurInst));
2300  // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes
2301  // some assertion failures.
2302  ICF->removeInstruction(CurInst);
2303  CurInst->eraseFromParent();
2304  ++NumGVNInstr;
2305 
2306  return true;
2307 }
2308 
2309 /// Perform a purely local form of PRE that looks for diamond
2310 /// control flow patterns and attempts to perform simple PRE at the join point.
2311 bool GVN::performPRE(Function &F) {
2312  bool Changed = false;
2313  for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2314  // Nothing to PRE in the entry block.
2315  if (CurrentBlock == &F.getEntryBlock())
2316  continue;
2317 
2318  // Don't perform PRE on an EH pad.
2319  if (CurrentBlock->isEHPad())
2320  continue;
2321 
2322  for (BasicBlock::iterator BI = CurrentBlock->begin(),
2323  BE = CurrentBlock->end();
2324  BI != BE;) {
2325  Instruction *CurInst = &*BI++;
2326  Changed |= performScalarPRE(CurInst);
2327  }
2328  }
2329 
2330  if (splitCriticalEdges())
2331  Changed = true;
2332 
2333  return Changed;
2334 }
2335 
2336 /// Split the critical edge connecting the given two blocks, and return
2337 /// the block inserted to the critical edge.
2338 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2339  BasicBlock *BB =
2341  if (MD)
2342  MD->invalidateCachedPredecessors();
2343  InvalidBlockRPONumbers = true;
2344  return BB;
2345 }
2346 
2347 /// Split critical edges found during the previous
2348 /// iteration that may enable further optimization.
2349 bool GVN::splitCriticalEdges() {
2350  if (toSplit.empty())
2351  return false;
2352  do {
2353  std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val();
2354  SplitCriticalEdge(Edge.first, Edge.second,
2356  } while (!toSplit.empty());
2357  if (MD) MD->invalidateCachedPredecessors();
2358  InvalidBlockRPONumbers = true;
2359  return true;
2360 }
2361 
2362 /// Executes one iteration of GVN
2363 bool GVN::iterateOnFunction(Function &F) {
2364  cleanupGlobalSets();
2365 
2366  // Top-down walk of the dominator tree
2367  bool Changed = false;
2368  // Needed for value numbering with phi construction to work.
2369  // RPOT walks the graph in its constructor and will not be invalidated during
2370  // processBlock.
2372 
2373  for (BasicBlock *BB : RPOT)
2374  Changed |= processBlock(BB);
2375 
2376  return Changed;
2377 }
2378 
2379 void GVN::cleanupGlobalSets() {
2380  VN.clear();
2381  LeaderTable.clear();
2382  BlockRPONumber.clear();
2383  TableAllocator.Reset();
2384  ICF->clear();
2385  InvalidBlockRPONumbers = true;
2386 }
2387 
2388 /// Verify that the specified instruction does not occur in our
2389 /// internal data structures.
2390 void GVN::verifyRemoved(const Instruction *Inst) const {
2391  VN.verifyRemoved(Inst);
2392 
2393  // Walk through the value number scope to make sure the instruction isn't
2394  // ferreted away in it.
2396  I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2397  const LeaderTableEntry *Node = &I->second;
2398  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2399 
2400  while (Node->Next) {
2401  Node = Node->Next;
2402  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2403  }
2404  }
2405 }
2406 
2407 /// BB is declared dead, which implied other blocks become dead as well. This
2408 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2409 /// live successors, update their phi nodes by replacing the operands
2410 /// corresponding to dead blocks with UndefVal.
2411 void GVN::addDeadBlock(BasicBlock *BB) {
2414 
2415  NewDead.push_back(BB);
2416  while (!NewDead.empty()) {
2417  BasicBlock *D = NewDead.pop_back_val();
2418  if (DeadBlocks.count(D))
2419  continue;
2420 
2421  // All blocks dominated by D are dead.
2423  DT->getDescendants(D, Dom);
2424  DeadBlocks.insert(Dom.begin(), Dom.end());
2425 
2426  // Figure out the dominance-frontier(D).
2427  for (BasicBlock *B : Dom) {
2428  for (BasicBlock *S : successors(B)) {
2429  if (DeadBlocks.count(S))
2430  continue;
2431 
2432  bool AllPredDead = true;
2433  for (BasicBlock *P : predecessors(S))
2434  if (!DeadBlocks.count(P)) {
2435  AllPredDead = false;
2436  break;
2437  }
2438 
2439  if (!AllPredDead) {
2440  // S could be proved dead later on. That is why we don't update phi
2441  // operands at this moment.
2442  DF.insert(S);
2443  } else {
2444  // While S is not dominated by D, it is dead by now. This could take
2445  // place if S already have a dead predecessor before D is declared
2446  // dead.
2447  NewDead.push_back(S);
2448  }
2449  }
2450  }
2451  }
2452 
2453  // For the dead blocks' live successors, update their phi nodes by replacing
2454  // the operands corresponding to dead blocks with UndefVal.
2456  I != E; I++) {
2457  BasicBlock *B = *I;
2458  if (DeadBlocks.count(B))
2459  continue;
2460 
2462  for (BasicBlock *P : Preds) {
2463  if (!DeadBlocks.count(P))
2464  continue;
2465 
2466  if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2467  if (BasicBlock *S = splitCriticalEdges(P, B))
2468  DeadBlocks.insert(P = S);
2469  }
2470 
2471  for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2472  PHINode &Phi = cast<PHINode>(*II);
2474  UndefValue::get(Phi.getType()));
2475  if (MD)
2476  MD->invalidateCachedPointerInfo(&Phi);
2477  }
2478  }
2479  }
2480 }
2481 
2482 // If the given branch is recognized as a foldable branch (i.e. conditional
2483 // branch with constant condition), it will perform following analyses and
2484 // transformation.
2485 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2486 // R be the target of the dead out-coming edge.
2487 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2488 // edge. The result of this step will be {X| X is dominated by R}
2489 // 2) Identify those blocks which haves at least one dead predecessor. The
2490 // result of this step will be dominance-frontier(R).
2491 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2492 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2493 //
2494 // Return true iff *NEW* dead code are found.
2495 bool GVN::processFoldableCondBr(BranchInst *BI) {
2496  if (!BI || BI->isUnconditional())
2497  return false;
2498 
2499  // If a branch has two identical successors, we cannot declare either dead.
2500  if (BI->getSuccessor(0) == BI->getSuccessor(1))
2501  return false;
2502 
2503  ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2504  if (!Cond)
2505  return false;
2506 
2507  BasicBlock *DeadRoot =
2508  Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2509  if (DeadBlocks.count(DeadRoot))
2510  return false;
2511 
2512  if (!DeadRoot->getSinglePredecessor())
2513  DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2514 
2515  addDeadBlock(DeadRoot);
2516  return true;
2517 }
2518 
2519 // performPRE() will trigger assert if it comes across an instruction without
2520 // associated val-num. As it normally has far more live instructions than dead
2521 // instructions, it makes more sense just to "fabricate" a val-number for the
2522 // dead code than checking if instruction involved is dead or not.
2523 void GVN::assignValNumForDeadCode() {
2524  for (BasicBlock *BB : DeadBlocks) {
2525  for (Instruction &Inst : *BB) {
2526  unsigned ValNum = VN.lookupOrAdd(&Inst);
2527  addToLeaderTable(ValNum, &Inst, BB);
2528  }
2529  }
2530 }
2531 
2533 public:
2534  static char ID; // Pass identification, replacement for typeid
2535 
2536  explicit GVNLegacyPass(bool NoMemDepAnalysis = !EnableMemDep)
2537  : FunctionPass(ID), NoMemDepAnalysis(NoMemDepAnalysis) {
2539  }
2540 
2541  bool runOnFunction(Function &F) override {
2542  if (skipFunction(F))
2543  return false;
2544 
2545  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2546 
2547  return Impl.runImpl(
2548  F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2549  getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2550  getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2551  getAnalysis<AAResultsWrapperPass>().getAAResults(),
2552  NoMemDepAnalysis ? nullptr
2553  : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2554  LIWP ? &LIWP->getLoopInfo() : nullptr,
2555  &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2556  }
2557 
2558  void getAnalysisUsage(AnalysisUsage &AU) const override {
2562  if (!NoMemDepAnalysis)
2565 
2570  }
2571 
2572 private:
2573  bool NoMemDepAnalysis;
2574  GVN Impl;
2575 };
2576 
2577 char GVNLegacyPass::ID = 0;
2578 
2579 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2587 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2588 
2589 // The public interface to this file...
2590 FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) {
2591  return new GVNLegacyPass(NoMemDepAnalysis);
2592 }
Legacy wrapper pass to provide the GlobalsAAResult object.
static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV)
Definition: GVN.cpp:244
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:748
uint64_t CallInst * C
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:67
FunctionPass * createGVNPass(bool NoLoads=false)
Create a legacy GVN pass.
Definition: GVN.cpp:2590
static cl::opt< bool > EnableLoadPRE("enable-load-pre", cl::init(true))
void eraseTranslateCacheEntry(uint32_t Num, const BasicBlock &CurrBlock)
Erase stale entry from phiTranslate cache so phiTranslate can be computed again.
Definition: GVN.cpp:1580
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
bool isUndefValue() const
Definition: GVN.cpp:212
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:594
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:70
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:636
static bool runImpl(Function &F, TargetLibraryInfo &TLI, DominatorTree &DT)
This is the entry point for all transforms.
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
raw_ostream & errs()
This returns a reference to a raw_ostream for standard error.
Helper class for SSA formation on a set of values defined in multiple blocks.
Definition: SSAUpdater.h:38
Diagnostic information for missed-optimization remarks.
Provides a lazy, caching interface for making common memory aliasing information queries, backed by LLVM&#39;s alias analysis passes.
int analyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, LoadInst *DepLI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the load at De...
Definition: VNCoercion.cpp:242
DILocation * get() const
Get the underlying DILocation.
Definition: DebugLoc.cpp:21
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
This instruction extracts a struct member or array element value from an aggregate value...
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset=0)
Definition: GVN.cpp:185
size_type size() const
Definition: MapVector.h:60
unsigned Offset
Offset - The byte offset in Val that is interesting for the load query.
Definition: GVN.cpp:175
DiagnosticInfoOptimizationBase::Argument NV
Value * getRHS() const
static Type * makeCmpResultType(Type *opnd_type)
Create a result type for fcmp/icmp.
Definition: InstrTypes.h:889
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:769
This class represents lattice values for constants.
Definition: AllocatorList.h:23
PointerTy getPointer() const
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:473
bool isAtomic() const
Return true if this instruction has an AtomicOrdering of unordered or higher.
This is the interface for a simple mod/ref and alias analysis over globals.
void Initialize(Type *Ty, StringRef Name)
Reset this object to get ready for a new set of SSA updates with type &#39;Ty&#39;.
Definition: SSAUpdater.cpp:53
This class represents a op.with.overflow intrinsic.
bool MergeBlockIntoPredecessor(BasicBlock *BB, DomTreeUpdater *DTU=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, MemoryDependenceResults *MemDep=nullptr)
Attempts to merge a block into its predecessor, if possible.
uint32_t lookupOrAddCmp(unsigned Opcode, CmpInst::Predicate Pred, Value *LHS, Value *RHS)
Returns the value number of the given comparison, assigning it a new number if it did not have one be...
Definition: GVN.cpp:567
Instruction::BinaryOps getBinaryOp() const
Returns the binary operation underlying the intrinsic.
iterator end()
Definition: Function.h:660
void AddAvailableValue(BasicBlock *BB, Value *V)
Indicate that a rewritten value is available in the specified block with the specified value...
Definition: SSAUpdater.cpp:71
bool operator==(const Expression &other) const
Definition: GVN.cpp:120
This class represents a function call, abstracting a target machine&#39;s calling convention.
bool isNonLocal() const
Tests if this MemDepResult represents a query that is transparent to the start of the block...
This file contains the declarations for metadata subclasses.
An immutable pass that tracks lazily created AssumptionCache objects.
A cache of @llvm.assume calls within a function.
bool salvageDebugInfo(Instruction &I)
Assuming the instruction I is going to be deleted, attempt to salvage debug users of I by writing the...
Definition: Local.cpp:1603
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:247
uint32_t phiTranslate(const BasicBlock *BB, const BasicBlock *PhiBlock, uint32_t Num, GVN &Gvn)
Wrap phiTranslateImpl to provide caching functionality.
Definition: GVN.cpp:1516
bool isTerminator() const
Definition: Instruction.h:128
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:662
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:98
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.h:320
unsigned second
This class implements a map that also provides access to all stored values in a deterministic order...
Definition: MapVector.h:37
BasicBlock * getSuccessor(unsigned i) const
bool properlyDominates(const DomTreeNodeBase< NodeT > *A, const DomTreeNodeBase< NodeT > *B) const
properlyDominates - Returns true iff A dominates B and A != B.
STATISTIC(NumFunctions, "Total number of functions")
A debug info location.
Definition: DebugLoc.h:33
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:230
F(f)
bool isCoercedLoadValue() const
Definition: GVN.cpp:210
An instruction for reading from memory.
Definition: Instructions.h:167
const BasicBlock * getEnd() const
Definition: Dominators.h:94
Hexagon Common GEP
Value * getCondition() const
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:137
This defines the Use class.
idx_iterator idx_end() const
unsigned replaceDominatedUsesWith(Value *From, Value *To, DominatorTree &DT, const BasicBlockEdge &Edge)
Replace each use of &#39;From&#39; with &#39;To&#39; if that use is dominated by the given edge.
Definition: Local.cpp:2436
Use * op_iterator
Definition: User.h:224
iterator end()
Get an iterator to the end of the SetVector.
Definition: SetVector.h:92
Value * getMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingMemInst returned an offset, this function can be used to actually perform...
Definition: VNCoercion.cpp:515
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:32
op_iterator op_begin()
Definition: User.h:229
gvn Early GVN Hoisting of Expressions
Definition: GVNHoist.cpp:1203
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:274
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:268
uint32_t lookup(Value *V, bool Verify=true) const
Returns the value number of the specified value.
Definition: GVN.cpp:554
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1155
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:221
void dump() const
Support for debugging, callable in GDB: V->dump()
Definition: AsmWriter.cpp:4361
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:47
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:50
bool isVolatile() const
Return true if this is a load from a volatile memory location.
Definition: Instructions.h:231
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: GVN.cpp:2558
static cl::opt< bool > EnablePRE("enable-pre", cl::init(true), cl::Hidden)
void patchReplacementInstruction(Instruction *I, Value *Repl)
Patch the replacement so that it is not more restrictive than the value being replaced.
Definition: Local.cpp:2367
bool isDef() const
Tests if this MemDepResult represents a query that is an instruction definition dependency.
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:369
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: GVN.cpp:2541
Option class for critical edge splitting.
int getBasicBlockIndex(const BasicBlock *BB) const
Return the first index of the specified basic block in the value list for this PHI.
void clear()
Remove all entries from the ValueTable.
Definition: GVN.cpp:575
bool isClobber() const
Tests if this MemDepResult represents a query that is an instruction clobber dependency.
A Use represents the edge between a Value definition and its users.
Definition: Use.h:55
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:651
int analyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, MemIntrinsic *DepMI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the memory int...
Definition: VNCoercion.cpp:279
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:41
MemoryDependenceResults & getMemDep() const
Definition: GVN.h:84
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:196
This file contains the simple types necessary to represent the attributes associated with functions a...
An analysis that produces MemoryDependenceResults for a function.
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:285
Analysis pass that exposes the LoopInfo for a function.
Definition: LoopInfo.h:949
static const uint16_t * lookup(unsigned opcode, unsigned domain, ArrayRef< uint16_t[3]> Table)
bool isSimpleValue() const
Definition: GVN.cpp:209
Interval::succ_iterator succ_begin(Interval *I)
succ_begin/succ_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:102
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:244
ppc ctr loops PowerPC CTR Loops Verify
bool insert(const value_type &X)
Insert a new element into the SetVector.
Definition: SetVector.h:141
The core GVN pass object.
Definition: GVN.h:68
IntType getInt() const
bool canCoerceMustAliasedValueToLoad(Value *StoredVal, Type *LoadTy, const DataLayout &DL)
Return true if CoerceAvailableValueToLoadType would succeed if it was called.
Definition: VNCoercion.cpp:15
Expression(uint32_t o=~2U)
Definition: GVN.cpp:118
#define DEBUG_TYPE
Definition: GVN.cpp:88
iterator begin()
Get an iterator to the beginning of the SetVector.
Definition: SetVector.h:82
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:234
DiagnosticInfoOptimizationBase::setExtraArgs setExtraArgs
static AvailableValue getLoad(LoadInst *LI, unsigned Offset=0)
Definition: GVN.cpp:193
hash_code hash_value(const APFloat &Arg)
See friend declarations above.
Definition: APFloat.cpp:4433
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
BasicBlock * SplitCriticalEdge(Instruction *TI, unsigned SuccNum, const CriticalEdgeSplittingOptions &Options=CriticalEdgeSplittingOptions())
If this edge is a critical edge, insert a new node to split the critical edge.
LoadInst * getCoercedLoadValue() const
Definition: GVN.cpp:219
static GVN::Expression getEmptyKey()
Definition: GVN.cpp:142
An instruction for storing to memory.
Definition: Instructions.h:320
bool isMinusOne() const
This function will return true iff every bit in this constant is set to true.
Definition: Constants.h:208
void add(Value *V, uint32_t num)
add - Insert a value into the table with a specified value number.
Definition: GVN.cpp:368
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:429
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:291
iterator begin()
Definition: Function.h:658
static unsigned getHashValue(const GVN::Expression &e)
Definition: GVN.cpp:145
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:144
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Value * getOperand(unsigned i) const
Definition: User.h:169
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:105
int analyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, StoreInst *DepSI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the store at D...
Definition: VNCoercion.cpp:214
void initializeGVNLegacyPassPass(PassRegistry &)
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:140
const BasicBlock & getEntryBlock() const
Definition: Function.h:642
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:873
void getAAMetadata(AAMDNodes &N, bool Merge=false) const
Fills the AAMDNodes structure with AA metadata from this instruction.
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:427
Value * GetValueInMiddleOfBlock(BasicBlock *BB)
Construct SSA form, materializing a value that is live in the middle of the specified block...
Definition: SSAUpdater.cpp:99
SmallVector< uint32_t, 4 > varargs
Definition: GVN.cpp:116
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:148
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:321
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:233
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:73
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
PointerIntPair - This class implements a pair of a pointer and small integer.
PHITransAddr - An address value which tracks and handles phi translation.
Definition: PHITransAddr.h:35
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
Definition: PatternMatch.h:754
Conditional or Unconditional Branch instruction.
This file provides the interface for LLVM&#39;s Global Value Numbering pass which eliminates fully redund...
static GVN::Expression getTombstoneKey()
Definition: GVN.cpp:143
static Value * ConstructSSAForLoadSet(LoadInst *LI, SmallVectorImpl< AvailableValueInBlock > &ValuesPerBlock, GVN &gvn)
Given a set of loads specified by ValuesPerBlock, construct SSA form, allowing us to eliminate LI...
Definition: GVN.cpp:725
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:41
static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS)
Definition: GVN.cpp:151
This file contains the declarations for the subclasses of Constant, which represent the different fla...
static cl::opt< uint32_t > MaxRecurseDepth("gvn-max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, cl::desc("Max recurse depth in GVN (default = 1000)"))
const Instruction & front() const
Definition: BasicBlock.h:280
A manager for alias analyses.
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:575
Diagnostic information for applied optimization remarks.
Interval::pred_iterator pred_begin(Interval *I)
pred_begin/pred_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:112
unsigned getNumIndices() const
bool isUnordered() const
Definition: Instructions.h:278
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:231
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1192
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:646
PointerIntPair< Value *, 2, ValType > Val
V - The value that is live out of the block.
Definition: GVN.cpp:172
MemIntrinsic * getMemIntrinValue() const
Definition: GVN.cpp:224
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:115
op_range operands()
Definition: User.h:237
Value * getPointerOperand()
Definition: Instructions.h:284
bool isCriticalEdge(const Instruction *TI, unsigned SuccNum, bool AllowIdenticalEdges=false)
Return true if the specified edge is a critical edge.
Definition: CFG.cpp:88
Value * getLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingLoad returned an offset, this function can be used to actually perform th...
Definition: VNCoercion.cpp:403
static void reportLoadElim(LoadInst *LI, Value *AvailableValue, OptimizationRemarkEmitter *ORE)
Definition: GVN.cpp:1274
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1424
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
size_t size() const
Definition: SmallVector.h:52
static cl::opt< bool > EnableMemDep("enable-gvn-memdep", cl::init(true))
A wrapper analysis pass for the legacy pass manager that exposes a MemoryDepnedenceResults instance...
void printAsOperand(raw_ostream &O, bool PrintType=true, const Module *M=nullptr) const
Print the name of this Value out to the specified raw_ostream.
Definition: AsmWriter.cpp:4288
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
A memory dependence query can return one of three different answers.
DominatorTree & getDominatorTree() const
Definition: GVN.h:82
unsigned first
static cl::opt< uint32_t > MaxNumDeps("gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore, cl::desc("Max number of dependences to attempt Load PRE (default = 100)"))
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:50
static void reportMayClobberedLoad(LoadInst *LI, MemDepResult DepInfo, DominatorTree *DT, OptimizationRemarkEmitter *ORE)
Try to locate the three instruction involved in a missed load-elimination case that is due to an inte...
Definition: GVN.cpp:822
A function analysis which provides an AssumptionCache.
bool isPtrOrPtrVectorTy() const
Return true if this is a pointer type or a vector of pointer types.
Definition: Type.h:226
Value * MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const
Emit code at the end of this block to adjust the value defined here to the specified type...
Definition: GVN.cpp:262
A SetVector that performs no allocations if smaller than a certain size.
Definition: SetVector.h:297
This is the common base class for memset/memcpy/memmove.
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:191
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:417
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
void emit(DiagnosticInfoOptimizationBase &OptDiag)
Output the remark via the diagnostic handler and to the optimization record file. ...
iterator end()
Definition: BasicBlock.h:270
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:248
Module.h This file contains the declarations for the Module class.
Provides information about what library functions are available for the current target.
const MemDepResult & getResult() const
size_type count(const KeyT &Key) const
Definition: MapVector.h:142
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:47
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:643
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:374
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:631
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:124
Value * MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt, GVN &gvn) const
Emit code at the specified insertion point to adjust the value defined here to the specified type...
Definition: GVN.cpp:765
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:587
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:491
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:940
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:55
hash_code hash_combine(const Ts &...args)
Combine values into a single hash_code.
Definition: Hashing.h:600
Represents an AvailableValue which can be rematerialized at the end of the associated BasicBlock...
Definition: GVN.cpp:237
iterator_range< user_iterator > users()
Definition: Value.h:399
hash_code hash_combine_range(InputIteratorT first, InputIteratorT last)
Compute a hash_code for a sequence of values.
Definition: Hashing.h:478
std::vector< NonLocalDepEntry > NonLocalDepInfo
An opaque object representing a hash code.
Definition: Hashing.h:71
bool isMallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates uninitialized memory (such ...
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:471
void verifyRemoved(const Value *) const
verifyRemoved - Verify that the value is removed from all internal data structures.
Definition: GVN.cpp:597
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:387
void erase(Value *v)
Remove a value from the value numbering.
Definition: GVN.cpp:587
static bool isLifetimeStart(const Instruction *Inst)
Definition: GVN.cpp:814
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:545
unsigned GetSuccessorNumber(const BasicBlock *BB, const BasicBlock *Succ)
Search for the specified successor of basic block BB and return its position in the terminator instru...
Definition: CFG.cpp:72
unsigned getNumArgOperands() const
Definition: InstrTypes.h:1153
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:324
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:240
Instruction * getInst() const
If this is a normal dependency, returns the instruction that is depended on.
void clear()
Definition: ilist.h:307
Value * getStoreValueForLoad(Value *SrcVal, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingStore returned an offset, this function can be used to actually perform t...
Definition: VNCoercion.cpp:383
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
GVNLegacyPass(bool NoMemDepAnalysis=!EnableMemDep)
Definition: GVN.cpp:2536
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:214
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
bool isCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates zero-filled memory (such as...
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this load instruction.
Definition: Instructions.h:259
#define I(x, y, z)
Definition: MD5.cpp:58
bool mayReadFromMemory() const
Return true if this instruction may read memory.
static AvailableValue get(Value *V, unsigned Offset=0)
Definition: GVN.cpp:177
uint32_t opcode
Definition: GVN.cpp:113
PassT::Result * getCachedResult(IRUnitT &IR) const
Get the cached result of an analysis pass for a given IR unit.
Definition: PassManager.h:788
bool exists(Value *V) const
Returns true if a value number exists for the specified value.
Definition: GVN.cpp:476
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:332
idx_iterator idx_begin() const
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
This class allows to keep track on instructions with implicit control flow.
bool isUnconditional() const
friend hash_code hash_value(const Expression &Value)
Definition: GVN.cpp:132
uint32_t lookupOrAdd(Value *V)
lookup_or_add - Returns the value number for the specified value, assigning it a new number if it did...
Definition: GVN.cpp:480
ValueT lookup(const_arg_type_t< KeyT > Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: DenseMap.h:211
Value * getSimpleValue() const
Definition: GVN.cpp:214
Analysis pass providing the TargetLibraryInfo.
iterator_range< df_iterator< T > > depth_first(const T &G)
Multiway switch.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
const BasicBlock * getStart() const
Definition: Dominators.h:90
Represents a particular available value that we know how to materialize.
Definition: GVN.cpp:162
bool isSafeToSpeculativelyExecute(const Value *V, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr)
Return true if the instruction does not have any effects besides calculating the result and does not ...
static bool IsValueFullyAvailableInBlock(BasicBlock *BB, DenseMap< BasicBlock *, char > &FullyAvailableBlocks, uint32_t RecurseDepth)
Return true if we can prove that the value we&#39;re analyzing is fully available in the specified block...
Definition: GVN.cpp:652
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:649
bool isInstructionTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction is not used, and the instruction has no side ef...
Definition: Local.cpp:349
LLVM Value Representation.
Definition: Value.h:72
static AvailableValueInBlock getUndef(BasicBlock *BB)
Definition: GVN.cpp:256
Value * getLHS() const
void removeInstruction(Instruction *InstToRemove)
Removes an instruction from the dependence analysis, updating the dependence of instructions that pre...
succ_range successors(Instruction *I)
Definition: CFG.h:259
OptimizationRemarkEmitter legacy analysis pass.
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: GVN.cpp:608
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:412
Predicate getSwappedPredicate() const
For example, EQ->EQ, SLE->SGE, ULT->UGT, OEQ->OEQ, ULE->UGE, OLT->OGT, etc.
Definition: InstrTypes.h:761
This is an entry in the NonLocalDepInfo cache.
A container for analyses that lazily runs them and caches their results.
BasicBlock * BB
BB - The basic block in question.
Definition: GVN.cpp:239
static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl)
Definition: GVN.cpp:1433
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:259
bool isMemIntrinValue() const
Definition: GVN.cpp:211
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
This header defines various interfaces for pass management in LLVM.
void setIncomingValue(unsigned i, Value *V)
AvailableValue AV
AV - The actual available value.
Definition: GVN.cpp:242
#define LLVM_DEBUG(X)
Definition: Debug.h:122
Value * SimplifyInstruction(Instruction *I, const SimplifyQuery &Q, OptimizationRemarkEmitter *ORE=nullptr)
See if we can compute a simplified version of this instruction.
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:173
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:322
static AvailableValue getUndef()
Definition: GVN.cpp:201
static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, DominatorTree *DT)
There is an edge from &#39;Src&#39; to &#39;Dst&#39;.
Definition: GVN.cpp:1620
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:43
const BasicBlock * getParent() const
Definition: Instruction.h:66
This instruction inserts a struct field of array element value into an aggregate value.
bool HasValueForBlock(BasicBlock *BB) const
Return true if the SSAUpdater already has a value for the specified block.
Definition: SSAUpdater.cpp:62