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::AddrSpaceCast:
527  case Instruction::BitCast:
528  case Instruction::Select:
529  case Instruction::ExtractElement:
530  case Instruction::InsertElement:
531  case Instruction::ShuffleVector:
532  case Instruction::InsertValue:
533  case Instruction::GetElementPtr:
534  exp = createExpr(I);
535  break;
536  case Instruction::ExtractValue:
537  exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
538  break;
539  case Instruction::PHI:
540  valueNumbering[V] = nextValueNumber;
541  NumberingPhi[nextValueNumber] = cast<PHINode>(V);
542  return nextValueNumber++;
543  default:
544  valueNumbering[V] = nextValueNumber;
545  return nextValueNumber++;
546  }
547 
548  uint32_t e = assignExpNewValueNum(exp).first;
549  valueNumbering[V] = e;
550  return e;
551 }
552 
553 /// Returns the value number of the specified value. Fails if
554 /// the value has not yet been numbered.
556  DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
557  if (Verify) {
558  assert(VI != valueNumbering.end() && "Value not numbered?");
559  return VI->second;
560  }
561  return (VI != valueNumbering.end()) ? VI->second : 0;
562 }
563 
564 /// Returns the value number of the given comparison,
565 /// assigning it a new number if it did not have one before. Useful when
566 /// we deduced the result of a comparison, but don't immediately have an
567 /// instruction realizing that comparison to hand.
569  CmpInst::Predicate Predicate,
570  Value *LHS, Value *RHS) {
571  Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
572  return assignExpNewValueNum(exp).first;
573 }
574 
575 /// Remove all entries from the ValueTable.
577  valueNumbering.clear();
578  expressionNumbering.clear();
579  NumberingPhi.clear();
580  PhiTranslateTable.clear();
581  nextValueNumber = 1;
582  Expressions.clear();
583  ExprIdx.clear();
584  nextExprNumber = 0;
585 }
586 
587 /// Remove a value from the value numbering.
589  uint32_t Num = valueNumbering.lookup(V);
590  valueNumbering.erase(V);
591  // If V is PHINode, V <--> value number is an one-to-one mapping.
592  if (isa<PHINode>(V))
593  NumberingPhi.erase(Num);
594 }
595 
596 /// verifyRemoved - Verify that the value is removed from all internal data
597 /// structures.
598 void GVN::ValueTable::verifyRemoved(const Value *V) const {
600  I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
601  assert(I->first != V && "Inst still occurs in value numbering map!");
602  }
603 }
604 
605 //===----------------------------------------------------------------------===//
606 // GVN Pass
607 //===----------------------------------------------------------------------===//
608 
610  // FIXME: The order of evaluation of these 'getResult' calls is very
611  // significant! Re-ordering these variables will cause GVN when run alone to
612  // be less effective! We should fix memdep and basic-aa to not exhibit this
613  // behavior, but until then don't change the order here.
614  auto &AC = AM.getResult<AssumptionAnalysis>(F);
615  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
616  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
617  auto &AA = AM.getResult<AAManager>(F);
618  auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
619  auto *LI = AM.getCachedResult<LoopAnalysis>(F);
621  bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
622  if (!Changed)
623  return PreservedAnalyses::all();
626  PA.preserve<GlobalsAA>();
628  return PA;
629 }
630 
631 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
632 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
633  errs() << "{\n";
635  E = d.end(); I != E; ++I) {
636  errs() << I->first << "\n";
637  I->second->dump();
638  }
639  errs() << "}\n";
640 }
641 #endif
642 
643 /// Return true if we can prove that the value
644 /// we're analyzing is fully available in the specified block. As we go, keep
645 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
646 /// map is actually a tri-state map with the following values:
647 /// 0) we know the block *is not* fully available.
648 /// 1) we know the block *is* fully available.
649 /// 2) we do not know whether the block is fully available or not, but we are
650 /// currently speculating that it will be.
651 /// 3) we are speculating for this block and have used that to speculate for
652 /// other blocks.
654  DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
655  uint32_t RecurseDepth) {
656  if (RecurseDepth > MaxRecurseDepth)
657  return false;
658 
659  // Optimistically assume that the block is fully available and check to see
660  // if we already know about this block in one lookup.
661  std::pair<DenseMap<BasicBlock*, char>::iterator, bool> IV =
662  FullyAvailableBlocks.insert(std::make_pair(BB, 2));
663 
664  // If the entry already existed for this block, return the precomputed value.
665  if (!IV.second) {
666  // If this is a speculative "available" value, mark it as being used for
667  // speculation of other blocks.
668  if (IV.first->second == 2)
669  IV.first->second = 3;
670  return IV.first->second != 0;
671  }
672 
673  // Otherwise, see if it is fully available in all predecessors.
674  pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
675 
676  // If this block has no predecessors, it isn't live-in here.
677  if (PI == PE)
678  goto SpeculationFailure;
679 
680  for (; PI != PE; ++PI)
681  // If the value isn't fully available in one of our predecessors, then it
682  // isn't fully available in this block either. Undo our previous
683  // optimistic assumption and bail out.
684  if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
685  goto SpeculationFailure;
686 
687  return true;
688 
689 // If we get here, we found out that this is not, after
690 // all, a fully-available block. We have a problem if we speculated on this and
691 // used the speculation to mark other blocks as available.
692 SpeculationFailure:
693  char &BBVal = FullyAvailableBlocks[BB];
694 
695  // If we didn't speculate on this, just return with it set to false.
696  if (BBVal == 2) {
697  BBVal = 0;
698  return false;
699  }
700 
701  // If we did speculate on this value, we could have blocks set to 1 that are
702  // incorrect. Walk the (transitive) successors of this block and mark them as
703  // 0 if set to one.
704  SmallVector<BasicBlock*, 32> BBWorklist;
705  BBWorklist.push_back(BB);
706 
707  do {
708  BasicBlock *Entry = BBWorklist.pop_back_val();
709  // Note that this sets blocks to 0 (unavailable) if they happen to not
710  // already be in FullyAvailableBlocks. This is safe.
711  char &EntryVal = FullyAvailableBlocks[Entry];
712  if (EntryVal == 0) continue; // Already unavailable.
713 
714  // Mark as unavailable.
715  EntryVal = 0;
716 
717  BBWorklist.append(succ_begin(Entry), succ_end(Entry));
718  } while (!BBWorklist.empty());
719 
720  return false;
721 }
722 
723 /// Given a set of loads specified by ValuesPerBlock,
724 /// construct SSA form, allowing us to eliminate LI. This returns the value
725 /// that should be used at LI's definition site.
728  GVN &gvn) {
729  // Check for the fully redundant, dominating load case. In this case, we can
730  // just use the dominating value directly.
731  if (ValuesPerBlock.size() == 1 &&
732  gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
733  LI->getParent())) {
734  assert(!ValuesPerBlock[0].AV.isUndefValue() &&
735  "Dead BB dominate this block");
736  return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
737  }
738 
739  // Otherwise, we have to construct SSA form.
740  SmallVector<PHINode*, 8> NewPHIs;
741  SSAUpdater SSAUpdate(&NewPHIs);
742  SSAUpdate.Initialize(LI->getType(), LI->getName());
743 
744  for (const AvailableValueInBlock &AV : ValuesPerBlock) {
745  BasicBlock *BB = AV.BB;
746 
747  if (SSAUpdate.HasValueForBlock(BB))
748  continue;
749 
750  // If the value is the load that we will be eliminating, and the block it's
751  // available in is the block that the load is in, then don't add it as
752  // SSAUpdater will resolve the value to the relevant phi which may let it
753  // avoid phi construction entirely if there's actually only one value.
754  if (BB == LI->getParent() &&
755  ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == LI) ||
756  (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == LI)))
757  continue;
758 
759  SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
760  }
761 
762  // Perform PHI construction.
763  return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
764 }
765 
767  Instruction *InsertPt,
768  GVN &gvn) const {
769  Value *Res;
770  Type *LoadTy = LI->getType();
771  const DataLayout &DL = LI->getModule()->getDataLayout();
772  if (isSimpleValue()) {
773  Res = getSimpleValue();
774  if (Res->getType() != LoadTy) {
775  Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
776 
777  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset
778  << " " << *getSimpleValue() << '\n'
779  << *Res << '\n'
780  << "\n\n\n");
781  }
782  } else if (isCoercedLoadValue()) {
783  LoadInst *Load = getCoercedLoadValue();
784  if (Load->getType() == LoadTy && Offset == 0) {
785  Res = Load;
786  } else {
787  Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
788  // We would like to use gvn.markInstructionForDeletion here, but we can't
789  // because the load is already memoized into the leader map table that GVN
790  // tracks. It is potentially possible to remove the load from the table,
791  // but then there all of the operations based on it would need to be
792  // rehashed. Just leave the dead load around.
793  gvn.getMemDep().removeInstruction(Load);
794  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset
795  << " " << *getCoercedLoadValue() << '\n'
796  << *Res << '\n'
797  << "\n\n\n");
798  }
799  } else if (isMemIntrinValue()) {
800  Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
801  InsertPt, DL);
802  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
803  << " " << *getMemIntrinValue() << '\n'
804  << *Res << '\n'
805  << "\n\n\n");
806  } else {
807  assert(isUndefValue() && "Should be UndefVal");
808  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
809  return UndefValue::get(LoadTy);
810  }
811  assert(Res && "failed to materialize?");
812  return Res;
813 }
814 
815 static bool isLifetimeStart(const Instruction *Inst) {
816  if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
817  return II->getIntrinsicID() == Intrinsic::lifetime_start;
818  return false;
819 }
820 
821 /// Try to locate the three instruction involved in a missed
822 /// load-elimination case that is due to an intervening store.
824  DominatorTree *DT,
826  using namespace ore;
827 
828  User *OtherAccess = nullptr;
829 
830  OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
831  R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
832  << setExtraArgs();
833 
834  for (auto *U : LI->getPointerOperand()->users())
835  if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
836  DT->dominates(cast<Instruction>(U), LI)) {
837  // FIXME: for now give up if there are multiple memory accesses that
838  // dominate the load. We need further analysis to decide which one is
839  // that we're forwarding from.
840  if (OtherAccess)
841  OtherAccess = nullptr;
842  else
843  OtherAccess = U;
844  }
845 
846  if (OtherAccess)
847  R << " in favor of " << NV("OtherAccess", OtherAccess);
848 
849  R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
850 
851  ORE->emit(R);
852 }
853 
854 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
855  Value *Address, AvailableValue &Res) {
856  assert((DepInfo.isDef() || DepInfo.isClobber()) &&
857  "expected a local dependence");
858  assert(LI->isUnordered() && "rules below are incorrect for ordered access");
859 
860  const DataLayout &DL = LI->getModule()->getDataLayout();
861 
862  Instruction *DepInst = DepInfo.getInst();
863  if (DepInfo.isClobber()) {
864  // If the dependence is to a store that writes to a superset of the bits
865  // read by the load, we can extract the bits we need for the load from the
866  // stored value.
867  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
868  // Can't forward from non-atomic to atomic without violating memory model.
869  if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
870  int Offset =
872  if (Offset != -1) {
873  Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
874  return true;
875  }
876  }
877  }
878 
879  // Check to see if we have something like this:
880  // load i32* P
881  // load i8* (P+1)
882  // if we have this, replace the later with an extraction from the former.
883  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
884  // If this is a clobber and L is the first instruction in its block, then
885  // we have the first instruction in the entry block.
886  // Can't forward from non-atomic to atomic without violating memory model.
887  if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
888  int Offset =
889  analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
890 
891  if (Offset != -1) {
892  Res = AvailableValue::getLoad(DepLI, Offset);
893  return true;
894  }
895  }
896  }
897 
898  // If the clobbering value is a memset/memcpy/memmove, see if we can
899  // forward a value on from it.
900  if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
901  if (Address && !LI->isAtomic()) {
903  DepMI, DL);
904  if (Offset != -1) {
905  Res = AvailableValue::getMI(DepMI, Offset);
906  return true;
907  }
908  }
909  }
910  // Nothing known about this clobber, have to be conservative
911  LLVM_DEBUG(
912  // fast print dep, using operator<< on instruction is too slow.
913  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
914  dbgs() << " is clobbered by " << *DepInst << '\n';);
915  if (ORE->allowExtraAnalysis(DEBUG_TYPE))
916  reportMayClobberedLoad(LI, DepInfo, DT, ORE);
917 
918  return false;
919  }
920  assert(DepInfo.isDef() && "follows from above");
921 
922  // Loading the allocation -> undef.
923  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
924  // Loading immediately after lifetime begin -> undef.
925  isLifetimeStart(DepInst)) {
927  return true;
928  }
929 
930  // Loading from calloc (which zero initializes memory) -> zero
931  if (isCallocLikeFn(DepInst, TLI)) {
933  return true;
934  }
935 
936  if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
937  // Reject loads and stores that are to the same address but are of
938  // different types if we have to. If the stored value is larger or equal to
939  // the loaded value, we can reuse it.
940  if (!canCoerceMustAliasedValueToLoad(S->getValueOperand(), LI->getType(),
941  DL))
942  return false;
943 
944  // Can't forward from non-atomic to atomic without violating memory model.
945  if (S->isAtomic() < LI->isAtomic())
946  return false;
947 
948  Res = AvailableValue::get(S->getValueOperand());
949  return true;
950  }
951 
952  if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
953  // If the types mismatch and we can't handle it, reject reuse of the load.
954  // If the stored value is larger or equal to the loaded value, we can reuse
955  // it.
956  if (!canCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
957  return false;
958 
959  // Can't forward from non-atomic to atomic without violating memory model.
960  if (LD->isAtomic() < LI->isAtomic())
961  return false;
962 
964  return true;
965  }
966 
967  // Unknown def - must be conservative
968  LLVM_DEBUG(
969  // fast print dep, using operator<< on instruction is too slow.
970  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
971  dbgs() << " has unknown def " << *DepInst << '\n';);
972  return false;
973 }
974 
975 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
976  AvailValInBlkVect &ValuesPerBlock,
977  UnavailBlkVect &UnavailableBlocks) {
978  // Filter out useless results (non-locals, etc). Keep track of the blocks
979  // where we have a value available in repl, also keep track of whether we see
980  // dependencies that produce an unknown value for the load (such as a call
981  // that could potentially clobber the load).
982  unsigned NumDeps = Deps.size();
983  for (unsigned i = 0, e = NumDeps; i != e; ++i) {
984  BasicBlock *DepBB = Deps[i].getBB();
985  MemDepResult DepInfo = Deps[i].getResult();
986 
987  if (DeadBlocks.count(DepBB)) {
988  // Dead dependent mem-op disguise as a load evaluating the same value
989  // as the load in question.
990  ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
991  continue;
992  }
993 
994  if (!DepInfo.isDef() && !DepInfo.isClobber()) {
995  UnavailableBlocks.push_back(DepBB);
996  continue;
997  }
998 
999  // The address being loaded in this non-local block may not be the same as
1000  // the pointer operand of the load if PHI translation occurs. Make sure
1001  // to consider the right address.
1002  Value *Address = Deps[i].getAddress();
1003 
1004  AvailableValue AV;
1005  if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1006  // subtlety: because we know this was a non-local dependency, we know
1007  // it's safe to materialize anywhere between the instruction within
1008  // DepInfo and the end of it's block.
1009  ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1010  std::move(AV)));
1011  } else {
1012  UnavailableBlocks.push_back(DepBB);
1013  }
1014  }
1015 
1016  assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1017  "post condition violation");
1018 }
1019 
1020 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1021  UnavailBlkVect &UnavailableBlocks) {
1022  // Okay, we have *some* definitions of the value. This means that the value
1023  // is available in some of our (transitive) predecessors. Lets think about
1024  // doing PRE of this load. This will involve inserting a new load into the
1025  // predecessor when it's not available. We could do this in general, but
1026  // prefer to not increase code size. As such, we only do this when we know
1027  // that we only have to insert *one* load (which means we're basically moving
1028  // the load, not inserting a new one).
1029 
1030  SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1031  UnavailableBlocks.end());
1032 
1033  // Let's find the first basic block with more than one predecessor. Walk
1034  // backwards through predecessors if needed.
1035  BasicBlock *LoadBB = LI->getParent();
1036  BasicBlock *TmpBB = LoadBB;
1037  bool IsSafeToSpeculativelyExecute = isSafeToSpeculativelyExecute(LI);
1038 
1039  // Check that there is no implicit control flow instructions above our load in
1040  // its block. If there is an instruction that doesn't always pass the
1041  // execution to the following instruction, then moving through it may become
1042  // invalid. For example:
1043  //
1044  // int arr[LEN];
1045  // int index = ???;
1046  // ...
1047  // guard(0 <= index && index < LEN);
1048  // use(arr[index]);
1049  //
1050  // It is illegal to move the array access to any point above the guard,
1051  // because if the index is out of bounds we should deoptimize rather than
1052  // access the array.
1053  // Check that there is no guard in this block above our instruction.
1054  if (!IsSafeToSpeculativelyExecute && ICF->isDominatedByICFIFromSameBlock(LI))
1055  return false;
1056  while (TmpBB->getSinglePredecessor()) {
1057  TmpBB = TmpBB->getSinglePredecessor();
1058  if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1059  return false;
1060  if (Blockers.count(TmpBB))
1061  return false;
1062 
1063  // If any of these blocks has more than one successor (i.e. if the edge we
1064  // just traversed was critical), then there are other paths through this
1065  // block along which the load may not be anticipated. Hoisting the load
1066  // above this block would be adding the load to execution paths along
1067  // which it was not previously executed.
1068  if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1069  return false;
1070 
1071  // Check that there is no implicit control flow in a block above.
1072  if (!IsSafeToSpeculativelyExecute && ICF->hasICF(TmpBB))
1073  return false;
1074  }
1075 
1076  assert(TmpBB);
1077  LoadBB = TmpBB;
1078 
1079  // Check to see how many predecessors have the loaded value fully
1080  // available.
1082  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1083  for (const AvailableValueInBlock &AV : ValuesPerBlock)
1084  FullyAvailableBlocks[AV.BB] = true;
1085  for (BasicBlock *UnavailableBB : UnavailableBlocks)
1086  FullyAvailableBlocks[UnavailableBB] = false;
1087 
1088  SmallVector<BasicBlock *, 4> CriticalEdgePred;
1089  for (BasicBlock *Pred : predecessors(LoadBB)) {
1090  // If any predecessor block is an EH pad that does not allow non-PHI
1091  // instructions before the terminator, we can't PRE the load.
1092  if (Pred->getTerminator()->isEHPad()) {
1093  LLVM_DEBUG(
1094  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1095  << Pred->getName() << "': " << *LI << '\n');
1096  return false;
1097  }
1098 
1099  if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1100  continue;
1101  }
1102 
1103  if (Pred->getTerminator()->getNumSuccessors() != 1) {
1104  if (isa<IndirectBrInst>(Pred->getTerminator())) {
1105  LLVM_DEBUG(
1106  dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1107  << Pred->getName() << "': " << *LI << '\n');
1108  return false;
1109  }
1110 
1111  // FIXME: Can we support the fallthrough edge?
1112  if (isa<CallBrInst>(Pred->getTerminator())) {
1113  LLVM_DEBUG(
1114  dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '"
1115  << Pred->getName() << "': " << *LI << '\n');
1116  return false;
1117  }
1118 
1119  if (LoadBB->isEHPad()) {
1120  LLVM_DEBUG(
1121  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1122  << Pred->getName() << "': " << *LI << '\n');
1123  return false;
1124  }
1125 
1126  CriticalEdgePred.push_back(Pred);
1127  } else {
1128  // Only add the predecessors that will not be split for now.
1129  PredLoads[Pred] = nullptr;
1130  }
1131  }
1132 
1133  // Decide whether PRE is profitable for this load.
1134  unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1135  assert(NumUnavailablePreds != 0 &&
1136  "Fully available value should already be eliminated!");
1137 
1138  // If this load is unavailable in multiple predecessors, reject it.
1139  // FIXME: If we could restructure the CFG, we could make a common pred with
1140  // all the preds that don't have an available LI and insert a new load into
1141  // that one block.
1142  if (NumUnavailablePreds != 1)
1143  return false;
1144 
1145  // Split critical edges, and update the unavailable predecessors accordingly.
1146  for (BasicBlock *OrigPred : CriticalEdgePred) {
1147  BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1148  assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1149  PredLoads[NewPred] = nullptr;
1150  LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1151  << LoadBB->getName() << '\n');
1152  }
1153 
1154  // Check if the load can safely be moved to all the unavailable predecessors.
1155  bool CanDoPRE = true;
1156  const DataLayout &DL = LI->getModule()->getDataLayout();
1158  for (auto &PredLoad : PredLoads) {
1159  BasicBlock *UnavailablePred = PredLoad.first;
1160 
1161  // Do PHI translation to get its value in the predecessor if necessary. The
1162  // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1163 
1164  // If all preds have a single successor, then we know it is safe to insert
1165  // the load on the pred (?!?), so we can insert code to materialize the
1166  // pointer if it is not available.
1167  PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1168  Value *LoadPtr = nullptr;
1169  LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1170  *DT, NewInsts);
1171 
1172  // If we couldn't find or insert a computation of this phi translated value,
1173  // we fail PRE.
1174  if (!LoadPtr) {
1175  LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1176  << *LI->getPointerOperand() << "\n");
1177  CanDoPRE = false;
1178  break;
1179  }
1180 
1181  PredLoad.second = LoadPtr;
1182  }
1183 
1184  if (!CanDoPRE) {
1185  while (!NewInsts.empty()) {
1186  Instruction *I = NewInsts.pop_back_val();
1187  markInstructionForDeletion(I);
1188  }
1189  // HINT: Don't revert the edge-splitting as following transformation may
1190  // also need to split these critical edges.
1191  return !CriticalEdgePred.empty();
1192  }
1193 
1194  // Okay, we can eliminate this load by inserting a reload in the predecessor
1195  // and using PHI construction to get the value in the other predecessors, do
1196  // it.
1197  LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1198  LLVM_DEBUG(if (!NewInsts.empty()) dbgs()
1199  << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back()
1200  << '\n');
1201 
1202  // Assign value numbers to the new instructions.
1203  for (Instruction *I : NewInsts) {
1204  // Instructions that have been inserted in predecessor(s) to materialize
1205  // the load address do not retain their original debug locations. Doing
1206  // so could lead to confusing (but correct) source attributions.
1207  if (const DebugLoc &DL = I->getDebugLoc())
1208  I->setDebugLoc(DebugLoc::get(0, 0, DL.getScope(), DL.getInlinedAt()));
1209 
1210  // FIXME: We really _ought_ to insert these value numbers into their
1211  // parent's availability map. However, in doing so, we risk getting into
1212  // ordering issues. If a block hasn't been processed yet, we would be
1213  // marking a value as AVAIL-IN, which isn't what we intend.
1214  VN.lookupOrAdd(I);
1215  }
1216 
1217  for (const auto &PredLoad : PredLoads) {
1218  BasicBlock *UnavailablePred = PredLoad.first;
1219  Value *LoadPtr = PredLoad.second;
1220 
1221  auto *NewLoad =
1222  new LoadInst(LI->getType(), LoadPtr, LI->getName() + ".pre",
1223  LI->isVolatile(), LI->getAlignment(), LI->getOrdering(),
1224  LI->getSyncScopeID(), UnavailablePred->getTerminator());
1225  NewLoad->setDebugLoc(LI->getDebugLoc());
1226 
1227  // Transfer the old load's AA tags to the new load.
1228  AAMDNodes Tags;
1229  LI->getAAMetadata(Tags);
1230  if (Tags)
1231  NewLoad->setAAMetadata(Tags);
1232 
1233  if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1234  NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1235  if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1236  NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1237  if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1238  NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1239 
1240  // We do not propagate the old load's debug location, because the new
1241  // load now lives in a different BB, and we want to avoid a jumpy line
1242  // table.
1243  // FIXME: How do we retain source locations without causing poor debugging
1244  // behavior?
1245 
1246  // Add the newly created load.
1247  ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1248  NewLoad));
1249  MD->invalidateCachedPointerInfo(LoadPtr);
1250  LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1251  }
1252 
1253  // Perform PHI construction.
1254  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1255  LI->replaceAllUsesWith(V);
1256  if (isa<PHINode>(V))
1257  V->takeName(LI);
1258  if (Instruction *I = dyn_cast<Instruction>(V))
1259  I->setDebugLoc(LI->getDebugLoc());
1260  if (V->getType()->isPtrOrPtrVectorTy())
1261  MD->invalidateCachedPointerInfo(V);
1262  markInstructionForDeletion(LI);
1263  ORE->emit([&]() {
1264  return OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1265  << "load eliminated by PRE";
1266  });
1267  ++NumPRELoad;
1268  return true;
1269 }
1270 
1273  using namespace ore;
1274 
1275  ORE->emit([&]() {
1276  return OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1277  << "load of type " << NV("Type", LI->getType()) << " eliminated"
1278  << setExtraArgs() << " in favor of "
1279  << NV("InfavorOfValue", AvailableValue);
1280  });
1281 }
1282 
1283 /// Attempt to eliminate a load whose dependencies are
1284 /// non-local by performing PHI construction.
1285 bool GVN::processNonLocalLoad(LoadInst *LI) {
1286  // non-local speculations are not allowed under asan.
1287  if (LI->getParent()->getParent()->hasFnAttribute(
1288  Attribute::SanitizeAddress) ||
1290  Attribute::SanitizeHWAddress))
1291  return false;
1292 
1293  // Step 1: Find the non-local dependencies of the load.
1294  LoadDepVect Deps;
1295  MD->getNonLocalPointerDependency(LI, Deps);
1296 
1297  // If we had to process more than one hundred blocks to find the
1298  // dependencies, this load isn't worth worrying about. Optimizing
1299  // it will be too expensive.
1300  unsigned NumDeps = Deps.size();
1301  if (NumDeps > MaxNumDeps)
1302  return false;
1303 
1304  // If we had a phi translation failure, we'll have a single entry which is a
1305  // clobber in the current block. Reject this early.
1306  if (NumDeps == 1 &&
1307  !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1308  LLVM_DEBUG(dbgs() << "GVN: non-local load "; LI->printAsOperand(dbgs());
1309  dbgs() << " has unknown dependencies\n";);
1310  return false;
1311  }
1312 
1313  // If this load follows a GEP, see if we can PRE the indices before analyzing.
1314  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1315  for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1316  OE = GEP->idx_end();
1317  OI != OE; ++OI)
1318  if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1319  performScalarPRE(I);
1320  }
1321 
1322  // Step 2: Analyze the availability of the load
1323  AvailValInBlkVect ValuesPerBlock;
1324  UnavailBlkVect UnavailableBlocks;
1325  AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1326 
1327  // If we have no predecessors that produce a known value for this load, exit
1328  // early.
1329  if (ValuesPerBlock.empty())
1330  return false;
1331 
1332  // Step 3: Eliminate fully redundancy.
1333  //
1334  // If all of the instructions we depend on produce a known value for this
1335  // load, then it is fully redundant and we can use PHI insertion to compute
1336  // its value. Insert PHIs and remove the fully redundant value now.
1337  if (UnavailableBlocks.empty()) {
1338  LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1339 
1340  // Perform PHI construction.
1341  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1342  LI->replaceAllUsesWith(V);
1343 
1344  if (isa<PHINode>(V))
1345  V->takeName(LI);
1346  if (Instruction *I = dyn_cast<Instruction>(V))
1347  // If instruction I has debug info, then we should not update it.
1348  // Also, if I has a null DebugLoc, then it is still potentially incorrect
1349  // to propagate LI's DebugLoc because LI may not post-dominate I.
1350  if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1351  I->setDebugLoc(LI->getDebugLoc());
1352  if (V->getType()->isPtrOrPtrVectorTy())
1353  MD->invalidateCachedPointerInfo(V);
1354  markInstructionForDeletion(LI);
1355  ++NumGVNLoad;
1356  reportLoadElim(LI, V, ORE);
1357  return true;
1358  }
1359 
1360  // Step 4: Eliminate partial redundancy.
1361  if (!EnablePRE || !EnableLoadPRE)
1362  return false;
1363 
1364  return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1365 }
1366 
1367 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1368  assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1369  "This function can only be called with llvm.assume intrinsic");
1370  Value *V = IntrinsicI->getArgOperand(0);
1371 
1372  if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1373  if (Cond->isZero()) {
1374  Type *Int8Ty = Type::getInt8Ty(V->getContext());
1375  // Insert a new store to null instruction before the load to indicate that
1376  // this code is not reachable. FIXME: We could insert unreachable
1377  // instruction directly because we can modify the CFG.
1378  new StoreInst(UndefValue::get(Int8Ty),
1380  IntrinsicI);
1381  }
1382  markInstructionForDeletion(IntrinsicI);
1383  return false;
1384  } else if (isa<Constant>(V)) {
1385  // If it's not false, and constant, it must evaluate to true. This means our
1386  // assume is assume(true), and thus, pointless, and we don't want to do
1387  // anything more here.
1388  return false;
1389  }
1390 
1391  Constant *True = ConstantInt::getTrue(V->getContext());
1392  bool Changed = false;
1393 
1394  for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1395  BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1396 
1397  // This property is only true in dominated successors, propagateEquality
1398  // will check dominance for us.
1399  Changed |= propagateEquality(V, True, Edge, false);
1400  }
1401 
1402  // We can replace assume value with true, which covers cases like this:
1403  // call void @llvm.assume(i1 %cmp)
1404  // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1405  ReplaceWithConstMap[V] = True;
1406 
1407  // If one of *cmp *eq operand is const, adding it to map will cover this:
1408  // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1409  // call void @llvm.assume(i1 %cmp)
1410  // ret float %0 ; will change it to ret float 3.000000e+00
1411  if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1412  if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1413  CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1414  (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1415  CmpI->getFastMathFlags().noNaNs())) {
1416  Value *CmpLHS = CmpI->getOperand(0);
1417  Value *CmpRHS = CmpI->getOperand(1);
1418  if (isa<Constant>(CmpLHS))
1419  std::swap(CmpLHS, CmpRHS);
1420  auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1421 
1422  // If only one operand is constant.
1423  if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1424  ReplaceWithConstMap[CmpLHS] = RHSConst;
1425  }
1426  }
1427  return Changed;
1428 }
1429 
1431  patchReplacementInstruction(I, Repl);
1432  I->replaceAllUsesWith(Repl);
1433 }
1434 
1435 /// Attempt to eliminate a load, first by eliminating it
1436 /// locally, and then attempting non-local elimination if that fails.
1437 bool GVN::processLoad(LoadInst *L) {
1438  if (!MD)
1439  return false;
1440 
1441  // This code hasn't been audited for ordered or volatile memory access
1442  if (!L->isUnordered())
1443  return false;
1444 
1445  if (L->use_empty()) {
1446  markInstructionForDeletion(L);
1447  return true;
1448  }
1449 
1450  // ... to a pointer that has been loaded from before...
1451  MemDepResult Dep = MD->getDependency(L);
1452 
1453  // If it is defined in another block, try harder.
1454  if (Dep.isNonLocal())
1455  return processNonLocalLoad(L);
1456 
1457  // Only handle the local case below
1458  if (!Dep.isDef() && !Dep.isClobber()) {
1459  // This might be a NonFuncLocal or an Unknown
1460  LLVM_DEBUG(
1461  // fast print dep, using operator<< on instruction is too slow.
1462  dbgs() << "GVN: load "; L->printAsOperand(dbgs());
1463  dbgs() << " has unknown dependence\n";);
1464  return false;
1465  }
1466 
1467  AvailableValue AV;
1468  if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1469  Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1470 
1471  // Replace the load!
1472  patchAndReplaceAllUsesWith(L, AvailableValue);
1473  markInstructionForDeletion(L);
1474  ++NumGVNLoad;
1475  reportLoadElim(L, AvailableValue, ORE);
1476  // Tell MDA to rexamine the reused pointer since we might have more
1477  // information after forwarding it.
1478  if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1479  MD->invalidateCachedPointerInfo(AvailableValue);
1480  return true;
1481  }
1482 
1483  return false;
1484 }
1485 
1486 /// Return a pair the first field showing the value number of \p Exp and the
1487 /// second field showing whether it is a value number newly created.
1488 std::pair<uint32_t, bool>
1489 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1490  uint32_t &e = expressionNumbering[Exp];
1491  bool CreateNewValNum = !e;
1492  if (CreateNewValNum) {
1493  Expressions.push_back(Exp);
1494  if (ExprIdx.size() < nextValueNumber + 1)
1495  ExprIdx.resize(nextValueNumber * 2);
1496  e = nextValueNumber;
1497  ExprIdx[nextValueNumber++] = nextExprNumber++;
1498  }
1499  return {e, CreateNewValNum};
1500 }
1501 
1502 /// Return whether all the values related with the same \p num are
1503 /// defined in \p BB.
1504 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1505  GVN &Gvn) {
1506  LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1507  while (Vals && Vals->BB == BB)
1508  Vals = Vals->Next;
1509  return !Vals;
1510 }
1511 
1512 /// Wrap phiTranslateImpl to provide caching functionality.
1514  const BasicBlock *PhiBlock, uint32_t Num,
1515  GVN &Gvn) {
1516  auto FindRes = PhiTranslateTable.find({Num, Pred});
1517  if (FindRes != PhiTranslateTable.end())
1518  return FindRes->second;
1519  uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1520  PhiTranslateTable.insert({{Num, Pred}, NewNum});
1521  return NewNum;
1522 }
1523 
1524 /// Translate value number \p Num using phis, so that it has the values of
1525 /// the phis in BB.
1526 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1527  const BasicBlock *PhiBlock,
1528  uint32_t Num, GVN &Gvn) {
1529  if (PHINode *PN = NumberingPhi[Num]) {
1530  for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1531  if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1532  if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1533  return TransVal;
1534  }
1535  return Num;
1536  }
1537 
1538  // If there is any value related with Num is defined in a BB other than
1539  // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1540  // a backedge. We can do an early exit in that case to save compile time.
1541  if (!areAllValsInBB(Num, PhiBlock, Gvn))
1542  return Num;
1543 
1544  if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1545  return Num;
1546  Expression Exp = Expressions[ExprIdx[Num]];
1547 
1548  for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1549  // For InsertValue and ExtractValue, some varargs are index numbers
1550  // instead of value numbers. Those index numbers should not be
1551  // translated.
1552  if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1553  (i > 0 && Exp.opcode == Instruction::ExtractValue))
1554  continue;
1555  Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1556  }
1557 
1558  if (Exp.commutative) {
1559  assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1560  if (Exp.varargs[0] > Exp.varargs[1]) {
1561  std::swap(Exp.varargs[0], Exp.varargs[1]);
1562  uint32_t Opcode = Exp.opcode >> 8;
1563  if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1564  Exp.opcode = (Opcode << 8) |
1566  static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1567  }
1568  }
1569 
1570  if (uint32_t NewNum = expressionNumbering[Exp])
1571  return NewNum;
1572  return Num;
1573 }
1574 
1575 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1576 /// again.
1578  const BasicBlock &CurrBlock) {
1579  for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1580  auto FindRes = PhiTranslateTable.find({Num, Pred});
1581  if (FindRes != PhiTranslateTable.end())
1582  PhiTranslateTable.erase(FindRes);
1583  }
1584 }
1585 
1586 // In order to find a leader for a given value number at a
1587 // specific basic block, we first obtain the list of all Values for that number,
1588 // and then scan the list to find one whose block dominates the block in
1589 // question. This is fast because dominator tree queries consist of only
1590 // a few comparisons of DFS numbers.
1591 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1592  LeaderTableEntry Vals = LeaderTable[num];
1593  if (!Vals.Val) return nullptr;
1594 
1595  Value *Val = nullptr;
1596  if (DT->dominates(Vals.BB, BB)) {
1597  Val = Vals.Val;
1598  if (isa<Constant>(Val)) return Val;
1599  }
1600 
1601  LeaderTableEntry* Next = Vals.Next;
1602  while (Next) {
1603  if (DT->dominates(Next->BB, BB)) {
1604  if (isa<Constant>(Next->Val)) return Next->Val;
1605  if (!Val) Val = Next->Val;
1606  }
1607 
1608  Next = Next->Next;
1609  }
1610 
1611  return Val;
1612 }
1613 
1614 /// There is an edge from 'Src' to 'Dst'. Return
1615 /// true if every path from the entry block to 'Dst' passes via this edge. In
1616 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1618  DominatorTree *DT) {
1619  // While in theory it is interesting to consider the case in which Dst has
1620  // more than one predecessor, because Dst might be part of a loop which is
1621  // only reachable from Src, in practice it is pointless since at the time
1622  // GVN runs all such loops have preheaders, which means that Dst will have
1623  // been changed to have only one predecessor, namely Src.
1624  const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1625  assert((!Pred || Pred == E.getStart()) &&
1626  "No edge between these basic blocks!");
1627  return Pred != nullptr;
1628 }
1629 
1630 void GVN::assignBlockRPONumber(Function &F) {
1631  BlockRPONumber.clear();
1632  uint32_t NextBlockNumber = 1;
1634  for (BasicBlock *BB : RPOT)
1635  BlockRPONumber[BB] = NextBlockNumber++;
1636  InvalidBlockRPONumbers = false;
1637 }
1638 
1639 // Tries to replace instruction with const, using information from
1640 // ReplaceWithConstMap.
1641 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1642  bool Changed = false;
1643  for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1644  Value *Operand = Instr->getOperand(OpNum);
1645  auto it = ReplaceWithConstMap.find(Operand);
1646  if (it != ReplaceWithConstMap.end()) {
1647  assert(!isa<Constant>(Operand) &&
1648  "Replacing constants with constants is invalid");
1649  LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "
1650  << *it->second << " in instruction " << *Instr << '\n');
1651  Instr->setOperand(OpNum, it->second);
1652  Changed = true;
1653  }
1654  }
1655  return Changed;
1656 }
1657 
1658 /// The given values are known to be equal in every block
1659 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1660 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1661 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1662 /// value starting from the end of Root.Start.
1663 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1664  bool DominatesByEdge) {
1666  Worklist.push_back(std::make_pair(LHS, RHS));
1667  bool Changed = false;
1668  // For speed, compute a conservative fast approximation to
1669  // DT->dominates(Root, Root.getEnd());
1670  const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1671 
1672  while (!Worklist.empty()) {
1673  std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1674  LHS = Item.first; RHS = Item.second;
1675 
1676  if (LHS == RHS)
1677  continue;
1678  assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1679 
1680  // Don't try to propagate equalities between constants.
1681  if (isa<Constant>(LHS) && isa<Constant>(RHS))
1682  continue;
1683 
1684  // Prefer a constant on the right-hand side, or an Argument if no constants.
1685  if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1686  std::swap(LHS, RHS);
1687  assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1688 
1689  // If there is no obvious reason to prefer the left-hand side over the
1690  // right-hand side, ensure the longest lived term is on the right-hand side,
1691  // so the shortest lived term will be replaced by the longest lived.
1692  // This tends to expose more simplifications.
1693  uint32_t LVN = VN.lookupOrAdd(LHS);
1694  if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1695  (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1696  // Move the 'oldest' value to the right-hand side, using the value number
1697  // as a proxy for age.
1698  uint32_t RVN = VN.lookupOrAdd(RHS);
1699  if (LVN < RVN) {
1700  std::swap(LHS, RHS);
1701  LVN = RVN;
1702  }
1703  }
1704 
1705  // If value numbering later sees that an instruction in the scope is equal
1706  // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1707  // the invariant that instructions only occur in the leader table for their
1708  // own value number (this is used by removeFromLeaderTable), do not do this
1709  // if RHS is an instruction (if an instruction in the scope is morphed into
1710  // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1711  // using the leader table is about compiling faster, not optimizing better).
1712  // The leader table only tracks basic blocks, not edges. Only add to if we
1713  // have the simple case where the edge dominates the end.
1714  if (RootDominatesEnd && !isa<Instruction>(RHS))
1715  addToLeaderTable(LVN, RHS, Root.getEnd());
1716 
1717  // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1718  // LHS always has at least one use that is not dominated by Root, this will
1719  // never do anything if LHS has only one use.
1720  if (!LHS->hasOneUse()) {
1721  unsigned NumReplacements =
1722  DominatesByEdge
1723  ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1724  : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1725 
1726  Changed |= NumReplacements > 0;
1727  NumGVNEqProp += NumReplacements;
1728  // Cached information for anything that uses LHS will be invalid.
1729  if (MD)
1730  MD->invalidateCachedPointerInfo(LHS);
1731  }
1732 
1733  // Now try to deduce additional equalities from this one. For example, if
1734  // the known equality was "(A != B)" == "false" then it follows that A and B
1735  // are equal in the scope. Only boolean equalities with an explicit true or
1736  // false RHS are currently supported.
1737  if (!RHS->getType()->isIntegerTy(1))
1738  // Not a boolean equality - bail out.
1739  continue;
1740  ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1741  if (!CI)
1742  // RHS neither 'true' nor 'false' - bail out.
1743  continue;
1744  // Whether RHS equals 'true'. Otherwise it equals 'false'.
1745  bool isKnownTrue = CI->isMinusOne();
1746  bool isKnownFalse = !isKnownTrue;
1747 
1748  // If "A && B" is known true then both A and B are known true. If "A || B"
1749  // is known false then both A and B are known false.
1750  Value *A, *B;
1751  if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1752  (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1753  Worklist.push_back(std::make_pair(A, RHS));
1754  Worklist.push_back(std::make_pair(B, RHS));
1755  continue;
1756  }
1757 
1758  // If we are propagating an equality like "(A == B)" == "true" then also
1759  // propagate the equality A == B. When propagating a comparison such as
1760  // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1761  if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1762  Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1763 
1764  // If "A == B" is known true, or "A != B" is known false, then replace
1765  // A with B everywhere in the scope.
1766  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1767  (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1768  Worklist.push_back(std::make_pair(Op0, Op1));
1769 
1770  // Handle the floating point versions of equality comparisons too.
1771  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1772  (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1773 
1774  // Floating point -0.0 and 0.0 compare equal, so we can only
1775  // propagate values if we know that we have a constant and that
1776  // its value is non-zero.
1777 
1778  // FIXME: We should do this optimization if 'no signed zeros' is
1779  // applicable via an instruction-level fast-math-flag or some other
1780  // indicator that relaxed FP semantics are being used.
1781 
1782  if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1783  Worklist.push_back(std::make_pair(Op0, Op1));
1784  }
1785 
1786  // If "A >= B" is known true, replace "A < B" with false everywhere.
1787  CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1788  Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1789  // Since we don't have the instruction "A < B" immediately to hand, work
1790  // out the value number that it would have and use that to find an
1791  // appropriate instruction (if any).
1792  uint32_t NextNum = VN.getNextUnusedValueNumber();
1793  uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1794  // If the number we were assigned was brand new then there is no point in
1795  // looking for an instruction realizing it: there cannot be one!
1796  if (Num < NextNum) {
1797  Value *NotCmp = findLeader(Root.getEnd(), Num);
1798  if (NotCmp && isa<Instruction>(NotCmp)) {
1799  unsigned NumReplacements =
1800  DominatesByEdge
1801  ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1802  : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1803  Root.getStart());
1804  Changed |= NumReplacements > 0;
1805  NumGVNEqProp += NumReplacements;
1806  // Cached information for anything that uses NotCmp will be invalid.
1807  if (MD)
1808  MD->invalidateCachedPointerInfo(NotCmp);
1809  }
1810  }
1811  // Ensure that any instruction in scope that gets the "A < B" value number
1812  // is replaced with false.
1813  // The leader table only tracks basic blocks, not edges. Only add to if we
1814  // have the simple case where the edge dominates the end.
1815  if (RootDominatesEnd)
1816  addToLeaderTable(Num, NotVal, Root.getEnd());
1817 
1818  continue;
1819  }
1820  }
1821 
1822  return Changed;
1823 }
1824 
1825 /// When calculating availability, handle an instruction
1826 /// by inserting it into the appropriate sets
1827 bool GVN::processInstruction(Instruction *I) {
1828  // Ignore dbg info intrinsics.
1829  if (isa<DbgInfoIntrinsic>(I))
1830  return false;
1831 
1832  // If the instruction can be easily simplified then do so now in preference
1833  // to value numbering it. Value numbering often exposes redundancies, for
1834  // example if it determines that %y is equal to %x then the instruction
1835  // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1836  const DataLayout &DL = I->getModule()->getDataLayout();
1837  if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1838  bool Changed = false;
1839  if (!I->use_empty()) {
1840  I->replaceAllUsesWith(V);
1841  Changed = true;
1842  }
1843  if (isInstructionTriviallyDead(I, TLI)) {
1844  markInstructionForDeletion(I);
1845  Changed = true;
1846  }
1847  if (Changed) {
1848  if (MD && V->getType()->isPtrOrPtrVectorTy())
1849  MD->invalidateCachedPointerInfo(V);
1850  ++NumGVNSimpl;
1851  return true;
1852  }
1853  }
1854 
1855  if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1856  if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1857  return processAssumeIntrinsic(IntrinsicI);
1858 
1859  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1860  if (processLoad(LI))
1861  return true;
1862 
1863  unsigned Num = VN.lookupOrAdd(LI);
1864  addToLeaderTable(Num, LI, LI->getParent());
1865  return false;
1866  }
1867 
1868  // For conditional branches, we can perform simple conditional propagation on
1869  // the condition value itself.
1870  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1871  if (!BI->isConditional())
1872  return false;
1873 
1874  if (isa<Constant>(BI->getCondition()))
1875  return processFoldableCondBr(BI);
1876 
1877  Value *BranchCond = BI->getCondition();
1878  BasicBlock *TrueSucc = BI->getSuccessor(0);
1879  BasicBlock *FalseSucc = BI->getSuccessor(1);
1880  // Avoid multiple edges early.
1881  if (TrueSucc == FalseSucc)
1882  return false;
1883 
1884  BasicBlock *Parent = BI->getParent();
1885  bool Changed = false;
1886 
1887  Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1888  BasicBlockEdge TrueE(Parent, TrueSucc);
1889  Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1890 
1891  Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1892  BasicBlockEdge FalseE(Parent, FalseSucc);
1893  Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1894 
1895  return Changed;
1896  }
1897 
1898  // For switches, propagate the case values into the case destinations.
1899  if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1900  Value *SwitchCond = SI->getCondition();
1901  BasicBlock *Parent = SI->getParent();
1902  bool Changed = false;
1903 
1904  // Remember how many outgoing edges there are to every successor.
1906  for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1907  ++SwitchEdges[SI->getSuccessor(i)];
1908 
1909  for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1910  i != e; ++i) {
1911  BasicBlock *Dst = i->getCaseSuccessor();
1912  // If there is only a single edge, propagate the case value into it.
1913  if (SwitchEdges.lookup(Dst) == 1) {
1914  BasicBlockEdge E(Parent, Dst);
1915  Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1916  }
1917  }
1918  return Changed;
1919  }
1920 
1921  // Instructions with void type don't return a value, so there's
1922  // no point in trying to find redundancies in them.
1923  if (I->getType()->isVoidTy())
1924  return false;
1925 
1926  uint32_t NextNum = VN.getNextUnusedValueNumber();
1927  unsigned Num = VN.lookupOrAdd(I);
1928 
1929  // Allocations are always uniquely numbered, so we can save time and memory
1930  // by fast failing them.
1931  if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) {
1932  addToLeaderTable(Num, I, I->getParent());
1933  return false;
1934  }
1935 
1936  // If the number we were assigned was a brand new VN, then we don't
1937  // need to do a lookup to see if the number already exists
1938  // somewhere in the domtree: it can't!
1939  if (Num >= NextNum) {
1940  addToLeaderTable(Num, I, I->getParent());
1941  return false;
1942  }
1943 
1944  // Perform fast-path value-number based elimination of values inherited from
1945  // dominators.
1946  Value *Repl = findLeader(I->getParent(), Num);
1947  if (!Repl) {
1948  // Failure, just remember this instance for future use.
1949  addToLeaderTable(Num, I, I->getParent());
1950  return false;
1951  } else if (Repl == I) {
1952  // If I was the result of a shortcut PRE, it might already be in the table
1953  // and the best replacement for itself. Nothing to do.
1954  return false;
1955  }
1956 
1957  // Remove it!
1958  patchAndReplaceAllUsesWith(I, Repl);
1959  if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1960  MD->invalidateCachedPointerInfo(Repl);
1961  markInstructionForDeletion(I);
1962  return true;
1963 }
1964 
1965 /// runOnFunction - This is the main transformation entry point for a function.
1966 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1967  const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1968  MemoryDependenceResults *RunMD, LoopInfo *LI,
1969  OptimizationRemarkEmitter *RunORE) {
1970  AC = &RunAC;
1971  DT = &RunDT;
1972  VN.setDomTree(DT);
1973  TLI = &RunTLI;
1974  VN.setAliasAnalysis(&RunAA);
1975  MD = RunMD;
1976  ImplicitControlFlowTracking ImplicitCFT(DT);
1977  ICF = &ImplicitCFT;
1978  VN.setMemDep(MD);
1979  ORE = RunORE;
1980  InvalidBlockRPONumbers = true;
1981 
1982  bool Changed = false;
1983  bool ShouldContinue = true;
1984 
1986  // Merge unconditional branches, allowing PRE to catch more
1987  // optimization opportunities.
1988  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1989  BasicBlock *BB = &*FI++;
1990 
1991  bool removedBlock = MergeBlockIntoPredecessor(BB, &DTU, LI, nullptr, MD);
1992  if (removedBlock)
1993  ++NumGVNBlocks;
1994 
1995  Changed |= removedBlock;
1996  }
1997 
1998  unsigned Iteration = 0;
1999  while (ShouldContinue) {
2000  LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2001  ShouldContinue = iterateOnFunction(F);
2002  Changed |= ShouldContinue;
2003  ++Iteration;
2004  }
2005 
2006  if (EnablePRE) {
2007  // Fabricate val-num for dead-code in order to suppress assertion in
2008  // performPRE().
2009  assignValNumForDeadCode();
2010  bool PREChanged = true;
2011  while (PREChanged) {
2012  PREChanged = performPRE(F);
2013  Changed |= PREChanged;
2014  }
2015  }
2016 
2017  // FIXME: Should perform GVN again after PRE does something. PRE can move
2018  // computations into blocks where they become fully redundant. Note that
2019  // we can't do this until PRE's critical edge splitting updates memdep.
2020  // Actually, when this happens, we should just fully integrate PRE into GVN.
2021 
2022  cleanupGlobalSets();
2023  // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2024  // iteration.
2025  DeadBlocks.clear();
2026 
2027  return Changed;
2028 }
2029 
2030 bool GVN::processBlock(BasicBlock *BB) {
2031  // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2032  // (and incrementing BI before processing an instruction).
2033  assert(InstrsToErase.empty() &&
2034  "We expect InstrsToErase to be empty across iterations");
2035  if (DeadBlocks.count(BB))
2036  return false;
2037 
2038  // Clearing map before every BB because it can be used only for single BB.
2039  ReplaceWithConstMap.clear();
2040  bool ChangedFunction = false;
2041 
2042  for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2043  BI != BE;) {
2044  if (!ReplaceWithConstMap.empty())
2045  ChangedFunction |= replaceOperandsWithConsts(&*BI);
2046  ChangedFunction |= processInstruction(&*BI);
2047 
2048  if (InstrsToErase.empty()) {
2049  ++BI;
2050  continue;
2051  }
2052 
2053  // If we need some instructions deleted, do it now.
2054  NumGVNInstr += InstrsToErase.size();
2055 
2056  // Avoid iterator invalidation.
2057  bool AtStart = BI == BB->begin();
2058  if (!AtStart)
2059  --BI;
2060 
2061  for (auto *I : InstrsToErase) {
2062  assert(I->getParent() == BB && "Removing instruction from wrong block?");
2063  LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n');
2064  salvageDebugInfo(*I);
2065  if (MD) MD->removeInstruction(I);
2066  LLVM_DEBUG(verifyRemoved(I));
2067  ICF->removeInstruction(I);
2068  I->eraseFromParent();
2069  }
2070  InstrsToErase.clear();
2071 
2072  if (AtStart)
2073  BI = BB->begin();
2074  else
2075  ++BI;
2076  }
2077 
2078  return ChangedFunction;
2079 }
2080 
2081 // Instantiate an expression in a predecessor that lacked it.
2082 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2083  BasicBlock *Curr, unsigned int ValNo) {
2084  // Because we are going top-down through the block, all value numbers
2085  // will be available in the predecessor by the time we need them. Any
2086  // that weren't originally present will have been instantiated earlier
2087  // in this loop.
2088  bool success = true;
2089  for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2090  Value *Op = Instr->getOperand(i);
2091  if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2092  continue;
2093  // This could be a newly inserted instruction, in which case, we won't
2094  // find a value number, and should give up before we hurt ourselves.
2095  // FIXME: Rewrite the infrastructure to let it easier to value number
2096  // and process newly inserted instructions.
2097  if (!VN.exists(Op)) {
2098  success = false;
2099  break;
2100  }
2101  uint32_t TValNo =
2102  VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2103  if (Value *V = findLeader(Pred, TValNo)) {
2104  Instr->setOperand(i, V);
2105  } else {
2106  success = false;
2107  break;
2108  }
2109  }
2110 
2111  // Fail out if we encounter an operand that is not available in
2112  // the PRE predecessor. This is typically because of loads which
2113  // are not value numbered precisely.
2114  if (!success)
2115  return false;
2116 
2117  Instr->insertBefore(Pred->getTerminator());
2118  Instr->setName(Instr->getName() + ".pre");
2119  Instr->setDebugLoc(Instr->getDebugLoc());
2120 
2121  unsigned Num = VN.lookupOrAdd(Instr);
2122  VN.add(Instr, Num);
2123 
2124  // Update the availability map to include the new instruction.
2125  addToLeaderTable(Num, Instr, Pred);
2126  return true;
2127 }
2128 
2129 bool GVN::performScalarPRE(Instruction *CurInst) {
2130  if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() ||
2131  isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2132  CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2133  isa<DbgInfoIntrinsic>(CurInst))
2134  return false;
2135 
2136  // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2137  // sinking the compare again, and it would force the code generator to
2138  // move the i1 from processor flags or predicate registers into a general
2139  // purpose register.
2140  if (isa<CmpInst>(CurInst))
2141  return false;
2142 
2143  // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from
2144  // sinking the addressing mode computation back to its uses. Extending the
2145  // GEP's live range increases the register pressure, and therefore it can
2146  // introduce unnecessary spills.
2147  //
2148  // This doesn't prevent Load PRE. PHI translation will make the GEP available
2149  // to the load by moving it to the predecessor block if necessary.
2150  if (isa<GetElementPtrInst>(CurInst))
2151  return false;
2152 
2153  // We don't currently value number ANY inline asm calls.
2154  if (auto *CallB = dyn_cast<CallBase>(CurInst))
2155  if (CallB->isInlineAsm())
2156  return false;
2157 
2158  uint32_t ValNo = VN.lookup(CurInst);
2159 
2160  // Look for the predecessors for PRE opportunities. We're
2161  // only trying to solve the basic diamond case, where
2162  // a value is computed in the successor and one predecessor,
2163  // but not the other. We also explicitly disallow cases
2164  // where the successor is its own predecessor, because they're
2165  // more complicated to get right.
2166  unsigned NumWith = 0;
2167  unsigned NumWithout = 0;
2168  BasicBlock *PREPred = nullptr;
2169  BasicBlock *CurrentBlock = CurInst->getParent();
2170 
2171  // Update the RPO numbers for this function.
2172  if (InvalidBlockRPONumbers)
2173  assignBlockRPONumber(*CurrentBlock->getParent());
2174 
2176  for (BasicBlock *P : predecessors(CurrentBlock)) {
2177  // We're not interested in PRE where blocks with predecessors that are
2178  // not reachable.
2179  if (!DT->isReachableFromEntry(P)) {
2180  NumWithout = 2;
2181  break;
2182  }
2183  // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2184  // when CurInst has operand defined in CurrentBlock (so it may be defined
2185  // by phi in the loop header).
2186  assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) &&
2187  "Invalid BlockRPONumber map.");
2188  if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2189  llvm::any_of(CurInst->operands(), [&](const Use &U) {
2190  if (auto *Inst = dyn_cast<Instruction>(U.get()))
2191  return Inst->getParent() == CurrentBlock;
2192  return false;
2193  })) {
2194  NumWithout = 2;
2195  break;
2196  }
2197 
2198  uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2199  Value *predV = findLeader(P, TValNo);
2200  if (!predV) {
2201  predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2202  PREPred = P;
2203  ++NumWithout;
2204  } else if (predV == CurInst) {
2205  /* CurInst dominates this predecessor. */
2206  NumWithout = 2;
2207  break;
2208  } else {
2209  predMap.push_back(std::make_pair(predV, P));
2210  ++NumWith;
2211  }
2212  }
2213 
2214  // Don't do PRE when it might increase code size, i.e. when
2215  // we would need to insert instructions in more than one pred.
2216  if (NumWithout > 1 || NumWith == 0)
2217  return false;
2218 
2219  // We may have a case where all predecessors have the instruction,
2220  // and we just need to insert a phi node. Otherwise, perform
2221  // insertion.
2222  Instruction *PREInstr = nullptr;
2223 
2224  if (NumWithout != 0) {
2225  if (!isSafeToSpeculativelyExecute(CurInst)) {
2226  // It is only valid to insert a new instruction if the current instruction
2227  // is always executed. An instruction with implicit control flow could
2228  // prevent us from doing it. If we cannot speculate the execution, then
2229  // PRE should be prohibited.
2230  if (ICF->isDominatedByICFIFromSameBlock(CurInst))
2231  return false;
2232  }
2233 
2234  // Don't do PRE across indirect branch.
2235  if (isa<IndirectBrInst>(PREPred->getTerminator()))
2236  return false;
2237 
2238  // Don't do PRE across callbr.
2239  // FIXME: Can we do this across the fallthrough edge?
2240  if (isa<CallBrInst>(PREPred->getTerminator()))
2241  return false;
2242 
2243  // We can't do PRE safely on a critical edge, so instead we schedule
2244  // the edge to be split and perform the PRE the next time we iterate
2245  // on the function.
2246  unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2247  if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2248  toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2249  return false;
2250  }
2251  // We need to insert somewhere, so let's give it a shot
2252  PREInstr = CurInst->clone();
2253  if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2254  // If we failed insertion, make sure we remove the instruction.
2255  LLVM_DEBUG(verifyRemoved(PREInstr));
2256  PREInstr->deleteValue();
2257  return false;
2258  }
2259  }
2260 
2261  // Either we should have filled in the PRE instruction, or we should
2262  // not have needed insertions.
2263  assert(PREInstr != nullptr || NumWithout == 0);
2264 
2265  ++NumGVNPRE;
2266 
2267  // Create a PHI to make the value available in this block.
2268  PHINode *Phi =
2269  PHINode::Create(CurInst->getType(), predMap.size(),
2270  CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2271  for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2272  if (Value *V = predMap[i].first) {
2273  // If we use an existing value in this phi, we have to patch the original
2274  // value because the phi will be used to replace a later value.
2275  patchReplacementInstruction(CurInst, V);
2276  Phi->addIncoming(V, predMap[i].second);
2277  } else
2278  Phi->addIncoming(PREInstr, PREPred);
2279  }
2280 
2281  VN.add(Phi, ValNo);
2282  // After creating a new PHI for ValNo, the phi translate result for ValNo will
2283  // be changed, so erase the related stale entries in phi translate cache.
2284  VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2285  addToLeaderTable(ValNo, Phi, CurrentBlock);
2286  Phi->setDebugLoc(CurInst->getDebugLoc());
2287  CurInst->replaceAllUsesWith(Phi);
2288  if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2289  MD->invalidateCachedPointerInfo(Phi);
2290  VN.erase(CurInst);
2291  removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2292 
2293  LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2294  if (MD)
2295  MD->removeInstruction(CurInst);
2296  LLVM_DEBUG(verifyRemoved(CurInst));
2297  // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes
2298  // some assertion failures.
2299  ICF->removeInstruction(CurInst);
2300  CurInst->eraseFromParent();
2301  ++NumGVNInstr;
2302 
2303  return true;
2304 }
2305 
2306 /// Perform a purely local form of PRE that looks for diamond
2307 /// control flow patterns and attempts to perform simple PRE at the join point.
2308 bool GVN::performPRE(Function &F) {
2309  bool Changed = false;
2310  for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2311  // Nothing to PRE in the entry block.
2312  if (CurrentBlock == &F.getEntryBlock())
2313  continue;
2314 
2315  // Don't perform PRE on an EH pad.
2316  if (CurrentBlock->isEHPad())
2317  continue;
2318 
2319  for (BasicBlock::iterator BI = CurrentBlock->begin(),
2320  BE = CurrentBlock->end();
2321  BI != BE;) {
2322  Instruction *CurInst = &*BI++;
2323  Changed |= performScalarPRE(CurInst);
2324  }
2325  }
2326 
2327  if (splitCriticalEdges())
2328  Changed = true;
2329 
2330  return Changed;
2331 }
2332 
2333 /// Split the critical edge connecting the given two blocks, and return
2334 /// the block inserted to the critical edge.
2335 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2336  BasicBlock *BB =
2338  if (MD)
2339  MD->invalidateCachedPredecessors();
2340  InvalidBlockRPONumbers = true;
2341  return BB;
2342 }
2343 
2344 /// Split critical edges found during the previous
2345 /// iteration that may enable further optimization.
2346 bool GVN::splitCriticalEdges() {
2347  if (toSplit.empty())
2348  return false;
2349  do {
2350  std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val();
2351  SplitCriticalEdge(Edge.first, Edge.second,
2353  } while (!toSplit.empty());
2354  if (MD) MD->invalidateCachedPredecessors();
2355  InvalidBlockRPONumbers = true;
2356  return true;
2357 }
2358 
2359 /// Executes one iteration of GVN
2360 bool GVN::iterateOnFunction(Function &F) {
2361  cleanupGlobalSets();
2362 
2363  // Top-down walk of the dominator tree
2364  bool Changed = false;
2365  // Needed for value numbering with phi construction to work.
2366  // RPOT walks the graph in its constructor and will not be invalidated during
2367  // processBlock.
2369 
2370  for (BasicBlock *BB : RPOT)
2371  Changed |= processBlock(BB);
2372 
2373  return Changed;
2374 }
2375 
2376 void GVN::cleanupGlobalSets() {
2377  VN.clear();
2378  LeaderTable.clear();
2379  BlockRPONumber.clear();
2380  TableAllocator.Reset();
2381  ICF->clear();
2382  InvalidBlockRPONumbers = true;
2383 }
2384 
2385 /// Verify that the specified instruction does not occur in our
2386 /// internal data structures.
2387 void GVN::verifyRemoved(const Instruction *Inst) const {
2388  VN.verifyRemoved(Inst);
2389 
2390  // Walk through the value number scope to make sure the instruction isn't
2391  // ferreted away in it.
2393  I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2394  const LeaderTableEntry *Node = &I->second;
2395  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2396 
2397  while (Node->Next) {
2398  Node = Node->Next;
2399  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2400  }
2401  }
2402 }
2403 
2404 /// BB is declared dead, which implied other blocks become dead as well. This
2405 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2406 /// live successors, update their phi nodes by replacing the operands
2407 /// corresponding to dead blocks with UndefVal.
2408 void GVN::addDeadBlock(BasicBlock *BB) {
2411 
2412  NewDead.push_back(BB);
2413  while (!NewDead.empty()) {
2414  BasicBlock *D = NewDead.pop_back_val();
2415  if (DeadBlocks.count(D))
2416  continue;
2417 
2418  // All blocks dominated by D are dead.
2420  DT->getDescendants(D, Dom);
2421  DeadBlocks.insert(Dom.begin(), Dom.end());
2422 
2423  // Figure out the dominance-frontier(D).
2424  for (BasicBlock *B : Dom) {
2425  for (BasicBlock *S : successors(B)) {
2426  if (DeadBlocks.count(S))
2427  continue;
2428 
2429  bool AllPredDead = true;
2430  for (BasicBlock *P : predecessors(S))
2431  if (!DeadBlocks.count(P)) {
2432  AllPredDead = false;
2433  break;
2434  }
2435 
2436  if (!AllPredDead) {
2437  // S could be proved dead later on. That is why we don't update phi
2438  // operands at this moment.
2439  DF.insert(S);
2440  } else {
2441  // While S is not dominated by D, it is dead by now. This could take
2442  // place if S already have a dead predecessor before D is declared
2443  // dead.
2444  NewDead.push_back(S);
2445  }
2446  }
2447  }
2448  }
2449 
2450  // For the dead blocks' live successors, update their phi nodes by replacing
2451  // the operands corresponding to dead blocks with UndefVal.
2453  I != E; I++) {
2454  BasicBlock *B = *I;
2455  if (DeadBlocks.count(B))
2456  continue;
2457 
2459  for (BasicBlock *P : Preds) {
2460  if (!DeadBlocks.count(P))
2461  continue;
2462 
2463  if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2464  if (BasicBlock *S = splitCriticalEdges(P, B))
2465  DeadBlocks.insert(P = S);
2466  }
2467 
2468  for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2469  PHINode &Phi = cast<PHINode>(*II);
2471  if (MD)
2472  MD->invalidateCachedPointerInfo(&Phi);
2473  }
2474  }
2475  }
2476 }
2477 
2478 // If the given branch is recognized as a foldable branch (i.e. conditional
2479 // branch with constant condition), it will perform following analyses and
2480 // transformation.
2481 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2482 // R be the target of the dead out-coming edge.
2483 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2484 // edge. The result of this step will be {X| X is dominated by R}
2485 // 2) Identify those blocks which haves at least one dead predecessor. The
2486 // result of this step will be dominance-frontier(R).
2487 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2488 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2489 //
2490 // Return true iff *NEW* dead code are found.
2491 bool GVN::processFoldableCondBr(BranchInst *BI) {
2492  if (!BI || BI->isUnconditional())
2493  return false;
2494 
2495  // If a branch has two identical successors, we cannot declare either dead.
2496  if (BI->getSuccessor(0) == BI->getSuccessor(1))
2497  return false;
2498 
2499  ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2500  if (!Cond)
2501  return false;
2502 
2503  BasicBlock *DeadRoot =
2504  Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2505  if (DeadBlocks.count(DeadRoot))
2506  return false;
2507 
2508  if (!DeadRoot->getSinglePredecessor())
2509  DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2510 
2511  addDeadBlock(DeadRoot);
2512  return true;
2513 }
2514 
2515 // performPRE() will trigger assert if it comes across an instruction without
2516 // associated val-num. As it normally has far more live instructions than dead
2517 // instructions, it makes more sense just to "fabricate" a val-number for the
2518 // dead code than checking if instruction involved is dead or not.
2519 void GVN::assignValNumForDeadCode() {
2520  for (BasicBlock *BB : DeadBlocks) {
2521  for (Instruction &Inst : *BB) {
2522  unsigned ValNum = VN.lookupOrAdd(&Inst);
2523  addToLeaderTable(ValNum, &Inst, BB);
2524  }
2525  }
2526 }
2527 
2529 public:
2530  static char ID; // Pass identification, replacement for typeid
2531 
2532  explicit GVNLegacyPass(bool NoMemDepAnalysis = !EnableMemDep)
2533  : FunctionPass(ID), NoMemDepAnalysis(NoMemDepAnalysis) {
2535  }
2536 
2537  bool runOnFunction(Function &F) override {
2538  if (skipFunction(F))
2539  return false;
2540 
2541  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2542 
2543  return Impl.runImpl(
2544  F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2545  getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2546  getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2547  getAnalysis<AAResultsWrapperPass>().getAAResults(),
2548  NoMemDepAnalysis ? nullptr
2549  : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2550  LIWP ? &LIWP->getLoopInfo() : nullptr,
2551  &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2552  }
2553 
2554  void getAnalysisUsage(AnalysisUsage &AU) const override {
2558  if (!NoMemDepAnalysis)
2561 
2566  }
2567 
2568 private:
2569  bool NoMemDepAnalysis;
2570  GVN Impl;
2571 };
2572 
2573 char GVNLegacyPass::ID = 0;
2574 
2575 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2583 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2584 
2585 // The public interface to this file...
2586 FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) {
2587  return new GVNLegacyPass(NoMemDepAnalysis);
2588 }
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:756
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:2586
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:1577
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:722
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:246
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
static Type * makeCmpResultType(Type *opnd_type)
Create a result type for fcmp/icmp.
Definition: InstrTypes.h:975
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:776
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
Represents an 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:568
iterator end()
Definition: Function.h:674
void setIncomingValueForBlock(const BasicBlock *BB, Value *V)
Set every incoming value(s) for block BB to V.
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:1624
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:1513
bool isTerminator() const
Definition: Instruction.h:128
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:748
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:323
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:2461
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:519
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:555
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1241
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:4382
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:2554
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:2392
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:2537
Option class for critical edge splitting.
void clear()
Remove all entries from the ValueTable.
Definition: GVN.cpp:576
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:650
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:283
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:1113
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:4470
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
Value * getRHS() const
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:672
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:218
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:656
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:432
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:318
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:762
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:726
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:572
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:1199
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:732
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:407
static void reportLoadElim(LoadInst *LI, Value *AvailableValue, OptimizationRemarkEmitter *ORE)
Definition: GVN.cpp:1271
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:4309
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:823
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:766
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:587
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:488
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:467
void verifyRemoved(const Value *) const
verifyRemoved - Verify that the value is removed from all internal data structures.
Definition: GVN.cpp:598
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:588
static bool isLifetimeStart(const Instruction *Inst)
Definition: GVN.cpp:815
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:549
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:1239
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:321
Instruction::BinaryOps getBinaryOp() const
Returns the binary operation underlying the intrinsic.
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:387
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
GVNLegacyPass(bool NoMemDepAnalysis=!EnableMemDep)
Definition: GVN.cpp:2532
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:795
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:653
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:735
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:353
LLVM Value Representation.
Definition: Value.h:72
static AvailableValueInBlock getUndef(BasicBlock *BB)
Definition: GVN.cpp:256
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:609
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:847
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:1430
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.
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.
Value * getLHS() const
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:1617
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