LLVM  10.0.0svn
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"
73 #include "llvm/Transforms/Utils.h"
78 #include <algorithm>
79 #include <cassert>
80 #include <cstdint>
81 #include <utility>
82 #include <vector>
83 
84 using namespace llvm;
85 using namespace llvm::gvn;
86 using namespace llvm::VNCoercion;
87 using namespace PatternMatch;
88 
89 #define DEBUG_TYPE "gvn"
90 
91 STATISTIC(NumGVNInstr, "Number of instructions deleted");
92 STATISTIC(NumGVNLoad, "Number of loads deleted");
93 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
94 STATISTIC(NumGVNBlocks, "Number of blocks merged");
95 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
96 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
97 STATISTIC(NumPRELoad, "Number of loads PRE'd");
98 
99 static cl::opt<bool> EnablePRE("enable-pre",
100  cl::init(true), cl::Hidden);
101 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
102 static cl::opt<bool> EnableMemDep("enable-gvn-memdep", cl::init(true));
103 
104 // Maximum allowed recursion depth.
105 static cl::opt<uint32_t>
106 MaxRecurseDepth("gvn-max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
107  cl::desc("Max recurse depth in GVN (default = 1000)"));
108 
110  "gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore,
111  cl::desc("Max number of dependences to attempt Load PRE (default = 100)"));
112 
116  bool commutative = false;
118 
119  Expression(uint32_t o = ~2U) : opcode(o) {}
120 
121  bool operator==(const Expression &other) const {
122  if (opcode != other.opcode)
123  return false;
124  if (opcode == ~0U || opcode == ~1U)
125  return true;
126  if (type != other.type)
127  return false;
128  if (varargs != other.varargs)
129  return false;
130  return true;
131  }
132 
134  return hash_combine(
135  Value.opcode, Value.type,
136  hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
137  }
138 };
139 
140 namespace llvm {
141 
142 template <> struct DenseMapInfo<GVN::Expression> {
143  static inline GVN::Expression getEmptyKey() { return ~0U; }
144  static inline GVN::Expression getTombstoneKey() { return ~1U; }
145 
146  static unsigned getHashValue(const GVN::Expression &e) {
147  using llvm::hash_value;
148 
149  return static_cast<unsigned>(hash_value(e));
150  }
151 
152  static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
153  return LHS == RHS;
154  }
155 };
156 
157 } // end namespace llvm
158 
159 /// Represents a particular available value that we know how to materialize.
160 /// Materialization of an AvailableValue never fails. An AvailableValue is
161 /// implicitly associated with a rematerialization point which is the
162 /// location of the instruction from which it was formed.
164  enum ValType {
165  SimpleVal, // A simple offsetted value that is accessed.
166  LoadVal, // A value produced by a load.
167  MemIntrin, // A memory intrinsic which is loaded from.
168  UndefVal // A UndefValue representing a value from dead block (which
169  // is not yet physically removed from the CFG).
170  };
171 
172  /// V - The value that is live out of the block.
174 
175  /// Offset - The byte offset in Val that is interesting for the load query.
176  unsigned Offset;
177 
178  static AvailableValue get(Value *V, unsigned Offset = 0) {
179  AvailableValue Res;
180  Res.Val.setPointer(V);
181  Res.Val.setInt(SimpleVal);
182  Res.Offset = Offset;
183  return Res;
184  }
185 
186  static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
187  AvailableValue Res;
188  Res.Val.setPointer(MI);
189  Res.Val.setInt(MemIntrin);
190  Res.Offset = Offset;
191  return Res;
192  }
193 
194  static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
195  AvailableValue Res;
196  Res.Val.setPointer(LI);
197  Res.Val.setInt(LoadVal);
198  Res.Offset = Offset;
199  return Res;
200  }
201 
203  AvailableValue Res;
204  Res.Val.setPointer(nullptr);
205  Res.Val.setInt(UndefVal);
206  Res.Offset = 0;
207  return Res;
208  }
209 
210  bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
211  bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
212  bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
213  bool isUndefValue() const { return Val.getInt() == UndefVal; }
214 
216  assert(isSimpleValue() && "Wrong accessor");
217  return Val.getPointer();
218  }
219 
221  assert(isCoercedLoadValue() && "Wrong accessor");
222  return cast<LoadInst>(Val.getPointer());
223  }
224 
226  assert(isMemIntrinValue() && "Wrong accessor");
227  return cast<MemIntrinsic>(Val.getPointer());
228  }
229 
230  /// Emit code at the specified insertion point to adjust the value defined
231  /// here to the specified type. This handles various coercion cases.
232  Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
233  GVN &gvn) const;
234 };
235 
236 /// Represents an AvailableValue which can be rematerialized at the end of
237 /// the associated BasicBlock.
239  /// BB - The basic block in question.
241 
242  /// AV - The actual available value
244 
247  Res.BB = BB;
248  Res.AV = std::move(AV);
249  return Res;
250  }
251 
253  unsigned Offset = 0) {
254  return get(BB, AvailableValue::get(V, Offset));
255  }
256 
258  return get(BB, AvailableValue::getUndef());
259  }
260 
261  /// Emit code at the end of this block to adjust the value defined here to
262  /// the specified type. This handles various coercion cases.
264  return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
265  }
266 };
267 
268 //===----------------------------------------------------------------------===//
269 // ValueTable Internal Functions
270 //===----------------------------------------------------------------------===//
271 
272 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
273  Expression e;
274  e.type = I->getType();
275  e.opcode = I->getOpcode();
276  for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
277  OI != OE; ++OI)
278  e.varargs.push_back(lookupOrAdd(*OI));
279  if (I->isCommutative()) {
280  // Ensure that commutative instructions that only differ by a permutation
281  // of their operands get the same value number by sorting the operand value
282  // numbers. Since all commutative instructions have two operands it is more
283  // efficient to sort by hand rather than using, say, std::sort.
284  assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
285  if (e.varargs[0] > e.varargs[1])
286  std::swap(e.varargs[0], e.varargs[1]);
287  e.commutative = true;
288  }
289 
290  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
291  // Sort the operand value numbers so x<y and y>x get the same value number.
292  CmpInst::Predicate Predicate = C->getPredicate();
293  if (e.varargs[0] > e.varargs[1]) {
294  std::swap(e.varargs[0], e.varargs[1]);
295  Predicate = CmpInst::getSwappedPredicate(Predicate);
296  }
297  e.opcode = (C->getOpcode() << 8) | Predicate;
298  e.commutative = true;
299  } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
300  for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
301  II != IE; ++II)
302  e.varargs.push_back(*II);
303  }
304 
305  return e;
306 }
307 
308 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
310  Value *LHS, Value *RHS) {
311  assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
312  "Not a comparison!");
313  Expression e;
314  e.type = CmpInst::makeCmpResultType(LHS->getType());
315  e.varargs.push_back(lookupOrAdd(LHS));
316  e.varargs.push_back(lookupOrAdd(RHS));
317 
318  // Sort the operand value numbers so x<y and y>x get the same value number.
319  if (e.varargs[0] > e.varargs[1]) {
320  std::swap(e.varargs[0], e.varargs[1]);
321  Predicate = CmpInst::getSwappedPredicate(Predicate);
322  }
323  e.opcode = (Opcode << 8) | Predicate;
324  e.commutative = true;
325  return e;
326 }
327 
328 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
329  assert(EI && "Not an ExtractValueInst?");
330  Expression e;
331  e.type = EI->getType();
332  e.opcode = 0;
333 
335  if (WO != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
336  // EI is an extract from one of our with.overflow intrinsics. Synthesize
337  // a semantically equivalent expression instead of an extract value
338  // expression.
339  e.opcode = WO->getBinaryOp();
340  e.varargs.push_back(lookupOrAdd(WO->getLHS()));
341  e.varargs.push_back(lookupOrAdd(WO->getRHS()));
342  return e;
343  }
344 
345  // Not a recognised intrinsic. Fall back to producing an extract value
346  // expression.
347  e.opcode = EI->getOpcode();
348  for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
349  OI != OE; ++OI)
350  e.varargs.push_back(lookupOrAdd(*OI));
351 
352  for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
353  II != IE; ++II)
354  e.varargs.push_back(*II);
355 
356  return e;
357 }
358 
359 //===----------------------------------------------------------------------===//
360 // ValueTable External Functions
361 //===----------------------------------------------------------------------===//
362 
363 GVN::ValueTable::ValueTable() = default;
364 GVN::ValueTable::ValueTable(const ValueTable &) = default;
365 GVN::ValueTable::ValueTable(ValueTable &&) = default;
366 GVN::ValueTable::~ValueTable() = default;
367 
368 /// add - Insert a value into the table with a specified value number.
370  valueNumbering.insert(std::make_pair(V, num));
371  if (PHINode *PN = dyn_cast<PHINode>(V))
372  NumberingPhi[num] = PN;
373 }
374 
375 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
376  if (AA->doesNotAccessMemory(C)) {
377  Expression exp = createExpr(C);
378  uint32_t e = assignExpNewValueNum(exp).first;
379  valueNumbering[C] = e;
380  return e;
381  } else if (MD && AA->onlyReadsMemory(C)) {
382  Expression exp = createExpr(C);
383  auto ValNum = assignExpNewValueNum(exp);
384  if (ValNum.second) {
385  valueNumbering[C] = ValNum.first;
386  return ValNum.first;
387  }
388 
389  MemDepResult local_dep = MD->getDependency(C);
390 
391  if (!local_dep.isDef() && !local_dep.isNonLocal()) {
392  valueNumbering[C] = nextValueNumber;
393  return nextValueNumber++;
394  }
395 
396  if (local_dep.isDef()) {
397  CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
398 
399  if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
400  valueNumbering[C] = nextValueNumber;
401  return nextValueNumber++;
402  }
403 
404  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
405  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
406  uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
407  if (c_vn != cd_vn) {
408  valueNumbering[C] = nextValueNumber;
409  return nextValueNumber++;
410  }
411  }
412 
413  uint32_t v = lookupOrAdd(local_cdep);
414  valueNumbering[C] = v;
415  return v;
416  }
417 
418  // Non-local case.
420  MD->getNonLocalCallDependency(C);
421  // FIXME: Move the checking logic to MemDep!
422  CallInst* cdep = nullptr;
423 
424  // Check to see if we have a single dominating call instruction that is
425  // identical to C.
426  for (unsigned i = 0, e = deps.size(); i != e; ++i) {
427  const NonLocalDepEntry *I = &deps[i];
428  if (I->getResult().isNonLocal())
429  continue;
430 
431  // We don't handle non-definitions. If we already have a call, reject
432  // instruction dependencies.
433  if (!I->getResult().isDef() || cdep != nullptr) {
434  cdep = nullptr;
435  break;
436  }
437 
438  CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
439  // FIXME: All duplicated with non-local case.
440  if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
441  cdep = NonLocalDepCall;
442  continue;
443  }
444 
445  cdep = nullptr;
446  break;
447  }
448 
449  if (!cdep) {
450  valueNumbering[C] = nextValueNumber;
451  return nextValueNumber++;
452  }
453 
454  if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
455  valueNumbering[C] = nextValueNumber;
456  return nextValueNumber++;
457  }
458  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
459  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
460  uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
461  if (c_vn != cd_vn) {
462  valueNumbering[C] = nextValueNumber;
463  return nextValueNumber++;
464  }
465  }
466 
467  uint32_t v = lookupOrAdd(cdep);
468  valueNumbering[C] = v;
469  return v;
470  } else {
471  valueNumbering[C] = nextValueNumber;
472  return nextValueNumber++;
473  }
474 }
475 
476 /// Returns true if a value number exists for the specified value.
477 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
478 
479 /// lookup_or_add - Returns the value number for the specified value, assigning
480 /// it a new number if it did not have one before.
482  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
483  if (VI != valueNumbering.end())
484  return VI->second;
485 
486  if (!isa<Instruction>(V)) {
487  valueNumbering[V] = nextValueNumber;
488  return nextValueNumber++;
489  }
490 
491  Instruction* I = cast<Instruction>(V);
492  Expression exp;
493  switch (I->getOpcode()) {
494  case Instruction::Call:
495  return lookupOrAddCall(cast<CallInst>(I));
496  case Instruction::FNeg:
497  case Instruction::Add:
498  case Instruction::FAdd:
499  case Instruction::Sub:
500  case Instruction::FSub:
501  case Instruction::Mul:
502  case Instruction::FMul:
503  case Instruction::UDiv:
504  case Instruction::SDiv:
505  case Instruction::FDiv:
506  case Instruction::URem:
507  case Instruction::SRem:
508  case Instruction::FRem:
509  case Instruction::Shl:
510  case Instruction::LShr:
511  case Instruction::AShr:
512  case Instruction::And:
513  case Instruction::Or:
514  case Instruction::Xor:
515  case Instruction::ICmp:
516  case Instruction::FCmp:
517  case Instruction::Trunc:
518  case Instruction::ZExt:
519  case Instruction::SExt:
520  case Instruction::FPToUI:
521  case Instruction::FPToSI:
522  case Instruction::UIToFP:
523  case Instruction::SIToFP:
524  case Instruction::FPTrunc:
525  case Instruction::FPExt:
526  case Instruction::PtrToInt:
527  case Instruction::IntToPtr:
528  case Instruction::AddrSpaceCast:
529  case Instruction::BitCast:
530  case Instruction::Select:
531  case Instruction::ExtractElement:
532  case Instruction::InsertElement:
533  case Instruction::ShuffleVector:
534  case Instruction::InsertValue:
535  case Instruction::GetElementPtr:
536  exp = createExpr(I);
537  break;
538  case Instruction::ExtractValue:
539  exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
540  break;
541  case Instruction::PHI:
542  valueNumbering[V] = nextValueNumber;
543  NumberingPhi[nextValueNumber] = cast<PHINode>(V);
544  return nextValueNumber++;
545  default:
546  valueNumbering[V] = nextValueNumber;
547  return nextValueNumber++;
548  }
549 
550  uint32_t e = assignExpNewValueNum(exp).first;
551  valueNumbering[V] = e;
552  return e;
553 }
554 
555 /// Returns the value number of the specified value. Fails if
556 /// the value has not yet been numbered.
558  DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
559  if (Verify) {
560  assert(VI != valueNumbering.end() && "Value not numbered?");
561  return VI->second;
562  }
563  return (VI != valueNumbering.end()) ? VI->second : 0;
564 }
565 
566 /// Returns the value number of the given comparison,
567 /// assigning it a new number if it did not have one before. Useful when
568 /// we deduced the result of a comparison, but don't immediately have an
569 /// instruction realizing that comparison to hand.
571  CmpInst::Predicate Predicate,
572  Value *LHS, Value *RHS) {
573  Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
574  return assignExpNewValueNum(exp).first;
575 }
576 
577 /// Remove all entries from the ValueTable.
579  valueNumbering.clear();
580  expressionNumbering.clear();
581  NumberingPhi.clear();
582  PhiTranslateTable.clear();
583  nextValueNumber = 1;
584  Expressions.clear();
585  ExprIdx.clear();
586  nextExprNumber = 0;
587 }
588 
589 /// Remove a value from the value numbering.
591  uint32_t Num = valueNumbering.lookup(V);
592  valueNumbering.erase(V);
593  // If V is PHINode, V <--> value number is an one-to-one mapping.
594  if (isa<PHINode>(V))
595  NumberingPhi.erase(Num);
596 }
597 
598 /// verifyRemoved - Verify that the value is removed from all internal data
599 /// structures.
600 void GVN::ValueTable::verifyRemoved(const Value *V) const {
602  I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
603  assert(I->first != V && "Inst still occurs in value numbering map!");
604  }
605 }
606 
607 //===----------------------------------------------------------------------===//
608 // GVN Pass
609 //===----------------------------------------------------------------------===//
610 
612  // FIXME: The order of evaluation of these 'getResult' calls is very
613  // significant! Re-ordering these variables will cause GVN when run alone to
614  // be less effective! We should fix memdep and basic-aa to not exhibit this
615  // behavior, but until then don't change the order here.
616  auto &AC = AM.getResult<AssumptionAnalysis>(F);
617  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
618  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
619  auto &AA = AM.getResult<AAManager>(F);
620  auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
621  auto *LI = AM.getCachedResult<LoopAnalysis>(F);
623  bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
624  if (!Changed)
625  return PreservedAnalyses::all();
628  PA.preserve<GlobalsAA>();
630  if (LI)
631  PA.preserve<LoopAnalysis>();
632  return PA;
633 }
634 
635 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
636 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
637  errs() << "{\n";
639  E = d.end(); I != E; ++I) {
640  errs() << I->first << "\n";
641  I->second->dump();
642  }
643  errs() << "}\n";
644 }
645 #endif
646 
647 /// Return true if we can prove that the value
648 /// we're analyzing is fully available in the specified block. As we go, keep
649 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
650 /// map is actually a tri-state map with the following values:
651 /// 0) we know the block *is not* fully available.
652 /// 1) we know the block *is* fully available.
653 /// 2) we do not know whether the block is fully available or not, but we are
654 /// currently speculating that it will be.
655 /// 3) we are speculating for this block and have used that to speculate for
656 /// other blocks.
658  DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
659  uint32_t RecurseDepth) {
660  if (RecurseDepth > MaxRecurseDepth)
661  return false;
662 
663  // Optimistically assume that the block is fully available and check to see
664  // if we already know about this block in one lookup.
665  std::pair<DenseMap<BasicBlock*, char>::iterator, bool> IV =
666  FullyAvailableBlocks.insert(std::make_pair(BB, 2));
667 
668  // If the entry already existed for this block, return the precomputed value.
669  if (!IV.second) {
670  // If this is a speculative "available" value, mark it as being used for
671  // speculation of other blocks.
672  if (IV.first->second == 2)
673  IV.first->second = 3;
674  return IV.first->second != 0;
675  }
676 
677  // Otherwise, see if it is fully available in all predecessors.
678  pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
679 
680  // If this block has no predecessors, it isn't live-in here.
681  if (PI == PE)
682  goto SpeculationFailure;
683 
684  for (; PI != PE; ++PI)
685  // If the value isn't fully available in one of our predecessors, then it
686  // isn't fully available in this block either. Undo our previous
687  // optimistic assumption and bail out.
688  if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
689  goto SpeculationFailure;
690 
691  return true;
692 
693 // If we get here, we found out that this is not, after
694 // all, a fully-available block. We have a problem if we speculated on this and
695 // used the speculation to mark other blocks as available.
696 SpeculationFailure:
697  char &BBVal = FullyAvailableBlocks[BB];
698 
699  // If we didn't speculate on this, just return with it set to false.
700  if (BBVal == 2) {
701  BBVal = 0;
702  return false;
703  }
704 
705  // If we did speculate on this value, we could have blocks set to 1 that are
706  // incorrect. Walk the (transitive) successors of this block and mark them as
707  // 0 if set to one.
708  SmallVector<BasicBlock*, 32> BBWorklist;
709  BBWorklist.push_back(BB);
710 
711  do {
712  BasicBlock *Entry = BBWorklist.pop_back_val();
713  // Note that this sets blocks to 0 (unavailable) if they happen to not
714  // already be in FullyAvailableBlocks. This is safe.
715  char &EntryVal = FullyAvailableBlocks[Entry];
716  if (EntryVal == 0) continue; // Already unavailable.
717 
718  // Mark as unavailable.
719  EntryVal = 0;
720 
721  BBWorklist.append(succ_begin(Entry), succ_end(Entry));
722  } while (!BBWorklist.empty());
723 
724  return false;
725 }
726 
727 /// Given a set of loads specified by ValuesPerBlock,
728 /// construct SSA form, allowing us to eliminate LI. This returns the value
729 /// that should be used at LI's definition site.
732  GVN &gvn) {
733  // Check for the fully redundant, dominating load case. In this case, we can
734  // just use the dominating value directly.
735  if (ValuesPerBlock.size() == 1 &&
736  gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
737  LI->getParent())) {
738  assert(!ValuesPerBlock[0].AV.isUndefValue() &&
739  "Dead BB dominate this block");
740  return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
741  }
742 
743  // Otherwise, we have to construct SSA form.
744  SmallVector<PHINode*, 8> NewPHIs;
745  SSAUpdater SSAUpdate(&NewPHIs);
746  SSAUpdate.Initialize(LI->getType(), LI->getName());
747 
748  for (const AvailableValueInBlock &AV : ValuesPerBlock) {
749  BasicBlock *BB = AV.BB;
750 
751  if (SSAUpdate.HasValueForBlock(BB))
752  continue;
753 
754  // If the value is the load that we will be eliminating, and the block it's
755  // available in is the block that the load is in, then don't add it as
756  // SSAUpdater will resolve the value to the relevant phi which may let it
757  // avoid phi construction entirely if there's actually only one value.
758  if (BB == LI->getParent() &&
759  ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == LI) ||
760  (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == LI)))
761  continue;
762 
763  SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
764  }
765 
766  // Perform PHI construction.
767  return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
768 }
769 
771  Instruction *InsertPt,
772  GVN &gvn) const {
773  Value *Res;
774  Type *LoadTy = LI->getType();
775  const DataLayout &DL = LI->getModule()->getDataLayout();
776  if (isSimpleValue()) {
777  Res = getSimpleValue();
778  if (Res->getType() != LoadTy) {
779  Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
780 
781  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset
782  << " " << *getSimpleValue() << '\n'
783  << *Res << '\n'
784  << "\n\n\n");
785  }
786  } else if (isCoercedLoadValue()) {
787  LoadInst *Load = getCoercedLoadValue();
788  if (Load->getType() == LoadTy && Offset == 0) {
789  Res = Load;
790  } else {
791  Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
792  // We would like to use gvn.markInstructionForDeletion here, but we can't
793  // because the load is already memoized into the leader map table that GVN
794  // tracks. It is potentially possible to remove the load from the table,
795  // but then there all of the operations based on it would need to be
796  // rehashed. Just leave the dead load around.
797  gvn.getMemDep().removeInstruction(Load);
798  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset
799  << " " << *getCoercedLoadValue() << '\n'
800  << *Res << '\n'
801  << "\n\n\n");
802  }
803  } else if (isMemIntrinValue()) {
804  Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
805  InsertPt, DL);
806  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
807  << " " << *getMemIntrinValue() << '\n'
808  << *Res << '\n'
809  << "\n\n\n");
810  } else {
811  assert(isUndefValue() && "Should be UndefVal");
812  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
813  return UndefValue::get(LoadTy);
814  }
815  assert(Res && "failed to materialize?");
816  return Res;
817 }
818 
819 static bool isLifetimeStart(const Instruction *Inst) {
820  if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
821  return II->getIntrinsicID() == Intrinsic::lifetime_start;
822  return false;
823 }
824 
825 /// Try to locate the three instruction involved in a missed
826 /// load-elimination case that is due to an intervening store.
828  DominatorTree *DT,
830  using namespace ore;
831 
832  User *OtherAccess = nullptr;
833 
834  OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
835  R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
836  << setExtraArgs();
837 
838  for (auto *U : LI->getPointerOperand()->users())
839  if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
840  DT->dominates(cast<Instruction>(U), LI)) {
841  // FIXME: for now give up if there are multiple memory accesses that
842  // dominate the load. We need further analysis to decide which one is
843  // that we're forwarding from.
844  if (OtherAccess)
845  OtherAccess = nullptr;
846  else
847  OtherAccess = U;
848  }
849 
850  if (OtherAccess)
851  R << " in favor of " << NV("OtherAccess", OtherAccess);
852 
853  R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
854 
855  ORE->emit(R);
856 }
857 
858 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
859  Value *Address, AvailableValue &Res) {
860  assert((DepInfo.isDef() || DepInfo.isClobber()) &&
861  "expected a local dependence");
862  assert(LI->isUnordered() && "rules below are incorrect for ordered access");
863 
864  const DataLayout &DL = LI->getModule()->getDataLayout();
865 
866  Instruction *DepInst = DepInfo.getInst();
867  if (DepInfo.isClobber()) {
868  // If the dependence is to a store that writes to a superset of the bits
869  // read by the load, we can extract the bits we need for the load from the
870  // stored value.
871  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
872  // Can't forward from non-atomic to atomic without violating memory model.
873  if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
874  int Offset =
876  if (Offset != -1) {
877  Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
878  return true;
879  }
880  }
881  }
882 
883  // Check to see if we have something like this:
884  // load i32* P
885  // load i8* (P+1)
886  // if we have this, replace the later with an extraction from the former.
887  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
888  // If this is a clobber and L is the first instruction in its block, then
889  // we have the first instruction in the entry block.
890  // Can't forward from non-atomic to atomic without violating memory model.
891  if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
892  int Offset =
893  analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
894 
895  if (Offset != -1) {
896  Res = AvailableValue::getLoad(DepLI, Offset);
897  return true;
898  }
899  }
900  }
901 
902  // If the clobbering value is a memset/memcpy/memmove, see if we can
903  // forward a value on from it.
904  if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
905  if (Address && !LI->isAtomic()) {
907  DepMI, DL);
908  if (Offset != -1) {
909  Res = AvailableValue::getMI(DepMI, Offset);
910  return true;
911  }
912  }
913  }
914  // Nothing known about this clobber, have to be conservative
915  LLVM_DEBUG(
916  // fast print dep, using operator<< on instruction is too slow.
917  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
918  dbgs() << " is clobbered by " << *DepInst << '\n';);
919  if (ORE->allowExtraAnalysis(DEBUG_TYPE))
920  reportMayClobberedLoad(LI, DepInfo, DT, ORE);
921 
922  return false;
923  }
924  assert(DepInfo.isDef() && "follows from above");
925 
926  // Loading the allocation -> undef.
927  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
928  // Loading immediately after lifetime begin -> undef.
929  isLifetimeStart(DepInst)) {
931  return true;
932  }
933 
934  // Loading from calloc (which zero initializes memory) -> zero
935  if (isCallocLikeFn(DepInst, TLI)) {
937  return true;
938  }
939 
940  if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
941  // Reject loads and stores that are to the same address but are of
942  // different types if we have to. If the stored value is larger or equal to
943  // the loaded value, we can reuse it.
944  if (!canCoerceMustAliasedValueToLoad(S->getValueOperand(), LI->getType(),
945  DL))
946  return false;
947 
948  // Can't forward from non-atomic to atomic without violating memory model.
949  if (S->isAtomic() < LI->isAtomic())
950  return false;
951 
952  Res = AvailableValue::get(S->getValueOperand());
953  return true;
954  }
955 
956  if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
957  // If the types mismatch and we can't handle it, reject reuse of the load.
958  // If the stored value is larger or equal to the loaded value, we can reuse
959  // it.
960  if (!canCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
961  return false;
962 
963  // Can't forward from non-atomic to atomic without violating memory model.
964  if (LD->isAtomic() < LI->isAtomic())
965  return false;
966 
968  return true;
969  }
970 
971  // Unknown def - must be conservative
972  LLVM_DEBUG(
973  // fast print dep, using operator<< on instruction is too slow.
974  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
975  dbgs() << " has unknown def " << *DepInst << '\n';);
976  return false;
977 }
978 
979 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
980  AvailValInBlkVect &ValuesPerBlock,
981  UnavailBlkVect &UnavailableBlocks) {
982  // Filter out useless results (non-locals, etc). Keep track of the blocks
983  // where we have a value available in repl, also keep track of whether we see
984  // dependencies that produce an unknown value for the load (such as a call
985  // that could potentially clobber the load).
986  unsigned NumDeps = Deps.size();
987  for (unsigned i = 0, e = NumDeps; i != e; ++i) {
988  BasicBlock *DepBB = Deps[i].getBB();
989  MemDepResult DepInfo = Deps[i].getResult();
990 
991  if (DeadBlocks.count(DepBB)) {
992  // Dead dependent mem-op disguise as a load evaluating the same value
993  // as the load in question.
994  ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
995  continue;
996  }
997 
998  if (!DepInfo.isDef() && !DepInfo.isClobber()) {
999  UnavailableBlocks.push_back(DepBB);
1000  continue;
1001  }
1002 
1003  // The address being loaded in this non-local block may not be the same as
1004  // the pointer operand of the load if PHI translation occurs. Make sure
1005  // to consider the right address.
1006  Value *Address = Deps[i].getAddress();
1007 
1008  AvailableValue AV;
1009  if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1010  // subtlety: because we know this was a non-local dependency, we know
1011  // it's safe to materialize anywhere between the instruction within
1012  // DepInfo and the end of it's block.
1013  ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1014  std::move(AV)));
1015  } else {
1016  UnavailableBlocks.push_back(DepBB);
1017  }
1018  }
1019 
1020  assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1021  "post condition violation");
1022 }
1023 
1024 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1025  UnavailBlkVect &UnavailableBlocks) {
1026  // Okay, we have *some* definitions of the value. This means that the value
1027  // is available in some of our (transitive) predecessors. Lets think about
1028  // doing PRE of this load. This will involve inserting a new load into the
1029  // predecessor when it's not available. We could do this in general, but
1030  // prefer to not increase code size. As such, we only do this when we know
1031  // that we only have to insert *one* load (which means we're basically moving
1032  // the load, not inserting a new one).
1033 
1034  SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1035  UnavailableBlocks.end());
1036 
1037  // Let's find the first basic block with more than one predecessor. Walk
1038  // backwards through predecessors if needed.
1039  BasicBlock *LoadBB = LI->getParent();
1040  BasicBlock *TmpBB = LoadBB;
1041  bool IsSafeToSpeculativelyExecute = isSafeToSpeculativelyExecute(LI);
1042 
1043  // Check that there is no implicit control flow instructions above our load in
1044  // its block. If there is an instruction that doesn't always pass the
1045  // execution to the following instruction, then moving through it may become
1046  // invalid. For example:
1047  //
1048  // int arr[LEN];
1049  // int index = ???;
1050  // ...
1051  // guard(0 <= index && index < LEN);
1052  // use(arr[index]);
1053  //
1054  // It is illegal to move the array access to any point above the guard,
1055  // because if the index is out of bounds we should deoptimize rather than
1056  // access the array.
1057  // Check that there is no guard in this block above our instruction.
1058  if (!IsSafeToSpeculativelyExecute && ICF->isDominatedByICFIFromSameBlock(LI))
1059  return false;
1060  while (TmpBB->getSinglePredecessor()) {
1061  TmpBB = TmpBB->getSinglePredecessor();
1062  if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1063  return false;
1064  if (Blockers.count(TmpBB))
1065  return false;
1066 
1067  // If any of these blocks has more than one successor (i.e. if the edge we
1068  // just traversed was critical), then there are other paths through this
1069  // block along which the load may not be anticipated. Hoisting the load
1070  // above this block would be adding the load to execution paths along
1071  // which it was not previously executed.
1072  if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1073  return false;
1074 
1075  // Check that there is no implicit control flow in a block above.
1076  if (!IsSafeToSpeculativelyExecute && ICF->hasICF(TmpBB))
1077  return false;
1078  }
1079 
1080  assert(TmpBB);
1081  LoadBB = TmpBB;
1082 
1083  // Check to see how many predecessors have the loaded value fully
1084  // available.
1086  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1087  for (const AvailableValueInBlock &AV : ValuesPerBlock)
1088  FullyAvailableBlocks[AV.BB] = true;
1089  for (BasicBlock *UnavailableBB : UnavailableBlocks)
1090  FullyAvailableBlocks[UnavailableBB] = false;
1091 
1092  SmallVector<BasicBlock *, 4> CriticalEdgePred;
1093  for (BasicBlock *Pred : predecessors(LoadBB)) {
1094  // If any predecessor block is an EH pad that does not allow non-PHI
1095  // instructions before the terminator, we can't PRE the load.
1096  if (Pred->getTerminator()->isEHPad()) {
1097  LLVM_DEBUG(
1098  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1099  << Pred->getName() << "': " << *LI << '\n');
1100  return false;
1101  }
1102 
1103  if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1104  continue;
1105  }
1106 
1107  if (Pred->getTerminator()->getNumSuccessors() != 1) {
1108  if (isa<IndirectBrInst>(Pred->getTerminator())) {
1109  LLVM_DEBUG(
1110  dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1111  << Pred->getName() << "': " << *LI << '\n');
1112  return false;
1113  }
1114 
1115  // FIXME: Can we support the fallthrough edge?
1116  if (isa<CallBrInst>(Pred->getTerminator())) {
1117  LLVM_DEBUG(
1118  dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '"
1119  << Pred->getName() << "': " << *LI << '\n');
1120  return false;
1121  }
1122 
1123  if (LoadBB->isEHPad()) {
1124  LLVM_DEBUG(
1125  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1126  << Pred->getName() << "': " << *LI << '\n');
1127  return false;
1128  }
1129 
1130  CriticalEdgePred.push_back(Pred);
1131  } else {
1132  // Only add the predecessors that will not be split for now.
1133  PredLoads[Pred] = nullptr;
1134  }
1135  }
1136 
1137  // Decide whether PRE is profitable for this load.
1138  unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1139  assert(NumUnavailablePreds != 0 &&
1140  "Fully available value should already be eliminated!");
1141 
1142  // If this load is unavailable in multiple predecessors, reject it.
1143  // FIXME: If we could restructure the CFG, we could make a common pred with
1144  // all the preds that don't have an available LI and insert a new load into
1145  // that one block.
1146  if (NumUnavailablePreds != 1)
1147  return false;
1148 
1149  // Split critical edges, and update the unavailable predecessors accordingly.
1150  for (BasicBlock *OrigPred : CriticalEdgePred) {
1151  BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1152  assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1153  PredLoads[NewPred] = nullptr;
1154  LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1155  << LoadBB->getName() << '\n');
1156  }
1157 
1158  // Check if the load can safely be moved to all the unavailable predecessors.
1159  bool CanDoPRE = true;
1160  const DataLayout &DL = LI->getModule()->getDataLayout();
1162  for (auto &PredLoad : PredLoads) {
1163  BasicBlock *UnavailablePred = PredLoad.first;
1164 
1165  // Do PHI translation to get its value in the predecessor if necessary. The
1166  // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1167  // We do the translation for each edge we skipped by going from LI's block
1168  // to LoadBB, otherwise we might miss pieces needing translation.
1169 
1170  // If all preds have a single successor, then we know it is safe to insert
1171  // the load on the pred (?!?), so we can insert code to materialize the
1172  // pointer if it is not available.
1173  Value *LoadPtr = LI->getPointerOperand();
1174  BasicBlock *Cur = LI->getParent();
1175  while (Cur != LoadBB) {
1176  PHITransAddr Address(LoadPtr, DL, AC);
1177  LoadPtr = Address.PHITranslateWithInsertion(
1178  Cur, Cur->getSinglePredecessor(), *DT, NewInsts);
1179  if (!LoadPtr) {
1180  CanDoPRE = false;
1181  break;
1182  }
1183  Cur = Cur->getSinglePredecessor();
1184  }
1185 
1186  if (LoadPtr) {
1187  PHITransAddr Address(LoadPtr, DL, AC);
1188  LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, *DT,
1189  NewInsts);
1190  }
1191  // If we couldn't find or insert a computation of this phi translated value,
1192  // we fail PRE.
1193  if (!LoadPtr) {
1194  LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1195  << *LI->getPointerOperand() << "\n");
1196  CanDoPRE = false;
1197  break;
1198  }
1199 
1200  PredLoad.second = LoadPtr;
1201  }
1202 
1203  if (!CanDoPRE) {
1204  while (!NewInsts.empty()) {
1205  // Erase instructions generated by the failed PHI translation before
1206  // trying to number them. PHI translation might insert instructions
1207  // in basic blocks other than the current one, and we delete them
1208  // directly, as markInstructionForDeletion only allows removing from the
1209  // current basic block.
1210  NewInsts.pop_back_val()->eraseFromParent();
1211  }
1212  // HINT: Don't revert the edge-splitting as following transformation may
1213  // also need to split these critical edges.
1214  return !CriticalEdgePred.empty();
1215  }
1216 
1217  // Okay, we can eliminate this load by inserting a reload in the predecessor
1218  // and using PHI construction to get the value in the other predecessors, do
1219  // it.
1220  LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1221  LLVM_DEBUG(if (!NewInsts.empty()) dbgs()
1222  << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back()
1223  << '\n');
1224 
1225  // Assign value numbers to the new instructions.
1226  for (Instruction *I : NewInsts) {
1227  // Instructions that have been inserted in predecessor(s) to materialize
1228  // the load address do not retain their original debug locations. Doing
1229  // so could lead to confusing (but correct) source attributions.
1230  if (const DebugLoc &DL = I->getDebugLoc())
1231  I->setDebugLoc(DebugLoc::get(0, 0, DL.getScope(), DL.getInlinedAt()));
1232 
1233  // FIXME: We really _ought_ to insert these value numbers into their
1234  // parent's availability map. However, in doing so, we risk getting into
1235  // ordering issues. If a block hasn't been processed yet, we would be
1236  // marking a value as AVAIL-IN, which isn't what we intend.
1237  VN.lookupOrAdd(I);
1238  }
1239 
1240  for (const auto &PredLoad : PredLoads) {
1241  BasicBlock *UnavailablePred = PredLoad.first;
1242  Value *LoadPtr = PredLoad.second;
1243 
1244  auto *NewLoad =
1245  new LoadInst(LI->getType(), LoadPtr, LI->getName() + ".pre",
1246  LI->isVolatile(), LI->getAlignment(), LI->getOrdering(),
1247  LI->getSyncScopeID(), UnavailablePred->getTerminator());
1248  NewLoad->setDebugLoc(LI->getDebugLoc());
1249 
1250  // Transfer the old load's AA tags to the new load.
1251  AAMDNodes Tags;
1252  LI->getAAMetadata(Tags);
1253  if (Tags)
1254  NewLoad->setAAMetadata(Tags);
1255 
1256  if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1257  NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1258  if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1259  NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1260  if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1261  NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1262 
1263  // We do not propagate the old load's debug location, because the new
1264  // load now lives in a different BB, and we want to avoid a jumpy line
1265  // table.
1266  // FIXME: How do we retain source locations without causing poor debugging
1267  // behavior?
1268 
1269  // Add the newly created load.
1270  ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1271  NewLoad));
1272  MD->invalidateCachedPointerInfo(LoadPtr);
1273  LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1274  }
1275 
1276  // Perform PHI construction.
1277  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1278  LI->replaceAllUsesWith(V);
1279  if (isa<PHINode>(V))
1280  V->takeName(LI);
1281  if (Instruction *I = dyn_cast<Instruction>(V))
1282  I->setDebugLoc(LI->getDebugLoc());
1283  if (V->getType()->isPtrOrPtrVectorTy())
1284  MD->invalidateCachedPointerInfo(V);
1285  markInstructionForDeletion(LI);
1286  ORE->emit([&]() {
1287  return OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1288  << "load eliminated by PRE";
1289  });
1290  ++NumPRELoad;
1291  return true;
1292 }
1293 
1296  using namespace ore;
1297 
1298  ORE->emit([&]() {
1299  return OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1300  << "load of type " << NV("Type", LI->getType()) << " eliminated"
1301  << setExtraArgs() << " in favor of "
1302  << NV("InfavorOfValue", AvailableValue);
1303  });
1304 }
1305 
1306 /// Attempt to eliminate a load whose dependencies are
1307 /// non-local by performing PHI construction.
1308 bool GVN::processNonLocalLoad(LoadInst *LI) {
1309  // non-local speculations are not allowed under asan.
1310  if (LI->getParent()->getParent()->hasFnAttribute(
1311  Attribute::SanitizeAddress) ||
1313  Attribute::SanitizeHWAddress))
1314  return false;
1315 
1316  // Step 1: Find the non-local dependencies of the load.
1317  LoadDepVect Deps;
1318  MD->getNonLocalPointerDependency(LI, Deps);
1319 
1320  // If we had to process more than one hundred blocks to find the
1321  // dependencies, this load isn't worth worrying about. Optimizing
1322  // it will be too expensive.
1323  unsigned NumDeps = Deps.size();
1324  if (NumDeps > MaxNumDeps)
1325  return false;
1326 
1327  // If we had a phi translation failure, we'll have a single entry which is a
1328  // clobber in the current block. Reject this early.
1329  if (NumDeps == 1 &&
1330  !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1331  LLVM_DEBUG(dbgs() << "GVN: non-local load "; LI->printAsOperand(dbgs());
1332  dbgs() << " has unknown dependencies\n";);
1333  return false;
1334  }
1335 
1336  // If this load follows a GEP, see if we can PRE the indices before analyzing.
1337  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1338  for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1339  OE = GEP->idx_end();
1340  OI != OE; ++OI)
1341  if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1342  performScalarPRE(I);
1343  }
1344 
1345  // Step 2: Analyze the availability of the load
1346  AvailValInBlkVect ValuesPerBlock;
1347  UnavailBlkVect UnavailableBlocks;
1348  AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1349 
1350  // If we have no predecessors that produce a known value for this load, exit
1351  // early.
1352  if (ValuesPerBlock.empty())
1353  return false;
1354 
1355  // Step 3: Eliminate fully redundancy.
1356  //
1357  // If all of the instructions we depend on produce a known value for this
1358  // load, then it is fully redundant and we can use PHI insertion to compute
1359  // its value. Insert PHIs and remove the fully redundant value now.
1360  if (UnavailableBlocks.empty()) {
1361  LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1362 
1363  // Perform PHI construction.
1364  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1365  LI->replaceAllUsesWith(V);
1366 
1367  if (isa<PHINode>(V))
1368  V->takeName(LI);
1369  if (Instruction *I = dyn_cast<Instruction>(V))
1370  // If instruction I has debug info, then we should not update it.
1371  // Also, if I has a null DebugLoc, then it is still potentially incorrect
1372  // to propagate LI's DebugLoc because LI may not post-dominate I.
1373  if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1374  I->setDebugLoc(LI->getDebugLoc());
1375  if (V->getType()->isPtrOrPtrVectorTy())
1376  MD->invalidateCachedPointerInfo(V);
1377  markInstructionForDeletion(LI);
1378  ++NumGVNLoad;
1379  reportLoadElim(LI, V, ORE);
1380  return true;
1381  }
1382 
1383  // Step 4: Eliminate partial redundancy.
1384  if (!EnablePRE || !EnableLoadPRE)
1385  return false;
1386 
1387  return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1388 }
1389 
1390 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1391  assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1392  "This function can only be called with llvm.assume intrinsic");
1393  Value *V = IntrinsicI->getArgOperand(0);
1394 
1395  if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1396  if (Cond->isZero()) {
1397  Type *Int8Ty = Type::getInt8Ty(V->getContext());
1398  // Insert a new store to null instruction before the load to indicate that
1399  // this code is not reachable. FIXME: We could insert unreachable
1400  // instruction directly because we can modify the CFG.
1401  new StoreInst(UndefValue::get(Int8Ty),
1403  IntrinsicI);
1404  }
1405  markInstructionForDeletion(IntrinsicI);
1406  return false;
1407  } else if (isa<Constant>(V)) {
1408  // If it's not false, and constant, it must evaluate to true. This means our
1409  // assume is assume(true), and thus, pointless, and we don't want to do
1410  // anything more here.
1411  return false;
1412  }
1413 
1414  Constant *True = ConstantInt::getTrue(V->getContext());
1415  bool Changed = false;
1416 
1417  for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1418  BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1419 
1420  // This property is only true in dominated successors, propagateEquality
1421  // will check dominance for us.
1422  Changed |= propagateEquality(V, True, Edge, false);
1423  }
1424 
1425  // We can replace assume value with true, which covers cases like this:
1426  // call void @llvm.assume(i1 %cmp)
1427  // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1428  ReplaceWithConstMap[V] = True;
1429 
1430  // If one of *cmp *eq operand is const, adding it to map will cover this:
1431  // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1432  // call void @llvm.assume(i1 %cmp)
1433  // ret float %0 ; will change it to ret float 3.000000e+00
1434  if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1435  if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1436  CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1437  (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1438  CmpI->getFastMathFlags().noNaNs())) {
1439  Value *CmpLHS = CmpI->getOperand(0);
1440  Value *CmpRHS = CmpI->getOperand(1);
1441  if (isa<Constant>(CmpLHS))
1442  std::swap(CmpLHS, CmpRHS);
1443  auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1444 
1445  // If only one operand is constant.
1446  if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1447  ReplaceWithConstMap[CmpLHS] = RHSConst;
1448  }
1449  }
1450  return Changed;
1451 }
1452 
1454  patchReplacementInstruction(I, Repl);
1455  I->replaceAllUsesWith(Repl);
1456 }
1457 
1458 /// Attempt to eliminate a load, first by eliminating it
1459 /// locally, and then attempting non-local elimination if that fails.
1460 bool GVN::processLoad(LoadInst *L) {
1461  if (!MD)
1462  return false;
1463 
1464  // This code hasn't been audited for ordered or volatile memory access
1465  if (!L->isUnordered())
1466  return false;
1467 
1468  if (L->use_empty()) {
1469  markInstructionForDeletion(L);
1470  return true;
1471  }
1472 
1473  // ... to a pointer that has been loaded from before...
1474  MemDepResult Dep = MD->getDependency(L);
1475 
1476  // If it is defined in another block, try harder.
1477  if (Dep.isNonLocal())
1478  return processNonLocalLoad(L);
1479 
1480  // Only handle the local case below
1481  if (!Dep.isDef() && !Dep.isClobber()) {
1482  // This might be a NonFuncLocal or an Unknown
1483  LLVM_DEBUG(
1484  // fast print dep, using operator<< on instruction is too slow.
1485  dbgs() << "GVN: load "; L->printAsOperand(dbgs());
1486  dbgs() << " has unknown dependence\n";);
1487  return false;
1488  }
1489 
1490  AvailableValue AV;
1491  if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1492  Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1493 
1494  // Replace the load!
1495  patchAndReplaceAllUsesWith(L, AvailableValue);
1496  markInstructionForDeletion(L);
1497  ++NumGVNLoad;
1498  reportLoadElim(L, AvailableValue, ORE);
1499  // Tell MDA to rexamine the reused pointer since we might have more
1500  // information after forwarding it.
1501  if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1502  MD->invalidateCachedPointerInfo(AvailableValue);
1503  return true;
1504  }
1505 
1506  return false;
1507 }
1508 
1509 /// Return a pair the first field showing the value number of \p Exp and the
1510 /// second field showing whether it is a value number newly created.
1511 std::pair<uint32_t, bool>
1512 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1513  uint32_t &e = expressionNumbering[Exp];
1514  bool CreateNewValNum = !e;
1515  if (CreateNewValNum) {
1516  Expressions.push_back(Exp);
1517  if (ExprIdx.size() < nextValueNumber + 1)
1518  ExprIdx.resize(nextValueNumber * 2);
1519  e = nextValueNumber;
1520  ExprIdx[nextValueNumber++] = nextExprNumber++;
1521  }
1522  return {e, CreateNewValNum};
1523 }
1524 
1525 /// Return whether all the values related with the same \p num are
1526 /// defined in \p BB.
1527 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1528  GVN &Gvn) {
1529  LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1530  while (Vals && Vals->BB == BB)
1531  Vals = Vals->Next;
1532  return !Vals;
1533 }
1534 
1535 /// Wrap phiTranslateImpl to provide caching functionality.
1537  const BasicBlock *PhiBlock, uint32_t Num,
1538  GVN &Gvn) {
1539  auto FindRes = PhiTranslateTable.find({Num, Pred});
1540  if (FindRes != PhiTranslateTable.end())
1541  return FindRes->second;
1542  uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1543  PhiTranslateTable.insert({{Num, Pred}, NewNum});
1544  return NewNum;
1545 }
1546 
1547 /// Translate value number \p Num using phis, so that it has the values of
1548 /// the phis in BB.
1549 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1550  const BasicBlock *PhiBlock,
1551  uint32_t Num, GVN &Gvn) {
1552  if (PHINode *PN = NumberingPhi[Num]) {
1553  for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1554  if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1555  if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1556  return TransVal;
1557  }
1558  return Num;
1559  }
1560 
1561  // If there is any value related with Num is defined in a BB other than
1562  // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1563  // a backedge. We can do an early exit in that case to save compile time.
1564  if (!areAllValsInBB(Num, PhiBlock, Gvn))
1565  return Num;
1566 
1567  if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1568  return Num;
1569  Expression Exp = Expressions[ExprIdx[Num]];
1570 
1571  for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1572  // For InsertValue and ExtractValue, some varargs are index numbers
1573  // instead of value numbers. Those index numbers should not be
1574  // translated.
1575  if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1576  (i > 0 && Exp.opcode == Instruction::ExtractValue))
1577  continue;
1578  Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1579  }
1580 
1581  if (Exp.commutative) {
1582  assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1583  if (Exp.varargs[0] > Exp.varargs[1]) {
1584  std::swap(Exp.varargs[0], Exp.varargs[1]);
1585  uint32_t Opcode = Exp.opcode >> 8;
1586  if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1587  Exp.opcode = (Opcode << 8) |
1589  static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1590  }
1591  }
1592 
1593  if (uint32_t NewNum = expressionNumbering[Exp])
1594  return NewNum;
1595  return Num;
1596 }
1597 
1598 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1599 /// again.
1601  const BasicBlock &CurrBlock) {
1602  for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1603  auto FindRes = PhiTranslateTable.find({Num, Pred});
1604  if (FindRes != PhiTranslateTable.end())
1605  PhiTranslateTable.erase(FindRes);
1606  }
1607 }
1608 
1609 // In order to find a leader for a given value number at a
1610 // specific basic block, we first obtain the list of all Values for that number,
1611 // and then scan the list to find one whose block dominates the block in
1612 // question. This is fast because dominator tree queries consist of only
1613 // a few comparisons of DFS numbers.
1614 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1615  LeaderTableEntry Vals = LeaderTable[num];
1616  if (!Vals.Val) return nullptr;
1617 
1618  Value *Val = nullptr;
1619  if (DT->dominates(Vals.BB, BB)) {
1620  Val = Vals.Val;
1621  if (isa<Constant>(Val)) return Val;
1622  }
1623 
1624  LeaderTableEntry* Next = Vals.Next;
1625  while (Next) {
1626  if (DT->dominates(Next->BB, BB)) {
1627  if (isa<Constant>(Next->Val)) return Next->Val;
1628  if (!Val) Val = Next->Val;
1629  }
1630 
1631  Next = Next->Next;
1632  }
1633 
1634  return Val;
1635 }
1636 
1637 /// There is an edge from 'Src' to 'Dst'. Return
1638 /// true if every path from the entry block to 'Dst' passes via this edge. In
1639 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1641  DominatorTree *DT) {
1642  // While in theory it is interesting to consider the case in which Dst has
1643  // more than one predecessor, because Dst might be part of a loop which is
1644  // only reachable from Src, in practice it is pointless since at the time
1645  // GVN runs all such loops have preheaders, which means that Dst will have
1646  // been changed to have only one predecessor, namely Src.
1647  const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1648  assert((!Pred || Pred == E.getStart()) &&
1649  "No edge between these basic blocks!");
1650  return Pred != nullptr;
1651 }
1652 
1653 void GVN::assignBlockRPONumber(Function &F) {
1654  BlockRPONumber.clear();
1655  uint32_t NextBlockNumber = 1;
1657  for (BasicBlock *BB : RPOT)
1658  BlockRPONumber[BB] = NextBlockNumber++;
1659  InvalidBlockRPONumbers = false;
1660 }
1661 
1662 // Tries to replace instruction with const, using information from
1663 // ReplaceWithConstMap.
1664 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1665  bool Changed = false;
1666  for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1667  Value *Operand = Instr->getOperand(OpNum);
1668  auto it = ReplaceWithConstMap.find(Operand);
1669  if (it != ReplaceWithConstMap.end()) {
1670  assert(!isa<Constant>(Operand) &&
1671  "Replacing constants with constants is invalid");
1672  LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "
1673  << *it->second << " in instruction " << *Instr << '\n');
1674  Instr->setOperand(OpNum, it->second);
1675  Changed = true;
1676  }
1677  }
1678  return Changed;
1679 }
1680 
1681 /// The given values are known to be equal in every block
1682 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1683 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1684 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1685 /// value starting from the end of Root.Start.
1686 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1687  bool DominatesByEdge) {
1689  Worklist.push_back(std::make_pair(LHS, RHS));
1690  bool Changed = false;
1691  // For speed, compute a conservative fast approximation to
1692  // DT->dominates(Root, Root.getEnd());
1693  const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1694 
1695  while (!Worklist.empty()) {
1696  std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1697  LHS = Item.first; RHS = Item.second;
1698 
1699  if (LHS == RHS)
1700  continue;
1701  assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1702 
1703  // Don't try to propagate equalities between constants.
1704  if (isa<Constant>(LHS) && isa<Constant>(RHS))
1705  continue;
1706 
1707  // Prefer a constant on the right-hand side, or an Argument if no constants.
1708  if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1709  std::swap(LHS, RHS);
1710  assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1711 
1712  // If there is no obvious reason to prefer the left-hand side over the
1713  // right-hand side, ensure the longest lived term is on the right-hand side,
1714  // so the shortest lived term will be replaced by the longest lived.
1715  // This tends to expose more simplifications.
1716  uint32_t LVN = VN.lookupOrAdd(LHS);
1717  if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1718  (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1719  // Move the 'oldest' value to the right-hand side, using the value number
1720  // as a proxy for age.
1721  uint32_t RVN = VN.lookupOrAdd(RHS);
1722  if (LVN < RVN) {
1723  std::swap(LHS, RHS);
1724  LVN = RVN;
1725  }
1726  }
1727 
1728  // If value numbering later sees that an instruction in the scope is equal
1729  // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1730  // the invariant that instructions only occur in the leader table for their
1731  // own value number (this is used by removeFromLeaderTable), do not do this
1732  // if RHS is an instruction (if an instruction in the scope is morphed into
1733  // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1734  // using the leader table is about compiling faster, not optimizing better).
1735  // The leader table only tracks basic blocks, not edges. Only add to if we
1736  // have the simple case where the edge dominates the end.
1737  if (RootDominatesEnd && !isa<Instruction>(RHS))
1738  addToLeaderTable(LVN, RHS, Root.getEnd());
1739 
1740  // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1741  // LHS always has at least one use that is not dominated by Root, this will
1742  // never do anything if LHS has only one use.
1743  if (!LHS->hasOneUse()) {
1744  unsigned NumReplacements =
1745  DominatesByEdge
1746  ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1747  : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1748 
1749  Changed |= NumReplacements > 0;
1750  NumGVNEqProp += NumReplacements;
1751  // Cached information for anything that uses LHS will be invalid.
1752  if (MD)
1753  MD->invalidateCachedPointerInfo(LHS);
1754  }
1755 
1756  // Now try to deduce additional equalities from this one. For example, if
1757  // the known equality was "(A != B)" == "false" then it follows that A and B
1758  // are equal in the scope. Only boolean equalities with an explicit true or
1759  // false RHS are currently supported.
1760  if (!RHS->getType()->isIntegerTy(1))
1761  // Not a boolean equality - bail out.
1762  continue;
1763  ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1764  if (!CI)
1765  // RHS neither 'true' nor 'false' - bail out.
1766  continue;
1767  // Whether RHS equals 'true'. Otherwise it equals 'false'.
1768  bool isKnownTrue = CI->isMinusOne();
1769  bool isKnownFalse = !isKnownTrue;
1770 
1771  // If "A && B" is known true then both A and B are known true. If "A || B"
1772  // is known false then both A and B are known false.
1773  Value *A, *B;
1774  if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1775  (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1776  Worklist.push_back(std::make_pair(A, RHS));
1777  Worklist.push_back(std::make_pair(B, RHS));
1778  continue;
1779  }
1780 
1781  // If we are propagating an equality like "(A == B)" == "true" then also
1782  // propagate the equality A == B. When propagating a comparison such as
1783  // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1784  if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1785  Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1786 
1787  // If "A == B" is known true, or "A != B" is known false, then replace
1788  // A with B everywhere in the scope.
1789  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1790  (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1791  Worklist.push_back(std::make_pair(Op0, Op1));
1792 
1793  // Handle the floating point versions of equality comparisons too.
1794  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1795  (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1796 
1797  // Floating point -0.0 and 0.0 compare equal, so we can only
1798  // propagate values if we know that we have a constant and that
1799  // its value is non-zero.
1800 
1801  // FIXME: We should do this optimization if 'no signed zeros' is
1802  // applicable via an instruction-level fast-math-flag or some other
1803  // indicator that relaxed FP semantics are being used.
1804 
1805  if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1806  Worklist.push_back(std::make_pair(Op0, Op1));
1807  }
1808 
1809  // If "A >= B" is known true, replace "A < B" with false everywhere.
1810  CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1811  Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1812  // Since we don't have the instruction "A < B" immediately to hand, work
1813  // out the value number that it would have and use that to find an
1814  // appropriate instruction (if any).
1815  uint32_t NextNum = VN.getNextUnusedValueNumber();
1816  uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1817  // If the number we were assigned was brand new then there is no point in
1818  // looking for an instruction realizing it: there cannot be one!
1819  if (Num < NextNum) {
1820  Value *NotCmp = findLeader(Root.getEnd(), Num);
1821  if (NotCmp && isa<Instruction>(NotCmp)) {
1822  unsigned NumReplacements =
1823  DominatesByEdge
1824  ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1825  : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1826  Root.getStart());
1827  Changed |= NumReplacements > 0;
1828  NumGVNEqProp += NumReplacements;
1829  // Cached information for anything that uses NotCmp will be invalid.
1830  if (MD)
1831  MD->invalidateCachedPointerInfo(NotCmp);
1832  }
1833  }
1834  // Ensure that any instruction in scope that gets the "A < B" value number
1835  // is replaced with false.
1836  // The leader table only tracks basic blocks, not edges. Only add to if we
1837  // have the simple case where the edge dominates the end.
1838  if (RootDominatesEnd)
1839  addToLeaderTable(Num, NotVal, Root.getEnd());
1840 
1841  continue;
1842  }
1843  }
1844 
1845  return Changed;
1846 }
1847 
1848 /// When calculating availability, handle an instruction
1849 /// by inserting it into the appropriate sets
1850 bool GVN::processInstruction(Instruction *I) {
1851  // Ignore dbg info intrinsics.
1852  if (isa<DbgInfoIntrinsic>(I))
1853  return false;
1854 
1855  // If the instruction can be easily simplified then do so now in preference
1856  // to value numbering it. Value numbering often exposes redundancies, for
1857  // example if it determines that %y is equal to %x then the instruction
1858  // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1859  const DataLayout &DL = I->getModule()->getDataLayout();
1860  if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1861  bool Changed = false;
1862  if (!I->use_empty()) {
1863  I->replaceAllUsesWith(V);
1864  Changed = true;
1865  }
1866  if (isInstructionTriviallyDead(I, TLI)) {
1867  markInstructionForDeletion(I);
1868  Changed = true;
1869  }
1870  if (Changed) {
1871  if (MD && V->getType()->isPtrOrPtrVectorTy())
1872  MD->invalidateCachedPointerInfo(V);
1873  ++NumGVNSimpl;
1874  return true;
1875  }
1876  }
1877 
1878  if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1879  if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1880  return processAssumeIntrinsic(IntrinsicI);
1881 
1882  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1883  if (processLoad(LI))
1884  return true;
1885 
1886  unsigned Num = VN.lookupOrAdd(LI);
1887  addToLeaderTable(Num, LI, LI->getParent());
1888  return false;
1889  }
1890 
1891  // For conditional branches, we can perform simple conditional propagation on
1892  // the condition value itself.
1893  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1894  if (!BI->isConditional())
1895  return false;
1896 
1897  if (isa<Constant>(BI->getCondition()))
1898  return processFoldableCondBr(BI);
1899 
1900  Value *BranchCond = BI->getCondition();
1901  BasicBlock *TrueSucc = BI->getSuccessor(0);
1902  BasicBlock *FalseSucc = BI->getSuccessor(1);
1903  // Avoid multiple edges early.
1904  if (TrueSucc == FalseSucc)
1905  return false;
1906 
1907  BasicBlock *Parent = BI->getParent();
1908  bool Changed = false;
1909 
1910  Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1911  BasicBlockEdge TrueE(Parent, TrueSucc);
1912  Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1913 
1914  Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1915  BasicBlockEdge FalseE(Parent, FalseSucc);
1916  Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1917 
1918  return Changed;
1919  }
1920 
1921  // For switches, propagate the case values into the case destinations.
1922  if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1923  Value *SwitchCond = SI->getCondition();
1924  BasicBlock *Parent = SI->getParent();
1925  bool Changed = false;
1926 
1927  // Remember how many outgoing edges there are to every successor.
1929  for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1930  ++SwitchEdges[SI->getSuccessor(i)];
1931 
1932  for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1933  i != e; ++i) {
1934  BasicBlock *Dst = i->getCaseSuccessor();
1935  // If there is only a single edge, propagate the case value into it.
1936  if (SwitchEdges.lookup(Dst) == 1) {
1937  BasicBlockEdge E(Parent, Dst);
1938  Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1939  }
1940  }
1941  return Changed;
1942  }
1943 
1944  // Instructions with void type don't return a value, so there's
1945  // no point in trying to find redundancies in them.
1946  if (I->getType()->isVoidTy())
1947  return false;
1948 
1949  uint32_t NextNum = VN.getNextUnusedValueNumber();
1950  unsigned Num = VN.lookupOrAdd(I);
1951 
1952  // Allocations are always uniquely numbered, so we can save time and memory
1953  // by fast failing them.
1954  if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) {
1955  addToLeaderTable(Num, I, I->getParent());
1956  return false;
1957  }
1958 
1959  // If the number we were assigned was a brand new VN, then we don't
1960  // need to do a lookup to see if the number already exists
1961  // somewhere in the domtree: it can't!
1962  if (Num >= NextNum) {
1963  addToLeaderTable(Num, I, I->getParent());
1964  return false;
1965  }
1966 
1967  // Perform fast-path value-number based elimination of values inherited from
1968  // dominators.
1969  Value *Repl = findLeader(I->getParent(), Num);
1970  if (!Repl) {
1971  // Failure, just remember this instance for future use.
1972  addToLeaderTable(Num, I, I->getParent());
1973  return false;
1974  } else if (Repl == I) {
1975  // If I was the result of a shortcut PRE, it might already be in the table
1976  // and the best replacement for itself. Nothing to do.
1977  return false;
1978  }
1979 
1980  // Remove it!
1981  patchAndReplaceAllUsesWith(I, Repl);
1982  if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1983  MD->invalidateCachedPointerInfo(Repl);
1984  markInstructionForDeletion(I);
1985  return true;
1986 }
1987 
1988 /// runOnFunction - This is the main transformation entry point for a function.
1989 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1990  const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1991  MemoryDependenceResults *RunMD, LoopInfo *LI,
1992  OptimizationRemarkEmitter *RunORE) {
1993  AC = &RunAC;
1994  DT = &RunDT;
1995  VN.setDomTree(DT);
1996  TLI = &RunTLI;
1997  VN.setAliasAnalysis(&RunAA);
1998  MD = RunMD;
1999  ImplicitControlFlowTracking ImplicitCFT(DT);
2000  ICF = &ImplicitCFT;
2001  this->LI = LI;
2002  VN.setMemDep(MD);
2003  ORE = RunORE;
2004  InvalidBlockRPONumbers = true;
2005 
2006  bool Changed = false;
2007  bool ShouldContinue = true;
2008 
2010  // Merge unconditional branches, allowing PRE to catch more
2011  // optimization opportunities.
2012  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2013  BasicBlock *BB = &*FI++;
2014 
2015  bool removedBlock = MergeBlockIntoPredecessor(BB, &DTU, LI, nullptr, MD);
2016  if (removedBlock)
2017  ++NumGVNBlocks;
2018 
2019  Changed |= removedBlock;
2020  }
2021 
2022  unsigned Iteration = 0;
2023  while (ShouldContinue) {
2024  LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2025  ShouldContinue = iterateOnFunction(F);
2026  Changed |= ShouldContinue;
2027  ++Iteration;
2028  }
2029 
2030  if (EnablePRE) {
2031  // Fabricate val-num for dead-code in order to suppress assertion in
2032  // performPRE().
2033  assignValNumForDeadCode();
2034  bool PREChanged = true;
2035  while (PREChanged) {
2036  PREChanged = performPRE(F);
2037  Changed |= PREChanged;
2038  }
2039  }
2040 
2041  // FIXME: Should perform GVN again after PRE does something. PRE can move
2042  // computations into blocks where they become fully redundant. Note that
2043  // we can't do this until PRE's critical edge splitting updates memdep.
2044  // Actually, when this happens, we should just fully integrate PRE into GVN.
2045 
2046  cleanupGlobalSets();
2047  // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2048  // iteration.
2049  DeadBlocks.clear();
2050 
2051  return Changed;
2052 }
2053 
2054 bool GVN::processBlock(BasicBlock *BB) {
2055  // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2056  // (and incrementing BI before processing an instruction).
2057  assert(InstrsToErase.empty() &&
2058  "We expect InstrsToErase to be empty across iterations");
2059  if (DeadBlocks.count(BB))
2060  return false;
2061 
2062  // Clearing map before every BB because it can be used only for single BB.
2063  ReplaceWithConstMap.clear();
2064  bool ChangedFunction = false;
2065 
2066  for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2067  BI != BE;) {
2068  if (!ReplaceWithConstMap.empty())
2069  ChangedFunction |= replaceOperandsWithConsts(&*BI);
2070  ChangedFunction |= processInstruction(&*BI);
2071 
2072  if (InstrsToErase.empty()) {
2073  ++BI;
2074  continue;
2075  }
2076 
2077  // If we need some instructions deleted, do it now.
2078  NumGVNInstr += InstrsToErase.size();
2079 
2080  // Avoid iterator invalidation.
2081  bool AtStart = BI == BB->begin();
2082  if (!AtStart)
2083  --BI;
2084 
2085  for (auto *I : InstrsToErase) {
2086  assert(I->getParent() == BB && "Removing instruction from wrong block?");
2087  LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n');
2088  salvageDebugInfo(*I);
2089  if (MD) MD->removeInstruction(I);
2090  LLVM_DEBUG(verifyRemoved(I));
2091  ICF->removeInstruction(I);
2092  I->eraseFromParent();
2093  }
2094  InstrsToErase.clear();
2095 
2096  if (AtStart)
2097  BI = BB->begin();
2098  else
2099  ++BI;
2100  }
2101 
2102  return ChangedFunction;
2103 }
2104 
2105 // Instantiate an expression in a predecessor that lacked it.
2106 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2107  BasicBlock *Curr, unsigned int ValNo) {
2108  // Because we are going top-down through the block, all value numbers
2109  // will be available in the predecessor by the time we need them. Any
2110  // that weren't originally present will have been instantiated earlier
2111  // in this loop.
2112  bool success = true;
2113  for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2114  Value *Op = Instr->getOperand(i);
2115  if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2116  continue;
2117  // This could be a newly inserted instruction, in which case, we won't
2118  // find a value number, and should give up before we hurt ourselves.
2119  // FIXME: Rewrite the infrastructure to let it easier to value number
2120  // and process newly inserted instructions.
2121  if (!VN.exists(Op)) {
2122  success = false;
2123  break;
2124  }
2125  uint32_t TValNo =
2126  VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2127  if (Value *V = findLeader(Pred, TValNo)) {
2128  Instr->setOperand(i, V);
2129  } else {
2130  success = false;
2131  break;
2132  }
2133  }
2134 
2135  // Fail out if we encounter an operand that is not available in
2136  // the PRE predecessor. This is typically because of loads which
2137  // are not value numbered precisely.
2138  if (!success)
2139  return false;
2140 
2141  Instr->insertBefore(Pred->getTerminator());
2142  Instr->setName(Instr->getName() + ".pre");
2143  Instr->setDebugLoc(Instr->getDebugLoc());
2144 
2145  unsigned Num = VN.lookupOrAdd(Instr);
2146  VN.add(Instr, Num);
2147 
2148  // Update the availability map to include the new instruction.
2149  addToLeaderTable(Num, Instr, Pred);
2150  return true;
2151 }
2152 
2153 bool GVN::performScalarPRE(Instruction *CurInst) {
2154  if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() ||
2155  isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2156  CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2157  isa<DbgInfoIntrinsic>(CurInst))
2158  return false;
2159 
2160  // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2161  // sinking the compare again, and it would force the code generator to
2162  // move the i1 from processor flags or predicate registers into a general
2163  // purpose register.
2164  if (isa<CmpInst>(CurInst))
2165  return false;
2166 
2167  // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from
2168  // sinking the addressing mode computation back to its uses. Extending the
2169  // GEP's live range increases the register pressure, and therefore it can
2170  // introduce unnecessary spills.
2171  //
2172  // This doesn't prevent Load PRE. PHI translation will make the GEP available
2173  // to the load by moving it to the predecessor block if necessary.
2174  if (isa<GetElementPtrInst>(CurInst))
2175  return false;
2176 
2177  // We don't currently value number ANY inline asm calls.
2178  if (auto *CallB = dyn_cast<CallBase>(CurInst))
2179  if (CallB->isInlineAsm())
2180  return false;
2181 
2182  uint32_t ValNo = VN.lookup(CurInst);
2183 
2184  // Look for the predecessors for PRE opportunities. We're
2185  // only trying to solve the basic diamond case, where
2186  // a value is computed in the successor and one predecessor,
2187  // but not the other. We also explicitly disallow cases
2188  // where the successor is its own predecessor, because they're
2189  // more complicated to get right.
2190  unsigned NumWith = 0;
2191  unsigned NumWithout = 0;
2192  BasicBlock *PREPred = nullptr;
2193  BasicBlock *CurrentBlock = CurInst->getParent();
2194 
2195  // Update the RPO numbers for this function.
2196  if (InvalidBlockRPONumbers)
2197  assignBlockRPONumber(*CurrentBlock->getParent());
2198 
2200  for (BasicBlock *P : predecessors(CurrentBlock)) {
2201  // We're not interested in PRE where blocks with predecessors that are
2202  // not reachable.
2203  if (!DT->isReachableFromEntry(P)) {
2204  NumWithout = 2;
2205  break;
2206  }
2207  // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2208  // when CurInst has operand defined in CurrentBlock (so it may be defined
2209  // by phi in the loop header).
2210  assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) &&
2211  "Invalid BlockRPONumber map.");
2212  if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2213  llvm::any_of(CurInst->operands(), [&](const Use &U) {
2214  if (auto *Inst = dyn_cast<Instruction>(U.get()))
2215  return Inst->getParent() == CurrentBlock;
2216  return false;
2217  })) {
2218  NumWithout = 2;
2219  break;
2220  }
2221 
2222  uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2223  Value *predV = findLeader(P, TValNo);
2224  if (!predV) {
2225  predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2226  PREPred = P;
2227  ++NumWithout;
2228  } else if (predV == CurInst) {
2229  /* CurInst dominates this predecessor. */
2230  NumWithout = 2;
2231  break;
2232  } else {
2233  predMap.push_back(std::make_pair(predV, P));
2234  ++NumWith;
2235  }
2236  }
2237 
2238  // Don't do PRE when it might increase code size, i.e. when
2239  // we would need to insert instructions in more than one pred.
2240  if (NumWithout > 1 || NumWith == 0)
2241  return false;
2242 
2243  // We may have a case where all predecessors have the instruction,
2244  // and we just need to insert a phi node. Otherwise, perform
2245  // insertion.
2246  Instruction *PREInstr = nullptr;
2247 
2248  if (NumWithout != 0) {
2249  if (!isSafeToSpeculativelyExecute(CurInst)) {
2250  // It is only valid to insert a new instruction if the current instruction
2251  // is always executed. An instruction with implicit control flow could
2252  // prevent us from doing it. If we cannot speculate the execution, then
2253  // PRE should be prohibited.
2254  if (ICF->isDominatedByICFIFromSameBlock(CurInst))
2255  return false;
2256  }
2257 
2258  // Don't do PRE across indirect branch.
2259  if (isa<IndirectBrInst>(PREPred->getTerminator()))
2260  return false;
2261 
2262  // Don't do PRE across callbr.
2263  // FIXME: Can we do this across the fallthrough edge?
2264  if (isa<CallBrInst>(PREPred->getTerminator()))
2265  return false;
2266 
2267  // We can't do PRE safely on a critical edge, so instead we schedule
2268  // the edge to be split and perform the PRE the next time we iterate
2269  // on the function.
2270  unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2271  if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2272  toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2273  return false;
2274  }
2275  // We need to insert somewhere, so let's give it a shot
2276  PREInstr = CurInst->clone();
2277  if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2278  // If we failed insertion, make sure we remove the instruction.
2279  LLVM_DEBUG(verifyRemoved(PREInstr));
2280  PREInstr->deleteValue();
2281  return false;
2282  }
2283  }
2284 
2285  // Either we should have filled in the PRE instruction, or we should
2286  // not have needed insertions.
2287  assert(PREInstr != nullptr || NumWithout == 0);
2288 
2289  ++NumGVNPRE;
2290 
2291  // Create a PHI to make the value available in this block.
2292  PHINode *Phi =
2293  PHINode::Create(CurInst->getType(), predMap.size(),
2294  CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2295  for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2296  if (Value *V = predMap[i].first) {
2297  // If we use an existing value in this phi, we have to patch the original
2298  // value because the phi will be used to replace a later value.
2299  patchReplacementInstruction(CurInst, V);
2300  Phi->addIncoming(V, predMap[i].second);
2301  } else
2302  Phi->addIncoming(PREInstr, PREPred);
2303  }
2304 
2305  VN.add(Phi, ValNo);
2306  // After creating a new PHI for ValNo, the phi translate result for ValNo will
2307  // be changed, so erase the related stale entries in phi translate cache.
2308  VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2309  addToLeaderTable(ValNo, Phi, CurrentBlock);
2310  Phi->setDebugLoc(CurInst->getDebugLoc());
2311  CurInst->replaceAllUsesWith(Phi);
2312  if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2313  MD->invalidateCachedPointerInfo(Phi);
2314  VN.erase(CurInst);
2315  removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2316 
2317  LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2318  if (MD)
2319  MD->removeInstruction(CurInst);
2320  LLVM_DEBUG(verifyRemoved(CurInst));
2321  // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes
2322  // some assertion failures.
2323  ICF->removeInstruction(CurInst);
2324  CurInst->eraseFromParent();
2325  ++NumGVNInstr;
2326 
2327  return true;
2328 }
2329 
2330 /// Perform a purely local form of PRE that looks for diamond
2331 /// control flow patterns and attempts to perform simple PRE at the join point.
2332 bool GVN::performPRE(Function &F) {
2333  bool Changed = false;
2334  for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2335  // Nothing to PRE in the entry block.
2336  if (CurrentBlock == &F.getEntryBlock())
2337  continue;
2338 
2339  // Don't perform PRE on an EH pad.
2340  if (CurrentBlock->isEHPad())
2341  continue;
2342 
2343  for (BasicBlock::iterator BI = CurrentBlock->begin(),
2344  BE = CurrentBlock->end();
2345  BI != BE;) {
2346  Instruction *CurInst = &*BI++;
2347  Changed |= performScalarPRE(CurInst);
2348  }
2349  }
2350 
2351  if (splitCriticalEdges())
2352  Changed = true;
2353 
2354  return Changed;
2355 }
2356 
2357 /// Split the critical edge connecting the given two blocks, and return
2358 /// the block inserted to the critical edge.
2359 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2360  BasicBlock *BB =
2361  SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT, LI));
2362  if (MD)
2363  MD->invalidateCachedPredecessors();
2364  InvalidBlockRPONumbers = true;
2365  return BB;
2366 }
2367 
2368 /// Split critical edges found during the previous
2369 /// iteration that may enable further optimization.
2370 bool GVN::splitCriticalEdges() {
2371  if (toSplit.empty())
2372  return false;
2373  do {
2374  std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val();
2375  SplitCriticalEdge(Edge.first, Edge.second,
2377  } while (!toSplit.empty());
2378  if (MD) MD->invalidateCachedPredecessors();
2379  InvalidBlockRPONumbers = true;
2380  return true;
2381 }
2382 
2383 /// Executes one iteration of GVN
2384 bool GVN::iterateOnFunction(Function &F) {
2385  cleanupGlobalSets();
2386 
2387  // Top-down walk of the dominator tree
2388  bool Changed = false;
2389  // Needed for value numbering with phi construction to work.
2390  // RPOT walks the graph in its constructor and will not be invalidated during
2391  // processBlock.
2393 
2394  for (BasicBlock *BB : RPOT)
2395  Changed |= processBlock(BB);
2396 
2397  return Changed;
2398 }
2399 
2400 void GVN::cleanupGlobalSets() {
2401  VN.clear();
2402  LeaderTable.clear();
2403  BlockRPONumber.clear();
2404  TableAllocator.Reset();
2405  ICF->clear();
2406  InvalidBlockRPONumbers = true;
2407 }
2408 
2409 /// Verify that the specified instruction does not occur in our
2410 /// internal data structures.
2411 void GVN::verifyRemoved(const Instruction *Inst) const {
2412  VN.verifyRemoved(Inst);
2413 
2414  // Walk through the value number scope to make sure the instruction isn't
2415  // ferreted away in it.
2417  I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2418  const LeaderTableEntry *Node = &I->second;
2419  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2420 
2421  while (Node->Next) {
2422  Node = Node->Next;
2423  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2424  }
2425  }
2426 }
2427 
2428 /// BB is declared dead, which implied other blocks become dead as well. This
2429 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2430 /// live successors, update their phi nodes by replacing the operands
2431 /// corresponding to dead blocks with UndefVal.
2432 void GVN::addDeadBlock(BasicBlock *BB) {
2435 
2436  NewDead.push_back(BB);
2437  while (!NewDead.empty()) {
2438  BasicBlock *D = NewDead.pop_back_val();
2439  if (DeadBlocks.count(D))
2440  continue;
2441 
2442  // All blocks dominated by D are dead.
2444  DT->getDescendants(D, Dom);
2445  DeadBlocks.insert(Dom.begin(), Dom.end());
2446 
2447  // Figure out the dominance-frontier(D).
2448  for (BasicBlock *B : Dom) {
2449  for (BasicBlock *S : successors(B)) {
2450  if (DeadBlocks.count(S))
2451  continue;
2452 
2453  bool AllPredDead = true;
2454  for (BasicBlock *P : predecessors(S))
2455  if (!DeadBlocks.count(P)) {
2456  AllPredDead = false;
2457  break;
2458  }
2459 
2460  if (!AllPredDead) {
2461  // S could be proved dead later on. That is why we don't update phi
2462  // operands at this moment.
2463  DF.insert(S);
2464  } else {
2465  // While S is not dominated by D, it is dead by now. This could take
2466  // place if S already have a dead predecessor before D is declared
2467  // dead.
2468  NewDead.push_back(S);
2469  }
2470  }
2471  }
2472  }
2473 
2474  // For the dead blocks' live successors, update their phi nodes by replacing
2475  // the operands corresponding to dead blocks with UndefVal.
2477  I != E; I++) {
2478  BasicBlock *B = *I;
2479  if (DeadBlocks.count(B))
2480  continue;
2481 
2482  // First, split the critical edges. This might also create additional blocks
2483  // to preserve LoopSimplify form and adjust edges accordingly.
2485  for (BasicBlock *P : Preds) {
2486  if (!DeadBlocks.count(P))
2487  continue;
2488 
2489  if (llvm::any_of(successors(P),
2490  [B](BasicBlock *Succ) { return Succ == B; }) &&
2491  isCriticalEdge(P->getTerminator(), B)) {
2492  if (BasicBlock *S = splitCriticalEdges(P, B))
2493  DeadBlocks.insert(P = S);
2494  }
2495  }
2496 
2497  // Now undef the incoming values from the dead predecessors.
2498  for (BasicBlock *P : predecessors(B)) {
2499  if (!DeadBlocks.count(P))
2500  continue;
2501  for (PHINode &Phi : B->phis()) {
2502  Phi.setIncomingValueForBlock(P, UndefValue::get(Phi.getType()));
2503  if (MD)
2504  MD->invalidateCachedPointerInfo(&Phi);
2505  }
2506  }
2507  }
2508 }
2509 
2510 // If the given branch is recognized as a foldable branch (i.e. conditional
2511 // branch with constant condition), it will perform following analyses and
2512 // transformation.
2513 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2514 // R be the target of the dead out-coming edge.
2515 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2516 // edge. The result of this step will be {X| X is dominated by R}
2517 // 2) Identify those blocks which haves at least one dead predecessor. The
2518 // result of this step will be dominance-frontier(R).
2519 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2520 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2521 //
2522 // Return true iff *NEW* dead code are found.
2523 bool GVN::processFoldableCondBr(BranchInst *BI) {
2524  if (!BI || BI->isUnconditional())
2525  return false;
2526 
2527  // If a branch has two identical successors, we cannot declare either dead.
2528  if (BI->getSuccessor(0) == BI->getSuccessor(1))
2529  return false;
2530 
2531  ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2532  if (!Cond)
2533  return false;
2534 
2535  BasicBlock *DeadRoot =
2536  Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2537  if (DeadBlocks.count(DeadRoot))
2538  return false;
2539 
2540  if (!DeadRoot->getSinglePredecessor())
2541  DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2542 
2543  addDeadBlock(DeadRoot);
2544  return true;
2545 }
2546 
2547 // performPRE() will trigger assert if it comes across an instruction without
2548 // associated val-num. As it normally has far more live instructions than dead
2549 // instructions, it makes more sense just to "fabricate" a val-number for the
2550 // dead code than checking if instruction involved is dead or not.
2551 void GVN::assignValNumForDeadCode() {
2552  for (BasicBlock *BB : DeadBlocks) {
2553  for (Instruction &Inst : *BB) {
2554  unsigned ValNum = VN.lookupOrAdd(&Inst);
2555  addToLeaderTable(ValNum, &Inst, BB);
2556  }
2557  }
2558 }
2559 
2561 public:
2562  static char ID; // Pass identification, replacement for typeid
2563 
2564  explicit GVNLegacyPass(bool NoMemDepAnalysis = !EnableMemDep)
2565  : FunctionPass(ID), NoMemDepAnalysis(NoMemDepAnalysis) {
2567  }
2568 
2569  bool runOnFunction(Function &F) override {
2570  if (skipFunction(F))
2571  return false;
2572 
2573  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2574 
2575  return Impl.runImpl(
2576  F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2577  getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2578  getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2579  getAnalysis<AAResultsWrapperPass>().getAAResults(),
2580  NoMemDepAnalysis ? nullptr
2581  : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2582  LIWP ? &LIWP->getLoopInfo() : nullptr,
2583  &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2584  }
2585 
2586  void getAnalysisUsage(AnalysisUsage &AU) const override {
2591  if (!NoMemDepAnalysis)
2594 
2601  }
2602 
2603 private:
2604  bool NoMemDepAnalysis;
2605  GVN Impl;
2606 };
2607 
2608 char GVNLegacyPass::ID = 0;
2609 
2610 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2618 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2619 
2620 // The public interface to this file...
2621 FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) {
2622  return new GVNLegacyPass(NoMemDepAnalysis);
2623 }
Legacy wrapper pass to provide the GlobalsAAResult object.
static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV)
Definition: GVN.cpp:245
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:831
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:2621
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:1600
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:111
bool isUndefValue() const
Definition: GVN.cpp:213
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:616
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:186
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:176
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:476
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:570
iterator end()
Definition: Function.h:682
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:121
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:1600
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:1536
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:211
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:2497
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:289
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:557
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:4428
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:2586
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:2428
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:2569
Option class for critical edge splitting.
void clear()
Remove all entries from the ValueTable.
Definition: GVN.cpp:578
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:654
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:1183
static const uint16_t * lookup(unsigned opcode, unsigned domain, ArrayRef< uint16_t[3]> Table)
bool isSimpleValue() const
Definition: GVN.cpp:210
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:245
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:119
#define DEBUG_TYPE
Definition: GVN.cpp:89
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:194
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.
AnalysisUsage & addPreservedID(const void *ID)
LoadInst * getCoercedLoadValue() const
Definition: GVN.cpp:220
static GVN::Expression getEmptyKey()
Definition: GVN.cpp:143
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:369
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:680
static unsigned getHashValue(const GVN::Expression &e)
Definition: GVN.cpp:146
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:664
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:875
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:117
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:837
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:144
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:730
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:152
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:1172
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:173
MemIntrinsic * getMemIntrinValue() const
Definition: GVN.cpp:225
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:1294
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1446
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:4355
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
char & LoopSimplifyID
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:827
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:263
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
Value * PHITranslateWithInsertion(BasicBlock *CurBB, BasicBlock *PredBB, const DominatorTree &DT, SmallVectorImpl< Instruction *> &NewInsts)
PHITranslateWithInsertion - PHI translate this value into the specified predecessor block...
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:653
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:770
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:609
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:238
iterator_range< user_iterator > users()
Definition: Value.h:419
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:600
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:590
static bool isLifetimeStart(const Instruction *Inst)
Definition: GVN.cpp:819
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:2564
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:178
uint32_t opcode
Definition: GVN.cpp:114
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:477
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
iterator_range< const_phi_iterator > phis() const
Returns a range that iterates over the phis in the basic block.
Definition: BasicBlock.h:324
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:133
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:481
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:215
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:163
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:657
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:359
LLVM Value Representation.
Definition: Value.h:73
static AvailableValueInBlock getUndef(BasicBlock *BB)
Definition: GVN.cpp:257
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.
The legacy pass manager&#39;s analysis pass to compute loop information.
Definition: LoopInfo.h:1208
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: GVN.cpp:611
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:432
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.
BasicBlock * BB
BB - The basic block in question.
Definition: GVN.cpp:240
static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl)
Definition: GVN.cpp:1453
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:259
bool isMemIntrinValue() const
Definition: GVN.cpp:212
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:243
#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:342
static AvailableValue getUndef()
Definition: GVN.cpp:202
static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, DominatorTree *DT)
There is an edge from &#39;Src&#39; to &#39;Dst&#39;.
Definition: GVN.cpp:1640
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