LLVM  9.0.0svn
GVN.cpp
Go to the documentation of this file.
1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass performs global value numbering to eliminate fully redundant
10 // instructions. It also performs simple dead load elimination.
11 //
12 // Note that this pass does the value numbering itself; it does not use the
13 // ValueNumbering analysis passes.
14 //
15 //===----------------------------------------------------------------------===//
16 
18 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/Hashing.h"
21 #include "llvm/ADT/MapVector.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SetVector.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/CFG.h"
35 #include "llvm/Analysis/LoopInfo.h"
42 #include "llvm/Config/llvm-config.h"
43 #include "llvm/IR/Attributes.h"
44 #include "llvm/IR/BasicBlock.h"
45 #include "llvm/IR/CallSite.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
50 #include "llvm/IR/DebugLoc.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
53 #include "llvm/IR/InstrTypes.h"
54 #include "llvm/IR/Instruction.h"
55 #include "llvm/IR/Instructions.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/Intrinsics.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/Metadata.h"
60 #include "llvm/IR/Module.h"
61 #include "llvm/IR/Operator.h"
62 #include "llvm/IR/PassManager.h"
63 #include "llvm/IR/PatternMatch.h"
64 #include "llvm/IR/Type.h"
65 #include "llvm/IR/Use.h"
66 #include "llvm/IR/Value.h"
67 #include "llvm/Pass.h"
68 #include "llvm/Support/Casting.h"
70 #include "llvm/Support/Compiler.h"
71 #include "llvm/Support/Debug.h"
77 #include <algorithm>
78 #include <cassert>
79 #include <cstdint>
80 #include <utility>
81 #include <vector>
82 
83 using namespace llvm;
84 using namespace llvm::gvn;
85 using namespace llvm::VNCoercion;
86 using namespace PatternMatch;
87 
88 #define DEBUG_TYPE "gvn"
89 
90 STATISTIC(NumGVNInstr, "Number of instructions deleted");
91 STATISTIC(NumGVNLoad, "Number of loads deleted");
92 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
93 STATISTIC(NumGVNBlocks, "Number of blocks merged");
94 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
95 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
96 STATISTIC(NumPRELoad, "Number of loads PRE'd");
97 
98 static cl::opt<bool> EnablePRE("enable-pre",
99  cl::init(true), cl::Hidden);
100 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
101 static cl::opt<bool> EnableMemDep("enable-gvn-memdep", cl::init(true));
102 
103 // Maximum allowed recursion depth.
104 static cl::opt<uint32_t>
105 MaxRecurseDepth("gvn-max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
106  cl::desc("Max recurse depth in GVN (default = 1000)"));
107 
109  "gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore,
110  cl::desc("Max number of dependences to attempt Load PRE (default = 100)"));
111 
115  bool commutative = false;
117 
118  Expression(uint32_t o = ~2U) : opcode(o) {}
119 
120  bool operator==(const Expression &other) const {
121  if (opcode != other.opcode)
122  return false;
123  if (opcode == ~0U || opcode == ~1U)
124  return true;
125  if (type != other.type)
126  return false;
127  if (varargs != other.varargs)
128  return false;
129  return true;
130  }
131 
133  return hash_combine(
134  Value.opcode, Value.type,
135  hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
136  }
137 };
138 
139 namespace llvm {
140 
141 template <> struct DenseMapInfo<GVN::Expression> {
142  static inline GVN::Expression getEmptyKey() { return ~0U; }
143  static inline GVN::Expression getTombstoneKey() { return ~1U; }
144 
145  static unsigned getHashValue(const GVN::Expression &e) {
146  using llvm::hash_value;
147 
148  return static_cast<unsigned>(hash_value(e));
149  }
150 
151  static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
152  return LHS == RHS;
153  }
154 };
155 
156 } // end namespace llvm
157 
158 /// Represents a particular available value that we know how to materialize.
159 /// Materialization of an AvailableValue never fails. An AvailableValue is
160 /// implicitly associated with a rematerialization point which is the
161 /// location of the instruction from which it was formed.
163  enum ValType {
164  SimpleVal, // A simple offsetted value that is accessed.
165  LoadVal, // A value produced by a load.
166  MemIntrin, // A memory intrinsic which is loaded from.
167  UndefVal // A UndefValue representing a value from dead block (which
168  // is not yet physically removed from the CFG).
169  };
170 
171  /// V - The value that is live out of the block.
173 
174  /// Offset - The byte offset in Val that is interesting for the load query.
175  unsigned Offset;
176 
177  static AvailableValue get(Value *V, unsigned Offset = 0) {
178  AvailableValue Res;
179  Res.Val.setPointer(V);
180  Res.Val.setInt(SimpleVal);
181  Res.Offset = Offset;
182  return Res;
183  }
184 
185  static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
186  AvailableValue Res;
187  Res.Val.setPointer(MI);
188  Res.Val.setInt(MemIntrin);
189  Res.Offset = Offset;
190  return Res;
191  }
192 
193  static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
194  AvailableValue Res;
195  Res.Val.setPointer(LI);
196  Res.Val.setInt(LoadVal);
197  Res.Offset = Offset;
198  return Res;
199  }
200 
202  AvailableValue Res;
203  Res.Val.setPointer(nullptr);
204  Res.Val.setInt(UndefVal);
205  Res.Offset = 0;
206  return Res;
207  }
208 
209  bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
210  bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
211  bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
212  bool isUndefValue() const { return Val.getInt() == UndefVal; }
213 
215  assert(isSimpleValue() && "Wrong accessor");
216  return Val.getPointer();
217  }
218 
220  assert(isCoercedLoadValue() && "Wrong accessor");
221  return cast<LoadInst>(Val.getPointer());
222  }
223 
225  assert(isMemIntrinValue() && "Wrong accessor");
226  return cast<MemIntrinsic>(Val.getPointer());
227  }
228 
229  /// Emit code at the specified insertion point to adjust the value defined
230  /// here to the specified type. This handles various coercion cases.
231  Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
232  GVN &gvn) const;
233 };
234 
235 /// Represents an AvailableValue which can be rematerialized at the end of
236 /// the associated BasicBlock.
238  /// BB - The basic block in question.
240 
241  /// AV - The actual available value
243 
246  Res.BB = BB;
247  Res.AV = std::move(AV);
248  return Res;
249  }
250 
252  unsigned Offset = 0) {
253  return get(BB, AvailableValue::get(V, Offset));
254  }
255 
257  return get(BB, AvailableValue::getUndef());
258  }
259 
260  /// Emit code at the end of this block to adjust the value defined here to
261  /// the specified type. This handles various coercion cases.
263  return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
264  }
265 };
266 
267 //===----------------------------------------------------------------------===//
268 // ValueTable Internal Functions
269 //===----------------------------------------------------------------------===//
270 
271 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
272  Expression e;
273  e.type = I->getType();
274  e.opcode = I->getOpcode();
275  for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
276  OI != OE; ++OI)
277  e.varargs.push_back(lookupOrAdd(*OI));
278  if (I->isCommutative()) {
279  // Ensure that commutative instructions that only differ by a permutation
280  // of their operands get the same value number by sorting the operand value
281  // numbers. Since all commutative instructions have two operands it is more
282  // efficient to sort by hand rather than using, say, std::sort.
283  assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
284  if (e.varargs[0] > e.varargs[1])
285  std::swap(e.varargs[0], e.varargs[1]);
286  e.commutative = true;
287  }
288 
289  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
290  // Sort the operand value numbers so x<y and y>x get the same value number.
291  CmpInst::Predicate Predicate = C->getPredicate();
292  if (e.varargs[0] > e.varargs[1]) {
293  std::swap(e.varargs[0], e.varargs[1]);
294  Predicate = CmpInst::getSwappedPredicate(Predicate);
295  }
296  e.opcode = (C->getOpcode() << 8) | Predicate;
297  e.commutative = true;
298  } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
299  for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
300  II != IE; ++II)
301  e.varargs.push_back(*II);
302  }
303 
304  return e;
305 }
306 
307 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
309  Value *LHS, Value *RHS) {
310  assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
311  "Not a comparison!");
312  Expression e;
313  e.type = CmpInst::makeCmpResultType(LHS->getType());
314  e.varargs.push_back(lookupOrAdd(LHS));
315  e.varargs.push_back(lookupOrAdd(RHS));
316 
317  // Sort the operand value numbers so x<y and y>x get the same value number.
318  if (e.varargs[0] > e.varargs[1]) {
319  std::swap(e.varargs[0], e.varargs[1]);
320  Predicate = CmpInst::getSwappedPredicate(Predicate);
321  }
322  e.opcode = (Opcode << 8) | Predicate;
323  e.commutative = true;
324  return e;
325 }
326 
327 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
328  assert(EI && "Not an ExtractValueInst?");
329  Expression e;
330  e.type = EI->getType();
331  e.opcode = 0;
332 
334  if (WO != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
335  // EI is an extract from one of our with.overflow intrinsics. Synthesize
336  // a semantically equivalent expression instead of an extract value
337  // expression.
338  e.opcode = WO->getBinaryOp();
339  e.varargs.push_back(lookupOrAdd(WO->getLHS()));
340  e.varargs.push_back(lookupOrAdd(WO->getRHS()));
341  return e;
342  }
343 
344  // Not a recognised intrinsic. Fall back to producing an extract value
345  // expression.
346  e.opcode = EI->getOpcode();
347  for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
348  OI != OE; ++OI)
349  e.varargs.push_back(lookupOrAdd(*OI));
350 
351  for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
352  II != IE; ++II)
353  e.varargs.push_back(*II);
354 
355  return e;
356 }
357 
358 //===----------------------------------------------------------------------===//
359 // ValueTable External Functions
360 //===----------------------------------------------------------------------===//
361 
362 GVN::ValueTable::ValueTable() = default;
363 GVN::ValueTable::ValueTable(const ValueTable &) = default;
364 GVN::ValueTable::ValueTable(ValueTable &&) = default;
365 GVN::ValueTable::~ValueTable() = default;
366 
367 /// add - Insert a value into the table with a specified value number.
369  valueNumbering.insert(std::make_pair(V, num));
370  if (PHINode *PN = dyn_cast<PHINode>(V))
371  NumberingPhi[num] = PN;
372 }
373 
374 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
375  if (AA->doesNotAccessMemory(C)) {
376  Expression exp = createExpr(C);
377  uint32_t e = assignExpNewValueNum(exp).first;
378  valueNumbering[C] = e;
379  return e;
380  } else if (MD && AA->onlyReadsMemory(C)) {
381  Expression exp = createExpr(C);
382  auto ValNum = assignExpNewValueNum(exp);
383  if (ValNum.second) {
384  valueNumbering[C] = ValNum.first;
385  return ValNum.first;
386  }
387 
388  MemDepResult local_dep = MD->getDependency(C);
389 
390  if (!local_dep.isDef() && !local_dep.isNonLocal()) {
391  valueNumbering[C] = nextValueNumber;
392  return nextValueNumber++;
393  }
394 
395  if (local_dep.isDef()) {
396  CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
397 
398  if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
399  valueNumbering[C] = nextValueNumber;
400  return nextValueNumber++;
401  }
402 
403  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
404  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
405  uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
406  if (c_vn != cd_vn) {
407  valueNumbering[C] = nextValueNumber;
408  return nextValueNumber++;
409  }
410  }
411 
412  uint32_t v = lookupOrAdd(local_cdep);
413  valueNumbering[C] = v;
414  return v;
415  }
416 
417  // Non-local case.
419  MD->getNonLocalCallDependency(C);
420  // FIXME: Move the checking logic to MemDep!
421  CallInst* cdep = nullptr;
422 
423  // Check to see if we have a single dominating call instruction that is
424  // identical to C.
425  for (unsigned i = 0, e = deps.size(); i != e; ++i) {
426  const NonLocalDepEntry *I = &deps[i];
427  if (I->getResult().isNonLocal())
428  continue;
429 
430  // We don't handle non-definitions. If we already have a call, reject
431  // instruction dependencies.
432  if (!I->getResult().isDef() || cdep != nullptr) {
433  cdep = nullptr;
434  break;
435  }
436 
437  CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
438  // FIXME: All duplicated with non-local case.
439  if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
440  cdep = NonLocalDepCall;
441  continue;
442  }
443 
444  cdep = nullptr;
445  break;
446  }
447 
448  if (!cdep) {
449  valueNumbering[C] = nextValueNumber;
450  return nextValueNumber++;
451  }
452 
453  if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
454  valueNumbering[C] = nextValueNumber;
455  return nextValueNumber++;
456  }
457  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
458  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
459  uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
460  if (c_vn != cd_vn) {
461  valueNumbering[C] = nextValueNumber;
462  return nextValueNumber++;
463  }
464  }
465 
466  uint32_t v = lookupOrAdd(cdep);
467  valueNumbering[C] = v;
468  return v;
469  } else {
470  valueNumbering[C] = nextValueNumber;
471  return nextValueNumber++;
472  }
473 }
474 
475 /// Returns true if a value number exists for the specified value.
476 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
477 
478 /// lookup_or_add - Returns the value number for the specified value, assigning
479 /// it a new number if it did not have one before.
481  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
482  if (VI != valueNumbering.end())
483  return VI->second;
484 
485  if (!isa<Instruction>(V)) {
486  valueNumbering[V] = nextValueNumber;
487  return nextValueNumber++;
488  }
489 
490  Instruction* I = cast<Instruction>(V);
491  Expression exp;
492  switch (I->getOpcode()) {
493  case Instruction::Call:
494  return lookupOrAddCall(cast<CallInst>(I));
495  case Instruction::FNeg:
496  case Instruction::Add:
497  case Instruction::FAdd:
498  case Instruction::Sub:
499  case Instruction::FSub:
500  case Instruction::Mul:
501  case Instruction::FMul:
502  case Instruction::UDiv:
503  case Instruction::SDiv:
504  case Instruction::FDiv:
505  case Instruction::URem:
506  case Instruction::SRem:
507  case Instruction::FRem:
508  case Instruction::Shl:
509  case Instruction::LShr:
510  case Instruction::AShr:
511  case Instruction::And:
512  case Instruction::Or:
513  case Instruction::Xor:
514  case Instruction::ICmp:
515  case Instruction::FCmp:
516  case Instruction::Trunc:
517  case Instruction::ZExt:
518  case Instruction::SExt:
519  case Instruction::FPToUI:
520  case Instruction::FPToSI:
521  case Instruction::UIToFP:
522  case Instruction::SIToFP:
523  case Instruction::FPTrunc:
524  case Instruction::FPExt:
525  case Instruction::PtrToInt:
526  case Instruction::IntToPtr:
527  case Instruction::AddrSpaceCast:
528  case Instruction::BitCast:
529  case Instruction::Select:
530  case Instruction::ExtractElement:
531  case Instruction::InsertElement:
532  case Instruction::ShuffleVector:
533  case Instruction::InsertValue:
534  case Instruction::GetElementPtr:
535  exp = createExpr(I);
536  break;
537  case Instruction::ExtractValue:
538  exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
539  break;
540  case Instruction::PHI:
541  valueNumbering[V] = nextValueNumber;
542  NumberingPhi[nextValueNumber] = cast<PHINode>(V);
543  return nextValueNumber++;
544  default:
545  valueNumbering[V] = nextValueNumber;
546  return nextValueNumber++;
547  }
548 
549  uint32_t e = assignExpNewValueNum(exp).first;
550  valueNumbering[V] = e;
551  return e;
552 }
553 
554 /// Returns the value number of the specified value. Fails if
555 /// the value has not yet been numbered.
557  DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
558  if (Verify) {
559  assert(VI != valueNumbering.end() && "Value not numbered?");
560  return VI->second;
561  }
562  return (VI != valueNumbering.end()) ? VI->second : 0;
563 }
564 
565 /// Returns the value number of the given comparison,
566 /// assigning it a new number if it did not have one before. Useful when
567 /// we deduced the result of a comparison, but don't immediately have an
568 /// instruction realizing that comparison to hand.
570  CmpInst::Predicate Predicate,
571  Value *LHS, Value *RHS) {
572  Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
573  return assignExpNewValueNum(exp).first;
574 }
575 
576 /// Remove all entries from the ValueTable.
578  valueNumbering.clear();
579  expressionNumbering.clear();
580  NumberingPhi.clear();
581  PhiTranslateTable.clear();
582  nextValueNumber = 1;
583  Expressions.clear();
584  ExprIdx.clear();
585  nextExprNumber = 0;
586 }
587 
588 /// Remove a value from the value numbering.
590  uint32_t Num = valueNumbering.lookup(V);
591  valueNumbering.erase(V);
592  // If V is PHINode, V <--> value number is an one-to-one mapping.
593  if (isa<PHINode>(V))
594  NumberingPhi.erase(Num);
595 }
596 
597 /// verifyRemoved - Verify that the value is removed from all internal data
598 /// structures.
599 void GVN::ValueTable::verifyRemoved(const Value *V) const {
601  I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
602  assert(I->first != V && "Inst still occurs in value numbering map!");
603  }
604 }
605 
606 //===----------------------------------------------------------------------===//
607 // GVN Pass
608 //===----------------------------------------------------------------------===//
609 
611  // FIXME: The order of evaluation of these 'getResult' calls is very
612  // significant! Re-ordering these variables will cause GVN when run alone to
613  // be less effective! We should fix memdep and basic-aa to not exhibit this
614  // behavior, but until then don't change the order here.
615  auto &AC = AM.getResult<AssumptionAnalysis>(F);
616  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
617  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
618  auto &AA = AM.getResult<AAManager>(F);
619  auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
620  auto *LI = AM.getCachedResult<LoopAnalysis>(F);
622  bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
623  if (!Changed)
624  return PreservedAnalyses::all();
627  PA.preserve<GlobalsAA>();
629  return PA;
630 }
631 
632 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
633 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
634  errs() << "{\n";
636  E = d.end(); I != E; ++I) {
637  errs() << I->first << "\n";
638  I->second->dump();
639  }
640  errs() << "}\n";
641 }
642 #endif
643 
644 /// Return true if we can prove that the value
645 /// we're analyzing is fully available in the specified block. As we go, keep
646 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
647 /// map is actually a tri-state map with the following values:
648 /// 0) we know the block *is not* fully available.
649 /// 1) we know the block *is* fully available.
650 /// 2) we do not know whether the block is fully available or not, but we are
651 /// currently speculating that it will be.
652 /// 3) we are speculating for this block and have used that to speculate for
653 /// other blocks.
655  DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
656  uint32_t RecurseDepth) {
657  if (RecurseDepth > MaxRecurseDepth)
658  return false;
659 
660  // Optimistically assume that the block is fully available and check to see
661  // if we already know about this block in one lookup.
662  std::pair<DenseMap<BasicBlock*, char>::iterator, bool> IV =
663  FullyAvailableBlocks.insert(std::make_pair(BB, 2));
664 
665  // If the entry already existed for this block, return the precomputed value.
666  if (!IV.second) {
667  // If this is a speculative "available" value, mark it as being used for
668  // speculation of other blocks.
669  if (IV.first->second == 2)
670  IV.first->second = 3;
671  return IV.first->second != 0;
672  }
673 
674  // Otherwise, see if it is fully available in all predecessors.
675  pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
676 
677  // If this block has no predecessors, it isn't live-in here.
678  if (PI == PE)
679  goto SpeculationFailure;
680 
681  for (; PI != PE; ++PI)
682  // If the value isn't fully available in one of our predecessors, then it
683  // isn't fully available in this block either. Undo our previous
684  // optimistic assumption and bail out.
685  if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
686  goto SpeculationFailure;
687 
688  return true;
689 
690 // If we get here, we found out that this is not, after
691 // all, a fully-available block. We have a problem if we speculated on this and
692 // used the speculation to mark other blocks as available.
693 SpeculationFailure:
694  char &BBVal = FullyAvailableBlocks[BB];
695 
696  // If we didn't speculate on this, just return with it set to false.
697  if (BBVal == 2) {
698  BBVal = 0;
699  return false;
700  }
701 
702  // If we did speculate on this value, we could have blocks set to 1 that are
703  // incorrect. Walk the (transitive) successors of this block and mark them as
704  // 0 if set to one.
705  SmallVector<BasicBlock*, 32> BBWorklist;
706  BBWorklist.push_back(BB);
707 
708  do {
709  BasicBlock *Entry = BBWorklist.pop_back_val();
710  // Note that this sets blocks to 0 (unavailable) if they happen to not
711  // already be in FullyAvailableBlocks. This is safe.
712  char &EntryVal = FullyAvailableBlocks[Entry];
713  if (EntryVal == 0) continue; // Already unavailable.
714 
715  // Mark as unavailable.
716  EntryVal = 0;
717 
718  BBWorklist.append(succ_begin(Entry), succ_end(Entry));
719  } while (!BBWorklist.empty());
720 
721  return false;
722 }
723 
724 /// Given a set of loads specified by ValuesPerBlock,
725 /// construct SSA form, allowing us to eliminate LI. This returns the value
726 /// that should be used at LI's definition site.
729  GVN &gvn) {
730  // Check for the fully redundant, dominating load case. In this case, we can
731  // just use the dominating value directly.
732  if (ValuesPerBlock.size() == 1 &&
733  gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
734  LI->getParent())) {
735  assert(!ValuesPerBlock[0].AV.isUndefValue() &&
736  "Dead BB dominate this block");
737  return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
738  }
739 
740  // Otherwise, we have to construct SSA form.
741  SmallVector<PHINode*, 8> NewPHIs;
742  SSAUpdater SSAUpdate(&NewPHIs);
743  SSAUpdate.Initialize(LI->getType(), LI->getName());
744 
745  for (const AvailableValueInBlock &AV : ValuesPerBlock) {
746  BasicBlock *BB = AV.BB;
747 
748  if (SSAUpdate.HasValueForBlock(BB))
749  continue;
750 
751  // If the value is the load that we will be eliminating, and the block it's
752  // available in is the block that the load is in, then don't add it as
753  // SSAUpdater will resolve the value to the relevant phi which may let it
754  // avoid phi construction entirely if there's actually only one value.
755  if (BB == LI->getParent() &&
756  ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == LI) ||
757  (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == LI)))
758  continue;
759 
760  SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
761  }
762 
763  // Perform PHI construction.
764  return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
765 }
766 
768  Instruction *InsertPt,
769  GVN &gvn) const {
770  Value *Res;
771  Type *LoadTy = LI->getType();
772  const DataLayout &DL = LI->getModule()->getDataLayout();
773  if (isSimpleValue()) {
774  Res = getSimpleValue();
775  if (Res->getType() != LoadTy) {
776  Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
777 
778  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset
779  << " " << *getSimpleValue() << '\n'
780  << *Res << '\n'
781  << "\n\n\n");
782  }
783  } else if (isCoercedLoadValue()) {
784  LoadInst *Load = getCoercedLoadValue();
785  if (Load->getType() == LoadTy && Offset == 0) {
786  Res = Load;
787  } else {
788  Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
789  // We would like to use gvn.markInstructionForDeletion here, but we can't
790  // because the load is already memoized into the leader map table that GVN
791  // tracks. It is potentially possible to remove the load from the table,
792  // but then there all of the operations based on it would need to be
793  // rehashed. Just leave the dead load around.
794  gvn.getMemDep().removeInstruction(Load);
795  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset
796  << " " << *getCoercedLoadValue() << '\n'
797  << *Res << '\n'
798  << "\n\n\n");
799  }
800  } else if (isMemIntrinValue()) {
801  Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
802  InsertPt, DL);
803  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
804  << " " << *getMemIntrinValue() << '\n'
805  << *Res << '\n'
806  << "\n\n\n");
807  } else {
808  assert(isUndefValue() && "Should be UndefVal");
809  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
810  return UndefValue::get(LoadTy);
811  }
812  assert(Res && "failed to materialize?");
813  return Res;
814 }
815 
816 static bool isLifetimeStart(const Instruction *Inst) {
817  if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
818  return II->getIntrinsicID() == Intrinsic::lifetime_start;
819  return false;
820 }
821 
822 /// Try to locate the three instruction involved in a missed
823 /// load-elimination case that is due to an intervening store.
825  DominatorTree *DT,
827  using namespace ore;
828 
829  User *OtherAccess = nullptr;
830 
831  OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
832  R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
833  << setExtraArgs();
834 
835  for (auto *U : LI->getPointerOperand()->users())
836  if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
837  DT->dominates(cast<Instruction>(U), LI)) {
838  // FIXME: for now give up if there are multiple memory accesses that
839  // dominate the load. We need further analysis to decide which one is
840  // that we're forwarding from.
841  if (OtherAccess)
842  OtherAccess = nullptr;
843  else
844  OtherAccess = U;
845  }
846 
847  if (OtherAccess)
848  R << " in favor of " << NV("OtherAccess", OtherAccess);
849 
850  R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
851 
852  ORE->emit(R);
853 }
854 
855 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
856  Value *Address, AvailableValue &Res) {
857  assert((DepInfo.isDef() || DepInfo.isClobber()) &&
858  "expected a local dependence");
859  assert(LI->isUnordered() && "rules below are incorrect for ordered access");
860 
861  const DataLayout &DL = LI->getModule()->getDataLayout();
862 
863  Instruction *DepInst = DepInfo.getInst();
864  if (DepInfo.isClobber()) {
865  // If the dependence is to a store that writes to a superset of the bits
866  // read by the load, we can extract the bits we need for the load from the
867  // stored value.
868  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
869  // Can't forward from non-atomic to atomic without violating memory model.
870  if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
871  int Offset =
873  if (Offset != -1) {
874  Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
875  return true;
876  }
877  }
878  }
879 
880  // Check to see if we have something like this:
881  // load i32* P
882  // load i8* (P+1)
883  // if we have this, replace the later with an extraction from the former.
884  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
885  // If this is a clobber and L is the first instruction in its block, then
886  // we have the first instruction in the entry block.
887  // Can't forward from non-atomic to atomic without violating memory model.
888  if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
889  int Offset =
890  analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
891 
892  if (Offset != -1) {
893  Res = AvailableValue::getLoad(DepLI, Offset);
894  return true;
895  }
896  }
897  }
898 
899  // If the clobbering value is a memset/memcpy/memmove, see if we can
900  // forward a value on from it.
901  if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInst)) {
902  if (Address && !LI->isAtomic()) {
904  DepMI, DL);
905  if (Offset != -1) {
906  Res = AvailableValue::getMI(DepMI, Offset);
907  return true;
908  }
909  }
910  }
911  // Nothing known about this clobber, have to be conservative
912  LLVM_DEBUG(
913  // fast print dep, using operator<< on instruction is too slow.
914  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
915  dbgs() << " is clobbered by " << *DepInst << '\n';);
916  if (ORE->allowExtraAnalysis(DEBUG_TYPE))
917  reportMayClobberedLoad(LI, DepInfo, DT, ORE);
918 
919  return false;
920  }
921  assert(DepInfo.isDef() && "follows from above");
922 
923  // Loading the allocation -> undef.
924  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
925  // Loading immediately after lifetime begin -> undef.
926  isLifetimeStart(DepInst)) {
928  return true;
929  }
930 
931  // Loading from calloc (which zero initializes memory) -> zero
932  if (isCallocLikeFn(DepInst, TLI)) {
934  return true;
935  }
936 
937  if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
938  // Reject loads and stores that are to the same address but are of
939  // different types if we have to. If the stored value is larger or equal to
940  // the loaded value, we can reuse it.
941  if (!canCoerceMustAliasedValueToLoad(S->getValueOperand(), LI->getType(),
942  DL))
943  return false;
944 
945  // Can't forward from non-atomic to atomic without violating memory model.
946  if (S->isAtomic() < LI->isAtomic())
947  return false;
948 
949  Res = AvailableValue::get(S->getValueOperand());
950  return true;
951  }
952 
953  if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
954  // If the types mismatch and we can't handle it, reject reuse of the load.
955  // If the stored value is larger or equal to the loaded value, we can reuse
956  // it.
957  if (!canCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
958  return false;
959 
960  // Can't forward from non-atomic to atomic without violating memory model.
961  if (LD->isAtomic() < LI->isAtomic())
962  return false;
963 
965  return true;
966  }
967 
968  // Unknown def - must be conservative
969  LLVM_DEBUG(
970  // fast print dep, using operator<< on instruction is too slow.
971  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
972  dbgs() << " has unknown def " << *DepInst << '\n';);
973  return false;
974 }
975 
976 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
977  AvailValInBlkVect &ValuesPerBlock,
978  UnavailBlkVect &UnavailableBlocks) {
979  // Filter out useless results (non-locals, etc). Keep track of the blocks
980  // where we have a value available in repl, also keep track of whether we see
981  // dependencies that produce an unknown value for the load (such as a call
982  // that could potentially clobber the load).
983  unsigned NumDeps = Deps.size();
984  for (unsigned i = 0, e = NumDeps; i != e; ++i) {
985  BasicBlock *DepBB = Deps[i].getBB();
986  MemDepResult DepInfo = Deps[i].getResult();
987 
988  if (DeadBlocks.count(DepBB)) {
989  // Dead dependent mem-op disguise as a load evaluating the same value
990  // as the load in question.
991  ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
992  continue;
993  }
994 
995  if (!DepInfo.isDef() && !DepInfo.isClobber()) {
996  UnavailableBlocks.push_back(DepBB);
997  continue;
998  }
999 
1000  // The address being loaded in this non-local block may not be the same as
1001  // the pointer operand of the load if PHI translation occurs. Make sure
1002  // to consider the right address.
1003  Value *Address = Deps[i].getAddress();
1004 
1005  AvailableValue AV;
1006  if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1007  // subtlety: because we know this was a non-local dependency, we know
1008  // it's safe to materialize anywhere between the instruction within
1009  // DepInfo and the end of it's block.
1010  ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1011  std::move(AV)));
1012  } else {
1013  UnavailableBlocks.push_back(DepBB);
1014  }
1015  }
1016 
1017  assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1018  "post condition violation");
1019 }
1020 
1021 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1022  UnavailBlkVect &UnavailableBlocks) {
1023  // Okay, we have *some* definitions of the value. This means that the value
1024  // is available in some of our (transitive) predecessors. Lets think about
1025  // doing PRE of this load. This will involve inserting a new load into the
1026  // predecessor when it's not available. We could do this in general, but
1027  // prefer to not increase code size. As such, we only do this when we know
1028  // that we only have to insert *one* load (which means we're basically moving
1029  // the load, not inserting a new one).
1030 
1031  SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1032  UnavailableBlocks.end());
1033 
1034  // Let's find the first basic block with more than one predecessor. Walk
1035  // backwards through predecessors if needed.
1036  BasicBlock *LoadBB = LI->getParent();
1037  BasicBlock *TmpBB = LoadBB;
1038  bool IsSafeToSpeculativelyExecute = isSafeToSpeculativelyExecute(LI);
1039 
1040  // Check that there is no implicit control flow instructions above our load in
1041  // its block. If there is an instruction that doesn't always pass the
1042  // execution to the following instruction, then moving through it may become
1043  // invalid. For example:
1044  //
1045  // int arr[LEN];
1046  // int index = ???;
1047  // ...
1048  // guard(0 <= index && index < LEN);
1049  // use(arr[index]);
1050  //
1051  // It is illegal to move the array access to any point above the guard,
1052  // because if the index is out of bounds we should deoptimize rather than
1053  // access the array.
1054  // Check that there is no guard in this block above our instruction.
1055  if (!IsSafeToSpeculativelyExecute && ICF->isDominatedByICFIFromSameBlock(LI))
1056  return false;
1057  while (TmpBB->getSinglePredecessor()) {
1058  TmpBB = TmpBB->getSinglePredecessor();
1059  if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1060  return false;
1061  if (Blockers.count(TmpBB))
1062  return false;
1063 
1064  // If any of these blocks has more than one successor (i.e. if the edge we
1065  // just traversed was critical), then there are other paths through this
1066  // block along which the load may not be anticipated. Hoisting the load
1067  // above this block would be adding the load to execution paths along
1068  // which it was not previously executed.
1069  if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1070  return false;
1071 
1072  // Check that there is no implicit control flow in a block above.
1073  if (!IsSafeToSpeculativelyExecute && ICF->hasICF(TmpBB))
1074  return false;
1075  }
1076 
1077  assert(TmpBB);
1078  LoadBB = TmpBB;
1079 
1080  // Check to see how many predecessors have the loaded value fully
1081  // available.
1083  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1084  for (const AvailableValueInBlock &AV : ValuesPerBlock)
1085  FullyAvailableBlocks[AV.BB] = true;
1086  for (BasicBlock *UnavailableBB : UnavailableBlocks)
1087  FullyAvailableBlocks[UnavailableBB] = false;
1088 
1089  SmallVector<BasicBlock *, 4> CriticalEdgePred;
1090  for (BasicBlock *Pred : predecessors(LoadBB)) {
1091  // If any predecessor block is an EH pad that does not allow non-PHI
1092  // instructions before the terminator, we can't PRE the load.
1093  if (Pred->getTerminator()->isEHPad()) {
1094  LLVM_DEBUG(
1095  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1096  << Pred->getName() << "': " << *LI << '\n');
1097  return false;
1098  }
1099 
1100  if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1101  continue;
1102  }
1103 
1104  if (Pred->getTerminator()->getNumSuccessors() != 1) {
1105  if (isa<IndirectBrInst>(Pred->getTerminator())) {
1106  LLVM_DEBUG(
1107  dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1108  << Pred->getName() << "': " << *LI << '\n');
1109  return false;
1110  }
1111 
1112  // FIXME: Can we support the fallthrough edge?
1113  if (isa<CallBrInst>(Pred->getTerminator())) {
1114  LLVM_DEBUG(
1115  dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '"
1116  << Pred->getName() << "': " << *LI << '\n');
1117  return false;
1118  }
1119 
1120  if (LoadBB->isEHPad()) {
1121  LLVM_DEBUG(
1122  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1123  << Pred->getName() << "': " << *LI << '\n');
1124  return false;
1125  }
1126 
1127  CriticalEdgePred.push_back(Pred);
1128  } else {
1129  // Only add the predecessors that will not be split for now.
1130  PredLoads[Pred] = nullptr;
1131  }
1132  }
1133 
1134  // Decide whether PRE is profitable for this load.
1135  unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1136  assert(NumUnavailablePreds != 0 &&
1137  "Fully available value should already be eliminated!");
1138 
1139  // If this load is unavailable in multiple predecessors, reject it.
1140  // FIXME: If we could restructure the CFG, we could make a common pred with
1141  // all the preds that don't have an available LI and insert a new load into
1142  // that one block.
1143  if (NumUnavailablePreds != 1)
1144  return false;
1145 
1146  // Split critical edges, and update the unavailable predecessors accordingly.
1147  for (BasicBlock *OrigPred : CriticalEdgePred) {
1148  BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1149  assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1150  PredLoads[NewPred] = nullptr;
1151  LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1152  << LoadBB->getName() << '\n');
1153  }
1154 
1155  // Check if the load can safely be moved to all the unavailable predecessors.
1156  bool CanDoPRE = true;
1157  const DataLayout &DL = LI->getModule()->getDataLayout();
1159  for (auto &PredLoad : PredLoads) {
1160  BasicBlock *UnavailablePred = PredLoad.first;
1161 
1162  // Do PHI translation to get its value in the predecessor if necessary. The
1163  // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1164 
1165  // If all preds have a single successor, then we know it is safe to insert
1166  // the load on the pred (?!?), so we can insert code to materialize the
1167  // pointer if it is not available.
1168  PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1169  Value *LoadPtr = nullptr;
1170  LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1171  *DT, NewInsts);
1172 
1173  // If we couldn't find or insert a computation of this phi translated value,
1174  // we fail PRE.
1175  if (!LoadPtr) {
1176  LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1177  << *LI->getPointerOperand() << "\n");
1178  CanDoPRE = false;
1179  break;
1180  }
1181 
1182  PredLoad.second = LoadPtr;
1183  }
1184 
1185  if (!CanDoPRE) {
1186  while (!NewInsts.empty()) {
1187  Instruction *I = NewInsts.pop_back_val();
1188  markInstructionForDeletion(I);
1189  }
1190  // HINT: Don't revert the edge-splitting as following transformation may
1191  // also need to split these critical edges.
1192  return !CriticalEdgePred.empty();
1193  }
1194 
1195  // Okay, we can eliminate this load by inserting a reload in the predecessor
1196  // and using PHI construction to get the value in the other predecessors, do
1197  // it.
1198  LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1199  LLVM_DEBUG(if (!NewInsts.empty()) dbgs()
1200  << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back()
1201  << '\n');
1202 
1203  // Assign value numbers to the new instructions.
1204  for (Instruction *I : NewInsts) {
1205  // Instructions that have been inserted in predecessor(s) to materialize
1206  // the load address do not retain their original debug locations. Doing
1207  // so could lead to confusing (but correct) source attributions.
1208  if (const DebugLoc &DL = I->getDebugLoc())
1209  I->setDebugLoc(DebugLoc::get(0, 0, DL.getScope(), DL.getInlinedAt()));
1210 
1211  // FIXME: We really _ought_ to insert these value numbers into their
1212  // parent's availability map. However, in doing so, we risk getting into
1213  // ordering issues. If a block hasn't been processed yet, we would be
1214  // marking a value as AVAIL-IN, which isn't what we intend.
1215  VN.lookupOrAdd(I);
1216  }
1217 
1218  for (const auto &PredLoad : PredLoads) {
1219  BasicBlock *UnavailablePred = PredLoad.first;
1220  Value *LoadPtr = PredLoad.second;
1221 
1222  auto *NewLoad =
1223  new LoadInst(LI->getType(), LoadPtr, LI->getName() + ".pre",
1224  LI->isVolatile(), LI->getAlignment(), LI->getOrdering(),
1225  LI->getSyncScopeID(), UnavailablePred->getTerminator());
1226  NewLoad->setDebugLoc(LI->getDebugLoc());
1227 
1228  // Transfer the old load's AA tags to the new load.
1229  AAMDNodes Tags;
1230  LI->getAAMetadata(Tags);
1231  if (Tags)
1232  NewLoad->setAAMetadata(Tags);
1233 
1234  if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1235  NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1236  if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1237  NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1238  if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1239  NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1240 
1241  // We do not propagate the old load's debug location, because the new
1242  // load now lives in a different BB, and we want to avoid a jumpy line
1243  // table.
1244  // FIXME: How do we retain source locations without causing poor debugging
1245  // behavior?
1246 
1247  // Add the newly created load.
1248  ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1249  NewLoad));
1250  MD->invalidateCachedPointerInfo(LoadPtr);
1251  LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1252  }
1253 
1254  // Perform PHI construction.
1255  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1256  LI->replaceAllUsesWith(V);
1257  if (isa<PHINode>(V))
1258  V->takeName(LI);
1259  if (Instruction *I = dyn_cast<Instruction>(V))
1260  I->setDebugLoc(LI->getDebugLoc());
1261  if (V->getType()->isPtrOrPtrVectorTy())
1262  MD->invalidateCachedPointerInfo(V);
1263  markInstructionForDeletion(LI);
1264  ORE->emit([&]() {
1265  return OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1266  << "load eliminated by PRE";
1267  });
1268  ++NumPRELoad;
1269  return true;
1270 }
1271 
1274  using namespace ore;
1275 
1276  ORE->emit([&]() {
1277  return OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1278  << "load of type " << NV("Type", LI->getType()) << " eliminated"
1279  << setExtraArgs() << " in favor of "
1280  << NV("InfavorOfValue", AvailableValue);
1281  });
1282 }
1283 
1284 /// Attempt to eliminate a load whose dependencies are
1285 /// non-local by performing PHI construction.
1286 bool GVN::processNonLocalLoad(LoadInst *LI) {
1287  // non-local speculations are not allowed under asan.
1288  if (LI->getParent()->getParent()->hasFnAttribute(
1289  Attribute::SanitizeAddress) ||
1291  Attribute::SanitizeHWAddress))
1292  return false;
1293 
1294  // Step 1: Find the non-local dependencies of the load.
1295  LoadDepVect Deps;
1296  MD->getNonLocalPointerDependency(LI, Deps);
1297 
1298  // If we had to process more than one hundred blocks to find the
1299  // dependencies, this load isn't worth worrying about. Optimizing
1300  // it will be too expensive.
1301  unsigned NumDeps = Deps.size();
1302  if (NumDeps > MaxNumDeps)
1303  return false;
1304 
1305  // If we had a phi translation failure, we'll have a single entry which is a
1306  // clobber in the current block. Reject this early.
1307  if (NumDeps == 1 &&
1308  !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1309  LLVM_DEBUG(dbgs() << "GVN: non-local load "; LI->printAsOperand(dbgs());
1310  dbgs() << " has unknown dependencies\n";);
1311  return false;
1312  }
1313 
1314  // If this load follows a GEP, see if we can PRE the indices before analyzing.
1315  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1316  for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1317  OE = GEP->idx_end();
1318  OI != OE; ++OI)
1319  if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1320  performScalarPRE(I);
1321  }
1322 
1323  // Step 2: Analyze the availability of the load
1324  AvailValInBlkVect ValuesPerBlock;
1325  UnavailBlkVect UnavailableBlocks;
1326  AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1327 
1328  // If we have no predecessors that produce a known value for this load, exit
1329  // early.
1330  if (ValuesPerBlock.empty())
1331  return false;
1332 
1333  // Step 3: Eliminate fully redundancy.
1334  //
1335  // If all of the instructions we depend on produce a known value for this
1336  // load, then it is fully redundant and we can use PHI insertion to compute
1337  // its value. Insert PHIs and remove the fully redundant value now.
1338  if (UnavailableBlocks.empty()) {
1339  LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1340 
1341  // Perform PHI construction.
1342  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1343  LI->replaceAllUsesWith(V);
1344 
1345  if (isa<PHINode>(V))
1346  V->takeName(LI);
1347  if (Instruction *I = dyn_cast<Instruction>(V))
1348  // If instruction I has debug info, then we should not update it.
1349  // Also, if I has a null DebugLoc, then it is still potentially incorrect
1350  // to propagate LI's DebugLoc because LI may not post-dominate I.
1351  if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1352  I->setDebugLoc(LI->getDebugLoc());
1353  if (V->getType()->isPtrOrPtrVectorTy())
1354  MD->invalidateCachedPointerInfo(V);
1355  markInstructionForDeletion(LI);
1356  ++NumGVNLoad;
1357  reportLoadElim(LI, V, ORE);
1358  return true;
1359  }
1360 
1361  // Step 4: Eliminate partial redundancy.
1362  if (!EnablePRE || !EnableLoadPRE)
1363  return false;
1364 
1365  return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1366 }
1367 
1368 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1369  assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1370  "This function can only be called with llvm.assume intrinsic");
1371  Value *V = IntrinsicI->getArgOperand(0);
1372 
1373  if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1374  if (Cond->isZero()) {
1375  Type *Int8Ty = Type::getInt8Ty(V->getContext());
1376  // Insert a new store to null instruction before the load to indicate that
1377  // this code is not reachable. FIXME: We could insert unreachable
1378  // instruction directly because we can modify the CFG.
1379  new StoreInst(UndefValue::get(Int8Ty),
1381  IntrinsicI);
1382  }
1383  markInstructionForDeletion(IntrinsicI);
1384  return false;
1385  } else if (isa<Constant>(V)) {
1386  // If it's not false, and constant, it must evaluate to true. This means our
1387  // assume is assume(true), and thus, pointless, and we don't want to do
1388  // anything more here.
1389  return false;
1390  }
1391 
1392  Constant *True = ConstantInt::getTrue(V->getContext());
1393  bool Changed = false;
1394 
1395  for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1396  BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1397 
1398  // This property is only true in dominated successors, propagateEquality
1399  // will check dominance for us.
1400  Changed |= propagateEquality(V, True, Edge, false);
1401  }
1402 
1403  // We can replace assume value with true, which covers cases like this:
1404  // call void @llvm.assume(i1 %cmp)
1405  // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1406  ReplaceWithConstMap[V] = True;
1407 
1408  // If one of *cmp *eq operand is const, adding it to map will cover this:
1409  // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1410  // call void @llvm.assume(i1 %cmp)
1411  // ret float %0 ; will change it to ret float 3.000000e+00
1412  if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1413  if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1414  CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1415  (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1416  CmpI->getFastMathFlags().noNaNs())) {
1417  Value *CmpLHS = CmpI->getOperand(0);
1418  Value *CmpRHS = CmpI->getOperand(1);
1419  if (isa<Constant>(CmpLHS))
1420  std::swap(CmpLHS, CmpRHS);
1421  auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1422 
1423  // If only one operand is constant.
1424  if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1425  ReplaceWithConstMap[CmpLHS] = RHSConst;
1426  }
1427  }
1428  return Changed;
1429 }
1430 
1432  patchReplacementInstruction(I, Repl);
1433  I->replaceAllUsesWith(Repl);
1434 }
1435 
1436 /// Attempt to eliminate a load, first by eliminating it
1437 /// locally, and then attempting non-local elimination if that fails.
1438 bool GVN::processLoad(LoadInst *L) {
1439  if (!MD)
1440  return false;
1441 
1442  // This code hasn't been audited for ordered or volatile memory access
1443  if (!L->isUnordered())
1444  return false;
1445 
1446  if (L->use_empty()) {
1447  markInstructionForDeletion(L);
1448  return true;
1449  }
1450 
1451  // ... to a pointer that has been loaded from before...
1452  MemDepResult Dep = MD->getDependency(L);
1453 
1454  // If it is defined in another block, try harder.
1455  if (Dep.isNonLocal())
1456  return processNonLocalLoad(L);
1457 
1458  // Only handle the local case below
1459  if (!Dep.isDef() && !Dep.isClobber()) {
1460  // This might be a NonFuncLocal or an Unknown
1461  LLVM_DEBUG(
1462  // fast print dep, using operator<< on instruction is too slow.
1463  dbgs() << "GVN: load "; L->printAsOperand(dbgs());
1464  dbgs() << " has unknown dependence\n";);
1465  return false;
1466  }
1467 
1468  AvailableValue AV;
1469  if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1470  Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1471 
1472  // Replace the load!
1473  patchAndReplaceAllUsesWith(L, AvailableValue);
1474  markInstructionForDeletion(L);
1475  ++NumGVNLoad;
1476  reportLoadElim(L, AvailableValue, ORE);
1477  // Tell MDA to rexamine the reused pointer since we might have more
1478  // information after forwarding it.
1479  if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1480  MD->invalidateCachedPointerInfo(AvailableValue);
1481  return true;
1482  }
1483 
1484  return false;
1485 }
1486 
1487 /// Return a pair the first field showing the value number of \p Exp and the
1488 /// second field showing whether it is a value number newly created.
1489 std::pair<uint32_t, bool>
1490 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1491  uint32_t &e = expressionNumbering[Exp];
1492  bool CreateNewValNum = !e;
1493  if (CreateNewValNum) {
1494  Expressions.push_back(Exp);
1495  if (ExprIdx.size() < nextValueNumber + 1)
1496  ExprIdx.resize(nextValueNumber * 2);
1497  e = nextValueNumber;
1498  ExprIdx[nextValueNumber++] = nextExprNumber++;
1499  }
1500  return {e, CreateNewValNum};
1501 }
1502 
1503 /// Return whether all the values related with the same \p num are
1504 /// defined in \p BB.
1505 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1506  GVN &Gvn) {
1507  LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1508  while (Vals && Vals->BB == BB)
1509  Vals = Vals->Next;
1510  return !Vals;
1511 }
1512 
1513 /// Wrap phiTranslateImpl to provide caching functionality.
1515  const BasicBlock *PhiBlock, uint32_t Num,
1516  GVN &Gvn) {
1517  auto FindRes = PhiTranslateTable.find({Num, Pred});
1518  if (FindRes != PhiTranslateTable.end())
1519  return FindRes->second;
1520  uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1521  PhiTranslateTable.insert({{Num, Pred}, NewNum});
1522  return NewNum;
1523 }
1524 
1525 /// Translate value number \p Num using phis, so that it has the values of
1526 /// the phis in BB.
1527 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1528  const BasicBlock *PhiBlock,
1529  uint32_t Num, GVN &Gvn) {
1530  if (PHINode *PN = NumberingPhi[Num]) {
1531  for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1532  if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1533  if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1534  return TransVal;
1535  }
1536  return Num;
1537  }
1538 
1539  // If there is any value related with Num is defined in a BB other than
1540  // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1541  // a backedge. We can do an early exit in that case to save compile time.
1542  if (!areAllValsInBB(Num, PhiBlock, Gvn))
1543  return Num;
1544 
1545  if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1546  return Num;
1547  Expression Exp = Expressions[ExprIdx[Num]];
1548 
1549  for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1550  // For InsertValue and ExtractValue, some varargs are index numbers
1551  // instead of value numbers. Those index numbers should not be
1552  // translated.
1553  if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1554  (i > 0 && Exp.opcode == Instruction::ExtractValue))
1555  continue;
1556  Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1557  }
1558 
1559  if (Exp.commutative) {
1560  assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1561  if (Exp.varargs[0] > Exp.varargs[1]) {
1562  std::swap(Exp.varargs[0], Exp.varargs[1]);
1563  uint32_t Opcode = Exp.opcode >> 8;
1564  if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1565  Exp.opcode = (Opcode << 8) |
1567  static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1568  }
1569  }
1570 
1571  if (uint32_t NewNum = expressionNumbering[Exp])
1572  return NewNum;
1573  return Num;
1574 }
1575 
1576 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1577 /// again.
1579  const BasicBlock &CurrBlock) {
1580  for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1581  auto FindRes = PhiTranslateTable.find({Num, Pred});
1582  if (FindRes != PhiTranslateTable.end())
1583  PhiTranslateTable.erase(FindRes);
1584  }
1585 }
1586 
1587 // In order to find a leader for a given value number at a
1588 // specific basic block, we first obtain the list of all Values for that number,
1589 // and then scan the list to find one whose block dominates the block in
1590 // question. This is fast because dominator tree queries consist of only
1591 // a few comparisons of DFS numbers.
1592 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1593  LeaderTableEntry Vals = LeaderTable[num];
1594  if (!Vals.Val) return nullptr;
1595 
1596  Value *Val = nullptr;
1597  if (DT->dominates(Vals.BB, BB)) {
1598  Val = Vals.Val;
1599  if (isa<Constant>(Val)) return Val;
1600  }
1601 
1602  LeaderTableEntry* Next = Vals.Next;
1603  while (Next) {
1604  if (DT->dominates(Next->BB, BB)) {
1605  if (isa<Constant>(Next->Val)) return Next->Val;
1606  if (!Val) Val = Next->Val;
1607  }
1608 
1609  Next = Next->Next;
1610  }
1611 
1612  return Val;
1613 }
1614 
1615 /// There is an edge from 'Src' to 'Dst'. Return
1616 /// true if every path from the entry block to 'Dst' passes via this edge. In
1617 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1619  DominatorTree *DT) {
1620  // While in theory it is interesting to consider the case in which Dst has
1621  // more than one predecessor, because Dst might be part of a loop which is
1622  // only reachable from Src, in practice it is pointless since at the time
1623  // GVN runs all such loops have preheaders, which means that Dst will have
1624  // been changed to have only one predecessor, namely Src.
1625  const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1626  assert((!Pred || Pred == E.getStart()) &&
1627  "No edge between these basic blocks!");
1628  return Pred != nullptr;
1629 }
1630 
1631 void GVN::assignBlockRPONumber(Function &F) {
1632  BlockRPONumber.clear();
1633  uint32_t NextBlockNumber = 1;
1635  for (BasicBlock *BB : RPOT)
1636  BlockRPONumber[BB] = NextBlockNumber++;
1637  InvalidBlockRPONumbers = false;
1638 }
1639 
1640 // Tries to replace instruction with const, using information from
1641 // ReplaceWithConstMap.
1642 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1643  bool Changed = false;
1644  for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1645  Value *Operand = Instr->getOperand(OpNum);
1646  auto it = ReplaceWithConstMap.find(Operand);
1647  if (it != ReplaceWithConstMap.end()) {
1648  assert(!isa<Constant>(Operand) &&
1649  "Replacing constants with constants is invalid");
1650  LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "
1651  << *it->second << " in instruction " << *Instr << '\n');
1652  Instr->setOperand(OpNum, it->second);
1653  Changed = true;
1654  }
1655  }
1656  return Changed;
1657 }
1658 
1659 /// The given values are known to be equal in every block
1660 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1661 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1662 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1663 /// value starting from the end of Root.Start.
1664 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1665  bool DominatesByEdge) {
1667  Worklist.push_back(std::make_pair(LHS, RHS));
1668  bool Changed = false;
1669  // For speed, compute a conservative fast approximation to
1670  // DT->dominates(Root, Root.getEnd());
1671  const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1672 
1673  while (!Worklist.empty()) {
1674  std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1675  LHS = Item.first; RHS = Item.second;
1676 
1677  if (LHS == RHS)
1678  continue;
1679  assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1680 
1681  // Don't try to propagate equalities between constants.
1682  if (isa<Constant>(LHS) && isa<Constant>(RHS))
1683  continue;
1684 
1685  // Prefer a constant on the right-hand side, or an Argument if no constants.
1686  if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1687  std::swap(LHS, RHS);
1688  assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1689 
1690  // If there is no obvious reason to prefer the left-hand side over the
1691  // right-hand side, ensure the longest lived term is on the right-hand side,
1692  // so the shortest lived term will be replaced by the longest lived.
1693  // This tends to expose more simplifications.
1694  uint32_t LVN = VN.lookupOrAdd(LHS);
1695  if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1696  (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1697  // Move the 'oldest' value to the right-hand side, using the value number
1698  // as a proxy for age.
1699  uint32_t RVN = VN.lookupOrAdd(RHS);
1700  if (LVN < RVN) {
1701  std::swap(LHS, RHS);
1702  LVN = RVN;
1703  }
1704  }
1705 
1706  // If value numbering later sees that an instruction in the scope is equal
1707  // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1708  // the invariant that instructions only occur in the leader table for their
1709  // own value number (this is used by removeFromLeaderTable), do not do this
1710  // if RHS is an instruction (if an instruction in the scope is morphed into
1711  // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1712  // using the leader table is about compiling faster, not optimizing better).
1713  // The leader table only tracks basic blocks, not edges. Only add to if we
1714  // have the simple case where the edge dominates the end.
1715  if (RootDominatesEnd && !isa<Instruction>(RHS))
1716  addToLeaderTable(LVN, RHS, Root.getEnd());
1717 
1718  // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1719  // LHS always has at least one use that is not dominated by Root, this will
1720  // never do anything if LHS has only one use.
1721  if (!LHS->hasOneUse()) {
1722  unsigned NumReplacements =
1723  DominatesByEdge
1724  ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1725  : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1726 
1727  Changed |= NumReplacements > 0;
1728  NumGVNEqProp += NumReplacements;
1729  // Cached information for anything that uses LHS will be invalid.
1730  if (MD)
1731  MD->invalidateCachedPointerInfo(LHS);
1732  }
1733 
1734  // Now try to deduce additional equalities from this one. For example, if
1735  // the known equality was "(A != B)" == "false" then it follows that A and B
1736  // are equal in the scope. Only boolean equalities with an explicit true or
1737  // false RHS are currently supported.
1738  if (!RHS->getType()->isIntegerTy(1))
1739  // Not a boolean equality - bail out.
1740  continue;
1741  ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1742  if (!CI)
1743  // RHS neither 'true' nor 'false' - bail out.
1744  continue;
1745  // Whether RHS equals 'true'. Otherwise it equals 'false'.
1746  bool isKnownTrue = CI->isMinusOne();
1747  bool isKnownFalse = !isKnownTrue;
1748 
1749  // If "A && B" is known true then both A and B are known true. If "A || B"
1750  // is known false then both A and B are known false.
1751  Value *A, *B;
1752  if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1753  (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1754  Worklist.push_back(std::make_pair(A, RHS));
1755  Worklist.push_back(std::make_pair(B, RHS));
1756  continue;
1757  }
1758 
1759  // If we are propagating an equality like "(A == B)" == "true" then also
1760  // propagate the equality A == B. When propagating a comparison such as
1761  // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1762  if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1763  Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1764 
1765  // If "A == B" is known true, or "A != B" is known false, then replace
1766  // A with B everywhere in the scope.
1767  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1768  (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1769  Worklist.push_back(std::make_pair(Op0, Op1));
1770 
1771  // Handle the floating point versions of equality comparisons too.
1772  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1773  (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1774 
1775  // Floating point -0.0 and 0.0 compare equal, so we can only
1776  // propagate values if we know that we have a constant and that
1777  // its value is non-zero.
1778 
1779  // FIXME: We should do this optimization if 'no signed zeros' is
1780  // applicable via an instruction-level fast-math-flag or some other
1781  // indicator that relaxed FP semantics are being used.
1782 
1783  if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1784  Worklist.push_back(std::make_pair(Op0, Op1));
1785  }
1786 
1787  // If "A >= B" is known true, replace "A < B" with false everywhere.
1788  CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1789  Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1790  // Since we don't have the instruction "A < B" immediately to hand, work
1791  // out the value number that it would have and use that to find an
1792  // appropriate instruction (if any).
1793  uint32_t NextNum = VN.getNextUnusedValueNumber();
1794  uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1795  // If the number we were assigned was brand new then there is no point in
1796  // looking for an instruction realizing it: there cannot be one!
1797  if (Num < NextNum) {
1798  Value *NotCmp = findLeader(Root.getEnd(), Num);
1799  if (NotCmp && isa<Instruction>(NotCmp)) {
1800  unsigned NumReplacements =
1801  DominatesByEdge
1802  ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1803  : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1804  Root.getStart());
1805  Changed |= NumReplacements > 0;
1806  NumGVNEqProp += NumReplacements;
1807  // Cached information for anything that uses NotCmp will be invalid.
1808  if (MD)
1809  MD->invalidateCachedPointerInfo(NotCmp);
1810  }
1811  }
1812  // Ensure that any instruction in scope that gets the "A < B" value number
1813  // is replaced with false.
1814  // The leader table only tracks basic blocks, not edges. Only add to if we
1815  // have the simple case where the edge dominates the end.
1816  if (RootDominatesEnd)
1817  addToLeaderTable(Num, NotVal, Root.getEnd());
1818 
1819  continue;
1820  }
1821  }
1822 
1823  return Changed;
1824 }
1825 
1826 /// When calculating availability, handle an instruction
1827 /// by inserting it into the appropriate sets
1828 bool GVN::processInstruction(Instruction *I) {
1829  // Ignore dbg info intrinsics.
1830  if (isa<DbgInfoIntrinsic>(I))
1831  return false;
1832 
1833  // If the instruction can be easily simplified then do so now in preference
1834  // to value numbering it. Value numbering often exposes redundancies, for
1835  // example if it determines that %y is equal to %x then the instruction
1836  // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1837  const DataLayout &DL = I->getModule()->getDataLayout();
1838  if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1839  bool Changed = false;
1840  if (!I->use_empty()) {
1841  I->replaceAllUsesWith(V);
1842  Changed = true;
1843  }
1844  if (isInstructionTriviallyDead(I, TLI)) {
1845  markInstructionForDeletion(I);
1846  Changed = true;
1847  }
1848  if (Changed) {
1849  if (MD && V->getType()->isPtrOrPtrVectorTy())
1850  MD->invalidateCachedPointerInfo(V);
1851  ++NumGVNSimpl;
1852  return true;
1853  }
1854  }
1855 
1856  if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1857  if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1858  return processAssumeIntrinsic(IntrinsicI);
1859 
1860  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1861  if (processLoad(LI))
1862  return true;
1863 
1864  unsigned Num = VN.lookupOrAdd(LI);
1865  addToLeaderTable(Num, LI, LI->getParent());
1866  return false;
1867  }
1868 
1869  // For conditional branches, we can perform simple conditional propagation on
1870  // the condition value itself.
1871  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1872  if (!BI->isConditional())
1873  return false;
1874 
1875  if (isa<Constant>(BI->getCondition()))
1876  return processFoldableCondBr(BI);
1877 
1878  Value *BranchCond = BI->getCondition();
1879  BasicBlock *TrueSucc = BI->getSuccessor(0);
1880  BasicBlock *FalseSucc = BI->getSuccessor(1);
1881  // Avoid multiple edges early.
1882  if (TrueSucc == FalseSucc)
1883  return false;
1884 
1885  BasicBlock *Parent = BI->getParent();
1886  bool Changed = false;
1887 
1888  Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1889  BasicBlockEdge TrueE(Parent, TrueSucc);
1890  Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1891 
1892  Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1893  BasicBlockEdge FalseE(Parent, FalseSucc);
1894  Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1895 
1896  return Changed;
1897  }
1898 
1899  // For switches, propagate the case values into the case destinations.
1900  if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1901  Value *SwitchCond = SI->getCondition();
1902  BasicBlock *Parent = SI->getParent();
1903  bool Changed = false;
1904 
1905  // Remember how many outgoing edges there are to every successor.
1907  for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1908  ++SwitchEdges[SI->getSuccessor(i)];
1909 
1910  for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1911  i != e; ++i) {
1912  BasicBlock *Dst = i->getCaseSuccessor();
1913  // If there is only a single edge, propagate the case value into it.
1914  if (SwitchEdges.lookup(Dst) == 1) {
1915  BasicBlockEdge E(Parent, Dst);
1916  Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1917  }
1918  }
1919  return Changed;
1920  }
1921 
1922  // Instructions with void type don't return a value, so there's
1923  // no point in trying to find redundancies in them.
1924  if (I->getType()->isVoidTy())
1925  return false;
1926 
1927  uint32_t NextNum = VN.getNextUnusedValueNumber();
1928  unsigned Num = VN.lookupOrAdd(I);
1929 
1930  // Allocations are always uniquely numbered, so we can save time and memory
1931  // by fast failing them.
1932  if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) {
1933  addToLeaderTable(Num, I, I->getParent());
1934  return false;
1935  }
1936 
1937  // If the number we were assigned was a brand new VN, then we don't
1938  // need to do a lookup to see if the number already exists
1939  // somewhere in the domtree: it can't!
1940  if (Num >= NextNum) {
1941  addToLeaderTable(Num, I, I->getParent());
1942  return false;
1943  }
1944 
1945  // Perform fast-path value-number based elimination of values inherited from
1946  // dominators.
1947  Value *Repl = findLeader(I->getParent(), Num);
1948  if (!Repl) {
1949  // Failure, just remember this instance for future use.
1950  addToLeaderTable(Num, I, I->getParent());
1951  return false;
1952  } else if (Repl == I) {
1953  // If I was the result of a shortcut PRE, it might already be in the table
1954  // and the best replacement for itself. Nothing to do.
1955  return false;
1956  }
1957 
1958  // Remove it!
1959  patchAndReplaceAllUsesWith(I, Repl);
1960  if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1961  MD->invalidateCachedPointerInfo(Repl);
1962  markInstructionForDeletion(I);
1963  return true;
1964 }
1965 
1966 /// runOnFunction - This is the main transformation entry point for a function.
1967 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1968  const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1969  MemoryDependenceResults *RunMD, LoopInfo *LI,
1970  OptimizationRemarkEmitter *RunORE) {
1971  AC = &RunAC;
1972  DT = &RunDT;
1973  VN.setDomTree(DT);
1974  TLI = &RunTLI;
1975  VN.setAliasAnalysis(&RunAA);
1976  MD = RunMD;
1977  ImplicitControlFlowTracking ImplicitCFT(DT);
1978  ICF = &ImplicitCFT;
1979  VN.setMemDep(MD);
1980  ORE = RunORE;
1981  InvalidBlockRPONumbers = true;
1982 
1983  bool Changed = false;
1984  bool ShouldContinue = true;
1985 
1987  // Merge unconditional branches, allowing PRE to catch more
1988  // optimization opportunities.
1989  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1990  BasicBlock *BB = &*FI++;
1991 
1992  bool removedBlock = MergeBlockIntoPredecessor(BB, &DTU, LI, nullptr, MD);
1993  if (removedBlock)
1994  ++NumGVNBlocks;
1995 
1996  Changed |= removedBlock;
1997  }
1998 
1999  unsigned Iteration = 0;
2000  while (ShouldContinue) {
2001  LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2002  ShouldContinue = iterateOnFunction(F);
2003  Changed |= ShouldContinue;
2004  ++Iteration;
2005  }
2006 
2007  if (EnablePRE) {
2008  // Fabricate val-num for dead-code in order to suppress assertion in
2009  // performPRE().
2010  assignValNumForDeadCode();
2011  bool PREChanged = true;
2012  while (PREChanged) {
2013  PREChanged = performPRE(F);
2014  Changed |= PREChanged;
2015  }
2016  }
2017 
2018  // FIXME: Should perform GVN again after PRE does something. PRE can move
2019  // computations into blocks where they become fully redundant. Note that
2020  // we can't do this until PRE's critical edge splitting updates memdep.
2021  // Actually, when this happens, we should just fully integrate PRE into GVN.
2022 
2023  cleanupGlobalSets();
2024  // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2025  // iteration.
2026  DeadBlocks.clear();
2027 
2028  return Changed;
2029 }
2030 
2031 bool GVN::processBlock(BasicBlock *BB) {
2032  // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2033  // (and incrementing BI before processing an instruction).
2034  assert(InstrsToErase.empty() &&
2035  "We expect InstrsToErase to be empty across iterations");
2036  if (DeadBlocks.count(BB))
2037  return false;
2038 
2039  // Clearing map before every BB because it can be used only for single BB.
2040  ReplaceWithConstMap.clear();
2041  bool ChangedFunction = false;
2042 
2043  for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2044  BI != BE;) {
2045  if (!ReplaceWithConstMap.empty())
2046  ChangedFunction |= replaceOperandsWithConsts(&*BI);
2047  ChangedFunction |= processInstruction(&*BI);
2048 
2049  if (InstrsToErase.empty()) {
2050  ++BI;
2051  continue;
2052  }
2053 
2054  // If we need some instructions deleted, do it now.
2055  NumGVNInstr += InstrsToErase.size();
2056 
2057  // Avoid iterator invalidation.
2058  bool AtStart = BI == BB->begin();
2059  if (!AtStart)
2060  --BI;
2061 
2062  for (auto *I : InstrsToErase) {
2063  assert(I->getParent() == BB && "Removing instruction from wrong block?");
2064  LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n');
2065  salvageDebugInfo(*I);
2066  if (MD) MD->removeInstruction(I);
2067  LLVM_DEBUG(verifyRemoved(I));
2068  ICF->removeInstruction(I);
2069  I->eraseFromParent();
2070  }
2071  InstrsToErase.clear();
2072 
2073  if (AtStart)
2074  BI = BB->begin();
2075  else
2076  ++BI;
2077  }
2078 
2079  return ChangedFunction;
2080 }
2081 
2082 // Instantiate an expression in a predecessor that lacked it.
2083 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2084  BasicBlock *Curr, unsigned int ValNo) {
2085  // Because we are going top-down through the block, all value numbers
2086  // will be available in the predecessor by the time we need them. Any
2087  // that weren't originally present will have been instantiated earlier
2088  // in this loop.
2089  bool success = true;
2090  for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2091  Value *Op = Instr->getOperand(i);
2092  if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2093  continue;
2094  // This could be a newly inserted instruction, in which case, we won't
2095  // find a value number, and should give up before we hurt ourselves.
2096  // FIXME: Rewrite the infrastructure to let it easier to value number
2097  // and process newly inserted instructions.
2098  if (!VN.exists(Op)) {
2099  success = false;
2100  break;
2101  }
2102  uint32_t TValNo =
2103  VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2104  if (Value *V = findLeader(Pred, TValNo)) {
2105  Instr->setOperand(i, V);
2106  } else {
2107  success = false;
2108  break;
2109  }
2110  }
2111 
2112  // Fail out if we encounter an operand that is not available in
2113  // the PRE predecessor. This is typically because of loads which
2114  // are not value numbered precisely.
2115  if (!success)
2116  return false;
2117 
2118  Instr->insertBefore(Pred->getTerminator());
2119  Instr->setName(Instr->getName() + ".pre");
2120  Instr->setDebugLoc(Instr->getDebugLoc());
2121 
2122  unsigned Num = VN.lookupOrAdd(Instr);
2123  VN.add(Instr, Num);
2124 
2125  // Update the availability map to include the new instruction.
2126  addToLeaderTable(Num, Instr, Pred);
2127  return true;
2128 }
2129 
2130 bool GVN::performScalarPRE(Instruction *CurInst) {
2131  if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() ||
2132  isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2133  CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2134  isa<DbgInfoIntrinsic>(CurInst))
2135  return false;
2136 
2137  // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2138  // sinking the compare again, and it would force the code generator to
2139  // move the i1 from processor flags or predicate registers into a general
2140  // purpose register.
2141  if (isa<CmpInst>(CurInst))
2142  return false;
2143 
2144  // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from
2145  // sinking the addressing mode computation back to its uses. Extending the
2146  // GEP's live range increases the register pressure, and therefore it can
2147  // introduce unnecessary spills.
2148  //
2149  // This doesn't prevent Load PRE. PHI translation will make the GEP available
2150  // to the load by moving it to the predecessor block if necessary.
2151  if (isa<GetElementPtrInst>(CurInst))
2152  return false;
2153 
2154  // We don't currently value number ANY inline asm calls.
2155  if (auto *CallB = dyn_cast<CallBase>(CurInst))
2156  if (CallB->isInlineAsm())
2157  return false;
2158 
2159  uint32_t ValNo = VN.lookup(CurInst);
2160 
2161  // Look for the predecessors for PRE opportunities. We're
2162  // only trying to solve the basic diamond case, where
2163  // a value is computed in the successor and one predecessor,
2164  // but not the other. We also explicitly disallow cases
2165  // where the successor is its own predecessor, because they're
2166  // more complicated to get right.
2167  unsigned NumWith = 0;
2168  unsigned NumWithout = 0;
2169  BasicBlock *PREPred = nullptr;
2170  BasicBlock *CurrentBlock = CurInst->getParent();
2171 
2172  // Update the RPO numbers for this function.
2173  if (InvalidBlockRPONumbers)
2174  assignBlockRPONumber(*CurrentBlock->getParent());
2175 
2177  for (BasicBlock *P : predecessors(CurrentBlock)) {
2178  // We're not interested in PRE where blocks with predecessors that are
2179  // not reachable.
2180  if (!DT->isReachableFromEntry(P)) {
2181  NumWithout = 2;
2182  break;
2183  }
2184  // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2185  // when CurInst has operand defined in CurrentBlock (so it may be defined
2186  // by phi in the loop header).
2187  assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) &&
2188  "Invalid BlockRPONumber map.");
2189  if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2190  llvm::any_of(CurInst->operands(), [&](const Use &U) {
2191  if (auto *Inst = dyn_cast<Instruction>(U.get()))
2192  return Inst->getParent() == CurrentBlock;
2193  return false;
2194  })) {
2195  NumWithout = 2;
2196  break;
2197  }
2198 
2199  uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2200  Value *predV = findLeader(P, TValNo);
2201  if (!predV) {
2202  predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2203  PREPred = P;
2204  ++NumWithout;
2205  } else if (predV == CurInst) {
2206  /* CurInst dominates this predecessor. */
2207  NumWithout = 2;
2208  break;
2209  } else {
2210  predMap.push_back(std::make_pair(predV, P));
2211  ++NumWith;
2212  }
2213  }
2214 
2215  // Don't do PRE when it might increase code size, i.e. when
2216  // we would need to insert instructions in more than one pred.
2217  if (NumWithout > 1 || NumWith == 0)
2218  return false;
2219 
2220  // We may have a case where all predecessors have the instruction,
2221  // and we just need to insert a phi node. Otherwise, perform
2222  // insertion.
2223  Instruction *PREInstr = nullptr;
2224 
2225  if (NumWithout != 0) {
2226  if (!isSafeToSpeculativelyExecute(CurInst)) {
2227  // It is only valid to insert a new instruction if the current instruction
2228  // is always executed. An instruction with implicit control flow could
2229  // prevent us from doing it. If we cannot speculate the execution, then
2230  // PRE should be prohibited.
2231  if (ICF->isDominatedByICFIFromSameBlock(CurInst))
2232  return false;
2233  }
2234 
2235  // Don't do PRE across indirect branch.
2236  if (isa<IndirectBrInst>(PREPred->getTerminator()))
2237  return false;
2238 
2239  // Don't do PRE across callbr.
2240  // FIXME: Can we do this across the fallthrough edge?
2241  if (isa<CallBrInst>(PREPred->getTerminator()))
2242  return false;
2243 
2244  // We can't do PRE safely on a critical edge, so instead we schedule
2245  // the edge to be split and perform the PRE the next time we iterate
2246  // on the function.
2247  unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2248  if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2249  toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2250  return false;
2251  }
2252  // We need to insert somewhere, so let's give it a shot
2253  PREInstr = CurInst->clone();
2254  if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2255  // If we failed insertion, make sure we remove the instruction.
2256  LLVM_DEBUG(verifyRemoved(PREInstr));
2257  PREInstr->deleteValue();
2258  return false;
2259  }
2260  }
2261 
2262  // Either we should have filled in the PRE instruction, or we should
2263  // not have needed insertions.
2264  assert(PREInstr != nullptr || NumWithout == 0);
2265 
2266  ++NumGVNPRE;
2267 
2268  // Create a PHI to make the value available in this block.
2269  PHINode *Phi =
2270  PHINode::Create(CurInst->getType(), predMap.size(),
2271  CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2272  for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2273  if (Value *V = predMap[i].first) {
2274  // If we use an existing value in this phi, we have to patch the original
2275  // value because the phi will be used to replace a later value.
2276  patchReplacementInstruction(CurInst, V);
2277  Phi->addIncoming(V, predMap[i].second);
2278  } else
2279  Phi->addIncoming(PREInstr, PREPred);
2280  }
2281 
2282  VN.add(Phi, ValNo);
2283  // After creating a new PHI for ValNo, the phi translate result for ValNo will
2284  // be changed, so erase the related stale entries in phi translate cache.
2285  VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2286  addToLeaderTable(ValNo, Phi, CurrentBlock);
2287  Phi->setDebugLoc(CurInst->getDebugLoc());
2288  CurInst->replaceAllUsesWith(Phi);
2289  if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2290  MD->invalidateCachedPointerInfo(Phi);
2291  VN.erase(CurInst);
2292  removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2293 
2294  LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2295  if (MD)
2296  MD->removeInstruction(CurInst);
2297  LLVM_DEBUG(verifyRemoved(CurInst));
2298  // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes
2299  // some assertion failures.
2300  ICF->removeInstruction(CurInst);
2301  CurInst->eraseFromParent();
2302  ++NumGVNInstr;
2303 
2304  return true;
2305 }
2306 
2307 /// Perform a purely local form of PRE that looks for diamond
2308 /// control flow patterns and attempts to perform simple PRE at the join point.
2309 bool GVN::performPRE(Function &F) {
2310  bool Changed = false;
2311  for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2312  // Nothing to PRE in the entry block.
2313  if (CurrentBlock == &F.getEntryBlock())
2314  continue;
2315 
2316  // Don't perform PRE on an EH pad.
2317  if (CurrentBlock->isEHPad())
2318  continue;
2319 
2320  for (BasicBlock::iterator BI = CurrentBlock->begin(),
2321  BE = CurrentBlock->end();
2322  BI != BE;) {
2323  Instruction *CurInst = &*BI++;
2324  Changed |= performScalarPRE(CurInst);
2325  }
2326  }
2327 
2328  if (splitCriticalEdges())
2329  Changed = true;
2330 
2331  return Changed;
2332 }
2333 
2334 /// Split the critical edge connecting the given two blocks, and return
2335 /// the block inserted to the critical edge.
2336 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2337  BasicBlock *BB =
2339  if (MD)
2340  MD->invalidateCachedPredecessors();
2341  InvalidBlockRPONumbers = true;
2342  return BB;
2343 }
2344 
2345 /// Split critical edges found during the previous
2346 /// iteration that may enable further optimization.
2347 bool GVN::splitCriticalEdges() {
2348  if (toSplit.empty())
2349  return false;
2350  do {
2351  std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val();
2352  SplitCriticalEdge(Edge.first, Edge.second,
2354  } while (!toSplit.empty());
2355  if (MD) MD->invalidateCachedPredecessors();
2356  InvalidBlockRPONumbers = true;
2357  return true;
2358 }
2359 
2360 /// Executes one iteration of GVN
2361 bool GVN::iterateOnFunction(Function &F) {
2362  cleanupGlobalSets();
2363 
2364  // Top-down walk of the dominator tree
2365  bool Changed = false;
2366  // Needed for value numbering with phi construction to work.
2367  // RPOT walks the graph in its constructor and will not be invalidated during
2368  // processBlock.
2370 
2371  for (BasicBlock *BB : RPOT)
2372  Changed |= processBlock(BB);
2373 
2374  return Changed;
2375 }
2376 
2377 void GVN::cleanupGlobalSets() {
2378  VN.clear();
2379  LeaderTable.clear();
2380  BlockRPONumber.clear();
2381  TableAllocator.Reset();
2382  ICF->clear();
2383  InvalidBlockRPONumbers = true;
2384 }
2385 
2386 /// Verify that the specified instruction does not occur in our
2387 /// internal data structures.
2388 void GVN::verifyRemoved(const Instruction *Inst) const {
2389  VN.verifyRemoved(Inst);
2390 
2391  // Walk through the value number scope to make sure the instruction isn't
2392  // ferreted away in it.
2394  I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2395  const LeaderTableEntry *Node = &I->second;
2396  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2397 
2398  while (Node->Next) {
2399  Node = Node->Next;
2400  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2401  }
2402  }
2403 }
2404 
2405 /// BB is declared dead, which implied other blocks become dead as well. This
2406 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2407 /// live successors, update their phi nodes by replacing the operands
2408 /// corresponding to dead blocks with UndefVal.
2409 void GVN::addDeadBlock(BasicBlock *BB) {
2412 
2413  NewDead.push_back(BB);
2414  while (!NewDead.empty()) {
2415  BasicBlock *D = NewDead.pop_back_val();
2416  if (DeadBlocks.count(D))
2417  continue;
2418 
2419  // All blocks dominated by D are dead.
2421  DT->getDescendants(D, Dom);
2422  DeadBlocks.insert(Dom.begin(), Dom.end());
2423 
2424  // Figure out the dominance-frontier(D).
2425  for (BasicBlock *B : Dom) {
2426  for (BasicBlock *S : successors(B)) {
2427  if (DeadBlocks.count(S))
2428  continue;
2429 
2430  bool AllPredDead = true;
2431  for (BasicBlock *P : predecessors(S))
2432  if (!DeadBlocks.count(P)) {
2433  AllPredDead = false;
2434  break;
2435  }
2436 
2437  if (!AllPredDead) {
2438  // S could be proved dead later on. That is why we don't update phi
2439  // operands at this moment.
2440  DF.insert(S);
2441  } else {
2442  // While S is not dominated by D, it is dead by now. This could take
2443  // place if S already have a dead predecessor before D is declared
2444  // dead.
2445  NewDead.push_back(S);
2446  }
2447  }
2448  }
2449  }
2450 
2451  // For the dead blocks' live successors, update their phi nodes by replacing
2452  // the operands corresponding to dead blocks with UndefVal.
2454  I != E; I++) {
2455  BasicBlock *B = *I;
2456  if (DeadBlocks.count(B))
2457  continue;
2458 
2460  for (BasicBlock *P : Preds) {
2461  if (!DeadBlocks.count(P))
2462  continue;
2463 
2464  if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2465  if (BasicBlock *S = splitCriticalEdges(P, B))
2466  DeadBlocks.insert(P = S);
2467  }
2468 
2469  for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2470  PHINode &Phi = cast<PHINode>(*II);
2472  if (MD)
2473  MD->invalidateCachedPointerInfo(&Phi);
2474  }
2475  }
2476  }
2477 }
2478 
2479 // If the given branch is recognized as a foldable branch (i.e. conditional
2480 // branch with constant condition), it will perform following analyses and
2481 // transformation.
2482 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2483 // R be the target of the dead out-coming edge.
2484 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2485 // edge. The result of this step will be {X| X is dominated by R}
2486 // 2) Identify those blocks which haves at least one dead predecessor. The
2487 // result of this step will be dominance-frontier(R).
2488 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2489 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2490 //
2491 // Return true iff *NEW* dead code are found.
2492 bool GVN::processFoldableCondBr(BranchInst *BI) {
2493  if (!BI || BI->isUnconditional())
2494  return false;
2495 
2496  // If a branch has two identical successors, we cannot declare either dead.
2497  if (BI->getSuccessor(0) == BI->getSuccessor(1))
2498  return false;
2499 
2500  ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2501  if (!Cond)
2502  return false;
2503 
2504  BasicBlock *DeadRoot =
2505  Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2506  if (DeadBlocks.count(DeadRoot))
2507  return false;
2508 
2509  if (!DeadRoot->getSinglePredecessor())
2510  DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2511 
2512  addDeadBlock(DeadRoot);
2513  return true;
2514 }
2515 
2516 // performPRE() will trigger assert if it comes across an instruction without
2517 // associated val-num. As it normally has far more live instructions than dead
2518 // instructions, it makes more sense just to "fabricate" a val-number for the
2519 // dead code than checking if instruction involved is dead or not.
2520 void GVN::assignValNumForDeadCode() {
2521  for (BasicBlock *BB : DeadBlocks) {
2522  for (Instruction &Inst : *BB) {
2523  unsigned ValNum = VN.lookupOrAdd(&Inst);
2524  addToLeaderTable(ValNum, &Inst, BB);
2525  }
2526  }
2527 }
2528 
2530 public:
2531  static char ID; // Pass identification, replacement for typeid
2532 
2533  explicit GVNLegacyPass(bool NoMemDepAnalysis = !EnableMemDep)
2534  : FunctionPass(ID), NoMemDepAnalysis(NoMemDepAnalysis) {
2536  }
2537 
2538  bool runOnFunction(Function &F) override {
2539  if (skipFunction(F))
2540  return false;
2541 
2542  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2543 
2544  return Impl.runImpl(
2545  F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2546  getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2547  getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2548  getAnalysis<AAResultsWrapperPass>().getAAResults(),
2549  NoMemDepAnalysis ? nullptr
2550  : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2551  LIWP ? &LIWP->getLoopInfo() : nullptr,
2552  &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2553  }
2554 
2555  void getAnalysisUsage(AnalysisUsage &AU) const override {
2559  if (!NoMemDepAnalysis)
2562 
2567  }
2568 
2569 private:
2570  bool NoMemDepAnalysis;
2571  GVN Impl;
2572 };
2573 
2574 char GVNLegacyPass::ID = 0;
2575 
2576 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2584 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2585 
2586 // The public interface to this file...
2587 FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) {
2588  return new GVNLegacyPass(NoMemDepAnalysis);
2589 }
Legacy wrapper pass to provide the GlobalsAAResult object.
static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV)
Definition: GVN.cpp:244
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:796
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:2587
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:1578
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
bool isUndefValue() const
Definition: GVN.cpp:212
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:594
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:70
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:722
static bool runImpl(Function &F, TargetLibraryInfo &TLI, DominatorTree &DT)
This is the entry point for all transforms.
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
raw_ostream & errs()
This returns a reference to a raw_ostream for standard error.
Helper class for SSA formation on a set of values defined in multiple blocks.
Definition: SSAUpdater.h:38
Diagnostic information for missed-optimization remarks.
Provides a lazy, caching interface for making common memory aliasing information queries, backed by LLVM&#39;s alias analysis passes.
int analyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, LoadInst *DepLI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the load at De...
Definition: VNCoercion.cpp:246
DILocation * get() const
Get the underlying DILocation.
Definition: DebugLoc.cpp:21
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
This instruction extracts a struct member or array element value from an aggregate value...
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset=0)
Definition: GVN.cpp:185
size_type size() const
Definition: MapVector.h:60
unsigned Offset
Offset - The byte offset in Val that is interesting for the load query.
Definition: GVN.cpp:175
DiagnosticInfoOptimizationBase::Argument NV
static Type * makeCmpResultType(Type *opnd_type)
Create a result type for fcmp/icmp.
Definition: InstrTypes.h:975
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:776
This class represents lattice values for constants.
Definition: AllocatorList.h:23
PointerTy getPointer() const
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:473
bool isAtomic() const
Return true if this instruction has an AtomicOrdering of unordered or higher.
This is the interface for a simple mod/ref and alias analysis over globals.
void Initialize(Type *Ty, StringRef Name)
Reset this object to get ready for a new set of SSA updates with type &#39;Ty&#39;.
Definition: SSAUpdater.cpp:53
Represents an op.with.overflow intrinsic.
bool MergeBlockIntoPredecessor(BasicBlock *BB, DomTreeUpdater *DTU=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, MemoryDependenceResults *MemDep=nullptr)
Attempts to merge a block into its predecessor, if possible.
uint32_t lookupOrAddCmp(unsigned Opcode, CmpInst::Predicate Pred, Value *LHS, Value *RHS)
Returns the value number of the given comparison, assigning it a new number if it did not have one be...
Definition: GVN.cpp:569
iterator end()
Definition: Function.h:682
void setIncomingValueForBlock(const BasicBlock *BB, Value *V)
Set every incoming value(s) for block BB to V.
void AddAvailableValue(BasicBlock *BB, Value *V)
Indicate that a rewritten value is available in the specified block with the specified value...
Definition: SSAUpdater.cpp:71
bool operator==(const Expression &other) const
Definition: GVN.cpp:120
This class represents a function call, abstracting a target machine&#39;s calling convention.
bool isNonLocal() const
Tests if this MemDepResult represents a query that is transparent to the start of the block...
This file contains the declarations for metadata subclasses.
An immutable pass that tracks lazily created AssumptionCache objects.
A cache of @llvm.assume calls within a function.
bool salvageDebugInfo(Instruction &I)
Assuming the instruction I is going to be deleted, attempt to salvage debug users of I by writing the...
Definition: Local.cpp:1624
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:247
uint32_t phiTranslate(const BasicBlock *BB, const BasicBlock *PhiBlock, uint32_t Num, GVN &Gvn)
Wrap phiTranslateImpl to provide caching functionality.
Definition: GVN.cpp:1514
bool isTerminator() const
Definition: Instruction.h:128
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:748
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:98
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.h:323
unsigned second
This class implements a map that also provides access to all stored values in a deterministic order...
Definition: MapVector.h:37
BasicBlock * getSuccessor(unsigned i) const
bool properlyDominates(const DomTreeNodeBase< NodeT > *A, const DomTreeNodeBase< NodeT > *B) const
properlyDominates - Returns true iff A dominates B and A != B.
STATISTIC(NumFunctions, "Total number of functions")
A debug info location.
Definition: DebugLoc.h:33
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:230
F(f)
bool isCoercedLoadValue() const
Definition: GVN.cpp:210
An instruction for reading from memory.
Definition: Instructions.h:167
const BasicBlock * getEnd() const
Definition: Dominators.h:94
Hexagon Common GEP
Value * getCondition() const
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:137
This defines the Use class.
idx_iterator idx_end() const
unsigned replaceDominatedUsesWith(Value *From, Value *To, DominatorTree &DT, const BasicBlockEdge &Edge)
Replace each use of &#39;From&#39; with &#39;To&#39; if that use is dominated by the given edge.
Definition: Local.cpp:2461
Use * op_iterator
Definition: User.h:224
iterator end()
Get an iterator to the end of the SetVector.
Definition: SetVector.h:92
Value * getMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingMemInst returned an offset, this function can be used to actually perform...
Definition: VNCoercion.cpp:519
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:32
op_iterator op_begin()
Definition: User.h:229
gvn Early GVN Hoisting of Expressions
Definition: GVNHoist.cpp:1203
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:274
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:268
uint32_t lookup(Value *V, bool Verify=true) const
Returns the value number of the specified value.
Definition: GVN.cpp:556
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:4424
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:2555
static cl::opt< bool > EnablePRE("enable-pre", cl::init(true), cl::Hidden)
void patchReplacementInstruction(Instruction *I, Value *Repl)
Patch the replacement so that it is not more restrictive than the value being replaced.
Definition: Local.cpp:2392
bool isDef() const
Tests if this MemDepResult represents a query that is an instruction definition dependency.
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:369
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: GVN.cpp:2538
Option class for critical edge splitting.
void clear()
Remove all entries from the ValueTable.
Definition: GVN.cpp:577
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:1152
static const uint16_t * lookup(unsigned opcode, unsigned domain, ArrayRef< uint16_t[3]> Table)
bool isSimpleValue() const
Definition: GVN.cpp:209
Interval::succ_iterator succ_begin(Interval *I)
succ_begin/succ_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:102
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:244
ppc ctr loops PowerPC CTR Loops Verify
bool insert(const value_type &X)
Insert a new element into the SetVector.
Definition: SetVector.h:141
The core GVN pass object.
Definition: GVN.h:68
IntType getInt() const
bool canCoerceMustAliasedValueToLoad(Value *StoredVal, Type *LoadTy, const DataLayout &DL)
Return true if CoerceAvailableValueToLoadType would succeed if it was called.
Definition: VNCoercion.cpp:15
Expression(uint32_t o=~2U)
Definition: GVN.cpp:118
#define DEBUG_TYPE
Definition: GVN.cpp:88
iterator begin()
Get an iterator to the beginning of the SetVector.
Definition: SetVector.h:82
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:234
DiagnosticInfoOptimizationBase::setExtraArgs setExtraArgs
static AvailableValue getLoad(LoadInst *LI, unsigned Offset=0)
Definition: GVN.cpp:193
hash_code hash_value(const APFloat &Arg)
See friend declarations above.
Definition: APFloat.cpp:4470
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
BasicBlock * SplitCriticalEdge(Instruction *TI, unsigned SuccNum, const CriticalEdgeSplittingOptions &Options=CriticalEdgeSplittingOptions())
If this edge is a critical edge, insert a new node to split the critical edge.
LoadInst * getCoercedLoadValue() const
Definition: GVN.cpp:219
static GVN::Expression getEmptyKey()
Definition: GVN.cpp:142
An instruction for storing to memory.
Definition: Instructions.h:320
bool isMinusOne() const
This function will return true iff every bit in this constant is set to true.
Definition: Constants.h:208
void add(Value *V, uint32_t num)
add - Insert a value into the table with a specified value number.
Definition: GVN.cpp:368
Value * getRHS() const
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:429
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:291
iterator begin()
Definition: Function.h:680
static unsigned getHashValue(const GVN::Expression &e)
Definition: GVN.cpp:145
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:144
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Value * getOperand(unsigned i) const
Definition: User.h:169
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:105
int analyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, StoreInst *DepSI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the store at D...
Definition: VNCoercion.cpp:218
void initializeGVNLegacyPassPass(PassRegistry &)
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:140
const BasicBlock & getEntryBlock() const
Definition: Function.h: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:116
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:148
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:318
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:233
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:73
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
PointerIntPair - This class implements a pair of a pointer and small integer.
PHITransAddr - An address value which tracks and handles phi translation.
Definition: PHITransAddr.h:35
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
Definition: PatternMatch.h:802
Conditional or Unconditional Branch instruction.
This file provides the interface for LLVM&#39;s Global Value Numbering pass which eliminates fully redund...
static GVN::Expression getTombstoneKey()
Definition: GVN.cpp:143
static Value * ConstructSSAForLoadSet(LoadInst *LI, SmallVectorImpl< AvailableValueInBlock > &ValuesPerBlock, GVN &gvn)
Given a set of loads specified by ValuesPerBlock, construct SSA form, allowing us to eliminate LI...
Definition: GVN.cpp:727
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:41
static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS)
Definition: GVN.cpp:151
This file contains the declarations for the subclasses of Constant, which represent the different fla...
static cl::opt< uint32_t > MaxRecurseDepth("gvn-max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, cl::desc("Max recurse depth in GVN (default = 1000)"))
const Instruction & front() const
Definition: BasicBlock.h:280
A manager for alias analyses.
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:572
Diagnostic information for applied optimization remarks.
Interval::pred_iterator pred_begin(Interval *I)
pred_begin/pred_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:112
unsigned getNumIndices() const
bool isUnordered() const
Definition: Instructions.h:278
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:231
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1199
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:732
PointerIntPair< Value *, 2, ValType > Val
V - The value that is live out of the block.
Definition: GVN.cpp:172
MemIntrinsic * getMemIntrinValue() const
Definition: GVN.cpp:224
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:115
op_range operands()
Definition: User.h:237
Value * getPointerOperand()
Definition: Instructions.h:284
bool isCriticalEdge(const Instruction *TI, unsigned SuccNum, bool AllowIdenticalEdges=false)
Return true if the specified edge is a critical edge.
Definition: CFG.cpp:88
Value * getLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingLoad returned an offset, this function can be used to actually perform th...
Definition: VNCoercion.cpp:407
static void reportLoadElim(LoadInst *LI, Value *AvailableValue, OptimizationRemarkEmitter *ORE)
Definition: GVN.cpp:1272
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1424
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
size_t size() const
Definition: SmallVector.h:52
static cl::opt< bool > EnableMemDep("enable-gvn-memdep", cl::init(true))
A wrapper analysis pass for the legacy pass manager that exposes a MemoryDepnedenceResults instance...
void printAsOperand(raw_ostream &O, bool PrintType=true, const Module *M=nullptr) const
Print the name of this Value out to the specified raw_ostream.
Definition: AsmWriter.cpp:4351
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
A memory dependence query can return one of three different answers.
DominatorTree & getDominatorTree() const
Definition: GVN.h:82
unsigned first
static cl::opt< uint32_t > MaxNumDeps("gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore, cl::desc("Max number of dependences to attempt Load PRE (default = 100)"))
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:50
static void reportMayClobberedLoad(LoadInst *LI, MemDepResult DepInfo, DominatorTree *DT, OptimizationRemarkEmitter *ORE)
Try to locate the three instruction involved in a missed load-elimination case that is due to an inte...
Definition: GVN.cpp:824
A function analysis which provides an AssumptionCache.
bool isPtrOrPtrVectorTy() const
Return true if this is a pointer type or a vector of pointer types.
Definition: Type.h:226
Value * MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const
Emit code at the end of this block to adjust the value defined here to the specified type...
Definition: GVN.cpp:262
A SetVector that performs no allocations if smaller than a certain size.
Definition: SetVector.h:297
This is the common base class for memset/memcpy/memmove.
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:191
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:417
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
void emit(DiagnosticInfoOptimizationBase &OptDiag)
Output the remark via the diagnostic handler and to the optimization record file. ...
iterator end()
Definition: BasicBlock.h:270
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:248
Module.h This file contains the declarations for the Module class.
Provides information about what library functions are available for the current target.
const MemDepResult & getResult() const
size_type count(const KeyT &Key) const
Definition: MapVector.h:142
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:47
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:643
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:374
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:631
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:124
Value * MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt, GVN &gvn) const
Emit code at the specified insertion point to adjust the value defined here to the specified type...
Definition: GVN.cpp:767
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:587
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:488
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:940
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:55
hash_code hash_combine(const Ts &...args)
Combine values into a single hash_code.
Definition: Hashing.h:600
Represents an AvailableValue which can be rematerialized at the end of the associated BasicBlock...
Definition: GVN.cpp:237
iterator_range< user_iterator > users()
Definition: Value.h:399
hash_code hash_combine_range(InputIteratorT first, InputIteratorT last)
Compute a hash_code for a sequence of values.
Definition: Hashing.h:478
std::vector< NonLocalDepEntry > NonLocalDepInfo
An opaque object representing a hash code.
Definition: Hashing.h:71
bool isMallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates uninitialized memory (such ...
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:467
void verifyRemoved(const Value *) const
verifyRemoved - Verify that the value is removed from all internal data structures.
Definition: GVN.cpp:599
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:589
static bool isLifetimeStart(const Instruction *Inst)
Definition: GVN.cpp:816
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:2533
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:214
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
bool isCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates zero-filled memory (such as...
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this load instruction.
Definition: Instructions.h:259
#define I(x, y, z)
Definition: MD5.cpp:58
bool mayReadFromMemory() const
Return true if this instruction may read memory.
static AvailableValue get(Value *V, unsigned Offset=0)
Definition: GVN.cpp:177
uint32_t opcode
Definition: GVN.cpp:113
PassT::Result * getCachedResult(IRUnitT &IR) const
Get the cached result of an analysis pass for a given IR unit.
Definition: PassManager.h:795
bool exists(Value *V) const
Returns true if a value number exists for the specified value.
Definition: GVN.cpp:476
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:332
idx_iterator idx_begin() const
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
This class allows to keep track on instructions with implicit control flow.
bool isUnconditional() const
friend hash_code hash_value(const Expression &Value)
Definition: GVN.cpp:132
uint32_t lookupOrAdd(Value *V)
lookup_or_add - Returns the value number for the specified value, assigning it a new number if it did...
Definition: GVN.cpp:480
ValueT lookup(const_arg_type_t< KeyT > Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: DenseMap.h:211
Value * getSimpleValue() const
Definition: GVN.cpp:214
Analysis pass providing the TargetLibraryInfo.
iterator_range< df_iterator< T > > depth_first(const T &G)
Multiway switch.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
const BasicBlock * getStart() const
Definition: Dominators.h:90
Represents a particular available value that we know how to materialize.
Definition: GVN.cpp:162
bool isSafeToSpeculativelyExecute(const Value *V, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr)
Return true if the instruction does not have any effects besides calculating the result and does not ...
static bool IsValueFullyAvailableInBlock(BasicBlock *BB, DenseMap< BasicBlock *, char > &FullyAvailableBlocks, uint32_t RecurseDepth)
Return true if we can prove that the value we&#39;re analyzing is fully available in the specified block...
Definition: GVN.cpp:654
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:735
bool isInstructionTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction is not used, and the instruction has no side ef...
Definition: Local.cpp:353
LLVM Value Representation.
Definition: Value.h:72
static AvailableValueInBlock getUndef(BasicBlock *BB)
Definition: GVN.cpp:256
void removeInstruction(Instruction *InstToRemove)
Removes an instruction from the dependence analysis, updating the dependence of instructions that pre...
succ_range successors(Instruction *I)
Definition: CFG.h:259
OptimizationRemarkEmitter legacy analysis pass.
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: GVN.cpp:610
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:412
Predicate getSwappedPredicate() const
For example, EQ->EQ, SLE->SGE, ULT->UGT, OEQ->OEQ, ULE->UGE, OLT->OGT, etc.
Definition: InstrTypes.h:847
This is an entry in the NonLocalDepInfo cache.
A container for analyses that lazily runs them and caches their results.
BasicBlock * BB
BB - The basic block in question.
Definition: GVN.cpp:239
static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl)
Definition: GVN.cpp:1431
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:259
bool isMemIntrinValue() const
Definition: GVN.cpp:211
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
This header defines various interfaces for pass management in LLVM.
AvailableValue AV
AV - The actual available value.
Definition: GVN.cpp:242
#define LLVM_DEBUG(X)
Definition: Debug.h:122
Value * SimplifyInstruction(Instruction *I, const SimplifyQuery &Q, OptimizationRemarkEmitter *ORE=nullptr)
See if we can compute a simplified version of this instruction.
Value * getLHS() const
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:173
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:322
static AvailableValue getUndef()
Definition: GVN.cpp:201
static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, DominatorTree *DT)
There is an edge from &#39;Src&#39; to &#39;Dst&#39;.
Definition: GVN.cpp:1618
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