LLVM  8.0.0svn
GVN.cpp
Go to the documentation of this file.
1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
12 //
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
15 //
16 //===----------------------------------------------------------------------===//
17 
19 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/Hashing.h"
22 #include "llvm/ADT/MapVector.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/ADT/SetVector.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/Statistic.h"
32 #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"
49 #include "llvm/IR/DebugLoc.h"
50 #include "llvm/IR/DomTreeUpdater.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 (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
335  // EI might be an extract from one of our recognised intrinsics. If it
336  // is we'll synthesize a semantically equivalent expression instead on
337  // an extract value expression.
338  switch (I->getIntrinsicID()) {
339  case Intrinsic::sadd_with_overflow:
340  case Intrinsic::uadd_with_overflow:
341  e.opcode = Instruction::Add;
342  break;
343  case Intrinsic::ssub_with_overflow:
344  case Intrinsic::usub_with_overflow:
345  e.opcode = Instruction::Sub;
346  break;
347  case Intrinsic::smul_with_overflow:
348  case Intrinsic::umul_with_overflow:
349  e.opcode = Instruction::Mul;
350  break;
351  default:
352  break;
353  }
354 
355  if (e.opcode != 0) {
356  // Intrinsic recognized. Grab its args to finish building the expression.
357  assert(I->getNumArgOperands() == 2 &&
358  "Expect two args for recognised intrinsics.");
359  e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
360  e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
361  return e;
362  }
363  }
364 
365  // Not a recognised intrinsic. Fall back to producing an extract value
366  // expression.
367  e.opcode = EI->getOpcode();
368  for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
369  OI != OE; ++OI)
370  e.varargs.push_back(lookupOrAdd(*OI));
371 
372  for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
373  II != IE; ++II)
374  e.varargs.push_back(*II);
375 
376  return e;
377 }
378 
379 //===----------------------------------------------------------------------===//
380 // ValueTable External Functions
381 //===----------------------------------------------------------------------===//
382 
383 GVN::ValueTable::ValueTable() = default;
384 GVN::ValueTable::ValueTable(const ValueTable &) = default;
385 GVN::ValueTable::ValueTable(ValueTable &&) = default;
386 GVN::ValueTable::~ValueTable() = default;
387 
388 /// add - Insert a value into the table with a specified value number.
390  valueNumbering.insert(std::make_pair(V, num));
391  if (PHINode *PN = dyn_cast<PHINode>(V))
392  NumberingPhi[num] = PN;
393 }
394 
395 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
396  if (AA->doesNotAccessMemory(C)) {
397  Expression exp = createExpr(C);
398  uint32_t e = assignExpNewValueNum(exp).first;
399  valueNumbering[C] = e;
400  return e;
401  } else if (MD && AA->onlyReadsMemory(C)) {
402  Expression exp = createExpr(C);
403  auto ValNum = assignExpNewValueNum(exp);
404  if (ValNum.second) {
405  valueNumbering[C] = ValNum.first;
406  return ValNum.first;
407  }
408 
409  MemDepResult local_dep = MD->getDependency(C);
410 
411  if (!local_dep.isDef() && !local_dep.isNonLocal()) {
412  valueNumbering[C] = nextValueNumber;
413  return nextValueNumber++;
414  }
415 
416  if (local_dep.isDef()) {
417  CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
418 
419  if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
420  valueNumbering[C] = nextValueNumber;
421  return nextValueNumber++;
422  }
423 
424  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
425  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
426  uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
427  if (c_vn != cd_vn) {
428  valueNumbering[C] = nextValueNumber;
429  return nextValueNumber++;
430  }
431  }
432 
433  uint32_t v = lookupOrAdd(local_cdep);
434  valueNumbering[C] = v;
435  return v;
436  }
437 
438  // Non-local case.
440  MD->getNonLocalCallDependency(CallSite(C));
441  // FIXME: Move the checking logic to MemDep!
442  CallInst* cdep = nullptr;
443 
444  // Check to see if we have a single dominating call instruction that is
445  // identical to C.
446  for (unsigned i = 0, e = deps.size(); i != e; ++i) {
447  const NonLocalDepEntry *I = &deps[i];
448  if (I->getResult().isNonLocal())
449  continue;
450 
451  // We don't handle non-definitions. If we already have a call, reject
452  // instruction dependencies.
453  if (!I->getResult().isDef() || cdep != nullptr) {
454  cdep = nullptr;
455  break;
456  }
457 
458  CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
459  // FIXME: All duplicated with non-local case.
460  if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
461  cdep = NonLocalDepCall;
462  continue;
463  }
464 
465  cdep = nullptr;
466  break;
467  }
468 
469  if (!cdep) {
470  valueNumbering[C] = nextValueNumber;
471  return nextValueNumber++;
472  }
473 
474  if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
475  valueNumbering[C] = nextValueNumber;
476  return nextValueNumber++;
477  }
478  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
479  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
480  uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
481  if (c_vn != cd_vn) {
482  valueNumbering[C] = nextValueNumber;
483  return nextValueNumber++;
484  }
485  }
486 
487  uint32_t v = lookupOrAdd(cdep);
488  valueNumbering[C] = v;
489  return v;
490  } else {
491  valueNumbering[C] = nextValueNumber;
492  return nextValueNumber++;
493  }
494 }
495 
496 /// Returns true if a value number exists for the specified value.
497 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
498 
499 /// lookup_or_add - Returns the value number for the specified value, assigning
500 /// it a new number if it did not have one before.
502  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
503  if (VI != valueNumbering.end())
504  return VI->second;
505 
506  if (!isa<Instruction>(V)) {
507  valueNumbering[V] = nextValueNumber;
508  return nextValueNumber++;
509  }
510 
511  Instruction* I = cast<Instruction>(V);
512  Expression exp;
513  switch (I->getOpcode()) {
514  case Instruction::Call:
515  return lookupOrAddCall(cast<CallInst>(I));
516  case Instruction::Add:
517  case Instruction::FAdd:
518  case Instruction::Sub:
519  case Instruction::FSub:
520  case Instruction::Mul:
521  case Instruction::FMul:
522  case Instruction::UDiv:
523  case Instruction::SDiv:
524  case Instruction::FDiv:
525  case Instruction::URem:
526  case Instruction::SRem:
527  case Instruction::FRem:
528  case Instruction::Shl:
529  case Instruction::LShr:
530  case Instruction::AShr:
531  case Instruction::And:
532  case Instruction::Or:
533  case Instruction::Xor:
534  case Instruction::ICmp:
535  case Instruction::FCmp:
536  case Instruction::Trunc:
537  case Instruction::ZExt:
538  case Instruction::SExt:
539  case Instruction::FPToUI:
540  case Instruction::FPToSI:
541  case Instruction::UIToFP:
542  case Instruction::SIToFP:
543  case Instruction::FPTrunc:
544  case Instruction::FPExt:
545  case Instruction::PtrToInt:
546  case Instruction::IntToPtr:
547  case Instruction::BitCast:
548  case Instruction::Select:
549  case Instruction::ExtractElement:
550  case Instruction::InsertElement:
551  case Instruction::ShuffleVector:
552  case Instruction::InsertValue:
553  case Instruction::GetElementPtr:
554  exp = createExpr(I);
555  break;
556  case Instruction::ExtractValue:
557  exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
558  break;
559  case Instruction::PHI:
560  valueNumbering[V] = nextValueNumber;
561  NumberingPhi[nextValueNumber] = cast<PHINode>(V);
562  return nextValueNumber++;
563  default:
564  valueNumbering[V] = nextValueNumber;
565  return nextValueNumber++;
566  }
567 
568  uint32_t e = assignExpNewValueNum(exp).first;
569  valueNumbering[V] = e;
570  return e;
571 }
572 
573 /// Returns the value number of the specified value. Fails if
574 /// the value has not yet been numbered.
576  DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
577  if (Verify) {
578  assert(VI != valueNumbering.end() && "Value not numbered?");
579  return VI->second;
580  }
581  return (VI != valueNumbering.end()) ? VI->second : 0;
582 }
583 
584 /// Returns the value number of the given comparison,
585 /// assigning it a new number if it did not have one before. Useful when
586 /// we deduced the result of a comparison, but don't immediately have an
587 /// instruction realizing that comparison to hand.
589  CmpInst::Predicate Predicate,
590  Value *LHS, Value *RHS) {
591  Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
592  return assignExpNewValueNum(exp).first;
593 }
594 
595 /// Remove all entries from the ValueTable.
597  valueNumbering.clear();
598  expressionNumbering.clear();
599  NumberingPhi.clear();
600  PhiTranslateTable.clear();
601  nextValueNumber = 1;
602  Expressions.clear();
603  ExprIdx.clear();
604  nextExprNumber = 0;
605 }
606 
607 /// Remove a value from the value numbering.
609  uint32_t Num = valueNumbering.lookup(V);
610  valueNumbering.erase(V);
611  // If V is PHINode, V <--> value number is an one-to-one mapping.
612  if (isa<PHINode>(V))
613  NumberingPhi.erase(Num);
614 }
615 
616 /// verifyRemoved - Verify that the value is removed from all internal data
617 /// structures.
618 void GVN::ValueTable::verifyRemoved(const Value *V) const {
620  I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
621  assert(I->first != V && "Inst still occurs in value numbering map!");
622  }
623 }
624 
625 //===----------------------------------------------------------------------===//
626 // GVN Pass
627 //===----------------------------------------------------------------------===//
628 
630  // FIXME: The order of evaluation of these 'getResult' calls is very
631  // significant! Re-ordering these variables will cause GVN when run alone to
632  // be less effective! We should fix memdep and basic-aa to not exhibit this
633  // behavior, but until then don't change the order here.
634  auto &AC = AM.getResult<AssumptionAnalysis>(F);
635  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
636  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
637  auto &AA = AM.getResult<AAManager>(F);
638  auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
639  auto *LI = AM.getCachedResult<LoopAnalysis>(F);
641  bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
642  if (!Changed)
643  return PreservedAnalyses::all();
646  PA.preserve<GlobalsAA>();
648  return PA;
649 }
650 
651 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
652 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
653  errs() << "{\n";
655  E = d.end(); I != E; ++I) {
656  errs() << I->first << "\n";
657  I->second->dump();
658  }
659  errs() << "}\n";
660 }
661 #endif
662 
663 /// Return true if we can prove that the value
664 /// we're analyzing is fully available in the specified block. As we go, keep
665 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
666 /// map is actually a tri-state map with the following values:
667 /// 0) we know the block *is not* fully available.
668 /// 1) we know the block *is* fully available.
669 /// 2) we do not know whether the block is fully available or not, but we are
670 /// currently speculating that it will be.
671 /// 3) we are speculating for this block and have used that to speculate for
672 /// other blocks.
674  DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
675  uint32_t RecurseDepth) {
676  if (RecurseDepth > MaxRecurseDepth)
677  return false;
678 
679  // Optimistically assume that the block is fully available and check to see
680  // if we already know about this block in one lookup.
681  std::pair<DenseMap<BasicBlock*, char>::iterator, bool> IV =
682  FullyAvailableBlocks.insert(std::make_pair(BB, 2));
683 
684  // If the entry already existed for this block, return the precomputed value.
685  if (!IV.second) {
686  // If this is a speculative "available" value, mark it as being used for
687  // speculation of other blocks.
688  if (IV.first->second == 2)
689  IV.first->second = 3;
690  return IV.first->second != 0;
691  }
692 
693  // Otherwise, see if it is fully available in all predecessors.
694  pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
695 
696  // If this block has no predecessors, it isn't live-in here.
697  if (PI == PE)
698  goto SpeculationFailure;
699 
700  for (; PI != PE; ++PI)
701  // If the value isn't fully available in one of our predecessors, then it
702  // isn't fully available in this block either. Undo our previous
703  // optimistic assumption and bail out.
704  if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
705  goto SpeculationFailure;
706 
707  return true;
708 
709 // If we get here, we found out that this is not, after
710 // all, a fully-available block. We have a problem if we speculated on this and
711 // used the speculation to mark other blocks as available.
712 SpeculationFailure:
713  char &BBVal = FullyAvailableBlocks[BB];
714 
715  // If we didn't speculate on this, just return with it set to false.
716  if (BBVal == 2) {
717  BBVal = 0;
718  return false;
719  }
720 
721  // If we did speculate on this value, we could have blocks set to 1 that are
722  // incorrect. Walk the (transitive) successors of this block and mark them as
723  // 0 if set to one.
724  SmallVector<BasicBlock*, 32> BBWorklist;
725  BBWorklist.push_back(BB);
726 
727  do {
728  BasicBlock *Entry = BBWorklist.pop_back_val();
729  // Note that this sets blocks to 0 (unavailable) if they happen to not
730  // already be in FullyAvailableBlocks. This is safe.
731  char &EntryVal = FullyAvailableBlocks[Entry];
732  if (EntryVal == 0) continue; // Already unavailable.
733 
734  // Mark as unavailable.
735  EntryVal = 0;
736 
737  BBWorklist.append(succ_begin(Entry), succ_end(Entry));
738  } while (!BBWorklist.empty());
739 
740  return false;
741 }
742 
743 /// Given a set of loads specified by ValuesPerBlock,
744 /// construct SSA form, allowing us to eliminate LI. This returns the value
745 /// that should be used at LI's definition site.
748  GVN &gvn) {
749  // Check for the fully redundant, dominating load case. In this case, we can
750  // just use the dominating value directly.
751  if (ValuesPerBlock.size() == 1 &&
752  gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
753  LI->getParent())) {
754  assert(!ValuesPerBlock[0].AV.isUndefValue() &&
755  "Dead BB dominate this block");
756  return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
757  }
758 
759  // Otherwise, we have to construct SSA form.
760  SmallVector<PHINode*, 8> NewPHIs;
761  SSAUpdater SSAUpdate(&NewPHIs);
762  SSAUpdate.Initialize(LI->getType(), LI->getName());
763 
764  for (const AvailableValueInBlock &AV : ValuesPerBlock) {
765  BasicBlock *BB = AV.BB;
766 
767  if (SSAUpdate.HasValueForBlock(BB))
768  continue;
769 
770  // If the value is the load that we will be eliminating, and the block it's
771  // available in is the block that the load is in, then don't add it as
772  // SSAUpdater will resolve the value to the relevant phi which may let it
773  // avoid phi construction entirely if there's actually only one value.
774  if (BB == LI->getParent() &&
775  ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == LI) ||
776  (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == LI)))
777  continue;
778 
779  SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
780  }
781 
782  // Perform PHI construction.
783  return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
784 }
785 
787  Instruction *InsertPt,
788  GVN &gvn) const {
789  Value *Res;
790  Type *LoadTy = LI->getType();
791  const DataLayout &DL = LI->getModule()->getDataLayout();
792  if (isSimpleValue()) {
793  Res = getSimpleValue();
794  if (Res->getType() != LoadTy) {
795  Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
796 
797  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset
798  << " " << *getSimpleValue() << '\n'
799  << *Res << '\n'
800  << "\n\n\n");
801  }
802  } else if (isCoercedLoadValue()) {
803  LoadInst *Load = getCoercedLoadValue();
804  if (Load->getType() == LoadTy && Offset == 0) {
805  Res = Load;
806  } else {
807  Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
808  // We would like to use gvn.markInstructionForDeletion here, but we can't
809  // because the load is already memoized into the leader map table that GVN
810  // tracks. It is potentially possible to remove the load from the table,
811  // but then there all of the operations based on it would need to be
812  // rehashed. Just leave the dead load around.
813  gvn.getMemDep().removeInstruction(Load);
814  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset
815  << " " << *getCoercedLoadValue() << '\n'
816  << *Res << '\n'
817  << "\n\n\n");
818  }
819  } else if (isMemIntrinValue()) {
820  Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
821  InsertPt, DL);
822  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
823  << " " << *getMemIntrinValue() << '\n'
824  << *Res << '\n'
825  << "\n\n\n");
826  } else {
827  assert(isUndefValue() && "Should be UndefVal");
828  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
829  return UndefValue::get(LoadTy);
830  }
831  assert(Res && "failed to materialize?");
832  return Res;
833 }
834 
835 static bool isLifetimeStart(const Instruction *Inst) {
836  if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
837  return II->getIntrinsicID() == Intrinsic::lifetime_start;
838  return false;
839 }
840 
841 /// Try to locate the three instruction involved in a missed
842 /// load-elimination case that is due to an intervening store.
844  DominatorTree *DT,
846  using namespace ore;
847 
848  User *OtherAccess = nullptr;
849 
850  OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
851  R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
852  << setExtraArgs();
853 
854  for (auto *U : LI->getPointerOperand()->users())
855  if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
856  DT->dominates(cast<Instruction>(U), LI)) {
857  // FIXME: for now give up if there are multiple memory accesses that
858  // dominate the load. We need further analysis to decide which one is
859  // that we're forwarding from.
860  if (OtherAccess)
861  OtherAccess = nullptr;
862  else
863  OtherAccess = U;
864  }
865 
866  if (OtherAccess)
867  R << " in favor of " << NV("OtherAccess", OtherAccess);
868 
869  R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
870 
871  ORE->emit(R);
872 }
873 
874 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
875  Value *Address, AvailableValue &Res) {
876  assert((DepInfo.isDef() || DepInfo.isClobber()) &&
877  "expected a local dependence");
878  assert(LI->isUnordered() && "rules below are incorrect for ordered access");
879 
880  const DataLayout &DL = LI->getModule()->getDataLayout();
881 
882  if (DepInfo.isClobber()) {
883  // If the dependence is to a store that writes to a superset of the bits
884  // read by the load, we can extract the bits we need for the load from the
885  // stored value.
886  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
887  // Can't forward from non-atomic to atomic without violating memory model.
888  if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
889  int Offset =
891  if (Offset != -1) {
892  Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
893  return true;
894  }
895  }
896  }
897 
898  // Check to see if we have something like this:
899  // load i32* P
900  // load i8* (P+1)
901  // if we have this, replace the later with an extraction from the former.
902  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
903  // If this is a clobber and L is the first instruction in its block, then
904  // we have the first instruction in the entry block.
905  // Can't forward from non-atomic to atomic without violating memory model.
906  if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
907  int Offset =
908  analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
909 
910  if (Offset != -1) {
911  Res = AvailableValue::getLoad(DepLI, Offset);
912  return true;
913  }
914  }
915  }
916 
917  // If the clobbering value is a memset/memcpy/memmove, see if we can
918  // forward a value on from it.
919  if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
920  if (Address && !LI->isAtomic()) {
922  DepMI, DL);
923  if (Offset != -1) {
924  Res = AvailableValue::getMI(DepMI, Offset);
925  return true;
926  }
927  }
928  }
929  // Nothing known about this clobber, have to be conservative
930  LLVM_DEBUG(
931  // fast print dep, using operator<< on instruction is too slow.
932  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
933  Instruction *I = DepInfo.getInst();
934  dbgs() << " is clobbered by " << *I << '\n';);
935  if (ORE->allowExtraAnalysis(DEBUG_TYPE))
936  reportMayClobberedLoad(LI, DepInfo, DT, ORE);
937 
938  return false;
939  }
940  assert(DepInfo.isDef() && "follows from above");
941 
942  Instruction *DepInst = DepInfo.getInst();
943 
944  // Loading the allocation -> undef.
945  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
946  // Loading immediately after lifetime begin -> undef.
947  isLifetimeStart(DepInst)) {
949  return true;
950  }
951 
952  // Loading from calloc (which zero initializes memory) -> zero
953  if (isCallocLikeFn(DepInst, TLI)) {
955  return true;
956  }
957 
958  if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
959  // Reject loads and stores that are to the same address but are of
960  // different types if we have to. If the stored value is larger or equal to
961  // the loaded value, we can reuse it.
962  if (S->getValueOperand()->getType() != LI->getType() &&
963  !canCoerceMustAliasedValueToLoad(S->getValueOperand(),
964  LI->getType(), DL))
965  return false;
966 
967  // Can't forward from non-atomic to atomic without violating memory model.
968  if (S->isAtomic() < LI->isAtomic())
969  return false;
970 
971  Res = AvailableValue::get(S->getValueOperand());
972  return true;
973  }
974 
975  if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
976  // If the types mismatch and we can't handle it, reject reuse of the load.
977  // If the stored value is larger or equal to the loaded value, we can reuse
978  // it.
979  if (LD->getType() != LI->getType() &&
981  return false;
982 
983  // Can't forward from non-atomic to atomic without violating memory model.
984  if (LD->isAtomic() < LI->isAtomic())
985  return false;
986 
988  return true;
989  }
990 
991  // Unknown def - must be conservative
992  LLVM_DEBUG(
993  // fast print dep, using operator<< on instruction is too slow.
994  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
995  dbgs() << " has unknown def " << *DepInst << '\n';);
996  return false;
997 }
998 
999 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1000  AvailValInBlkVect &ValuesPerBlock,
1001  UnavailBlkVect &UnavailableBlocks) {
1002  // Filter out useless results (non-locals, etc). Keep track of the blocks
1003  // where we have a value available in repl, also keep track of whether we see
1004  // dependencies that produce an unknown value for the load (such as a call
1005  // that could potentially clobber the load).
1006  unsigned NumDeps = Deps.size();
1007  for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1008  BasicBlock *DepBB = Deps[i].getBB();
1009  MemDepResult DepInfo = Deps[i].getResult();
1010 
1011  if (DeadBlocks.count(DepBB)) {
1012  // Dead dependent mem-op disguise as a load evaluating the same value
1013  // as the load in question.
1014  ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1015  continue;
1016  }
1017 
1018  if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1019  UnavailableBlocks.push_back(DepBB);
1020  continue;
1021  }
1022 
1023  // The address being loaded in this non-local block may not be the same as
1024  // the pointer operand of the load if PHI translation occurs. Make sure
1025  // to consider the right address.
1026  Value *Address = Deps[i].getAddress();
1027 
1028  AvailableValue AV;
1029  if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1030  // subtlety: because we know this was a non-local dependency, we know
1031  // it's safe to materialize anywhere between the instruction within
1032  // DepInfo and the end of it's block.
1033  ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1034  std::move(AV)));
1035  } else {
1036  UnavailableBlocks.push_back(DepBB);
1037  }
1038  }
1039 
1040  assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1041  "post condition violation");
1042 }
1043 
1044 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1045  UnavailBlkVect &UnavailableBlocks) {
1046  // Okay, we have *some* definitions of the value. This means that the value
1047  // is available in some of our (transitive) predecessors. Lets think about
1048  // doing PRE of this load. This will involve inserting a new load into the
1049  // predecessor when it's not available. We could do this in general, but
1050  // prefer to not increase code size. As such, we only do this when we know
1051  // that we only have to insert *one* load (which means we're basically moving
1052  // the load, not inserting a new one).
1053 
1054  SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1055  UnavailableBlocks.end());
1056 
1057  // Let's find the first basic block with more than one predecessor. Walk
1058  // backwards through predecessors if needed.
1059  BasicBlock *LoadBB = LI->getParent();
1060  BasicBlock *TmpBB = LoadBB;
1061  bool IsSafeToSpeculativelyExecute = isSafeToSpeculativelyExecute(LI);
1062 
1063  // Check that there is no implicit control flow instructions above our load in
1064  // its block. If there is an instruction that doesn't always pass the
1065  // execution to the following instruction, then moving through it may become
1066  // invalid. For example:
1067  //
1068  // int arr[LEN];
1069  // int index = ???;
1070  // ...
1071  // guard(0 <= index && index < LEN);
1072  // use(arr[index]);
1073  //
1074  // It is illegal to move the array access to any point above the guard,
1075  // because if the index is out of bounds we should deoptimize rather than
1076  // access the array.
1077  // Check that there is no guard in this block above our instruction.
1078  if (!IsSafeToSpeculativelyExecute && ICF->isDominatedByICFIFromSameBlock(LI))
1079  return false;
1080  while (TmpBB->getSinglePredecessor()) {
1081  TmpBB = TmpBB->getSinglePredecessor();
1082  if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1083  return false;
1084  if (Blockers.count(TmpBB))
1085  return false;
1086 
1087  // If any of these blocks has more than one successor (i.e. if the edge we
1088  // just traversed was critical), then there are other paths through this
1089  // block along which the load may not be anticipated. Hoisting the load
1090  // above this block would be adding the load to execution paths along
1091  // which it was not previously executed.
1092  if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1093  return false;
1094 
1095  // Check that there is no implicit control flow in a block above.
1096  if (!IsSafeToSpeculativelyExecute && ICF->hasICF(TmpBB))
1097  return false;
1098  }
1099 
1100  assert(TmpBB);
1101  LoadBB = TmpBB;
1102 
1103  // Check to see how many predecessors have the loaded value fully
1104  // available.
1106  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1107  for (const AvailableValueInBlock &AV : ValuesPerBlock)
1108  FullyAvailableBlocks[AV.BB] = true;
1109  for (BasicBlock *UnavailableBB : UnavailableBlocks)
1110  FullyAvailableBlocks[UnavailableBB] = false;
1111 
1112  SmallVector<BasicBlock *, 4> CriticalEdgePred;
1113  for (BasicBlock *Pred : predecessors(LoadBB)) {
1114  // If any predecessor block is an EH pad that does not allow non-PHI
1115  // instructions before the terminator, we can't PRE the load.
1116  if (Pred->getTerminator()->isEHPad()) {
1117  LLVM_DEBUG(
1118  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1119  << Pred->getName() << "': " << *LI << '\n');
1120  return false;
1121  }
1122 
1123  if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1124  continue;
1125  }
1126 
1127  if (Pred->getTerminator()->getNumSuccessors() != 1) {
1128  if (isa<IndirectBrInst>(Pred->getTerminator())) {
1129  LLVM_DEBUG(
1130  dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1131  << Pred->getName() << "': " << *LI << '\n');
1132  return false;
1133  }
1134 
1135  if (LoadBB->isEHPad()) {
1136  LLVM_DEBUG(
1137  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1138  << Pred->getName() << "': " << *LI << '\n');
1139  return false;
1140  }
1141 
1142  CriticalEdgePred.push_back(Pred);
1143  } else {
1144  // Only add the predecessors that will not be split for now.
1145  PredLoads[Pred] = nullptr;
1146  }
1147  }
1148 
1149  // Decide whether PRE is profitable for this load.
1150  unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1151  assert(NumUnavailablePreds != 0 &&
1152  "Fully available value should already be eliminated!");
1153 
1154  // If this load is unavailable in multiple predecessors, reject it.
1155  // FIXME: If we could restructure the CFG, we could make a common pred with
1156  // all the preds that don't have an available LI and insert a new load into
1157  // that one block.
1158  if (NumUnavailablePreds != 1)
1159  return false;
1160 
1161  // Split critical edges, and update the unavailable predecessors accordingly.
1162  for (BasicBlock *OrigPred : CriticalEdgePred) {
1163  BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1164  assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1165  PredLoads[NewPred] = nullptr;
1166  LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1167  << LoadBB->getName() << '\n');
1168  }
1169 
1170  // Check if the load can safely be moved to all the unavailable predecessors.
1171  bool CanDoPRE = true;
1172  const DataLayout &DL = LI->getModule()->getDataLayout();
1174  for (auto &PredLoad : PredLoads) {
1175  BasicBlock *UnavailablePred = PredLoad.first;
1176 
1177  // Do PHI translation to get its value in the predecessor if necessary. The
1178  // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1179 
1180  // If all preds have a single successor, then we know it is safe to insert
1181  // the load on the pred (?!?), so we can insert code to materialize the
1182  // pointer if it is not available.
1183  PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1184  Value *LoadPtr = nullptr;
1185  LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1186  *DT, NewInsts);
1187 
1188  // If we couldn't find or insert a computation of this phi translated value,
1189  // we fail PRE.
1190  if (!LoadPtr) {
1191  LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1192  << *LI->getPointerOperand() << "\n");
1193  CanDoPRE = false;
1194  break;
1195  }
1196 
1197  PredLoad.second = LoadPtr;
1198  }
1199 
1200  if (!CanDoPRE) {
1201  while (!NewInsts.empty()) {
1202  Instruction *I = NewInsts.pop_back_val();
1203  markInstructionForDeletion(I);
1204  }
1205  // HINT: Don't revert the edge-splitting as following transformation may
1206  // also need to split these critical edges.
1207  return !CriticalEdgePred.empty();
1208  }
1209 
1210  // Okay, we can eliminate this load by inserting a reload in the predecessor
1211  // and using PHI construction to get the value in the other predecessors, do
1212  // it.
1213  LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1214  LLVM_DEBUG(if (!NewInsts.empty()) dbgs()
1215  << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back()
1216  << '\n');
1217 
1218  // Assign value numbers to the new instructions.
1219  for (Instruction *I : NewInsts) {
1220  // Instructions that have been inserted in predecessor(s) to materialize
1221  // the load address do not retain their original debug locations. Doing
1222  // so could lead to confusing (but correct) source attributions.
1223  // FIXME: How do we retain source locations without causing poor debugging
1224  // behavior?
1225  I->setDebugLoc(DebugLoc());
1226 
1227  // FIXME: We really _ought_ to insert these value numbers into their
1228  // parent's availability map. However, in doing so, we risk getting into
1229  // ordering issues. If a block hasn't been processed yet, we would be
1230  // marking a value as AVAIL-IN, which isn't what we intend.
1231  VN.lookupOrAdd(I);
1232  }
1233 
1234  for (const auto &PredLoad : PredLoads) {
1235  BasicBlock *UnavailablePred = PredLoad.first;
1236  Value *LoadPtr = PredLoad.second;
1237 
1238  auto *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre",
1239  LI->isVolatile(), LI->getAlignment(),
1240  LI->getOrdering(), LI->getSyncScopeID(),
1241  UnavailablePred->getTerminator());
1242  NewLoad->setDebugLoc(LI->getDebugLoc());
1243 
1244  // Transfer the old load's AA tags to the new load.
1245  AAMDNodes Tags;
1246  LI->getAAMetadata(Tags);
1247  if (Tags)
1248  NewLoad->setAAMetadata(Tags);
1249 
1250  if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1251  NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1252  if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1253  NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1254  if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1255  NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1256 
1257  // We do not propagate the old load's debug location, because the new
1258  // load now lives in a different BB, and we want to avoid a jumpy line
1259  // table.
1260  // FIXME: How do we retain source locations without causing poor debugging
1261  // behavior?
1262 
1263  // Add the newly created load.
1264  ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1265  NewLoad));
1266  MD->invalidateCachedPointerInfo(LoadPtr);
1267  LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1268  }
1269 
1270  // Perform PHI construction.
1271  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1272  LI->replaceAllUsesWith(V);
1273  if (isa<PHINode>(V))
1274  V->takeName(LI);
1275  if (Instruction *I = dyn_cast<Instruction>(V))
1276  I->setDebugLoc(LI->getDebugLoc());
1277  if (V->getType()->isPtrOrPtrVectorTy())
1278  MD->invalidateCachedPointerInfo(V);
1279  markInstructionForDeletion(LI);
1280  ORE->emit([&]() {
1281  return OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1282  << "load eliminated by PRE";
1283  });
1284  ++NumPRELoad;
1285  return true;
1286 }
1287 
1290  using namespace ore;
1291 
1292  ORE->emit([&]() {
1293  return OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1294  << "load of type " << NV("Type", LI->getType()) << " eliminated"
1295  << setExtraArgs() << " in favor of "
1296  << NV("InfavorOfValue", AvailableValue);
1297  });
1298 }
1299 
1300 /// Attempt to eliminate a load whose dependencies are
1301 /// non-local by performing PHI construction.
1302 bool GVN::processNonLocalLoad(LoadInst *LI) {
1303  // non-local speculations are not allowed under asan.
1304  if (LI->getParent()->getParent()->hasFnAttribute(
1305  Attribute::SanitizeAddress) ||
1307  Attribute::SanitizeHWAddress))
1308  return false;
1309 
1310  // Step 1: Find the non-local dependencies of the load.
1311  LoadDepVect Deps;
1312  MD->getNonLocalPointerDependency(LI, Deps);
1313 
1314  // If we had to process more than one hundred blocks to find the
1315  // dependencies, this load isn't worth worrying about. Optimizing
1316  // it will be too expensive.
1317  unsigned NumDeps = Deps.size();
1318  if (NumDeps > MaxNumDeps)
1319  return false;
1320 
1321  // If we had a phi translation failure, we'll have a single entry which is a
1322  // clobber in the current block. Reject this early.
1323  if (NumDeps == 1 &&
1324  !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1325  LLVM_DEBUG(dbgs() << "GVN: non-local load "; LI->printAsOperand(dbgs());
1326  dbgs() << " has unknown dependencies\n";);
1327  return false;
1328  }
1329 
1330  // If this load follows a GEP, see if we can PRE the indices before analyzing.
1331  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1332  for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1333  OE = GEP->idx_end();
1334  OI != OE; ++OI)
1335  if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1336  performScalarPRE(I);
1337  }
1338 
1339  // Step 2: Analyze the availability of the load
1340  AvailValInBlkVect ValuesPerBlock;
1341  UnavailBlkVect UnavailableBlocks;
1342  AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1343 
1344  // If we have no predecessors that produce a known value for this load, exit
1345  // early.
1346  if (ValuesPerBlock.empty())
1347  return false;
1348 
1349  // Step 3: Eliminate fully redundancy.
1350  //
1351  // If all of the instructions we depend on produce a known value for this
1352  // load, then it is fully redundant and we can use PHI insertion to compute
1353  // its value. Insert PHIs and remove the fully redundant value now.
1354  if (UnavailableBlocks.empty()) {
1355  LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1356 
1357  // Perform PHI construction.
1358  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1359  LI->replaceAllUsesWith(V);
1360 
1361  if (isa<PHINode>(V))
1362  V->takeName(LI);
1363  if (Instruction *I = dyn_cast<Instruction>(V))
1364  // If instruction I has debug info, then we should not update it.
1365  // Also, if I has a null DebugLoc, then it is still potentially incorrect
1366  // to propagate LI's DebugLoc because LI may not post-dominate I.
1367  if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1368  I->setDebugLoc(LI->getDebugLoc());
1369  if (V->getType()->isPtrOrPtrVectorTy())
1370  MD->invalidateCachedPointerInfo(V);
1371  markInstructionForDeletion(LI);
1372  ++NumGVNLoad;
1373  reportLoadElim(LI, V, ORE);
1374  return true;
1375  }
1376 
1377  // Step 4: Eliminate partial redundancy.
1378  if (!EnablePRE || !EnableLoadPRE)
1379  return false;
1380 
1381  return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1382 }
1383 
1384 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1385  assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1386  "This function can only be called with llvm.assume intrinsic");
1387  Value *V = IntrinsicI->getArgOperand(0);
1388 
1389  if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1390  if (Cond->isZero()) {
1391  Type *Int8Ty = Type::getInt8Ty(V->getContext());
1392  // Insert a new store to null instruction before the load to indicate that
1393  // this code is not reachable. FIXME: We could insert unreachable
1394  // instruction directly because we can modify the CFG.
1395  new StoreInst(UndefValue::get(Int8Ty),
1397  IntrinsicI);
1398  }
1399  markInstructionForDeletion(IntrinsicI);
1400  return false;
1401  } else if (isa<Constant>(V)) {
1402  // If it's not false, and constant, it must evaluate to true. This means our
1403  // assume is assume(true), and thus, pointless, and we don't want to do
1404  // anything more here.
1405  return false;
1406  }
1407 
1408  Constant *True = ConstantInt::getTrue(V->getContext());
1409  bool Changed = false;
1410 
1411  for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1412  BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1413 
1414  // This property is only true in dominated successors, propagateEquality
1415  // will check dominance for us.
1416  Changed |= propagateEquality(V, True, Edge, false);
1417  }
1418 
1419  // We can replace assume value with true, which covers cases like this:
1420  // call void @llvm.assume(i1 %cmp)
1421  // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1422  ReplaceWithConstMap[V] = True;
1423 
1424  // If one of *cmp *eq operand is const, adding it to map will cover this:
1425  // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1426  // call void @llvm.assume(i1 %cmp)
1427  // ret float %0 ; will change it to ret float 3.000000e+00
1428  if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1429  if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1430  CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1431  (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1432  CmpI->getFastMathFlags().noNaNs())) {
1433  Value *CmpLHS = CmpI->getOperand(0);
1434  Value *CmpRHS = CmpI->getOperand(1);
1435  if (isa<Constant>(CmpLHS))
1436  std::swap(CmpLHS, CmpRHS);
1437  auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1438 
1439  // If only one operand is constant.
1440  if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1441  ReplaceWithConstMap[CmpLHS] = RHSConst;
1442  }
1443  }
1444  return Changed;
1445 }
1446 
1448  patchReplacementInstruction(I, Repl);
1449  I->replaceAllUsesWith(Repl);
1450 }
1451 
1452 /// Attempt to eliminate a load, first by eliminating it
1453 /// locally, and then attempting non-local elimination if that fails.
1454 bool GVN::processLoad(LoadInst *L) {
1455  if (!MD)
1456  return false;
1457 
1458  // This code hasn't been audited for ordered or volatile memory access
1459  if (!L->isUnordered())
1460  return false;
1461 
1462  if (L->use_empty()) {
1463  markInstructionForDeletion(L);
1464  return true;
1465  }
1466 
1467  // ... to a pointer that has been loaded from before...
1468  MemDepResult Dep = MD->getDependency(L);
1469 
1470  // If it is defined in another block, try harder.
1471  if (Dep.isNonLocal())
1472  return processNonLocalLoad(L);
1473 
1474  // Only handle the local case below
1475  if (!Dep.isDef() && !Dep.isClobber()) {
1476  // This might be a NonFuncLocal or an Unknown
1477  LLVM_DEBUG(
1478  // fast print dep, using operator<< on instruction is too slow.
1479  dbgs() << "GVN: load "; L->printAsOperand(dbgs());
1480  dbgs() << " has unknown dependence\n";);
1481  return false;
1482  }
1483 
1484  AvailableValue AV;
1485  if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1486  Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1487 
1488  // Replace the load!
1489  patchAndReplaceAllUsesWith(L, AvailableValue);
1490  markInstructionForDeletion(L);
1491  ++NumGVNLoad;
1492  reportLoadElim(L, AvailableValue, ORE);
1493  // Tell MDA to rexamine the reused pointer since we might have more
1494  // information after forwarding it.
1495  if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1496  MD->invalidateCachedPointerInfo(AvailableValue);
1497  return true;
1498  }
1499 
1500  return false;
1501 }
1502 
1503 /// Return a pair the first field showing the value number of \p Exp and the
1504 /// second field showing whether it is a value number newly created.
1505 std::pair<uint32_t, bool>
1506 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1507  uint32_t &e = expressionNumbering[Exp];
1508  bool CreateNewValNum = !e;
1509  if (CreateNewValNum) {
1510  Expressions.push_back(Exp);
1511  if (ExprIdx.size() < nextValueNumber + 1)
1512  ExprIdx.resize(nextValueNumber * 2);
1513  e = nextValueNumber;
1514  ExprIdx[nextValueNumber++] = nextExprNumber++;
1515  }
1516  return {e, CreateNewValNum};
1517 }
1518 
1519 /// Return whether all the values related with the same \p num are
1520 /// defined in \p BB.
1521 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1522  GVN &Gvn) {
1523  LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1524  while (Vals && Vals->BB == BB)
1525  Vals = Vals->Next;
1526  return !Vals;
1527 }
1528 
1529 /// Wrap phiTranslateImpl to provide caching functionality.
1531  const BasicBlock *PhiBlock, uint32_t Num,
1532  GVN &Gvn) {
1533  auto FindRes = PhiTranslateTable.find({Num, Pred});
1534  if (FindRes != PhiTranslateTable.end())
1535  return FindRes->second;
1536  uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1537  PhiTranslateTable.insert({{Num, Pred}, NewNum});
1538  return NewNum;
1539 }
1540 
1541 /// Translate value number \p Num using phis, so that it has the values of
1542 /// the phis in BB.
1543 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1544  const BasicBlock *PhiBlock,
1545  uint32_t Num, GVN &Gvn) {
1546  if (PHINode *PN = NumberingPhi[Num]) {
1547  for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1548  if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1549  if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1550  return TransVal;
1551  }
1552  return Num;
1553  }
1554 
1555  // If there is any value related with Num is defined in a BB other than
1556  // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1557  // a backedge. We can do an early exit in that case to save compile time.
1558  if (!areAllValsInBB(Num, PhiBlock, Gvn))
1559  return Num;
1560 
1561  if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1562  return Num;
1563  Expression Exp = Expressions[ExprIdx[Num]];
1564 
1565  for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1566  // For InsertValue and ExtractValue, some varargs are index numbers
1567  // instead of value numbers. Those index numbers should not be
1568  // translated.
1569  if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1570  (i > 0 && Exp.opcode == Instruction::ExtractValue))
1571  continue;
1572  Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1573  }
1574 
1575  if (Exp.commutative) {
1576  assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1577  if (Exp.varargs[0] > Exp.varargs[1]) {
1578  std::swap(Exp.varargs[0], Exp.varargs[1]);
1579  uint32_t Opcode = Exp.opcode >> 8;
1580  if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1581  Exp.opcode = (Opcode << 8) |
1583  static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1584  }
1585  }
1586 
1587  if (uint32_t NewNum = expressionNumbering[Exp])
1588  return NewNum;
1589  return Num;
1590 }
1591 
1592 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1593 /// again.
1595  const BasicBlock &CurrBlock) {
1596  for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1597  auto FindRes = PhiTranslateTable.find({Num, Pred});
1598  if (FindRes != PhiTranslateTable.end())
1599  PhiTranslateTable.erase(FindRes);
1600  }
1601 }
1602 
1603 // In order to find a leader for a given value number at a
1604 // specific basic block, we first obtain the list of all Values for that number,
1605 // and then scan the list to find one whose block dominates the block in
1606 // question. This is fast because dominator tree queries consist of only
1607 // a few comparisons of DFS numbers.
1608 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1609  LeaderTableEntry Vals = LeaderTable[num];
1610  if (!Vals.Val) return nullptr;
1611 
1612  Value *Val = nullptr;
1613  if (DT->dominates(Vals.BB, BB)) {
1614  Val = Vals.Val;
1615  if (isa<Constant>(Val)) return Val;
1616  }
1617 
1618  LeaderTableEntry* Next = Vals.Next;
1619  while (Next) {
1620  if (DT->dominates(Next->BB, BB)) {
1621  if (isa<Constant>(Next->Val)) return Next->Val;
1622  if (!Val) Val = Next->Val;
1623  }
1624 
1625  Next = Next->Next;
1626  }
1627 
1628  return Val;
1629 }
1630 
1631 /// There is an edge from 'Src' to 'Dst'. Return
1632 /// true if every path from the entry block to 'Dst' passes via this edge. In
1633 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1635  DominatorTree *DT) {
1636  // While in theory it is interesting to consider the case in which Dst has
1637  // more than one predecessor, because Dst might be part of a loop which is
1638  // only reachable from Src, in practice it is pointless since at the time
1639  // GVN runs all such loops have preheaders, which means that Dst will have
1640  // been changed to have only one predecessor, namely Src.
1641  const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1642  assert((!Pred || Pred == E.getStart()) &&
1643  "No edge between these basic blocks!");
1644  return Pred != nullptr;
1645 }
1646 
1647 void GVN::assignBlockRPONumber(Function &F) {
1648  uint32_t NextBlockNumber = 1;
1650  for (BasicBlock *BB : RPOT)
1651  BlockRPONumber[BB] = NextBlockNumber++;
1652 }
1653 
1654 // Tries to replace instruction with const, using information from
1655 // ReplaceWithConstMap.
1656 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1657  bool Changed = false;
1658  for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1659  Value *Operand = Instr->getOperand(OpNum);
1660  auto it = ReplaceWithConstMap.find(Operand);
1661  if (it != ReplaceWithConstMap.end()) {
1662  assert(!isa<Constant>(Operand) &&
1663  "Replacing constants with constants is invalid");
1664  LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "
1665  << *it->second << " in instruction " << *Instr << '\n');
1666  Instr->setOperand(OpNum, it->second);
1667  Changed = true;
1668  }
1669  }
1670  return Changed;
1671 }
1672 
1673 /// The given values are known to be equal in every block
1674 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1675 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1676 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1677 /// value starting from the end of Root.Start.
1678 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1679  bool DominatesByEdge) {
1681  Worklist.push_back(std::make_pair(LHS, RHS));
1682  bool Changed = false;
1683  // For speed, compute a conservative fast approximation to
1684  // DT->dominates(Root, Root.getEnd());
1685  const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1686 
1687  while (!Worklist.empty()) {
1688  std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1689  LHS = Item.first; RHS = Item.second;
1690 
1691  if (LHS == RHS)
1692  continue;
1693  assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1694 
1695  // Don't try to propagate equalities between constants.
1696  if (isa<Constant>(LHS) && isa<Constant>(RHS))
1697  continue;
1698 
1699  // Prefer a constant on the right-hand side, or an Argument if no constants.
1700  if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1701  std::swap(LHS, RHS);
1702  assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1703 
1704  // If there is no obvious reason to prefer the left-hand side over the
1705  // right-hand side, ensure the longest lived term is on the right-hand side,
1706  // so the shortest lived term will be replaced by the longest lived.
1707  // This tends to expose more simplifications.
1708  uint32_t LVN = VN.lookupOrAdd(LHS);
1709  if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1710  (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1711  // Move the 'oldest' value to the right-hand side, using the value number
1712  // as a proxy for age.
1713  uint32_t RVN = VN.lookupOrAdd(RHS);
1714  if (LVN < RVN) {
1715  std::swap(LHS, RHS);
1716  LVN = RVN;
1717  }
1718  }
1719 
1720  // If value numbering later sees that an instruction in the scope is equal
1721  // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1722  // the invariant that instructions only occur in the leader table for their
1723  // own value number (this is used by removeFromLeaderTable), do not do this
1724  // if RHS is an instruction (if an instruction in the scope is morphed into
1725  // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1726  // using the leader table is about compiling faster, not optimizing better).
1727  // The leader table only tracks basic blocks, not edges. Only add to if we
1728  // have the simple case where the edge dominates the end.
1729  if (RootDominatesEnd && !isa<Instruction>(RHS))
1730  addToLeaderTable(LVN, RHS, Root.getEnd());
1731 
1732  // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1733  // LHS always has at least one use that is not dominated by Root, this will
1734  // never do anything if LHS has only one use.
1735  if (!LHS->hasOneUse()) {
1736  unsigned NumReplacements =
1737  DominatesByEdge
1738  ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1739  : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1740 
1741  Changed |= NumReplacements > 0;
1742  NumGVNEqProp += NumReplacements;
1743  // Cached information for anything that uses LHS will be invalid.
1744  if (MD)
1745  MD->invalidateCachedPointerInfo(LHS);
1746  }
1747 
1748  // Now try to deduce additional equalities from this one. For example, if
1749  // the known equality was "(A != B)" == "false" then it follows that A and B
1750  // are equal in the scope. Only boolean equalities with an explicit true or
1751  // false RHS are currently supported.
1752  if (!RHS->getType()->isIntegerTy(1))
1753  // Not a boolean equality - bail out.
1754  continue;
1755  ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1756  if (!CI)
1757  // RHS neither 'true' nor 'false' - bail out.
1758  continue;
1759  // Whether RHS equals 'true'. Otherwise it equals 'false'.
1760  bool isKnownTrue = CI->isMinusOne();
1761  bool isKnownFalse = !isKnownTrue;
1762 
1763  // If "A && B" is known true then both A and B are known true. If "A || B"
1764  // is known false then both A and B are known false.
1765  Value *A, *B;
1766  if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1767  (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1768  Worklist.push_back(std::make_pair(A, RHS));
1769  Worklist.push_back(std::make_pair(B, RHS));
1770  continue;
1771  }
1772 
1773  // If we are propagating an equality like "(A == B)" == "true" then also
1774  // propagate the equality A == B. When propagating a comparison such as
1775  // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1776  if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1777  Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1778 
1779  // If "A == B" is known true, or "A != B" is known false, then replace
1780  // A with B everywhere in the scope.
1781  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1782  (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1783  Worklist.push_back(std::make_pair(Op0, Op1));
1784 
1785  // Handle the floating point versions of equality comparisons too.
1786  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1787  (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1788 
1789  // Floating point -0.0 and 0.0 compare equal, so we can only
1790  // propagate values if we know that we have a constant and that
1791  // its value is non-zero.
1792 
1793  // FIXME: We should do this optimization if 'no signed zeros' is
1794  // applicable via an instruction-level fast-math-flag or some other
1795  // indicator that relaxed FP semantics are being used.
1796 
1797  if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1798  Worklist.push_back(std::make_pair(Op0, Op1));
1799  }
1800 
1801  // If "A >= B" is known true, replace "A < B" with false everywhere.
1802  CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1803  Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1804  // Since we don't have the instruction "A < B" immediately to hand, work
1805  // out the value number that it would have and use that to find an
1806  // appropriate instruction (if any).
1807  uint32_t NextNum = VN.getNextUnusedValueNumber();
1808  uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1809  // If the number we were assigned was brand new then there is no point in
1810  // looking for an instruction realizing it: there cannot be one!
1811  if (Num < NextNum) {
1812  Value *NotCmp = findLeader(Root.getEnd(), Num);
1813  if (NotCmp && isa<Instruction>(NotCmp)) {
1814  unsigned NumReplacements =
1815  DominatesByEdge
1816  ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1817  : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1818  Root.getStart());
1819  Changed |= NumReplacements > 0;
1820  NumGVNEqProp += NumReplacements;
1821  // Cached information for anything that uses NotCmp will be invalid.
1822  if (MD)
1823  MD->invalidateCachedPointerInfo(NotCmp);
1824  }
1825  }
1826  // Ensure that any instruction in scope that gets the "A < B" value number
1827  // is replaced with false.
1828  // The leader table only tracks basic blocks, not edges. Only add to if we
1829  // have the simple case where the edge dominates the end.
1830  if (RootDominatesEnd)
1831  addToLeaderTable(Num, NotVal, Root.getEnd());
1832 
1833  continue;
1834  }
1835  }
1836 
1837  return Changed;
1838 }
1839 
1840 /// When calculating availability, handle an instruction
1841 /// by inserting it into the appropriate sets
1842 bool GVN::processInstruction(Instruction *I) {
1843  // Ignore dbg info intrinsics.
1844  if (isa<DbgInfoIntrinsic>(I))
1845  return false;
1846 
1847  // If the instruction can be easily simplified then do so now in preference
1848  // to value numbering it. Value numbering often exposes redundancies, for
1849  // example if it determines that %y is equal to %x then the instruction
1850  // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1851  const DataLayout &DL = I->getModule()->getDataLayout();
1852  if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1853  bool Changed = false;
1854  if (!I->use_empty()) {
1855  I->replaceAllUsesWith(V);
1856  Changed = true;
1857  }
1858  if (isInstructionTriviallyDead(I, TLI)) {
1859  markInstructionForDeletion(I);
1860  Changed = true;
1861  }
1862  if (Changed) {
1863  if (MD && V->getType()->isPtrOrPtrVectorTy())
1864  MD->invalidateCachedPointerInfo(V);
1865  ++NumGVNSimpl;
1866  return true;
1867  }
1868  }
1869 
1870  if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1871  if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1872  return processAssumeIntrinsic(IntrinsicI);
1873 
1874  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1875  if (processLoad(LI))
1876  return true;
1877 
1878  unsigned Num = VN.lookupOrAdd(LI);
1879  addToLeaderTable(Num, LI, LI->getParent());
1880  return false;
1881  }
1882 
1883  // For conditional branches, we can perform simple conditional propagation on
1884  // the condition value itself.
1885  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1886  if (!BI->isConditional())
1887  return false;
1888 
1889  if (isa<Constant>(BI->getCondition()))
1890  return processFoldableCondBr(BI);
1891 
1892  Value *BranchCond = BI->getCondition();
1893  BasicBlock *TrueSucc = BI->getSuccessor(0);
1894  BasicBlock *FalseSucc = BI->getSuccessor(1);
1895  // Avoid multiple edges early.
1896  if (TrueSucc == FalseSucc)
1897  return false;
1898 
1899  BasicBlock *Parent = BI->getParent();
1900  bool Changed = false;
1901 
1902  Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1903  BasicBlockEdge TrueE(Parent, TrueSucc);
1904  Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1905 
1906  Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1907  BasicBlockEdge FalseE(Parent, FalseSucc);
1908  Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1909 
1910  return Changed;
1911  }
1912 
1913  // For switches, propagate the case values into the case destinations.
1914  if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1915  Value *SwitchCond = SI->getCondition();
1916  BasicBlock *Parent = SI->getParent();
1917  bool Changed = false;
1918 
1919  // Remember how many outgoing edges there are to every successor.
1921  for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1922  ++SwitchEdges[SI->getSuccessor(i)];
1923 
1924  for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1925  i != e; ++i) {
1926  BasicBlock *Dst = i->getCaseSuccessor();
1927  // If there is only a single edge, propagate the case value into it.
1928  if (SwitchEdges.lookup(Dst) == 1) {
1929  BasicBlockEdge E(Parent, Dst);
1930  Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1931  }
1932  }
1933  return Changed;
1934  }
1935 
1936  // Instructions with void type don't return a value, so there's
1937  // no point in trying to find redundancies in them.
1938  if (I->getType()->isVoidTy())
1939  return false;
1940 
1941  uint32_t NextNum = VN.getNextUnusedValueNumber();
1942  unsigned Num = VN.lookupOrAdd(I);
1943 
1944  // Allocations are always uniquely numbered, so we can save time and memory
1945  // by fast failing them.
1946  if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) {
1947  addToLeaderTable(Num, I, I->getParent());
1948  return false;
1949  }
1950 
1951  // If the number we were assigned was a brand new VN, then we don't
1952  // need to do a lookup to see if the number already exists
1953  // somewhere in the domtree: it can't!
1954  if (Num >= NextNum) {
1955  addToLeaderTable(Num, I, I->getParent());
1956  return false;
1957  }
1958 
1959  // Perform fast-path value-number based elimination of values inherited from
1960  // dominators.
1961  Value *Repl = findLeader(I->getParent(), Num);
1962  if (!Repl) {
1963  // Failure, just remember this instance for future use.
1964  addToLeaderTable(Num, I, I->getParent());
1965  return false;
1966  } else if (Repl == I) {
1967  // If I was the result of a shortcut PRE, it might already be in the table
1968  // and the best replacement for itself. Nothing to do.
1969  return false;
1970  }
1971 
1972  // Remove it!
1973  patchAndReplaceAllUsesWith(I, Repl);
1974  if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1975  MD->invalidateCachedPointerInfo(Repl);
1976  markInstructionForDeletion(I);
1977  return true;
1978 }
1979 
1980 /// runOnFunction - This is the main transformation entry point for a function.
1981 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1982  const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1983  MemoryDependenceResults *RunMD, LoopInfo *LI,
1984  OptimizationRemarkEmitter *RunORE) {
1985  AC = &RunAC;
1986  DT = &RunDT;
1987  VN.setDomTree(DT);
1988  TLI = &RunTLI;
1989  VN.setAliasAnalysis(&RunAA);
1990  MD = RunMD;
1991  ImplicitControlFlowTracking ImplicitCFT(DT);
1992  ICF = &ImplicitCFT;
1993  VN.setMemDep(MD);
1994  ORE = RunORE;
1995 
1996  bool Changed = false;
1997  bool ShouldContinue = true;
1998 
2000  // Merge unconditional branches, allowing PRE to catch more
2001  // optimization opportunities.
2002  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2003  BasicBlock *BB = &*FI++;
2004 
2005  bool removedBlock = MergeBlockIntoPredecessor(BB, &DTU, LI, nullptr, MD);
2006  if (removedBlock)
2007  ++NumGVNBlocks;
2008 
2009  Changed |= removedBlock;
2010  }
2011 
2012  unsigned Iteration = 0;
2013  while (ShouldContinue) {
2014  LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2015  ShouldContinue = iterateOnFunction(F);
2016  Changed |= ShouldContinue;
2017  ++Iteration;
2018  }
2019 
2020  if (EnablePRE) {
2021  // Fabricate val-num for dead-code in order to suppress assertion in
2022  // performPRE().
2023  assignValNumForDeadCode();
2024  assignBlockRPONumber(F);
2025  bool PREChanged = true;
2026  while (PREChanged) {
2027  PREChanged = performPRE(F);
2028  Changed |= PREChanged;
2029  }
2030  }
2031 
2032  // FIXME: Should perform GVN again after PRE does something. PRE can move
2033  // computations into blocks where they become fully redundant. Note that
2034  // we can't do this until PRE's critical edge splitting updates memdep.
2035  // Actually, when this happens, we should just fully integrate PRE into GVN.
2036 
2037  cleanupGlobalSets();
2038  // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2039  // iteration.
2040  DeadBlocks.clear();
2041 
2042  return Changed;
2043 }
2044 
2045 bool GVN::processBlock(BasicBlock *BB) {
2046  // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2047  // (and incrementing BI before processing an instruction).
2048  assert(InstrsToErase.empty() &&
2049  "We expect InstrsToErase to be empty across iterations");
2050  if (DeadBlocks.count(BB))
2051  return false;
2052 
2053  // Clearing map before every BB because it can be used only for single BB.
2054  ReplaceWithConstMap.clear();
2055  bool ChangedFunction = false;
2056 
2057  for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2058  BI != BE;) {
2059  if (!ReplaceWithConstMap.empty())
2060  ChangedFunction |= replaceOperandsWithConsts(&*BI);
2061  ChangedFunction |= processInstruction(&*BI);
2062 
2063  if (InstrsToErase.empty()) {
2064  ++BI;
2065  continue;
2066  }
2067 
2068  // If we need some instructions deleted, do it now.
2069  NumGVNInstr += InstrsToErase.size();
2070 
2071  // Avoid iterator invalidation.
2072  bool AtStart = BI == BB->begin();
2073  if (!AtStart)
2074  --BI;
2075 
2076  for (auto *I : InstrsToErase) {
2077  assert(I->getParent() == BB && "Removing instruction from wrong block?");
2078  LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n');
2079  salvageDebugInfo(*I);
2080  if (MD) MD->removeInstruction(I);
2081  LLVM_DEBUG(verifyRemoved(I));
2082  I->eraseFromParent();
2083  }
2084 
2085  ICF->invalidateBlock(BB);
2086  InstrsToErase.clear();
2087 
2088  if (AtStart)
2089  BI = BB->begin();
2090  else
2091  ++BI;
2092  }
2093 
2094  return ChangedFunction;
2095 }
2096 
2097 // Instantiate an expression in a predecessor that lacked it.
2098 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2099  BasicBlock *Curr, unsigned int ValNo) {
2100  // Because we are going top-down through the block, all value numbers
2101  // will be available in the predecessor by the time we need them. Any
2102  // that weren't originally present will have been instantiated earlier
2103  // in this loop.
2104  bool success = true;
2105  for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2106  Value *Op = Instr->getOperand(i);
2107  if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2108  continue;
2109  // This could be a newly inserted instruction, in which case, we won't
2110  // find a value number, and should give up before we hurt ourselves.
2111  // FIXME: Rewrite the infrastructure to let it easier to value number
2112  // and process newly inserted instructions.
2113  if (!VN.exists(Op)) {
2114  success = false;
2115  break;
2116  }
2117  uint32_t TValNo =
2118  VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2119  if (Value *V = findLeader(Pred, TValNo)) {
2120  Instr->setOperand(i, V);
2121  } else {
2122  success = false;
2123  break;
2124  }
2125  }
2126 
2127  // Fail out if we encounter an operand that is not available in
2128  // the PRE predecessor. This is typically because of loads which
2129  // are not value numbered precisely.
2130  if (!success)
2131  return false;
2132 
2133  Instr->insertBefore(Pred->getTerminator());
2134  Instr->setName(Instr->getName() + ".pre");
2135  Instr->setDebugLoc(Instr->getDebugLoc());
2136 
2137  unsigned Num = VN.lookupOrAdd(Instr);
2138  VN.add(Instr, Num);
2139 
2140  // Update the availability map to include the new instruction.
2141  addToLeaderTable(Num, Instr, Pred);
2142  return true;
2143 }
2144 
2145 bool GVN::performScalarPRE(Instruction *CurInst) {
2146  if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() ||
2147  isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2148  CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2149  isa<DbgInfoIntrinsic>(CurInst))
2150  return false;
2151 
2152  // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2153  // sinking the compare again, and it would force the code generator to
2154  // move the i1 from processor flags or predicate registers into a general
2155  // purpose register.
2156  if (isa<CmpInst>(CurInst))
2157  return false;
2158 
2159  // We don't currently value number ANY inline asm calls.
2160  if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2161  if (CallI->isInlineAsm())
2162  return false;
2163 
2164  uint32_t ValNo = VN.lookup(CurInst);
2165 
2166  // Look for the predecessors for PRE opportunities. We're
2167  // only trying to solve the basic diamond case, where
2168  // a value is computed in the successor and one predecessor,
2169  // but not the other. We also explicitly disallow cases
2170  // where the successor is its own predecessor, because they're
2171  // more complicated to get right.
2172  unsigned NumWith = 0;
2173  unsigned NumWithout = 0;
2174  BasicBlock *PREPred = nullptr;
2175  BasicBlock *CurrentBlock = CurInst->getParent();
2176 
2178  for (BasicBlock *P : predecessors(CurrentBlock)) {
2179  // We're not interested in PRE where blocks with predecessors that are
2180  // not reachable.
2181  if (!DT->isReachableFromEntry(P)) {
2182  NumWithout = 2;
2183  break;
2184  }
2185  // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2186  // when CurInst has operand defined in CurrentBlock (so it may be defined
2187  // by phi in the loop header).
2188  if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2189  llvm::any_of(CurInst->operands(), [&](const Use &U) {
2190  if (auto *Inst = dyn_cast<Instruction>(U.get()))
2191  return Inst->getParent() == CurrentBlock;
2192  return false;
2193  })) {
2194  NumWithout = 2;
2195  break;
2196  }
2197 
2198  uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2199  Value *predV = findLeader(P, TValNo);
2200  if (!predV) {
2201  predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2202  PREPred = P;
2203  ++NumWithout;
2204  } else if (predV == CurInst) {
2205  /* CurInst dominates this predecessor. */
2206  NumWithout = 2;
2207  break;
2208  } else {
2209  predMap.push_back(std::make_pair(predV, P));
2210  ++NumWith;
2211  }
2212  }
2213 
2214  // Don't do PRE when it might increase code size, i.e. when
2215  // we would need to insert instructions in more than one pred.
2216  if (NumWithout > 1 || NumWith == 0)
2217  return false;
2218 
2219  // We may have a case where all predecessors have the instruction,
2220  // and we just need to insert a phi node. Otherwise, perform
2221  // insertion.
2222  Instruction *PREInstr = nullptr;
2223 
2224  if (NumWithout != 0) {
2225  if (!isSafeToSpeculativelyExecute(CurInst)) {
2226  // It is only valid to insert a new instruction if the current instruction
2227  // is always executed. An instruction with implicit control flow could
2228  // prevent us from doing it. If we cannot speculate the execution, then
2229  // PRE should be prohibited.
2230  if (ICF->isDominatedByICFIFromSameBlock(CurInst))
2231  return false;
2232  }
2233 
2234  // Don't do PRE across indirect branch.
2235  if (isa<IndirectBrInst>(PREPred->getTerminator()))
2236  return false;
2237 
2238  // We can't do PRE safely on a critical edge, so instead we schedule
2239  // the edge to be split and perform the PRE the next time we iterate
2240  // on the function.
2241  unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2242  if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2243  toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2244  return false;
2245  }
2246  // We need to insert somewhere, so let's give it a shot
2247  PREInstr = CurInst->clone();
2248  if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2249  // If we failed insertion, make sure we remove the instruction.
2250  LLVM_DEBUG(verifyRemoved(PREInstr));
2251  PREInstr->deleteValue();
2252  return false;
2253  }
2254  }
2255 
2256  // Either we should have filled in the PRE instruction, or we should
2257  // not have needed insertions.
2258  assert(PREInstr != nullptr || NumWithout == 0);
2259 
2260  ++NumGVNPRE;
2261 
2262  // Create a PHI to make the value available in this block.
2263  PHINode *Phi =
2264  PHINode::Create(CurInst->getType(), predMap.size(),
2265  CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2266  for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2267  if (Value *V = predMap[i].first) {
2268  // If we use an existing value in this phi, we have to patch the original
2269  // value because the phi will be used to replace a later value.
2270  patchReplacementInstruction(CurInst, V);
2271  Phi->addIncoming(V, predMap[i].second);
2272  } else
2273  Phi->addIncoming(PREInstr, PREPred);
2274  }
2275 
2276  VN.add(Phi, ValNo);
2277  // After creating a new PHI for ValNo, the phi translate result for ValNo will
2278  // be changed, so erase the related stale entries in phi translate cache.
2279  VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2280  addToLeaderTable(ValNo, Phi, CurrentBlock);
2281  Phi->setDebugLoc(CurInst->getDebugLoc());
2282  CurInst->replaceAllUsesWith(Phi);
2283  if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2284  MD->invalidateCachedPointerInfo(Phi);
2285  VN.erase(CurInst);
2286  removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2287 
2288  LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2289  if (MD)
2290  MD->removeInstruction(CurInst);
2291  LLVM_DEBUG(verifyRemoved(CurInst));
2292  // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes
2293  // some assertion failures.
2294  ICF->invalidateBlock(CurrentBlock);
2295  CurInst->eraseFromParent();
2296  ++NumGVNInstr;
2297 
2298  return true;
2299 }
2300 
2301 /// Perform a purely local form of PRE that looks for diamond
2302 /// control flow patterns and attempts to perform simple PRE at the join point.
2303 bool GVN::performPRE(Function &F) {
2304  bool Changed = false;
2305  for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2306  // Nothing to PRE in the entry block.
2307  if (CurrentBlock == &F.getEntryBlock())
2308  continue;
2309 
2310  // Don't perform PRE on an EH pad.
2311  if (CurrentBlock->isEHPad())
2312  continue;
2313 
2314  for (BasicBlock::iterator BI = CurrentBlock->begin(),
2315  BE = CurrentBlock->end();
2316  BI != BE;) {
2317  Instruction *CurInst = &*BI++;
2318  Changed |= performScalarPRE(CurInst);
2319  }
2320  }
2321 
2322  if (splitCriticalEdges())
2323  Changed = true;
2324 
2325  return Changed;
2326 }
2327 
2328 /// Split the critical edge connecting the given two blocks, and return
2329 /// the block inserted to the critical edge.
2330 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2331  BasicBlock *BB =
2333  if (MD)
2334  MD->invalidateCachedPredecessors();
2335  return BB;
2336 }
2337 
2338 /// Split critical edges found during the previous
2339 /// iteration that may enable further optimization.
2340 bool GVN::splitCriticalEdges() {
2341  if (toSplit.empty())
2342  return false;
2343  do {
2344  std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val();
2345  SplitCriticalEdge(Edge.first, Edge.second,
2347  } while (!toSplit.empty());
2348  if (MD) MD->invalidateCachedPredecessors();
2349  return true;
2350 }
2351 
2352 /// Executes one iteration of GVN
2353 bool GVN::iterateOnFunction(Function &F) {
2354  cleanupGlobalSets();
2355 
2356  // Top-down walk of the dominator tree
2357  bool Changed = false;
2358  // Needed for value numbering with phi construction to work.
2359  // RPOT walks the graph in its constructor and will not be invalidated during
2360  // processBlock.
2362 
2363  for (BasicBlock *BB : RPOT)
2364  Changed |= processBlock(BB);
2365 
2366  return Changed;
2367 }
2368 
2369 void GVN::cleanupGlobalSets() {
2370  VN.clear();
2371  LeaderTable.clear();
2372  BlockRPONumber.clear();
2373  TableAllocator.Reset();
2374  ICF->clear();
2375 }
2376 
2377 /// Verify that the specified instruction does not occur in our
2378 /// internal data structures.
2379 void GVN::verifyRemoved(const Instruction *Inst) const {
2380  VN.verifyRemoved(Inst);
2381 
2382  // Walk through the value number scope to make sure the instruction isn't
2383  // ferreted away in it.
2385  I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2386  const LeaderTableEntry *Node = &I->second;
2387  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2388 
2389  while (Node->Next) {
2390  Node = Node->Next;
2391  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2392  }
2393  }
2394 }
2395 
2396 /// BB is declared dead, which implied other blocks become dead as well. This
2397 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2398 /// live successors, update their phi nodes by replacing the operands
2399 /// corresponding to dead blocks with UndefVal.
2400 void GVN::addDeadBlock(BasicBlock *BB) {
2403 
2404  NewDead.push_back(BB);
2405  while (!NewDead.empty()) {
2406  BasicBlock *D = NewDead.pop_back_val();
2407  if (DeadBlocks.count(D))
2408  continue;
2409 
2410  // All blocks dominated by D are dead.
2412  DT->getDescendants(D, Dom);
2413  DeadBlocks.insert(Dom.begin(), Dom.end());
2414 
2415  // Figure out the dominance-frontier(D).
2416  for (BasicBlock *B : Dom) {
2417  for (BasicBlock *S : successors(B)) {
2418  if (DeadBlocks.count(S))
2419  continue;
2420 
2421  bool AllPredDead = true;
2422  for (BasicBlock *P : predecessors(S))
2423  if (!DeadBlocks.count(P)) {
2424  AllPredDead = false;
2425  break;
2426  }
2427 
2428  if (!AllPredDead) {
2429  // S could be proved dead later on. That is why we don't update phi
2430  // operands at this moment.
2431  DF.insert(S);
2432  } else {
2433  // While S is not dominated by D, it is dead by now. This could take
2434  // place if S already have a dead predecessor before D is declared
2435  // dead.
2436  NewDead.push_back(S);
2437  }
2438  }
2439  }
2440  }
2441 
2442  // For the dead blocks' live successors, update their phi nodes by replacing
2443  // the operands corresponding to dead blocks with UndefVal.
2445  I != E; I++) {
2446  BasicBlock *B = *I;
2447  if (DeadBlocks.count(B))
2448  continue;
2449 
2451  for (BasicBlock *P : Preds) {
2452  if (!DeadBlocks.count(P))
2453  continue;
2454 
2455  if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2456  if (BasicBlock *S = splitCriticalEdges(P, B))
2457  DeadBlocks.insert(P = S);
2458  }
2459 
2460  for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2461  PHINode &Phi = cast<PHINode>(*II);
2463  UndefValue::get(Phi.getType()));
2464  if (MD)
2465  MD->invalidateCachedPointerInfo(&Phi);
2466  }
2467  }
2468  }
2469 }
2470 
2471 // If the given branch is recognized as a foldable branch (i.e. conditional
2472 // branch with constant condition), it will perform following analyses and
2473 // transformation.
2474 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2475 // R be the target of the dead out-coming edge.
2476 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2477 // edge. The result of this step will be {X| X is dominated by R}
2478 // 2) Identify those blocks which haves at least one dead predecessor. The
2479 // result of this step will be dominance-frontier(R).
2480 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2481 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2482 //
2483 // Return true iff *NEW* dead code are found.
2484 bool GVN::processFoldableCondBr(BranchInst *BI) {
2485  if (!BI || BI->isUnconditional())
2486  return false;
2487 
2488  // If a branch has two identical successors, we cannot declare either dead.
2489  if (BI->getSuccessor(0) == BI->getSuccessor(1))
2490  return false;
2491 
2492  ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2493  if (!Cond)
2494  return false;
2495 
2496  BasicBlock *DeadRoot =
2497  Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2498  if (DeadBlocks.count(DeadRoot))
2499  return false;
2500 
2501  if (!DeadRoot->getSinglePredecessor())
2502  DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2503 
2504  addDeadBlock(DeadRoot);
2505  return true;
2506 }
2507 
2508 // performPRE() will trigger assert if it comes across an instruction without
2509 // associated val-num. As it normally has far more live instructions than dead
2510 // instructions, it makes more sense just to "fabricate" a val-number for the
2511 // dead code than checking if instruction involved is dead or not.
2512 void GVN::assignValNumForDeadCode() {
2513  for (BasicBlock *BB : DeadBlocks) {
2514  for (Instruction &Inst : *BB) {
2515  unsigned ValNum = VN.lookupOrAdd(&Inst);
2516  addToLeaderTable(ValNum, &Inst, BB);
2517  }
2518  }
2519 }
2520 
2522 public:
2523  static char ID; // Pass identification, replacement for typeid
2524 
2525  explicit GVNLegacyPass(bool NoMemDepAnalysis = !EnableMemDep)
2526  : FunctionPass(ID), NoMemDepAnalysis(NoMemDepAnalysis) {
2528  }
2529 
2530  bool runOnFunction(Function &F) override {
2531  if (skipFunction(F))
2532  return false;
2533 
2534  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2535 
2536  return Impl.runImpl(
2537  F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2538  getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2539  getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2540  getAnalysis<AAResultsWrapperPass>().getAAResults(),
2541  NoMemDepAnalysis ? nullptr
2542  : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2543  LIWP ? &LIWP->getLoopInfo() : nullptr,
2544  &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2545  }
2546 
2547  void getAnalysisUsage(AnalysisUsage &AU) const override {
2551  if (!NoMemDepAnalysis)
2554 
2559  }
2560 
2561 private:
2562  bool NoMemDepAnalysis;
2563  GVN Impl;
2564 };
2565 
2566 char GVNLegacyPass::ID = 0;
2567 
2568 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2576 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2577 
2578 // The public interface to this file...
2579 FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) {
2580  return new GVNLegacyPass(NoMemDepAnalysis);
2581 }
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:750
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:68
FunctionPass * createGVNPass(bool NoLoads=false)
Create a legacy GVN pass.
Definition: GVN.cpp:2579
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:1594
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:111
bool isUndefValue() const
Definition: GVN.cpp:212
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:584
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:631
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:39
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:225
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:61
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:884
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:770
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
PointerTy getPointer() const
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:54
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:588
iterator end()
Definition: Function.h:658
void AddAvailableValue(BasicBlock *BB, Value *V)
Indicate that a rewritten value is available in the specified block with the specified value...
Definition: SSAUpdater.cpp:72
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:1607
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:237
uint32_t phiTranslate(const BasicBlock *BB, const BasicBlock *PhiBlock, uint32_t Num, GVN &Gvn)
Wrap phiTranslateImpl to provide caching functionality.
Definition: GVN.cpp:1530
bool isTerminator() const
Definition: Instruction.h:129
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:657
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:99
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.h:321
unsigned second
This class implements a map that also provides access to all stored values in a deterministic order...
Definition: MapVector.h:38
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:34
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:231
F(f)
bool isCoercedLoadValue() const
Definition: GVN.cpp:210
An instruction for reading from memory.
Definition: Instructions.h:168
const BasicBlock * getEnd() const
Definition: Dominators.h:95
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:138
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:2443
Use * op_iterator
Definition: User.h:225
iterator end()
Get an iterator to the end of the SetVector.
Definition: SetVector.h:93
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:481
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:33
op_iterator op_begin()
Definition: User.h:230
gvn Early GVN Hoisting of Expressions
Definition: GVNHoist.cpp:1204
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:268
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:269
uint32_t lookup(Value *V, bool Verify=true) const
Returns the value number of the specified value.
Definition: GVN.cpp:575
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:4272
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
bool isVolatile() const
Return true if this is a load from a volatile memory location.
Definition: Instructions.h:221
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: GVN.cpp:2547
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:2375
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:364
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: GVN.cpp:2530
Option class for critical edge splitting.
int getBasicBlockIndex(const BasicBlock *BB) const
Return the first index of the specified basic block in the value list for this PHI.
void clear()
Remove all entries from the ValueTable.
Definition: GVN.cpp:596
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:56
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:652
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:257
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
MemoryDependenceResults & getMemDep() const
Definition: GVN.h:84
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
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:286
Analysis pass that exposes the LoopInfo for a function.
Definition: LoopInfo.h:939
#define LLVM_DUMP_METHOD
Definition: Compiler.h:74
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:103
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
ppc ctr loops PowerPC CTR Loops Verify
bool insert(const value_type &X)
Insert a new element into the SetVector.
Definition: SetVector.h:142
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:83
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:221
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:4431
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:126
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:310
bool isMinusOne() const
This function will return true iff every bit in this constant is set to true.
Definition: Constants.h:209
void add(Value *V, uint32_t num)
add - Insert a value into the table with a specified value number.
Definition: GVN.cpp:389
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:430
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:292
iterator begin()
Definition: Function.h:656
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:145
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Value * getOperand(unsigned i) const
Definition: User.h:170
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:106
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:208
void initializeGVNLegacyPassPass(PassRegistry &)
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:141
const BasicBlock & getEntryBlock() const
Definition: Function.h:640
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:843
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:419
Value * GetValueInMiddleOfBlock(BasicBlock *BB)
Construct SSA form, materializing a value that is live in the middle of the specified block...
Definition: SSAUpdater.cpp:100
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:149
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:154
* 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:308
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:236
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:74
LLVM Basic Block Representation.
Definition: BasicBlock.h:58
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:36
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
Definition: PatternMatch.h:756
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:746
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:129
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:281
A manager for alias analyses.
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:562
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:113
unsigned getNumIndices() const
bool isUnordered() const
Definition: Instructions.h:268
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:232
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:1055
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:641
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:285
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:116
op_range operands()
Definition: User.h:238
Value * getPointerOperand()
Definition: Instructions.h:274
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:369
static void reportLoadElim(LoadInst *LI, Value *AvailableValue, OptimizationRemarkEmitter *ORE)
Definition: GVN.cpp:1288
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1392
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:160
size_t size() const
Definition: SmallVector.h:53
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:4199
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:51
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:843
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:227
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:298
This is the common base class for memset/memcpy/memmove.
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:192
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:418
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
void emit(DiagnosticInfoOptimizationBase &OptDiag)
Output the remark via the diagnostic handler and to the optimization record file. ...
iterator end()
Definition: BasicBlock.h:271
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:249
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:143
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:48
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:644
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:381
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:621
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:125
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:786
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:577
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:478
void setOperand(unsigned i, Value *Val)
Definition: User.h:175
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:133
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:941
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:56
hash_code hash_combine(const Ts &...args)
Combine values into a single hash_code.
Definition: Hashing.h:601
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:400
hash_code hash_combine_range(InputIteratorT first, InputIteratorT last)
Compute a hash_code for a sequence of values.
Definition: Hashing.h:479
std::vector< NonLocalDepEntry > NonLocalDepInfo
An opaque object representing a hash code.
Definition: Hashing.h:72
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:478
void verifyRemoved(const Value *) const
verifyRemoved - Verify that the value is removed from all internal data structures.
Definition: GVN.cpp:618
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:394
void erase(Value *v)
Remove a value from the value numbering.
Definition: GVN.cpp:608
static bool isLifetimeStart(const Instruction *Inst)
Definition: GVN.cpp:835
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:546
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:133
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
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:311
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:230
Instruction * getInst() const
If this is a normal dependency, returns the instruction that is depended on.
void clear()
Definition: ilist.h:309
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:349
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:56
unsigned getNumArgOperands() const
Return the number of call arguments.
GVNLegacyPass(bool NoMemDepAnalysis=!EnableMemDep)
Definition: GVN.cpp:2525
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:215
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:107
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:249
#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:789
bool exists(Value *V) const
Returns true if a value number exists for the specified value.
Definition: GVN.cpp:497
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:323
idx_iterator idx_begin() const
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:175
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:501
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:91
Represents a particular available value that we know how to materialize.
Definition: GVN.cpp:162
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
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:673
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:644
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:348
LLVM Value Representation.
Definition: Value.h:73
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:264
OptimizationRemarkEmitter legacy analysis pass.
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: GVN.cpp:629
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:413
Predicate getSwappedPredicate() const
For example, EQ->EQ, SLE->SGE, ULT->UGT, OEQ->OEQ, ULE->UGE, OLT->OGT, etc.
Definition: InstrTypes.h:756
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:1447
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:260
bool isMemIntrinValue() const
Definition: GVN.cpp:211
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
This header defines various interfaces for pass management in LLVM.
void setIncomingValue(unsigned i, Value *V)
AvailableValue AV
AV - The actual available value.
Definition: GVN.cpp:242
#define LLVM_DEBUG(X)
Definition: Debug.h:123
Value * SimplifyInstruction(Instruction *I, const SimplifyQuery &Q, OptimizationRemarkEmitter *ORE=nullptr)
See if we can compute a simplified version of this instruction.
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:174
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:323
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:1634
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:67
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:63