LLVM  6.0.0svn
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"
41 #include "llvm/IR/Attributes.h"
42 #include "llvm/IR/BasicBlock.h"
43 #include "llvm/IR/CallSite.h"
44 #include "llvm/IR/Constant.h"
45 #include "llvm/IR/Constants.h"
46 #include "llvm/IR/DataLayout.h"
47 #include "llvm/IR/DebugLoc.h"
48 #include "llvm/IR/Dominators.h"
49 #include "llvm/IR/Function.h"
50 #include "llvm/IR/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/Intrinsics.h"
55 #include "llvm/IR/LLVMContext.h"
56 #include "llvm/IR/Metadata.h"
57 #include "llvm/IR/Module.h"
58 #include "llvm/IR/Operator.h"
59 #include "llvm/IR/PassManager.h"
60 #include "llvm/IR/PatternMatch.h"
61 #include "llvm/IR/Type.h"
62 #include "llvm/IR/Use.h"
63 #include "llvm/IR/Value.h"
64 #include "llvm/Pass.h"
65 #include "llvm/Support/Casting.h"
67 #include "llvm/Support/Compiler.h"
68 #include "llvm/Support/Debug.h"
74 #include <algorithm>
75 #include <cassert>
76 #include <cstdint>
77 #include <utility>
78 #include <vector>
79 
80 using namespace llvm;
81 using namespace llvm::gvn;
82 using namespace llvm::VNCoercion;
83 using namespace PatternMatch;
84 
85 #define DEBUG_TYPE "gvn"
86 
87 STATISTIC(NumGVNInstr, "Number of instructions deleted");
88 STATISTIC(NumGVNLoad, "Number of loads deleted");
89 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
90 STATISTIC(NumGVNBlocks, "Number of blocks merged");
91 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
92 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
93 STATISTIC(NumPRELoad, "Number of loads PRE'd");
94 
95 static cl::opt<bool> EnablePRE("enable-pre",
96  cl::init(true), cl::Hidden);
97 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
98 
99 // Maximum allowed recursion depth.
100 static cl::opt<uint32_t>
101 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
102  cl::desc("Max recurse depth (default = 1000)"));
103 
107  bool commutative = false;
109 
110  Expression(uint32_t o = ~2U) : opcode(o) {}
111 
112  bool operator==(const Expression &other) const {
113  if (opcode != other.opcode)
114  return false;
115  if (opcode == ~0U || opcode == ~1U)
116  return true;
117  if (type != other.type)
118  return false;
119  if (varargs != other.varargs)
120  return false;
121  return true;
122  }
123 
125  return hash_combine(
126  Value.opcode, Value.type,
127  hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
128  }
129 };
130 
131 namespace llvm {
132 
133 template <> struct DenseMapInfo<GVN::Expression> {
134  static inline GVN::Expression getEmptyKey() { return ~0U; }
135  static inline GVN::Expression getTombstoneKey() { return ~1U; }
136 
137  static unsigned getHashValue(const GVN::Expression &e) {
138  using llvm::hash_value;
139 
140  return static_cast<unsigned>(hash_value(e));
141  }
142 
143  static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
144  return LHS == RHS;
145  }
146 };
147 
148 } // end namespace llvm
149 
150 /// Represents a particular available value that we know how to materialize.
151 /// Materialization of an AvailableValue never fails. An AvailableValue is
152 /// implicitly associated with a rematerialization point which is the
153 /// location of the instruction from which it was formed.
155  enum ValType {
156  SimpleVal, // A simple offsetted value that is accessed.
157  LoadVal, // A value produced by a load.
158  MemIntrin, // A memory intrinsic which is loaded from.
159  UndefVal // A UndefValue representing a value from dead block (which
160  // is not yet physically removed from the CFG).
161  };
162 
163  /// V - The value that is live out of the block.
165 
166  /// Offset - The byte offset in Val that is interesting for the load query.
167  unsigned Offset;
168 
169  static AvailableValue get(Value *V, unsigned Offset = 0) {
170  AvailableValue Res;
171  Res.Val.setPointer(V);
172  Res.Val.setInt(SimpleVal);
173  Res.Offset = Offset;
174  return Res;
175  }
176 
177  static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
178  AvailableValue Res;
179  Res.Val.setPointer(MI);
180  Res.Val.setInt(MemIntrin);
181  Res.Offset = Offset;
182  return Res;
183  }
184 
185  static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
186  AvailableValue Res;
187  Res.Val.setPointer(LI);
188  Res.Val.setInt(LoadVal);
189  Res.Offset = Offset;
190  return Res;
191  }
192 
194  AvailableValue Res;
195  Res.Val.setPointer(nullptr);
196  Res.Val.setInt(UndefVal);
197  Res.Offset = 0;
198  return Res;
199  }
200 
201  bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
202  bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
203  bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
204  bool isUndefValue() const { return Val.getInt() == UndefVal; }
205 
207  assert(isSimpleValue() && "Wrong accessor");
208  return Val.getPointer();
209  }
210 
212  assert(isCoercedLoadValue() && "Wrong accessor");
213  return cast<LoadInst>(Val.getPointer());
214  }
215 
217  assert(isMemIntrinValue() && "Wrong accessor");
218  return cast<MemIntrinsic>(Val.getPointer());
219  }
220 
221  /// Emit code at the specified insertion point to adjust the value defined
222  /// here to the specified type. This handles various coercion cases.
223  Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
224  GVN &gvn) const;
225 };
226 
227 /// Represents an AvailableValue which can be rematerialized at the end of
228 /// the associated BasicBlock.
230  /// BB - The basic block in question.
232 
233  /// AV - The actual available value
235 
238  Res.BB = BB;
239  Res.AV = std::move(AV);
240  return Res;
241  }
242 
244  unsigned Offset = 0) {
245  return get(BB, AvailableValue::get(V, Offset));
246  }
247 
249  return get(BB, AvailableValue::getUndef());
250  }
251 
252  /// Emit code at the end of this block to adjust the value defined here to
253  /// the specified type. This handles various coercion cases.
255  return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
256  }
257 };
258 
259 //===----------------------------------------------------------------------===//
260 // ValueTable Internal Functions
261 //===----------------------------------------------------------------------===//
262 
263 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
264  Expression e;
265  e.type = I->getType();
266  e.opcode = I->getOpcode();
267  for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
268  OI != OE; ++OI)
269  e.varargs.push_back(lookupOrAdd(*OI));
270  if (I->isCommutative()) {
271  // Ensure that commutative instructions that only differ by a permutation
272  // of their operands get the same value number by sorting the operand value
273  // numbers. Since all commutative instructions have two operands it is more
274  // efficient to sort by hand rather than using, say, std::sort.
275  assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
276  if (e.varargs[0] > e.varargs[1])
277  std::swap(e.varargs[0], e.varargs[1]);
278  e.commutative = true;
279  }
280 
281  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
282  // Sort the operand value numbers so x<y and y>x get the same value number.
283  CmpInst::Predicate Predicate = C->getPredicate();
284  if (e.varargs[0] > e.varargs[1]) {
285  std::swap(e.varargs[0], e.varargs[1]);
286  Predicate = CmpInst::getSwappedPredicate(Predicate);
287  }
288  e.opcode = (C->getOpcode() << 8) | Predicate;
289  e.commutative = true;
290  } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
291  for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
292  II != IE; ++II)
293  e.varargs.push_back(*II);
294  }
295 
296  return e;
297 }
298 
299 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
301  Value *LHS, Value *RHS) {
302  assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
303  "Not a comparison!");
304  Expression e;
305  e.type = CmpInst::makeCmpResultType(LHS->getType());
306  e.varargs.push_back(lookupOrAdd(LHS));
307  e.varargs.push_back(lookupOrAdd(RHS));
308 
309  // Sort the operand value numbers so x<y and y>x get the same value number.
310  if (e.varargs[0] > e.varargs[1]) {
311  std::swap(e.varargs[0], e.varargs[1]);
312  Predicate = CmpInst::getSwappedPredicate(Predicate);
313  }
314  e.opcode = (Opcode << 8) | Predicate;
315  e.commutative = true;
316  return e;
317 }
318 
319 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
320  assert(EI && "Not an ExtractValueInst?");
321  Expression e;
322  e.type = EI->getType();
323  e.opcode = 0;
324 
326  if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
327  // EI might be an extract from one of our recognised intrinsics. If it
328  // is we'll synthesize a semantically equivalent expression instead on
329  // an extract value expression.
330  switch (I->getIntrinsicID()) {
331  case Intrinsic::sadd_with_overflow:
332  case Intrinsic::uadd_with_overflow:
333  e.opcode = Instruction::Add;
334  break;
335  case Intrinsic::ssub_with_overflow:
336  case Intrinsic::usub_with_overflow:
337  e.opcode = Instruction::Sub;
338  break;
339  case Intrinsic::smul_with_overflow:
340  case Intrinsic::umul_with_overflow:
341  e.opcode = Instruction::Mul;
342  break;
343  default:
344  break;
345  }
346 
347  if (e.opcode != 0) {
348  // Intrinsic recognized. Grab its args to finish building the expression.
349  assert(I->getNumArgOperands() == 2 &&
350  "Expect two args for recognised intrinsics.");
351  e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
352  e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
353  return e;
354  }
355  }
356 
357  // Not a recognised intrinsic. Fall back to producing an extract value
358  // expression.
359  e.opcode = EI->getOpcode();
360  for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
361  OI != OE; ++OI)
362  e.varargs.push_back(lookupOrAdd(*OI));
363 
364  for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
365  II != IE; ++II)
366  e.varargs.push_back(*II);
367 
368  return e;
369 }
370 
371 //===----------------------------------------------------------------------===//
372 // ValueTable External Functions
373 //===----------------------------------------------------------------------===//
374 
375 GVN::ValueTable::ValueTable() = default;
376 GVN::ValueTable::ValueTable(const ValueTable &) = default;
377 GVN::ValueTable::ValueTable(ValueTable &&) = default;
378 GVN::ValueTable::~ValueTable() = default;
379 
380 /// add - Insert a value into the table with a specified value number.
382  valueNumbering.insert(std::make_pair(V, num));
383  if (PHINode *PN = dyn_cast<PHINode>(V))
384  NumberingPhi[num] = PN;
385 }
386 
387 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
388  if (AA->doesNotAccessMemory(C)) {
389  Expression exp = createExpr(C);
390  uint32_t e = assignExpNewValueNum(exp).first;
391  valueNumbering[C] = e;
392  return e;
393  } else if (AA->onlyReadsMemory(C)) {
394  Expression exp = createExpr(C);
395  auto ValNum = assignExpNewValueNum(exp);
396  if (ValNum.second) {
397  valueNumbering[C] = ValNum.first;
398  return ValNum.first;
399  }
400  if (!MD) {
401  uint32_t e = assignExpNewValueNum(exp).first;
402  valueNumbering[C] = e;
403  return e;
404  }
405 
406  MemDepResult local_dep = MD->getDependency(C);
407 
408  if (!local_dep.isDef() && !local_dep.isNonLocal()) {
409  valueNumbering[C] = nextValueNumber;
410  return nextValueNumber++;
411  }
412 
413  if (local_dep.isDef()) {
414  CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
415 
416  if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
417  valueNumbering[C] = nextValueNumber;
418  return nextValueNumber++;
419  }
420 
421  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
422  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
423  uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
424  if (c_vn != cd_vn) {
425  valueNumbering[C] = nextValueNumber;
426  return nextValueNumber++;
427  }
428  }
429 
430  uint32_t v = lookupOrAdd(local_cdep);
431  valueNumbering[C] = v;
432  return v;
433  }
434 
435  // Non-local case.
437  MD->getNonLocalCallDependency(CallSite(C));
438  // FIXME: Move the checking logic to MemDep!
439  CallInst* cdep = nullptr;
440 
441  // Check to see if we have a single dominating call instruction that is
442  // identical to C.
443  for (unsigned i = 0, e = deps.size(); i != e; ++i) {
444  const NonLocalDepEntry *I = &deps[i];
445  if (I->getResult().isNonLocal())
446  continue;
447 
448  // We don't handle non-definitions. If we already have a call, reject
449  // instruction dependencies.
450  if (!I->getResult().isDef() || cdep != nullptr) {
451  cdep = nullptr;
452  break;
453  }
454 
455  CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
456  // FIXME: All duplicated with non-local case.
457  if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
458  cdep = NonLocalDepCall;
459  continue;
460  }
461 
462  cdep = nullptr;
463  break;
464  }
465 
466  if (!cdep) {
467  valueNumbering[C] = nextValueNumber;
468  return nextValueNumber++;
469  }
470 
471  if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
472  valueNumbering[C] = nextValueNumber;
473  return nextValueNumber++;
474  }
475  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
476  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
477  uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
478  if (c_vn != cd_vn) {
479  valueNumbering[C] = nextValueNumber;
480  return nextValueNumber++;
481  }
482  }
483 
484  uint32_t v = lookupOrAdd(cdep);
485  valueNumbering[C] = v;
486  return v;
487  } else {
488  valueNumbering[C] = nextValueNumber;
489  return nextValueNumber++;
490  }
491 }
492 
493 /// Returns true if a value number exists for the specified value.
494 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
495 
496 /// lookup_or_add - Returns the value number for the specified value, assigning
497 /// it a new number if it did not have one before.
499  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
500  if (VI != valueNumbering.end())
501  return VI->second;
502 
503  if (!isa<Instruction>(V)) {
504  valueNumbering[V] = nextValueNumber;
505  return nextValueNumber++;
506  }
507 
508  Instruction* I = cast<Instruction>(V);
509  Expression exp;
510  switch (I->getOpcode()) {
511  case Instruction::Call:
512  return lookupOrAddCall(cast<CallInst>(I));
513  case Instruction::Add:
514  case Instruction::FAdd:
515  case Instruction::Sub:
516  case Instruction::FSub:
517  case Instruction::Mul:
518  case Instruction::FMul:
519  case Instruction::UDiv:
520  case Instruction::SDiv:
521  case Instruction::FDiv:
522  case Instruction::URem:
523  case Instruction::SRem:
524  case Instruction::FRem:
525  case Instruction::Shl:
526  case Instruction::LShr:
527  case Instruction::AShr:
528  case Instruction::And:
529  case Instruction::Or:
530  case Instruction::Xor:
531  case Instruction::ICmp:
532  case Instruction::FCmp:
533  case Instruction::Trunc:
534  case Instruction::ZExt:
535  case Instruction::SExt:
536  case Instruction::FPToUI:
537  case Instruction::FPToSI:
538  case Instruction::UIToFP:
539  case Instruction::SIToFP:
540  case Instruction::FPTrunc:
541  case Instruction::FPExt:
542  case Instruction::PtrToInt:
543  case Instruction::IntToPtr:
544  case Instruction::BitCast:
545  case Instruction::Select:
546  case Instruction::ExtractElement:
547  case Instruction::InsertElement:
548  case Instruction::ShuffleVector:
549  case Instruction::InsertValue:
550  case Instruction::GetElementPtr:
551  exp = createExpr(I);
552  break;
553  case Instruction::ExtractValue:
554  exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
555  break;
556  case Instruction::PHI:
557  valueNumbering[V] = nextValueNumber;
558  NumberingPhi[nextValueNumber] = cast<PHINode>(V);
559  return nextValueNumber++;
560  default:
561  valueNumbering[V] = nextValueNumber;
562  return nextValueNumber++;
563  }
564 
565  uint32_t e = assignExpNewValueNum(exp).first;
566  valueNumbering[V] = e;
567  return e;
568 }
569 
570 /// Returns the value number of the specified value. Fails if
571 /// the value has not yet been numbered.
573  DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
574  if (Verify) {
575  assert(VI != valueNumbering.end() && "Value not numbered?");
576  return VI->second;
577  }
578  return (VI != valueNumbering.end()) ? VI->second : 0;
579 }
580 
581 /// Returns the value number of the given comparison,
582 /// assigning it a new number if it did not have one before. Useful when
583 /// we deduced the result of a comparison, but don't immediately have an
584 /// instruction realizing that comparison to hand.
586  CmpInst::Predicate Predicate,
587  Value *LHS, Value *RHS) {
588  Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
589  return assignExpNewValueNum(exp).first;
590 }
591 
592 /// Remove all entries from the ValueTable.
594  valueNumbering.clear();
595  expressionNumbering.clear();
596  NumberingPhi.clear();
597  PhiTranslateTable.clear();
598  nextValueNumber = 1;
599  Expressions.clear();
600  ExprIdx.clear();
601  nextExprNumber = 0;
602 }
603 
604 /// Remove a value from the value numbering.
606  uint32_t Num = valueNumbering.lookup(V);
607  valueNumbering.erase(V);
608  // If V is PHINode, V <--> value number is an one-to-one mapping.
609  if (isa<PHINode>(V))
610  NumberingPhi.erase(Num);
611 }
612 
613 /// verifyRemoved - Verify that the value is removed from all internal data
614 /// structures.
615 void GVN::ValueTable::verifyRemoved(const Value *V) const {
617  I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
618  assert(I->first != V && "Inst still occurs in value numbering map!");
619  }
620 }
621 
622 //===----------------------------------------------------------------------===//
623 // GVN Pass
624 //===----------------------------------------------------------------------===//
625 
627  // FIXME: The order of evaluation of these 'getResult' calls is very
628  // significant! Re-ordering these variables will cause GVN when run alone to
629  // be less effective! We should fix memdep and basic-aa to not exhibit this
630  // behavior, but until then don't change the order here.
631  auto &AC = AM.getResult<AssumptionAnalysis>(F);
632  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
633  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
634  auto &AA = AM.getResult<AAManager>(F);
635  auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
636  auto *LI = AM.getCachedResult<LoopAnalysis>(F);
638  bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
639  if (!Changed)
640  return PreservedAnalyses::all();
643  PA.preserve<GlobalsAA>();
645  return PA;
646 }
647 
648 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
649 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
650  errs() << "{\n";
652  E = d.end(); I != E; ++I) {
653  errs() << I->first << "\n";
654  I->second->dump();
655  }
656  errs() << "}\n";
657 }
658 #endif
659 
660 /// Return true if we can prove that the value
661 /// we're analyzing is fully available in the specified block. As we go, keep
662 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
663 /// map is actually a tri-state map with the following values:
664 /// 0) we know the block *is not* fully available.
665 /// 1) we know the block *is* fully available.
666 /// 2) we do not know whether the block is fully available or not, but we are
667 /// currently speculating that it will be.
668 /// 3) we are speculating for this block and have used that to speculate for
669 /// other blocks.
671  DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
672  uint32_t RecurseDepth) {
673  if (RecurseDepth > MaxRecurseDepth)
674  return false;
675 
676  // Optimistically assume that the block is fully available and check to see
677  // if we already know about this block in one lookup.
678  std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
679  FullyAvailableBlocks.insert(std::make_pair(BB, 2));
680 
681  // If the entry already existed for this block, return the precomputed value.
682  if (!IV.second) {
683  // If this is a speculative "available" value, mark it as being used for
684  // speculation of other blocks.
685  if (IV.first->second == 2)
686  IV.first->second = 3;
687  return IV.first->second != 0;
688  }
689 
690  // Otherwise, see if it is fully available in all predecessors.
691  pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
692 
693  // If this block has no predecessors, it isn't live-in here.
694  if (PI == PE)
695  goto SpeculationFailure;
696 
697  for (; PI != PE; ++PI)
698  // If the value isn't fully available in one of our predecessors, then it
699  // isn't fully available in this block either. Undo our previous
700  // optimistic assumption and bail out.
701  if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
702  goto SpeculationFailure;
703 
704  return true;
705 
706 // If we get here, we found out that this is not, after
707 // all, a fully-available block. We have a problem if we speculated on this and
708 // used the speculation to mark other blocks as available.
709 SpeculationFailure:
710  char &BBVal = FullyAvailableBlocks[BB];
711 
712  // If we didn't speculate on this, just return with it set to false.
713  if (BBVal == 2) {
714  BBVal = 0;
715  return false;
716  }
717 
718  // If we did speculate on this value, we could have blocks set to 1 that are
719  // incorrect. Walk the (transitive) successors of this block and mark them as
720  // 0 if set to one.
721  SmallVector<BasicBlock*, 32> BBWorklist;
722  BBWorklist.push_back(BB);
723 
724  do {
725  BasicBlock *Entry = BBWorklist.pop_back_val();
726  // Note that this sets blocks to 0 (unavailable) if they happen to not
727  // already be in FullyAvailableBlocks. This is safe.
728  char &EntryVal = FullyAvailableBlocks[Entry];
729  if (EntryVal == 0) continue; // Already unavailable.
730 
731  // Mark as unavailable.
732  EntryVal = 0;
733 
734  BBWorklist.append(succ_begin(Entry), succ_end(Entry));
735  } while (!BBWorklist.empty());
736 
737  return false;
738 }
739 
740 /// Given a set of loads specified by ValuesPerBlock,
741 /// construct SSA form, allowing us to eliminate LI. This returns the value
742 /// that should be used at LI's definition site.
745  GVN &gvn) {
746  // Check for the fully redundant, dominating load case. In this case, we can
747  // just use the dominating value directly.
748  if (ValuesPerBlock.size() == 1 &&
749  gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
750  LI->getParent())) {
751  assert(!ValuesPerBlock[0].AV.isUndefValue() &&
752  "Dead BB dominate this block");
753  return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
754  }
755 
756  // Otherwise, we have to construct SSA form.
757  SmallVector<PHINode*, 8> NewPHIs;
758  SSAUpdater SSAUpdate(&NewPHIs);
759  SSAUpdate.Initialize(LI->getType(), LI->getName());
760 
761  for (const AvailableValueInBlock &AV : ValuesPerBlock) {
762  BasicBlock *BB = AV.BB;
763 
764  if (SSAUpdate.HasValueForBlock(BB))
765  continue;
766 
767  SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
768  }
769 
770  // Perform PHI construction.
771  return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
772 }
773 
775  Instruction *InsertPt,
776  GVN &gvn) const {
777  Value *Res;
778  Type *LoadTy = LI->getType();
779  const DataLayout &DL = LI->getModule()->getDataLayout();
780  if (isSimpleValue()) {
781  Res = getSimpleValue();
782  if (Res->getType() != LoadTy) {
783  Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
784 
785  DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
786  << *getSimpleValue() << '\n'
787  << *Res << '\n' << "\n\n\n");
788  }
789  } else if (isCoercedLoadValue()) {
790  LoadInst *Load = getCoercedLoadValue();
791  if (Load->getType() == LoadTy && Offset == 0) {
792  Res = Load;
793  } else {
794  Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
795  // We would like to use gvn.markInstructionForDeletion here, but we can't
796  // because the load is already memoized into the leader map table that GVN
797  // tracks. It is potentially possible to remove the load from the table,
798  // but then there all of the operations based on it would need to be
799  // rehashed. Just leave the dead load around.
800  gvn.getMemDep().removeInstruction(Load);
801  DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
802  << *getCoercedLoadValue() << '\n'
803  << *Res << '\n'
804  << "\n\n\n");
805  }
806  } else if (isMemIntrinValue()) {
807  Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
808  InsertPt, DL);
809  DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
810  << " " << *getMemIntrinValue() << '\n'
811  << *Res << '\n' << "\n\n\n");
812  } else {
813  assert(isUndefValue() && "Should be UndefVal");
814  DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
815  return UndefValue::get(LoadTy);
816  }
817  assert(Res && "failed to materialize?");
818  return Res;
819 }
820 
821 static bool isLifetimeStart(const Instruction *Inst) {
822  if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
823  return II->getIntrinsicID() == Intrinsic::lifetime_start;
824  return false;
825 }
826 
827 /// \brief Try to locate the three instruction involved in a missed
828 /// load-elimination case that is due to an intervening store.
830  DominatorTree *DT,
832  using namespace ore;
833 
834  User *OtherAccess = nullptr;
835 
836  OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
837  R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
838  << setExtraArgs();
839 
840  for (auto *U : LI->getPointerOperand()->users())
841  if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
842  DT->dominates(cast<Instruction>(U), LI)) {
843  // FIXME: for now give up if there are multiple memory accesses that
844  // dominate the load. We need further analysis to decide which one is
845  // that we're forwarding from.
846  if (OtherAccess)
847  OtherAccess = nullptr;
848  else
849  OtherAccess = U;
850  }
851 
852  if (OtherAccess)
853  R << " in favor of " << NV("OtherAccess", OtherAccess);
854 
855  R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
856 
857  ORE->emit(R);
858 }
859 
860 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
861  Value *Address, AvailableValue &Res) {
862  assert((DepInfo.isDef() || DepInfo.isClobber()) &&
863  "expected a local dependence");
864  assert(LI->isUnordered() && "rules below are incorrect for ordered access");
865 
866  const DataLayout &DL = LI->getModule()->getDataLayout();
867 
868  if (DepInfo.isClobber()) {
869  // If the dependence is to a store that writes to a superset of the bits
870  // read by the load, we can extract the bits we need for the load from the
871  // stored value.
872  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
873  // Can't forward from non-atomic to atomic without violating memory model.
874  if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
875  int Offset =
877  if (Offset != -1) {
878  Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
879  return true;
880  }
881  }
882  }
883 
884  // Check to see if we have something like this:
885  // load i32* P
886  // load i8* (P+1)
887  // if we have this, replace the later with an extraction from the former.
888  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
889  // If this is a clobber and L is the first instruction in its block, then
890  // we have the first instruction in the entry block.
891  // Can't forward from non-atomic to atomic without violating memory model.
892  if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
893  int Offset =
894  analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
895 
896  if (Offset != -1) {
897  Res = AvailableValue::getLoad(DepLI, Offset);
898  return true;
899  }
900  }
901  }
902 
903  // If the clobbering value is a memset/memcpy/memmove, see if we can
904  // forward a value on from it.
905  if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
906  if (Address && !LI->isAtomic()) {
908  DepMI, DL);
909  if (Offset != -1) {
910  Res = AvailableValue::getMI(DepMI, Offset);
911  return true;
912  }
913  }
914  }
915  // Nothing known about this clobber, have to be conservative
916  DEBUG(
917  // fast print dep, using operator<< on instruction is too slow.
918  dbgs() << "GVN: load ";
919  LI->printAsOperand(dbgs());
920  Instruction *I = DepInfo.getInst();
921  dbgs() << " is clobbered by " << *I << '\n';
922  );
923  if (ORE->allowExtraAnalysis(DEBUG_TYPE))
924  reportMayClobberedLoad(LI, DepInfo, DT, ORE);
925 
926  return false;
927  }
928  assert(DepInfo.isDef() && "follows from above");
929 
930  Instruction *DepInst = DepInfo.getInst();
931 
932  // Loading the allocation -> undef.
933  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
934  // Loading immediately after lifetime begin -> undef.
935  isLifetimeStart(DepInst)) {
937  return true;
938  }
939 
940  // Loading from calloc (which zero initializes memory) -> zero
941  if (isCallocLikeFn(DepInst, TLI)) {
943  return true;
944  }
945 
946  if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
947  // Reject loads and stores that are to the same address but are of
948  // different types if we have to. If the stored value is larger or equal to
949  // the loaded value, we can reuse it.
950  if (S->getValueOperand()->getType() != LI->getType() &&
951  !canCoerceMustAliasedValueToLoad(S->getValueOperand(),
952  LI->getType(), DL))
953  return false;
954 
955  // Can't forward from non-atomic to atomic without violating memory model.
956  if (S->isAtomic() < LI->isAtomic())
957  return false;
958 
959  Res = AvailableValue::get(S->getValueOperand());
960  return true;
961  }
962 
963  if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
964  // If the types mismatch and we can't handle it, reject reuse of the load.
965  // If the stored value is larger or equal to the loaded value, we can reuse
966  // it.
967  if (LD->getType() != LI->getType() &&
969  return false;
970 
971  // Can't forward from non-atomic to atomic without violating memory model.
972  if (LD->isAtomic() < LI->isAtomic())
973  return false;
974 
976  return true;
977  }
978 
979  // Unknown def - must be conservative
980  DEBUG(
981  // fast print dep, using operator<< on instruction is too slow.
982  dbgs() << "GVN: load ";
983  LI->printAsOperand(dbgs());
984  dbgs() << " has unknown def " << *DepInst << '\n';
985  );
986  return false;
987 }
988 
989 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
990  AvailValInBlkVect &ValuesPerBlock,
991  UnavailBlkVect &UnavailableBlocks) {
992  // Filter out useless results (non-locals, etc). Keep track of the blocks
993  // where we have a value available in repl, also keep track of whether we see
994  // dependencies that produce an unknown value for the load (such as a call
995  // that could potentially clobber the load).
996  unsigned NumDeps = Deps.size();
997  for (unsigned i = 0, e = NumDeps; i != e; ++i) {
998  BasicBlock *DepBB = Deps[i].getBB();
999  MemDepResult DepInfo = Deps[i].getResult();
1000 
1001  if (DeadBlocks.count(DepBB)) {
1002  // Dead dependent mem-op disguise as a load evaluating the same value
1003  // as the load in question.
1004  ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1005  continue;
1006  }
1007 
1008  if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1009  UnavailableBlocks.push_back(DepBB);
1010  continue;
1011  }
1012 
1013  // The address being loaded in this non-local block may not be the same as
1014  // the pointer operand of the load if PHI translation occurs. Make sure
1015  // to consider the right address.
1016  Value *Address = Deps[i].getAddress();
1017 
1018  AvailableValue AV;
1019  if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1020  // subtlety: because we know this was a non-local dependency, we know
1021  // it's safe to materialize anywhere between the instruction within
1022  // DepInfo and the end of it's block.
1023  ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1024  std::move(AV)));
1025  } else {
1026  UnavailableBlocks.push_back(DepBB);
1027  }
1028  }
1029 
1030  assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1031  "post condition violation");
1032 }
1033 
1034 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1035  UnavailBlkVect &UnavailableBlocks) {
1036  // Okay, we have *some* definitions of the value. This means that the value
1037  // is available in some of our (transitive) predecessors. Lets think about
1038  // doing PRE of this load. This will involve inserting a new load into the
1039  // predecessor when it's not available. We could do this in general, but
1040  // prefer to not increase code size. As such, we only do this when we know
1041  // that we only have to insert *one* load (which means we're basically moving
1042  // the load, not inserting a new one).
1043 
1044  SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1045  UnavailableBlocks.end());
1046 
1047  // Let's find the first basic block with more than one predecessor. Walk
1048  // backwards through predecessors if needed.
1049  BasicBlock *LoadBB = LI->getParent();
1050  BasicBlock *TmpBB = LoadBB;
1051 
1052  while (TmpBB->getSinglePredecessor()) {
1053  TmpBB = TmpBB->getSinglePredecessor();
1054  if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1055  return false;
1056  if (Blockers.count(TmpBB))
1057  return false;
1058 
1059  // If any of these blocks has more than one successor (i.e. if the edge we
1060  // just traversed was critical), then there are other paths through this
1061  // block along which the load may not be anticipated. Hoisting the load
1062  // above this block would be adding the load to execution paths along
1063  // which it was not previously executed.
1064  if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1065  return false;
1066  }
1067 
1068  assert(TmpBB);
1069  LoadBB = TmpBB;
1070 
1071  // Check to see how many predecessors have the loaded value fully
1072  // available.
1074  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1075  for (const AvailableValueInBlock &AV : ValuesPerBlock)
1076  FullyAvailableBlocks[AV.BB] = true;
1077  for (BasicBlock *UnavailableBB : UnavailableBlocks)
1078  FullyAvailableBlocks[UnavailableBB] = false;
1079 
1080  SmallVector<BasicBlock *, 4> CriticalEdgePred;
1081  for (BasicBlock *Pred : predecessors(LoadBB)) {
1082  // If any predecessor block is an EH pad that does not allow non-PHI
1083  // instructions before the terminator, we can't PRE the load.
1084  if (Pred->getTerminator()->isEHPad()) {
1085  DEBUG(dbgs()
1086  << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1087  << Pred->getName() << "': " << *LI << '\n');
1088  return false;
1089  }
1090 
1091  if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1092  continue;
1093  }
1094 
1095  if (Pred->getTerminator()->getNumSuccessors() != 1) {
1096  if (isa<IndirectBrInst>(Pred->getTerminator())) {
1097  DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1098  << Pred->getName() << "': " << *LI << '\n');
1099  return false;
1100  }
1101 
1102  if (LoadBB->isEHPad()) {
1103  DEBUG(dbgs()
1104  << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1105  << Pred->getName() << "': " << *LI << '\n');
1106  return false;
1107  }
1108 
1109  CriticalEdgePred.push_back(Pred);
1110  } else {
1111  // Only add the predecessors that will not be split for now.
1112  PredLoads[Pred] = nullptr;
1113  }
1114  }
1115 
1116  // Decide whether PRE is profitable for this load.
1117  unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1118  assert(NumUnavailablePreds != 0 &&
1119  "Fully available value should already be eliminated!");
1120 
1121  // If this load is unavailable in multiple predecessors, reject it.
1122  // FIXME: If we could restructure the CFG, we could make a common pred with
1123  // all the preds that don't have an available LI and insert a new load into
1124  // that one block.
1125  if (NumUnavailablePreds != 1)
1126  return false;
1127 
1128  // Split critical edges, and update the unavailable predecessors accordingly.
1129  for (BasicBlock *OrigPred : CriticalEdgePred) {
1130  BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1131  assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1132  PredLoads[NewPred] = nullptr;
1133  DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1134  << LoadBB->getName() << '\n');
1135  }
1136 
1137  // Check if the load can safely be moved to all the unavailable predecessors.
1138  bool CanDoPRE = true;
1139  const DataLayout &DL = LI->getModule()->getDataLayout();
1141  for (auto &PredLoad : PredLoads) {
1142  BasicBlock *UnavailablePred = PredLoad.first;
1143 
1144  // Do PHI translation to get its value in the predecessor if necessary. The
1145  // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1146 
1147  // If all preds have a single successor, then we know it is safe to insert
1148  // the load on the pred (?!?), so we can insert code to materialize the
1149  // pointer if it is not available.
1150  PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1151  Value *LoadPtr = nullptr;
1152  LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1153  *DT, NewInsts);
1154 
1155  // If we couldn't find or insert a computation of this phi translated value,
1156  // we fail PRE.
1157  if (!LoadPtr) {
1158  DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1159  << *LI->getPointerOperand() << "\n");
1160  CanDoPRE = false;
1161  break;
1162  }
1163 
1164  PredLoad.second = LoadPtr;
1165  }
1166 
1167  if (!CanDoPRE) {
1168  while (!NewInsts.empty()) {
1169  Instruction *I = NewInsts.pop_back_val();
1170  if (MD) MD->removeInstruction(I);
1171  I->eraseFromParent();
1172  }
1173  // HINT: Don't revert the edge-splitting as following transformation may
1174  // also need to split these critical edges.
1175  return !CriticalEdgePred.empty();
1176  }
1177 
1178  // Okay, we can eliminate this load by inserting a reload in the predecessor
1179  // and using PHI construction to get the value in the other predecessors, do
1180  // it.
1181  DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1182  DEBUG(if (!NewInsts.empty())
1183  dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1184  << *NewInsts.back() << '\n');
1185 
1186  // Assign value numbers to the new instructions.
1187  for (Instruction *I : NewInsts) {
1188  // Instructions that have been inserted in predecessor(s) to materialize
1189  // the load address do not retain their original debug locations. Doing
1190  // so could lead to confusing (but correct) source attributions.
1191  // FIXME: How do we retain source locations without causing poor debugging
1192  // behavior?
1193  I->setDebugLoc(DebugLoc());
1194 
1195  // FIXME: We really _ought_ to insert these value numbers into their
1196  // parent's availability map. However, in doing so, we risk getting into
1197  // ordering issues. If a block hasn't been processed yet, we would be
1198  // marking a value as AVAIL-IN, which isn't what we intend.
1199  VN.lookupOrAdd(I);
1200  }
1201 
1202  for (const auto &PredLoad : PredLoads) {
1203  BasicBlock *UnavailablePred = PredLoad.first;
1204  Value *LoadPtr = PredLoad.second;
1205 
1206  auto *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre",
1207  LI->isVolatile(), LI->getAlignment(),
1208  LI->getOrdering(), LI->getSyncScopeID(),
1209  UnavailablePred->getTerminator());
1210  NewLoad->setDebugLoc(LI->getDebugLoc());
1211 
1212  // Transfer the old load's AA tags to the new load.
1213  AAMDNodes Tags;
1214  LI->getAAMetadata(Tags);
1215  if (Tags)
1216  NewLoad->setAAMetadata(Tags);
1217 
1218  if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1219  NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1220  if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1221  NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1222  if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1223  NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1224 
1225  // We do not propagate the old load's debug location, because the new
1226  // load now lives in a different BB, and we want to avoid a jumpy line
1227  // table.
1228  // FIXME: How do we retain source locations without causing poor debugging
1229  // behavior?
1230 
1231  // Add the newly created load.
1232  ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1233  NewLoad));
1234  MD->invalidateCachedPointerInfo(LoadPtr);
1235  DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1236  }
1237 
1238  // Perform PHI construction.
1239  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1240  LI->replaceAllUsesWith(V);
1241  if (isa<PHINode>(V))
1242  V->takeName(LI);
1243  if (Instruction *I = dyn_cast<Instruction>(V))
1244  I->setDebugLoc(LI->getDebugLoc());
1245  if (V->getType()->isPtrOrPtrVectorTy())
1246  MD->invalidateCachedPointerInfo(V);
1247  markInstructionForDeletion(LI);
1248  ORE->emit(OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1249  << "load eliminated by PRE");
1250  ++NumPRELoad;
1251  return true;
1252 }
1253 
1256  using namespace ore;
1257 
1258  ORE->emit(OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1259  << "load of type " << NV("Type", LI->getType()) << " eliminated"
1260  << setExtraArgs() << " in favor of "
1261  << NV("InfavorOfValue", AvailableValue));
1262 }
1263 
1264 /// Attempt to eliminate a load whose dependencies are
1265 /// non-local by performing PHI construction.
1266 bool GVN::processNonLocalLoad(LoadInst *LI) {
1267  // non-local speculations are not allowed under asan.
1268  if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeAddress))
1269  return false;
1270 
1271  // Step 1: Find the non-local dependencies of the load.
1272  LoadDepVect Deps;
1273  MD->getNonLocalPointerDependency(LI, Deps);
1274 
1275  // If we had to process more than one hundred blocks to find the
1276  // dependencies, this load isn't worth worrying about. Optimizing
1277  // it will be too expensive.
1278  unsigned NumDeps = Deps.size();
1279  if (NumDeps > 100)
1280  return false;
1281 
1282  // If we had a phi translation failure, we'll have a single entry which is a
1283  // clobber in the current block. Reject this early.
1284  if (NumDeps == 1 &&
1285  !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1286  DEBUG(
1287  dbgs() << "GVN: non-local load ";
1288  LI->printAsOperand(dbgs());
1289  dbgs() << " has unknown dependencies\n";
1290  );
1291  return false;
1292  }
1293 
1294  // If this load follows a GEP, see if we can PRE the indices before analyzing.
1295  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1296  for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1297  OE = GEP->idx_end();
1298  OI != OE; ++OI)
1299  if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1300  performScalarPRE(I);
1301  }
1302 
1303  // Step 2: Analyze the availability of the load
1304  AvailValInBlkVect ValuesPerBlock;
1305  UnavailBlkVect UnavailableBlocks;
1306  AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1307 
1308  // If we have no predecessors that produce a known value for this load, exit
1309  // early.
1310  if (ValuesPerBlock.empty())
1311  return false;
1312 
1313  // Step 3: Eliminate fully redundancy.
1314  //
1315  // If all of the instructions we depend on produce a known value for this
1316  // load, then it is fully redundant and we can use PHI insertion to compute
1317  // its value. Insert PHIs and remove the fully redundant value now.
1318  if (UnavailableBlocks.empty()) {
1319  DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1320 
1321  // Perform PHI construction.
1322  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1323  LI->replaceAllUsesWith(V);
1324 
1325  if (isa<PHINode>(V))
1326  V->takeName(LI);
1327  if (Instruction *I = dyn_cast<Instruction>(V))
1328  // If instruction I has debug info, then we should not update it.
1329  // Also, if I has a null DebugLoc, then it is still potentially incorrect
1330  // to propagate LI's DebugLoc because LI may not post-dominate I.
1331  if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1332  I->setDebugLoc(LI->getDebugLoc());
1333  if (V->getType()->isPtrOrPtrVectorTy())
1334  MD->invalidateCachedPointerInfo(V);
1335  markInstructionForDeletion(LI);
1336  ++NumGVNLoad;
1337  reportLoadElim(LI, V, ORE);
1338  return true;
1339  }
1340 
1341  // Step 4: Eliminate partial redundancy.
1342  if (!EnablePRE || !EnableLoadPRE)
1343  return false;
1344 
1345  return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1346 }
1347 
1348 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1349  assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1350  "This function can only be called with llvm.assume intrinsic");
1351  Value *V = IntrinsicI->getArgOperand(0);
1352 
1353  if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1354  if (Cond->isZero()) {
1355  Type *Int8Ty = Type::getInt8Ty(V->getContext());
1356  // Insert a new store to null instruction before the load to indicate that
1357  // this code is not reachable. FIXME: We could insert unreachable
1358  // instruction directly because we can modify the CFG.
1359  new StoreInst(UndefValue::get(Int8Ty),
1361  IntrinsicI);
1362  }
1363  markInstructionForDeletion(IntrinsicI);
1364  return false;
1365  }
1366 
1367  Constant *True = ConstantInt::getTrue(V->getContext());
1368  bool Changed = false;
1369 
1370  for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1371  BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1372 
1373  // This property is only true in dominated successors, propagateEquality
1374  // will check dominance for us.
1375  Changed |= propagateEquality(V, True, Edge, false);
1376  }
1377 
1378  // We can replace assume value with true, which covers cases like this:
1379  // call void @llvm.assume(i1 %cmp)
1380  // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1381  ReplaceWithConstMap[V] = True;
1382 
1383  // If one of *cmp *eq operand is const, adding it to map will cover this:
1384  // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1385  // call void @llvm.assume(i1 %cmp)
1386  // ret float %0 ; will change it to ret float 3.000000e+00
1387  if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1388  if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1389  CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1390  (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1391  CmpI->getFastMathFlags().noNaNs())) {
1392  Value *CmpLHS = CmpI->getOperand(0);
1393  Value *CmpRHS = CmpI->getOperand(1);
1394  if (isa<Constant>(CmpLHS))
1395  std::swap(CmpLHS, CmpRHS);
1396  auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1397 
1398  // If only one operand is constant.
1399  if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1400  ReplaceWithConstMap[CmpLHS] = RHSConst;
1401  }
1402  }
1403  return Changed;
1404 }
1405 
1407  auto *ReplInst = dyn_cast<Instruction>(Repl);
1408  if (!ReplInst)
1409  return;
1410 
1411  // Patch the replacement so that it is not more restrictive than the value
1412  // being replaced.
1413  // Note that if 'I' is a load being replaced by some operation,
1414  // for example, by an arithmetic operation, then andIRFlags()
1415  // would just erase all math flags from the original arithmetic
1416  // operation, which is clearly not wanted and not needed.
1417  if (!isa<LoadInst>(I))
1418  ReplInst->andIRFlags(I);
1419 
1420  // FIXME: If both the original and replacement value are part of the
1421  // same control-flow region (meaning that the execution of one
1422  // guarantees the execution of the other), then we can combine the
1423  // noalias scopes here and do better than the general conservative
1424  // answer used in combineMetadata().
1425 
1426  // In general, GVN unifies expressions over different control-flow
1427  // regions, and so we need a conservative combination of the noalias
1428  // scopes.
1429  static const unsigned KnownIDs[] = {
1434  combineMetadata(ReplInst, I, KnownIDs);
1435 }
1436 
1438  patchReplacementInstruction(I, Repl);
1439  I->replaceAllUsesWith(Repl);
1440 }
1441 
1442 /// Attempt to eliminate a load, first by eliminating it
1443 /// locally, and then attempting non-local elimination if that fails.
1444 bool GVN::processLoad(LoadInst *L) {
1445  if (!MD)
1446  return false;
1447 
1448  // This code hasn't been audited for ordered or volatile memory access
1449  if (!L->isUnordered())
1450  return false;
1451 
1452  if (L->use_empty()) {
1453  markInstructionForDeletion(L);
1454  return true;
1455  }
1456 
1457  // ... to a pointer that has been loaded from before...
1458  MemDepResult Dep = MD->getDependency(L);
1459 
1460  // If it is defined in another block, try harder.
1461  if (Dep.isNonLocal())
1462  return processNonLocalLoad(L);
1463 
1464  // Only handle the local case below
1465  if (!Dep.isDef() && !Dep.isClobber()) {
1466  // This might be a NonFuncLocal or an Unknown
1467  DEBUG(
1468  // fast print dep, using operator<< on instruction is too slow.
1469  dbgs() << "GVN: load ";
1470  L->printAsOperand(dbgs());
1471  dbgs() << " has unknown dependence\n";
1472  );
1473  return false;
1474  }
1475 
1476  AvailableValue AV;
1477  if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1478  Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1479 
1480  // Replace the load!
1481  patchAndReplaceAllUsesWith(L, AvailableValue);
1482  markInstructionForDeletion(L);
1483  ++NumGVNLoad;
1484  reportLoadElim(L, AvailableValue, ORE);
1485  // Tell MDA to rexamine the reused pointer since we might have more
1486  // information after forwarding it.
1487  if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1488  MD->invalidateCachedPointerInfo(AvailableValue);
1489  return true;
1490  }
1491 
1492  return false;
1493 }
1494 
1495 /// Return a pair the first field showing the value number of \p Exp and the
1496 /// second field showing whether it is a value number newly created.
1497 std::pair<uint32_t, bool>
1498 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1499  uint32_t &e = expressionNumbering[Exp];
1500  bool CreateNewValNum = !e;
1501  if (CreateNewValNum) {
1502  Expressions.push_back(Exp);
1503  if (ExprIdx.size() < nextValueNumber + 1)
1504  ExprIdx.resize(nextValueNumber * 2);
1505  e = nextValueNumber;
1506  ExprIdx[nextValueNumber++] = nextExprNumber++;
1507  }
1508  return {e, CreateNewValNum};
1509 }
1510 
1511 /// Return whether all the values related with the same \p num are
1512 /// defined in \p BB.
1513 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1514  GVN &Gvn) {
1515  LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1516  while (Vals && Vals->BB == BB)
1517  Vals = Vals->Next;
1518  return !Vals;
1519 }
1520 
1521 /// Wrap phiTranslateImpl to provide caching functionality.
1523  const BasicBlock *PhiBlock, uint32_t Num,
1524  GVN &Gvn) {
1525  auto FindRes = PhiTranslateTable.find({Num, Pred});
1526  if (FindRes != PhiTranslateTable.end())
1527  return FindRes->second;
1528  uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1529  PhiTranslateTable.insert({{Num, Pred}, NewNum});
1530  return NewNum;
1531 }
1532 
1533 /// Translate value number \p Num using phis, so that it has the values of
1534 /// the phis in BB.
1535 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1536  const BasicBlock *PhiBlock,
1537  uint32_t Num, GVN &Gvn) {
1538  if (PHINode *PN = NumberingPhi[Num]) {
1539  for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1540  if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1541  if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1542  return TransVal;
1543  }
1544  return Num;
1545  }
1546 
1547  // If there is any value related with Num is defined in a BB other than
1548  // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1549  // a backedge. We can do an early exit in that case to save compile time.
1550  if (!areAllValsInBB(Num, PhiBlock, Gvn))
1551  return Num;
1552 
1553  if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1554  return Num;
1555  Expression Exp = Expressions[ExprIdx[Num]];
1556 
1557  for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1558  // For InsertValue and ExtractValue, some varargs are index numbers
1559  // instead of value numbers. Those index numbers should not be
1560  // translated.
1561  if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1562  (i > 0 && Exp.opcode == Instruction::ExtractValue))
1563  continue;
1564  Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1565  }
1566 
1567  if (Exp.commutative) {
1568  assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1569  if (Exp.varargs[0] > Exp.varargs[1]) {
1570  std::swap(Exp.varargs[0], Exp.varargs[1]);
1571  uint32_t Opcode = Exp.opcode >> 8;
1572  if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1573  Exp.opcode = (Opcode << 8) |
1575  static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1576  }
1577  }
1578 
1579  if (uint32_t NewNum = expressionNumbering[Exp])
1580  return NewNum;
1581  return Num;
1582 }
1583 
1584 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1585 /// again.
1587  const BasicBlock &CurrBlock) {
1588  for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1589  auto FindRes = PhiTranslateTable.find({Num, Pred});
1590  if (FindRes != PhiTranslateTable.end())
1591  PhiTranslateTable.erase(FindRes);
1592  }
1593 }
1594 
1595 // In order to find a leader for a given value number at a
1596 // specific basic block, we first obtain the list of all Values for that number,
1597 // and then scan the list to find one whose block dominates the block in
1598 // question. This is fast because dominator tree queries consist of only
1599 // a few comparisons of DFS numbers.
1600 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1601  LeaderTableEntry Vals = LeaderTable[num];
1602  if (!Vals.Val) return nullptr;
1603 
1604  Value *Val = nullptr;
1605  if (DT->dominates(Vals.BB, BB)) {
1606  Val = Vals.Val;
1607  if (isa<Constant>(Val)) return Val;
1608  }
1609 
1610  LeaderTableEntry* Next = Vals.Next;
1611  while (Next) {
1612  if (DT->dominates(Next->BB, BB)) {
1613  if (isa<Constant>(Next->Val)) return Next->Val;
1614  if (!Val) Val = Next->Val;
1615  }
1616 
1617  Next = Next->Next;
1618  }
1619 
1620  return Val;
1621 }
1622 
1623 /// There is an edge from 'Src' to 'Dst'. Return
1624 /// true if every path from the entry block to 'Dst' passes via this edge. In
1625 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1627  DominatorTree *DT) {
1628  // While in theory it is interesting to consider the case in which Dst has
1629  // more than one predecessor, because Dst might be part of a loop which is
1630  // only reachable from Src, in practice it is pointless since at the time
1631  // GVN runs all such loops have preheaders, which means that Dst will have
1632  // been changed to have only one predecessor, namely Src.
1633  const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1634  assert((!Pred || Pred == E.getStart()) &&
1635  "No edge between these basic blocks!");
1636  return Pred != nullptr;
1637 }
1638 
1639 void GVN::assignBlockRPONumber(Function &F) {
1640  uint32_t NextBlockNumber = 1;
1642  for (BasicBlock *BB : RPOT)
1643  BlockRPONumber[BB] = NextBlockNumber++;
1644 }
1645 
1646 // Tries to replace instruction with const, using information from
1647 // ReplaceWithConstMap.
1648 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1649  bool Changed = false;
1650  for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1651  Value *Operand = Instr->getOperand(OpNum);
1652  auto it = ReplaceWithConstMap.find(Operand);
1653  if (it != ReplaceWithConstMap.end()) {
1654  assert(!isa<Constant>(Operand) &&
1655  "Replacing constants with constants is invalid");
1656  DEBUG(dbgs() << "GVN replacing: " << *Operand << " with " << *it->second
1657  << " in instruction " << *Instr << '\n');
1658  Instr->setOperand(OpNum, it->second);
1659  Changed = true;
1660  }
1661  }
1662  return Changed;
1663 }
1664 
1665 /// The given values are known to be equal in every block
1666 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1667 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1668 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1669 /// value starting from the end of Root.Start.
1670 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1671  bool DominatesByEdge) {
1673  Worklist.push_back(std::make_pair(LHS, RHS));
1674  bool Changed = false;
1675  // For speed, compute a conservative fast approximation to
1676  // DT->dominates(Root, Root.getEnd());
1677  const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1678 
1679  while (!Worklist.empty()) {
1680  std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1681  LHS = Item.first; RHS = Item.second;
1682 
1683  if (LHS == RHS)
1684  continue;
1685  assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1686 
1687  // Don't try to propagate equalities between constants.
1688  if (isa<Constant>(LHS) && isa<Constant>(RHS))
1689  continue;
1690 
1691  // Prefer a constant on the right-hand side, or an Argument if no constants.
1692  if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1693  std::swap(LHS, RHS);
1694  assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1695 
1696  // If there is no obvious reason to prefer the left-hand side over the
1697  // right-hand side, ensure the longest lived term is on the right-hand side,
1698  // so the shortest lived term will be replaced by the longest lived.
1699  // This tends to expose more simplifications.
1700  uint32_t LVN = VN.lookupOrAdd(LHS);
1701  if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1702  (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1703  // Move the 'oldest' value to the right-hand side, using the value number
1704  // as a proxy for age.
1705  uint32_t RVN = VN.lookupOrAdd(RHS);
1706  if (LVN < RVN) {
1707  std::swap(LHS, RHS);
1708  LVN = RVN;
1709  }
1710  }
1711 
1712  // If value numbering later sees that an instruction in the scope is equal
1713  // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1714  // the invariant that instructions only occur in the leader table for their
1715  // own value number (this is used by removeFromLeaderTable), do not do this
1716  // if RHS is an instruction (if an instruction in the scope is morphed into
1717  // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1718  // using the leader table is about compiling faster, not optimizing better).
1719  // The leader table only tracks basic blocks, not edges. Only add to if we
1720  // have the simple case where the edge dominates the end.
1721  if (RootDominatesEnd && !isa<Instruction>(RHS))
1722  addToLeaderTable(LVN, RHS, Root.getEnd());
1723 
1724  // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1725  // LHS always has at least one use that is not dominated by Root, this will
1726  // never do anything if LHS has only one use.
1727  if (!LHS->hasOneUse()) {
1728  unsigned NumReplacements =
1729  DominatesByEdge
1730  ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1731  : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1732 
1733  Changed |= NumReplacements > 0;
1734  NumGVNEqProp += NumReplacements;
1735  }
1736 
1737  // Now try to deduce additional equalities from this one. For example, if
1738  // the known equality was "(A != B)" == "false" then it follows that A and B
1739  // are equal in the scope. Only boolean equalities with an explicit true or
1740  // false RHS are currently supported.
1741  if (!RHS->getType()->isIntegerTy(1))
1742  // Not a boolean equality - bail out.
1743  continue;
1744  ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1745  if (!CI)
1746  // RHS neither 'true' nor 'false' - bail out.
1747  continue;
1748  // Whether RHS equals 'true'. Otherwise it equals 'false'.
1749  bool isKnownTrue = CI->isMinusOne();
1750  bool isKnownFalse = !isKnownTrue;
1751 
1752  // If "A && B" is known true then both A and B are known true. If "A || B"
1753  // is known false then both A and B are known false.
1754  Value *A, *B;
1755  if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1756  (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1757  Worklist.push_back(std::make_pair(A, RHS));
1758  Worklist.push_back(std::make_pair(B, RHS));
1759  continue;
1760  }
1761 
1762  // If we are propagating an equality like "(A == B)" == "true" then also
1763  // propagate the equality A == B. When propagating a comparison such as
1764  // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1765  if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1766  Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1767 
1768  // If "A == B" is known true, or "A != B" is known false, then replace
1769  // A with B everywhere in the scope.
1770  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1771  (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1772  Worklist.push_back(std::make_pair(Op0, Op1));
1773 
1774  // Handle the floating point versions of equality comparisons too.
1775  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1776  (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1777 
1778  // Floating point -0.0 and 0.0 compare equal, so we can only
1779  // propagate values if we know that we have a constant and that
1780  // its value is non-zero.
1781 
1782  // FIXME: We should do this optimization if 'no signed zeros' is
1783  // applicable via an instruction-level fast-math-flag or some other
1784  // indicator that relaxed FP semantics are being used.
1785 
1786  if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1787  Worklist.push_back(std::make_pair(Op0, Op1));
1788  }
1789 
1790  // If "A >= B" is known true, replace "A < B" with false everywhere.
1791  CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1792  Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1793  // Since we don't have the instruction "A < B" immediately to hand, work
1794  // out the value number that it would have and use that to find an
1795  // appropriate instruction (if any).
1796  uint32_t NextNum = VN.getNextUnusedValueNumber();
1797  uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1798  // If the number we were assigned was brand new then there is no point in
1799  // looking for an instruction realizing it: there cannot be one!
1800  if (Num < NextNum) {
1801  Value *NotCmp = findLeader(Root.getEnd(), Num);
1802  if (NotCmp && isa<Instruction>(NotCmp)) {
1803  unsigned NumReplacements =
1804  DominatesByEdge
1805  ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1806  : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1807  Root.getStart());
1808  Changed |= NumReplacements > 0;
1809  NumGVNEqProp += NumReplacements;
1810  }
1811  }
1812  // Ensure that any instruction in scope that gets the "A < B" value number
1813  // is replaced with false.
1814  // The leader table only tracks basic blocks, not edges. Only add to if we
1815  // have the simple case where the edge dominates the end.
1816  if (RootDominatesEnd)
1817  addToLeaderTable(Num, NotVal, Root.getEnd());
1818 
1819  continue;
1820  }
1821  }
1822 
1823  return Changed;
1824 }
1825 
1826 /// When calculating availability, handle an instruction
1827 /// by inserting it into the appropriate sets
1828 bool GVN::processInstruction(Instruction *I) {
1829  // Ignore dbg info intrinsics.
1830  if (isa<DbgInfoIntrinsic>(I))
1831  return false;
1832 
1833  // If the instruction can be easily simplified then do so now in preference
1834  // to value numbering it. Value numbering often exposes redundancies, for
1835  // example if it determines that %y is equal to %x then the instruction
1836  // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1837  const DataLayout &DL = I->getModule()->getDataLayout();
1838  if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1839  bool Changed = false;
1840  if (!I->use_empty()) {
1841  I->replaceAllUsesWith(V);
1842  Changed = true;
1843  }
1844  if (isInstructionTriviallyDead(I, TLI)) {
1845  markInstructionForDeletion(I);
1846  Changed = true;
1847  }
1848  if (Changed) {
1849  if (MD && V->getType()->isPtrOrPtrVectorTy())
1850  MD->invalidateCachedPointerInfo(V);
1851  ++NumGVNSimpl;
1852  return true;
1853  }
1854  }
1855 
1856  if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1857  if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1858  return processAssumeIntrinsic(IntrinsicI);
1859 
1860  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1861  if (processLoad(LI))
1862  return true;
1863 
1864  unsigned Num = VN.lookupOrAdd(LI);
1865  addToLeaderTable(Num, LI, LI->getParent());
1866  return false;
1867  }
1868 
1869  // For conditional branches, we can perform simple conditional propagation on
1870  // the condition value itself.
1871  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1872  if (!BI->isConditional())
1873  return false;
1874 
1875  if (isa<Constant>(BI->getCondition()))
1876  return processFoldableCondBr(BI);
1877 
1878  Value *BranchCond = BI->getCondition();
1879  BasicBlock *TrueSucc = BI->getSuccessor(0);
1880  BasicBlock *FalseSucc = BI->getSuccessor(1);
1881  // Avoid multiple edges early.
1882  if (TrueSucc == FalseSucc)
1883  return false;
1884 
1885  BasicBlock *Parent = BI->getParent();
1886  bool Changed = false;
1887 
1888  Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1889  BasicBlockEdge TrueE(Parent, TrueSucc);
1890  Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1891 
1892  Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1893  BasicBlockEdge FalseE(Parent, FalseSucc);
1894  Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1895 
1896  return Changed;
1897  }
1898 
1899  // For switches, propagate the case values into the case destinations.
1900  if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1901  Value *SwitchCond = SI->getCondition();
1902  BasicBlock *Parent = SI->getParent();
1903  bool Changed = false;
1904 
1905  // Remember how many outgoing edges there are to every successor.
1907  for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1908  ++SwitchEdges[SI->getSuccessor(i)];
1909 
1910  for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1911  i != e; ++i) {
1912  BasicBlock *Dst = i->getCaseSuccessor();
1913  // If there is only a single edge, propagate the case value into it.
1914  if (SwitchEdges.lookup(Dst) == 1) {
1915  BasicBlockEdge E(Parent, Dst);
1916  Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1917  }
1918  }
1919  return Changed;
1920  }
1921 
1922  // Instructions with void type don't return a value, so there's
1923  // no point in trying to find redundancies in them.
1924  if (I->getType()->isVoidTy())
1925  return false;
1926 
1927  uint32_t NextNum = VN.getNextUnusedValueNumber();
1928  unsigned Num = VN.lookupOrAdd(I);
1929 
1930  // Allocations are always uniquely numbered, so we can save time and memory
1931  // by fast failing them.
1932  if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
1933  addToLeaderTable(Num, I, I->getParent());
1934  return false;
1935  }
1936 
1937  // If the number we were assigned was a brand new VN, then we don't
1938  // need to do a lookup to see if the number already exists
1939  // somewhere in the domtree: it can't!
1940  if (Num >= NextNum) {
1941  addToLeaderTable(Num, I, I->getParent());
1942  return false;
1943  }
1944 
1945  // Perform fast-path value-number based elimination of values inherited from
1946  // dominators.
1947  Value *Repl = findLeader(I->getParent(), Num);
1948  if (!Repl) {
1949  // Failure, just remember this instance for future use.
1950  addToLeaderTable(Num, I, I->getParent());
1951  return false;
1952  } else if (Repl == I) {
1953  // If I was the result of a shortcut PRE, it might already be in the table
1954  // and the best replacement for itself. Nothing to do.
1955  return false;
1956  }
1957 
1958  // Remove it!
1959  patchAndReplaceAllUsesWith(I, Repl);
1960  if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1961  MD->invalidateCachedPointerInfo(Repl);
1962  markInstructionForDeletion(I);
1963  return true;
1964 }
1965 
1966 /// runOnFunction - This is the main transformation entry point for a function.
1967 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1968  const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1969  MemoryDependenceResults *RunMD, LoopInfo *LI,
1970  OptimizationRemarkEmitter *RunORE) {
1971  AC = &RunAC;
1972  DT = &RunDT;
1973  VN.setDomTree(DT);
1974  TLI = &RunTLI;
1975  VN.setAliasAnalysis(&RunAA);
1976  MD = RunMD;
1977  VN.setMemDep(MD);
1978  ORE = RunORE;
1979 
1980  bool Changed = false;
1981  bool ShouldContinue = true;
1982 
1983  // Merge unconditional branches, allowing PRE to catch more
1984  // optimization opportunities.
1985  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1986  BasicBlock *BB = &*FI++;
1987 
1988  bool removedBlock = MergeBlockIntoPredecessor(BB, DT, LI, MD);
1989  if (removedBlock)
1990  ++NumGVNBlocks;
1991 
1992  Changed |= removedBlock;
1993  }
1994 
1995  unsigned Iteration = 0;
1996  while (ShouldContinue) {
1997  DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
1998  ShouldContinue = iterateOnFunction(F);
1999  Changed |= ShouldContinue;
2000  ++Iteration;
2001  }
2002 
2003  if (EnablePRE) {
2004  // Fabricate val-num for dead-code in order to suppress assertion in
2005  // performPRE().
2006  assignValNumForDeadCode();
2007  assignBlockRPONumber(F);
2008  bool PREChanged = true;
2009  while (PREChanged) {
2010  PREChanged = performPRE(F);
2011  Changed |= PREChanged;
2012  }
2013  }
2014 
2015  // FIXME: Should perform GVN again after PRE does something. PRE can move
2016  // computations into blocks where they become fully redundant. Note that
2017  // we can't do this until PRE's critical edge splitting updates memdep.
2018  // Actually, when this happens, we should just fully integrate PRE into GVN.
2019 
2020  cleanupGlobalSets();
2021  // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2022  // iteration.
2023  DeadBlocks.clear();
2024 
2025  return Changed;
2026 }
2027 
2028 bool GVN::processBlock(BasicBlock *BB) {
2029  // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2030  // (and incrementing BI before processing an instruction).
2031  assert(InstrsToErase.empty() &&
2032  "We expect InstrsToErase to be empty across iterations");
2033  if (DeadBlocks.count(BB))
2034  return false;
2035 
2036  // Clearing map before every BB because it can be used only for single BB.
2037  ReplaceWithConstMap.clear();
2038  bool ChangedFunction = false;
2039 
2040  for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2041  BI != BE;) {
2042  if (!ReplaceWithConstMap.empty())
2043  ChangedFunction |= replaceOperandsWithConsts(&*BI);
2044  ChangedFunction |= processInstruction(&*BI);
2045 
2046  if (InstrsToErase.empty()) {
2047  ++BI;
2048  continue;
2049  }
2050 
2051  // If we need some instructions deleted, do it now.
2052  NumGVNInstr += InstrsToErase.size();
2053 
2054  // Avoid iterator invalidation.
2055  bool AtStart = BI == BB->begin();
2056  if (!AtStart)
2057  --BI;
2058 
2059  for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2060  E = InstrsToErase.end(); I != E; ++I) {
2061  DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2062  if (MD) MD->removeInstruction(*I);
2063  DEBUG(verifyRemoved(*I));
2064  (*I)->eraseFromParent();
2065  }
2066  InstrsToErase.clear();
2067 
2068  if (AtStart)
2069  BI = BB->begin();
2070  else
2071  ++BI;
2072  }
2073 
2074  return ChangedFunction;
2075 }
2076 
2077 // Instantiate an expression in a predecessor that lacked it.
2078 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2079  BasicBlock *Curr, unsigned int ValNo) {
2080  // Because we are going top-down through the block, all value numbers
2081  // will be available in the predecessor by the time we need them. Any
2082  // that weren't originally present will have been instantiated earlier
2083  // in this loop.
2084  bool success = true;
2085  for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2086  Value *Op = Instr->getOperand(i);
2087  if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2088  continue;
2089  // This could be a newly inserted instruction, in which case, we won't
2090  // find a value number, and should give up before we hurt ourselves.
2091  // FIXME: Rewrite the infrastructure to let it easier to value number
2092  // and process newly inserted instructions.
2093  if (!VN.exists(Op)) {
2094  success = false;
2095  break;
2096  }
2097  uint32_t TValNo =
2098  VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2099  if (Value *V = findLeader(Pred, TValNo)) {
2100  Instr->setOperand(i, V);
2101  } else {
2102  success = false;
2103  break;
2104  }
2105  }
2106 
2107  // Fail out if we encounter an operand that is not available in
2108  // the PRE predecessor. This is typically because of loads which
2109  // are not value numbered precisely.
2110  if (!success)
2111  return false;
2112 
2113  Instr->insertBefore(Pred->getTerminator());
2114  Instr->setName(Instr->getName() + ".pre");
2115  Instr->setDebugLoc(Instr->getDebugLoc());
2116 
2117  unsigned Num = VN.lookupOrAdd(Instr);
2118  VN.add(Instr, Num);
2119 
2120  // Update the availability map to include the new instruction.
2121  addToLeaderTable(Num, Instr, Pred);
2122  return true;
2123 }
2124 
2125 bool GVN::performScalarPRE(Instruction *CurInst) {
2126  if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2127  isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2128  CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2129  isa<DbgInfoIntrinsic>(CurInst))
2130  return false;
2131 
2132  // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2133  // sinking the compare again, and it would force the code generator to
2134  // move the i1 from processor flags or predicate registers into a general
2135  // purpose register.
2136  if (isa<CmpInst>(CurInst))
2137  return false;
2138 
2139  // We don't currently value number ANY inline asm calls.
2140  if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2141  if (CallI->isInlineAsm())
2142  return false;
2143 
2144  uint32_t ValNo = VN.lookup(CurInst);
2145 
2146  // Look for the predecessors for PRE opportunities. We're
2147  // only trying to solve the basic diamond case, where
2148  // a value is computed in the successor and one predecessor,
2149  // but not the other. We also explicitly disallow cases
2150  // where the successor is its own predecessor, because they're
2151  // more complicated to get right.
2152  unsigned NumWith = 0;
2153  unsigned NumWithout = 0;
2154  BasicBlock *PREPred = nullptr;
2155  BasicBlock *CurrentBlock = CurInst->getParent();
2156 
2158  for (BasicBlock *P : predecessors(CurrentBlock)) {
2159  // We're not interested in PRE where blocks with predecessors that are
2160  // not reachable.
2161  if (!DT->isReachableFromEntry(P)) {
2162  NumWithout = 2;
2163  break;
2164  }
2165  // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2166  // when CurInst has operand defined in CurrentBlock (so it may be defined
2167  // by phi in the loop header).
2168  if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2169  llvm::any_of(CurInst->operands(), [&](const Use &U) {
2170  if (auto *Inst = dyn_cast<Instruction>(U.get()))
2171  return Inst->getParent() == CurrentBlock;
2172  return false;
2173  })) {
2174  NumWithout = 2;
2175  break;
2176  }
2177 
2178  uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2179  Value *predV = findLeader(P, TValNo);
2180  if (!predV) {
2181  predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2182  PREPred = P;
2183  ++NumWithout;
2184  } else if (predV == CurInst) {
2185  /* CurInst dominates this predecessor. */
2186  NumWithout = 2;
2187  break;
2188  } else {
2189  predMap.push_back(std::make_pair(predV, P));
2190  ++NumWith;
2191  }
2192  }
2193 
2194  // Don't do PRE when it might increase code size, i.e. when
2195  // we would need to insert instructions in more than one pred.
2196  if (NumWithout > 1 || NumWith == 0)
2197  return false;
2198 
2199  // We may have a case where all predecessors have the instruction,
2200  // and we just need to insert a phi node. Otherwise, perform
2201  // insertion.
2202  Instruction *PREInstr = nullptr;
2203 
2204  if (NumWithout != 0) {
2205  // Don't do PRE across indirect branch.
2206  if (isa<IndirectBrInst>(PREPred->getTerminator()))
2207  return false;
2208 
2209  // We can't do PRE safely on a critical edge, so instead we schedule
2210  // the edge to be split and perform the PRE the next time we iterate
2211  // on the function.
2212  unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2213  if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2214  toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2215  return false;
2216  }
2217  // We need to insert somewhere, so let's give it a shot
2218  PREInstr = CurInst->clone();
2219  if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2220  // If we failed insertion, make sure we remove the instruction.
2221  DEBUG(verifyRemoved(PREInstr));
2222  PREInstr->deleteValue();
2223  return false;
2224  }
2225  }
2226 
2227  // Either we should have filled in the PRE instruction, or we should
2228  // not have needed insertions.
2229  assert(PREInstr != nullptr || NumWithout == 0);
2230 
2231  ++NumGVNPRE;
2232 
2233  // Create a PHI to make the value available in this block.
2234  PHINode *Phi =
2235  PHINode::Create(CurInst->getType(), predMap.size(),
2236  CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2237  for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2238  if (Value *V = predMap[i].first)
2239  Phi->addIncoming(V, predMap[i].second);
2240  else
2241  Phi->addIncoming(PREInstr, PREPred);
2242  }
2243 
2244  VN.add(Phi, ValNo);
2245  // After creating a new PHI for ValNo, the phi translate result for ValNo will
2246  // be changed, so erase the related stale entries in phi translate cache.
2247  VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2248  addToLeaderTable(ValNo, Phi, CurrentBlock);
2249  Phi->setDebugLoc(CurInst->getDebugLoc());
2250  CurInst->replaceAllUsesWith(Phi);
2251  if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2252  MD->invalidateCachedPointerInfo(Phi);
2253  VN.erase(CurInst);
2254  removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2255 
2256  DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2257  if (MD)
2258  MD->removeInstruction(CurInst);
2259  DEBUG(verifyRemoved(CurInst));
2260  CurInst->eraseFromParent();
2261  ++NumGVNInstr;
2262 
2263  return true;
2264 }
2265 
2266 /// Perform a purely local form of PRE that looks for diamond
2267 /// control flow patterns and attempts to perform simple PRE at the join point.
2268 bool GVN::performPRE(Function &F) {
2269  bool Changed = false;
2270  for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2271  // Nothing to PRE in the entry block.
2272  if (CurrentBlock == &F.getEntryBlock())
2273  continue;
2274 
2275  // Don't perform PRE on an EH pad.
2276  if (CurrentBlock->isEHPad())
2277  continue;
2278 
2279  for (BasicBlock::iterator BI = CurrentBlock->begin(),
2280  BE = CurrentBlock->end();
2281  BI != BE;) {
2282  Instruction *CurInst = &*BI++;
2283  Changed |= performScalarPRE(CurInst);
2284  }
2285  }
2286 
2287  if (splitCriticalEdges())
2288  Changed = true;
2289 
2290  return Changed;
2291 }
2292 
2293 /// Split the critical edge connecting the given two blocks, and return
2294 /// the block inserted to the critical edge.
2295 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2296  BasicBlock *BB =
2298  if (MD)
2299  MD->invalidateCachedPredecessors();
2300  return BB;
2301 }
2302 
2303 /// Split critical edges found during the previous
2304 /// iteration that may enable further optimization.
2305 bool GVN::splitCriticalEdges() {
2306  if (toSplit.empty())
2307  return false;
2308  do {
2309  std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2310  SplitCriticalEdge(Edge.first, Edge.second,
2312  } while (!toSplit.empty());
2313  if (MD) MD->invalidateCachedPredecessors();
2314  return true;
2315 }
2316 
2317 /// Executes one iteration of GVN
2318 bool GVN::iterateOnFunction(Function &F) {
2319  cleanupGlobalSets();
2320 
2321  // Top-down walk of the dominator tree
2322  bool Changed = false;
2323  // Needed for value numbering with phi construction to work.
2324  // RPOT walks the graph in its constructor and will not be invalidated during
2325  // processBlock.
2327  for (BasicBlock *BB : RPOT)
2328  Changed |= processBlock(BB);
2329 
2330  return Changed;
2331 }
2332 
2333 void GVN::cleanupGlobalSets() {
2334  VN.clear();
2335  LeaderTable.clear();
2336  BlockRPONumber.clear();
2337  TableAllocator.Reset();
2338 }
2339 
2340 /// Verify that the specified instruction does not occur in our
2341 /// internal data structures.
2342 void GVN::verifyRemoved(const Instruction *Inst) const {
2343  VN.verifyRemoved(Inst);
2344 
2345  // Walk through the value number scope to make sure the instruction isn't
2346  // ferreted away in it.
2348  I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2349  const LeaderTableEntry *Node = &I->second;
2350  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2351 
2352  while (Node->Next) {
2353  Node = Node->Next;
2354  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2355  }
2356  }
2357 }
2358 
2359 /// BB is declared dead, which implied other blocks become dead as well. This
2360 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2361 /// live successors, update their phi nodes by replacing the operands
2362 /// corresponding to dead blocks with UndefVal.
2363 void GVN::addDeadBlock(BasicBlock *BB) {
2366 
2367  NewDead.push_back(BB);
2368  while (!NewDead.empty()) {
2369  BasicBlock *D = NewDead.pop_back_val();
2370  if (DeadBlocks.count(D))
2371  continue;
2372 
2373  // All blocks dominated by D are dead.
2375  DT->getDescendants(D, Dom);
2376  DeadBlocks.insert(Dom.begin(), Dom.end());
2377 
2378  // Figure out the dominance-frontier(D).
2379  for (BasicBlock *B : Dom) {
2380  for (BasicBlock *S : successors(B)) {
2381  if (DeadBlocks.count(S))
2382  continue;
2383 
2384  bool AllPredDead = true;
2385  for (BasicBlock *P : predecessors(S))
2386  if (!DeadBlocks.count(P)) {
2387  AllPredDead = false;
2388  break;
2389  }
2390 
2391  if (!AllPredDead) {
2392  // S could be proved dead later on. That is why we don't update phi
2393  // operands at this moment.
2394  DF.insert(S);
2395  } else {
2396  // While S is not dominated by D, it is dead by now. This could take
2397  // place if S already have a dead predecessor before D is declared
2398  // dead.
2399  NewDead.push_back(S);
2400  }
2401  }
2402  }
2403  }
2404 
2405  // For the dead blocks' live successors, update their phi nodes by replacing
2406  // the operands corresponding to dead blocks with UndefVal.
2408  I != E; I++) {
2409  BasicBlock *B = *I;
2410  if (DeadBlocks.count(B))
2411  continue;
2412 
2414  for (BasicBlock *P : Preds) {
2415  if (!DeadBlocks.count(P))
2416  continue;
2417 
2418  if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2419  if (BasicBlock *S = splitCriticalEdges(P, B))
2420  DeadBlocks.insert(P = S);
2421  }
2422 
2423  for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2424  PHINode &Phi = cast<PHINode>(*II);
2426  UndefValue::get(Phi.getType()));
2427  }
2428  }
2429  }
2430 }
2431 
2432 // If the given branch is recognized as a foldable branch (i.e. conditional
2433 // branch with constant condition), it will perform following analyses and
2434 // transformation.
2435 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2436 // R be the target of the dead out-coming edge.
2437 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2438 // edge. The result of this step will be {X| X is dominated by R}
2439 // 2) Identify those blocks which haves at least one dead predecessor. The
2440 // result of this step will be dominance-frontier(R).
2441 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2442 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2443 //
2444 // Return true iff *NEW* dead code are found.
2445 bool GVN::processFoldableCondBr(BranchInst *BI) {
2446  if (!BI || BI->isUnconditional())
2447  return false;
2448 
2449  // If a branch has two identical successors, we cannot declare either dead.
2450  if (BI->getSuccessor(0) == BI->getSuccessor(1))
2451  return false;
2452 
2453  ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2454  if (!Cond)
2455  return false;
2456 
2457  BasicBlock *DeadRoot =
2458  Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2459  if (DeadBlocks.count(DeadRoot))
2460  return false;
2461 
2462  if (!DeadRoot->getSinglePredecessor())
2463  DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2464 
2465  addDeadBlock(DeadRoot);
2466  return true;
2467 }
2468 
2469 // performPRE() will trigger assert if it comes across an instruction without
2470 // associated val-num. As it normally has far more live instructions than dead
2471 // instructions, it makes more sense just to "fabricate" a val-number for the
2472 // dead code than checking if instruction involved is dead or not.
2473 void GVN::assignValNumForDeadCode() {
2474  for (BasicBlock *BB : DeadBlocks) {
2475  for (Instruction &Inst : *BB) {
2476  unsigned ValNum = VN.lookupOrAdd(&Inst);
2477  addToLeaderTable(ValNum, &Inst, BB);
2478  }
2479  }
2480 }
2481 
2483 public:
2484  static char ID; // Pass identification, replacement for typeid
2485 
2486  explicit GVNLegacyPass(bool NoLoads = false)
2487  : FunctionPass(ID), NoLoads(NoLoads) {
2489  }
2490 
2491  bool runOnFunction(Function &F) override {
2492  if (skipFunction(F))
2493  return false;
2494 
2495  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2496 
2497  return Impl.runImpl(
2498  F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2499  getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2500  getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2501  getAnalysis<AAResultsWrapperPass>().getAAResults(),
2502  NoLoads ? nullptr
2503  : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2504  LIWP ? &LIWP->getLoopInfo() : nullptr,
2505  &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2506  }
2507 
2508  void getAnalysisUsage(AnalysisUsage &AU) const override {
2512  if (!NoLoads)
2515 
2520  }
2521 
2522 private:
2523  bool NoLoads;
2524  GVN Impl;
2525 };
2526 
2527 char GVNLegacyPass::ID = 0;
2528 
2529 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2537 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2538 
2539 // The public interface to this file...
2541  return new GVNLegacyPass(NoLoads);
2542 }
Legacy wrapper pass to provide the GlobalsAAResult object.
static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV)
Definition: GVN.cpp:236
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:574
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:69
FunctionPass * createGVNPass(bool NoLoads=false)
Create a legacy GVN pass.
Definition: GVN.cpp:2540
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:1586
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:109
bool isUndefValue() const
Definition: GVN.cpp:204
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:523
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:850
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:219
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:177
size_type size() const
Definition: MapVector.h:57
unsigned Offset
Offset - The byte offset in Val that is interesting for the load query.
Definition: GVN.cpp:167
DiagnosticInfoOptimizationBase::Argument NV
static Type * makeCmpResultType(Type *opnd_type)
Create a result type for fcmp/icmp.
Definition: InstrTypes.h:1074
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
PointerTy getPointer() const
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:449
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
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
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:585
iterator end()
Definition: Function.h:590
void AddAvailableValue(BasicBlock *BB, Value *V)
Indicate that a rewritten value is available in the specified block with the specified value...
Definition: SSAUpdater.cpp:67
bool operator==(const Expression &other) const
Definition: GVN.cpp:112
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 .assume calls within a function.
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:233
uint32_t phiTranslate(const BasicBlock *BB, const BasicBlock *PhiBlock, uint32_t Num, GVN &Gvn)
Wrap phiTranslateImpl to provide caching functionality.
Definition: GVN.cpp:1522
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:876
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:93
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.h:262
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:238
F(f)
bool isCoercedLoadValue() const
Definition: GVN.cpp:202
An instruction for reading from memory.
Definition: Instructions.h:164
const BasicBlock * getEnd() const
Definition: Dominators.h:90
Hexagon Common GEP
Value * getCondition() const
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:1824
Use * op_iterator
Definition: User.h:209
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:475
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:33
op_iterator op_begin()
Definition: User.h:214
gvn Early GVN Hoisting of Expressions
Definition: GVNHoist.cpp:1170
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:207
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
uint32_t lookup(Value *V, bool Verify=true) const
Returns the value number of the specified value.
Definition: GVN.cpp:572
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:191
void dump() const
Support for debugging, callable in GDB: V->dump()
Definition: AsmWriter.cpp:3627
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:217
static cl::opt< uint32_t > MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, cl::desc("Max recurse depth (default = 1000)"))
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: GVN.cpp:2508
static cl::opt< bool > EnablePRE("enable-pre", cl::init(true), cl::Hidden)
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:361
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: GVN.cpp:2491
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:593
unsigned getNumArgOperands() const
Return the number of call arguments.
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:639
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:251
static bool runImpl(CallGraphSCC &SCC, AARGetterT AARGetter)
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:82
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:284
Analysis pass that exposes the LoopInfo for a function.
Definition: LoopInfo.h:871
static const uint16_t * lookup(unsigned opcode, unsigned domain, ArrayRef< uint16_t[3]> Table)
bool isSimpleValue() const
Definition: GVN.cpp:201
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: ...
static void patchReplacementInstruction(Instruction *I, Value *Repl)
Definition: GVN.cpp:1406
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:66
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:110
void andIRFlags(const Value *V)
Logical &#39;and&#39; of any supported wrapping, exact, and fast-math flags of V and this instruction...
#define DEBUG_TYPE
Definition: GVN.cpp:85
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:194
DiagnosticInfoOptimizationBase::setExtraArgs setExtraArgs
static AvailableValue getLoad(LoadInst *LI, unsigned Offset=0)
Definition: GVN.cpp:185
hash_code hash_value(const APFloat &Arg)
See friend declarations above.
Definition: APFloat.cpp:4425
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
LoadInst * getCoercedLoadValue() const
Definition: GVN.cpp:211
static GVN::Expression getEmptyKey()
Definition: GVN.cpp:134
An instruction for storing to memory.
Definition: Instructions.h:306
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:381
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:428
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:290
iterator begin()
Definition: Function.h:588
static unsigned getHashValue(const GVN::Expression &e)
Definition: GVN.cpp:137
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:140
Value * getOperand(unsigned i) const
Definition: User.h:154
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:202
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:572
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:837
void getAAMetadata(AAMDNodes &N, bool Merge=false) const
Fills the AAMDNodes structure with AA metadata from this instruction.
BasicBlock * SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum, const CriticalEdgeSplittingOptions &Options=CriticalEdgeSplittingOptions())
If this edge is a critical edge, insert a new node to split the critical edge.
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
Value * GetValueInMiddleOfBlock(BasicBlock *BB)
Construct SSA form, materializing a value that is live in the middle of the specified block...
Definition: SSAUpdater.cpp:95
SmallVector< uint32_t, 4 > varargs
Definition: GVN.cpp:108
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:153
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:281
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:217
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:75
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
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:580
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:135
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:743
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:116
static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS)
Definition: GVN.cpp:143
This file contains the declarations for the subclasses of Constant, which represent the different fla...
const Instruction & front() const
Definition: BasicBlock.h:264
A manager for alias analyses.
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:504
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:264
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:216
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:823
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:860
PointerIntPair< Value *, 2, ValType > Val
V - The value that is live out of the block.
Definition: GVN.cpp:164
MemIntrinsic * getMemIntrinValue() const
Definition: GVN.cpp:216
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:222
Value * getPointerOperand()
Definition: Instructions.h:270
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:363
GVNLegacyPass(bool NoLoads=false)
Definition: GVN.cpp:2486
static void reportLoadElim(LoadInst *LI, Value *AvailableValue, OptimizationRemarkEmitter *ORE)
Definition: GVN.cpp:1254
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1320
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
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:3559
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:80
unsigned first
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:829
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:224
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:254
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:176
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:410
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:254
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:239
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:132
#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:642
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:385
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:560
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:110
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:774
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:516
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:420
void setOperand(unsigned i, Value *Val)
Definition: User.h:159
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:923
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:57
hash_code hash_combine(const Ts &...args)
Combine values into a single hash_code.
Definition: Hashing.h:602
Represents an AvailableValue which can be rematerialized at the end of the associated BasicBlock...
Definition: GVN.cpp:229
typename SuperClass::iterator iterator
Definition: SmallVector.h:328
iterator_range< user_iterator > users()
Definition: Value.h:395
hash_code hash_combine_range(InputIteratorT first, InputIteratorT last)
Compute a hash_code for a sequence of values.
Definition: Hashing.h:480
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:482
void verifyRemoved(const Value *) const
verifyRemoved - Verify that the value is removed from all internal data structures.
Definition: GVN.cpp:615
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:398
void erase(Value *v)
Remove a value from the value numbering.
Definition: GVN.cpp:605
static bool isLifetimeStart(const Instruction *Inst)
Definition: GVN.cpp:821
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:529
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:120
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:284
bool MergeBlockIntoPredecessor(BasicBlock *BB, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, MemoryDependenceResults *MemDep=nullptr)
Attempts to merge a block into its predecessor, if possible.
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:226
Instruction * getInst() const
If this is a normal dependency, returns the instruction that is depended on.
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:343
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:218
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
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:245
#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:169
uint32_t opcode
Definition: GVN.cpp:105
PassT::Result * getCachedResult(IRUnitT &IR) const
Get the cached result of an analysis pass for a given IR unit.
Definition: PassManager.h:706
bool exists(Value *V) const
Returns true if a value number exists for the specified value.
Definition: GVN.cpp:494
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:174
bool isUnconditional() const
friend hash_code hash_value(const Expression &Value)
Definition: GVN.cpp:124
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:498
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:181
Value * getSimpleValue() const
Definition: GVN.cpp:206
Analysis pass providing the TargetLibraryInfo.
iterator_range< df_iterator< T > > depth_first(const T &G)
Multiway switch.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
unsigned getNumSuccessors() const
Return the number of successors that this terminator has.
const BasicBlock * getStart() const
Definition: Dominators.h:86
Represents a particular available value that we know how to materialize.
Definition: GVN.cpp:154
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:670
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:863
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:293
LLVM Value Representation.
Definition: Value.h:73
succ_range successors(BasicBlock *BB)
Definition: CFG.h:143
static AvailableValueInBlock getUndef(BasicBlock *BB)
Definition: GVN.cpp:248
void removeInstruction(Instruction *InstToRemove)
Removes an instruction from the dependence analysis, updating the dependence of instructions that pre...
OptimizationRemarkEmitter legacy analysis pass.
#define DEBUG(X)
Definition: Debug.h:118
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: GVN.cpp:626
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:408
Predicate getSwappedPredicate() const
For example, EQ->EQ, SLE->SGE, ULT->UGT, OEQ->OEQ, ULE->UGE, OLT->OGT, etc.
Definition: InstrTypes.h:974
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:231
static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl)
Definition: GVN.cpp:1437
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:267
bool isMemIntrinValue() const
Definition: GVN.cpp:203
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
const TerminatorInst * 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:120
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:234
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
void combineMetadata(Instruction *K, const Instruction *J, ArrayRef< unsigned > KnownIDs)
Combine the metadata of two instructions so that K can replace J.
Definition: Local.cpp:1711
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:322
static AvailableValue getUndef()
Definition: GVN.cpp:193
static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, DominatorTree *DT)
There is an edge from &#39;Src&#39; to &#39;Dst&#39;.
Definition: GVN.cpp:1626
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:66
This instruction inserts a struct field of array element value into an aggregate value.
bool HasValueForBlock(BasicBlock *BB) const
Return true if the SSAUpdater already has a value for the specified block.
Definition: SSAUpdater.cpp:63
bool isCriticalEdge(const TerminatorInst *TI, unsigned SuccNum, bool AllowIdenticalEdges=false)
Return true if the specified edge is a critical edge.
Definition: CFG.cpp:88