LLVM  7.0.0svn
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"
43 #include "llvm/Config/llvm-config.h"
44 #include "llvm/IR/Attributes.h"
45 #include "llvm/IR/BasicBlock.h"
46 #include "llvm/IR/CallSite.h"
47 #include "llvm/IR/Constant.h"
48 #include "llvm/IR/Constants.h"
49 #include "llvm/IR/DataLayout.h"
50 #include "llvm/IR/DebugLoc.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
53 #include "llvm/IR/InstrTypes.h"
54 #include "llvm/IR/Instruction.h"
55 #include "llvm/IR/Instructions.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/Intrinsics.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/Metadata.h"
60 #include "llvm/IR/Module.h"
61 #include "llvm/IR/Operator.h"
62 #include "llvm/IR/PassManager.h"
63 #include "llvm/IR/PatternMatch.h"
64 #include "llvm/IR/Type.h"
65 #include "llvm/IR/Use.h"
66 #include "llvm/IR/Value.h"
67 #include "llvm/Pass.h"
68 #include "llvm/Support/Casting.h"
70 #include "llvm/Support/Compiler.h"
71 #include "llvm/Support/Debug.h"
76 #include <algorithm>
77 #include <cassert>
78 #include <cstdint>
79 #include <utility>
80 #include <vector>
81 
82 using namespace llvm;
83 using namespace llvm::gvn;
84 using namespace llvm::VNCoercion;
85 using namespace PatternMatch;
86 
87 #define DEBUG_TYPE "gvn"
88 
89 STATISTIC(NumGVNInstr, "Number of instructions deleted");
90 STATISTIC(NumGVNLoad, "Number of loads deleted");
91 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
92 STATISTIC(NumGVNBlocks, "Number of blocks merged");
93 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
94 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
95 STATISTIC(NumPRELoad, "Number of loads PRE'd");
96 
97 static cl::opt<bool> EnablePRE("enable-pre",
98  cl::init(true), cl::Hidden);
99 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
100 
101 // Maximum allowed recursion depth.
102 static cl::opt<uint32_t>
103 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
104  cl::desc("Max recurse depth (default = 1000)"));
105 
109  bool commutative = false;
111 
112  Expression(uint32_t o = ~2U) : opcode(o) {}
113 
114  bool operator==(const Expression &other) const {
115  if (opcode != other.opcode)
116  return false;
117  if (opcode == ~0U || opcode == ~1U)
118  return true;
119  if (type != other.type)
120  return false;
121  if (varargs != other.varargs)
122  return false;
123  return true;
124  }
125 
127  return hash_combine(
128  Value.opcode, Value.type,
129  hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
130  }
131 };
132 
133 namespace llvm {
134 
135 template <> struct DenseMapInfo<GVN::Expression> {
136  static inline GVN::Expression getEmptyKey() { return ~0U; }
137  static inline GVN::Expression getTombstoneKey() { return ~1U; }
138 
139  static unsigned getHashValue(const GVN::Expression &e) {
140  using llvm::hash_value;
141 
142  return static_cast<unsigned>(hash_value(e));
143  }
144 
145  static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
146  return LHS == RHS;
147  }
148 };
149 
150 } // end namespace llvm
151 
152 /// Represents a particular available value that we know how to materialize.
153 /// Materialization of an AvailableValue never fails. An AvailableValue is
154 /// implicitly associated with a rematerialization point which is the
155 /// location of the instruction from which it was formed.
157  enum ValType {
158  SimpleVal, // A simple offsetted value that is accessed.
159  LoadVal, // A value produced by a load.
160  MemIntrin, // A memory intrinsic which is loaded from.
161  UndefVal // A UndefValue representing a value from dead block (which
162  // is not yet physically removed from the CFG).
163  };
164 
165  /// V - The value that is live out of the block.
167 
168  /// Offset - The byte offset in Val that is interesting for the load query.
169  unsigned Offset;
170 
171  static AvailableValue get(Value *V, unsigned Offset = 0) {
172  AvailableValue Res;
173  Res.Val.setPointer(V);
174  Res.Val.setInt(SimpleVal);
175  Res.Offset = Offset;
176  return Res;
177  }
178 
179  static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
180  AvailableValue Res;
181  Res.Val.setPointer(MI);
182  Res.Val.setInt(MemIntrin);
183  Res.Offset = Offset;
184  return Res;
185  }
186 
187  static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
188  AvailableValue Res;
189  Res.Val.setPointer(LI);
190  Res.Val.setInt(LoadVal);
191  Res.Offset = Offset;
192  return Res;
193  }
194 
196  AvailableValue Res;
197  Res.Val.setPointer(nullptr);
198  Res.Val.setInt(UndefVal);
199  Res.Offset = 0;
200  return Res;
201  }
202 
203  bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
204  bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
205  bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
206  bool isUndefValue() const { return Val.getInt() == UndefVal; }
207 
209  assert(isSimpleValue() && "Wrong accessor");
210  return Val.getPointer();
211  }
212 
214  assert(isCoercedLoadValue() && "Wrong accessor");
215  return cast<LoadInst>(Val.getPointer());
216  }
217 
219  assert(isMemIntrinValue() && "Wrong accessor");
220  return cast<MemIntrinsic>(Val.getPointer());
221  }
222 
223  /// Emit code at the specified insertion point to adjust the value defined
224  /// here to the specified type. This handles various coercion cases.
225  Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
226  GVN &gvn) const;
227 };
228 
229 /// Represents an AvailableValue which can be rematerialized at the end of
230 /// the associated BasicBlock.
232  /// BB - The basic block in question.
234 
235  /// AV - The actual available value
237 
240  Res.BB = BB;
241  Res.AV = std::move(AV);
242  return Res;
243  }
244 
246  unsigned Offset = 0) {
247  return get(BB, AvailableValue::get(V, Offset));
248  }
249 
251  return get(BB, AvailableValue::getUndef());
252  }
253 
254  /// Emit code at the end of this block to adjust the value defined here to
255  /// the specified type. This handles various coercion cases.
257  return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
258  }
259 };
260 
261 //===----------------------------------------------------------------------===//
262 // ValueTable Internal Functions
263 //===----------------------------------------------------------------------===//
264 
265 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
266  Expression e;
267  e.type = I->getType();
268  e.opcode = I->getOpcode();
269  for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
270  OI != OE; ++OI)
271  e.varargs.push_back(lookupOrAdd(*OI));
272  if (I->isCommutative()) {
273  // Ensure that commutative instructions that only differ by a permutation
274  // of their operands get the same value number by sorting the operand value
275  // numbers. Since all commutative instructions have two operands it is more
276  // efficient to sort by hand rather than using, say, std::sort.
277  assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
278  if (e.varargs[0] > e.varargs[1])
279  std::swap(e.varargs[0], e.varargs[1]);
280  e.commutative = true;
281  }
282 
283  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
284  // Sort the operand value numbers so x<y and y>x get the same value number.
285  CmpInst::Predicate Predicate = C->getPredicate();
286  if (e.varargs[0] > e.varargs[1]) {
287  std::swap(e.varargs[0], e.varargs[1]);
288  Predicate = CmpInst::getSwappedPredicate(Predicate);
289  }
290  e.opcode = (C->getOpcode() << 8) | Predicate;
291  e.commutative = true;
292  } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
293  for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
294  II != IE; ++II)
295  e.varargs.push_back(*II);
296  }
297 
298  return e;
299 }
300 
301 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
303  Value *LHS, Value *RHS) {
304  assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
305  "Not a comparison!");
306  Expression e;
307  e.type = CmpInst::makeCmpResultType(LHS->getType());
308  e.varargs.push_back(lookupOrAdd(LHS));
309  e.varargs.push_back(lookupOrAdd(RHS));
310 
311  // Sort the operand value numbers so x<y and y>x get the same value number.
312  if (e.varargs[0] > e.varargs[1]) {
313  std::swap(e.varargs[0], e.varargs[1]);
314  Predicate = CmpInst::getSwappedPredicate(Predicate);
315  }
316  e.opcode = (Opcode << 8) | Predicate;
317  e.commutative = true;
318  return e;
319 }
320 
321 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
322  assert(EI && "Not an ExtractValueInst?");
323  Expression e;
324  e.type = EI->getType();
325  e.opcode = 0;
326 
328  if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
329  // EI might be an extract from one of our recognised intrinsics. If it
330  // is we'll synthesize a semantically equivalent expression instead on
331  // an extract value expression.
332  switch (I->getIntrinsicID()) {
333  case Intrinsic::sadd_with_overflow:
334  case Intrinsic::uadd_with_overflow:
335  e.opcode = Instruction::Add;
336  break;
337  case Intrinsic::ssub_with_overflow:
338  case Intrinsic::usub_with_overflow:
339  e.opcode = Instruction::Sub;
340  break;
341  case Intrinsic::smul_with_overflow:
342  case Intrinsic::umul_with_overflow:
343  e.opcode = Instruction::Mul;
344  break;
345  default:
346  break;
347  }
348 
349  if (e.opcode != 0) {
350  // Intrinsic recognized. Grab its args to finish building the expression.
351  assert(I->getNumArgOperands() == 2 &&
352  "Expect two args for recognised intrinsics.");
353  e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
354  e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
355  return e;
356  }
357  }
358 
359  // Not a recognised intrinsic. Fall back to producing an extract value
360  // expression.
361  e.opcode = EI->getOpcode();
362  for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
363  OI != OE; ++OI)
364  e.varargs.push_back(lookupOrAdd(*OI));
365 
366  for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
367  II != IE; ++II)
368  e.varargs.push_back(*II);
369 
370  return e;
371 }
372 
373 //===----------------------------------------------------------------------===//
374 // ValueTable External Functions
375 //===----------------------------------------------------------------------===//
376 
377 GVN::ValueTable::ValueTable() = default;
378 GVN::ValueTable::ValueTable(const ValueTable &) = default;
379 GVN::ValueTable::ValueTable(ValueTable &&) = default;
380 GVN::ValueTable::~ValueTable() = default;
381 
382 /// add - Insert a value into the table with a specified value number.
384  valueNumbering.insert(std::make_pair(V, num));
385  if (PHINode *PN = dyn_cast<PHINode>(V))
386  NumberingPhi[num] = PN;
387 }
388 
389 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
390  if (AA->doesNotAccessMemory(C)) {
391  Expression exp = createExpr(C);
392  uint32_t e = assignExpNewValueNum(exp).first;
393  valueNumbering[C] = e;
394  return e;
395  } else if (AA->onlyReadsMemory(C)) {
396  Expression exp = createExpr(C);
397  auto ValNum = assignExpNewValueNum(exp);
398  if (ValNum.second) {
399  valueNumbering[C] = ValNum.first;
400  return ValNum.first;
401  }
402  if (!MD) {
403  uint32_t e = assignExpNewValueNum(exp).first;
404  valueNumbering[C] = e;
405  return e;
406  }
407 
408  MemDepResult local_dep = MD->getDependency(C);
409 
410  if (!local_dep.isDef() && !local_dep.isNonLocal()) {
411  valueNumbering[C] = nextValueNumber;
412  return nextValueNumber++;
413  }
414 
415  if (local_dep.isDef()) {
416  CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
417 
418  if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
419  valueNumbering[C] = nextValueNumber;
420  return nextValueNumber++;
421  }
422 
423  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
424  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
425  uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
426  if (c_vn != cd_vn) {
427  valueNumbering[C] = nextValueNumber;
428  return nextValueNumber++;
429  }
430  }
431 
432  uint32_t v = lookupOrAdd(local_cdep);
433  valueNumbering[C] = v;
434  return v;
435  }
436 
437  // Non-local case.
439  MD->getNonLocalCallDependency(CallSite(C));
440  // FIXME: Move the checking logic to MemDep!
441  CallInst* cdep = nullptr;
442 
443  // Check to see if we have a single dominating call instruction that is
444  // identical to C.
445  for (unsigned i = 0, e = deps.size(); i != e; ++i) {
446  const NonLocalDepEntry *I = &deps[i];
447  if (I->getResult().isNonLocal())
448  continue;
449 
450  // We don't handle non-definitions. If we already have a call, reject
451  // instruction dependencies.
452  if (!I->getResult().isDef() || cdep != nullptr) {
453  cdep = nullptr;
454  break;
455  }
456 
457  CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
458  // FIXME: All duplicated with non-local case.
459  if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
460  cdep = NonLocalDepCall;
461  continue;
462  }
463 
464  cdep = nullptr;
465  break;
466  }
467 
468  if (!cdep) {
469  valueNumbering[C] = nextValueNumber;
470  return nextValueNumber++;
471  }
472 
473  if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
474  valueNumbering[C] = nextValueNumber;
475  return nextValueNumber++;
476  }
477  for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
478  uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
479  uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
480  if (c_vn != cd_vn) {
481  valueNumbering[C] = nextValueNumber;
482  return nextValueNumber++;
483  }
484  }
485 
486  uint32_t v = lookupOrAdd(cdep);
487  valueNumbering[C] = v;
488  return v;
489  } else {
490  valueNumbering[C] = nextValueNumber;
491  return nextValueNumber++;
492  }
493 }
494 
495 /// Returns true if a value number exists for the specified value.
496 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
497 
498 /// lookup_or_add - Returns the value number for the specified value, assigning
499 /// it a new number if it did not have one before.
501  DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
502  if (VI != valueNumbering.end())
503  return VI->second;
504 
505  if (!isa<Instruction>(V)) {
506  valueNumbering[V] = nextValueNumber;
507  return nextValueNumber++;
508  }
509 
510  Instruction* I = cast<Instruction>(V);
511  Expression exp;
512  switch (I->getOpcode()) {
513  case Instruction::Call:
514  return lookupOrAddCall(cast<CallInst>(I));
515  case Instruction::Add:
516  case Instruction::FAdd:
517  case Instruction::Sub:
518  case Instruction::FSub:
519  case Instruction::Mul:
520  case Instruction::FMul:
521  case Instruction::UDiv:
522  case Instruction::SDiv:
523  case Instruction::FDiv:
524  case Instruction::URem:
525  case Instruction::SRem:
526  case Instruction::FRem:
527  case Instruction::Shl:
528  case Instruction::LShr:
529  case Instruction::AShr:
530  case Instruction::And:
531  case Instruction::Or:
532  case Instruction::Xor:
533  case Instruction::ICmp:
534  case Instruction::FCmp:
535  case Instruction::Trunc:
536  case Instruction::ZExt:
537  case Instruction::SExt:
538  case Instruction::FPToUI:
539  case Instruction::FPToSI:
540  case Instruction::UIToFP:
541  case Instruction::SIToFP:
542  case Instruction::FPTrunc:
543  case Instruction::FPExt:
544  case Instruction::PtrToInt:
545  case Instruction::IntToPtr:
546  case Instruction::BitCast:
547  case Instruction::Select:
548  case Instruction::ExtractElement:
549  case Instruction::InsertElement:
550  case Instruction::ShuffleVector:
551  case Instruction::InsertValue:
552  case Instruction::GetElementPtr:
553  exp = createExpr(I);
554  break;
555  case Instruction::ExtractValue:
556  exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
557  break;
558  case Instruction::PHI:
559  valueNumbering[V] = nextValueNumber;
560  NumberingPhi[nextValueNumber] = cast<PHINode>(V);
561  return nextValueNumber++;
562  default:
563  valueNumbering[V] = nextValueNumber;
564  return nextValueNumber++;
565  }
566 
567  uint32_t e = assignExpNewValueNum(exp).first;
568  valueNumbering[V] = e;
569  return e;
570 }
571 
572 /// Returns the value number of the specified value. Fails if
573 /// the value has not yet been numbered.
575  DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
576  if (Verify) {
577  assert(VI != valueNumbering.end() && "Value not numbered?");
578  return VI->second;
579  }
580  return (VI != valueNumbering.end()) ? VI->second : 0;
581 }
582 
583 /// Returns the value number of the given comparison,
584 /// assigning it a new number if it did not have one before. Useful when
585 /// we deduced the result of a comparison, but don't immediately have an
586 /// instruction realizing that comparison to hand.
588  CmpInst::Predicate Predicate,
589  Value *LHS, Value *RHS) {
590  Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
591  return assignExpNewValueNum(exp).first;
592 }
593 
594 /// Remove all entries from the ValueTable.
596  valueNumbering.clear();
597  expressionNumbering.clear();
598  NumberingPhi.clear();
599  PhiTranslateTable.clear();
600  nextValueNumber = 1;
601  Expressions.clear();
602  ExprIdx.clear();
603  nextExprNumber = 0;
604 }
605 
606 /// Remove a value from the value numbering.
608  uint32_t Num = valueNumbering.lookup(V);
609  valueNumbering.erase(V);
610  // If V is PHINode, V <--> value number is an one-to-one mapping.
611  if (isa<PHINode>(V))
612  NumberingPhi.erase(Num);
613 }
614 
615 /// verifyRemoved - Verify that the value is removed from all internal data
616 /// structures.
617 void GVN::ValueTable::verifyRemoved(const Value *V) const {
619  I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
620  assert(I->first != V && "Inst still occurs in value numbering map!");
621  }
622 }
623 
624 //===----------------------------------------------------------------------===//
625 // GVN Pass
626 //===----------------------------------------------------------------------===//
627 
629  // FIXME: The order of evaluation of these 'getResult' calls is very
630  // significant! Re-ordering these variables will cause GVN when run alone to
631  // be less effective! We should fix memdep and basic-aa to not exhibit this
632  // behavior, but until then don't change the order here.
633  auto &AC = AM.getResult<AssumptionAnalysis>(F);
634  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
635  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
636  auto &AA = AM.getResult<AAManager>(F);
637  auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
638  auto *LI = AM.getCachedResult<LoopAnalysis>(F);
640  bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
641  if (!Changed)
642  return PreservedAnalyses::all();
645  PA.preserve<GlobalsAA>();
647  return PA;
648 }
649 
650 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
651 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
652  errs() << "{\n";
654  E = d.end(); I != E; ++I) {
655  errs() << I->first << "\n";
656  I->second->dump();
657  }
658  errs() << "}\n";
659 }
660 #endif
661 
662 /// Return true if we can prove that the value
663 /// we're analyzing is fully available in the specified block. As we go, keep
664 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
665 /// map is actually a tri-state map with the following values:
666 /// 0) we know the block *is not* fully available.
667 /// 1) we know the block *is* fully available.
668 /// 2) we do not know whether the block is fully available or not, but we are
669 /// currently speculating that it will be.
670 /// 3) we are speculating for this block and have used that to speculate for
671 /// other blocks.
673  DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
674  uint32_t RecurseDepth) {
675  if (RecurseDepth > MaxRecurseDepth)
676  return false;
677 
678  // Optimistically assume that the block is fully available and check to see
679  // if we already know about this block in one lookup.
680  std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
681  FullyAvailableBlocks.insert(std::make_pair(BB, 2));
682 
683  // If the entry already existed for this block, return the precomputed value.
684  if (!IV.second) {
685  // If this is a speculative "available" value, mark it as being used for
686  // speculation of other blocks.
687  if (IV.first->second == 2)
688  IV.first->second = 3;
689  return IV.first->second != 0;
690  }
691 
692  // Otherwise, see if it is fully available in all predecessors.
693  pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
694 
695  // If this block has no predecessors, it isn't live-in here.
696  if (PI == PE)
697  goto SpeculationFailure;
698 
699  for (; PI != PE; ++PI)
700  // If the value isn't fully available in one of our predecessors, then it
701  // isn't fully available in this block either. Undo our previous
702  // optimistic assumption and bail out.
703  if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
704  goto SpeculationFailure;
705 
706  return true;
707 
708 // If we get here, we found out that this is not, after
709 // all, a fully-available block. We have a problem if we speculated on this and
710 // used the speculation to mark other blocks as available.
711 SpeculationFailure:
712  char &BBVal = FullyAvailableBlocks[BB];
713 
714  // If we didn't speculate on this, just return with it set to false.
715  if (BBVal == 2) {
716  BBVal = 0;
717  return false;
718  }
719 
720  // If we did speculate on this value, we could have blocks set to 1 that are
721  // incorrect. Walk the (transitive) successors of this block and mark them as
722  // 0 if set to one.
723  SmallVector<BasicBlock*, 32> BBWorklist;
724  BBWorklist.push_back(BB);
725 
726  do {
727  BasicBlock *Entry = BBWorklist.pop_back_val();
728  // Note that this sets blocks to 0 (unavailable) if they happen to not
729  // already be in FullyAvailableBlocks. This is safe.
730  char &EntryVal = FullyAvailableBlocks[Entry];
731  if (EntryVal == 0) continue; // Already unavailable.
732 
733  // Mark as unavailable.
734  EntryVal = 0;
735 
736  BBWorklist.append(succ_begin(Entry), succ_end(Entry));
737  } while (!BBWorklist.empty());
738 
739  return false;
740 }
741 
742 /// Given a set of loads specified by ValuesPerBlock,
743 /// construct SSA form, allowing us to eliminate LI. This returns the value
744 /// that should be used at LI's definition site.
747  GVN &gvn) {
748  // Check for the fully redundant, dominating load case. In this case, we can
749  // just use the dominating value directly.
750  if (ValuesPerBlock.size() == 1 &&
751  gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
752  LI->getParent())) {
753  assert(!ValuesPerBlock[0].AV.isUndefValue() &&
754  "Dead BB dominate this block");
755  return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
756  }
757 
758  // Otherwise, we have to construct SSA form.
759  SmallVector<PHINode*, 8> NewPHIs;
760  SSAUpdater SSAUpdate(&NewPHIs);
761  SSAUpdate.Initialize(LI->getType(), LI->getName());
762 
763  for (const AvailableValueInBlock &AV : ValuesPerBlock) {
764  BasicBlock *BB = AV.BB;
765 
766  if (SSAUpdate.HasValueForBlock(BB))
767  continue;
768 
769  SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
770  }
771 
772  // Perform PHI construction.
773  return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
774 }
775 
777  Instruction *InsertPt,
778  GVN &gvn) const {
779  Value *Res;
780  Type *LoadTy = LI->getType();
781  const DataLayout &DL = LI->getModule()->getDataLayout();
782  if (isSimpleValue()) {
783  Res = getSimpleValue();
784  if (Res->getType() != LoadTy) {
785  Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
786 
787  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset
788  << " " << *getSimpleValue() << '\n'
789  << *Res << '\n'
790  << "\n\n\n");
791  }
792  } else if (isCoercedLoadValue()) {
793  LoadInst *Load = getCoercedLoadValue();
794  if (Load->getType() == LoadTy && Offset == 0) {
795  Res = Load;
796  } else {
797  Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
798  // We would like to use gvn.markInstructionForDeletion here, but we can't
799  // because the load is already memoized into the leader map table that GVN
800  // tracks. It is potentially possible to remove the load from the table,
801  // but then there all of the operations based on it would need to be
802  // rehashed. Just leave the dead load around.
803  gvn.getMemDep().removeInstruction(Load);
804  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset
805  << " " << *getCoercedLoadValue() << '\n'
806  << *Res << '\n'
807  << "\n\n\n");
808  }
809  } else if (isMemIntrinValue()) {
810  Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
811  InsertPt, DL);
812  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
813  << " " << *getMemIntrinValue() << '\n'
814  << *Res << '\n'
815  << "\n\n\n");
816  } else {
817  assert(isUndefValue() && "Should be UndefVal");
818  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
819  return UndefValue::get(LoadTy);
820  }
821  assert(Res && "failed to materialize?");
822  return Res;
823 }
824 
825 static bool isLifetimeStart(const Instruction *Inst) {
826  if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
827  return II->getIntrinsicID() == Intrinsic::lifetime_start;
828  return false;
829 }
830 
831 /// Try to locate the three instruction involved in a missed
832 /// load-elimination case that is due to an intervening store.
834  DominatorTree *DT,
836  using namespace ore;
837 
838  User *OtherAccess = nullptr;
839 
840  OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
841  R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
842  << setExtraArgs();
843 
844  for (auto *U : LI->getPointerOperand()->users())
845  if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
846  DT->dominates(cast<Instruction>(U), LI)) {
847  // FIXME: for now give up if there are multiple memory accesses that
848  // dominate the load. We need further analysis to decide which one is
849  // that we're forwarding from.
850  if (OtherAccess)
851  OtherAccess = nullptr;
852  else
853  OtherAccess = U;
854  }
855 
856  if (OtherAccess)
857  R << " in favor of " << NV("OtherAccess", OtherAccess);
858 
859  R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
860 
861  ORE->emit(R);
862 }
863 
864 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
865  Value *Address, AvailableValue &Res) {
866  assert((DepInfo.isDef() || DepInfo.isClobber()) &&
867  "expected a local dependence");
868  assert(LI->isUnordered() && "rules below are incorrect for ordered access");
869 
870  const DataLayout &DL = LI->getModule()->getDataLayout();
871 
872  if (DepInfo.isClobber()) {
873  // If the dependence is to a store that writes to a superset of the bits
874  // read by the load, we can extract the bits we need for the load from the
875  // stored value.
876  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
877  // Can't forward from non-atomic to atomic without violating memory model.
878  if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
879  int Offset =
881  if (Offset != -1) {
882  Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
883  return true;
884  }
885  }
886  }
887 
888  // Check to see if we have something like this:
889  // load i32* P
890  // load i8* (P+1)
891  // if we have this, replace the later with an extraction from the former.
892  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
893  // If this is a clobber and L is the first instruction in its block, then
894  // we have the first instruction in the entry block.
895  // Can't forward from non-atomic to atomic without violating memory model.
896  if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
897  int Offset =
898  analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
899 
900  if (Offset != -1) {
901  Res = AvailableValue::getLoad(DepLI, Offset);
902  return true;
903  }
904  }
905  }
906 
907  // If the clobbering value is a memset/memcpy/memmove, see if we can
908  // forward a value on from it.
909  if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
910  if (Address && !LI->isAtomic()) {
912  DepMI, DL);
913  if (Offset != -1) {
914  Res = AvailableValue::getMI(DepMI, Offset);
915  return true;
916  }
917  }
918  }
919  // Nothing known about this clobber, have to be conservative
920  LLVM_DEBUG(
921  // fast print dep, using operator<< on instruction is too slow.
922  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
923  Instruction *I = DepInfo.getInst();
924  dbgs() << " is clobbered by " << *I << '\n';);
925  if (ORE->allowExtraAnalysis(DEBUG_TYPE))
926  reportMayClobberedLoad(LI, DepInfo, DT, ORE);
927 
928  return false;
929  }
930  assert(DepInfo.isDef() && "follows from above");
931 
932  Instruction *DepInst = DepInfo.getInst();
933 
934  // Loading the allocation -> undef.
935  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
936  // Loading immediately after lifetime begin -> undef.
937  isLifetimeStart(DepInst)) {
939  return true;
940  }
941 
942  // Loading from calloc (which zero initializes memory) -> zero
943  if (isCallocLikeFn(DepInst, TLI)) {
945  return true;
946  }
947 
948  if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
949  // Reject loads and stores that are to the same address but are of
950  // different types if we have to. If the stored value is larger or equal to
951  // the loaded value, we can reuse it.
952  if (S->getValueOperand()->getType() != LI->getType() &&
953  !canCoerceMustAliasedValueToLoad(S->getValueOperand(),
954  LI->getType(), DL))
955  return false;
956 
957  // Can't forward from non-atomic to atomic without violating memory model.
958  if (S->isAtomic() < LI->isAtomic())
959  return false;
960 
961  Res = AvailableValue::get(S->getValueOperand());
962  return true;
963  }
964 
965  if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
966  // If the types mismatch and we can't handle it, reject reuse of the load.
967  // If the stored value is larger or equal to the loaded value, we can reuse
968  // it.
969  if (LD->getType() != LI->getType() &&
971  return false;
972 
973  // Can't forward from non-atomic to atomic without violating memory model.
974  if (LD->isAtomic() < LI->isAtomic())
975  return false;
976 
978  return true;
979  }
980 
981  // Unknown def - must be conservative
982  LLVM_DEBUG(
983  // fast print dep, using operator<< on instruction is too slow.
984  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
985  dbgs() << " has unknown def " << *DepInst << '\n';);
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  bool IsSafeToSpeculativelyExecute = isSafeToSpeculativelyExecute(LI);
1052 
1053  // Check that there is no implicit control flow instructions above our load in
1054  // its block. If there is an instruction that doesn't always pass the
1055  // execution to the following instruction, then moving through it may become
1056  // invalid. For example:
1057  //
1058  // int arr[LEN];
1059  // int index = ???;
1060  // ...
1061  // guard(0 <= index && index < LEN);
1062  // use(arr[index]);
1063  //
1064  // It is illegal to move the array access to any point above the guard,
1065  // because if the index is out of bounds we should deoptimize rather than
1066  // access the array.
1067  // Check that there is no guard in this block above our instruction.
1068  if (!IsSafeToSpeculativelyExecute) {
1069  auto It = FirstImplicitControlFlowInsts.find(TmpBB);
1070  if (It != FirstImplicitControlFlowInsts.end()) {
1071  assert(It->second->getParent() == TmpBB &&
1072  "Implicit control flow map broken?");
1073  if (OI->dominates(It->second, LI))
1074  return false;
1075  }
1076  }
1077  while (TmpBB->getSinglePredecessor()) {
1078  TmpBB = TmpBB->getSinglePredecessor();
1079  if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1080  return false;
1081  if (Blockers.count(TmpBB))
1082  return false;
1083 
1084  // If any of these blocks has more than one successor (i.e. if the edge we
1085  // just traversed was critical), then there are other paths through this
1086  // block along which the load may not be anticipated. Hoisting the load
1087  // above this block would be adding the load to execution paths along
1088  // which it was not previously executed.
1089  if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1090  return false;
1091 
1092  // Check that there is no implicit control flow in a block above.
1093  if (!IsSafeToSpeculativelyExecute &&
1094  FirstImplicitControlFlowInsts.count(TmpBB))
1095  return false;
1096  }
1097 
1098  assert(TmpBB);
1099  LoadBB = TmpBB;
1100 
1101  // Check to see how many predecessors have the loaded value fully
1102  // available.
1104  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1105  for (const AvailableValueInBlock &AV : ValuesPerBlock)
1106  FullyAvailableBlocks[AV.BB] = true;
1107  for (BasicBlock *UnavailableBB : UnavailableBlocks)
1108  FullyAvailableBlocks[UnavailableBB] = false;
1109 
1110  SmallVector<BasicBlock *, 4> CriticalEdgePred;
1111  for (BasicBlock *Pred : predecessors(LoadBB)) {
1112  // If any predecessor block is an EH pad that does not allow non-PHI
1113  // instructions before the terminator, we can't PRE the load.
1114  if (Pred->getTerminator()->isEHPad()) {
1115  LLVM_DEBUG(
1116  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1117  << Pred->getName() << "': " << *LI << '\n');
1118  return false;
1119  }
1120 
1121  if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1122  continue;
1123  }
1124 
1125  if (Pred->getTerminator()->getNumSuccessors() != 1) {
1126  if (isa<IndirectBrInst>(Pred->getTerminator())) {
1127  LLVM_DEBUG(
1128  dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1129  << Pred->getName() << "': " << *LI << '\n');
1130  return false;
1131  }
1132 
1133  if (LoadBB->isEHPad()) {
1134  LLVM_DEBUG(
1135  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1136  << Pred->getName() << "': " << *LI << '\n');
1137  return false;
1138  }
1139 
1140  CriticalEdgePred.push_back(Pred);
1141  } else {
1142  // Only add the predecessors that will not be split for now.
1143  PredLoads[Pred] = nullptr;
1144  }
1145  }
1146 
1147  // Decide whether PRE is profitable for this load.
1148  unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1149  assert(NumUnavailablePreds != 0 &&
1150  "Fully available value should already be eliminated!");
1151 
1152  // If this load is unavailable in multiple predecessors, reject it.
1153  // FIXME: If we could restructure the CFG, we could make a common pred with
1154  // all the preds that don't have an available LI and insert a new load into
1155  // that one block.
1156  if (NumUnavailablePreds != 1)
1157  return false;
1158 
1159  // Split critical edges, and update the unavailable predecessors accordingly.
1160  for (BasicBlock *OrigPred : CriticalEdgePred) {
1161  BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1162  assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1163  PredLoads[NewPred] = nullptr;
1164  LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1165  << LoadBB->getName() << '\n');
1166  }
1167 
1168  // Check if the load can safely be moved to all the unavailable predecessors.
1169  bool CanDoPRE = true;
1170  const DataLayout &DL = LI->getModule()->getDataLayout();
1172  for (auto &PredLoad : PredLoads) {
1173  BasicBlock *UnavailablePred = PredLoad.first;
1174 
1175  // Do PHI translation to get its value in the predecessor if necessary. The
1176  // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1177 
1178  // If all preds have a single successor, then we know it is safe to insert
1179  // the load on the pred (?!?), so we can insert code to materialize the
1180  // pointer if it is not available.
1181  PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1182  Value *LoadPtr = nullptr;
1183  LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1184  *DT, NewInsts);
1185 
1186  // If we couldn't find or insert a computation of this phi translated value,
1187  // we fail PRE.
1188  if (!LoadPtr) {
1189  LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1190  << *LI->getPointerOperand() << "\n");
1191  CanDoPRE = false;
1192  break;
1193  }
1194 
1195  PredLoad.second = LoadPtr;
1196  }
1197 
1198  if (!CanDoPRE) {
1199  while (!NewInsts.empty()) {
1200  Instruction *I = NewInsts.pop_back_val();
1201  markInstructionForDeletion(I);
1202  }
1203  // HINT: Don't revert the edge-splitting as following transformation may
1204  // also need to split these critical edges.
1205  return !CriticalEdgePred.empty();
1206  }
1207 
1208  // Okay, we can eliminate this load by inserting a reload in the predecessor
1209  // and using PHI construction to get the value in the other predecessors, do
1210  // it.
1211  LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1212  LLVM_DEBUG(if (!NewInsts.empty()) dbgs()
1213  << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back()
1214  << '\n');
1215 
1216  // Assign value numbers to the new instructions.
1217  for (Instruction *I : NewInsts) {
1218  // Instructions that have been inserted in predecessor(s) to materialize
1219  // the load address do not retain their original debug locations. Doing
1220  // so could lead to confusing (but correct) source attributions.
1221  // FIXME: How do we retain source locations without causing poor debugging
1222  // behavior?
1223  I->setDebugLoc(DebugLoc());
1224 
1225  // FIXME: We really _ought_ to insert these value numbers into their
1226  // parent's availability map. However, in doing so, we risk getting into
1227  // ordering issues. If a block hasn't been processed yet, we would be
1228  // marking a value as AVAIL-IN, which isn't what we intend.
1229  VN.lookupOrAdd(I);
1230  }
1231 
1232  for (const auto &PredLoad : PredLoads) {
1233  BasicBlock *UnavailablePred = PredLoad.first;
1234  Value *LoadPtr = PredLoad.second;
1235 
1236  auto *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre",
1237  LI->isVolatile(), LI->getAlignment(),
1238  LI->getOrdering(), LI->getSyncScopeID(),
1239  UnavailablePred->getTerminator());
1240  NewLoad->setDebugLoc(LI->getDebugLoc());
1241 
1242  // Transfer the old load's AA tags to the new load.
1243  AAMDNodes Tags;
1244  LI->getAAMetadata(Tags);
1245  if (Tags)
1246  NewLoad->setAAMetadata(Tags);
1247 
1248  if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1249  NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1250  if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1251  NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1252  if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1253  NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1254 
1255  // We do not propagate the old load's debug location, because the new
1256  // load now lives in a different BB, and we want to avoid a jumpy line
1257  // table.
1258  // FIXME: How do we retain source locations without causing poor debugging
1259  // behavior?
1260 
1261  // Add the newly created load.
1262  ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1263  NewLoad));
1264  MD->invalidateCachedPointerInfo(LoadPtr);
1265  LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1266  }
1267 
1268  // Perform PHI construction.
1269  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1270  LI->replaceAllUsesWith(V);
1271  if (isa<PHINode>(V))
1272  V->takeName(LI);
1273  if (Instruction *I = dyn_cast<Instruction>(V))
1274  I->setDebugLoc(LI->getDebugLoc());
1275  if (V->getType()->isPtrOrPtrVectorTy())
1276  MD->invalidateCachedPointerInfo(V);
1277  markInstructionForDeletion(LI);
1278  ORE->emit([&]() {
1279  return OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1280  << "load eliminated by PRE";
1281  });
1282  ++NumPRELoad;
1283  return true;
1284 }
1285 
1288  using namespace ore;
1289 
1290  ORE->emit([&]() {
1291  return OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1292  << "load of type " << NV("Type", LI->getType()) << " eliminated"
1293  << setExtraArgs() << " in favor of "
1294  << NV("InfavorOfValue", AvailableValue);
1295  });
1296 }
1297 
1298 /// Attempt to eliminate a load whose dependencies are
1299 /// non-local by performing PHI construction.
1300 bool GVN::processNonLocalLoad(LoadInst *LI) {
1301  // non-local speculations are not allowed under asan.
1302  if (LI->getParent()->getParent()->hasFnAttribute(
1303  Attribute::SanitizeAddress) ||
1305  Attribute::SanitizeHWAddress))
1306  return false;
1307 
1308  // Step 1: Find the non-local dependencies of the load.
1309  LoadDepVect Deps;
1310  MD->getNonLocalPointerDependency(LI, Deps);
1311 
1312  // If we had to process more than one hundred blocks to find the
1313  // dependencies, this load isn't worth worrying about. Optimizing
1314  // it will be too expensive.
1315  unsigned NumDeps = Deps.size();
1316  if (NumDeps > 100)
1317  return false;
1318 
1319  // If we had a phi translation failure, we'll have a single entry which is a
1320  // clobber in the current block. Reject this early.
1321  if (NumDeps == 1 &&
1322  !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1323  LLVM_DEBUG(dbgs() << "GVN: non-local load "; LI->printAsOperand(dbgs());
1324  dbgs() << " has unknown dependencies\n";);
1325  return false;
1326  }
1327 
1328  // If this load follows a GEP, see if we can PRE the indices before analyzing.
1329  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1330  for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1331  OE = GEP->idx_end();
1332  OI != OE; ++OI)
1333  if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1334  performScalarPRE(I);
1335  }
1336 
1337  // Step 2: Analyze the availability of the load
1338  AvailValInBlkVect ValuesPerBlock;
1339  UnavailBlkVect UnavailableBlocks;
1340  AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1341 
1342  // If we have no predecessors that produce a known value for this load, exit
1343  // early.
1344  if (ValuesPerBlock.empty())
1345  return false;
1346 
1347  // Step 3: Eliminate fully redundancy.
1348  //
1349  // If all of the instructions we depend on produce a known value for this
1350  // load, then it is fully redundant and we can use PHI insertion to compute
1351  // its value. Insert PHIs and remove the fully redundant value now.
1352  if (UnavailableBlocks.empty()) {
1353  LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1354 
1355  // Perform PHI construction.
1356  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1357  LI->replaceAllUsesWith(V);
1358 
1359  if (isa<PHINode>(V))
1360  V->takeName(LI);
1361  if (Instruction *I = dyn_cast<Instruction>(V))
1362  // If instruction I has debug info, then we should not update it.
1363  // Also, if I has a null DebugLoc, then it is still potentially incorrect
1364  // to propagate LI's DebugLoc because LI may not post-dominate I.
1365  if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1366  I->setDebugLoc(LI->getDebugLoc());
1367  if (V->getType()->isPtrOrPtrVectorTy())
1368  MD->invalidateCachedPointerInfo(V);
1369  markInstructionForDeletion(LI);
1370  ++NumGVNLoad;
1371  reportLoadElim(LI, V, ORE);
1372  return true;
1373  }
1374 
1375  // Step 4: Eliminate partial redundancy.
1376  if (!EnablePRE || !EnableLoadPRE)
1377  return false;
1378 
1379  return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1380 }
1381 
1382 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1383  assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1384  "This function can only be called with llvm.assume intrinsic");
1385  Value *V = IntrinsicI->getArgOperand(0);
1386 
1387  if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1388  if (Cond->isZero()) {
1389  Type *Int8Ty = Type::getInt8Ty(V->getContext());
1390  // Insert a new store to null instruction before the load to indicate that
1391  // this code is not reachable. FIXME: We could insert unreachable
1392  // instruction directly because we can modify the CFG.
1393  new StoreInst(UndefValue::get(Int8Ty),
1395  IntrinsicI);
1396  }
1397  markInstructionForDeletion(IntrinsicI);
1398  return false;
1399  } else if (isa<Constant>(V)) {
1400  // If it's not false, and constant, it must evaluate to true. This means our
1401  // assume is assume(true), and thus, pointless, and we don't want to do
1402  // anything more here.
1403  return false;
1404  }
1405 
1406  Constant *True = ConstantInt::getTrue(V->getContext());
1407  bool Changed = false;
1408 
1409  for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1410  BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1411 
1412  // This property is only true in dominated successors, propagateEquality
1413  // will check dominance for us.
1414  Changed |= propagateEquality(V, True, Edge, false);
1415  }
1416 
1417  // We can replace assume value with true, which covers cases like this:
1418  // call void @llvm.assume(i1 %cmp)
1419  // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1420  ReplaceWithConstMap[V] = True;
1421 
1422  // If one of *cmp *eq operand is const, adding it to map will cover this:
1423  // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1424  // call void @llvm.assume(i1 %cmp)
1425  // ret float %0 ; will change it to ret float 3.000000e+00
1426  if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1427  if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1428  CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1429  (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1430  CmpI->getFastMathFlags().noNaNs())) {
1431  Value *CmpLHS = CmpI->getOperand(0);
1432  Value *CmpRHS = CmpI->getOperand(1);
1433  if (isa<Constant>(CmpLHS))
1434  std::swap(CmpLHS, CmpRHS);
1435  auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1436 
1437  // If only one operand is constant.
1438  if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1439  ReplaceWithConstMap[CmpLHS] = RHSConst;
1440  }
1441  }
1442  return Changed;
1443 }
1444 
1446  auto *ReplInst = dyn_cast<Instruction>(Repl);
1447  if (!ReplInst)
1448  return;
1449 
1450  // Patch the replacement so that it is not more restrictive than the value
1451  // being replaced.
1452  // Note that if 'I' is a load being replaced by some operation,
1453  // for example, by an arithmetic operation, then andIRFlags()
1454  // would just erase all math flags from the original arithmetic
1455  // operation, which is clearly not wanted and not needed.
1456  if (!isa<LoadInst>(I))
1457  ReplInst->andIRFlags(I);
1458 
1459  // FIXME: If both the original and replacement value are part of the
1460  // same control-flow region (meaning that the execution of one
1461  // guarantees the execution of the other), then we can combine the
1462  // noalias scopes here and do better than the general conservative
1463  // answer used in combineMetadata().
1464 
1465  // In general, GVN unifies expressions over different control-flow
1466  // regions, and so we need a conservative combination of the noalias
1467  // scopes.
1468  static const unsigned KnownIDs[] = {
1473  combineMetadata(ReplInst, I, KnownIDs);
1474 }
1475 
1477  patchReplacementInstruction(I, Repl);
1478  I->replaceAllUsesWith(Repl);
1479 }
1480 
1481 /// Attempt to eliminate a load, first by eliminating it
1482 /// locally, and then attempting non-local elimination if that fails.
1483 bool GVN::processLoad(LoadInst *L) {
1484  if (!MD)
1485  return false;
1486 
1487  // This code hasn't been audited for ordered or volatile memory access
1488  if (!L->isUnordered())
1489  return false;
1490 
1491  if (L->use_empty()) {
1492  markInstructionForDeletion(L);
1493  return true;
1494  }
1495 
1496  // ... to a pointer that has been loaded from before...
1497  MemDepResult Dep = MD->getDependency(L);
1498 
1499  // If it is defined in another block, try harder.
1500  if (Dep.isNonLocal())
1501  return processNonLocalLoad(L);
1502 
1503  // Only handle the local case below
1504  if (!Dep.isDef() && !Dep.isClobber()) {
1505  // This might be a NonFuncLocal or an Unknown
1506  LLVM_DEBUG(
1507  // fast print dep, using operator<< on instruction is too slow.
1508  dbgs() << "GVN: load "; L->printAsOperand(dbgs());
1509  dbgs() << " has unknown dependence\n";);
1510  return false;
1511  }
1512 
1513  AvailableValue AV;
1514  if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1515  Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1516 
1517  // Replace the load!
1518  patchAndReplaceAllUsesWith(L, AvailableValue);
1519  markInstructionForDeletion(L);
1520  ++NumGVNLoad;
1521  reportLoadElim(L, AvailableValue, ORE);
1522  // Tell MDA to rexamine the reused pointer since we might have more
1523  // information after forwarding it.
1524  if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1525  MD->invalidateCachedPointerInfo(AvailableValue);
1526  return true;
1527  }
1528 
1529  return false;
1530 }
1531 
1532 /// Return a pair the first field showing the value number of \p Exp and the
1533 /// second field showing whether it is a value number newly created.
1534 std::pair<uint32_t, bool>
1535 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1536  uint32_t &e = expressionNumbering[Exp];
1537  bool CreateNewValNum = !e;
1538  if (CreateNewValNum) {
1539  Expressions.push_back(Exp);
1540  if (ExprIdx.size() < nextValueNumber + 1)
1541  ExprIdx.resize(nextValueNumber * 2);
1542  e = nextValueNumber;
1543  ExprIdx[nextValueNumber++] = nextExprNumber++;
1544  }
1545  return {e, CreateNewValNum};
1546 }
1547 
1548 /// Return whether all the values related with the same \p num are
1549 /// defined in \p BB.
1550 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1551  GVN &Gvn) {
1552  LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1553  while (Vals && Vals->BB == BB)
1554  Vals = Vals->Next;
1555  return !Vals;
1556 }
1557 
1558 /// Wrap phiTranslateImpl to provide caching functionality.
1560  const BasicBlock *PhiBlock, uint32_t Num,
1561  GVN &Gvn) {
1562  auto FindRes = PhiTranslateTable.find({Num, Pred});
1563  if (FindRes != PhiTranslateTable.end())
1564  return FindRes->second;
1565  uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1566  PhiTranslateTable.insert({{Num, Pred}, NewNum});
1567  return NewNum;
1568 }
1569 
1570 /// Translate value number \p Num using phis, so that it has the values of
1571 /// the phis in BB.
1572 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1573  const BasicBlock *PhiBlock,
1574  uint32_t Num, GVN &Gvn) {
1575  if (PHINode *PN = NumberingPhi[Num]) {
1576  for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1577  if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1578  if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1579  return TransVal;
1580  }
1581  return Num;
1582  }
1583 
1584  // If there is any value related with Num is defined in a BB other than
1585  // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1586  // a backedge. We can do an early exit in that case to save compile time.
1587  if (!areAllValsInBB(Num, PhiBlock, Gvn))
1588  return Num;
1589 
1590  if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1591  return Num;
1592  Expression Exp = Expressions[ExprIdx[Num]];
1593 
1594  for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1595  // For InsertValue and ExtractValue, some varargs are index numbers
1596  // instead of value numbers. Those index numbers should not be
1597  // translated.
1598  if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1599  (i > 0 && Exp.opcode == Instruction::ExtractValue))
1600  continue;
1601  Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1602  }
1603 
1604  if (Exp.commutative) {
1605  assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1606  if (Exp.varargs[0] > Exp.varargs[1]) {
1607  std::swap(Exp.varargs[0], Exp.varargs[1]);
1608  uint32_t Opcode = Exp.opcode >> 8;
1609  if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1610  Exp.opcode = (Opcode << 8) |
1612  static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1613  }
1614  }
1615 
1616  if (uint32_t NewNum = expressionNumbering[Exp])
1617  return NewNum;
1618  return Num;
1619 }
1620 
1621 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1622 /// again.
1624  const BasicBlock &CurrBlock) {
1625  for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1626  auto FindRes = PhiTranslateTable.find({Num, Pred});
1627  if (FindRes != PhiTranslateTable.end())
1628  PhiTranslateTable.erase(FindRes);
1629  }
1630 }
1631 
1632 // In order to find a leader for a given value number at a
1633 // specific basic block, we first obtain the list of all Values for that number,
1634 // and then scan the list to find one whose block dominates the block in
1635 // question. This is fast because dominator tree queries consist of only
1636 // a few comparisons of DFS numbers.
1637 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1638  LeaderTableEntry Vals = LeaderTable[num];
1639  if (!Vals.Val) return nullptr;
1640 
1641  Value *Val = nullptr;
1642  if (DT->dominates(Vals.BB, BB)) {
1643  Val = Vals.Val;
1644  if (isa<Constant>(Val)) return Val;
1645  }
1646 
1647  LeaderTableEntry* Next = Vals.Next;
1648  while (Next) {
1649  if (DT->dominates(Next->BB, BB)) {
1650  if (isa<Constant>(Next->Val)) return Next->Val;
1651  if (!Val) Val = Next->Val;
1652  }
1653 
1654  Next = Next->Next;
1655  }
1656 
1657  return Val;
1658 }
1659 
1660 /// There is an edge from 'Src' to 'Dst'. Return
1661 /// true if every path from the entry block to 'Dst' passes via this edge. In
1662 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1664  DominatorTree *DT) {
1665  // While in theory it is interesting to consider the case in which Dst has
1666  // more than one predecessor, because Dst might be part of a loop which is
1667  // only reachable from Src, in practice it is pointless since at the time
1668  // GVN runs all such loops have preheaders, which means that Dst will have
1669  // been changed to have only one predecessor, namely Src.
1670  const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1671  assert((!Pred || Pred == E.getStart()) &&
1672  "No edge between these basic blocks!");
1673  return Pred != nullptr;
1674 }
1675 
1676 void GVN::assignBlockRPONumber(Function &F) {
1677  uint32_t NextBlockNumber = 1;
1679  for (BasicBlock *BB : RPOT)
1680  BlockRPONumber[BB] = NextBlockNumber++;
1681 }
1682 
1683 // Tries to replace instruction with const, using information from
1684 // ReplaceWithConstMap.
1685 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1686  bool Changed = false;
1687  for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1688  Value *Operand = Instr->getOperand(OpNum);
1689  auto it = ReplaceWithConstMap.find(Operand);
1690  if (it != ReplaceWithConstMap.end()) {
1691  assert(!isa<Constant>(Operand) &&
1692  "Replacing constants with constants is invalid");
1693  LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "
1694  << *it->second << " in instruction " << *Instr << '\n');
1695  Instr->setOperand(OpNum, it->second);
1696  Changed = true;
1697  }
1698  }
1699  return Changed;
1700 }
1701 
1702 /// The given values are known to be equal in every block
1703 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1704 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1705 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1706 /// value starting from the end of Root.Start.
1707 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1708  bool DominatesByEdge) {
1710  Worklist.push_back(std::make_pair(LHS, RHS));
1711  bool Changed = false;
1712  // For speed, compute a conservative fast approximation to
1713  // DT->dominates(Root, Root.getEnd());
1714  const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1715 
1716  while (!Worklist.empty()) {
1717  std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1718  LHS = Item.first; RHS = Item.second;
1719 
1720  if (LHS == RHS)
1721  continue;
1722  assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1723 
1724  // Don't try to propagate equalities between constants.
1725  if (isa<Constant>(LHS) && isa<Constant>(RHS))
1726  continue;
1727 
1728  // Prefer a constant on the right-hand side, or an Argument if no constants.
1729  if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1730  std::swap(LHS, RHS);
1731  assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1732 
1733  // If there is no obvious reason to prefer the left-hand side over the
1734  // right-hand side, ensure the longest lived term is on the right-hand side,
1735  // so the shortest lived term will be replaced by the longest lived.
1736  // This tends to expose more simplifications.
1737  uint32_t LVN = VN.lookupOrAdd(LHS);
1738  if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1739  (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1740  // Move the 'oldest' value to the right-hand side, using the value number
1741  // as a proxy for age.
1742  uint32_t RVN = VN.lookupOrAdd(RHS);
1743  if (LVN < RVN) {
1744  std::swap(LHS, RHS);
1745  LVN = RVN;
1746  }
1747  }
1748 
1749  // If value numbering later sees that an instruction in the scope is equal
1750  // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1751  // the invariant that instructions only occur in the leader table for their
1752  // own value number (this is used by removeFromLeaderTable), do not do this
1753  // if RHS is an instruction (if an instruction in the scope is morphed into
1754  // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1755  // using the leader table is about compiling faster, not optimizing better).
1756  // The leader table only tracks basic blocks, not edges. Only add to if we
1757  // have the simple case where the edge dominates the end.
1758  if (RootDominatesEnd && !isa<Instruction>(RHS))
1759  addToLeaderTable(LVN, RHS, Root.getEnd());
1760 
1761  // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1762  // LHS always has at least one use that is not dominated by Root, this will
1763  // never do anything if LHS has only one use.
1764  if (!LHS->hasOneUse()) {
1765  unsigned NumReplacements =
1766  DominatesByEdge
1767  ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1768  : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1769 
1770  Changed |= NumReplacements > 0;
1771  NumGVNEqProp += NumReplacements;
1772  }
1773 
1774  // Now try to deduce additional equalities from this one. For example, if
1775  // the known equality was "(A != B)" == "false" then it follows that A and B
1776  // are equal in the scope. Only boolean equalities with an explicit true or
1777  // false RHS are currently supported.
1778  if (!RHS->getType()->isIntegerTy(1))
1779  // Not a boolean equality - bail out.
1780  continue;
1781  ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1782  if (!CI)
1783  // RHS neither 'true' nor 'false' - bail out.
1784  continue;
1785  // Whether RHS equals 'true'. Otherwise it equals 'false'.
1786  bool isKnownTrue = CI->isMinusOne();
1787  bool isKnownFalse = !isKnownTrue;
1788 
1789  // If "A && B" is known true then both A and B are known true. If "A || B"
1790  // is known false then both A and B are known false.
1791  Value *A, *B;
1792  if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1793  (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1794  Worklist.push_back(std::make_pair(A, RHS));
1795  Worklist.push_back(std::make_pair(B, RHS));
1796  continue;
1797  }
1798 
1799  // If we are propagating an equality like "(A == B)" == "true" then also
1800  // propagate the equality A == B. When propagating a comparison such as
1801  // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1802  if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1803  Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1804 
1805  // If "A == B" is known true, or "A != B" is known false, then replace
1806  // A with B everywhere in the scope.
1807  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1808  (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1809  Worklist.push_back(std::make_pair(Op0, Op1));
1810 
1811  // Handle the floating point versions of equality comparisons too.
1812  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1813  (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1814 
1815  // Floating point -0.0 and 0.0 compare equal, so we can only
1816  // propagate values if we know that we have a constant and that
1817  // its value is non-zero.
1818 
1819  // FIXME: We should do this optimization if 'no signed zeros' is
1820  // applicable via an instruction-level fast-math-flag or some other
1821  // indicator that relaxed FP semantics are being used.
1822 
1823  if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1824  Worklist.push_back(std::make_pair(Op0, Op1));
1825  }
1826 
1827  // If "A >= B" is known true, replace "A < B" with false everywhere.
1828  CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1829  Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1830  // Since we don't have the instruction "A < B" immediately to hand, work
1831  // out the value number that it would have and use that to find an
1832  // appropriate instruction (if any).
1833  uint32_t NextNum = VN.getNextUnusedValueNumber();
1834  uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1835  // If the number we were assigned was brand new then there is no point in
1836  // looking for an instruction realizing it: there cannot be one!
1837  if (Num < NextNum) {
1838  Value *NotCmp = findLeader(Root.getEnd(), Num);
1839  if (NotCmp && isa<Instruction>(NotCmp)) {
1840  unsigned NumReplacements =
1841  DominatesByEdge
1842  ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1843  : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1844  Root.getStart());
1845  Changed |= NumReplacements > 0;
1846  NumGVNEqProp += NumReplacements;
1847  }
1848  }
1849  // Ensure that any instruction in scope that gets the "A < B" value number
1850  // is replaced with false.
1851  // The leader table only tracks basic blocks, not edges. Only add to if we
1852  // have the simple case where the edge dominates the end.
1853  if (RootDominatesEnd)
1854  addToLeaderTable(Num, NotVal, Root.getEnd());
1855 
1856  continue;
1857  }
1858  }
1859 
1860  return Changed;
1861 }
1862 
1863 /// When calculating availability, handle an instruction
1864 /// by inserting it into the appropriate sets
1865 bool GVN::processInstruction(Instruction *I) {
1866  // Ignore dbg info intrinsics.
1867  if (isa<DbgInfoIntrinsic>(I))
1868  return false;
1869 
1870  // If the instruction can be easily simplified then do so now in preference
1871  // to value numbering it. Value numbering often exposes redundancies, for
1872  // example if it determines that %y is equal to %x then the instruction
1873  // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1874  const DataLayout &DL = I->getModule()->getDataLayout();
1875  if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1876  bool Changed = false;
1877  if (!I->use_empty()) {
1878  I->replaceAllUsesWith(V);
1879  Changed = true;
1880  }
1881  if (isInstructionTriviallyDead(I, TLI)) {
1882  markInstructionForDeletion(I);
1883  Changed = true;
1884  }
1885  if (Changed) {
1886  if (MD && V->getType()->isPtrOrPtrVectorTy())
1887  MD->invalidateCachedPointerInfo(V);
1888  ++NumGVNSimpl;
1889  return true;
1890  }
1891  }
1892 
1893  if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1894  if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1895  return processAssumeIntrinsic(IntrinsicI);
1896 
1897  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1898  if (processLoad(LI))
1899  return true;
1900 
1901  unsigned Num = VN.lookupOrAdd(LI);
1902  addToLeaderTable(Num, LI, LI->getParent());
1903  return false;
1904  }
1905 
1906  // For conditional branches, we can perform simple conditional propagation on
1907  // the condition value itself.
1908  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1909  if (!BI->isConditional())
1910  return false;
1911 
1912  if (isa<Constant>(BI->getCondition()))
1913  return processFoldableCondBr(BI);
1914 
1915  Value *BranchCond = BI->getCondition();
1916  BasicBlock *TrueSucc = BI->getSuccessor(0);
1917  BasicBlock *FalseSucc = BI->getSuccessor(1);
1918  // Avoid multiple edges early.
1919  if (TrueSucc == FalseSucc)
1920  return false;
1921 
1922  BasicBlock *Parent = BI->getParent();
1923  bool Changed = false;
1924 
1925  Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1926  BasicBlockEdge TrueE(Parent, TrueSucc);
1927  Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1928 
1929  Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1930  BasicBlockEdge FalseE(Parent, FalseSucc);
1931  Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1932 
1933  return Changed;
1934  }
1935 
1936  // For switches, propagate the case values into the case destinations.
1937  if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1938  Value *SwitchCond = SI->getCondition();
1939  BasicBlock *Parent = SI->getParent();
1940  bool Changed = false;
1941 
1942  // Remember how many outgoing edges there are to every successor.
1944  for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1945  ++SwitchEdges[SI->getSuccessor(i)];
1946 
1947  for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1948  i != e; ++i) {
1949  BasicBlock *Dst = i->getCaseSuccessor();
1950  // If there is only a single edge, propagate the case value into it.
1951  if (SwitchEdges.lookup(Dst) == 1) {
1952  BasicBlockEdge E(Parent, Dst);
1953  Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1954  }
1955  }
1956  return Changed;
1957  }
1958 
1959  // Instructions with void type don't return a value, so there's
1960  // no point in trying to find redundancies in them.
1961  if (I->getType()->isVoidTy())
1962  return false;
1963 
1964  uint32_t NextNum = VN.getNextUnusedValueNumber();
1965  unsigned Num = VN.lookupOrAdd(I);
1966 
1967  // Allocations are always uniquely numbered, so we can save time and memory
1968  // by fast failing them.
1969  if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
1970  addToLeaderTable(Num, I, I->getParent());
1971  return false;
1972  }
1973 
1974  // If the number we were assigned was a brand new VN, then we don't
1975  // need to do a lookup to see if the number already exists
1976  // somewhere in the domtree: it can't!
1977  if (Num >= NextNum) {
1978  addToLeaderTable(Num, I, I->getParent());
1979  return false;
1980  }
1981 
1982  // Perform fast-path value-number based elimination of values inherited from
1983  // dominators.
1984  Value *Repl = findLeader(I->getParent(), Num);
1985  if (!Repl) {
1986  // Failure, just remember this instance for future use.
1987  addToLeaderTable(Num, I, I->getParent());
1988  return false;
1989  } else if (Repl == I) {
1990  // If I was the result of a shortcut PRE, it might already be in the table
1991  // and the best replacement for itself. Nothing to do.
1992  return false;
1993  }
1994 
1995  // Remove it!
1996  patchAndReplaceAllUsesWith(I, Repl);
1997  if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1998  MD->invalidateCachedPointerInfo(Repl);
1999  markInstructionForDeletion(I);
2000  return true;
2001 }
2002 
2003 /// runOnFunction - This is the main transformation entry point for a function.
2004 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
2005  const TargetLibraryInfo &RunTLI, AAResults &RunAA,
2006  MemoryDependenceResults *RunMD, LoopInfo *LI,
2007  OptimizationRemarkEmitter *RunORE) {
2008  AC = &RunAC;
2009  DT = &RunDT;
2010  VN.setDomTree(DT);
2011  TLI = &RunTLI;
2012  VN.setAliasAnalysis(&RunAA);
2013  MD = RunMD;
2014  OrderedInstructions OrderedInstrs(DT);
2015  OI = &OrderedInstrs;
2016  VN.setMemDep(MD);
2017  ORE = RunORE;
2018 
2019  bool Changed = false;
2020  bool ShouldContinue = true;
2021 
2022  // Merge unconditional branches, allowing PRE to catch more
2023  // optimization opportunities.
2024  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2025  BasicBlock *BB = &*FI++;
2026 
2027  bool removedBlock = MergeBlockIntoPredecessor(BB, DT, LI, MD);
2028  if (removedBlock)
2029  ++NumGVNBlocks;
2030 
2031  Changed |= removedBlock;
2032  }
2033 
2034  unsigned Iteration = 0;
2035  while (ShouldContinue) {
2036  LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2037  ShouldContinue = iterateOnFunction(F);
2038  Changed |= ShouldContinue;
2039  ++Iteration;
2040  }
2041 
2042  if (EnablePRE) {
2043  // Fabricate val-num for dead-code in order to suppress assertion in
2044  // performPRE().
2045  assignValNumForDeadCode();
2046  assignBlockRPONumber(F);
2047  bool PREChanged = true;
2048  while (PREChanged) {
2049  PREChanged = performPRE(F);
2050  Changed |= PREChanged;
2051  }
2052  }
2053 
2054  // FIXME: Should perform GVN again after PRE does something. PRE can move
2055  // computations into blocks where they become fully redundant. Note that
2056  // we can't do this until PRE's critical edge splitting updates memdep.
2057  // Actually, when this happens, we should just fully integrate PRE into GVN.
2058 
2059  cleanupGlobalSets();
2060  // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2061  // iteration.
2062  DeadBlocks.clear();
2063 
2064  return Changed;
2065 }
2066 
2067 bool GVN::processBlock(BasicBlock *BB) {
2068  // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2069  // (and incrementing BI before processing an instruction).
2070  assert(InstrsToErase.empty() &&
2071  "We expect InstrsToErase to be empty across iterations");
2072  if (DeadBlocks.count(BB))
2073  return false;
2074 
2075  // Clearing map before every BB because it can be used only for single BB.
2076  ReplaceWithConstMap.clear();
2077  bool ChangedFunction = false;
2078 
2079  for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2080  BI != BE;) {
2081  if (!ReplaceWithConstMap.empty())
2082  ChangedFunction |= replaceOperandsWithConsts(&*BI);
2083  ChangedFunction |= processInstruction(&*BI);
2084 
2085  if (InstrsToErase.empty()) {
2086  ++BI;
2087  continue;
2088  }
2089 
2090  // If we need some instructions deleted, do it now.
2091  NumGVNInstr += InstrsToErase.size();
2092 
2093  // Avoid iterator invalidation.
2094  bool AtStart = BI == BB->begin();
2095  if (!AtStart)
2096  --BI;
2097 
2098  bool InvalidateImplicitCF = false;
2099  const Instruction *MaybeFirstICF = FirstImplicitControlFlowInsts.lookup(BB);
2100  for (auto *I : InstrsToErase) {
2101  assert(I->getParent() == BB && "Removing instruction from wrong block?");
2102  LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n');
2103  salvageDebugInfo(*I);
2104  if (MD) MD->removeInstruction(I);
2105  LLVM_DEBUG(verifyRemoved(I));
2106  if (MaybeFirstICF == I) {
2107  // We have erased the first ICF in block. The map needs to be updated.
2108  InvalidateImplicitCF = true;
2109  // Do not keep dangling pointer on the erased instruction.
2110  MaybeFirstICF = nullptr;
2111  }
2112  I->eraseFromParent();
2113  }
2114 
2115  OI->invalidateBlock(BB);
2116  InstrsToErase.clear();
2117  if (InvalidateImplicitCF)
2118  fillImplicitControlFlowInfo(BB);
2119 
2120  if (AtStart)
2121  BI = BB->begin();
2122  else
2123  ++BI;
2124  }
2125 
2126  return ChangedFunction;
2127 }
2128 
2129 // Instantiate an expression in a predecessor that lacked it.
2130 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2131  BasicBlock *Curr, unsigned int ValNo) {
2132  // Because we are going top-down through the block, all value numbers
2133  // will be available in the predecessor by the time we need them. Any
2134  // that weren't originally present will have been instantiated earlier
2135  // in this loop.
2136  bool success = true;
2137  for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2138  Value *Op = Instr->getOperand(i);
2139  if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2140  continue;
2141  // This could be a newly inserted instruction, in which case, we won't
2142  // find a value number, and should give up before we hurt ourselves.
2143  // FIXME: Rewrite the infrastructure to let it easier to value number
2144  // and process newly inserted instructions.
2145  if (!VN.exists(Op)) {
2146  success = false;
2147  break;
2148  }
2149  uint32_t TValNo =
2150  VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2151  if (Value *V = findLeader(Pred, TValNo)) {
2152  Instr->setOperand(i, V);
2153  } else {
2154  success = false;
2155  break;
2156  }
2157  }
2158 
2159  // Fail out if we encounter an operand that is not available in
2160  // the PRE predecessor. This is typically because of loads which
2161  // are not value numbered precisely.
2162  if (!success)
2163  return false;
2164 
2165  Instr->insertBefore(Pred->getTerminator());
2166  Instr->setName(Instr->getName() + ".pre");
2167  Instr->setDebugLoc(Instr->getDebugLoc());
2168 
2169  unsigned Num = VN.lookupOrAdd(Instr);
2170  VN.add(Instr, Num);
2171 
2172  // Update the availability map to include the new instruction.
2173  addToLeaderTable(Num, Instr, Pred);
2174  return true;
2175 }
2176 
2177 bool GVN::performScalarPRE(Instruction *CurInst) {
2178  if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2179  isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2180  CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2181  isa<DbgInfoIntrinsic>(CurInst))
2182  return false;
2183 
2184  // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2185  // sinking the compare again, and it would force the code generator to
2186  // move the i1 from processor flags or predicate registers into a general
2187  // purpose register.
2188  if (isa<CmpInst>(CurInst))
2189  return false;
2190 
2191  // We don't currently value number ANY inline asm calls.
2192  if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2193  if (CallI->isInlineAsm())
2194  return false;
2195 
2196  uint32_t ValNo = VN.lookup(CurInst);
2197 
2198  // Look for the predecessors for PRE opportunities. We're
2199  // only trying to solve the basic diamond case, where
2200  // a value is computed in the successor and one predecessor,
2201  // but not the other. We also explicitly disallow cases
2202  // where the successor is its own predecessor, because they're
2203  // more complicated to get right.
2204  unsigned NumWith = 0;
2205  unsigned NumWithout = 0;
2206  BasicBlock *PREPred = nullptr;
2207  BasicBlock *CurrentBlock = CurInst->getParent();
2208 
2210  for (BasicBlock *P : predecessors(CurrentBlock)) {
2211  // We're not interested in PRE where blocks with predecessors that are
2212  // not reachable.
2213  if (!DT->isReachableFromEntry(P)) {
2214  NumWithout = 2;
2215  break;
2216  }
2217  // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2218  // when CurInst has operand defined in CurrentBlock (so it may be defined
2219  // by phi in the loop header).
2220  if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2221  llvm::any_of(CurInst->operands(), [&](const Use &U) {
2222  if (auto *Inst = dyn_cast<Instruction>(U.get()))
2223  return Inst->getParent() == CurrentBlock;
2224  return false;
2225  })) {
2226  NumWithout = 2;
2227  break;
2228  }
2229 
2230  uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2231  Value *predV = findLeader(P, TValNo);
2232  if (!predV) {
2233  predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2234  PREPred = P;
2235  ++NumWithout;
2236  } else if (predV == CurInst) {
2237  /* CurInst dominates this predecessor. */
2238  NumWithout = 2;
2239  break;
2240  } else {
2241  predMap.push_back(std::make_pair(predV, P));
2242  ++NumWith;
2243  }
2244  }
2245 
2246  // Don't do PRE when it might increase code size, i.e. when
2247  // we would need to insert instructions in more than one pred.
2248  if (NumWithout > 1 || NumWith == 0)
2249  return false;
2250 
2251  // We may have a case where all predecessors have the instruction,
2252  // and we just need to insert a phi node. Otherwise, perform
2253  // insertion.
2254  Instruction *PREInstr = nullptr;
2255 
2256  if (NumWithout != 0) {
2257  if (!isSafeToSpeculativelyExecute(CurInst)) {
2258  // It is only valid to insert a new instruction if the current instruction
2259  // is always executed. An instruction with implicit control flow could
2260  // prevent us from doing it. If we cannot speculate the execution, then
2261  // PRE should be prohibited.
2262  auto It = FirstImplicitControlFlowInsts.find(CurrentBlock);
2263  if (It != FirstImplicitControlFlowInsts.end()) {
2264  assert(It->second->getParent() == CurrentBlock &&
2265  "Implicit control flow map broken?");
2266  if (OI->dominates(It->second, CurInst))
2267  return false;
2268  }
2269  }
2270 
2271  // Don't do PRE across indirect branch.
2272  if (isa<IndirectBrInst>(PREPred->getTerminator()))
2273  return false;
2274 
2275  // We can't do PRE safely on a critical edge, so instead we schedule
2276  // the edge to be split and perform the PRE the next time we iterate
2277  // on the function.
2278  unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2279  if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2280  toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2281  return false;
2282  }
2283  // We need to insert somewhere, so let's give it a shot
2284  PREInstr = CurInst->clone();
2285  if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2286  // If we failed insertion, make sure we remove the instruction.
2287  LLVM_DEBUG(verifyRemoved(PREInstr));
2288  PREInstr->deleteValue();
2289  return false;
2290  }
2291  }
2292 
2293  // Either we should have filled in the PRE instruction, or we should
2294  // not have needed insertions.
2295  assert(PREInstr != nullptr || NumWithout == 0);
2296 
2297  ++NumGVNPRE;
2298 
2299  // Create a PHI to make the value available in this block.
2300  PHINode *Phi =
2301  PHINode::Create(CurInst->getType(), predMap.size(),
2302  CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2303  for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2304  if (Value *V = predMap[i].first) {
2305  // If we use an existing value in this phi, we have to patch the original
2306  // value because the phi will be used to replace a later value.
2307  patchReplacementInstruction(CurInst, V);
2308  Phi->addIncoming(V, predMap[i].second);
2309  } else
2310  Phi->addIncoming(PREInstr, PREPred);
2311  }
2312 
2313  VN.add(Phi, ValNo);
2314  // After creating a new PHI for ValNo, the phi translate result for ValNo will
2315  // be changed, so erase the related stale entries in phi translate cache.
2316  VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2317  addToLeaderTable(ValNo, Phi, CurrentBlock);
2318  Phi->setDebugLoc(CurInst->getDebugLoc());
2319  CurInst->replaceAllUsesWith(Phi);
2320  if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2321  MD->invalidateCachedPointerInfo(Phi);
2322  VN.erase(CurInst);
2323  removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2324 
2325  LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2326  if (MD)
2327  MD->removeInstruction(CurInst);
2328  LLVM_DEBUG(verifyRemoved(CurInst));
2329  bool InvalidateImplicitCF =
2330  FirstImplicitControlFlowInsts.lookup(CurInst->getParent()) == CurInst;
2331  // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes
2332  // some assertion failures.
2333  OI->invalidateBlock(CurrentBlock);
2334  CurInst->eraseFromParent();
2335  if (InvalidateImplicitCF)
2336  fillImplicitControlFlowInfo(CurrentBlock);
2337  ++NumGVNInstr;
2338 
2339  return true;
2340 }
2341 
2342 /// Perform a purely local form of PRE that looks for diamond
2343 /// control flow patterns and attempts to perform simple PRE at the join point.
2344 bool GVN::performPRE(Function &F) {
2345  bool Changed = false;
2346  for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2347  // Nothing to PRE in the entry block.
2348  if (CurrentBlock == &F.getEntryBlock())
2349  continue;
2350 
2351  // Don't perform PRE on an EH pad.
2352  if (CurrentBlock->isEHPad())
2353  continue;
2354 
2355  for (BasicBlock::iterator BI = CurrentBlock->begin(),
2356  BE = CurrentBlock->end();
2357  BI != BE;) {
2358  Instruction *CurInst = &*BI++;
2359  Changed |= performScalarPRE(CurInst);
2360  }
2361  }
2362 
2363  if (splitCriticalEdges())
2364  Changed = true;
2365 
2366  return Changed;
2367 }
2368 
2369 /// Split the critical edge connecting the given two blocks, and return
2370 /// the block inserted to the critical edge.
2371 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2372  BasicBlock *BB =
2374  if (MD)
2375  MD->invalidateCachedPredecessors();
2376  return BB;
2377 }
2378 
2379 /// Split critical edges found during the previous
2380 /// iteration that may enable further optimization.
2381 bool GVN::splitCriticalEdges() {
2382  if (toSplit.empty())
2383  return false;
2384  do {
2385  std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2386  SplitCriticalEdge(Edge.first, Edge.second,
2388  } while (!toSplit.empty());
2389  if (MD) MD->invalidateCachedPredecessors();
2390  return true;
2391 }
2392 
2393 /// Executes one iteration of GVN
2394 bool GVN::iterateOnFunction(Function &F) {
2395  cleanupGlobalSets();
2396 
2397  // Top-down walk of the dominator tree
2398  bool Changed = false;
2399  // Needed for value numbering with phi construction to work.
2400  // RPOT walks the graph in its constructor and will not be invalidated during
2401  // processBlock.
2403 
2404  for (BasicBlock *BB : RPOT)
2405  fillImplicitControlFlowInfo(BB);
2406  for (BasicBlock *BB : RPOT)
2407  Changed |= processBlock(BB);
2408 
2409  return Changed;
2410 }
2411 
2412 void GVN::cleanupGlobalSets() {
2413  VN.clear();
2414  LeaderTable.clear();
2415  BlockRPONumber.clear();
2416  TableAllocator.Reset();
2417  FirstImplicitControlFlowInsts.clear();
2418 }
2419 
2420 void
2421 GVN::fillImplicitControlFlowInfo(BasicBlock *BB) {
2422  // Make sure that all marked instructions are actually deleted by this point,
2423  // so that we don't need to care about omitting them.
2424  assert(InstrsToErase.empty() && "Filling before removed all marked insns?");
2425  auto MayNotTransferExecutionToSuccessor = [&](const Instruction *I) {
2426  // If a block's instruction doesn't always pass the control to its successor
2427  // instruction, mark the block as having implicit control flow. We use them
2428  // to avoid wrong assumptions of sort "if A is executed and B post-dominates
2429  // A, then B is also executed". This is not true is there is an implicit
2430  // control flow instruction (e.g. a guard) between them.
2431  //
2432  // TODO: Currently, isGuaranteedToTransferExecutionToSuccessor returns false
2433  // for volatile stores and loads because they can trap. The discussion on
2434  // whether or not it is correct is still ongoing. We might want to get rid
2435  // of this logic in the future. Anyways, trapping instructions shouldn't
2436  // introduce implicit control flow, so we explicitly allow them here. This
2437  // must be removed once isGuaranteedToTransferExecutionToSuccessor is fixed.
2439  return false;
2440  if (isa<LoadInst>(I)) {
2441  assert(cast<LoadInst>(I)->isVolatile() &&
2442  "Non-volatile load should transfer execution to successor!");
2443  return false;
2444  }
2445  if (isa<StoreInst>(I)) {
2446  assert(cast<StoreInst>(I)->isVolatile() &&
2447  "Non-volatile store should transfer execution to successor!");
2448  return false;
2449  }
2450  return true;
2451  };
2452  FirstImplicitControlFlowInsts.erase(BB);
2453 
2454  for (auto &I : *BB)
2455  if (MayNotTransferExecutionToSuccessor(&I)) {
2456  FirstImplicitControlFlowInsts[BB] = &I;
2457  break;
2458  }
2459 }
2460 
2461 /// Verify that the specified instruction does not occur in our
2462 /// internal data structures.
2463 void GVN::verifyRemoved(const Instruction *Inst) const {
2464  VN.verifyRemoved(Inst);
2465 
2466  // Walk through the value number scope to make sure the instruction isn't
2467  // ferreted away in it.
2469  I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2470  const LeaderTableEntry *Node = &I->second;
2471  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2472 
2473  while (Node->Next) {
2474  Node = Node->Next;
2475  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2476  }
2477  }
2478 }
2479 
2480 /// BB is declared dead, which implied other blocks become dead as well. This
2481 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2482 /// live successors, update their phi nodes by replacing the operands
2483 /// corresponding to dead blocks with UndefVal.
2484 void GVN::addDeadBlock(BasicBlock *BB) {
2487 
2488  NewDead.push_back(BB);
2489  while (!NewDead.empty()) {
2490  BasicBlock *D = NewDead.pop_back_val();
2491  if (DeadBlocks.count(D))
2492  continue;
2493 
2494  // All blocks dominated by D are dead.
2496  DT->getDescendants(D, Dom);
2497  DeadBlocks.insert(Dom.begin(), Dom.end());
2498 
2499  // Figure out the dominance-frontier(D).
2500  for (BasicBlock *B : Dom) {
2501  for (BasicBlock *S : successors(B)) {
2502  if (DeadBlocks.count(S))
2503  continue;
2504 
2505  bool AllPredDead = true;
2506  for (BasicBlock *P : predecessors(S))
2507  if (!DeadBlocks.count(P)) {
2508  AllPredDead = false;
2509  break;
2510  }
2511 
2512  if (!AllPredDead) {
2513  // S could be proved dead later on. That is why we don't update phi
2514  // operands at this moment.
2515  DF.insert(S);
2516  } else {
2517  // While S is not dominated by D, it is dead by now. This could take
2518  // place if S already have a dead predecessor before D is declared
2519  // dead.
2520  NewDead.push_back(S);
2521  }
2522  }
2523  }
2524  }
2525 
2526  // For the dead blocks' live successors, update their phi nodes by replacing
2527  // the operands corresponding to dead blocks with UndefVal.
2529  I != E; I++) {
2530  BasicBlock *B = *I;
2531  if (DeadBlocks.count(B))
2532  continue;
2533 
2535  for (BasicBlock *P : Preds) {
2536  if (!DeadBlocks.count(P))
2537  continue;
2538 
2539  if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2540  if (BasicBlock *S = splitCriticalEdges(P, B))
2541  DeadBlocks.insert(P = S);
2542  }
2543 
2544  for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2545  PHINode &Phi = cast<PHINode>(*II);
2547  UndefValue::get(Phi.getType()));
2548  }
2549  }
2550  }
2551 }
2552 
2553 // If the given branch is recognized as a foldable branch (i.e. conditional
2554 // branch with constant condition), it will perform following analyses and
2555 // transformation.
2556 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2557 // R be the target of the dead out-coming edge.
2558 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2559 // edge. The result of this step will be {X| X is dominated by R}
2560 // 2) Identify those blocks which haves at least one dead predecessor. The
2561 // result of this step will be dominance-frontier(R).
2562 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2563 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2564 //
2565 // Return true iff *NEW* dead code are found.
2566 bool GVN::processFoldableCondBr(BranchInst *BI) {
2567  if (!BI || BI->isUnconditional())
2568  return false;
2569 
2570  // If a branch has two identical successors, we cannot declare either dead.
2571  if (BI->getSuccessor(0) == BI->getSuccessor(1))
2572  return false;
2573 
2574  ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2575  if (!Cond)
2576  return false;
2577 
2578  BasicBlock *DeadRoot =
2579  Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2580  if (DeadBlocks.count(DeadRoot))
2581  return false;
2582 
2583  if (!DeadRoot->getSinglePredecessor())
2584  DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2585 
2586  addDeadBlock(DeadRoot);
2587  return true;
2588 }
2589 
2590 // performPRE() will trigger assert if it comes across an instruction without
2591 // associated val-num. As it normally has far more live instructions than dead
2592 // instructions, it makes more sense just to "fabricate" a val-number for the
2593 // dead code than checking if instruction involved is dead or not.
2594 void GVN::assignValNumForDeadCode() {
2595  for (BasicBlock *BB : DeadBlocks) {
2596  for (Instruction &Inst : *BB) {
2597  unsigned ValNum = VN.lookupOrAdd(&Inst);
2598  addToLeaderTable(ValNum, &Inst, BB);
2599  }
2600  }
2601 }
2602 
2604 public:
2605  static char ID; // Pass identification, replacement for typeid
2606 
2607  explicit GVNLegacyPass(bool NoLoads = false)
2608  : FunctionPass(ID), NoLoads(NoLoads) {
2610  }
2611 
2612  bool runOnFunction(Function &F) override {
2613  if (skipFunction(F))
2614  return false;
2615 
2616  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2617 
2618  return Impl.runImpl(
2619  F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2620  getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2621  getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2622  getAnalysis<AAResultsWrapperPass>().getAAResults(),
2623  NoLoads ? nullptr
2624  : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2625  LIWP ? &LIWP->getLoopInfo() : nullptr,
2626  &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2627  }
2628 
2629  void getAnalysisUsage(AnalysisUsage &AU) const override {
2633  if (!NoLoads)
2636 
2641  }
2642 
2643 private:
2644  bool NoLoads;
2645  GVN Impl;
2646 };
2647 
2648 char GVNLegacyPass::ID = 0;
2649 
2650 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2658 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2659 
2660 // The public interface to this file...
2662  return new GVNLegacyPass(NoLoads);
2663 }
Legacy wrapper pass to provide the GlobalsAAResult object.
static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV)
Definition: GVN.cpp:238
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:718
uint64_t CallInst * C
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:68
FunctionPass * createGVNPass(bool NoLoads=false)
Create a legacy GVN pass.
Definition: GVN.cpp:2661
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:1623
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:111
bool isUndefValue() const
Definition: GVN.cpp:206
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:584
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:875
static bool runImpl(Function &F, TargetLibraryInfo &TLI, DominatorTree &DT)
This is the entry point for all transforms.
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
raw_ostream & errs()
This returns a reference to a raw_ostream for standard error.
bool MergeBlockIntoPredecessor(BasicBlock *BB, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, MemoryDependenceResults *MemDep=nullptr, DeferredDominance *DDT=nullptr)
Attempts to merge a block into its predecessor, if possible.
Helper class for SSA formation on a set of values defined in multiple blocks.
Definition: SSAUpdater.h:39
Diagnostic information for missed-optimization remarks.
Provides a lazy, caching interface for making common memory aliasing information queries, backed by LLVM&#39;s alias analysis passes.
int analyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, LoadInst *DepLI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the load at De...
Definition: VNCoercion.cpp:225
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
This instruction extracts a struct member or array element value from an aggregate value...
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset=0)
Definition: GVN.cpp:179
size_type size() const
Definition: MapVector.h:61
unsigned Offset
Offset - The byte offset in Val that is interesting for the load query.
Definition: GVN.cpp:169
DiagnosticInfoOptimizationBase::Argument NV
static Type * makeCmpResultType(Type *opnd_type)
Create a result type for fcmp/icmp.
Definition: InstrTypes.h:1127
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
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
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:587
iterator end()
Definition: Function.h:644
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:114
This class represents a function call, abstracting a target machine&#39;s calling convention.
bool isNonLocal() const
Tests if this MemDepResult represents a query that is transparent to the start of the block...
This file contains the declarations for metadata subclasses.
An immutable pass that tracks lazily created AssumptionCache objects.
A cache of @llvm.assume calls within a function.
bool salvageDebugInfo(Instruction &I)
Assuming the instruction I is going to be deleted, attempt to salvage debug users of I by writing the...
Definition: Local.cpp:1598
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:237
uint32_t phiTranslate(const BasicBlock *BB, const BasicBlock *PhiBlock, uint32_t Num, GVN &Gvn)
Wrap phiTranslateImpl to provide caching functionality.
Definition: GVN.cpp:1559
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:901
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:99
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.h:307
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:225
F(f)
bool isCoercedLoadValue() const
Definition: GVN.cpp:204
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:452
An instruction for reading from memory.
Definition: Instructions.h:168
const BasicBlock * getEnd() const
Definition: Dominators.h:92
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:2368
Use * op_iterator
Definition: User.h:225
iterator end()
Get an iterator to the end of the SetVector.
Definition: SetVector.h:93
Value * getMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingMemInst returned an offset, this function can be used to actually perform...
Definition: VNCoercion.cpp:481
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:33
op_iterator op_begin()
Definition: User.h:230
gvn Early GVN Hoisting of Expressions
Definition: GVNHoist.cpp:1205
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:268
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:264
uint32_t lookup(Value *V, bool Verify=true) const
Returns the value number of the specified value.
Definition: GVN.cpp:574
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:4216
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
bool isVolatile() const
Return true if this is a load from a volatile memory location.
Definition: Instructions.h:221
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:2629
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:2612
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:595
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:257
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
MemoryDependenceResults & getMemDep() const
Definition: GVN.h:84
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
This file contains the simple types necessary to represent the attributes associated with functions a...
An analysis that produces MemoryDependenceResults for a function.
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:295
Analysis pass that exposes the LoopInfo for a function.
Definition: LoopInfo.h:939
static const uint16_t * lookup(unsigned opcode, unsigned domain, ArrayRef< uint16_t[3]> Table)
bool isSimpleValue() const
Definition: GVN.cpp:203
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:1445
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
ppc ctr loops PowerPC CTR Loops Verify
bool insert(const value_type &X)
Insert a new element into the SetVector.
Definition: SetVector.h:142
The core GVN pass object.
Definition: GVN.h:68
IntType getInt() const
bool canCoerceMustAliasedValueToLoad(Value *StoredVal, Type *LoadTy, const DataLayout &DL)
Return true if CoerceAvailableValueToLoadType would succeed if it was called.
Definition: VNCoercion.cpp:15
Expression(uint32_t o=~2U)
Definition: GVN.cpp:112
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:87
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:200
DiagnosticInfoOptimizationBase::setExtraArgs setExtraArgs
static AvailableValue getLoad(LoadInst *LI, unsigned Offset=0)
Definition: GVN.cpp:187
hash_code hash_value(const APFloat &Arg)
See friend declarations above.
Definition: APFloat.cpp:4431
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:126
LoadInst * getCoercedLoadValue() const
Definition: GVN.cpp:213
static GVN::Expression getEmptyKey()
Definition: GVN.cpp:136
An instruction for storing to memory.
Definition: Instructions.h:310
bool isMinusOne() const
This function will return true iff every bit in this constant is set to true.
Definition: Constants.h:209
void add(Value *V, uint32_t num)
add - Insert a value into the table with a specified value number.
Definition: GVN.cpp:383
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:439
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:301
iterator begin()
Definition: Function.h:642
static unsigned getHashValue(const GVN::Expression &e)
Definition: GVN.cpp:139
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:142
Value * getOperand(unsigned i) const
Definition: User.h:170
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:106
int analyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, StoreInst *DepSI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the store at D...
Definition: VNCoercion.cpp:208
void initializeGVNLegacyPassPass(PassRegistry &)
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:141
const BasicBlock & getEntryBlock() const
Definition: Function.h:626
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:841
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:410
bool isGuaranteedToTransferExecutionToSuccessor(const Instruction *I)
Return true if this function can prove that the instruction I will always transfer execution to one o...
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:110
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:287
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:235
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:74
LLVM Basic Block Representation.
Definition: BasicBlock.h: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:724
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:137
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:745
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:124
static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS)
Definition: GVN.cpp:145
This file contains the declarations for the subclasses of Constant, which represent the different fla...
const Instruction & front() const
Definition: BasicBlock.h:276
A manager for alias analyses.
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:541
Diagnostic information for applied optimization remarks.
Interval::pred_iterator pred_begin(Interval *I)
pred_begin/pred_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:113
unsigned getNumIndices() const
bool isUnordered() const
Definition: Instructions.h:268
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:232
bool any_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:928
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:885
PointerIntPair< Value *, 2, ValType > Val
V - The value that is live out of the block.
Definition: GVN.cpp:166
MemIntrinsic * getMemIntrinValue() const
Definition: GVN.cpp:218
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:116
op_range operands()
Definition: User.h:238
Value * getPointerOperand()
Definition: Instructions.h:274
Value * getLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingLoad returned an offset, this function can be used to actually perform th...
Definition: VNCoercion.cpp:369
GVNLegacyPass(bool NoLoads=false)
Definition: GVN.cpp:2607
static void reportLoadElim(LoadInst *LI, Value *AvailableValue, OptimizationRemarkEmitter *ORE)
Definition: GVN.cpp:1286
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1392
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
size_t size() const
Definition: SmallVector.h:53
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:4143
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
A memory dependence query can return one of three different answers.
DominatorTree & getDominatorTree() const
Definition: GVN.h:82
unsigned first
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:833
A function analysis which provides an AssumptionCache.
bool isPtrOrPtrVectorTy() const
Return true if this is a pointer type or a vector of pointer types.
Definition: Type.h:227
Value * MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const
Emit code at the end of this block to adjust the value defined here to the specified type...
Definition: GVN.cpp:256
A SetVector that performs no allocations if smaller than a certain size.
Definition: SetVector.h:298
This is the common base class for memset/memcpy/memmove.
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:192
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:418
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
void emit(DiagnosticInfoOptimizationBase &OptDiag)
Output the remark via the diagnostic handler and to the optimization record file. ...
iterator end()
Definition: BasicBlock.h:266
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:244
Module.h This file contains the declarations for the Module class.
Provides information about what library functions are available for the current target.
const MemDepResult & getResult() const
size_type count(const KeyT &Key) const
Definition: MapVector.h:143
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:48
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:642
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:376
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:621
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:113
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:776
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:577
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:457
void setOperand(unsigned i, Value *Val)
Definition: User.h:175
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:133
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:924
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:56
hash_code hash_combine(const Ts &...args)
Combine values into a single hash_code.
Definition: Hashing.h:602
Represents an AvailableValue which can be rematerialized at the end of the associated BasicBlock...
Definition: GVN.cpp:231
iterator_range< user_iterator > users()
Definition: Value.h:399
hash_code hash_combine_range(InputIteratorT first, InputIteratorT last)
Compute a hash_code for a sequence of values.
Definition: Hashing.h: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:473
void verifyRemoved(const Value *) const
verifyRemoved - Verify that the value is removed from all internal data structures.
Definition: GVN.cpp:617
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:389
void erase(Value *v)
Remove a value from the value numbering.
Definition: GVN.cpp:607
static bool isLifetimeStart(const Instruction *Inst)
Definition: GVN.cpp:825
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:546
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:128
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:290
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:230
Instruction * getInst() const
If this is a normal dependency, returns the instruction that is depended on.
void clear()
Definition: ilist.h:309
Value * getStoreValueForLoad(Value *SrcVal, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingStore returned an offset, this function can be used to actually perform t...
Definition: VNCoercion.cpp:349
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:56
unsigned getNumArgOperands() const
Return the number of call arguments.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:224
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:249
#define I(x, y, z)
Definition: MD5.cpp:58
bool mayReadFromMemory() const
Return true if this instruction may read memory.
static AvailableValue get(Value *V, unsigned Offset=0)
Definition: GVN.cpp:171
uint32_t opcode
Definition: GVN.cpp:107
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:496
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:126
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:500
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:208
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:88
Represents a particular available value that we know how to materialize.
Definition: GVN.cpp:156
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
bool isSafeToSpeculativelyExecute(const Value *V, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr)
Return true if the instruction does not have any effects besides calculating the result and does not ...
static bool IsValueFullyAvailableInBlock(BasicBlock *BB, DenseMap< BasicBlock *, char > &FullyAvailableBlocks, uint32_t RecurseDepth)
Return true if we can prove that the value we&#39;re analyzing is fully available in the specified block...
Definition: GVN.cpp:672
0 0 0 1 True if ordered and equal
Definition: InstrTypes.h:888
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:346
LLVM Value Representation.
Definition: Value.h:73
succ_range successors(BasicBlock *BB)
Definition: CFG.h:149
static AvailableValueInBlock getUndef(BasicBlock *BB)
Definition: GVN.cpp:250
void removeInstruction(Instruction *InstToRemove)
Removes an instruction from the dependence analysis, updating the dependence of instructions that pre...
OptimizationRemarkEmitter legacy analysis pass.
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: GVN.cpp:628
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:412
Predicate getSwappedPredicate() const
For example, EQ->EQ, SLE->SGE, ULT->UGT, OEQ->OEQ, ULE->UGE, OLT->OGT, etc.
Definition: InstrTypes.h:999
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:233
static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl)
Definition: GVN.cpp:1476
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:254
bool isMemIntrinValue() const
Definition: GVN.cpp:205
static bool isVolatile(Instruction *Inst)
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:138
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:236
#define LLVM_DEBUG(X)
Definition: Debug.h:119
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:2255
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:322
static AvailableValue getUndef()
Definition: GVN.cpp:195
static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, DominatorTree *DT)
There is an edge from &#39;Src&#39; to &#39;Dst&#39;.
Definition: GVN.cpp:1663
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:67
This instruction inserts a struct field of array element value into an aggregate value.
bool HasValueForBlock(BasicBlock *BB) const
Return true if the SSAUpdater already has a value for the specified block.
Definition: SSAUpdater.cpp:63
bool isCriticalEdge(const TerminatorInst *TI, unsigned SuccNum, bool AllowIdenticalEdges=false)
Return true if the specified edge is a critical edge.
Definition: CFG.cpp:88