LLVM  9.0.0svn
GVN.cpp
Go to the documentation of this file.
1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass performs global value numbering to eliminate fully redundant
10 // instructions. It also performs simple dead load elimination.
11 //
12 // Note that this pass does the value numbering itself; it does not use the
13 // ValueNumbering analysis passes.
14 //
15 //===----------------------------------------------------------------------===//
16 
18 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/Hashing.h"
21 #include "llvm/ADT/MapVector.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SetVector.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/CFG.h"
35 #include "llvm/Analysis/LoopInfo.h"
42 #include "llvm/Config/llvm-config.h"
43 #include "llvm/IR/Attributes.h"
44 #include "llvm/IR/BasicBlock.h"
45 #include "llvm/IR/CallSite.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
49 #include "llvm/IR/DebugLoc.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/IR/InstrTypes.h"
53 #include "llvm/IR/Instruction.h"
54 #include "llvm/IR/Instructions.h"
55 #include "llvm/IR/IntrinsicInst.h"
56 #include "llvm/IR/Intrinsics.h"
57 #include "llvm/IR/LLVMContext.h"
58 #include "llvm/IR/Metadata.h"
59 #include "llvm/IR/Module.h"
60 #include "llvm/IR/Operator.h"
61 #include "llvm/IR/PassManager.h"
62 #include "llvm/IR/PatternMatch.h"
63 #include "llvm/IR/Type.h"
64 #include "llvm/IR/Use.h"
65 #include "llvm/IR/Value.h"
66 #include "llvm/Pass.h"
67 #include "llvm/Support/Casting.h"
69 #include "llvm/Support/Compiler.h"
70 #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 static cl::opt<bool> EnableMemDep("enable-gvn-memdep", cl::init(true));
101 
102 // Maximum allowed recursion depth.
103 static cl::opt<uint32_t>
104 MaxRecurseDepth("gvn-max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
105  cl::desc("Max recurse depth in GVN (default = 1000)"));
106 
108  "gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore,
109  cl::desc("Max number of dependences to attempt Load PRE (default = 100)"));
110 
114  bool commutative = false;
116 
117  Expression(uint32_t o = ~2U) : opcode(o) {}
118 
119  bool operator==(const Expression &other) const {
120  if (opcode != other.opcode)
121  return false;
122  if (opcode == ~0U || opcode == ~1U)
123  return true;
124  if (type != other.type)
125  return false;
126  if (varargs != other.varargs)
127  return false;
128  return true;
129  }
130 
132  return hash_combine(
133  Value.opcode, Value.type,
134  hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
135  }
136 };
137 
138 namespace llvm {
139 
140 template <> struct DenseMapInfo<GVN::Expression> {
141  static inline GVN::Expression getEmptyKey() { return ~0U; }
142  static inline GVN::Expression getTombstoneKey() { return ~1U; }
143 
144  static unsigned getHashValue(const GVN::Expression &e) {
145  using llvm::hash_value;
146 
147  return static_cast<unsigned>(hash_value(e));
148  }
149 
150  static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
151  return LHS == RHS;
152  }
153 };
154 
155 } // end namespace llvm
156 
157 /// Represents a particular available value that we know how to materialize.
158 /// Materialization of an AvailableValue never fails. An AvailableValue is
159 /// implicitly associated with a rematerialization point which is the
160 /// location of the instruction from which it was formed.
162  enum ValType {
163  SimpleVal, // A simple offsetted value that is accessed.
164  LoadVal, // A value produced by a load.
165  MemIntrin, // A memory intrinsic which is loaded from.
166  UndefVal // A UndefValue representing a value from dead block (which
167  // is not yet physically removed from the CFG).
168  };
169 
170  /// V - The value that is live out of the block.
172 
173  /// Offset - The byte offset in Val that is interesting for the load query.
174  unsigned Offset;
175 
176  static AvailableValue get(Value *V, unsigned Offset = 0) {
177  AvailableValue Res;
178  Res.Val.setPointer(V);
179  Res.Val.setInt(SimpleVal);
180  Res.Offset = Offset;
181  return Res;
182  }
183 
184  static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
185  AvailableValue Res;
186  Res.Val.setPointer(MI);
187  Res.Val.setInt(MemIntrin);
188  Res.Offset = Offset;
189  return Res;
190  }
191 
192  static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
193  AvailableValue Res;
194  Res.Val.setPointer(LI);
195  Res.Val.setInt(LoadVal);
196  Res.Offset = Offset;
197  return Res;
198  }
199 
201  AvailableValue Res;
202  Res.Val.setPointer(nullptr);
203  Res.Val.setInt(UndefVal);
204  Res.Offset = 0;
205  return Res;
206  }
207 
208  bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
209  bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
210  bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
211  bool isUndefValue() const { return Val.getInt() == UndefVal; }
212 
214  assert(isSimpleValue() && "Wrong accessor");
215  return Val.getPointer();
216  }
217 
219  assert(isCoercedLoadValue() && "Wrong accessor");
220  return cast<LoadInst>(Val.getPointer());
221  }
222 
224  assert(isMemIntrinValue() && "Wrong accessor");
225  return cast<MemIntrinsic>(Val.getPointer());
226  }
227 
228  /// Emit code at the specified insertion point to adjust the value defined
229  /// here to the specified type. This handles various coercion cases.
230  Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
231  GVN &gvn) const;
232 };
233 
234 /// Represents an AvailableValue which can be rematerialized at the end of
235 /// the associated BasicBlock.
237  /// BB - The basic block in question.
239 
240  /// AV - The actual available value
242 
245  Res.BB = BB;
246  Res.AV = std::move(AV);
247  return Res;
248  }
249 
251  unsigned Offset = 0) {
252  return get(BB, AvailableValue::get(V, Offset));
253  }
254 
256  return get(BB, AvailableValue::getUndef());
257  }
258 
259  /// Emit code at the end of this block to adjust the value defined here to
260  /// the specified type. This handles various coercion cases.
262  return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
263  }
264 };
265 
266 //===----------------------------------------------------------------------===//
267 // ValueTable Internal Functions
268 //===----------------------------------------------------------------------===//
269 
270 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
271  Expression e;
272  e.type = I->getType();
273  e.opcode = I->getOpcode();
274  for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
275  OI != OE; ++OI)
276  e.varargs.push_back(lookupOrAdd(*OI));
277  if (I->isCommutative()) {
278  // Ensure that commutative instructions that only differ by a permutation
279  // of their operands get the same value number by sorting the operand value
280  // numbers. Since all commutative instructions have two operands it is more
281  // efficient to sort by hand rather than using, say, std::sort.
282  assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
283  if (e.varargs[0] > e.varargs[1])
284  std::swap(e.varargs[0], e.varargs[1]);
285  e.commutative = true;
286  }
287 
288  if (CmpInst *C = dyn_cast<CmpInst>(I)) {
289  // Sort the operand value numbers so x<y and y>x get the same value number.
290  CmpInst::Predicate Predicate = C->getPredicate();
291  if (e.varargs[0] > e.varargs[1]) {
292  std::swap(e.varargs[0], e.varargs[1]);
293  Predicate = CmpInst::getSwappedPredicate(Predicate);
294  }
295  e.opcode = (C->getOpcode() << 8) | Predicate;
296  e.commutative = true;
297  } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
298  for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
299  II != IE; ++II)
300  e.varargs.push_back(*II);
301  }
302 
303  return e;
304 }
305 
306 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
308  Value *LHS, Value *RHS) {
309  assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
310  "Not a comparison!");
311  Expression e;
312  e.type = CmpInst::makeCmpResultType(LHS->getType());
313  e.varargs.push_back(lookupOrAdd(LHS));
314  e.varargs.push_back(lookupOrAdd(RHS));
315 
316  // Sort the operand value numbers so x<y and y>x get the same value number.
317  if (e.varargs[0] > e.varargs[1]) {
318  std::swap(e.varargs[0], e.varargs[1]);
319  Predicate = CmpInst::getSwappedPredicate(Predicate);
320  }
321  e.opcode = (Opcode << 8) | Predicate;
322  e.commutative = true;
323  return e;
324 }
325 
326 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
327  assert(EI && "Not an ExtractValueInst?");
328  Expression e;
329  e.type = EI->getType();
330  e.opcode = 0;
331 
333  if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
334  // EI might be an extract from one of our recognised intrinsics. If it
335  // is we'll synthesize a semantically equivalent expression instead on
336  // an extract value expression.
337  switch (I->getIntrinsicID()) {
338  case Intrinsic::sadd_with_overflow:
339  case Intrinsic::uadd_with_overflow:
340  e.opcode = Instruction::Add;
341  break;
342  case Intrinsic::ssub_with_overflow:
343  case Intrinsic::usub_with_overflow:
344  e.opcode = Instruction::Sub;
345  break;
346  case Intrinsic::smul_with_overflow:
347  case Intrinsic::umul_with_overflow:
348  e.opcode = Instruction::Mul;
349  break;
350  default:
351  break;
352  }
353 
354  if (e.opcode != 0) {
355  // Intrinsic recognized. Grab its args to finish building the expression.
356  assert(I->getNumArgOperands() == 2 &&
357  "Expect two args for recognised intrinsics.");
358  e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
359  e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
360  return e;
361  }
362  }
363 
364  // Not a recognised intrinsic. Fall back to producing an extract value
365  // expression.
366  e.opcode = EI->getOpcode();
367  for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
368  OI != OE; ++OI)
369  e.varargs.push_back(lookupOrAdd(*OI));
370 
371  for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
372  II != IE; ++II)
373  e.varargs.push_back(*II);
374 
375  return e;
376 }
377 
378 //===----------------------------------------------------------------------===//
379 // ValueTable External Functions
380 //===----------------------------------------------------------------------===//
381 
382 GVN::ValueTable::ValueTable() = default;
383 GVN::ValueTable::ValueTable(const ValueTable &) = default;
384 GVN::ValueTable::ValueTable(ValueTable &&) = default;
385 GVN::ValueTable::~ValueTable() = default;
386 
387 /// add - Insert a value into the table with a specified value number.
389  valueNumbering.insert(std::make_pair(V, num));
390  if (PHINode *PN = dyn_cast<PHINode>(V))
391  NumberingPhi[num] = PN;
392 }
393 
394 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
395  if (AA->doesNotAccessMemory(C)) {
396  Expression exp = createExpr(C);
397  uint32_t e = assignExpNewValueNum(exp).first;
398  valueNumbering[C] = e;
399  return e;
400  } else if (MD && AA->onlyReadsMemory(C)) {
401  Expression exp = createExpr(C);
402  auto ValNum = assignExpNewValueNum(exp);
403  if (ValNum.second) {
404  valueNumbering[C] = ValNum.first;
405  return ValNum.first;
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(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, bool> 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  // If the value is the load that we will be eliminating, and the block it's
770  // available in is the block that the load is in, then don't add it as
771  // SSAUpdater will resolve the value to the relevant phi which may let it
772  // avoid phi construction entirely if there's actually only one value.
773  if (BB == LI->getParent() &&
774  ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == LI) ||
775  (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == LI)))
776  continue;
777 
778  SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
779  }
780 
781  // Perform PHI construction.
782  return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
783 }
784 
786  Instruction *InsertPt,
787  GVN &gvn) const {
788  Value *Res;
789  Type *LoadTy = LI->getType();
790  const DataLayout &DL = LI->getModule()->getDataLayout();
791  if (isSimpleValue()) {
792  Res = getSimpleValue();
793  if (Res->getType() != LoadTy) {
794  Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
795 
796  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset
797  << " " << *getSimpleValue() << '\n'
798  << *Res << '\n'
799  << "\n\n\n");
800  }
801  } else if (isCoercedLoadValue()) {
802  LoadInst *Load = getCoercedLoadValue();
803  if (Load->getType() == LoadTy && Offset == 0) {
804  Res = Load;
805  } else {
806  Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
807  // We would like to use gvn.markInstructionForDeletion here, but we can't
808  // because the load is already memoized into the leader map table that GVN
809  // tracks. It is potentially possible to remove the load from the table,
810  // but then there all of the operations based on it would need to be
811  // rehashed. Just leave the dead load around.
812  gvn.getMemDep().removeInstruction(Load);
813  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset
814  << " " << *getCoercedLoadValue() << '\n'
815  << *Res << '\n'
816  << "\n\n\n");
817  }
818  } else if (isMemIntrinValue()) {
819  Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
820  InsertPt, DL);
821  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
822  << " " << *getMemIntrinValue() << '\n'
823  << *Res << '\n'
824  << "\n\n\n");
825  } else {
826  assert(isUndefValue() && "Should be UndefVal");
827  LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
828  return UndefValue::get(LoadTy);
829  }
830  assert(Res && "failed to materialize?");
831  return Res;
832 }
833 
834 static bool isLifetimeStart(const Instruction *Inst) {
835  if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
836  return II->getIntrinsicID() == Intrinsic::lifetime_start;
837  return false;
838 }
839 
840 /// Try to locate the three instruction involved in a missed
841 /// load-elimination case that is due to an intervening store.
843  DominatorTree *DT,
845  using namespace ore;
846 
847  User *OtherAccess = nullptr;
848 
849  OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
850  R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
851  << setExtraArgs();
852 
853  for (auto *U : LI->getPointerOperand()->users())
854  if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
855  DT->dominates(cast<Instruction>(U), LI)) {
856  // FIXME: for now give up if there are multiple memory accesses that
857  // dominate the load. We need further analysis to decide which one is
858  // that we're forwarding from.
859  if (OtherAccess)
860  OtherAccess = nullptr;
861  else
862  OtherAccess = U;
863  }
864 
865  if (OtherAccess)
866  R << " in favor of " << NV("OtherAccess", OtherAccess);
867 
868  R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
869 
870  ORE->emit(R);
871 }
872 
873 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
874  Value *Address, AvailableValue &Res) {
875  assert((DepInfo.isDef() || DepInfo.isClobber()) &&
876  "expected a local dependence");
877  assert(LI->isUnordered() && "rules below are incorrect for ordered access");
878 
879  const DataLayout &DL = LI->getModule()->getDataLayout();
880 
881  if (DepInfo.isClobber()) {
882  // If the dependence is to a store that writes to a superset of the bits
883  // read by the load, we can extract the bits we need for the load from the
884  // stored value.
885  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
886  // Can't forward from non-atomic to atomic without violating memory model.
887  if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
888  int Offset =
890  if (Offset != -1) {
891  Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
892  return true;
893  }
894  }
895  }
896 
897  // Check to see if we have something like this:
898  // load i32* P
899  // load i8* (P+1)
900  // if we have this, replace the later with an extraction from the former.
901  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
902  // If this is a clobber and L is the first instruction in its block, then
903  // we have the first instruction in the entry block.
904  // Can't forward from non-atomic to atomic without violating memory model.
905  if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
906  int Offset =
907  analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
908 
909  if (Offset != -1) {
910  Res = AvailableValue::getLoad(DepLI, Offset);
911  return true;
912  }
913  }
914  }
915 
916  // If the clobbering value is a memset/memcpy/memmove, see if we can
917  // forward a value on from it.
918  if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
919  if (Address && !LI->isAtomic()) {
921  DepMI, DL);
922  if (Offset != -1) {
923  Res = AvailableValue::getMI(DepMI, Offset);
924  return true;
925  }
926  }
927  }
928  // Nothing known about this clobber, have to be conservative
929  LLVM_DEBUG(
930  // fast print dep, using operator<< on instruction is too slow.
931  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
932  Instruction *I = DepInfo.getInst();
933  dbgs() << " is clobbered by " << *I << '\n';);
934  if (ORE->allowExtraAnalysis(DEBUG_TYPE))
935  reportMayClobberedLoad(LI, DepInfo, DT, ORE);
936 
937  return false;
938  }
939  assert(DepInfo.isDef() && "follows from above");
940 
941  Instruction *DepInst = DepInfo.getInst();
942 
943  // Loading the allocation -> undef.
944  if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
945  // Loading immediately after lifetime begin -> undef.
946  isLifetimeStart(DepInst)) {
948  return true;
949  }
950 
951  // Loading from calloc (which zero initializes memory) -> zero
952  if (isCallocLikeFn(DepInst, TLI)) {
954  return true;
955  }
956 
957  if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
958  // Reject loads and stores that are to the same address but are of
959  // different types if we have to. If the stored value is larger or equal to
960  // the loaded value, we can reuse it.
961  if (S->getValueOperand()->getType() != LI->getType() &&
962  !canCoerceMustAliasedValueToLoad(S->getValueOperand(),
963  LI->getType(), DL))
964  return false;
965 
966  // Can't forward from non-atomic to atomic without violating memory model.
967  if (S->isAtomic() < LI->isAtomic())
968  return false;
969 
970  Res = AvailableValue::get(S->getValueOperand());
971  return true;
972  }
973 
974  if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
975  // If the types mismatch and we can't handle it, reject reuse of the load.
976  // If the stored value is larger or equal to the loaded value, we can reuse
977  // it.
978  if (LD->getType() != LI->getType() &&
980  return false;
981 
982  // Can't forward from non-atomic to atomic without violating memory model.
983  if (LD->isAtomic() < LI->isAtomic())
984  return false;
985 
987  return true;
988  }
989 
990  // Unknown def - must be conservative
991  LLVM_DEBUG(
992  // fast print dep, using operator<< on instruction is too slow.
993  dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
994  dbgs() << " has unknown def " << *DepInst << '\n';);
995  return false;
996 }
997 
998 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
999  AvailValInBlkVect &ValuesPerBlock,
1000  UnavailBlkVect &UnavailableBlocks) {
1001  // Filter out useless results (non-locals, etc). Keep track of the blocks
1002  // where we have a value available in repl, also keep track of whether we see
1003  // dependencies that produce an unknown value for the load (such as a call
1004  // that could potentially clobber the load).
1005  unsigned NumDeps = Deps.size();
1006  for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1007  BasicBlock *DepBB = Deps[i].getBB();
1008  MemDepResult DepInfo = Deps[i].getResult();
1009 
1010  if (DeadBlocks.count(DepBB)) {
1011  // Dead dependent mem-op disguise as a load evaluating the same value
1012  // as the load in question.
1013  ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1014  continue;
1015  }
1016 
1017  if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1018  UnavailableBlocks.push_back(DepBB);
1019  continue;
1020  }
1021 
1022  // The address being loaded in this non-local block may not be the same as
1023  // the pointer operand of the load if PHI translation occurs. Make sure
1024  // to consider the right address.
1025  Value *Address = Deps[i].getAddress();
1026 
1027  AvailableValue AV;
1028  if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1029  // subtlety: because we know this was a non-local dependency, we know
1030  // it's safe to materialize anywhere between the instruction within
1031  // DepInfo and the end of it's block.
1032  ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1033  std::move(AV)));
1034  } else {
1035  UnavailableBlocks.push_back(DepBB);
1036  }
1037  }
1038 
1039  assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1040  "post condition violation");
1041 }
1042 
1043 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1044  UnavailBlkVect &UnavailableBlocks) {
1045  // Okay, we have *some* definitions of the value. This means that the value
1046  // is available in some of our (transitive) predecessors. Lets think about
1047  // doing PRE of this load. This will involve inserting a new load into the
1048  // predecessor when it's not available. We could do this in general, but
1049  // prefer to not increase code size. As such, we only do this when we know
1050  // that we only have to insert *one* load (which means we're basically moving
1051  // the load, not inserting a new one).
1052 
1053  SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1054  UnavailableBlocks.end());
1055 
1056  // Let's find the first basic block with more than one predecessor. Walk
1057  // backwards through predecessors if needed.
1058  BasicBlock *LoadBB = LI->getParent();
1059  BasicBlock *TmpBB = LoadBB;
1060  bool IsSafeToSpeculativelyExecute = isSafeToSpeculativelyExecute(LI);
1061 
1062  // Check that there is no implicit control flow instructions above our load in
1063  // its block. If there is an instruction that doesn't always pass the
1064  // execution to the following instruction, then moving through it may become
1065  // invalid. For example:
1066  //
1067  // int arr[LEN];
1068  // int index = ???;
1069  // ...
1070  // guard(0 <= index && index < LEN);
1071  // use(arr[index]);
1072  //
1073  // It is illegal to move the array access to any point above the guard,
1074  // because if the index is out of bounds we should deoptimize rather than
1075  // access the array.
1076  // Check that there is no guard in this block above our instruction.
1077  if (!IsSafeToSpeculativelyExecute && ICF->isDominatedByICFIFromSameBlock(LI))
1078  return false;
1079  while (TmpBB->getSinglePredecessor()) {
1080  TmpBB = TmpBB->getSinglePredecessor();
1081  if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1082  return false;
1083  if (Blockers.count(TmpBB))
1084  return false;
1085 
1086  // If any of these blocks has more than one successor (i.e. if the edge we
1087  // just traversed was critical), then there are other paths through this
1088  // block along which the load may not be anticipated. Hoisting the load
1089  // above this block would be adding the load to execution paths along
1090  // which it was not previously executed.
1091  if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1092  return false;
1093 
1094  // Check that there is no implicit control flow in a block above.
1095  if (!IsSafeToSpeculativelyExecute && ICF->hasICF(TmpBB))
1096  return false;
1097  }
1098 
1099  assert(TmpBB);
1100  LoadBB = TmpBB;
1101 
1102  // Check to see how many predecessors have the loaded value fully
1103  // available.
1105  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1106  for (const AvailableValueInBlock &AV : ValuesPerBlock)
1107  FullyAvailableBlocks[AV.BB] = true;
1108  for (BasicBlock *UnavailableBB : UnavailableBlocks)
1109  FullyAvailableBlocks[UnavailableBB] = false;
1110 
1111  SmallVector<BasicBlock *, 4> CriticalEdgePred;
1112  for (BasicBlock *Pred : predecessors(LoadBB)) {
1113  // If any predecessor block is an EH pad that does not allow non-PHI
1114  // instructions before the terminator, we can't PRE the load.
1115  if (Pred->getTerminator()->isEHPad()) {
1116  LLVM_DEBUG(
1117  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1118  << Pred->getName() << "': " << *LI << '\n');
1119  return false;
1120  }
1121 
1122  if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1123  continue;
1124  }
1125 
1126  if (Pred->getTerminator()->getNumSuccessors() != 1) {
1127  if (isa<IndirectBrInst>(Pred->getTerminator())) {
1128  LLVM_DEBUG(
1129  dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1130  << Pred->getName() << "': " << *LI << '\n');
1131  return false;
1132  }
1133 
1134  // FIXME: Can we support the fallthrough edge?
1135  if (isa<CallBrInst>(Pred->getTerminator())) {
1136  LLVM_DEBUG(
1137  dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '"
1138  << Pred->getName() << "': " << *LI << '\n');
1139  return false;
1140  }
1141 
1142  if (LoadBB->isEHPad()) {
1143  LLVM_DEBUG(
1144  dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1145  << Pred->getName() << "': " << *LI << '\n');
1146  return false;
1147  }
1148 
1149  CriticalEdgePred.push_back(Pred);
1150  } else {
1151  // Only add the predecessors that will not be split for now.
1152  PredLoads[Pred] = nullptr;
1153  }
1154  }
1155 
1156  // Decide whether PRE is profitable for this load.
1157  unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1158  assert(NumUnavailablePreds != 0 &&
1159  "Fully available value should already be eliminated!");
1160 
1161  // If this load is unavailable in multiple predecessors, reject it.
1162  // FIXME: If we could restructure the CFG, we could make a common pred with
1163  // all the preds that don't have an available LI and insert a new load into
1164  // that one block.
1165  if (NumUnavailablePreds != 1)
1166  return false;
1167 
1168  // Split critical edges, and update the unavailable predecessors accordingly.
1169  for (BasicBlock *OrigPred : CriticalEdgePred) {
1170  BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1171  assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1172  PredLoads[NewPred] = nullptr;
1173  LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1174  << LoadBB->getName() << '\n');
1175  }
1176 
1177  // Check if the load can safely be moved to all the unavailable predecessors.
1178  bool CanDoPRE = true;
1179  const DataLayout &DL = LI->getModule()->getDataLayout();
1181  for (auto &PredLoad : PredLoads) {
1182  BasicBlock *UnavailablePred = PredLoad.first;
1183 
1184  // Do PHI translation to get its value in the predecessor if necessary. The
1185  // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1186 
1187  // If all preds have a single successor, then we know it is safe to insert
1188  // the load on the pred (?!?), so we can insert code to materialize the
1189  // pointer if it is not available.
1190  PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1191  Value *LoadPtr = nullptr;
1192  LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1193  *DT, NewInsts);
1194 
1195  // If we couldn't find or insert a computation of this phi translated value,
1196  // we fail PRE.
1197  if (!LoadPtr) {
1198  LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1199  << *LI->getPointerOperand() << "\n");
1200  CanDoPRE = false;
1201  break;
1202  }
1203 
1204  PredLoad.second = LoadPtr;
1205  }
1206 
1207  if (!CanDoPRE) {
1208  while (!NewInsts.empty()) {
1209  Instruction *I = NewInsts.pop_back_val();
1210  markInstructionForDeletion(I);
1211  }
1212  // HINT: Don't revert the edge-splitting as following transformation may
1213  // also need to split these critical edges.
1214  return !CriticalEdgePred.empty();
1215  }
1216 
1217  // Okay, we can eliminate this load by inserting a reload in the predecessor
1218  // and using PHI construction to get the value in the other predecessors, do
1219  // it.
1220  LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1221  LLVM_DEBUG(if (!NewInsts.empty()) dbgs()
1222  << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back()
1223  << '\n');
1224 
1225  // Assign value numbers to the new instructions.
1226  for (Instruction *I : NewInsts) {
1227  // Instructions that have been inserted in predecessor(s) to materialize
1228  // the load address do not retain their original debug locations. Doing
1229  // so could lead to confusing (but correct) source attributions.
1230  // FIXME: How do we retain source locations without causing poor debugging
1231  // behavior?
1232  I->setDebugLoc(DebugLoc());
1233 
1234  // FIXME: We really _ought_ to insert these value numbers into their
1235  // parent's availability map. However, in doing so, we risk getting into
1236  // ordering issues. If a block hasn't been processed yet, we would be
1237  // marking a value as AVAIL-IN, which isn't what we intend.
1238  VN.lookupOrAdd(I);
1239  }
1240 
1241  for (const auto &PredLoad : PredLoads) {
1242  BasicBlock *UnavailablePred = PredLoad.first;
1243  Value *LoadPtr = PredLoad.second;
1244 
1245  auto *NewLoad =
1246  new LoadInst(LI->getType(), LoadPtr, LI->getName() + ".pre",
1247  LI->isVolatile(), LI->getAlignment(), LI->getOrdering(),
1248  LI->getSyncScopeID(), UnavailablePred->getTerminator());
1249  NewLoad->setDebugLoc(LI->getDebugLoc());
1250 
1251  // Transfer the old load's AA tags to the new load.
1252  AAMDNodes Tags;
1253  LI->getAAMetadata(Tags);
1254  if (Tags)
1255  NewLoad->setAAMetadata(Tags);
1256 
1257  if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1258  NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1259  if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1260  NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1261  if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1262  NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1263 
1264  // We do not propagate the old load's debug location, because the new
1265  // load now lives in a different BB, and we want to avoid a jumpy line
1266  // table.
1267  // FIXME: How do we retain source locations without causing poor debugging
1268  // behavior?
1269 
1270  // Add the newly created load.
1271  ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1272  NewLoad));
1273  MD->invalidateCachedPointerInfo(LoadPtr);
1274  LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1275  }
1276 
1277  // Perform PHI construction.
1278  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1279  LI->replaceAllUsesWith(V);
1280  if (isa<PHINode>(V))
1281  V->takeName(LI);
1282  if (Instruction *I = dyn_cast<Instruction>(V))
1283  I->setDebugLoc(LI->getDebugLoc());
1284  if (V->getType()->isPtrOrPtrVectorTy())
1285  MD->invalidateCachedPointerInfo(V);
1286  markInstructionForDeletion(LI);
1287  ORE->emit([&]() {
1288  return OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1289  << "load eliminated by PRE";
1290  });
1291  ++NumPRELoad;
1292  return true;
1293 }
1294 
1297  using namespace ore;
1298 
1299  ORE->emit([&]() {
1300  return OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1301  << "load of type " << NV("Type", LI->getType()) << " eliminated"
1302  << setExtraArgs() << " in favor of "
1303  << NV("InfavorOfValue", AvailableValue);
1304  });
1305 }
1306 
1307 /// Attempt to eliminate a load whose dependencies are
1308 /// non-local by performing PHI construction.
1309 bool GVN::processNonLocalLoad(LoadInst *LI) {
1310  // non-local speculations are not allowed under asan.
1311  if (LI->getParent()->getParent()->hasFnAttribute(
1312  Attribute::SanitizeAddress) ||
1314  Attribute::SanitizeHWAddress))
1315  return false;
1316 
1317  // Step 1: Find the non-local dependencies of the load.
1318  LoadDepVect Deps;
1319  MD->getNonLocalPointerDependency(LI, Deps);
1320 
1321  // If we had to process more than one hundred blocks to find the
1322  // dependencies, this load isn't worth worrying about. Optimizing
1323  // it will be too expensive.
1324  unsigned NumDeps = Deps.size();
1325  if (NumDeps > MaxNumDeps)
1326  return false;
1327 
1328  // If we had a phi translation failure, we'll have a single entry which is a
1329  // clobber in the current block. Reject this early.
1330  if (NumDeps == 1 &&
1331  !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1332  LLVM_DEBUG(dbgs() << "GVN: non-local load "; LI->printAsOperand(dbgs());
1333  dbgs() << " has unknown dependencies\n";);
1334  return false;
1335  }
1336 
1337  // If this load follows a GEP, see if we can PRE the indices before analyzing.
1338  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1339  for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1340  OE = GEP->idx_end();
1341  OI != OE; ++OI)
1342  if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1343  performScalarPRE(I);
1344  }
1345 
1346  // Step 2: Analyze the availability of the load
1347  AvailValInBlkVect ValuesPerBlock;
1348  UnavailBlkVect UnavailableBlocks;
1349  AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1350 
1351  // If we have no predecessors that produce a known value for this load, exit
1352  // early.
1353  if (ValuesPerBlock.empty())
1354  return false;
1355 
1356  // Step 3: Eliminate fully redundancy.
1357  //
1358  // If all of the instructions we depend on produce a known value for this
1359  // load, then it is fully redundant and we can use PHI insertion to compute
1360  // its value. Insert PHIs and remove the fully redundant value now.
1361  if (UnavailableBlocks.empty()) {
1362  LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1363 
1364  // Perform PHI construction.
1365  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1366  LI->replaceAllUsesWith(V);
1367 
1368  if (isa<PHINode>(V))
1369  V->takeName(LI);
1370  if (Instruction *I = dyn_cast<Instruction>(V))
1371  // If instruction I has debug info, then we should not update it.
1372  // Also, if I has a null DebugLoc, then it is still potentially incorrect
1373  // to propagate LI's DebugLoc because LI may not post-dominate I.
1374  if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1375  I->setDebugLoc(LI->getDebugLoc());
1376  if (V->getType()->isPtrOrPtrVectorTy())
1377  MD->invalidateCachedPointerInfo(V);
1378  markInstructionForDeletion(LI);
1379  ++NumGVNLoad;
1380  reportLoadElim(LI, V, ORE);
1381  return true;
1382  }
1383 
1384  // Step 4: Eliminate partial redundancy.
1385  if (!EnablePRE || !EnableLoadPRE)
1386  return false;
1387 
1388  return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1389 }
1390 
1391 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1392  assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1393  "This function can only be called with llvm.assume intrinsic");
1394  Value *V = IntrinsicI->getArgOperand(0);
1395 
1396  if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1397  if (Cond->isZero()) {
1398  Type *Int8Ty = Type::getInt8Ty(V->getContext());
1399  // Insert a new store to null instruction before the load to indicate that
1400  // this code is not reachable. FIXME: We could insert unreachable
1401  // instruction directly because we can modify the CFG.
1402  new StoreInst(UndefValue::get(Int8Ty),
1404  IntrinsicI);
1405  }
1406  markInstructionForDeletion(IntrinsicI);
1407  return false;
1408  } else if (isa<Constant>(V)) {
1409  // If it's not false, and constant, it must evaluate to true. This means our
1410  // assume is assume(true), and thus, pointless, and we don't want to do
1411  // anything more here.
1412  return false;
1413  }
1414 
1415  Constant *True = ConstantInt::getTrue(V->getContext());
1416  bool Changed = false;
1417 
1418  for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1419  BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1420 
1421  // This property is only true in dominated successors, propagateEquality
1422  // will check dominance for us.
1423  Changed |= propagateEquality(V, True, Edge, false);
1424  }
1425 
1426  // We can replace assume value with true, which covers cases like this:
1427  // call void @llvm.assume(i1 %cmp)
1428  // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1429  ReplaceWithConstMap[V] = True;
1430 
1431  // If one of *cmp *eq operand is const, adding it to map will cover this:
1432  // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1433  // call void @llvm.assume(i1 %cmp)
1434  // ret float %0 ; will change it to ret float 3.000000e+00
1435  if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1436  if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1437  CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1438  (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1439  CmpI->getFastMathFlags().noNaNs())) {
1440  Value *CmpLHS = CmpI->getOperand(0);
1441  Value *CmpRHS = CmpI->getOperand(1);
1442  if (isa<Constant>(CmpLHS))
1443  std::swap(CmpLHS, CmpRHS);
1444  auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1445 
1446  // If only one operand is constant.
1447  if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1448  ReplaceWithConstMap[CmpLHS] = RHSConst;
1449  }
1450  }
1451  return Changed;
1452 }
1453 
1455  patchReplacementInstruction(I, Repl);
1456  I->replaceAllUsesWith(Repl);
1457 }
1458 
1459 /// Attempt to eliminate a load, first by eliminating it
1460 /// locally, and then attempting non-local elimination if that fails.
1461 bool GVN::processLoad(LoadInst *L) {
1462  if (!MD)
1463  return false;
1464 
1465  // This code hasn't been audited for ordered or volatile memory access
1466  if (!L->isUnordered())
1467  return false;
1468 
1469  if (L->use_empty()) {
1470  markInstructionForDeletion(L);
1471  return true;
1472  }
1473 
1474  // ... to a pointer that has been loaded from before...
1475  MemDepResult Dep = MD->getDependency(L);
1476 
1477  // If it is defined in another block, try harder.
1478  if (Dep.isNonLocal())
1479  return processNonLocalLoad(L);
1480 
1481  // Only handle the local case below
1482  if (!Dep.isDef() && !Dep.isClobber()) {
1483  // This might be a NonFuncLocal or an Unknown
1484  LLVM_DEBUG(
1485  // fast print dep, using operator<< on instruction is too slow.
1486  dbgs() << "GVN: load "; L->printAsOperand(dbgs());
1487  dbgs() << " has unknown dependence\n";);
1488  return false;
1489  }
1490 
1491  AvailableValue AV;
1492  if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1493  Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1494 
1495  // Replace the load!
1496  patchAndReplaceAllUsesWith(L, AvailableValue);
1497  markInstructionForDeletion(L);
1498  ++NumGVNLoad;
1499  reportLoadElim(L, AvailableValue, ORE);
1500  // Tell MDA to rexamine the reused pointer since we might have more
1501  // information after forwarding it.
1502  if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1503  MD->invalidateCachedPointerInfo(AvailableValue);
1504  return true;
1505  }
1506 
1507  return false;
1508 }
1509 
1510 /// Return a pair the first field showing the value number of \p Exp and the
1511 /// second field showing whether it is a value number newly created.
1512 std::pair<uint32_t, bool>
1513 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1514  uint32_t &e = expressionNumbering[Exp];
1515  bool CreateNewValNum = !e;
1516  if (CreateNewValNum) {
1517  Expressions.push_back(Exp);
1518  if (ExprIdx.size() < nextValueNumber + 1)
1519  ExprIdx.resize(nextValueNumber * 2);
1520  e = nextValueNumber;
1521  ExprIdx[nextValueNumber++] = nextExprNumber++;
1522  }
1523  return {e, CreateNewValNum};
1524 }
1525 
1526 /// Return whether all the values related with the same \p num are
1527 /// defined in \p BB.
1528 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1529  GVN &Gvn) {
1530  LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1531  while (Vals && Vals->BB == BB)
1532  Vals = Vals->Next;
1533  return !Vals;
1534 }
1535 
1536 /// Wrap phiTranslateImpl to provide caching functionality.
1538  const BasicBlock *PhiBlock, uint32_t Num,
1539  GVN &Gvn) {
1540  auto FindRes = PhiTranslateTable.find({Num, Pred});
1541  if (FindRes != PhiTranslateTable.end())
1542  return FindRes->second;
1543  uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1544  PhiTranslateTable.insert({{Num, Pred}, NewNum});
1545  return NewNum;
1546 }
1547 
1548 /// Translate value number \p Num using phis, so that it has the values of
1549 /// the phis in BB.
1550 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1551  const BasicBlock *PhiBlock,
1552  uint32_t Num, GVN &Gvn) {
1553  if (PHINode *PN = NumberingPhi[Num]) {
1554  for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1555  if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1556  if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1557  return TransVal;
1558  }
1559  return Num;
1560  }
1561 
1562  // If there is any value related with Num is defined in a BB other than
1563  // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1564  // a backedge. We can do an early exit in that case to save compile time.
1565  if (!areAllValsInBB(Num, PhiBlock, Gvn))
1566  return Num;
1567 
1568  if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1569  return Num;
1570  Expression Exp = Expressions[ExprIdx[Num]];
1571 
1572  for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1573  // For InsertValue and ExtractValue, some varargs are index numbers
1574  // instead of value numbers. Those index numbers should not be
1575  // translated.
1576  if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1577  (i > 0 && Exp.opcode == Instruction::ExtractValue))
1578  continue;
1579  Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1580  }
1581 
1582  if (Exp.commutative) {
1583  assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1584  if (Exp.varargs[0] > Exp.varargs[1]) {
1585  std::swap(Exp.varargs[0], Exp.varargs[1]);
1586  uint32_t Opcode = Exp.opcode >> 8;
1587  if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1588  Exp.opcode = (Opcode << 8) |
1590  static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1591  }
1592  }
1593 
1594  if (uint32_t NewNum = expressionNumbering[Exp])
1595  return NewNum;
1596  return Num;
1597 }
1598 
1599 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1600 /// again.
1602  const BasicBlock &CurrBlock) {
1603  for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1604  auto FindRes = PhiTranslateTable.find({Num, Pred});
1605  if (FindRes != PhiTranslateTable.end())
1606  PhiTranslateTable.erase(FindRes);
1607  }
1608 }
1609 
1610 // In order to find a leader for a given value number at a
1611 // specific basic block, we first obtain the list of all Values for that number,
1612 // and then scan the list to find one whose block dominates the block in
1613 // question. This is fast because dominator tree queries consist of only
1614 // a few comparisons of DFS numbers.
1615 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1616  LeaderTableEntry Vals = LeaderTable[num];
1617  if (!Vals.Val) return nullptr;
1618 
1619  Value *Val = nullptr;
1620  if (DT->dominates(Vals.BB, BB)) {
1621  Val = Vals.Val;
1622  if (isa<Constant>(Val)) return Val;
1623  }
1624 
1625  LeaderTableEntry* Next = Vals.Next;
1626  while (Next) {
1627  if (DT->dominates(Next->BB, BB)) {
1628  if (isa<Constant>(Next->Val)) return Next->Val;
1629  if (!Val) Val = Next->Val;
1630  }
1631 
1632  Next = Next->Next;
1633  }
1634 
1635  return Val;
1636 }
1637 
1638 /// There is an edge from 'Src' to 'Dst'. Return
1639 /// true if every path from the entry block to 'Dst' passes via this edge. In
1640 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1642  DominatorTree *DT) {
1643  // While in theory it is interesting to consider the case in which Dst has
1644  // more than one predecessor, because Dst might be part of a loop which is
1645  // only reachable from Src, in practice it is pointless since at the time
1646  // GVN runs all such loops have preheaders, which means that Dst will have
1647  // been changed to have only one predecessor, namely Src.
1648  const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1649  assert((!Pred || Pred == E.getStart()) &&
1650  "No edge between these basic blocks!");
1651  return Pred != nullptr;
1652 }
1653 
1654 void GVN::assignBlockRPONumber(Function &F) {
1655  BlockRPONumber.clear();
1656  uint32_t NextBlockNumber = 1;
1658  for (BasicBlock *BB : RPOT)
1659  BlockRPONumber[BB] = NextBlockNumber++;
1660  InvalidBlockRPONumbers = false;
1661 }
1662 
1663 // Tries to replace instruction with const, using information from
1664 // ReplaceWithConstMap.
1665 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1666  bool Changed = false;
1667  for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1668  Value *Operand = Instr->getOperand(OpNum);
1669  auto it = ReplaceWithConstMap.find(Operand);
1670  if (it != ReplaceWithConstMap.end()) {
1671  assert(!isa<Constant>(Operand) &&
1672  "Replacing constants with constants is invalid");
1673  LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "
1674  << *it->second << " in instruction " << *Instr << '\n');
1675  Instr->setOperand(OpNum, it->second);
1676  Changed = true;
1677  }
1678  }
1679  return Changed;
1680 }
1681 
1682 /// The given values are known to be equal in every block
1683 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1684 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1685 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1686 /// value starting from the end of Root.Start.
1687 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1688  bool DominatesByEdge) {
1690  Worklist.push_back(std::make_pair(LHS, RHS));
1691  bool Changed = false;
1692  // For speed, compute a conservative fast approximation to
1693  // DT->dominates(Root, Root.getEnd());
1694  const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1695 
1696  while (!Worklist.empty()) {
1697  std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1698  LHS = Item.first; RHS = Item.second;
1699 
1700  if (LHS == RHS)
1701  continue;
1702  assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1703 
1704  // Don't try to propagate equalities between constants.
1705  if (isa<Constant>(LHS) && isa<Constant>(RHS))
1706  continue;
1707 
1708  // Prefer a constant on the right-hand side, or an Argument if no constants.
1709  if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1710  std::swap(LHS, RHS);
1711  assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1712 
1713  // If there is no obvious reason to prefer the left-hand side over the
1714  // right-hand side, ensure the longest lived term is on the right-hand side,
1715  // so the shortest lived term will be replaced by the longest lived.
1716  // This tends to expose more simplifications.
1717  uint32_t LVN = VN.lookupOrAdd(LHS);
1718  if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1719  (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1720  // Move the 'oldest' value to the right-hand side, using the value number
1721  // as a proxy for age.
1722  uint32_t RVN = VN.lookupOrAdd(RHS);
1723  if (LVN < RVN) {
1724  std::swap(LHS, RHS);
1725  LVN = RVN;
1726  }
1727  }
1728 
1729  // If value numbering later sees that an instruction in the scope is equal
1730  // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1731  // the invariant that instructions only occur in the leader table for their
1732  // own value number (this is used by removeFromLeaderTable), do not do this
1733  // if RHS is an instruction (if an instruction in the scope is morphed into
1734  // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1735  // using the leader table is about compiling faster, not optimizing better).
1736  // The leader table only tracks basic blocks, not edges. Only add to if we
1737  // have the simple case where the edge dominates the end.
1738  if (RootDominatesEnd && !isa<Instruction>(RHS))
1739  addToLeaderTable(LVN, RHS, Root.getEnd());
1740 
1741  // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1742  // LHS always has at least one use that is not dominated by Root, this will
1743  // never do anything if LHS has only one use.
1744  if (!LHS->hasOneUse()) {
1745  unsigned NumReplacements =
1746  DominatesByEdge
1747  ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1748  : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1749 
1750  Changed |= NumReplacements > 0;
1751  NumGVNEqProp += NumReplacements;
1752  // Cached information for anything that uses LHS will be invalid.
1753  if (MD)
1754  MD->invalidateCachedPointerInfo(LHS);
1755  }
1756 
1757  // Now try to deduce additional equalities from this one. For example, if
1758  // the known equality was "(A != B)" == "false" then it follows that A and B
1759  // are equal in the scope. Only boolean equalities with an explicit true or
1760  // false RHS are currently supported.
1761  if (!RHS->getType()->isIntegerTy(1))
1762  // Not a boolean equality - bail out.
1763  continue;
1764  ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1765  if (!CI)
1766  // RHS neither 'true' nor 'false' - bail out.
1767  continue;
1768  // Whether RHS equals 'true'. Otherwise it equals 'false'.
1769  bool isKnownTrue = CI->isMinusOne();
1770  bool isKnownFalse = !isKnownTrue;
1771 
1772  // If "A && B" is known true then both A and B are known true. If "A || B"
1773  // is known false then both A and B are known false.
1774  Value *A, *B;
1775  if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1776  (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1777  Worklist.push_back(std::make_pair(A, RHS));
1778  Worklist.push_back(std::make_pair(B, RHS));
1779  continue;
1780  }
1781 
1782  // If we are propagating an equality like "(A == B)" == "true" then also
1783  // propagate the equality A == B. When propagating a comparison such as
1784  // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1785  if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1786  Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1787 
1788  // If "A == B" is known true, or "A != B" is known false, then replace
1789  // A with B everywhere in the scope.
1790  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1791  (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1792  Worklist.push_back(std::make_pair(Op0, Op1));
1793 
1794  // Handle the floating point versions of equality comparisons too.
1795  if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1796  (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1797 
1798  // Floating point -0.0 and 0.0 compare equal, so we can only
1799  // propagate values if we know that we have a constant and that
1800  // its value is non-zero.
1801 
1802  // FIXME: We should do this optimization if 'no signed zeros' is
1803  // applicable via an instruction-level fast-math-flag or some other
1804  // indicator that relaxed FP semantics are being used.
1805 
1806  if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1807  Worklist.push_back(std::make_pair(Op0, Op1));
1808  }
1809 
1810  // If "A >= B" is known true, replace "A < B" with false everywhere.
1811  CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1812  Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1813  // Since we don't have the instruction "A < B" immediately to hand, work
1814  // out the value number that it would have and use that to find an
1815  // appropriate instruction (if any).
1816  uint32_t NextNum = VN.getNextUnusedValueNumber();
1817  uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1818  // If the number we were assigned was brand new then there is no point in
1819  // looking for an instruction realizing it: there cannot be one!
1820  if (Num < NextNum) {
1821  Value *NotCmp = findLeader(Root.getEnd(), Num);
1822  if (NotCmp && isa<Instruction>(NotCmp)) {
1823  unsigned NumReplacements =
1824  DominatesByEdge
1825  ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1826  : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1827  Root.getStart());
1828  Changed |= NumReplacements > 0;
1829  NumGVNEqProp += NumReplacements;
1830  // Cached information for anything that uses NotCmp will be invalid.
1831  if (MD)
1832  MD->invalidateCachedPointerInfo(NotCmp);
1833  }
1834  }
1835  // Ensure that any instruction in scope that gets the "A < B" value number
1836  // is replaced with false.
1837  // The leader table only tracks basic blocks, not edges. Only add to if we
1838  // have the simple case where the edge dominates the end.
1839  if (RootDominatesEnd)
1840  addToLeaderTable(Num, NotVal, Root.getEnd());
1841 
1842  continue;
1843  }
1844  }
1845 
1846  return Changed;
1847 }
1848 
1849 /// When calculating availability, handle an instruction
1850 /// by inserting it into the appropriate sets
1851 bool GVN::processInstruction(Instruction *I) {
1852  // Ignore dbg info intrinsics.
1853  if (isa<DbgInfoIntrinsic>(I))
1854  return false;
1855 
1856  // If the instruction can be easily simplified then do so now in preference
1857  // to value numbering it. Value numbering often exposes redundancies, for
1858  // example if it determines that %y is equal to %x then the instruction
1859  // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1860  const DataLayout &DL = I->getModule()->getDataLayout();
1861  if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1862  bool Changed = false;
1863  if (!I->use_empty()) {
1864  I->replaceAllUsesWith(V);
1865  Changed = true;
1866  }
1867  if (isInstructionTriviallyDead(I, TLI)) {
1868  markInstructionForDeletion(I);
1869  Changed = true;
1870  }
1871  if (Changed) {
1872  if (MD && V->getType()->isPtrOrPtrVectorTy())
1873  MD->invalidateCachedPointerInfo(V);
1874  ++NumGVNSimpl;
1875  return true;
1876  }
1877  }
1878 
1879  if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1880  if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1881  return processAssumeIntrinsic(IntrinsicI);
1882 
1883  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1884  if (processLoad(LI))
1885  return true;
1886 
1887  unsigned Num = VN.lookupOrAdd(LI);
1888  addToLeaderTable(Num, LI, LI->getParent());
1889  return false;
1890  }
1891 
1892  // For conditional branches, we can perform simple conditional propagation on
1893  // the condition value itself.
1894  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1895  if (!BI->isConditional())
1896  return false;
1897 
1898  if (isa<Constant>(BI->getCondition()))
1899  return processFoldableCondBr(BI);
1900 
1901  Value *BranchCond = BI->getCondition();
1902  BasicBlock *TrueSucc = BI->getSuccessor(0);
1903  BasicBlock *FalseSucc = BI->getSuccessor(1);
1904  // Avoid multiple edges early.
1905  if (TrueSucc == FalseSucc)
1906  return false;
1907 
1908  BasicBlock *Parent = BI->getParent();
1909  bool Changed = false;
1910 
1911  Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1912  BasicBlockEdge TrueE(Parent, TrueSucc);
1913  Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1914 
1915  Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1916  BasicBlockEdge FalseE(Parent, FalseSucc);
1917  Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1918 
1919  return Changed;
1920  }
1921 
1922  // For switches, propagate the case values into the case destinations.
1923  if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1924  Value *SwitchCond = SI->getCondition();
1925  BasicBlock *Parent = SI->getParent();
1926  bool Changed = false;
1927 
1928  // Remember how many outgoing edges there are to every successor.
1930  for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1931  ++SwitchEdges[SI->getSuccessor(i)];
1932 
1933  for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1934  i != e; ++i) {
1935  BasicBlock *Dst = i->getCaseSuccessor();
1936  // If there is only a single edge, propagate the case value into it.
1937  if (SwitchEdges.lookup(Dst) == 1) {
1938  BasicBlockEdge E(Parent, Dst);
1939  Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1940  }
1941  }
1942  return Changed;
1943  }
1944 
1945  // Instructions with void type don't return a value, so there's
1946  // no point in trying to find redundancies in them.
1947  if (I->getType()->isVoidTy())
1948  return false;
1949 
1950  uint32_t NextNum = VN.getNextUnusedValueNumber();
1951  unsigned Num = VN.lookupOrAdd(I);
1952 
1953  // Allocations are always uniquely numbered, so we can save time and memory
1954  // by fast failing them.
1955  if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) {
1956  addToLeaderTable(Num, I, I->getParent());
1957  return false;
1958  }
1959 
1960  // If the number we were assigned was a brand new VN, then we don't
1961  // need to do a lookup to see if the number already exists
1962  // somewhere in the domtree: it can't!
1963  if (Num >= NextNum) {
1964  addToLeaderTable(Num, I, I->getParent());
1965  return false;
1966  }
1967 
1968  // Perform fast-path value-number based elimination of values inherited from
1969  // dominators.
1970  Value *Repl = findLeader(I->getParent(), Num);
1971  if (!Repl) {
1972  // Failure, just remember this instance for future use.
1973  addToLeaderTable(Num, I, I->getParent());
1974  return false;
1975  } else if (Repl == I) {
1976  // If I was the result of a shortcut PRE, it might already be in the table
1977  // and the best replacement for itself. Nothing to do.
1978  return false;
1979  }
1980 
1981  // Remove it!
1982  patchAndReplaceAllUsesWith(I, Repl);
1983  if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1984  MD->invalidateCachedPointerInfo(Repl);
1985  markInstructionForDeletion(I);
1986  return true;
1987 }
1988 
1989 /// runOnFunction - This is the main transformation entry point for a function.
1990 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1991  const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1992  MemoryDependenceResults *RunMD, LoopInfo *LI,
1993  OptimizationRemarkEmitter *RunORE) {
1994  AC = &RunAC;
1995  DT = &RunDT;
1996  VN.setDomTree(DT);
1997  TLI = &RunTLI;
1998  VN.setAliasAnalysis(&RunAA);
1999  MD = RunMD;
2000  ImplicitControlFlowTracking ImplicitCFT(DT);
2001  ICF = &ImplicitCFT;
2002  VN.setMemDep(MD);
2003  ORE = RunORE;
2004  InvalidBlockRPONumbers = true;
2005 
2006  bool Changed = false;
2007  bool ShouldContinue = true;
2008 
2010  // Merge unconditional branches, allowing PRE to catch more
2011  // optimization opportunities.
2012  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2013  BasicBlock *BB = &*FI++;
2014 
2015  bool removedBlock = MergeBlockIntoPredecessor(BB, &DTU, LI, nullptr, MD);
2016  if (removedBlock)
2017  ++NumGVNBlocks;
2018 
2019  Changed |= removedBlock;
2020  }
2021 
2022  unsigned Iteration = 0;
2023  while (ShouldContinue) {
2024  LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2025  ShouldContinue = iterateOnFunction(F);
2026  Changed |= ShouldContinue;
2027  ++Iteration;
2028  }
2029 
2030  if (EnablePRE) {
2031  // Fabricate val-num for dead-code in order to suppress assertion in
2032  // performPRE().
2033  assignValNumForDeadCode();
2034  bool PREChanged = true;
2035  while (PREChanged) {
2036  PREChanged = performPRE(F);
2037  Changed |= PREChanged;
2038  }
2039  }
2040 
2041  // FIXME: Should perform GVN again after PRE does something. PRE can move
2042  // computations into blocks where they become fully redundant. Note that
2043  // we can't do this until PRE's critical edge splitting updates memdep.
2044  // Actually, when this happens, we should just fully integrate PRE into GVN.
2045 
2046  cleanupGlobalSets();
2047  // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2048  // iteration.
2049  DeadBlocks.clear();
2050 
2051  return Changed;
2052 }
2053 
2054 bool GVN::processBlock(BasicBlock *BB) {
2055  // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2056  // (and incrementing BI before processing an instruction).
2057  assert(InstrsToErase.empty() &&
2058  "We expect InstrsToErase to be empty across iterations");
2059  if (DeadBlocks.count(BB))
2060  return false;
2061 
2062  // Clearing map before every BB because it can be used only for single BB.
2063  ReplaceWithConstMap.clear();
2064  bool ChangedFunction = false;
2065 
2066  for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2067  BI != BE;) {
2068  if (!ReplaceWithConstMap.empty())
2069  ChangedFunction |= replaceOperandsWithConsts(&*BI);
2070  ChangedFunction |= processInstruction(&*BI);
2071 
2072  if (InstrsToErase.empty()) {
2073  ++BI;
2074  continue;
2075  }
2076 
2077  // If we need some instructions deleted, do it now.
2078  NumGVNInstr += InstrsToErase.size();
2079 
2080  // Avoid iterator invalidation.
2081  bool AtStart = BI == BB->begin();
2082  if (!AtStart)
2083  --BI;
2084 
2085  for (auto *I : InstrsToErase) {
2086  assert(I->getParent() == BB && "Removing instruction from wrong block?");
2087  LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n');
2088  salvageDebugInfo(*I);
2089  if (MD) MD->removeInstruction(I);
2090  LLVM_DEBUG(verifyRemoved(I));
2091  ICF->removeInstruction(I);
2092  I->eraseFromParent();
2093  }
2094  InstrsToErase.clear();
2095 
2096  if (AtStart)
2097  BI = BB->begin();
2098  else
2099  ++BI;
2100  }
2101 
2102  return ChangedFunction;
2103 }
2104 
2105 // Instantiate an expression in a predecessor that lacked it.
2106 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2107  BasicBlock *Curr, unsigned int ValNo) {
2108  // Because we are going top-down through the block, all value numbers
2109  // will be available in the predecessor by the time we need them. Any
2110  // that weren't originally present will have been instantiated earlier
2111  // in this loop.
2112  bool success = true;
2113  for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2114  Value *Op = Instr->getOperand(i);
2115  if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2116  continue;
2117  // This could be a newly inserted instruction, in which case, we won't
2118  // find a value number, and should give up before we hurt ourselves.
2119  // FIXME: Rewrite the infrastructure to let it easier to value number
2120  // and process newly inserted instructions.
2121  if (!VN.exists(Op)) {
2122  success = false;
2123  break;
2124  }
2125  uint32_t TValNo =
2126  VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2127  if (Value *V = findLeader(Pred, TValNo)) {
2128  Instr->setOperand(i, V);
2129  } else {
2130  success = false;
2131  break;
2132  }
2133  }
2134 
2135  // Fail out if we encounter an operand that is not available in
2136  // the PRE predecessor. This is typically because of loads which
2137  // are not value numbered precisely.
2138  if (!success)
2139  return false;
2140 
2141  Instr->insertBefore(Pred->getTerminator());
2142  Instr->setName(Instr->getName() + ".pre");
2143  Instr->setDebugLoc(Instr->getDebugLoc());
2144 
2145  unsigned Num = VN.lookupOrAdd(Instr);
2146  VN.add(Instr, Num);
2147 
2148  // Update the availability map to include the new instruction.
2149  addToLeaderTable(Num, Instr, Pred);
2150  return true;
2151 }
2152 
2153 bool GVN::performScalarPRE(Instruction *CurInst) {
2154  if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() ||
2155  isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2156  CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2157  isa<DbgInfoIntrinsic>(CurInst))
2158  return false;
2159 
2160  // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2161  // sinking the compare again, and it would force the code generator to
2162  // move the i1 from processor flags or predicate registers into a general
2163  // purpose register.
2164  if (isa<CmpInst>(CurInst))
2165  return false;
2166 
2167  // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from
2168  // sinking the addressing mode computation back to its uses. Extending the
2169  // GEP's live range increases the register pressure, and therefore it can
2170  // introduce unnecessary spills.
2171  //
2172  // This doesn't prevent Load PRE. PHI translation will make the GEP available
2173  // to the load by moving it to the predecessor block if necessary.
2174  if (isa<GetElementPtrInst>(CurInst))
2175  return false;
2176 
2177  // We don't currently value number ANY inline asm calls.
2178  if (auto *CallB = dyn_cast<CallBase>(CurInst))
2179  if (CallB->isInlineAsm())
2180  return false;
2181 
2182  uint32_t ValNo = VN.lookup(CurInst);
2183 
2184  // Look for the predecessors for PRE opportunities. We're
2185  // only trying to solve the basic diamond case, where
2186  // a value is computed in the successor and one predecessor,
2187  // but not the other. We also explicitly disallow cases
2188  // where the successor is its own predecessor, because they're
2189  // more complicated to get right.
2190  unsigned NumWith = 0;
2191  unsigned NumWithout = 0;
2192  BasicBlock *PREPred = nullptr;
2193  BasicBlock *CurrentBlock = CurInst->getParent();
2194 
2195  // Update the RPO numbers for this function.
2196  if (InvalidBlockRPONumbers)
2197  assignBlockRPONumber(*CurrentBlock->getParent());
2198 
2200  for (BasicBlock *P : predecessors(CurrentBlock)) {
2201  // We're not interested in PRE where blocks with predecessors that are
2202  // not reachable.
2203  if (!DT->isReachableFromEntry(P)) {
2204  NumWithout = 2;
2205  break;
2206  }
2207  // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2208  // when CurInst has operand defined in CurrentBlock (so it may be defined
2209  // by phi in the loop header).
2210  assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) &&
2211  "Invalid BlockRPONumber map.");
2212  if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2213  llvm::any_of(CurInst->operands(), [&](const Use &U) {
2214  if (auto *Inst = dyn_cast<Instruction>(U.get()))
2215  return Inst->getParent() == CurrentBlock;
2216  return false;
2217  })) {
2218  NumWithout = 2;
2219  break;
2220  }
2221 
2222  uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2223  Value *predV = findLeader(P, TValNo);
2224  if (!predV) {
2225  predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2226  PREPred = P;
2227  ++NumWithout;
2228  } else if (predV == CurInst) {
2229  /* CurInst dominates this predecessor. */
2230  NumWithout = 2;
2231  break;
2232  } else {
2233  predMap.push_back(std::make_pair(predV, P));
2234  ++NumWith;
2235  }
2236  }
2237 
2238  // Don't do PRE when it might increase code size, i.e. when
2239  // we would need to insert instructions in more than one pred.
2240  if (NumWithout > 1 || NumWith == 0)
2241  return false;
2242 
2243  // We may have a case where all predecessors have the instruction,
2244  // and we just need to insert a phi node. Otherwise, perform
2245  // insertion.
2246  Instruction *PREInstr = nullptr;
2247 
2248  if (NumWithout != 0) {
2249  if (!isSafeToSpeculativelyExecute(CurInst)) {
2250  // It is only valid to insert a new instruction if the current instruction
2251  // is always executed. An instruction with implicit control flow could
2252  // prevent us from doing it. If we cannot speculate the execution, then
2253  // PRE should be prohibited.
2254  if (ICF->isDominatedByICFIFromSameBlock(CurInst))
2255  return false;
2256  }
2257 
2258  // Don't do PRE across indirect branch.
2259  if (isa<IndirectBrInst>(PREPred->getTerminator()))
2260  return false;
2261 
2262  // Don't do PRE across callbr.
2263  // FIXME: Can we do this across the fallthrough edge?
2264  if (isa<CallBrInst>(PREPred->getTerminator()))
2265  return false;
2266 
2267  // We can't do PRE safely on a critical edge, so instead we schedule
2268  // the edge to be split and perform the PRE the next time we iterate
2269  // on the function.
2270  unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2271  if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2272  toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2273  return false;
2274  }
2275  // We need to insert somewhere, so let's give it a shot
2276  PREInstr = CurInst->clone();
2277  if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2278  // If we failed insertion, make sure we remove the instruction.
2279  LLVM_DEBUG(verifyRemoved(PREInstr));
2280  PREInstr->deleteValue();
2281  return false;
2282  }
2283  }
2284 
2285  // Either we should have filled in the PRE instruction, or we should
2286  // not have needed insertions.
2287  assert(PREInstr != nullptr || NumWithout == 0);
2288 
2289  ++NumGVNPRE;
2290 
2291  // Create a PHI to make the value available in this block.
2292  PHINode *Phi =
2293  PHINode::Create(CurInst->getType(), predMap.size(),
2294  CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2295  for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2296  if (Value *V = predMap[i].first) {
2297  // If we use an existing value in this phi, we have to patch the original
2298  // value because the phi will be used to replace a later value.
2299  patchReplacementInstruction(CurInst, V);
2300  Phi->addIncoming(V, predMap[i].second);
2301  } else
2302  Phi->addIncoming(PREInstr, PREPred);
2303  }
2304 
2305  VN.add(Phi, ValNo);
2306  // After creating a new PHI for ValNo, the phi translate result for ValNo will
2307  // be changed, so erase the related stale entries in phi translate cache.
2308  VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2309  addToLeaderTable(ValNo, Phi, CurrentBlock);
2310  Phi->setDebugLoc(CurInst->getDebugLoc());
2311  CurInst->replaceAllUsesWith(Phi);
2312  if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2313  MD->invalidateCachedPointerInfo(Phi);
2314  VN.erase(CurInst);
2315  removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2316 
2317  LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2318  if (MD)
2319  MD->removeInstruction(CurInst);
2320  LLVM_DEBUG(verifyRemoved(CurInst));
2321  // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes
2322  // some assertion failures.
2323  ICF->removeInstruction(CurInst);
2324  CurInst->eraseFromParent();
2325  ++NumGVNInstr;
2326 
2327  return true;
2328 }
2329 
2330 /// Perform a purely local form of PRE that looks for diamond
2331 /// control flow patterns and attempts to perform simple PRE at the join point.
2332 bool GVN::performPRE(Function &F) {
2333  bool Changed = false;
2334  for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2335  // Nothing to PRE in the entry block.
2336  if (CurrentBlock == &F.getEntryBlock())
2337  continue;
2338 
2339  // Don't perform PRE on an EH pad.
2340  if (CurrentBlock->isEHPad())
2341  continue;
2342 
2343  for (BasicBlock::iterator BI = CurrentBlock->begin(),
2344  BE = CurrentBlock->end();
2345  BI != BE;) {
2346  Instruction *CurInst = &*BI++;
2347  Changed |= performScalarPRE(CurInst);
2348  }
2349  }
2350 
2351  if (splitCriticalEdges())
2352  Changed = true;
2353 
2354  return Changed;
2355 }
2356 
2357 /// Split the critical edge connecting the given two blocks, and return
2358 /// the block inserted to the critical edge.
2359 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2360  BasicBlock *BB =
2362  if (MD)
2363  MD->invalidateCachedPredecessors();
2364  InvalidBlockRPONumbers = true;
2365  return BB;
2366 }
2367 
2368 /// Split critical edges found during the previous
2369 /// iteration that may enable further optimization.
2370 bool GVN::splitCriticalEdges() {
2371  if (toSplit.empty())
2372  return false;
2373  do {
2374  std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val();
2375  SplitCriticalEdge(Edge.first, Edge.second,
2377  } while (!toSplit.empty());
2378  if (MD) MD->invalidateCachedPredecessors();
2379  InvalidBlockRPONumbers = true;
2380  return true;
2381 }
2382 
2383 /// Executes one iteration of GVN
2384 bool GVN::iterateOnFunction(Function &F) {
2385  cleanupGlobalSets();
2386 
2387  // Top-down walk of the dominator tree
2388  bool Changed = false;
2389  // Needed for value numbering with phi construction to work.
2390  // RPOT walks the graph in its constructor and will not be invalidated during
2391  // processBlock.
2393 
2394  for (BasicBlock *BB : RPOT)
2395  Changed |= processBlock(BB);
2396 
2397  return Changed;
2398 }
2399 
2400 void GVN::cleanupGlobalSets() {
2401  VN.clear();
2402  LeaderTable.clear();
2403  BlockRPONumber.clear();
2404  TableAllocator.Reset();
2405  ICF->clear();
2406  InvalidBlockRPONumbers = true;
2407 }
2408 
2409 /// Verify that the specified instruction does not occur in our
2410 /// internal data structures.
2411 void GVN::verifyRemoved(const Instruction *Inst) const {
2412  VN.verifyRemoved(Inst);
2413 
2414  // Walk through the value number scope to make sure the instruction isn't
2415  // ferreted away in it.
2417  I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2418  const LeaderTableEntry *Node = &I->second;
2419  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2420 
2421  while (Node->Next) {
2422  Node = Node->Next;
2423  assert(Node->Val != Inst && "Inst still in value numbering scope!");
2424  }
2425  }
2426 }
2427 
2428 /// BB is declared dead, which implied other blocks become dead as well. This
2429 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2430 /// live successors, update their phi nodes by replacing the operands
2431 /// corresponding to dead blocks with UndefVal.
2432 void GVN::addDeadBlock(BasicBlock *BB) {
2435 
2436  NewDead.push_back(BB);
2437  while (!NewDead.empty()) {
2438  BasicBlock *D = NewDead.pop_back_val();
2439  if (DeadBlocks.count(D))
2440  continue;
2441 
2442  // All blocks dominated by D are dead.
2444  DT->getDescendants(D, Dom);
2445  DeadBlocks.insert(Dom.begin(), Dom.end());
2446 
2447  // Figure out the dominance-frontier(D).
2448  for (BasicBlock *B : Dom) {
2449  for (BasicBlock *S : successors(B)) {
2450  if (DeadBlocks.count(S))
2451  continue;
2452 
2453  bool AllPredDead = true;
2454  for (BasicBlock *P : predecessors(S))
2455  if (!DeadBlocks.count(P)) {
2456  AllPredDead = false;
2457  break;
2458  }
2459 
2460  if (!AllPredDead) {
2461  // S could be proved dead later on. That is why we don't update phi
2462  // operands at this moment.
2463  DF.insert(S);
2464  } else {
2465  // While S is not dominated by D, it is dead by now. This could take
2466  // place if S already have a dead predecessor before D is declared
2467  // dead.
2468  NewDead.push_back(S);
2469  }
2470  }
2471  }
2472  }
2473 
2474  // For the dead blocks' live successors, update their phi nodes by replacing
2475  // the operands corresponding to dead blocks with UndefVal.
2477  I != E; I++) {
2478  BasicBlock *B = *I;
2479  if (DeadBlocks.count(B))
2480  continue;
2481 
2483  for (BasicBlock *P : Preds) {
2484  if (!DeadBlocks.count(P))
2485  continue;
2486 
2487  if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2488  if (BasicBlock *S = splitCriticalEdges(P, B))
2489  DeadBlocks.insert(P = S);
2490  }
2491 
2492  for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2493  PHINode &Phi = cast<PHINode>(*II);
2495  UndefValue::get(Phi.getType()));
2496  if (MD)
2497  MD->invalidateCachedPointerInfo(&Phi);
2498  }
2499  }
2500  }
2501 }
2502 
2503 // If the given branch is recognized as a foldable branch (i.e. conditional
2504 // branch with constant condition), it will perform following analyses and
2505 // transformation.
2506 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2507 // R be the target of the dead out-coming edge.
2508 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2509 // edge. The result of this step will be {X| X is dominated by R}
2510 // 2) Identify those blocks which haves at least one dead predecessor. The
2511 // result of this step will be dominance-frontier(R).
2512 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2513 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2514 //
2515 // Return true iff *NEW* dead code are found.
2516 bool GVN::processFoldableCondBr(BranchInst *BI) {
2517  if (!BI || BI->isUnconditional())
2518  return false;
2519 
2520  // If a branch has two identical successors, we cannot declare either dead.
2521  if (BI->getSuccessor(0) == BI->getSuccessor(1))
2522  return false;
2523 
2524  ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2525  if (!Cond)
2526  return false;
2527 
2528  BasicBlock *DeadRoot =
2529  Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2530  if (DeadBlocks.count(DeadRoot))
2531  return false;
2532 
2533  if (!DeadRoot->getSinglePredecessor())
2534  DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2535 
2536  addDeadBlock(DeadRoot);
2537  return true;
2538 }
2539 
2540 // performPRE() will trigger assert if it comes across an instruction without
2541 // associated val-num. As it normally has far more live instructions than dead
2542 // instructions, it makes more sense just to "fabricate" a val-number for the
2543 // dead code than checking if instruction involved is dead or not.
2544 void GVN::assignValNumForDeadCode() {
2545  for (BasicBlock *BB : DeadBlocks) {
2546  for (Instruction &Inst : *BB) {
2547  unsigned ValNum = VN.lookupOrAdd(&Inst);
2548  addToLeaderTable(ValNum, &Inst, BB);
2549  }
2550  }
2551 }
2552 
2554 public:
2555  static char ID; // Pass identification, replacement for typeid
2556 
2557  explicit GVNLegacyPass(bool NoMemDepAnalysis = !EnableMemDep)
2558  : FunctionPass(ID), NoMemDepAnalysis(NoMemDepAnalysis) {
2560  }
2561 
2562  bool runOnFunction(Function &F) override {
2563  if (skipFunction(F))
2564  return false;
2565 
2566  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2567 
2568  return Impl.runImpl(
2569  F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2570  getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2571  getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2572  getAnalysis<AAResultsWrapperPass>().getAAResults(),
2573  NoMemDepAnalysis ? nullptr
2574  : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2575  LIWP ? &LIWP->getLoopInfo() : nullptr,
2576  &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2577  }
2578 
2579  void getAnalysisUsage(AnalysisUsage &AU) const override {
2583  if (!NoMemDepAnalysis)
2586 
2591  }
2592 
2593 private:
2594  bool NoMemDepAnalysis;
2595  GVN Impl;
2596 };
2597 
2598 char GVNLegacyPass::ID = 0;
2599 
2600 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2608 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2609 
2610 // The public interface to this file...
2611 FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) {
2612  return new GVNLegacyPass(NoMemDepAnalysis);
2613 }
Legacy wrapper pass to provide the GlobalsAAResult object.
static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV)
Definition: GVN.cpp:243
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:748
uint64_t CallInst * C
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:67
FunctionPass * createGVNPass(bool NoLoads=false)
Create a legacy GVN pass.
Definition: GVN.cpp:2611
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:1601
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
bool isUndefValue() const
Definition: GVN.cpp:211
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:594
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:70
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:636
static bool runImpl(Function &F, TargetLibraryInfo &TLI, DominatorTree &DT)
This is the entry point for all transforms.
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
raw_ostream & errs()
This returns a reference to a raw_ostream for standard error.
Helper class for SSA formation on a set of values defined in multiple blocks.
Definition: SSAUpdater.h:38
Diagnostic information for missed-optimization remarks.
Provides a lazy, caching interface for making common memory aliasing information queries, backed by LLVM&#39;s alias analysis passes.
int analyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, LoadInst *DepLI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the load at De...
Definition: VNCoercion.cpp:242
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:184
size_type size() const
Definition: MapVector.h:60
unsigned Offset
Offset - The byte offset in Val that is interesting for the load query.
Definition: GVN.cpp:174
DiagnosticInfoOptimizationBase::Argument NV
static Type * makeCmpResultType(Type *opnd_type)
Create a result type for fcmp/icmp.
Definition: InstrTypes.h:889
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:769
This class represents lattice values for constants.
Definition: AllocatorList.h:23
PointerTy getPointer() const
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:464
bool isAtomic() const
Return true if this instruction has an AtomicOrdering of unordered or higher.
This is the interface for a simple mod/ref and alias analysis over globals.
void Initialize(Type *Ty, StringRef Name)
Reset this object to get ready for a new set of SSA updates with type &#39;Ty&#39;.
Definition: SSAUpdater.cpp:53
bool MergeBlockIntoPredecessor(BasicBlock *BB, DomTreeUpdater *DTU=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, MemoryDependenceResults *MemDep=nullptr)
Attempts to merge a block into its predecessor, if possible.
uint32_t lookupOrAddCmp(unsigned Opcode, CmpInst::Predicate Pred, Value *LHS, Value *RHS)
Returns the value number of the given comparison, assigning it a new number if it did not have one be...
Definition: GVN.cpp:587
iterator end()
Definition: Function.h:657
void AddAvailableValue(BasicBlock *BB, Value *V)
Indicate that a rewritten value is available in the specified block with the specified value...
Definition: SSAUpdater.cpp:71
bool operator==(const Expression &other) const
Definition: GVN.cpp:119
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:1603
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:247
uint32_t phiTranslate(const BasicBlock *BB, const BasicBlock *PhiBlock, uint32_t Num, GVN &Gvn)
Wrap phiTranslateImpl to provide caching functionality.
Definition: GVN.cpp:1537
bool isTerminator() const
Definition: Instruction.h:128
1 1 1 0 True if unordered or not equal
Definition: InstrTypes.h:662
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:98
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.h:320
unsigned second
This class implements a map that also provides access to all stored values in a deterministic order...
Definition: MapVector.h:37
BasicBlock * getSuccessor(unsigned i) const
bool properlyDominates(const DomTreeNodeBase< NodeT > *A, const DomTreeNodeBase< NodeT > *B) const
properlyDominates - Returns true iff A dominates B and A != B.
STATISTIC(NumFunctions, "Total number of functions")
A debug info location.
Definition: DebugLoc.h:33
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:230
F(f)
bool isCoercedLoadValue() const
Definition: GVN.cpp:209
An instruction for reading from memory.
Definition: Instructions.h:167
const BasicBlock * getEnd() const
Definition: Dominators.h:94
Hexagon Common GEP
Value * getCondition() const
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:137
This defines the Use class.
idx_iterator idx_end() const
unsigned replaceDominatedUsesWith(Value *From, Value *To, DominatorTree &DT, const BasicBlockEdge &Edge)
Replace each use of &#39;From&#39; with &#39;To&#39; if that use is dominated by the given edge.
Definition: Local.cpp:2437
Use * op_iterator
Definition: User.h:224
iterator end()
Get an iterator to the end of the SetVector.
Definition: SetVector.h:92
Value * getMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingMemInst returned an offset, this function can be used to actually perform...
Definition: VNCoercion.cpp:515
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:32
op_iterator op_begin()
Definition: User.h:229
gvn Early GVN Hoisting of Expressions
Definition: GVNHoist.cpp:1203
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:274
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:268
uint32_t lookup(Value *V, bool Verify=true) const
Returns the value number of the specified value.
Definition: GVN.cpp:574
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1155
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:221
void dump() const
Support for debugging, callable in GDB: V->dump()
Definition: AsmWriter.cpp:4347
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:47
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:50
bool isVolatile() const
Return true if this is a load from a volatile memory location.
Definition: Instructions.h:231
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: GVN.cpp:2579
static cl::opt< bool > EnablePRE("enable-pre", cl::init(true), cl::Hidden)
void patchReplacementInstruction(Instruction *I, Value *Repl)
Patch the replacement so that it is not more restrictive than the value being replaced.
Definition: Local.cpp:2368
bool isDef() const
Tests if this MemDepResult represents a query that is an instruction definition dependency.
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:369
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: GVN.cpp:2562
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:55
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:651
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:279
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:41
MemoryDependenceResults & getMemDep() const
Definition: GVN.h:84
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:196
This file contains the simple types necessary to represent the attributes associated with functions a...
An analysis that produces MemoryDependenceResults for a function.
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:285
Analysis pass that exposes the LoopInfo for a function.
Definition: LoopInfo.h:944
static const uint16_t * lookup(unsigned opcode, unsigned domain, ArrayRef< uint16_t[3]> Table)
bool isSimpleValue() const
Definition: GVN.cpp:208
Interval::succ_iterator succ_begin(Interval *I)
succ_begin/succ_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:102
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:244
ppc ctr loops PowerPC CTR Loops Verify
bool insert(const value_type &X)
Insert a new element into the SetVector.
Definition: SetVector.h:141
The core GVN pass object.
Definition: GVN.h:68
IntType getInt() const
bool canCoerceMustAliasedValueToLoad(Value *StoredVal, Type *LoadTy, const DataLayout &DL)
Return true if CoerceAvailableValueToLoadType would succeed if it was called.
Definition: VNCoercion.cpp:15
Expression(uint32_t o=~2U)
Definition: GVN.cpp:117
#define DEBUG_TYPE
Definition: GVN.cpp:87
iterator begin()
Get an iterator to the beginning of the SetVector.
Definition: SetVector.h:82
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:234
DiagnosticInfoOptimizationBase::setExtraArgs setExtraArgs
static AvailableValue getLoad(LoadInst *LI, unsigned Offset=0)
Definition: GVN.cpp:192
hash_code hash_value(const APFloat &Arg)
See friend declarations above.
Definition: APFloat.cpp:4430
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
BasicBlock * SplitCriticalEdge(Instruction *TI, unsigned SuccNum, const CriticalEdgeSplittingOptions &Options=CriticalEdgeSplittingOptions())
If this edge is a critical edge, insert a new node to split the critical edge.
LoadInst * getCoercedLoadValue() const
Definition: GVN.cpp:218
static GVN::Expression getEmptyKey()
Definition: GVN.cpp:141
An instruction for storing to memory.
Definition: Instructions.h:320
bool isMinusOne() const
This function will return true iff every bit in this constant is set to true.
Definition: Constants.h:208
void add(Value *V, uint32_t num)
add - Insert a value into the table with a specified value number.
Definition: GVN.cpp:388
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:429
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:291
iterator begin()
Definition: Function.h:655
static unsigned getHashValue(const GVN::Expression &e)
Definition: GVN.cpp:144
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:144
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Value * getOperand(unsigned i) const
Definition: User.h:169
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:105
int analyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, StoreInst *DepSI, const DataLayout &DL)
This function determines whether a value for the pointer LoadPtr can be extracted from the store at D...
Definition: VNCoercion.cpp:214
void initializeGVNLegacyPassPass(PassRegistry &)
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:140
const BasicBlock & getEntryBlock() const
Definition: Function.h:639
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:873
void getAAMetadata(AAMDNodes &N, bool Merge=false) const
Fills the AAMDNodes structure with AA metadata from this instruction.
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:423
Value * GetValueInMiddleOfBlock(BasicBlock *BB)
Construct SSA form, materializing a value that is live in the middle of the specified block...
Definition: SSAUpdater.cpp:99
SmallVector< uint32_t, 4 > varargs
Definition: GVN.cpp:115
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:148
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:321
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:233
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:73
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
PointerIntPair - This class implements a pair of a pointer and small integer.
PHITransAddr - An address value which tracks and handles phi translation.
Definition: PHITransAddr.h:35
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
Definition: PatternMatch.h:754
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:142
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:41
static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS)
Definition: GVN.cpp:150
This file contains the declarations for the subclasses of Constant, which represent the different fla...
static cl::opt< uint32_t > MaxRecurseDepth("gvn-max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore, cl::desc("Max recurse depth in GVN (default = 1000)"))
const Instruction & front() const
Definition: BasicBlock.h:280
A manager for alias analyses.
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:575
Diagnostic information for applied optimization remarks.
Interval::pred_iterator pred_begin(Interval *I)
pred_begin/pred_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:112
unsigned getNumIndices() const
bool isUnordered() const
Definition: Instructions.h:278
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:231
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1192
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:646
PointerIntPair< Value *, 2, ValType > Val
V - The value that is live out of the block.
Definition: GVN.cpp:171
MemIntrinsic * getMemIntrinValue() const
Definition: GVN.cpp:223
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:115
op_range operands()
Definition: User.h:237
Value * getPointerOperand()
Definition: Instructions.h:284
bool isCriticalEdge(const Instruction *TI, unsigned SuccNum, bool AllowIdenticalEdges=false)
Return true if the specified edge is a critical edge.
Definition: CFG.cpp:87
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:403
static void reportLoadElim(LoadInst *LI, Value *AvailableValue, OptimizationRemarkEmitter *ORE)
Definition: GVN.cpp:1295
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1424
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
size_t size() const
Definition: SmallVector.h:52
static cl::opt< bool > EnableMemDep("enable-gvn-memdep", cl::init(true))
A wrapper analysis pass for the legacy pass manager that exposes a MemoryDepnedenceResults instance...
void printAsOperand(raw_ostream &O, bool PrintType=true, const Module *M=nullptr) const
Print the name of this Value out to the specified raw_ostream.
Definition: AsmWriter.cpp:4274
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
A memory dependence query can return one of three different answers.
DominatorTree & getDominatorTree() const
Definition: GVN.h:82
unsigned first
static cl::opt< uint32_t > MaxNumDeps("gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore, cl::desc("Max number of dependences to attempt Load PRE (default = 100)"))
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:50
static void reportMayClobberedLoad(LoadInst *LI, MemDepResult DepInfo, DominatorTree *DT, OptimizationRemarkEmitter *ORE)
Try to locate the three instruction involved in a missed load-elimination case that is due to an inte...
Definition: GVN.cpp:842
A function analysis which provides an AssumptionCache.
bool isPtrOrPtrVectorTy() const
Return true if this is a pointer type or a vector of pointer types.
Definition: Type.h:226
Value * MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const
Emit code at the end of this block to adjust the value defined here to the specified type...
Definition: GVN.cpp:261
A SetVector that performs no allocations if smaller than a certain size.
Definition: SetVector.h:297
This is the common base class for memset/memcpy/memmove.
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:191
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:417
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
void emit(DiagnosticInfoOptimizationBase &OptDiag)
Output the remark via the diagnostic handler and to the optimization record file. ...
iterator end()
Definition: BasicBlock.h:270
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:248
Module.h This file contains the declarations for the Module class.
Provides information about what library functions are available for the current target.
const MemDepResult & getResult() const
size_type count(const KeyT &Key) const
Definition: MapVector.h:142
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:47
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:643
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:374
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:631
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:124
Value * MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt, GVN &gvn) const
Emit code at the specified insertion point to adjust the value defined here to the specified type...
Definition: GVN.cpp:785
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:587
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:491
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:940
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:55
hash_code hash_combine(const Ts &...args)
Combine values into a single hash_code.
Definition: Hashing.h:600
Represents an AvailableValue which can be rematerialized at the end of the associated BasicBlock...
Definition: GVN.cpp:236
iterator_range< user_iterator > users()
Definition: Value.h:399
hash_code hash_combine_range(InputIteratorT first, InputIteratorT last)
Compute a hash_code for a sequence of values.
Definition: Hashing.h:478
std::vector< NonLocalDepEntry > NonLocalDepInfo
An opaque object representing a hash code.
Definition: Hashing.h:71
bool isMallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates uninitialized memory (such ...
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:471
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:387
void erase(Value *v)
Remove a value from the value numbering.
Definition: GVN.cpp:607
static bool isLifetimeStart(const Instruction *Inst)
Definition: GVN.cpp:834
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:545
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:71
unsigned getNumArgOperands() const
Definition: InstrTypes.h:1153
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:324
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:240
Instruction * getInst() const
If this is a normal dependency, returns the instruction that is depended on.
void clear()
Definition: ilist.h:307
Value * getStoreValueForLoad(Value *SrcVal, unsigned Offset, Type *LoadTy, Instruction *InsertPt, const DataLayout &DL)
If analyzeLoadFromClobberingStore returned an offset, this function can be used to actually perform t...
Definition: VNCoercion.cpp:383
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
GVNLegacyPass(bool NoMemDepAnalysis=!EnableMemDep)
Definition: GVN.cpp:2557
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:214
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
bool isCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates zero-filled memory (such as...
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this load instruction.
Definition: Instructions.h:259
#define I(x, y, z)
Definition: MD5.cpp:58
bool mayReadFromMemory() const
Return true if this instruction may read memory.
static AvailableValue get(Value *V, unsigned Offset=0)
Definition: GVN.cpp:176
uint32_t opcode
Definition: GVN.cpp:112
PassT::Result * getCachedResult(IRUnitT &IR) const
Get the cached result of an analysis pass for a given IR unit.
Definition: PassManager.h:788
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:322
idx_iterator idx_begin() const
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
This class allows to keep track on instructions with implicit control flow.
bool isUnconditional() const
friend hash_code hash_value(const Expression &Value)
Definition: GVN.cpp:131
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:211
Value * getSimpleValue() const
Definition: GVN.cpp:213
Analysis pass providing the TargetLibraryInfo.
iterator_range< df_iterator< T > > depth_first(const T &G)
Multiway switch.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
const BasicBlock * getStart() const
Definition: Dominators.h:90
Represents a particular available value that we know how to materialize.
Definition: GVN.cpp:161
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:649
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:349
LLVM Value Representation.
Definition: Value.h:72
static AvailableValueInBlock getUndef(BasicBlock *BB)
Definition: GVN.cpp:255
void removeInstruction(Instruction *InstToRemove)
Removes an instruction from the dependence analysis, updating the dependence of instructions that pre...
succ_range successors(Instruction *I)
Definition: CFG.h:259
OptimizationRemarkEmitter legacy analysis pass.
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: GVN.cpp: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:761
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:238
static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl)
Definition: GVN.cpp:1454
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:259
bool isMemIntrinValue() const
Definition: GVN.cpp:210
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
This header defines various interfaces for pass management in LLVM.
void setIncomingValue(unsigned i, Value *V)
AvailableValue AV
AV - The actual available value.
Definition: GVN.cpp:241
#define LLVM_DEBUG(X)
Definition: Debug.h:122
Value * SimplifyInstruction(Instruction *I, const SimplifyQuery &Q, OptimizationRemarkEmitter *ORE=nullptr)
See if we can compute a simplified version of this instruction.
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:173
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:322
static AvailableValue getUndef()
Definition: GVN.cpp:200
static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, DominatorTree *DT)
There is an edge from &#39;Src&#39; to &#39;Dst&#39;.
Definition: GVN.cpp:1641
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:43
const BasicBlock * getParent() const
Definition: Instruction.h:66
This instruction inserts a struct field of array element value into an aggregate value.
bool HasValueForBlock(BasicBlock *BB) const
Return true if the SSAUpdater already has a value for the specified block.
Definition: SSAUpdater.cpp:62