llvm/utils/TableGen/DAGISelMatcherOpt.cpp

//===- DAGISelMatcherOpt.cpp - Optimize a DAG Matcher ---------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements the DAG Matcher optimizer.
//
//===----------------------------------------------------------------------===//

#include "DAGISelMatcher.h"
#include "CodeGenDAGPatterns.h"
#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/StringSet.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;

#define DEBUG_TYPE "isel-opt"

/// ContractNodes - Turn multiple matcher node patterns like 'MoveChild+Record'
/// into single compound nodes like RecordChild.
static void ContractNodes(std::unique_ptr<Matcher> &MatcherPtr,
                          const CodeGenDAGPatterns &CGP) {
  // If we reached the end of the chain, we're done.
  Matcher *N = MatcherPtr.get();
  if (!N) return;
  
  // If we have a scope node, walk down all of the children.
  if (ScopeMatcher *Scope = dyn_cast<ScopeMatcher>(N)) {
    for (unsigned i = 0, e = Scope->getNumChildren(); i != e; ++i) {
      std::unique_ptr<Matcher> Child(Scope->takeChild(i));
      ContractNodes(Child, CGP);
      Scope->resetChild(i, Child.release());
    }
    return;
  }
  
  // If we found a movechild node with a node that comes in a 'foochild' form,
  // transform it.
  if (MoveChildMatcher *MC = dyn_cast<MoveChildMatcher>(N)) {
    Matcher *New = nullptr;
    if (RecordMatcher *RM = dyn_cast<RecordMatcher>(MC->getNext()))
      if (MC->getChildNo() < 8)  // Only have RecordChild0...7
        New = new RecordChildMatcher(MC->getChildNo(), RM->getWhatFor(),
                                     RM->getResultNo());

    if (CheckTypeMatcher *CT = dyn_cast<CheckTypeMatcher>(MC->getNext()))
      if (MC->getChildNo() < 8 &&  // Only have CheckChildType0...7
          CT->getResNo() == 0)     // CheckChildType checks res #0
        New = new CheckChildTypeMatcher(MC->getChildNo(), CT->getType());

    if (CheckSameMatcher *CS = dyn_cast<CheckSameMatcher>(MC->getNext()))
      if (MC->getChildNo() < 4)  // Only have CheckChildSame0...3
        New = new CheckChildSameMatcher(MC->getChildNo(), CS->getMatchNumber());

    if (CheckIntegerMatcher *CS = dyn_cast<CheckIntegerMatcher>(MC->getNext()))
      if (MC->getChildNo() < 5)  // Only have CheckChildInteger0...4
        New = new CheckChildIntegerMatcher(MC->getChildNo(), CS->getValue());

    if (New) {
      // Insert the new node.
      New->setNext(MatcherPtr.release());
      MatcherPtr.reset(New);
      // Remove the old one.
      MC->setNext(MC->getNext()->takeNext());
      return ContractNodes(MatcherPtr, CGP);
    }
  }
  
  // Zap movechild -> moveparent.
  if (MoveChildMatcher *MC = dyn_cast<MoveChildMatcher>(N))
    if (MoveParentMatcher *MP = 
          dyn_cast<MoveParentMatcher>(MC->getNext())) {
      MatcherPtr.reset(MP->takeNext());
      return ContractNodes(MatcherPtr, CGP);
    }

  // Turn EmitNode->MarkFlagResults->CompleteMatch into
  // MarkFlagResults->EmitNode->CompleteMatch when we can to encourage
  // MorphNodeTo formation.  This is safe because MarkFlagResults never refers
  // to the root of the pattern.
  if (isa<EmitNodeMatcher>(N) && isa<MarkGlueResultsMatcher>(N->getNext()) &&
      isa<CompleteMatchMatcher>(N->getNext()->getNext())) {
    // Unlink the two nodes from the list.
    Matcher *EmitNode = MatcherPtr.release();
    Matcher *MFR = EmitNode->takeNext();
    Matcher *Tail = MFR->takeNext();
        
    // Relink them.
    MatcherPtr.reset(MFR);
    MFR->setNext(EmitNode);
    EmitNode->setNext(Tail);
    return ContractNodes(MatcherPtr, CGP);
  }

  // Turn EmitNode->CompleteMatch into MorphNodeTo if we can.
  if (EmitNodeMatcher *EN = dyn_cast<EmitNodeMatcher>(N))
    if (CompleteMatchMatcher *CM =
          dyn_cast<CompleteMatchMatcher>(EN->getNext())) {
      // We can only use MorphNodeTo if the result values match up.
      unsigned RootResultFirst = EN->getFirstResultSlot();
      bool ResultsMatch = true;
      for (unsigned i = 0, e = CM->getNumResults(); i != e; ++i)
        if (CM->getResult(i) != RootResultFirst+i)
          ResultsMatch = false;
      
      // If the selected node defines a subset of the glue/chain results, we
      // can't use MorphNodeTo.  For example, we can't use MorphNodeTo if the
      // matched pattern has a chain but the root node doesn't.
      const PatternToMatch &Pattern = CM->getPattern();
      
      if (!EN->hasChain() &&
          Pattern.getSrcPattern()->NodeHasProperty(SDNPHasChain, CGP))
        ResultsMatch = false;

      // If the matched node has glue and the output root doesn't, we can't
      // use MorphNodeTo.
      //
      // NOTE: Strictly speaking, we don't have to check for glue here
      // because the code in the pattern generator doesn't handle it right.  We
      // do it anyway for thoroughness.
      if (!EN->hasOutFlag() &&
          Pattern.getSrcPattern()->NodeHasProperty(SDNPOutGlue, CGP))
        ResultsMatch = false;
      
      
      // If the root result node defines more results than the source root node
      // *and* has a chain or glue input, then we can't match it because it
      // would end up replacing the extra result with the chain/glue.
#if 0
      if ((EN->hasGlue() || EN->hasChain()) &&
          EN->getNumNonChainGlueVTs() > ... need to get no results reliably ...)
        ResultMatch = false;
#endif
          
      if (ResultsMatch) {
        const SmallVectorImpl<MVT::SimpleValueType> &VTs = EN->getVTList();
        const SmallVectorImpl<unsigned> &Operands = EN->getOperandList();
        MatcherPtr.reset(new MorphNodeToMatcher(EN->getOpcodeName(),
                                                VTs, Operands,
                                                EN->hasChain(), EN->hasInFlag(),
                                                EN->hasOutFlag(),
                                                EN->hasMemRefs(),
                                                EN->getNumFixedArityOperands(),
                                                Pattern));
        return;
      }

      // FIXME2: Kill off all the SelectionDAG::SelectNodeTo and getMachineNode
      // variants.
    }
  
  ContractNodes(N->getNextPtr(), CGP);
  
  
  // If we have a CheckType/CheckChildType/Record node followed by a
  // CheckOpcode, invert the two nodes.  We prefer to do structural checks
  // before type checks, as this opens opportunities for factoring on targets
  // like X86 where many operations are valid on multiple types.
  if ((isa<CheckTypeMatcher>(N) || isa<CheckChildTypeMatcher>(N) ||
       isa<RecordMatcher>(N)) &&
      isa<CheckOpcodeMatcher>(N->getNext())) {
    // Unlink the two nodes from the list.
    Matcher *CheckType = MatcherPtr.release();
    Matcher *CheckOpcode = CheckType->takeNext();
    Matcher *Tail = CheckOpcode->takeNext();
    
    // Relink them.
    MatcherPtr.reset(CheckOpcode);
    CheckOpcode->setNext(CheckType);
    CheckType->setNext(Tail);
    return ContractNodes(MatcherPtr, CGP);
  }
}

/// SinkPatternPredicates - Pattern predicates can be checked at any level of
/// the matching tree.  The generator dumps them at the top level of the pattern
/// though, which prevents factoring from being able to see past them.  This
/// optimization sinks them as far down into the pattern as possible.
///
/// Conceptually, we'd like to sink these predicates all the way to the last
/// matcher predicate in the series.  However, it turns out that some
/// ComplexPatterns have side effects on the graph, so we really don't want to
/// run a the complex pattern if the pattern predicate will fail.  For this
/// reason, we refuse to sink the pattern predicate past a ComplexPattern.
///
static void SinkPatternPredicates(std::unique_ptr<Matcher> &MatcherPtr) {
  // Recursively scan for a PatternPredicate.
  // If we reached the end of the chain, we're done.
  Matcher *N = MatcherPtr.get();
  if (!N) return;
  
  // Walk down all members of a scope node.
  if (ScopeMatcher *Scope = dyn_cast<ScopeMatcher>(N)) {
    for (unsigned i = 0, e = Scope->getNumChildren(); i != e; ++i) {
      std::unique_ptr<Matcher> Child(Scope->takeChild(i));
      SinkPatternPredicates(Child);
      Scope->resetChild(i, Child.release());
    }
    return;
  }
  
  // If this node isn't a CheckPatternPredicateMatcher we keep scanning until
  // we find one.
  CheckPatternPredicateMatcher *CPPM =dyn_cast<CheckPatternPredicateMatcher>(N);
  if (!CPPM)
    return SinkPatternPredicates(N->getNextPtr());
  
  // Ok, we found one, lets try to sink it. Check if we can sink it past the
  // next node in the chain.  If not, we won't be able to change anything and
  // might as well bail.
  if (!CPPM->getNext()->isSafeToReorderWithPatternPredicate())
    return;
  
  // Okay, we know we can sink it past at least one node.  Unlink it from the
  // chain and scan for the new insertion point.
  MatcherPtr.release();  // Don't delete CPPM.
  MatcherPtr.reset(CPPM->takeNext());
  
  N = MatcherPtr.get();
  while (N->getNext()->isSafeToReorderWithPatternPredicate())
    N = N->getNext();
  
  // At this point, we want to insert CPPM after N.
  CPPM->setNext(N->takeNext());
  N->setNext(CPPM);
}

/// FindNodeWithKind - Scan a series of matchers looking for a matcher with a
/// specified kind.  Return null if we didn't find one otherwise return the
/// matcher.
static Matcher *FindNodeWithKind(Matcher *M, Matcher::KindTy Kind) {
  for (; M; M = M->getNext())
    if (M->getKind() == Kind)
      return M;
  return nullptr;
}


/// FactorNodes - Turn matches like this:
///   Scope
///     OPC_CheckType i32
///       ABC
///     OPC_CheckType i32
///       XYZ
/// into:
///   OPC_CheckType i32
///     Scope
///       ABC
///       XYZ
///
static void FactorNodes(std::unique_ptr<Matcher> &MatcherPtr) {
  // If we reached the end of the chain, we're done.
  Matcher *N = MatcherPtr.get();
  if (!N) return;
  
  // If this is not a push node, just scan for one.
  ScopeMatcher *Scope = dyn_cast<ScopeMatcher>(N);
  if (!Scope)
    return FactorNodes(N->getNextPtr());
  
  // Okay, pull together the children of the scope node into a vector so we can
  // inspect it more easily.  While we're at it, bucket them up by the hash
  // code of their first predicate.
  SmallVector<Matcher*, 32> OptionsToMatch;
  
  for (unsigned i = 0, e = Scope->getNumChildren(); i != e; ++i) {
    // Factor the subexpression.
    std::unique_ptr<Matcher> Child(Scope->takeChild(i));
    FactorNodes(Child);
    
    if (Matcher *N = Child.release())
      OptionsToMatch.push_back(N);
  }
  
  SmallVector<Matcher*, 32> NewOptionsToMatch;
  
  // Loop over options to match, merging neighboring patterns with identical
  // starting nodes into a shared matcher.
  for (unsigned OptionIdx = 0, e = OptionsToMatch.size(); OptionIdx != e;) {
    // Find the set of matchers that start with this node.
    Matcher *Optn = OptionsToMatch[OptionIdx++];

    if (OptionIdx == e) {
      NewOptionsToMatch.push_back(Optn);
      continue;
    }
    
    // See if the next option starts with the same matcher.  If the two
    // neighbors *do* start with the same matcher, we can factor the matcher out
    // of at least these two patterns.  See what the maximal set we can merge
    // together is.
    SmallVector<Matcher*, 8> EqualMatchers;
    EqualMatchers.push_back(Optn);
    
    // Factor all of the known-equal matchers after this one into the same
    // group.
    while (OptionIdx != e && OptionsToMatch[OptionIdx]->isEqual(Optn))
      EqualMatchers.push_back(OptionsToMatch[OptionIdx++]);

    // If we found a non-equal matcher, see if it is contradictory with the
    // current node.  If so, we know that the ordering relation between the
    // current sets of nodes and this node don't matter.  Look past it to see if
    // we can merge anything else into this matching group.
    unsigned Scan = OptionIdx;
    while (1) {
      // If we ran out of stuff to scan, we're done.
      if (Scan == e) break;
      
      Matcher *ScanMatcher = OptionsToMatch[Scan];
      
      // If we found an entry that matches out matcher, merge it into the set to
      // handle.
      if (Optn->isEqual(ScanMatcher)) {
        // If is equal after all, add the option to EqualMatchers and remove it
        // from OptionsToMatch.
        EqualMatchers.push_back(ScanMatcher);
        OptionsToMatch.erase(OptionsToMatch.begin()+Scan);
        --e;
        continue;
      }
      
      // If the option we're checking for contradicts the start of the list,
      // skip over it.
      if (Optn->isContradictory(ScanMatcher)) {
        ++Scan;
        continue;
      }

      // If we're scanning for a simple node, see if it occurs later in the
      // sequence.  If so, and if we can move it up, it might be contradictory
      // or the same as what we're looking for.  If so, reorder it.
      if (Optn->isSimplePredicateOrRecordNode()) {
        Matcher *M2 = FindNodeWithKind(ScanMatcher, Optn->getKind());
        if (M2 && M2 != ScanMatcher &&
            M2->canMoveBefore(ScanMatcher) &&
            (M2->isEqual(Optn) || M2->isContradictory(Optn))) {
          Matcher *MatcherWithoutM2 = ScanMatcher->unlinkNode(M2);
          M2->setNext(MatcherWithoutM2);
          OptionsToMatch[Scan] = M2;
          continue;
        }
      }
      
      // Otherwise, we don't know how to handle this entry, we have to bail.
      break;
    }
      
    if (Scan != e &&
        // Don't print it's obvious nothing extra could be merged anyway.
        Scan+1 != e) {
      DEBUG(errs() << "Couldn't merge this:\n";
            Optn->print(errs(), 4);
            errs() << "into this:\n";
            OptionsToMatch[Scan]->print(errs(), 4);
            if (Scan+1 != e)
              OptionsToMatch[Scan+1]->printOne(errs());
            if (Scan+2 < e)
              OptionsToMatch[Scan+2]->printOne(errs());
            errs() << "\n");
    }
    
    // If we only found one option starting with this matcher, no factoring is
    // possible.
    if (EqualMatchers.size() == 1) {
      NewOptionsToMatch.push_back(EqualMatchers[0]);
      continue;
    }
    
    // Factor these checks by pulling the first node off each entry and
    // discarding it.  Take the first one off the first entry to reuse.
    Matcher *Shared = Optn;
    Optn = Optn->takeNext();
    EqualMatchers[0] = Optn;

    // Remove and delete the first node from the other matchers we're factoring.
    for (unsigned i = 1, e = EqualMatchers.size(); i != e; ++i) {
      Matcher *Tmp = EqualMatchers[i]->takeNext();
      delete EqualMatchers[i];
      EqualMatchers[i] = Tmp;
    }
    
    Shared->setNext(new ScopeMatcher(EqualMatchers));

    // Recursively factor the newly created node.
    FactorNodes(Shared->getNextPtr());
    
    NewOptionsToMatch.push_back(Shared);
  }
  
  // If we're down to a single pattern to match, then we don't need this scope
  // anymore.
  if (NewOptionsToMatch.size() == 1) {
    MatcherPtr.reset(NewOptionsToMatch[0]);
    return;
  }
  
  if (NewOptionsToMatch.empty()) {
    MatcherPtr.reset();
    return;
  }
  
  // If our factoring failed (didn't achieve anything) see if we can simplify in
  // other ways.
  
  // Check to see if all of the leading entries are now opcode checks.  If so,
  // we can convert this Scope to be a OpcodeSwitch instead.
  bool AllOpcodeChecks = true, AllTypeChecks = true;
  for (unsigned i = 0, e = NewOptionsToMatch.size(); i != e; ++i) {
    // Check to see if this breaks a series of CheckOpcodeMatchers.
    if (AllOpcodeChecks &&
        !isa<CheckOpcodeMatcher>(NewOptionsToMatch[i])) {
#if 0
      if (i > 3) {
        errs() << "FAILING OPC #" << i << "\n";
        NewOptionsToMatch[i]->dump();
      }
#endif
      AllOpcodeChecks = false;
    }

    // Check to see if this breaks a series of CheckTypeMatcher's.
    if (AllTypeChecks) {
      CheckTypeMatcher *CTM =
        cast_or_null<CheckTypeMatcher>(FindNodeWithKind(NewOptionsToMatch[i],
                                                        Matcher::CheckType));
      if (!CTM ||
          // iPTR checks could alias any other case without us knowing, don't
          // bother with them.
          CTM->getType() == MVT::iPTR ||
          // SwitchType only works for result #0.
          CTM->getResNo() != 0 ||
          // If the CheckType isn't at the start of the list, see if we can move
          // it there.
          !CTM->canMoveBefore(NewOptionsToMatch[i])) {
#if 0
        if (i > 3 && AllTypeChecks) {
          errs() << "FAILING TYPE #" << i << "\n";
          NewOptionsToMatch[i]->dump();
        }
#endif
        AllTypeChecks = false;
      }
    }
  }
  
  // If all the options are CheckOpcode's, we can form the SwitchOpcode, woot.
  if (AllOpcodeChecks) {
    StringSet<> Opcodes;
    SmallVector<std::pair<const SDNodeInfo*, Matcher*>, 8> Cases;
    for (unsigned i = 0, e = NewOptionsToMatch.size(); i != e; ++i) {
      CheckOpcodeMatcher *COM = cast<CheckOpcodeMatcher>(NewOptionsToMatch[i]);
      assert(Opcodes.insert(COM->getOpcode().getEnumName()) &&
             "Duplicate opcodes not factored?");
      Cases.push_back(std::make_pair(&COM->getOpcode(), COM->getNext()));
    }
    
    MatcherPtr.reset(new SwitchOpcodeMatcher(Cases));
    return;
  }
  
  // If all the options are CheckType's, we can form the SwitchType, woot.
  if (AllTypeChecks) {
    DenseMap<unsigned, unsigned> TypeEntry;
    SmallVector<std::pair<MVT::SimpleValueType, Matcher*>, 8> Cases;
    for (unsigned i = 0, e = NewOptionsToMatch.size(); i != e; ++i) {
      CheckTypeMatcher *CTM =
        cast_or_null<CheckTypeMatcher>(FindNodeWithKind(NewOptionsToMatch[i],
                                                        Matcher::CheckType));
      Matcher *MatcherWithoutCTM = NewOptionsToMatch[i]->unlinkNode(CTM);
      MVT::SimpleValueType CTMTy = CTM->getType();
      delete CTM;
      
      unsigned &Entry = TypeEntry[CTMTy];
      if (Entry != 0) {
        // If we have unfactored duplicate types, then we should factor them.
        Matcher *PrevMatcher = Cases[Entry-1].second;
        if (ScopeMatcher *SM = dyn_cast<ScopeMatcher>(PrevMatcher)) {
          SM->setNumChildren(SM->getNumChildren()+1);
          SM->resetChild(SM->getNumChildren()-1, MatcherWithoutCTM);
          continue;
        }
        
        Matcher *Entries[2] = { PrevMatcher, MatcherWithoutCTM };
        Cases[Entry-1].second = new ScopeMatcher(Entries);
        continue;
      }
      
      Entry = Cases.size()+1;
      Cases.push_back(std::make_pair(CTMTy, MatcherWithoutCTM));
    }
    
    if (Cases.size() != 1) {
      MatcherPtr.reset(new SwitchTypeMatcher(Cases));
    } else {
      // If we factored and ended up with one case, create it now.
      MatcherPtr.reset(new CheckTypeMatcher(Cases[0].first, 0));
      MatcherPtr->setNext(Cases[0].second);
    }
    return;
  }
  

  // Reassemble the Scope node with the adjusted children.
  Scope->setNumChildren(NewOptionsToMatch.size());
  for (unsigned i = 0, e = NewOptionsToMatch.size(); i != e; ++i)
    Scope->resetChild(i, NewOptionsToMatch[i]);
}

Matcher *llvm::OptimizeMatcher(Matcher *TheMatcher,
                               const CodeGenDAGPatterns &CGP) {
  std::unique_ptr<Matcher> MatcherPtr(TheMatcher);
  ContractNodes(MatcherPtr, CGP);
  SinkPatternPredicates(MatcherPtr);
  FactorNodes(MatcherPtr);
  return MatcherPtr.release();
}