lib/Support/FoldingSet.cpp

   1 //===-- Support/FoldingSet.cpp - Uniquing Hash Set --------------*- C++ -*-===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file was developed by James M. Laskey and is distributed under
   6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file implements a hash set that can be used to remove duplication of
  11 // nodes in a graph.  This code was originally created by Chris Lattner for use
  12 // with SelectionDAGCSEMap, but was isolated to provide use across the llvm code
  13 // set.
  14 //
  15 //===----------------------------------------------------------------------===//
  16
  17 #include "llvm/ADT/FoldingSet.h"
  18 #include "llvm/Support/MathExtras.h"
  19 #include <cassert>
  20 using namespace llvm;
  21
  22 //===----------------------------------------------------------------------===//
  23 // FoldingSetImpl::NodeID Implementation
  24
  25 /// Add* - Add various data types to Bit data.
  26 ///
  27 void FoldingSetImpl::NodeID::AddPointer(const void *Ptr) {
  28   // Note: this adds pointers to the hash using sizes and endianness that
  29   // depend on the host.  It doesn't matter however, because hashing on
  30   // pointer values in inherently unstable.  Nothing  should depend on the
  31   // ordering of nodes in the folding set.
  32   intptr_t PtrI = (intptr_t)Ptr;
  33   Bits.push_back(unsigned(PtrI));
  34   if (sizeof(intptr_t) > sizeof(unsigned))
  35     Bits.push_back(unsigned(uint64_t(PtrI) >> 32));
  36 }
  37 void FoldingSetImpl::NodeID::AddInteger(signed I) {
  38   Bits.push_back(I);
  39 }
  40 void FoldingSetImpl::NodeID::AddInteger(unsigned I) {
  41   Bits.push_back(I);
  42 }
  43 void FoldingSetImpl::NodeID::AddInteger(int64_t I) {
  44   AddInteger((uint64_t)I);
  45 }
  46 void FoldingSetImpl::NodeID::AddInteger(uint64_t I) {
  47   Bits.push_back(unsigned(I));
  48
  49   // If the integer is small, encode it just as 32-bits.
  50   if ((uint64_t)(int)I != I)
  51     Bits.push_back(unsigned(I >> 32));
  52 }
  53 void FoldingSetImpl::NodeID::AddFloat(float F) {
  54   Bits.push_back(FloatToBits(F));
  55 }
  56 void FoldingSetImpl::NodeID::AddDouble(double D) {
  57  AddInteger(DoubleToBits(D));
  58 }
  59 void FoldingSetImpl::NodeID::AddAPFloat(const APFloat& apf) {
  60   APInt api = apf.convertToAPInt();
  61   const uint64_t *p = api.getRawData();
  62   for (unsigned i=0; i<api.getNumWords(); i++)
  63     AddInteger(*p++);
  64 }
  65 void FoldingSetImpl::NodeID::AddString(const std::string &String) {
  66   unsigned Size = String.size();
  67   Bits.push_back(Size);
  68   if (!Size) return;
  69
  70   unsigned Units = Size / 4;
  71   unsigned Pos = 0;
  72   const unsigned *Base = (const unsigned *)String.data();
  73
  74   // If the string is aligned do a bulk transfer.
  75   if (!((intptr_t)Base & 3)) {
  76     Bits.append(Base, Base + Units);
  77     Pos = (Units + 1) * 4;
  78   } else {
  79     // Otherwise do it the hard way.
  80     for ( Pos += 4; Pos <= Size; Pos += 4) {
  81       unsigned V = ((unsigned char)String[Pos - 4] << 24) |
  82                    ((unsigned char)String[Pos - 3] << 16) |
  83                    ((unsigned char)String[Pos - 2] << 8) |
  84                     (unsigned char)String[Pos - 1];
  85       Bits.push_back(V);
  86     }
  87   }
  88
  89   // With the leftover bits.
  90   unsigned V = 0;
  91   // Pos will have overshot size by 4 - #bytes left over.
  92   switch (Pos - Size) {
  93   case 1: V = (V << 8) | (unsigned char)String[Size - 3]; // Fall thru.
  94   case 2: V = (V << 8) | (unsigned char)String[Size - 2]; // Fall thru.
  95   case 3: V = (V << 8) | (unsigned char)String[Size - 1]; break;
  96   default: return; // Nothing left.
  97   }
  98
  99   Bits.push_back(V);
 100 }
 101
 102 /// ComputeHash - Compute a strong hash value for this NodeID, used to
 103 /// lookup the node in the FoldingSetImpl.
 104 unsigned FoldingSetImpl::NodeID::ComputeHash() const {
 105   // This is adapted from SuperFastHash by Paul Hsieh.
 106   unsigned Hash = Bits.size();
 107   for (const unsigned *BP = &Bits[0], *E = BP+Bits.size(); BP != E; ++BP) {
 108     unsigned Data = *BP;
 109     Hash         += Data & 0xFFFF;
 110     unsigned Tmp  = ((Data >> 16) << 11) ^ Hash;
 111     Hash          = (Hash << 16) ^ Tmp;
 112     Hash         += Hash >> 11;
 113   }
 114
 115   // Force "avalanching" of final 127 bits.
 116   Hash ^= Hash << 3;
 117   Hash += Hash >> 5;
 118   Hash ^= Hash << 4;
 119   Hash += Hash >> 17;
 120   Hash ^= Hash << 25;
 121   Hash += Hash >> 6;
 122   return Hash;
 123 }
 124
 125 /// operator== - Used to compare two nodes to each other.
 126 ///
 127 bool FoldingSetImpl::NodeID::operator==(const FoldingSetImpl::NodeID &RHS)const{
 128   if (Bits.size() != RHS.Bits.size()) return false;
 129   return memcmp(&Bits[0], &RHS.Bits[0], Bits.size()*sizeof(Bits[0])) == 0;
 130 }
 131
 132
 133 //===----------------------------------------------------------------------===//
 134 /// Helper functions for FoldingSetImpl.
 135
 136 /// GetNextPtr - In order to save space, each bucket is a
 137 /// singly-linked-list. In order to make deletion more efficient, we make
 138 /// the list circular, so we can delete a node without computing its hash.
 139 /// The problem with this is that the start of the hash buckets are not
 140 /// Nodes.  If NextInBucketPtr is a bucket pointer, this method returns null:
 141 /// use GetBucketPtr when this happens.
 142 static FoldingSetImpl::Node *GetNextPtr(void *NextInBucketPtr,
 143                                         void **Buckets, unsigned NumBuckets) {
 144   if (NextInBucketPtr >= Buckets && NextInBucketPtr < Buckets + NumBuckets)
 145     return 0;
 146   return static_cast<FoldingSetImpl::Node*>(NextInBucketPtr);
 147 }
 148
 149 /// GetBucketPtr - Provides a casting of a bucket pointer for isNode
 150 /// testing.
 151 static void **GetBucketPtr(void *NextInBucketPtr) {
 152   return static_cast<void**>(NextInBucketPtr);
 153 }
 154
 155 /// GetBucketFor - Hash the specified node ID and return the hash bucket for
 156 /// the specified ID.
 157 static void **GetBucketFor(const FoldingSetImpl::NodeID &ID,
 158                            void **Buckets, unsigned NumBuckets) {
 159   // NumBuckets is always a power of 2.
 160   unsigned BucketNum = ID.ComputeHash() & (NumBuckets-1);
 161   return Buckets + BucketNum;
 162 }
 163
 164 //===----------------------------------------------------------------------===//
 165 // FoldingSetImpl Implementation
 166
 167 FoldingSetImpl::FoldingSetImpl(unsigned Log2InitSize) : NumNodes(0) {
 168   assert(5 < Log2InitSize && Log2InitSize < 32 &&
 169          "Initial hash table size out of range");
 170   NumBuckets = 1 << Log2InitSize;
 171   Buckets = new void*[NumBuckets];
 172   memset(Buckets, 0, NumBuckets*sizeof(void*));
 173 }
 174 FoldingSetImpl::~FoldingSetImpl() {
 175   delete [] Buckets;
 176 }
 177
 178 /// GrowHashTable - Double the size of the hash table and rehash everything.
 179 ///
 180 void FoldingSetImpl::GrowHashTable() {
 181   void **OldBuckets = Buckets;
 182   unsigned OldNumBuckets = NumBuckets;
 183   NumBuckets <<= 1;
 184
 185   // Reset the node count to zero: we're going to reinsert everything.
 186   NumNodes = 0;
 187
 188   // Clear out new buckets.
 189   Buckets = new void*[NumBuckets];
 190   memset(Buckets, 0, NumBuckets*sizeof(void*));
 191
 192   // Walk the old buckets, rehashing nodes into their new place.
 193   for (unsigned i = 0; i != OldNumBuckets; ++i) {
 194     void *Probe = OldBuckets[i];
 195     if (!Probe) continue;
 196     while (Node *NodeInBucket = GetNextPtr(Probe, OldBuckets, OldNumBuckets)) {
 197       // Figure out the next link, remove NodeInBucket from the old link.
 198       Probe = NodeInBucket->getNextInBucket();
 199       NodeInBucket->SetNextInBucket(0);
 200
 201       // Insert the node into the new bucket, after recomputing the hash.
 202       NodeID ID;
 203       GetNodeProfile(ID, NodeInBucket);
 204       InsertNode(NodeInBucket, GetBucketFor(ID, Buckets, NumBuckets));
 205     }
 206   }
 207
 208   delete[] OldBuckets;
 209 }
 210
 211 /// FindNodeOrInsertPos - Look up the node specified by ID.  If it exists,
 212 /// return it.  If not, return the insertion token that will make insertion
 213 /// faster.
 214 FoldingSetImpl::Node *FoldingSetImpl::FindNodeOrInsertPos(const NodeID &ID,
 215                                                           void *&InsertPos) {
 216   void **Bucket = GetBucketFor(ID, Buckets, NumBuckets);
 217   void *Probe = *Bucket;
 218
 219   InsertPos = 0;
 220
 221   while (Node *NodeInBucket = GetNextPtr(Probe, Buckets, NumBuckets)) {
 222     NodeID OtherID;
 223     GetNodeProfile(OtherID, NodeInBucket);
 224     if (OtherID == ID)
 225       return NodeInBucket;
 226
 227     Probe = NodeInBucket->getNextInBucket();
 228   }
 229
 230   // Didn't find the node, return null with the bucket as the InsertPos.
 231   InsertPos = Bucket;
 232   return 0;
 233 }
 234
 235 /// InsertNode - Insert the specified node into the folding set, knowing that it
 236 /// is not already in the map.  InsertPos must be obtained from
 237 /// FindNodeOrInsertPos.
 238 void FoldingSetImpl::InsertNode(Node *N, void *InsertPos) {
 239   assert(N->getNextInBucket() == 0);
 240   // Do we need to grow the hashtable?
 241   if (NumNodes+1 > NumBuckets*2) {
 242     GrowHashTable();
 243     NodeID ID;
 244     GetNodeProfile(ID, N);
 245     InsertPos = GetBucketFor(ID, Buckets, NumBuckets);
 246   }
 247
 248   ++NumNodes;
 249
 250   /// The insert position is actually a bucket pointer.
 251   void **Bucket = static_cast<void**>(InsertPos);
 252
 253   void *Next = *Bucket;
 254
 255   // If this is the first insertion into this bucket, its next pointer will be
 256   // null.  Pretend as if it pointed to itself.
 257   if (Next == 0)
 258     Next = Bucket;
 259
 260   // Set the node's next pointer, and make the bucket point to the node.
 261   N->SetNextInBucket(Next);
 262   *Bucket = N;
 263 }
 264
 265 /// RemoveNode - Remove a node from the folding set, returning true if one was
 266 /// removed or false if the node was not in the folding set.
 267 bool FoldingSetImpl::RemoveNode(Node *N) {
 268   // Because each bucket is a circular list, we don't need to compute N's hash
 269   // to remove it.
 270   void *Ptr = N->getNextInBucket();
 271   if (Ptr == 0) return false;  // Not in folding set.
 272
 273   --NumNodes;
 274   N->SetNextInBucket(0);
 275
 276   // Remember what N originally pointed to, either a bucket or another node.
 277   void *NodeNextPtr = Ptr;
 278
 279   // Chase around the list until we find the node (or bucket) which points to N.
 280   while (true) {
 281     if (Node *NodeInBucket = GetNextPtr(Ptr, Buckets, NumBuckets)) {
 282       // Advance pointer.
 283       Ptr = NodeInBucket->getNextInBucket();
 284
 285       // We found a node that points to N, change it to point to N's next node,
 286       // removing N from the list.
 287       if (Ptr == N) {
 288         NodeInBucket->SetNextInBucket(NodeNextPtr);
 289         return true;
 290       }
 291     } else {
 292       void **Bucket = GetBucketPtr(Ptr);
 293       Ptr = *Bucket;
 294
 295       // If we found that the bucket points to N, update the bucket to point to
 296       // whatever is next.
 297       if (Ptr == N) {
 298         *Bucket = NodeNextPtr;
 299         return true;
 300       }
 301     }
 302   }
 303 }
 304
 305 /// GetOrInsertNode - If there is an existing simple Node exactly
 306 /// equal to the specified node, return it.  Otherwise, insert 'N' and it
 307 /// instead.
 308 FoldingSetImpl::Node *FoldingSetImpl::GetOrInsertNode(FoldingSetImpl::Node *N) {
 309   NodeID ID;
 310   GetNodeProfile(ID, N);
 311   void *IP;
 312   if (Node *E = FindNodeOrInsertPos(ID, IP))
 313     return E;
 314   InsertNode(N, IP);
 315   return N;
 316 }