1 //===-- SpillPlacement.cpp - Optimal Spill Code Placement -----------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the spill code placement analysis. 11 // 12 // Each edge bundle corresponds to a node in a Hopfield network. Constraints on 13 // basic blocks are weighted by the block frequency and added to become the node 14 // bias. 15 // 16 // Transparent basic blocks have the variable live through, but don't care if it 17 // is spilled or in a register. These blocks become connections in the Hopfield 18 // network, again weighted by block frequency. 19 // 20 // The Hopfield network minimizes (possibly locally) its energy function: 21 // 22 // E = -sum_n V_n * ( B_n + sum_{n, m linked by b} V_m * F_b ) 23 // 24 // The energy function represents the expected spill code execution frequency, 25 // or the cost of spilling. This is a Lyapunov function which never increases 26 // when a node is updated. It is guaranteed to converge to a local minimum. 27 // 28 //===----------------------------------------------------------------------===// 29 30 #define DEBUG_TYPE "spillplacement" 31 #include "SpillPlacement.h" 32 #include "llvm/ADT/BitVector.h" 33 #include "llvm/CodeGen/EdgeBundles.h" 34 #include "llvm/CodeGen/MachineBasicBlock.h" 35 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 36 #include "llvm/CodeGen/MachineFunction.h" 37 #include "llvm/CodeGen/MachineLoopInfo.h" 38 #include "llvm/CodeGen/Passes.h" 39 #include "llvm/Support/Debug.h" 40 #include "llvm/Support/Format.h" 41 42 using namespace llvm; 43 44 char SpillPlacement::ID = 0; 45 INITIALIZE_PASS_BEGIN(SpillPlacement, "spill-code-placement", 46 "Spill Code Placement Analysis", true, true) 47 INITIALIZE_PASS_DEPENDENCY(EdgeBundles) 48 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 49 INITIALIZE_PASS_END(SpillPlacement, "spill-code-placement", 50 "Spill Code Placement Analysis", true, true) 51 52 char &llvm::SpillPlacementID = SpillPlacement::ID; 53 54 void SpillPlacement::getAnalysisUsage(AnalysisUsage &AU) const { 55 AU.setPreservesAll(); 56 AU.addRequired<MachineBlockFrequencyInfo>(); 57 AU.addRequiredTransitive<EdgeBundles>(); 58 AU.addRequiredTransitive<MachineLoopInfo>(); 59 MachineFunctionPass::getAnalysisUsage(AU); 60 } 61 62 /// Decision threshold. A node gets the output value 0 if the weighted sum of 63 /// its inputs falls in the open interval (-Threshold;Threshold). 64 static const BlockFrequency Threshold = 2; 65 66 /// Node - Each edge bundle corresponds to a Hopfield node. 67 /// 68 /// The node contains precomputed frequency data that only depends on the CFG, 69 /// but Bias and Links are computed each time placeSpills is called. 70 /// 71 /// The node Value is positive when the variable should be in a register. The 72 /// value can change when linked nodes change, but convergence is very fast 73 /// because all weights are positive. 74 /// 75 struct SpillPlacement::Node { 76 /// BiasN - Sum of blocks that prefer a spill. 77 BlockFrequency BiasN; 78 /// BiasP - Sum of blocks that prefer a register. 79 BlockFrequency BiasP; 80 81 /// Value - Output value of this node computed from the Bias and links. 82 /// This is always on of the values {-1, 0, 1}. A positive number means the 83 /// variable should go in a register through this bundle. 84 int Value; 85 86 typedef SmallVector<std::pair<BlockFrequency, unsigned>, 4> LinkVector; 87 88 /// Links - (Weight, BundleNo) for all transparent blocks connecting to other 89 /// bundles. The weights are all positive block frequencies. 90 LinkVector Links; 91 92 /// SumLinkWeights - Cached sum of the weights of all links + ThresHold. 93 BlockFrequency SumLinkWeights; 94 95 /// preferReg - Return true when this node prefers to be in a register. 96 bool preferReg() const { 97 // Undecided nodes (Value==0) go on the stack. 98 return Value > 0; 99 } 100 101 /// mustSpill - Return True if this node is so biased that it must spill. 102 bool mustSpill() const { 103 // We must spill if Bias < -sum(weights) or the MustSpill flag was set. 104 // BiasN is saturated when MustSpill is set, make sure this still returns 105 // true when the RHS saturates. Note that SumLinkWeights includes Threshold. 106 return BiasN >= BiasP + SumLinkWeights; 107 } 108 109 /// clear - Reset per-query data, but preserve frequencies that only depend on 110 // the CFG. 111 void clear() { 112 BiasN = BiasP = Value = 0; 113 SumLinkWeights = Threshold; 114 Links.clear(); 115 } 116 117 /// addLink - Add a link to bundle b with weight w. 118 void addLink(unsigned b, BlockFrequency w) { 119 // Update cached sum. 120 SumLinkWeights += w; 121 122 // There can be multiple links to the same bundle, add them up. 123 for (LinkVector::iterator I = Links.begin(), E = Links.end(); I != E; ++I) 124 if (I->second == b) { 125 I->first += w; 126 return; 127 } 128 // This must be the first link to b. 129 Links.push_back(std::make_pair(w, b)); 130 } 131 132 /// addBias - Bias this node. 133 void addBias(BlockFrequency freq, BorderConstraint direction) { 134 switch (direction) { 135 default: 136 break; 137 case PrefReg: 138 BiasP += freq; 139 break; 140 case PrefSpill: 141 BiasN += freq; 142 break; 143 case MustSpill: 144 BiasN = BlockFrequency::getMaxFrequency(); 145 break; 146 } 147 } 148 149 /// update - Recompute Value from Bias and Links. Return true when node 150 /// preference changes. 151 bool update(const Node nodes[]) { 152 // Compute the weighted sum of inputs. 153 BlockFrequency SumN = BiasN; 154 BlockFrequency SumP = BiasP; 155 for (LinkVector::iterator I = Links.begin(), E = Links.end(); I != E; ++I) { 156 if (nodes[I->second].Value == -1) 157 SumN += I->first; 158 else if (nodes[I->second].Value == 1) 159 SumP += I->first; 160 } 161 162 // Each weighted sum is going to be less than the total frequency of the 163 // bundle. Ideally, we should simply set Value = sign(SumP - SumN), but we 164 // will add a dead zone around 0 for two reasons: 165 // 166 // 1. It avoids arbitrary bias when all links are 0 as is possible during 167 // initial iterations. 168 // 2. It helps tame rounding errors when the links nominally sum to 0. 169 // 170 bool Before = preferReg(); 171 if (SumN >= SumP + Threshold) 172 Value = -1; 173 else if (SumP >= SumN + Threshold) 174 Value = 1; 175 else 176 Value = 0; 177 return Before != preferReg(); 178 } 179 }; 180 181 bool SpillPlacement::runOnMachineFunction(MachineFunction &mf) { 182 MF = &mf; 183 bundles = &getAnalysis<EdgeBundles>(); 184 loops = &getAnalysis<MachineLoopInfo>(); 185 186 assert(!nodes && "Leaking node array"); 187 nodes = new Node[bundles->getNumBundles()]; 188 189 // Compute total ingoing and outgoing block frequencies for all bundles. 190 BlockFrequencies.resize(mf.getNumBlockIDs()); 191 MachineBlockFrequencyInfo &MBFI = getAnalysis<MachineBlockFrequencyInfo>(); 192 for (MachineFunction::iterator I = mf.begin(), E = mf.end(); I != E; ++I) { 193 unsigned Num = I->getNumber(); 194 BlockFrequencies[Num] = MBFI.getBlockFreq(I); 195 } 196 197 // We never change the function. 198 return false; 199 } 200 201 void SpillPlacement::releaseMemory() { 202 delete[] nodes; 203 nodes = 0; 204 } 205 206 /// activate - mark node n as active if it wasn't already. 207 void SpillPlacement::activate(unsigned n) { 208 if (ActiveNodes->test(n)) 209 return; 210 ActiveNodes->set(n); 211 nodes[n].clear(); 212 213 // Very large bundles usually come from big switches, indirect branches, 214 // landing pads, or loops with many 'continue' statements. It is difficult to 215 // allocate registers when so many different blocks are involved. 216 // 217 // Give a small negative bias to large bundles such that a substantial 218 // fraction of the connected blocks need to be interested before we consider 219 // expanding the region through the bundle. This helps compile time by 220 // limiting the number of blocks visited and the number of links in the 221 // Hopfield network. 222 if (bundles->getBlocks(n).size() > 100) { 223 nodes[n].BiasP = 0; 224 nodes[n].BiasN = (BlockFrequency::getEntryFrequency() / 16); 225 } 226 } 227 228 229 /// addConstraints - Compute node biases and weights from a set of constraints. 230 /// Set a bit in NodeMask for each active node. 231 void SpillPlacement::addConstraints(ArrayRef<BlockConstraint> LiveBlocks) { 232 for (ArrayRef<BlockConstraint>::iterator I = LiveBlocks.begin(), 233 E = LiveBlocks.end(); I != E; ++I) { 234 BlockFrequency Freq = BlockFrequencies[I->Number]; 235 236 // Live-in to block? 237 if (I->Entry != DontCare) { 238 unsigned ib = bundles->getBundle(I->Number, 0); 239 activate(ib); 240 nodes[ib].addBias(Freq, I->Entry); 241 } 242 243 // Live-out from block? 244 if (I->Exit != DontCare) { 245 unsigned ob = bundles->getBundle(I->Number, 1); 246 activate(ob); 247 nodes[ob].addBias(Freq, I->Exit); 248 } 249 } 250 } 251 252 /// addPrefSpill - Same as addConstraints(PrefSpill) 253 void SpillPlacement::addPrefSpill(ArrayRef<unsigned> Blocks, bool Strong) { 254 for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end(); 255 I != E; ++I) { 256 BlockFrequency Freq = BlockFrequencies[*I]; 257 if (Strong) 258 Freq += Freq; 259 unsigned ib = bundles->getBundle(*I, 0); 260 unsigned ob = bundles->getBundle(*I, 1); 261 activate(ib); 262 activate(ob); 263 nodes[ib].addBias(Freq, PrefSpill); 264 nodes[ob].addBias(Freq, PrefSpill); 265 } 266 } 267 268 void SpillPlacement::addLinks(ArrayRef<unsigned> Links) { 269 for (ArrayRef<unsigned>::iterator I = Links.begin(), E = Links.end(); I != E; 270 ++I) { 271 unsigned Number = *I; 272 unsigned ib = bundles->getBundle(Number, 0); 273 unsigned ob = bundles->getBundle(Number, 1); 274 275 // Ignore self-loops. 276 if (ib == ob) 277 continue; 278 activate(ib); 279 activate(ob); 280 if (nodes[ib].Links.empty() && !nodes[ib].mustSpill()) 281 Linked.push_back(ib); 282 if (nodes[ob].Links.empty() && !nodes[ob].mustSpill()) 283 Linked.push_back(ob); 284 BlockFrequency Freq = BlockFrequencies[Number]; 285 nodes[ib].addLink(ob, Freq); 286 nodes[ob].addLink(ib, Freq); 287 } 288 } 289 290 bool SpillPlacement::scanActiveBundles() { 291 Linked.clear(); 292 RecentPositive.clear(); 293 for (int n = ActiveNodes->find_first(); n>=0; n = ActiveNodes->find_next(n)) { 294 nodes[n].update(nodes); 295 // A node that must spill, or a node without any links is not going to 296 // change its value ever again, so exclude it from iterations. 297 if (nodes[n].mustSpill()) 298 continue; 299 if (!nodes[n].Links.empty()) 300 Linked.push_back(n); 301 if (nodes[n].preferReg()) 302 RecentPositive.push_back(n); 303 } 304 return !RecentPositive.empty(); 305 } 306 307 /// iterate - Repeatedly update the Hopfield nodes until stability or the 308 /// maximum number of iterations is reached. 309 /// @param Linked - Numbers of linked nodes that need updating. 310 void SpillPlacement::iterate() { 311 // First update the recently positive nodes. They have likely received new 312 // negative bias that will turn them off. 313 while (!RecentPositive.empty()) 314 nodes[RecentPositive.pop_back_val()].update(nodes); 315 316 if (Linked.empty()) 317 return; 318 319 // Run up to 10 iterations. The edge bundle numbering is closely related to 320 // basic block numbering, so there is a strong tendency towards chains of 321 // linked nodes with sequential numbers. By scanning the linked nodes 322 // backwards and forwards, we make it very likely that a single node can 323 // affect the entire network in a single iteration. That means very fast 324 // convergence, usually in a single iteration. 325 for (unsigned iteration = 0; iteration != 10; ++iteration) { 326 // Scan backwards, skipping the last node which was just updated. 327 bool Changed = false; 328 for (SmallVectorImpl<unsigned>::const_reverse_iterator I = 329 llvm::next(Linked.rbegin()), E = Linked.rend(); I != E; ++I) { 330 unsigned n = *I; 331 if (nodes[n].update(nodes)) { 332 Changed = true; 333 if (nodes[n].preferReg()) 334 RecentPositive.push_back(n); 335 } 336 } 337 if (!Changed || !RecentPositive.empty()) 338 return; 339 340 // Scan forwards, skipping the first node which was just updated. 341 Changed = false; 342 for (SmallVectorImpl<unsigned>::const_iterator I = 343 llvm::next(Linked.begin()), E = Linked.end(); I != E; ++I) { 344 unsigned n = *I; 345 if (nodes[n].update(nodes)) { 346 Changed = true; 347 if (nodes[n].preferReg()) 348 RecentPositive.push_back(n); 349 } 350 } 351 if (!Changed || !RecentPositive.empty()) 352 return; 353 } 354 } 355 356 void SpillPlacement::prepare(BitVector &RegBundles) { 357 Linked.clear(); 358 RecentPositive.clear(); 359 // Reuse RegBundles as our ActiveNodes vector. 360 ActiveNodes = &RegBundles; 361 ActiveNodes->clear(); 362 ActiveNodes->resize(bundles->getNumBundles()); 363 } 364 365 bool 366 SpillPlacement::finish() { 367 assert(ActiveNodes && "Call prepare() first"); 368 369 // Write preferences back to ActiveNodes. 370 bool Perfect = true; 371 for (int n = ActiveNodes->find_first(); n>=0; n = ActiveNodes->find_next(n)) 372 if (!nodes[n].preferReg()) { 373 ActiveNodes->reset(n); 374 Perfect = false; 375 } 376 ActiveNodes = 0; 377 return Perfect; 378 } 379