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/CodeGen/EdgeBundles.h" 33 #include "llvm/CodeGen/LiveIntervalAnalysis.h" 34 #include "llvm/CodeGen/MachineBasicBlock.h" 35 #include "llvm/CodeGen/MachineFunction.h" 36 #include "llvm/CodeGen/MachineLoopInfo.h" 37 #include "llvm/CodeGen/Passes.h" 38 #include "llvm/Support/Debug.h" 39 #include "llvm/Support/Format.h" 40 41 using namespace llvm; 42 43 char SpillPlacement::ID = 0; 44 INITIALIZE_PASS_BEGIN(SpillPlacement, "spill-code-placement", 45 "Spill Code Placement Analysis", true, true) 46 INITIALIZE_PASS_DEPENDENCY(EdgeBundles) 47 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 48 INITIALIZE_PASS_END(SpillPlacement, "spill-code-placement", 49 "Spill Code Placement Analysis", true, true) 50 51 char &llvm::SpillPlacementID = SpillPlacement::ID; 52 53 void SpillPlacement::getAnalysisUsage(AnalysisUsage &AU) const { 54 AU.setPreservesAll(); 55 AU.addRequiredTransitive<EdgeBundles>(); 56 AU.addRequiredTransitive<MachineLoopInfo>(); 57 MachineFunctionPass::getAnalysisUsage(AU); 58 } 59 60 /// Node - Each edge bundle corresponds to a Hopfield node. 61 /// 62 /// The node contains precomputed frequency data that only depends on the CFG, 63 /// but Bias and Links are computed each time placeSpills is called. 64 /// 65 /// The node Value is positive when the variable should be in a register. The 66 /// value can change when linked nodes change, but convergence is very fast 67 /// because all weights are positive. 68 /// 69 struct SpillPlacement::Node { 70 /// Scale - Inverse block frequency feeding into[0] or out of[1] the bundle. 71 /// Ideally, these two numbers should be identical, but inaccuracies in the 72 /// block frequency estimates means that we need to normalize ingoing and 73 /// outgoing frequencies separately so they are commensurate. 74 float Scale[2]; 75 76 /// Bias - Normalized contributions from non-transparent blocks. 77 /// A bundle connected to a MustSpill block has a huge negative bias, 78 /// otherwise it is a number in the range [-2;2]. 79 float Bias; 80 81 /// Value - Output value of this node computed from the Bias and links. 82 /// This is always in the range [-1;1]. A positive number means the variable 83 /// should go in a register through this bundle. 84 float Value; 85 86 typedef SmallVector<std::pair<float, unsigned>, 4> LinkVector; 87 88 /// Links - (Weight, BundleNo) for all transparent blocks connecting to other 89 /// bundles. The weights are all positive and add up to at most 2, weights 90 /// from ingoing and outgoing nodes separately add up to a most 1. The weight 91 /// sum can be less than 2 when the variable is not live into / out of some 92 /// connected basic blocks. 93 LinkVector Links; 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 // Actually, we must spill if Bias < sum(weights). 104 // It may be worth it to compute the weight sum here? 105 return Bias < -2.0f; 106 } 107 108 /// Node - Create a blank Node. 109 Node() { 110 Scale[0] = Scale[1] = 0; 111 } 112 113 /// clear - Reset per-query data, but preserve frequencies that only depend on 114 // the CFG. 115 void clear() { 116 Bias = Value = 0; 117 Links.clear(); 118 } 119 120 /// addLink - Add a link to bundle b with weight w. 121 /// out=0 for an ingoing link, and 1 for an outgoing link. 122 void addLink(unsigned b, float w, bool out) { 123 // Normalize w relative to all connected blocks from that direction. 124 w *= Scale[out]; 125 126 // There can be multiple links to the same bundle, add them up. 127 for (LinkVector::iterator I = Links.begin(), E = Links.end(); I != E; ++I) 128 if (I->second == b) { 129 I->first += w; 130 return; 131 } 132 // This must be the first link to b. 133 Links.push_back(std::make_pair(w, b)); 134 } 135 136 /// addBias - Bias this node from an ingoing[0] or outgoing[1] link. 137 /// Return the change to the total number of positive biases. 138 void addBias(float w, bool out) { 139 // Normalize w relative to all connected blocks from that direction. 140 w *= Scale[out]; 141 Bias += w; 142 } 143 144 /// update - Recompute Value from Bias and Links. Return true when node 145 /// preference changes. 146 bool update(const Node nodes[]) { 147 // Compute the weighted sum of inputs. 148 float Sum = Bias; 149 for (LinkVector::iterator I = Links.begin(), E = Links.end(); I != E; ++I) 150 Sum += I->first * nodes[I->second].Value; 151 152 // The weighted sum is going to be in the range [-2;2]. Ideally, we should 153 // simply set Value = sign(Sum), but we will add a dead zone around 0 for 154 // two reasons: 155 // 1. It avoids arbitrary bias when all links are 0 as is possible during 156 // initial iterations. 157 // 2. It helps tame rounding errors when the links nominally sum to 0. 158 const float Thres = 1e-4f; 159 bool Before = preferReg(); 160 if (Sum < -Thres) 161 Value = -1; 162 else if (Sum > Thres) 163 Value = 1; 164 else 165 Value = 0; 166 return Before != preferReg(); 167 } 168 }; 169 170 bool SpillPlacement::runOnMachineFunction(MachineFunction &mf) { 171 MF = &mf; 172 bundles = &getAnalysis<EdgeBundles>(); 173 loops = &getAnalysis<MachineLoopInfo>(); 174 175 assert(!nodes && "Leaking node array"); 176 nodes = new Node[bundles->getNumBundles()]; 177 178 // Compute total ingoing and outgoing block frequencies for all bundles. 179 BlockFrequency.resize(mf.getNumBlockIDs()); 180 for (MachineFunction::iterator I = mf.begin(), E = mf.end(); I != E; ++I) { 181 float Freq = LiveIntervals::getSpillWeight(true, false, 182 loops->getLoopDepth(I)); 183 unsigned Num = I->getNumber(); 184 BlockFrequency[Num] = Freq; 185 nodes[bundles->getBundle(Num, 1)].Scale[0] += Freq; 186 nodes[bundles->getBundle(Num, 0)].Scale[1] += Freq; 187 } 188 189 // Scales are reciprocal frequencies. 190 for (unsigned i = 0, e = bundles->getNumBundles(); i != e; ++i) 191 for (unsigned d = 0; d != 2; ++d) 192 if (nodes[i].Scale[d] > 0) 193 nodes[i].Scale[d] = 1 / nodes[i].Scale[d]; 194 195 // We never change the function. 196 return false; 197 } 198 199 void SpillPlacement::releaseMemory() { 200 delete[] nodes; 201 nodes = 0; 202 } 203 204 /// activate - mark node n as active if it wasn't already. 205 void SpillPlacement::activate(unsigned n) { 206 if (ActiveNodes->test(n)) 207 return; 208 ActiveNodes->set(n); 209 nodes[n].clear(); 210 211 // Very large bundles usually come from big switches, indirect branches, 212 // landing pads, or loops with many 'continue' statements. It is difficult to 213 // allocate registers when so many different blocks are involved. 214 // 215 // Give a small negative bias to large bundles such that 1/32 of the 216 // connected blocks need to be interested before we consider expanding the 217 // region through the bundle. This helps compile time by limiting the number 218 // of blocks visited and the number of links in the Hopfield network. 219 if (bundles->getBlocks(n).size() > 100) 220 nodes[n].Bias = -0.0625f; 221 } 222 223 224 /// addConstraints - Compute node biases and weights from a set of constraints. 225 /// Set a bit in NodeMask for each active node. 226 void SpillPlacement::addConstraints(ArrayRef<BlockConstraint> LiveBlocks) { 227 for (ArrayRef<BlockConstraint>::iterator I = LiveBlocks.begin(), 228 E = LiveBlocks.end(); I != E; ++I) { 229 float Freq = getBlockFrequency(I->Number); 230 const float Bias[] = { 231 0, // DontCare, 232 1, // PrefReg, 233 -1, // PrefSpill 234 0, // PrefBoth 235 -HUGE_VALF // MustSpill 236 }; 237 238 // Live-in to block? 239 if (I->Entry != DontCare) { 240 unsigned ib = bundles->getBundle(I->Number, 0); 241 activate(ib); 242 nodes[ib].addBias(Freq * Bias[I->Entry], 1); 243 } 244 245 // Live-out from block? 246 if (I->Exit != DontCare) { 247 unsigned ob = bundles->getBundle(I->Number, 1); 248 activate(ob); 249 nodes[ob].addBias(Freq * Bias[I->Exit], 0); 250 } 251 } 252 } 253 254 /// addPrefSpill - Same as addConstraints(PrefSpill) 255 void SpillPlacement::addPrefSpill(ArrayRef<unsigned> Blocks, bool Strong) { 256 for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end(); 257 I != E; ++I) { 258 float Freq = getBlockFrequency(*I); 259 if (Strong) 260 Freq += Freq; 261 unsigned ib = bundles->getBundle(*I, 0); 262 unsigned ob = bundles->getBundle(*I, 1); 263 activate(ib); 264 activate(ob); 265 nodes[ib].addBias(-Freq, 1); 266 nodes[ob].addBias(-Freq, 0); 267 } 268 } 269 270 void SpillPlacement::addLinks(ArrayRef<unsigned> Links) { 271 for (ArrayRef<unsigned>::iterator I = Links.begin(), E = Links.end(); I != E; 272 ++I) { 273 unsigned Number = *I; 274 unsigned ib = bundles->getBundle(Number, 0); 275 unsigned ob = bundles->getBundle(Number, 1); 276 277 // Ignore self-loops. 278 if (ib == ob) 279 continue; 280 activate(ib); 281 activate(ob); 282 if (nodes[ib].Links.empty() && !nodes[ib].mustSpill()) 283 Linked.push_back(ib); 284 if (nodes[ob].Links.empty() && !nodes[ob].mustSpill()) 285 Linked.push_back(ob); 286 float Freq = getBlockFrequency(Number); 287 nodes[ib].addLink(ob, Freq, 1); 288 nodes[ob].addLink(ib, Freq, 0); 289 } 290 } 291 292 bool SpillPlacement::scanActiveBundles() { 293 Linked.clear(); 294 RecentPositive.clear(); 295 for (int n = ActiveNodes->find_first(); n>=0; n = ActiveNodes->find_next(n)) { 296 nodes[n].update(nodes); 297 // A node that must spill, or a node without any links is not going to 298 // change its value ever again, so exclude it from iterations. 299 if (nodes[n].mustSpill()) 300 continue; 301 if (!nodes[n].Links.empty()) 302 Linked.push_back(n); 303 if (nodes[n].preferReg()) 304 RecentPositive.push_back(n); 305 } 306 return !RecentPositive.empty(); 307 } 308 309 /// iterate - Repeatedly update the Hopfield nodes until stability or the 310 /// maximum number of iterations is reached. 311 /// @param Linked - Numbers of linked nodes that need updating. 312 void SpillPlacement::iterate() { 313 // First update the recently positive nodes. They have likely received new 314 // negative bias that will turn them off. 315 while (!RecentPositive.empty()) 316 nodes[RecentPositive.pop_back_val()].update(nodes); 317 318 if (Linked.empty()) 319 return; 320 321 // Run up to 10 iterations. The edge bundle numbering is closely related to 322 // basic block numbering, so there is a strong tendency towards chains of 323 // linked nodes with sequential numbers. By scanning the linked nodes 324 // backwards and forwards, we make it very likely that a single node can 325 // affect the entire network in a single iteration. That means very fast 326 // convergence, usually in a single iteration. 327 for (unsigned iteration = 0; iteration != 10; ++iteration) { 328 // Scan backwards, skipping the last node which was just updated. 329 bool Changed = false; 330 for (SmallVectorImpl<unsigned>::const_reverse_iterator I = 331 llvm::next(Linked.rbegin()), E = Linked.rend(); I != E; ++I) { 332 unsigned n = *I; 333 if (nodes[n].update(nodes)) { 334 Changed = true; 335 if (nodes[n].preferReg()) 336 RecentPositive.push_back(n); 337 } 338 } 339 if (!Changed || !RecentPositive.empty()) 340 return; 341 342 // Scan forwards, skipping the first node which was just updated. 343 Changed = false; 344 for (SmallVectorImpl<unsigned>::const_iterator I = 345 llvm::next(Linked.begin()), E = Linked.end(); I != E; ++I) { 346 unsigned n = *I; 347 if (nodes[n].update(nodes)) { 348 Changed = true; 349 if (nodes[n].preferReg()) 350 RecentPositive.push_back(n); 351 } 352 } 353 if (!Changed || !RecentPositive.empty()) 354 return; 355 } 356 } 357 358 void SpillPlacement::prepare(BitVector &RegBundles) { 359 Linked.clear(); 360 RecentPositive.clear(); 361 // Reuse RegBundles as our ActiveNodes vector. 362 ActiveNodes = &RegBundles; 363 ActiveNodes->clear(); 364 ActiveNodes->resize(bundles->getNumBundles()); 365 } 366 367 bool 368 SpillPlacement::finish() { 369 assert(ActiveNodes && "Call prepare() first"); 370 371 // Write preferences back to ActiveNodes. 372 bool Perfect = true; 373 for (int n = ActiveNodes->find_first(); n>=0; n = ActiveNodes->find_next(n)) 374 if (!nodes[n].preferReg()) { 375 ActiveNodes->reset(n); 376 Perfect = false; 377 } 378 ActiveNodes = 0; 379 return Perfect; 380 } 381