SpillPlacement.cpp revision 218885
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 41using namespace llvm; 42 43char SpillPlacement::ID = 0; 44INITIALIZE_PASS_BEGIN(SpillPlacement, "spill-code-placement", 45 "Spill Code Placement Analysis", true, true) 46INITIALIZE_PASS_DEPENDENCY(EdgeBundles) 47INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 48INITIALIZE_PASS_END(SpillPlacement, "spill-code-placement", 49 "Spill Code Placement Analysis", true, true) 50 51char &llvm::SpillPlacementID = SpillPlacement::ID; 52 53void 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/// 69struct SpillPlacement::Node { 70 /// Frequency - Total 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 Frequency[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 Frequency[0] = Frequency[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 /= Frequency[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 void addBias(float w, bool out) { 138 // Normalize w relative to all connected blocks from that direction. 139 w /= Frequency[out]; 140 Bias += w; 141 } 142 143 /// update - Recompute Value from Bias and Links. Return true when node 144 /// preference changes. 145 bool update(const Node nodes[]) { 146 // Compute the weighted sum of inputs. 147 float Sum = Bias; 148 for (LinkVector::iterator I = Links.begin(), E = Links.end(); I != E; ++I) 149 Sum += I->first * nodes[I->second].Value; 150 151 // The weighted sum is going to be in the range [-2;2]. Ideally, we should 152 // simply set Value = sign(Sum), but we will add a dead zone around 0 for 153 // two reasons: 154 // 1. It avoids arbitrary bias when all links are 0 as is possible during 155 // initial iterations. 156 // 2. It helps tame rounding errors when the links nominally sum to 0. 157 const float Thres = 1e-4f; 158 bool Before = preferReg(); 159 if (Sum < -Thres) 160 Value = -1; 161 else if (Sum > Thres) 162 Value = 1; 163 else 164 Value = 0; 165 return Before != preferReg(); 166 } 167}; 168 169bool SpillPlacement::runOnMachineFunction(MachineFunction &mf) { 170 MF = &mf; 171 bundles = &getAnalysis<EdgeBundles>(); 172 loops = &getAnalysis<MachineLoopInfo>(); 173 174 assert(!nodes && "Leaking node array"); 175 nodes = new Node[bundles->getNumBundles()]; 176 177 // Compute total ingoing and outgoing block frequencies for all bundles. 178 for (MachineFunction::iterator I = mf.begin(), E = mf.end(); I != E; ++I) { 179 float Freq = getBlockFrequency(I); 180 unsigned Num = I->getNumber(); 181 nodes[bundles->getBundle(Num, 1)].Frequency[0] += Freq; 182 nodes[bundles->getBundle(Num, 0)].Frequency[1] += Freq; 183 } 184 185 // We never change the function. 186 return false; 187} 188 189void SpillPlacement::releaseMemory() { 190 delete[] nodes; 191 nodes = 0; 192} 193 194/// activate - mark node n as active if it wasn't already. 195void SpillPlacement::activate(unsigned n) { 196 if (ActiveNodes->test(n)) 197 return; 198 ActiveNodes->set(n); 199 nodes[n].clear(); 200} 201 202 203/// prepareNodes - Compute node biases and weights from a set of constraints. 204/// Set a bit in NodeMask for each active node. 205void SpillPlacement:: 206prepareNodes(const SmallVectorImpl<BlockConstraint> &LiveBlocks) { 207 for (SmallVectorImpl<BlockConstraint>::const_iterator I = LiveBlocks.begin(), 208 E = LiveBlocks.end(); I != E; ++I) { 209 MachineBasicBlock *MBB = MF->getBlockNumbered(I->Number); 210 float Freq = getBlockFrequency(MBB); 211 212 // Is this a transparent block? Link ingoing and outgoing bundles. 213 if (I->Entry == DontCare && I->Exit == DontCare) { 214 unsigned ib = bundles->getBundle(I->Number, 0); 215 unsigned ob = bundles->getBundle(I->Number, 1); 216 217 // Ignore self-loops. 218 if (ib == ob) 219 continue; 220 activate(ib); 221 activate(ob); 222 nodes[ib].addLink(ob, Freq, 1); 223 nodes[ob].addLink(ib, Freq, 0); 224 continue; 225 } 226 227 // This block is not transparent, but it can still add bias. 228 const float Bias[] = { 229 0, // DontCare, 230 1, // PrefReg, 231 -1, // PrefSpill 232 -HUGE_VALF // MustSpill 233 }; 234 235 // Live-in to block? 236 if (I->Entry != DontCare) { 237 unsigned ib = bundles->getBundle(I->Number, 0); 238 activate(ib); 239 nodes[ib].addBias(Freq * Bias[I->Entry], 1); 240 } 241 242 // Live-out from block? 243 if (I->Exit != DontCare) { 244 unsigned ob = bundles->getBundle(I->Number, 1); 245 activate(ob); 246 nodes[ob].addBias(Freq * Bias[I->Exit], 0); 247 } 248 } 249} 250 251/// iterate - Repeatedly update the Hopfield nodes until stability or the 252/// maximum number of iterations is reached. 253/// @param Linked - Numbers of linked nodes that need updating. 254void SpillPlacement::iterate(const SmallVectorImpl<unsigned> &Linked) { 255 if (Linked.empty()) 256 return; 257 258 // Run up to 10 iterations. The edge bundle numbering is closely related to 259 // basic block numbering, so there is a strong tendency towards chains of 260 // linked nodes with sequential numbers. By scanning the linked nodes 261 // backwards and forwards, we make it very likely that a single node can 262 // affect the entire network in a single iteration. That means very fast 263 // convergence, usually in a single iteration. 264 for (unsigned iteration = 0; iteration != 10; ++iteration) { 265 // Scan backwards, skipping the last node which was just updated. 266 bool Changed = false; 267 for (SmallVectorImpl<unsigned>::const_reverse_iterator I = 268 llvm::next(Linked.rbegin()), E = Linked.rend(); I != E; ++I) { 269 unsigned n = *I; 270 bool C = nodes[n].update(nodes); 271 Changed |= C; 272 } 273 if (!Changed) 274 return; 275 276 // Scan forwards, skipping the first node which was just updated. 277 Changed = false; 278 for (SmallVectorImpl<unsigned>::const_iterator I = 279 llvm::next(Linked.begin()), E = Linked.end(); I != E; ++I) { 280 unsigned n = *I; 281 bool C = nodes[n].update(nodes); 282 Changed |= C; 283 } 284 if (!Changed) 285 return; 286 } 287} 288 289bool 290SpillPlacement::placeSpills(const SmallVectorImpl<BlockConstraint> &LiveBlocks, 291 BitVector &RegBundles) { 292 // Reuse RegBundles as our ActiveNodes vector. 293 ActiveNodes = &RegBundles; 294 ActiveNodes->clear(); 295 ActiveNodes->resize(bundles->getNumBundles()); 296 297 // Compute active nodes, links and biases. 298 prepareNodes(LiveBlocks); 299 300 // Update all active nodes, and find the ones that are actually linked to 301 // something so their value may change when iterating. 302 SmallVector<unsigned, 8> Linked; 303 for (int n = RegBundles.find_first(); n>=0; n = RegBundles.find_next(n)) { 304 nodes[n].update(nodes); 305 // A node that must spill, or a node without any links is not going to 306 // change its value ever again, so exclude it from iterations. 307 if (!nodes[n].Links.empty() && !nodes[n].mustSpill()) 308 Linked.push_back(n); 309 } 310 311 // Iterate the network to convergence. 312 iterate(Linked); 313 314 // Write preferences back to RegBundles. 315 bool Perfect = true; 316 for (int n = RegBundles.find_first(); n>=0; n = RegBundles.find_next(n)) 317 if (!nodes[n].preferReg()) { 318 RegBundles.reset(n); 319 Perfect = false; 320 } 321 return Perfect; 322} 323 324/// getBlockFrequency - Return our best estimate of the block frequency which is 325/// the expected number of block executions per function invocation. 326float SpillPlacement::getBlockFrequency(const MachineBasicBlock *MBB) { 327 // Use the unnormalized spill weight for real block frequencies. 328 return LiveIntervals::getSpillWeight(true, false, loops->getLoopDepth(MBB)); 329} 330 331