1 //===- SpillPlacement.cpp - Optimal Spill Code Placement ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the spill code placement analysis.
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
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.
20 // The Hopfield network minimizes (possibly locally) its energy function:
22 // E = -sum_n V_n * ( B_n + sum_{n, m linked by b} V_m * F_b )
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.
28 //===----------------------------------------------------------------------===//
30 #include "SpillPlacement.h"
31 #include "llvm/ADT/ArrayRef.h"
32 #include "llvm/ADT/BitVector.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/SparseSet.h"
35 #include "llvm/CodeGen/EdgeBundles.h"
36 #include "llvm/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineLoopInfo.h"
40 #include "llvm/CodeGen/Passes.h"
41 #include "llvm/Pass.h"
42 #include "llvm/Support/BlockFrequency.h"
50 #define DEBUG_TYPE "spill-code-placement"
52 char SpillPlacement::ID = 0;
54 char &llvm::SpillPlacementID = SpillPlacement::ID;
56 INITIALIZE_PASS_BEGIN(SpillPlacement, DEBUG_TYPE,
57 "Spill Code Placement Analysis", true, true)
58 INITIALIZE_PASS_DEPENDENCY(EdgeBundles)
59 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
60 INITIALIZE_PASS_END(SpillPlacement, DEBUG_TYPE,
61 "Spill Code Placement Analysis", true, true)
63 void SpillPlacement::getAnalysisUsage(AnalysisUsage &AU) const {
65 AU.addRequired<MachineBlockFrequencyInfo>();
66 AU.addRequiredTransitive<EdgeBundles>();
67 AU.addRequiredTransitive<MachineLoopInfo>();
68 MachineFunctionPass::getAnalysisUsage(AU);
71 /// Node - Each edge bundle corresponds to a Hopfield node.
73 /// The node contains precomputed frequency data that only depends on the CFG,
74 /// but Bias and Links are computed each time placeSpills is called.
76 /// The node Value is positive when the variable should be in a register. The
77 /// value can change when linked nodes change, but convergence is very fast
78 /// because all weights are positive.
79 struct SpillPlacement::Node {
80 /// BiasN - Sum of blocks that prefer a spill.
83 /// BiasP - Sum of blocks that prefer a register.
86 /// Value - Output value of this node computed from the Bias and links.
87 /// This is always on of the values {-1, 0, 1}. A positive number means the
88 /// variable should go in a register through this bundle.
91 using LinkVector = SmallVector<std::pair<BlockFrequency, unsigned>, 4>;
93 /// Links - (Weight, BundleNo) for all transparent blocks connecting to other
94 /// bundles. The weights are all positive block frequencies.
97 /// SumLinkWeights - Cached sum of the weights of all links + ThresHold.
98 BlockFrequency SumLinkWeights;
100 /// preferReg - Return true when this node prefers to be in a register.
101 bool preferReg() const {
102 // Undecided nodes (Value==0) go on the stack.
106 /// mustSpill - Return True if this node is so biased that it must spill.
107 bool mustSpill() const {
108 // We must spill if Bias < -sum(weights) or the MustSpill flag was set.
109 // BiasN is saturated when MustSpill is set, make sure this still returns
110 // true when the RHS saturates. Note that SumLinkWeights includes Threshold.
111 return BiasN >= BiasP + SumLinkWeights;
114 /// clear - Reset per-query data, but preserve frequencies that only depend on
116 void clear(const BlockFrequency &Threshold) {
117 BiasN = BiasP = Value = 0;
118 SumLinkWeights = Threshold;
122 /// addLink - Add a link to bundle b with weight w.
123 void addLink(unsigned b, BlockFrequency w) {
124 // Update cached sum.
127 // There can be multiple links to the same bundle, add them up.
128 for (LinkVector::iterator I = Links.begin(), E = Links.end(); I != E; ++I)
129 if (I->second == b) {
133 // This must be the first link to b.
134 Links.push_back(std::make_pair(w, b));
137 /// addBias - Bias this node.
138 void addBias(BlockFrequency freq, BorderConstraint direction) {
149 BiasN = BlockFrequency::getMaxFrequency();
154 /// update - Recompute Value from Bias and Links. Return true when node
155 /// preference changes.
156 bool update(const Node nodes[], const BlockFrequency &Threshold) {
157 // Compute the weighted sum of inputs.
158 BlockFrequency SumN = BiasN;
159 BlockFrequency SumP = BiasP;
160 for (LinkVector::iterator I = Links.begin(), E = Links.end(); I != E; ++I) {
161 if (nodes[I->second].Value == -1)
163 else if (nodes[I->second].Value == 1)
167 // Each weighted sum is going to be less than the total frequency of the
168 // bundle. Ideally, we should simply set Value = sign(SumP - SumN), but we
169 // will add a dead zone around 0 for two reasons:
171 // 1. It avoids arbitrary bias when all links are 0 as is possible during
172 // initial iterations.
173 // 2. It helps tame rounding errors when the links nominally sum to 0.
175 bool Before = preferReg();
176 if (SumN >= SumP + Threshold)
178 else if (SumP >= SumN + Threshold)
182 return Before != preferReg();
185 void getDissentingNeighbors(SparseSet<unsigned> &List,
186 const Node nodes[]) const {
187 for (const auto &Elt : Links) {
188 unsigned n = Elt.second;
189 // Neighbors that already have the same value are not going to
190 // change because of this node changing.
191 if (Value != nodes[n].Value)
197 bool SpillPlacement::runOnMachineFunction(MachineFunction &mf) {
199 bundles = &getAnalysis<EdgeBundles>();
200 loops = &getAnalysis<MachineLoopInfo>();
202 assert(!nodes && "Leaking node array");
203 nodes = new Node[bundles->getNumBundles()];
205 TodoList.setUniverse(bundles->getNumBundles());
207 // Compute total ingoing and outgoing block frequencies for all bundles.
208 BlockFrequencies.resize(mf.getNumBlockIDs());
209 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
210 setThreshold(MBFI->getEntryFreq());
212 unsigned Num = I.getNumber();
213 BlockFrequencies[Num] = MBFI->getBlockFreq(&I);
216 // We never change the function.
220 void SpillPlacement::releaseMemory() {
226 /// activate - mark node n as active if it wasn't already.
227 void SpillPlacement::activate(unsigned n) {
229 if (ActiveNodes->test(n))
232 nodes[n].clear(Threshold);
234 // Very large bundles usually come from big switches, indirect branches,
235 // landing pads, or loops with many 'continue' statements. It is difficult to
236 // allocate registers when so many different blocks are involved.
238 // Give a small negative bias to large bundles such that a substantial
239 // fraction of the connected blocks need to be interested before we consider
240 // expanding the region through the bundle. This helps compile time by
241 // limiting the number of blocks visited and the number of links in the
243 if (bundles->getBlocks(n).size() > 100) {
245 nodes[n].BiasN = (MBFI->getEntryFreq() / 16);
249 /// Set the threshold for a given entry frequency.
251 /// Set the threshold relative to \c Entry. Since the threshold is used as a
252 /// bound on the open interval (-Threshold;Threshold), 1 is the minimum
254 void SpillPlacement::setThreshold(const BlockFrequency &Entry) {
255 // Apparently 2 is a good threshold when Entry==2^14, but we need to scale
256 // it. Divide by 2^13, rounding as appropriate.
257 uint64_t Freq = Entry.getFrequency();
258 uint64_t Scaled = (Freq >> 13) + bool(Freq & (1 << 12));
259 Threshold = std::max(UINT64_C(1), Scaled);
262 /// addConstraints - Compute node biases and weights from a set of constraints.
263 /// Set a bit in NodeMask for each active node.
264 void SpillPlacement::addConstraints(ArrayRef<BlockConstraint> LiveBlocks) {
265 for (ArrayRef<BlockConstraint>::iterator I = LiveBlocks.begin(),
266 E = LiveBlocks.end(); I != E; ++I) {
267 BlockFrequency Freq = BlockFrequencies[I->Number];
270 if (I->Entry != DontCare) {
271 unsigned ib = bundles->getBundle(I->Number, false);
273 nodes[ib].addBias(Freq, I->Entry);
276 // Live-out from block?
277 if (I->Exit != DontCare) {
278 unsigned ob = bundles->getBundle(I->Number, true);
280 nodes[ob].addBias(Freq, I->Exit);
285 /// addPrefSpill - Same as addConstraints(PrefSpill)
286 void SpillPlacement::addPrefSpill(ArrayRef<unsigned> Blocks, bool Strong) {
287 for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end();
289 BlockFrequency Freq = BlockFrequencies[*I];
292 unsigned ib = bundles->getBundle(*I, false);
293 unsigned ob = bundles->getBundle(*I, true);
296 nodes[ib].addBias(Freq, PrefSpill);
297 nodes[ob].addBias(Freq, PrefSpill);
301 void SpillPlacement::addLinks(ArrayRef<unsigned> Links) {
302 for (ArrayRef<unsigned>::iterator I = Links.begin(), E = Links.end(); I != E;
304 unsigned Number = *I;
305 unsigned ib = bundles->getBundle(Number, false);
306 unsigned ob = bundles->getBundle(Number, true);
308 // Ignore self-loops.
313 BlockFrequency Freq = BlockFrequencies[Number];
314 nodes[ib].addLink(ob, Freq);
315 nodes[ob].addLink(ib, Freq);
319 bool SpillPlacement::scanActiveBundles() {
320 RecentPositive.clear();
321 for (unsigned n : ActiveNodes->set_bits()) {
323 // A node that must spill, or a node without any links is not going to
324 // change its value ever again, so exclude it from iterations.
325 if (nodes[n].mustSpill())
327 if (nodes[n].preferReg())
328 RecentPositive.push_back(n);
330 return !RecentPositive.empty();
333 bool SpillPlacement::update(unsigned n) {
334 if (!nodes[n].update(nodes, Threshold))
336 nodes[n].getDissentingNeighbors(TodoList, nodes);
340 /// iterate - Repeatedly update the Hopfield nodes until stability or the
341 /// maximum number of iterations is reached.
342 void SpillPlacement::iterate() {
343 // We do not need to push those node in the todolist.
344 // They are already been proceeded as part of the previous iteration.
345 RecentPositive.clear();
347 // Since the last iteration, the todolist have been augmented by calls
348 // to addConstraints, addLinks, and co.
349 // Update the network energy starting at this new frontier.
350 // The call to ::update will add the nodes that changed into the todolist.
351 unsigned Limit = bundles->getNumBundles() * 10;
352 while(Limit-- > 0 && !TodoList.empty()) {
353 unsigned n = TodoList.pop_back_val();
356 if (nodes[n].preferReg())
357 RecentPositive.push_back(n);
361 void SpillPlacement::prepare(BitVector &RegBundles) {
362 RecentPositive.clear();
364 // Reuse RegBundles as our ActiveNodes vector.
365 ActiveNodes = &RegBundles;
366 ActiveNodes->clear();
367 ActiveNodes->resize(bundles->getNumBundles());
371 SpillPlacement::finish() {
372 assert(ActiveNodes && "Call prepare() first");
374 // Write preferences back to ActiveNodes.
376 for (unsigned n : ActiveNodes->set_bits())
377 if (!nodes[n].preferReg()) {
378 ActiveNodes->reset(n);
381 ActiveNodes = nullptr;