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 #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"
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)
52 char &llvm::SpillPlacementID = SpillPlacement::ID;
54 void SpillPlacement::getAnalysisUsage(AnalysisUsage &AU) const {
56 AU.addRequired<MachineBlockFrequencyInfo>();
57 AU.addRequiredTransitive<EdgeBundles>();
58 AU.addRequiredTransitive<MachineLoopInfo>();
59 MachineFunctionPass::getAnalysisUsage(AU);
62 /// Node - Each edge bundle corresponds to a Hopfield node.
64 /// The node contains precomputed frequency data that only depends on the CFG,
65 /// but Bias and Links are computed each time placeSpills is called.
67 /// The node Value is positive when the variable should be in a register. The
68 /// value can change when linked nodes change, but convergence is very fast
69 /// because all weights are positive.
71 struct SpillPlacement::Node {
72 /// Scale - Inverse block frequency feeding into[0] or out of[1] the bundle.
73 /// Ideally, these two numbers should be identical, but inaccuracies in the
74 /// block frequency estimates means that we need to normalize ingoing and
75 /// outgoing frequencies separately so they are commensurate.
78 /// Bias - Normalized contributions from non-transparent blocks.
79 /// A bundle connected to a MustSpill block has a huge negative bias,
80 /// otherwise it is a number in the range [-2;2].
83 /// Value - Output value of this node computed from the Bias and links.
84 /// This is always in the range [-1;1]. A positive number means the variable
85 /// should go in a register through this bundle.
88 typedef SmallVector<std::pair<float, unsigned>, 4> LinkVector;
90 /// Links - (Weight, BundleNo) for all transparent blocks connecting to other
91 /// bundles. The weights are all positive and add up to at most 2, weights
92 /// from ingoing and outgoing nodes separately add up to a most 1. The weight
93 /// sum can be less than 2 when the variable is not live into / out of some
94 /// connected basic blocks.
97 /// preferReg - Return true when this node prefers to be in a register.
98 bool preferReg() const {
99 // Undecided nodes (Value==0) go on the stack.
103 /// mustSpill - Return True if this node is so biased that it must spill.
104 bool mustSpill() const {
105 // Actually, we must spill if Bias < sum(weights).
106 // It may be worth it to compute the weight sum here?
110 /// Node - Create a blank Node.
112 Scale[0] = Scale[1] = 0;
115 /// clear - Reset per-query data, but preserve frequencies that only depend on
122 /// addLink - Add a link to bundle b with weight w.
123 /// out=0 for an ingoing link, and 1 for an outgoing link.
124 void addLink(unsigned b, float w, bool out) {
125 // Normalize w relative to all connected blocks from that direction.
128 // There can be multiple links to the same bundle, add them up.
129 for (LinkVector::iterator I = Links.begin(), E = Links.end(); I != E; ++I)
130 if (I->second == b) {
134 // This must be the first link to b.
135 Links.push_back(std::make_pair(w, b));
138 /// addBias - Bias this node from an ingoing[0] or outgoing[1] link.
139 /// Return the change to the total number of positive biases.
140 void addBias(float w, bool out) {
141 // Normalize w relative to all connected blocks from that direction.
146 /// update - Recompute Value from Bias and Links. Return true when node
147 /// preference changes.
148 bool update(const Node nodes[]) {
149 // Compute the weighted sum of inputs.
151 for (LinkVector::iterator I = Links.begin(), E = Links.end(); I != E; ++I)
152 Sum += I->first * nodes[I->second].Value;
154 // The weighted sum is going to be in the range [-2;2]. Ideally, we should
155 // simply set Value = sign(Sum), but we will add a dead zone around 0 for
157 // 1. It avoids arbitrary bias when all links are 0 as is possible during
158 // initial iterations.
159 // 2. It helps tame rounding errors when the links nominally sum to 0.
160 const float Thres = 1e-4f;
161 bool Before = preferReg();
164 else if (Sum > Thres)
168 return Before != preferReg();
172 bool SpillPlacement::runOnMachineFunction(MachineFunction &mf) {
174 bundles = &getAnalysis<EdgeBundles>();
175 loops = &getAnalysis<MachineLoopInfo>();
177 assert(!nodes && "Leaking node array");
178 nodes = new Node[bundles->getNumBundles()];
180 // Compute total ingoing and outgoing block frequencies for all bundles.
181 BlockFrequency.resize(mf.getNumBlockIDs());
182 MachineBlockFrequencyInfo &MBFI = getAnalysis<MachineBlockFrequencyInfo>();
183 float EntryFreq = BlockFrequency::getEntryFrequency();
184 for (MachineFunction::iterator I = mf.begin(), E = mf.end(); I != E; ++I) {
185 float Freq = MBFI.getBlockFreq(I).getFrequency() / EntryFreq;
186 unsigned Num = I->getNumber();
187 BlockFrequency[Num] = Freq;
188 nodes[bundles->getBundle(Num, 1)].Scale[0] += Freq;
189 nodes[bundles->getBundle(Num, 0)].Scale[1] += Freq;
192 // Scales are reciprocal frequencies.
193 for (unsigned i = 0, e = bundles->getNumBundles(); i != e; ++i)
194 for (unsigned d = 0; d != 2; ++d)
195 if (nodes[i].Scale[d] > 0)
196 nodes[i].Scale[d] = 1 / nodes[i].Scale[d];
198 // We never change the function.
202 void SpillPlacement::releaseMemory() {
207 /// activate - mark node n as active if it wasn't already.
208 void SpillPlacement::activate(unsigned n) {
209 if (ActiveNodes->test(n))
214 // Very large bundles usually come from big switches, indirect branches,
215 // landing pads, or loops with many 'continue' statements. It is difficult to
216 // allocate registers when so many different blocks are involved.
218 // Give a small negative bias to large bundles such that 1/32 of the
219 // connected blocks need to be interested before we consider expanding the
220 // region through the bundle. This helps compile time by limiting the number
221 // of blocks visited and the number of links in the Hopfield network.
222 if (bundles->getBlocks(n).size() > 100)
223 nodes[n].Bias = -0.0625f;
227 /// addConstraints - Compute node biases and weights from a set of constraints.
228 /// Set a bit in NodeMask for each active node.
229 void SpillPlacement::addConstraints(ArrayRef<BlockConstraint> LiveBlocks) {
230 for (ArrayRef<BlockConstraint>::iterator I = LiveBlocks.begin(),
231 E = LiveBlocks.end(); I != E; ++I) {
232 float Freq = getBlockFrequency(I->Number);
233 const float Bias[] = {
238 -HUGE_VALF // MustSpill
242 if (I->Entry != DontCare) {
243 unsigned ib = bundles->getBundle(I->Number, 0);
245 nodes[ib].addBias(Freq * Bias[I->Entry], 1);
248 // Live-out from block?
249 if (I->Exit != DontCare) {
250 unsigned ob = bundles->getBundle(I->Number, 1);
252 nodes[ob].addBias(Freq * Bias[I->Exit], 0);
257 /// addPrefSpill - Same as addConstraints(PrefSpill)
258 void SpillPlacement::addPrefSpill(ArrayRef<unsigned> Blocks, bool Strong) {
259 for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end();
261 float Freq = getBlockFrequency(*I);
264 unsigned ib = bundles->getBundle(*I, 0);
265 unsigned ob = bundles->getBundle(*I, 1);
268 nodes[ib].addBias(-Freq, 1);
269 nodes[ob].addBias(-Freq, 0);
273 void SpillPlacement::addLinks(ArrayRef<unsigned> Links) {
274 for (ArrayRef<unsigned>::iterator I = Links.begin(), E = Links.end(); I != E;
276 unsigned Number = *I;
277 unsigned ib = bundles->getBundle(Number, 0);
278 unsigned ob = bundles->getBundle(Number, 1);
280 // Ignore self-loops.
285 if (nodes[ib].Links.empty() && !nodes[ib].mustSpill())
286 Linked.push_back(ib);
287 if (nodes[ob].Links.empty() && !nodes[ob].mustSpill())
288 Linked.push_back(ob);
289 float Freq = getBlockFrequency(Number);
290 nodes[ib].addLink(ob, Freq, 1);
291 nodes[ob].addLink(ib, Freq, 0);
295 bool SpillPlacement::scanActiveBundles() {
297 RecentPositive.clear();
298 for (int n = ActiveNodes->find_first(); n>=0; n = ActiveNodes->find_next(n)) {
299 nodes[n].update(nodes);
300 // A node that must spill, or a node without any links is not going to
301 // change its value ever again, so exclude it from iterations.
302 if (nodes[n].mustSpill())
304 if (!nodes[n].Links.empty())
306 if (nodes[n].preferReg())
307 RecentPositive.push_back(n);
309 return !RecentPositive.empty();
312 /// iterate - Repeatedly update the Hopfield nodes until stability or the
313 /// maximum number of iterations is reached.
314 /// @param Linked - Numbers of linked nodes that need updating.
315 void SpillPlacement::iterate() {
316 // First update the recently positive nodes. They have likely received new
317 // negative bias that will turn them off.
318 while (!RecentPositive.empty())
319 nodes[RecentPositive.pop_back_val()].update(nodes);
324 // Run up to 10 iterations. The edge bundle numbering is closely related to
325 // basic block numbering, so there is a strong tendency towards chains of
326 // linked nodes with sequential numbers. By scanning the linked nodes
327 // backwards and forwards, we make it very likely that a single node can
328 // affect the entire network in a single iteration. That means very fast
329 // convergence, usually in a single iteration.
330 for (unsigned iteration = 0; iteration != 10; ++iteration) {
331 // Scan backwards, skipping the last node which was just updated.
332 bool Changed = false;
333 for (SmallVectorImpl<unsigned>::const_reverse_iterator I =
334 llvm::next(Linked.rbegin()), E = Linked.rend(); I != E; ++I) {
336 if (nodes[n].update(nodes)) {
338 if (nodes[n].preferReg())
339 RecentPositive.push_back(n);
342 if (!Changed || !RecentPositive.empty())
345 // Scan forwards, skipping the first node which was just updated.
347 for (SmallVectorImpl<unsigned>::const_iterator I =
348 llvm::next(Linked.begin()), E = Linked.end(); I != E; ++I) {
350 if (nodes[n].update(nodes)) {
352 if (nodes[n].preferReg())
353 RecentPositive.push_back(n);
356 if (!Changed || !RecentPositive.empty())
361 void SpillPlacement::prepare(BitVector &RegBundles) {
363 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 (int n = ActiveNodes->find_first(); n>=0; n = ActiveNodes->find_next(n))
377 if (!nodes[n].preferReg()) {
378 ActiveNodes->reset(n);