1 //===-- Briggs.h --- Briggs Heuristic for PBQP ------------------*- C++ -*-===//
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 class implements the Briggs test for "allocability" of nodes in a
11 // PBQP graph representing a register allocation problem. Nodes which can be
12 // proven allocable (by a safe and relatively accurate test) are removed from
13 // the PBQP graph first. If no provably allocable node is present in the graph
14 // then the node with the minimal spill-cost to degree ratio is removed.
16 //===----------------------------------------------------------------------===//
18 #ifndef LLVM_CODEGEN_PBQP_HEURISTICS_BRIGGS_H
19 #define LLVM_CODEGEN_PBQP_HEURISTICS_BRIGGS_H
21 #include "llvm/Support/Compiler.h"
22 #include "../HeuristicSolver.h"
23 #include "../HeuristicBase.h"
29 namespace Heuristics {
31 /// \brief PBQP Heuristic which applies an allocability test based on
34 /// This heuristic assumes that the elements of cost vectors in the PBQP
35 /// problem represent storage options, with the first being the spill
36 /// option and subsequent elements representing legal registers for the
37 /// corresponding node. Edge cost matrices are likewise assumed to represent
38 /// register constraints.
39 /// If one or more nodes can be proven allocable by this heuristic (by
40 /// inspection of their constraint matrices) then the allocable node of
41 /// highest degree is selected for the next reduction and pushed to the
42 /// solver stack. If no nodes can be proven allocable then the node with
43 /// the lowest estimated spill cost is selected and push to the solver stack
46 /// This implementation is built on top of HeuristicBase.
47 class Briggs : public HeuristicBase<Briggs> {
50 class LinkDegreeComparator {
52 LinkDegreeComparator(HeuristicSolverImpl<Briggs> &s) : s(&s) {}
53 bool operator()(Graph::NodeItr n1Itr, Graph::NodeItr n2Itr) const {
54 if (s->getSolverDegree(n1Itr) > s->getSolverDegree(n2Itr))
56 if (s->getSolverDegree(n1Itr) < s->getSolverDegree(n2Itr))
58 return (&*n1Itr < &*n2Itr);
61 HeuristicSolverImpl<Briggs> *s;
64 class SpillCostComparator {
66 SpillCostComparator(HeuristicSolverImpl<Briggs> &s)
67 : s(&s), g(&s.getGraph()) {}
68 bool operator()(Graph::NodeItr n1Itr, Graph::NodeItr n2Itr) const {
69 PBQPNum cost1 = g->getNodeCosts(n1Itr)[0] / s->getSolverDegree(n1Itr),
70 cost2 = g->getNodeCosts(n2Itr)[0] / s->getSolverDegree(n2Itr);
75 return (&*n1Itr < &*n2Itr);
79 HeuristicSolverImpl<Briggs> *s;
83 typedef std::list<Graph::NodeItr> RNAllocableList;
84 typedef RNAllocableList::iterator RNAllocableListItr;
86 typedef std::list<Graph::NodeItr> RNUnallocableList;
87 typedef RNUnallocableList::iterator RNUnallocableListItr;
92 typedef std::vector<unsigned> UnsafeDegreesArray;
93 bool isHeuristic, isAllocable, isInitialized;
94 unsigned numDenied, numSafe;
95 UnsafeDegreesArray unsafeDegrees;
96 RNAllocableListItr rnaItr;
97 RNUnallocableListItr rnuItr;
100 : isHeuristic(false), isAllocable(false), isInitialized(false),
101 numDenied(0), numSafe(0) { }
105 typedef std::vector<unsigned> UnsafeArray;
106 unsigned worst, reverseWorst;
107 UnsafeArray unsafe, reverseUnsafe;
110 EdgeData() : worst(0), reverseWorst(0), isUpToDate(false) {}
113 /// \brief Construct an instance of the Briggs heuristic.
114 /// @param solver A reference to the solver which is using this heuristic.
115 Briggs(HeuristicSolverImpl<Briggs> &solver) :
116 HeuristicBase<Briggs>(solver) {}
118 /// \brief Determine whether a node should be reduced using optimal
120 /// @param nItr Node iterator to be considered.
121 /// @return True if the given node should be optimally reduced, false
124 /// Selects nodes of degree 0, 1 or 2 for optimal reduction, with one
125 /// exception. Nodes whose spill cost (element 0 of their cost vector) is
126 /// infinite are checked for allocability first. Allocable nodes may be
127 /// optimally reduced, but nodes whose allocability cannot be proven are
128 /// selected for heuristic reduction instead.
129 bool shouldOptimallyReduce(Graph::NodeItr nItr) {
130 if (getSolver().getSolverDegree(nItr) < 3) {
137 /// \brief Add a node to the heuristic reduce list.
138 /// @param nItr Node iterator to add to the heuristic reduce list.
139 void addToHeuristicReduceList(Graph::NodeItr nItr) {
140 NodeData &nd = getHeuristicNodeData(nItr);
141 initializeNode(nItr);
142 nd.isHeuristic = true;
143 if (nd.isAllocable) {
144 nd.rnaItr = rnAllocableList.insert(rnAllocableList.end(), nItr);
146 nd.rnuItr = rnUnallocableList.insert(rnUnallocableList.end(), nItr);
150 /// \brief Heuristically reduce one of the nodes in the heuristic
152 /// @return True if a reduction takes place, false if the heuristic reduce
155 /// If the list of allocable nodes is non-empty a node is selected
156 /// from it and pushed to the stack. Otherwise if the non-allocable list
157 /// is non-empty a node is selected from it and pushed to the stack.
158 /// If both lists are empty the method simply returns false with no action
160 bool heuristicReduce() {
161 if (!rnAllocableList.empty()) {
162 RNAllocableListItr rnaItr =
163 min_element(rnAllocableList.begin(), rnAllocableList.end(),
164 LinkDegreeComparator(getSolver()));
165 Graph::NodeItr nItr = *rnaItr;
166 rnAllocableList.erase(rnaItr);
167 handleRemoveNode(nItr);
168 getSolver().pushToStack(nItr);
170 } else if (!rnUnallocableList.empty()) {
171 RNUnallocableListItr rnuItr =
172 min_element(rnUnallocableList.begin(), rnUnallocableList.end(),
173 SpillCostComparator(getSolver()));
174 Graph::NodeItr nItr = *rnuItr;
175 rnUnallocableList.erase(rnuItr);
176 handleRemoveNode(nItr);
177 getSolver().pushToStack(nItr);
184 /// \brief Prepare a change in the costs on the given edge.
185 /// @param eItr Edge iterator.
186 void preUpdateEdgeCosts(Graph::EdgeItr eItr) {
187 Graph &g = getGraph();
188 Graph::NodeItr n1Itr = g.getEdgeNode1(eItr),
189 n2Itr = g.getEdgeNode2(eItr);
190 NodeData &n1 = getHeuristicNodeData(n1Itr),
191 &n2 = getHeuristicNodeData(n2Itr);
194 subtractEdgeContributions(eItr, getGraph().getEdgeNode1(eItr));
196 subtractEdgeContributions(eItr, getGraph().getEdgeNode2(eItr));
198 EdgeData &ed = getHeuristicEdgeData(eItr);
199 ed.isUpToDate = false;
202 /// \brief Handle the change in the costs on the given edge.
203 /// @param eItr Edge iterator.
204 void postUpdateEdgeCosts(Graph::EdgeItr eItr) {
205 // This is effectively the same as adding a new edge now, since
206 // we've factored out the costs of the old one.
210 /// \brief Handle the addition of a new edge into the PBQP graph.
211 /// @param eItr Edge iterator for the added edge.
213 /// Updates allocability of any nodes connected by this edge which are
214 /// being managed by the heuristic. If allocability changes they are
215 /// moved to the appropriate list.
216 void handleAddEdge(Graph::EdgeItr eItr) {
217 Graph &g = getGraph();
218 Graph::NodeItr n1Itr = g.getEdgeNode1(eItr),
219 n2Itr = g.getEdgeNode2(eItr);
220 NodeData &n1 = getHeuristicNodeData(n1Itr),
221 &n2 = getHeuristicNodeData(n2Itr);
223 // If neither node is managed by the heuristic there's nothing to be
225 if (!n1.isHeuristic && !n2.isHeuristic)
228 // Ok - we need to update at least one node.
229 computeEdgeContributions(eItr);
231 // Update node 1 if it's managed by the heuristic.
232 if (n1.isHeuristic) {
233 bool n1WasAllocable = n1.isAllocable;
234 addEdgeContributions(eItr, n1Itr);
235 updateAllocability(n1Itr);
236 if (n1WasAllocable && !n1.isAllocable) {
237 rnAllocableList.erase(n1.rnaItr);
239 rnUnallocableList.insert(rnUnallocableList.end(), n1Itr);
243 // Likewise for node 2.
244 if (n2.isHeuristic) {
245 bool n2WasAllocable = n2.isAllocable;
246 addEdgeContributions(eItr, n2Itr);
247 updateAllocability(n2Itr);
248 if (n2WasAllocable && !n2.isAllocable) {
249 rnAllocableList.erase(n2.rnaItr);
251 rnUnallocableList.insert(rnUnallocableList.end(), n2Itr);
256 /// \brief Handle disconnection of an edge from a node.
257 /// @param eItr Edge iterator for edge being disconnected.
258 /// @param nItr Node iterator for the node being disconnected from.
260 /// Updates allocability of the given node and, if appropriate, moves the
261 /// node to a new list.
262 void handleRemoveEdge(Graph::EdgeItr eItr, Graph::NodeItr nItr) {
263 NodeData &nd = getHeuristicNodeData(nItr);
265 // If the node is not managed by the heuristic there's nothing to be
270 EdgeData &ed ATTRIBUTE_UNUSED = getHeuristicEdgeData(eItr);
272 assert(ed.isUpToDate && "Edge data is not up to date.");
275 bool ndWasAllocable = nd.isAllocable;
276 subtractEdgeContributions(eItr, nItr);
277 updateAllocability(nItr);
279 // If the node has gone optimal...
280 if (shouldOptimallyReduce(nItr)) {
281 nd.isHeuristic = false;
282 addToOptimalReduceList(nItr);
283 if (ndWasAllocable) {
284 rnAllocableList.erase(nd.rnaItr);
286 rnUnallocableList.erase(nd.rnuItr);
289 // Node didn't go optimal, but we might have to move it
290 // from "unallocable" to "allocable".
291 if (!ndWasAllocable && nd.isAllocable) {
292 rnUnallocableList.erase(nd.rnuItr);
293 nd.rnaItr = rnAllocableList.insert(rnAllocableList.end(), nItr);
300 NodeData& getHeuristicNodeData(Graph::NodeItr nItr) {
301 return getSolver().getHeuristicNodeData(nItr);
304 EdgeData& getHeuristicEdgeData(Graph::EdgeItr eItr) {
305 return getSolver().getHeuristicEdgeData(eItr);
308 // Work out what this edge will contribute to the allocability of the
309 // nodes connected to it.
310 void computeEdgeContributions(Graph::EdgeItr eItr) {
311 EdgeData &ed = getHeuristicEdgeData(eItr);
314 return; // Edge data is already up to date.
316 Matrix &eCosts = getGraph().getEdgeCosts(eItr);
318 unsigned numRegs = eCosts.getRows() - 1,
319 numReverseRegs = eCosts.getCols() - 1;
321 std::vector<unsigned> rowInfCounts(numRegs, 0),
322 colInfCounts(numReverseRegs, 0);
327 ed.unsafe.resize(numRegs, 0);
328 ed.reverseUnsafe.clear();
329 ed.reverseUnsafe.resize(numReverseRegs, 0);
331 for (unsigned i = 0; i < numRegs; ++i) {
332 for (unsigned j = 0; j < numReverseRegs; ++j) {
333 if (eCosts[i + 1][j + 1] ==
334 std::numeric_limits<PBQPNum>::infinity()) {
336 ed.reverseUnsafe[j] = 1;
340 if (colInfCounts[j] > ed.worst) {
341 ed.worst = colInfCounts[j];
344 if (rowInfCounts[i] > ed.reverseWorst) {
345 ed.reverseWorst = rowInfCounts[i];
351 ed.isUpToDate = true;
354 // Add the contributions of the given edge to the given node's
355 // numDenied and safe members. No action is taken other than to update
356 // these member values. Once updated these numbers can be used by clients
357 // to update the node's allocability.
358 void addEdgeContributions(Graph::EdgeItr eItr, Graph::NodeItr nItr) {
359 EdgeData &ed = getHeuristicEdgeData(eItr);
361 assert(ed.isUpToDate && "Using out-of-date edge numbers.");
363 NodeData &nd = getHeuristicNodeData(nItr);
364 unsigned numRegs = getGraph().getNodeCosts(nItr).getLength() - 1;
366 bool nIsNode1 = nItr == getGraph().getEdgeNode1(eItr);
367 EdgeData::UnsafeArray &unsafe =
368 nIsNode1 ? ed.unsafe : ed.reverseUnsafe;
369 nd.numDenied += nIsNode1 ? ed.worst : ed.reverseWorst;
371 for (unsigned r = 0; r < numRegs; ++r) {
373 if (nd.unsafeDegrees[r]==0) {
376 ++nd.unsafeDegrees[r];
381 // Subtract the contributions of the given edge to the given node's
382 // numDenied and safe members. No action is taken other than to update
383 // these member values. Once updated these numbers can be used by clients
384 // to update the node's allocability.
385 void subtractEdgeContributions(Graph::EdgeItr eItr, Graph::NodeItr nItr) {
386 EdgeData &ed = getHeuristicEdgeData(eItr);
388 assert(ed.isUpToDate && "Using out-of-date edge numbers.");
390 NodeData &nd = getHeuristicNodeData(nItr);
391 unsigned numRegs = getGraph().getNodeCosts(nItr).getLength() - 1;
393 bool nIsNode1 = nItr == getGraph().getEdgeNode1(eItr);
394 EdgeData::UnsafeArray &unsafe =
395 nIsNode1 ? ed.unsafe : ed.reverseUnsafe;
396 nd.numDenied -= nIsNode1 ? ed.worst : ed.reverseWorst;
398 for (unsigned r = 0; r < numRegs; ++r) {
400 if (nd.unsafeDegrees[r] == 1) {
403 --nd.unsafeDegrees[r];
408 void updateAllocability(Graph::NodeItr nItr) {
409 NodeData &nd = getHeuristicNodeData(nItr);
410 unsigned numRegs = getGraph().getNodeCosts(nItr).getLength() - 1;
411 nd.isAllocable = nd.numDenied < numRegs || nd.numSafe > 0;
414 void initializeNode(Graph::NodeItr nItr) {
415 NodeData &nd = getHeuristicNodeData(nItr);
417 if (nd.isInitialized)
418 return; // Node data is already up to date.
420 unsigned numRegs = getGraph().getNodeCosts(nItr).getLength() - 1;
423 nd.numSafe = numRegs;
424 nd.unsafeDegrees.resize(numRegs, 0);
426 typedef HeuristicSolverImpl<Briggs>::SolverEdgeItr SolverEdgeItr;
428 for (SolverEdgeItr aeItr = getSolver().solverEdgesBegin(nItr),
429 aeEnd = getSolver().solverEdgesEnd(nItr);
430 aeItr != aeEnd; ++aeItr) {
432 Graph::EdgeItr eItr = *aeItr;
433 computeEdgeContributions(eItr);
434 addEdgeContributions(eItr, nItr);
437 updateAllocability(nItr);
438 nd.isInitialized = true;
441 void handleRemoveNode(Graph::NodeItr xnItr) {
442 typedef HeuristicSolverImpl<Briggs>::SolverEdgeItr SolverEdgeItr;
443 std::vector<Graph::EdgeItr> edgesToRemove;
444 for (SolverEdgeItr aeItr = getSolver().solverEdgesBegin(xnItr),
445 aeEnd = getSolver().solverEdgesEnd(xnItr);
446 aeItr != aeEnd; ++aeItr) {
447 Graph::NodeItr ynItr = getGraph().getEdgeOtherNode(*aeItr, xnItr);
448 handleRemoveEdge(*aeItr, ynItr);
449 edgesToRemove.push_back(*aeItr);
451 while (!edgesToRemove.empty()) {
452 getSolver().removeSolverEdge(edgesToRemove.back());
453 edgesToRemove.pop_back();
457 RNAllocableList rnAllocableList;
458 RNUnallocableList rnUnallocableList;
465 #endif // LLVM_CODEGEN_PBQP_HEURISTICS_BRIGGS_H