1 //===-- RegAllocSolver.h - Heuristic PBQP Solver for reg alloc --*- 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 // Heuristic PBQP solver for register allocation problems. This solver uses a
11 // graph reduction approach. Nodes of degree 0, 1 and 2 are eliminated with
12 // optimality-preserving rules (see ReductionRules.h). When no low-degree (<3)
13 // nodes are present, a heuristic derived from Brigg's graph coloring approach
16 //===----------------------------------------------------------------------===//
18 #ifndef LLVM_CODEGEN_PBQP_REGALLOCSOLVER_H
19 #define LLVM_CODEGEN_PBQP_REGALLOCSOLVER_H
21 #include "CostAllocator.h"
23 #include "ReductionRules.h"
25 #include "llvm/Support/ErrorHandling.h"
33 /// \brief Metadata to speed allocatability test.
35 /// Keeps track of the number of infinities in each row and column.
36 class MatrixMetadata {
38 MatrixMetadata(const MatrixMetadata&);
39 void operator=(const MatrixMetadata&);
41 MatrixMetadata(const PBQP::Matrix& M)
42 : WorstRow(0), WorstCol(0),
43 UnsafeRows(new bool[M.getRows() - 1]()),
44 UnsafeCols(new bool[M.getCols() - 1]()) {
46 unsigned* ColCounts = new unsigned[M.getCols() - 1]();
48 for (unsigned i = 1; i < M.getRows(); ++i) {
49 unsigned RowCount = 0;
50 for (unsigned j = 1; j < M.getCols(); ++j) {
51 if (M[i][j] == std::numeric_limits<PBQP::PBQPNum>::infinity()) {
54 UnsafeRows[i - 1] = true;
55 UnsafeCols[j - 1] = true;
58 WorstRow = std::max(WorstRow, RowCount);
60 unsigned WorstColCountForCurRow =
61 *std::max_element(ColCounts, ColCounts + M.getCols() - 1);
62 WorstCol = std::max(WorstCol, WorstColCountForCurRow);
71 unsigned getWorstRow() const { return WorstRow; }
72 unsigned getWorstCol() const { return WorstCol; }
73 const bool* getUnsafeRows() const { return UnsafeRows; }
74 const bool* getUnsafeCols() const { return UnsafeCols; }
77 unsigned WorstRow, WorstCol;
84 typedef enum { Unprocessed,
86 ConservativelyAllocatable,
87 NotProvablyAllocatable } ReductionState;
89 NodeMetadata() : RS(Unprocessed), DeniedOpts(0), OptUnsafeEdges(nullptr){}
90 ~NodeMetadata() { delete[] OptUnsafeEdges; }
92 void setup(const Vector& Costs) {
93 NumOpts = Costs.getLength() - 1;
94 OptUnsafeEdges = new unsigned[NumOpts]();
97 ReductionState getReductionState() const { return RS; }
98 void setReductionState(ReductionState RS) { this->RS = RS; }
100 void handleAddEdge(const MatrixMetadata& MD, bool Transpose) {
101 DeniedOpts += Transpose ? MD.getWorstCol() : MD.getWorstRow();
102 const bool* UnsafeOpts =
103 Transpose ? MD.getUnsafeCols() : MD.getUnsafeRows();
104 for (unsigned i = 0; i < NumOpts; ++i)
105 OptUnsafeEdges[i] += UnsafeOpts[i];
108 void handleRemoveEdge(const MatrixMetadata& MD, bool Transpose) {
109 DeniedOpts -= Transpose ? MD.getWorstCol() : MD.getWorstRow();
110 const bool* UnsafeOpts =
111 Transpose ? MD.getUnsafeCols() : MD.getUnsafeRows();
112 for (unsigned i = 0; i < NumOpts; ++i)
113 OptUnsafeEdges[i] -= UnsafeOpts[i];
116 bool isConservativelyAllocatable() const {
117 return (DeniedOpts < NumOpts) ||
118 (std::find(OptUnsafeEdges, OptUnsafeEdges + NumOpts, 0) !=
119 OptUnsafeEdges + NumOpts);
126 unsigned* OptUnsafeEdges;
129 class RegAllocSolverImpl {
131 typedef PBQP::MDMatrix<MatrixMetadata> RAMatrix;
133 typedef PBQP::Vector RawVector;
134 typedef PBQP::Matrix RawMatrix;
135 typedef PBQP::Vector Vector;
136 typedef RAMatrix Matrix;
137 typedef PBQP::PoolCostAllocator<
138 Vector, PBQP::VectorComparator,
139 Matrix, PBQP::MatrixComparator> CostAllocator;
141 typedef PBQP::GraphBase::NodeId NodeId;
142 typedef PBQP::GraphBase::EdgeId EdgeId;
144 typedef RegAlloc::NodeMetadata NodeMetadata;
146 struct EdgeMetadata { };
147 struct GraphMetadata { };
149 typedef PBQP::Graph<RegAllocSolverImpl> Graph;
151 RegAllocSolverImpl(Graph &G) : G(G) {}
157 S = backpropagate(G, reduce());
162 void handleAddNode(NodeId NId) {
163 G.getNodeMetadata(NId).setup(G.getNodeCosts(NId));
165 void handleRemoveNode(NodeId NId) {}
166 void handleSetNodeCosts(NodeId NId, const Vector& newCosts) {}
168 void handleAddEdge(EdgeId EId) {
169 handleReconnectEdge(EId, G.getEdgeNode1Id(EId));
170 handleReconnectEdge(EId, G.getEdgeNode2Id(EId));
173 void handleRemoveEdge(EdgeId EId) {
174 handleDisconnectEdge(EId, G.getEdgeNode1Id(EId));
175 handleDisconnectEdge(EId, G.getEdgeNode2Id(EId));
178 void handleDisconnectEdge(EdgeId EId, NodeId NId) {
179 NodeMetadata& NMd = G.getNodeMetadata(NId);
180 const MatrixMetadata& MMd = G.getEdgeCosts(EId).getMetadata();
181 NMd.handleRemoveEdge(MMd, NId == G.getEdgeNode2Id(EId));
182 if (G.getNodeDegree(NId) == 3) {
183 // This node is becoming optimally reducible.
184 moveToOptimallyReducibleNodes(NId);
185 } else if (NMd.getReductionState() ==
186 NodeMetadata::NotProvablyAllocatable &&
187 NMd.isConservativelyAllocatable()) {
188 // This node just became conservatively allocatable.
189 moveToConservativelyAllocatableNodes(NId);
193 void handleReconnectEdge(EdgeId EId, NodeId NId) {
194 NodeMetadata& NMd = G.getNodeMetadata(NId);
195 const MatrixMetadata& MMd = G.getEdgeCosts(EId).getMetadata();
196 NMd.handleAddEdge(MMd, NId == G.getEdgeNode2Id(EId));
199 void handleSetEdgeCosts(EdgeId EId, const Matrix& NewCosts) {
200 handleRemoveEdge(EId);
202 NodeId N1Id = G.getEdgeNode1Id(EId);
203 NodeId N2Id = G.getEdgeNode2Id(EId);
204 NodeMetadata& N1Md = G.getNodeMetadata(N1Id);
205 NodeMetadata& N2Md = G.getNodeMetadata(N2Id);
206 const MatrixMetadata& MMd = NewCosts.getMetadata();
207 N1Md.handleAddEdge(MMd, N1Id != G.getEdgeNode1Id(EId));
208 N2Md.handleAddEdge(MMd, N2Id != G.getEdgeNode1Id(EId));
213 void removeFromCurrentSet(NodeId NId) {
214 switch (G.getNodeMetadata(NId).getReductionState()) {
215 case NodeMetadata::Unprocessed: break;
216 case NodeMetadata::OptimallyReducible:
217 assert(OptimallyReducibleNodes.find(NId) !=
218 OptimallyReducibleNodes.end() &&
219 "Node not in optimally reducible set.");
220 OptimallyReducibleNodes.erase(NId);
222 case NodeMetadata::ConservativelyAllocatable:
223 assert(ConservativelyAllocatableNodes.find(NId) !=
224 ConservativelyAllocatableNodes.end() &&
225 "Node not in conservatively allocatable set.");
226 ConservativelyAllocatableNodes.erase(NId);
228 case NodeMetadata::NotProvablyAllocatable:
229 assert(NotProvablyAllocatableNodes.find(NId) !=
230 NotProvablyAllocatableNodes.end() &&
231 "Node not in not-provably-allocatable set.");
232 NotProvablyAllocatableNodes.erase(NId);
237 void moveToOptimallyReducibleNodes(NodeId NId) {
238 removeFromCurrentSet(NId);
239 OptimallyReducibleNodes.insert(NId);
240 G.getNodeMetadata(NId).setReductionState(
241 NodeMetadata::OptimallyReducible);
244 void moveToConservativelyAllocatableNodes(NodeId NId) {
245 removeFromCurrentSet(NId);
246 ConservativelyAllocatableNodes.insert(NId);
247 G.getNodeMetadata(NId).setReductionState(
248 NodeMetadata::ConservativelyAllocatable);
251 void moveToNotProvablyAllocatableNodes(NodeId NId) {
252 removeFromCurrentSet(NId);
253 NotProvablyAllocatableNodes.insert(NId);
254 G.getNodeMetadata(NId).setReductionState(
255 NodeMetadata::NotProvablyAllocatable);
260 for (auto NId : G.nodeIds()) {
261 if (G.getNodeDegree(NId) < 3)
262 moveToOptimallyReducibleNodes(NId);
263 else if (G.getNodeMetadata(NId).isConservativelyAllocatable())
264 moveToConservativelyAllocatableNodes(NId);
266 moveToNotProvablyAllocatableNodes(NId);
270 // Compute a reduction order for the graph by iteratively applying PBQP
271 // reduction rules. Locally optimal rules are applied whenever possible (R0,
272 // R1, R2). If no locally-optimal rules apply then any conservatively
273 // allocatable node is reduced. Finally, if no conservatively allocatable
274 // node exists then the node with the lowest spill-cost:degree ratio is
276 std::vector<GraphBase::NodeId> reduce() {
277 assert(!G.empty() && "Cannot reduce empty graph.");
279 typedef GraphBase::NodeId NodeId;
280 std::vector<NodeId> NodeStack;
282 // Consume worklists.
284 if (!OptimallyReducibleNodes.empty()) {
285 NodeSet::iterator NItr = OptimallyReducibleNodes.begin();
287 OptimallyReducibleNodes.erase(NItr);
288 NodeStack.push_back(NId);
289 switch (G.getNodeDegree(NId)) {
298 default: llvm_unreachable("Not an optimally reducible node.");
300 } else if (!ConservativelyAllocatableNodes.empty()) {
301 // Conservatively allocatable nodes will never spill. For now just
302 // take the first node in the set and push it on the stack. When we
303 // start optimizing more heavily for register preferencing, it may
304 // would be better to push nodes with lower 'expected' or worst-case
305 // register costs first (since early nodes are the most
307 NodeSet::iterator NItr = ConservativelyAllocatableNodes.begin();
309 ConservativelyAllocatableNodes.erase(NItr);
310 NodeStack.push_back(NId);
311 G.disconnectAllNeighborsFromNode(NId);
313 } else if (!NotProvablyAllocatableNodes.empty()) {
314 NodeSet::iterator NItr =
315 std::min_element(NotProvablyAllocatableNodes.begin(),
316 NotProvablyAllocatableNodes.end(),
317 SpillCostComparator(G));
319 NotProvablyAllocatableNodes.erase(NItr);
320 NodeStack.push_back(NId);
321 G.disconnectAllNeighborsFromNode(NId);
329 class SpillCostComparator {
331 SpillCostComparator(const Graph& G) : G(G) {}
332 bool operator()(NodeId N1Id, NodeId N2Id) {
333 PBQPNum N1SC = G.getNodeCosts(N1Id)[0] / G.getNodeDegree(N1Id);
334 PBQPNum N2SC = G.getNodeCosts(N2Id)[0] / G.getNodeDegree(N2Id);
342 typedef std::set<NodeId> NodeSet;
343 NodeSet OptimallyReducibleNodes;
344 NodeSet ConservativelyAllocatableNodes;
345 NodeSet NotProvablyAllocatableNodes;
348 typedef Graph<RegAllocSolverImpl> Graph;
350 inline Solution solve(Graph& G) {
353 RegAllocSolverImpl RegAllocSolver(G);
354 return RegAllocSolver.solve();
360 #endif // LLVM_CODEGEN_PBQP_REGALLOCSOLVER_H