xref: /llvm-project/mlir/lib/Conversion/PDLToPDLInterp/RootOrdering.cpp (revision 9eb8e7b176e9fc38c8df86bd927663c6409ac262)
1 //===- RootOrdering.cpp - Optimal root ordering ---------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // An implementation of Edmonds' optimal branching algorithm. This is a
10 // directed analogue of the minimum spanning tree problem for a given root.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "RootOrdering.h"
15 
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include <queue>
20 #include <utility>
21 
22 using namespace mlir;
23 using namespace mlir::pdl_to_pdl_interp;
24 
25 /// Returns the cycle implied by the specified parent relation, starting at the
26 /// given node.
getCycle(const DenseMap<Value,Value> & parents,Value rep)27 static SmallVector<Value> getCycle(const DenseMap<Value, Value> &parents,
28                                    Value rep) {
29   SmallVector<Value> cycle;
30   Value node = rep;
31   do {
32     cycle.push_back(node);
33     node = parents.lookup(node);
34     assert(node && "got an empty value in the cycle");
35   } while (node != rep);
36   return cycle;
37 }
38 
39 /// Contracts the specified cycle in the given graph in-place.
40 /// The parentsCost map specifies, for each node in the cycle, the lowest cost
41 /// among the edges entering that node. Then, the nodes in the cycle C are
42 /// replaced with a single node v_C (the first node in the cycle). All edges
43 /// (u, v) entering the cycle, v \in C, are replaced with a single edge
44 /// (u, v_C) with an appropriately chosen cost, and the selected node v is
45 /// marked in the output map actualTarget[u]. All edges (u, v) leaving the
46 /// cycle, u \in C, are replaced with a single edge (v_C, v), and the selected
47 /// node u is marked in the ouptut map actualSource[v].
contract(RootOrderingGraph & graph,ArrayRef<Value> cycle,const DenseMap<Value,unsigned> & parentDepths,DenseMap<Value,Value> & actualSource,DenseMap<Value,Value> & actualTarget)48 static void contract(RootOrderingGraph &graph, ArrayRef<Value> cycle,
49                      const DenseMap<Value, unsigned> &parentDepths,
50                      DenseMap<Value, Value> &actualSource,
51                      DenseMap<Value, Value> &actualTarget) {
52   Value rep = cycle.front();
53   DenseSet<Value> cycleSet(cycle.begin(), cycle.end());
54 
55   // Now, contract the cycle, marking the actual sources and targets.
56   DenseMap<Value, RootOrderingEntry> repEntries;
57   for (auto outer = graph.begin(), e = graph.end(); outer != e; ++outer) {
58     Value target = outer->first;
59     if (cycleSet.contains(target)) {
60       // Target in the cycle => edges incoming to the cycle or within the cycle.
61       unsigned parentDepth = parentDepths.lookup(target);
62       for (const auto &inner : outer->second) {
63         Value source = inner.first;
64         // Ignore edges within the cycle.
65         if (cycleSet.contains(source))
66           continue;
67 
68         // Edge incoming to the cycle.
69         std::pair<unsigned, unsigned> cost = inner.second.cost;
70         assert(parentDepth <= cost.first && "invalid parent depth");
71 
72         // Subtract the cost of the parent within the cycle from the cost of
73         // the edge incoming to the cycle. This update ensures that the cost
74         // of the minimum-weight spanning arborescence of the entire graph is
75         // the cost of arborescence for the contracted graph plus the cost of
76         // the cycle, no matter which edge in the cycle we choose to drop.
77         cost.first -= parentDepth;
78         auto it = repEntries.find(source);
79         if (it == repEntries.end() || it->second.cost > cost) {
80           actualTarget[source] = target;
81           // Do not bother populating the connector (the connector is only
82           // relevant for the final traversal, not for the optimal branching).
83           repEntries[source].cost = cost;
84         }
85       }
86       // Erase the node in the cycle.
87       graph.erase(outer);
88     } else {
89       // Target not in cycle => edges going away from or unrelated to the cycle.
90       DenseMap<Value, RootOrderingEntry> &entries = outer->second;
91       Value bestSource;
92       std::pair<unsigned, unsigned> bestCost;
93       auto inner = entries.begin(), innerE = entries.end();
94       while (inner != innerE) {
95         Value source = inner->first;
96         if (cycleSet.contains(source)) {
97           // Going-away edge => get its cost and erase it.
98           if (!bestSource || bestCost > inner->second.cost) {
99             bestSource = source;
100             bestCost = inner->second.cost;
101           }
102           entries.erase(inner++);
103         } else {
104           ++inner;
105         }
106       }
107 
108       // There were going-away edges, contract them.
109       if (bestSource) {
110         entries[rep].cost = bestCost;
111         actualSource[target] = bestSource;
112       }
113     }
114   }
115 
116   // Store the edges to the representative.
117   graph[rep] = std::move(repEntries);
118 }
119 
OptimalBranching(RootOrderingGraph graph,Value root)120 OptimalBranching::OptimalBranching(RootOrderingGraph graph, Value root)
121     : graph(std::move(graph)), root(root) {}
122 
solve()123 unsigned OptimalBranching::solve() {
124   // Initialize the parents and total cost.
125   parents.clear();
126   parents[root] = Value();
127   unsigned totalCost = 0;
128 
129   // A map that stores the cost of the optimal local choice for each node
130   // in a directed cycle. This map is cleared every time we seed the search.
131   DenseMap<Value, unsigned> parentDepths;
132   parentDepths.reserve(graph.size());
133 
134   // Determine if the optimal local choice results in an acyclic graph. This is
135   // done by computing the optimal local choice and traversing up the computed
136   // parents. On success, `parents` will contain the parent of each node.
137   for (const auto &outer : graph) {
138     Value node = outer.first;
139     if (parents.count(node)) // already visited
140       continue;
141 
142     // Follow the trail of best sources until we reach an already visited node.
143     // The code will assert if we cannot reach an already visited node, i.e.,
144     // the graph is not strongly connected.
145     parentDepths.clear();
146     do {
147       auto it = graph.find(node);
148       assert(it != graph.end() && "the graph is not strongly connected");
149 
150       // Find the best local parent, taking into account both the depth and the
151       // tie breaking rules.
152       Value &bestSource = parents[node];
153       std::pair<unsigned, unsigned> bestCost;
154       for (const auto &inner : it->second) {
155         const RootOrderingEntry &entry = inner.second;
156         if (!bestSource /* initial */ || bestCost > entry.cost) {
157           bestSource = inner.first;
158           bestCost = entry.cost;
159         }
160       }
161       assert(bestSource && "the graph is not strongly connected");
162       parentDepths[node] = bestCost.first;
163       node = bestSource;
164       totalCost += bestCost.first;
165     } while (!parents.count(node));
166 
167     // If we reached a non-root node, we have a cycle.
168     if (parentDepths.count(node)) {
169       // Determine the cycle starting at the representative node.
170       SmallVector<Value> cycle = getCycle(parents, node);
171 
172       // The following maps disambiguate the source / target of the edges
173       // going out of / into the cycle.
174       DenseMap<Value, Value> actualSource, actualTarget;
175 
176       // Contract the cycle and recurse.
177       contract(graph, cycle, parentDepths, actualSource, actualTarget);
178       totalCost = solve();
179 
180       // Redirect the going-away edges.
181       for (auto &p : parents)
182         if (p.second == node)
183           // The parent is the node representating the cycle; replace it
184           // with the actual (best) source in the cycle.
185           p.second = actualSource.lookup(p.first);
186 
187       // Redirect the unique incoming edge and copy the cycle.
188       Value parent = parents.lookup(node);
189       Value entry = actualTarget.lookup(parent);
190       cycle.push_back(node); // complete the cycle
191       for (size_t i = 0, e = cycle.size() - 1; i < e; ++i) {
192         totalCost += parentDepths.lookup(cycle[i]);
193         if (cycle[i] == entry)
194           parents[cycle[i]] = parent; // break the cycle
195         else
196           parents[cycle[i]] = cycle[i + 1];
197       }
198 
199       // `parents` has a complete solution.
200       break;
201     }
202   }
203 
204   return totalCost;
205 }
206 
207 OptimalBranching::EdgeList
preOrderTraversal(ArrayRef<Value> nodes) const208 OptimalBranching::preOrderTraversal(ArrayRef<Value> nodes) const {
209   // Invert the parent mapping.
210   DenseMap<Value, std::vector<Value>> children;
211   for (Value node : nodes) {
212     if (node != root) {
213       Value parent = parents.lookup(node);
214       assert(parent && "invalid parent");
215       children[parent].push_back(node);
216     }
217   }
218 
219   // The result which simultaneously acts as a queue.
220   EdgeList result;
221   result.reserve(nodes.size());
222   result.emplace_back(root, Value());
223 
224   // Perform a BFS, pushing into the queue.
225   for (size_t i = 0; i < result.size(); ++i) {
226     Value node = result[i].first;
227     for (Value child : children[node])
228       result.emplace_back(child, node);
229   }
230 
231   return result;
232 }
233