[RDF] Use hash-based containers, cache extra information

This improves performance.

Author: Krzysztof Parzyszek

llvm/include/llvm/CodeGen/RDFLiveness.h

@@ -18,6 +18,8 @@
 #include "llvm/MC/LaneBitmask.h"
 #include <map>
 #include <set>
+#include <unordered_map>
+#include <unordered_set>
 #include <utility>
 
 namespace llvm {
@@ -28,6 +30,30 @@ class MachineDominatorTree;
 class MachineRegisterInfo;
 class TargetRegisterInfo;
 
+} // namespace llvm
+
+namespace llvm {
+namespace rdf {
+namespace detail {
+
+using NodeRef = std::pair<NodeId, LaneBitmask>;
+
+} // namespace detail
+} // namespace rdf
+} // namespace llvm
+
+namespace std {
+
+template <> struct hash<llvm::rdf::detail::NodeRef> {
+  std::size_t operator()(llvm::rdf::detail::NodeRef R) const {
+    return std::hash<llvm::rdf::NodeId>{}(R.first) ^
+           std::hash<llvm::LaneBitmask::Type>{}(R.second.getAsInteger());
+  }
+};
+
+} // namespace std
+
+namespace llvm {
 namespace rdf {
 
   struct Liveness {
@@ -46,10 +72,9 @@ namespace rdf {
       std::map<MachineBasicBlock*,RegisterAggr> Map;
     };
 
-    using NodeRef = std::pair<NodeId, LaneBitmask>;
-    using NodeRefSet = std::set<NodeRef>;
-    // RegisterId in RefMap must be normalized.
-    using RefMap = std::map<RegisterId, NodeRefSet>;
+    using NodeRef = detail::NodeRef;
+    using NodeRefSet = std::unordered_set<NodeRef>;
+    using RefMap = std::unordered_map<RegisterId, NodeRefSet>;
 
     Liveness(MachineRegisterInfo &mri, const DataFlowGraph &g)
         : DFG(g), TRI(g.getTRI()), PRI(g.getPRI()), MDT(g.getDT()),
@@ -110,15 +135,14 @@ namespace rdf {
     // Cache of mapping from node ids (for RefNodes) to the containing
     // basic blocks. Not computing it each time for each node reduces
     // the liveness calculation time by a large fraction.
-    using NodeBlockMap = DenseMap<NodeId, MachineBasicBlock *>;
-    NodeBlockMap NBMap;
+    DenseMap<NodeId, MachineBasicBlock *> NBMap;
 
     // Phi information:
     //
     // RealUseMap
     // map: NodeId -> (map: RegisterId -> NodeRefSet)
     //      phi id -> (map: register -> set of reached non-phi uses)
-    std::map<NodeId, RefMap> RealUseMap;
+    DenseMap<NodeId, RefMap> RealUseMap;
 
     // Inverse iterated dominance frontier.
     std::map<MachineBasicBlock*,std::set<MachineBasicBlock*>> IIDF;

llvm/include/llvm/CodeGen/RDFRegisters.h

@@ -91,6 +91,11 @@ namespace rdf {
     bool operator< (const RegisterRef &RR) const {
       return Reg < RR.Reg || (Reg == RR.Reg && Mask < RR.Mask);
     }
+
+    size_t hash() const {
+      return std::hash<RegisterId>{}(Reg) ^
+             std::hash<LaneBitmask::Type>{}(Mask.getAsInteger());
+    }
   };
 
 
@@ -171,10 +176,15 @@ namespace rdf {
         : Units(pri.getTRI().getNumRegUnits()), PRI(pri) {}
     RegisterAggr(const RegisterAggr &RG) = default;
 
+    unsigned count() const { return Units.count(); }
     bool empty() const { return Units.none(); }
     bool hasAliasOf(RegisterRef RR) const;
     bool hasCoverOf(RegisterRef RR) const;
 
+    bool operator==(const RegisterAggr &A) const {
+      return DenseMapInfo<BitVector>::isEqual(Units, A.Units);
+    }
+
     static bool isCoverOf(RegisterRef RA, RegisterRef RB,
                           const PhysicalRegisterInfo &PRI) {
       return RegisterAggr(PRI).insert(RA).hasCoverOf(RB);
@@ -191,6 +201,10 @@ namespace rdf {
     RegisterRef clearIn(RegisterRef RR) const;
     RegisterRef makeRegRef() const;
 
+    size_t hash() const {
+      return DenseMapInfo<BitVector>::getHashValue(Units);
+    }
+
     void print(raw_ostream &OS) const;
 
     struct rr_iterator {
@@ -244,9 +258,26 @@ namespace rdf {
     LaneBitmask Mask;
   };
   raw_ostream &operator<< (raw_ostream &OS, const PrintLaneMaskOpt &P);
-
 } // end namespace rdf
 
 } // end namespace llvm
 
+namespace std {
+  template <> struct hash<llvm::rdf::RegisterRef> {
+    size_t operator()(llvm::rdf::RegisterRef A) const {
+      return A.hash();
+    }
+  };
+  template <> struct hash<llvm::rdf::RegisterAggr> {
+    size_t operator()(const llvm::rdf::RegisterAggr &A) const {
+      return A.hash();
+    }
+  };
+  template <> struct equal_to<llvm::rdf::RegisterAggr> {
+    bool operator()(const llvm::rdf::RegisterAggr &A,
+                    const llvm::rdf::RegisterAggr &B) const {
+      return A == B;
+    }
+  };
+}
 #endif // LLVM_LIB_TARGET_HEXAGON_RDFREGISTERS_H

llvm/lib/CodeGen/RDFLiveness.cpp

@@ -47,6 +47,7 @@
 #include <cstdint>
 #include <iterator>
 #include <map>
+#include <unordered_map>
 #include <utility>
 #include <vector>
 
@@ -476,7 +477,7 @@ void Liveness::computePhiInfo() {
   // phi use -> (map: reaching phi -> set of registers defined in between)
   std::map<NodeId,std::map<NodeId,RegisterAggr>> PhiUp;
   std::vector<NodeId> PhiUQ;  // Work list of phis for upward propagation.
-  std::map<NodeId,RegisterAggr> PhiDRs;  // Phi -> registers defined by it.
+  std::unordered_map<NodeId,RegisterAggr> PhiDRs;  // Phi -> registers defined by it.
 
   // Go over all phis.
   for (NodeAddr<PhiNode*> PhiA : Phis) {
@@ -652,6 +653,23 @@ void Liveness::computePhiInfo() {
   // is covered, or until reaching the final phi. Only assume that the
   // reference reaches the phi in the latter case.
 
+  // The operation "clearIn" can be expensive. For a given set of intervening
+  // defs, cache the result of subtracting these defs from a given register
+  // ref.
+  using SubMap = std::unordered_map<RegisterRef, RegisterRef>;
+  std::unordered_map<RegisterAggr, SubMap> Subs;
+  auto ClearIn = [] (RegisterRef RR, const RegisterAggr &Mid, SubMap &SM) {
+    if (Mid.empty())
+      return RR;
+    auto F = SM.find(RR);
+    if (F != SM.end())
+      return F->second;
+    RegisterRef S = Mid.clearIn(RR);
+    SM.insert({RR, S});
+    return S;
+  };
+
+  // Go over all phis.
   for (unsigned i = 0; i < PhiUQ.size(); ++i) {
     auto PA = DFG.addr<PhiNode*>(PhiUQ[i]);
     NodeList PUs = PA.Addr->members_if(DFG.IsRef<NodeAttrs::Use>, DFG);
@@ -663,13 +681,13 @@ void Liveness::computePhiInfo() {
       for (const std::pair<const NodeId, RegisterAggr> &P : PUM) {
         bool Changed = false;
         const RegisterAggr &MidDefs = P.second;
-
         // Collect the set PropUp of uses that are reached by the current
         // phi PA, and are not covered by any intervening def between the
         // currently visited use UA and the upward phi P.
 
         if (MidDefs.hasCoverOf(UR))
           continue;
+        SubMap &SM = Subs[MidDefs];
 
         // General algorithm:
         //   for each (R,U) : U is use node of R, U is reached by PA
@@ -689,7 +707,7 @@ void Liveness::computePhiInfo() {
             LaneBitmask M = R.Mask & V.second;
             if (M.none())
               continue;
-            if (RegisterRef SS = MidDefs.clearIn(RegisterRef(R.Reg, M))) {
+            if (RegisterRef SS = ClearIn(RegisterRef(R.Reg, M), MidDefs, SM)) {
               NodeRefSet &RS = RealUseMap[P.first][SS.Reg];
               Changed |= RS.insert({V.first,SS.Mask}).second;
             }