[llvm] r323475 - [SyntheticCounts] Rewrite the code using only graph traits.

Easwaran Raman via llvm-commits llvm-commits at lists.llvm.org
Thu Jan 25 14:02:29 PST 2018


Author: eraman
Date: Thu Jan 25 14:02:29 2018
New Revision: 323475

URL: http://llvm.org/viewvc/llvm-project?rev=323475&view=rev
Log:
[SyntheticCounts] Rewrite the code using only graph traits.

Summary:
The intent of this is to allow the code to be used with ThinLTO. In
Thinlink phase, a traditional Callgraph can not be computed even though
all the necessary information (nodes and edges of a call graph) is
available. This is due to the fact that CallGraph class is closely tied
to the IR. This patch first extends GraphTraits to add a CallGraphTraits
graph. This is then used to implement a version of counts propagation
on a generic callgraph.

Reviewers: davidxl

Subscribers: mehdi_amini, tejohnson, llvm-commits

Differential Revision: https://reviews.llvm.org/D42311

Modified:
    llvm/trunk/include/llvm/Analysis/CallGraph.h
    llvm/trunk/include/llvm/Analysis/SyntheticCountsUtils.h
    llvm/trunk/lib/Analysis/SyntheticCountsUtils.cpp
    llvm/trunk/lib/Transforms/IPO/SyntheticCountsPropagation.cpp

Modified: llvm/trunk/include/llvm/Analysis/CallGraph.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Analysis/CallGraph.h?rev=323475&r1=323474&r2=323475&view=diff
==============================================================================
--- llvm/trunk/include/llvm/Analysis/CallGraph.h (original)
+++ llvm/trunk/include/llvm/Analysis/CallGraph.h Thu Jan 25 14:02:29 2018
@@ -495,6 +495,56 @@ struct GraphTraits<const CallGraph *> :
   }
 };
 
+// FIXME: The traits here are not limited to callgraphs and can be moved
+// elsewhere including GraphTraits. They are left here because only algorithms
+// that operate on Callgraphs currently use them. If other algorithms operating
+// on a general graph need edge traversals, these can be moved.
+template <class CallGraphType>
+struct CallGraphTraits : public GraphTraits<CallGraphType> {
+  // Elements to provide:
+
+  // typedef EdgeRef           - Type of Edge token in the graph, which should
+  //                             be cheap to copy.
+  // typedef CallEdgeIteratorType - Type used to iterate over children edges in
+  //                             graph, dereference to a EdgeRef.
+
+  // static CallEdgeIteratorType call_edge_begin(NodeRef)
+  // static CallEdgeIteratorType call_edge_end  (NodeRef)
+  //     Return iterators that point to the beginning and ending of the call
+  //     edges list for the given callgraph node.
+  //
+  // static NodeRef edge_dest(EdgeRef)
+  //     Return the destination node of an edge.
+
+  // If anyone tries to use this class without having an appropriate
+  // specialization, make an error.  If you get this error, it's because you
+  // need to include the appropriate specialization of GraphTraits<> for your
+  // graph, or you need to define it for a new graph type. Either that or
+  // your argument to XXX_begin(...) is unknown or needs to have the proper .h
+  // file #include'd.
+  using CallEdgeIteratorType =
+      typename CallGraphType::UnknownCallGraphTypeError;
+};
+
+template <class GraphType>
+iterator_range<typename CallGraphTraits<GraphType>::CallEdgeIteratorType>
+call_edges(const typename CallGraphTraits<GraphType>::NodeRef &G) {
+  return make_range(CallGraphTraits<GraphType>::call_edge_begin(G),
+                    CallGraphTraits<GraphType>::call_edge_end(G));
+}
+
+template <>
+struct CallGraphTraits<const CallGraph *>
+    : public GraphTraits<const CallGraph *> {
+  using EdgeRef = const CallGraphNode::CallRecord &;
+  using CallEdgeIteratorType = CallGraphNode::const_iterator;
+
+  static CallEdgeIteratorType call_edge_begin(NodeRef N) { return N->begin(); }
+  static CallEdgeIteratorType call_edge_end(NodeRef N) { return N->end(); }
+
+  static NodeRef edge_dest(EdgeRef E) { return E.second; }
+};
+
 } // end namespace llvm
 
 #endif // LLVM_ANALYSIS_CALLGRAPH_H

Modified: llvm/trunk/include/llvm/Analysis/SyntheticCountsUtils.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Analysis/SyntheticCountsUtils.h?rev=323475&r1=323474&r2=323475&view=diff
==============================================================================
--- llvm/trunk/include/llvm/Analysis/SyntheticCountsUtils.h (original)
+++ llvm/trunk/include/llvm/Analysis/SyntheticCountsUtils.h Thu Jan 25 14:02:29 2018
@@ -15,7 +15,7 @@
 #define LLVM_ANALYSIS_SYNTHETIC_COUNTS_UTILS_H
 
 #include "llvm/ADT/STLExtras.h"
-#include "llvm/IR/CallSite.h"
+#include "llvm/Analysis/CallGraph.h"
 #include "llvm/Support/ScaledNumber.h"
 
 namespace llvm {
@@ -23,11 +23,30 @@ namespace llvm {
 class CallGraph;
 class Function;
 
-using Scaled64 = ScaledNumber<uint64_t>;
-void propagateSyntheticCounts(
-    const CallGraph &CG, function_ref<Scaled64(CallSite CS)> GetCallSiteRelFreq,
-    function_ref<uint64_t(Function *F)> GetCount,
-    function_ref<void(Function *F, uint64_t)> AddToCount);
+/// Class with methods to propagate synthetic entry counts.
+///
+/// This class is templated on the type of the call graph and designed to work
+/// with the traditional per-module callgraph and the summary callgraphs used in
+/// ThinLTO. This contains only static methods and alias templates.
+template <typename CallGraphType> class SyntheticCountsUtils {
+public:
+  using Scaled64 = ScaledNumber<uint64_t>;
+  using CGT = CallGraphTraits<CallGraphType>;
+  using NodeRef = typename CGT::NodeRef;
+  using EdgeRef = typename CGT::EdgeRef;
+  using SccTy = std::vector<NodeRef>;
+
+  using GetRelBBFreqTy = function_ref<Optional<Scaled64>(EdgeRef)>;
+  using GetCountTy = function_ref<uint64_t(NodeRef)>;
+  using AddCountTy = function_ref<void(NodeRef, uint64_t)>;
+
+  static void propagate(const CallGraphType &CG, GetRelBBFreqTy GetRelBBFreq,
+                        GetCountTy GetCount, AddCountTy AddCount);
+
+private:
+  static void propagateFromSCC(const SccTy &SCC, GetRelBBFreqTy GetRelBBFreq,
+                               GetCountTy GetCount, AddCountTy AddCount);
+};
 } // namespace llvm
 
 #endif

Modified: llvm/trunk/lib/Analysis/SyntheticCountsUtils.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/SyntheticCountsUtils.cpp?rev=323475&r1=323474&r2=323475&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/SyntheticCountsUtils.cpp (original)
+++ llvm/trunk/lib/Analysis/SyntheticCountsUtils.cpp Thu Jan 25 14:02:29 2018
@@ -23,100 +23,91 @@
 
 using namespace llvm;
 
-// Given a set of functions in an SCC, propagate entry counts to functions
-// called by the SCC.
-static void
-propagateFromSCC(const SmallPtrSetImpl<Function *> &SCCFunctions,
-                 function_ref<Scaled64(CallSite CS)> GetCallSiteRelFreq,
-                 function_ref<uint64_t(Function *F)> GetCount,
-                 function_ref<void(Function *F, uint64_t)> AddToCount) {
-
-  SmallVector<CallSite, 16> CallSites;
-
-  // Gather all callsites in the SCC.
-  auto GatherCallSites = [&]() {
-    for (auto *F : SCCFunctions) {
-      assert(F && !F->isDeclaration());
-      for (auto &I : instructions(F)) {
-        if (auto CS = CallSite(&I)) {
-          CallSites.push_back(CS);
-        }
-      }
+// Given an SCC, propagate entry counts along the edge of the SCC nodes.
+template <typename CallGraphType>
+void SyntheticCountsUtils<CallGraphType>::propagateFromSCC(
+    const SccTy &SCC, GetRelBBFreqTy GetRelBBFreq, GetCountTy GetCount,
+    AddCountTy AddCount) {
+
+  SmallPtrSet<NodeRef, 8> SCCNodes;
+  SmallVector<std::pair<NodeRef, EdgeRef>, 8> SCCEdges, NonSCCEdges;
+
+  for (auto &Node : SCC)
+    SCCNodes.insert(Node);
+
+  // Partition the edges coming out of the SCC into those whose destination is
+  // in the SCC and the rest.
+  for (const auto &Node : SCCNodes) {
+    for (auto &E : call_edges<CallGraphType>(Node)) {
+      if (SCCNodes.count(CGT::edge_dest(E)))
+        SCCEdges.emplace_back(Node, E);
+      else
+        NonSCCEdges.emplace_back(Node, E);
     }
-  };
-
-  GatherCallSites();
+  }
 
-  // Partition callsites so that the callsites that call functions in the same
-  // SCC come first.
-  auto Mid = partition(CallSites, [&](CallSite &CS) {
-    auto *Callee = CS.getCalledFunction();
-    if (Callee)
-      return SCCFunctions.count(Callee);
-    // FIXME: Use the !callees metadata to propagate counts through indirect
-    // calls.
-    return 0U;
-  });
-
-  // For functions in the same SCC, update the counts in two steps:
-  // 1. Compute the additional count for each function by propagating the counts
-  // along all incoming edges to the function that originate from the same SCC
-  // and summing them up.
-  // 2. Add the additional counts to the functions in the SCC.
+  // For nodes in the same SCC, update the counts in two steps:
+  // 1. Compute the additional count for each node by propagating the counts
+  // along all incoming edges to the node that originate from within the same
+  // SCC and summing them up.
+  // 2. Add the additional counts to the nodes in the SCC.
   // This ensures that the order of
-  // traversal of functions within the SCC doesn't change the final result.
+  // traversal of nodes within the SCC doesn't affect the final result.
 
-  DenseMap<Function *, uint64_t> AdditionalCounts;
-  for (auto It = CallSites.begin(); It != Mid; It++) {
-    auto &CS = *It;
-    auto RelFreq = GetCallSiteRelFreq(CS);
-    Function *Callee = CS.getCalledFunction();
-    Function *Caller = CS.getCaller();
+  DenseMap<NodeRef, uint64_t> AdditionalCounts;
+  for (auto &E : SCCEdges) {
+    auto OptRelFreq = GetRelBBFreq(E.second);
+    if (!OptRelFreq)
+      continue;
+    Scaled64 RelFreq = OptRelFreq.getValue();
+    auto Caller = E.first;
+    auto Callee = CGT::edge_dest(E.second);
     RelFreq *= Scaled64(GetCount(Caller), 0);
     uint64_t AdditionalCount = RelFreq.toInt<uint64_t>();
     AdditionalCounts[Callee] += AdditionalCount;
   }
 
-  // Update the counts for the functions in the SCC.
+  // Update the counts for the nodes in the SCC.
   for (auto &Entry : AdditionalCounts)
-    AddToCount(Entry.first, Entry.second);
+    AddCount(Entry.first, Entry.second);
 
-  // Now update the counts for functions not in SCC.
-  for (auto It = Mid; It != CallSites.end(); It++) {
-    auto &CS = *It;
-    auto Weight = GetCallSiteRelFreq(CS);
-    Function *Callee = CS.getCalledFunction();
-    Function *Caller = CS.getCaller();
-    Weight *= Scaled64(GetCount(Caller), 0);
-    AddToCount(Callee, Weight.toInt<uint64_t>());
+  // Now update the counts for nodes outside the SCC.
+  for (auto &E : NonSCCEdges) {
+    auto OptRelFreq = GetRelBBFreq(E.second);
+    if (!OptRelFreq)
+      continue;
+    Scaled64 RelFreq = OptRelFreq.getValue();
+    auto Caller = E.first;
+    auto Callee = CGT::edge_dest(E.second);
+    RelFreq *= Scaled64(GetCount(Caller), 0);
+    AddCount(Callee, RelFreq.toInt<uint64_t>());
   }
 }
 
-/// Propgate synthetic entry counts on a callgraph.
+/// Propgate synthetic entry counts on a callgraph \p CG.
 ///
 /// This performs a reverse post-order traversal of the callgraph SCC. For each
-/// SCC, it first propagates the entry counts to the functions within the SCC
+/// SCC, it first propagates the entry counts to the nodes within the SCC
 /// through call edges and updates them in one shot. Then the entry counts are
-/// propagated to functions outside the SCC.
-void llvm::propagateSyntheticCounts(
-    const CallGraph &CG, function_ref<Scaled64(CallSite CS)> GetCallSiteRelFreq,
-    function_ref<uint64_t(Function *F)> GetCount,
-    function_ref<void(Function *F, uint64_t)> AddToCount) {
-
-  SmallVector<SmallPtrSet<Function *, 8>, 16> SCCs;
-  for (auto I = scc_begin(&CG); !I.isAtEnd(); ++I) {
-    auto SCC = *I;
-
-    SmallPtrSet<Function *, 8> SCCFunctions;
-    for (auto *Node : SCC) {
-      Function *F = Node->getFunction();
-      if (F && !F->isDeclaration()) {
-        SCCFunctions.insert(F);
-      }
-    }
-    SCCs.push_back(SCCFunctions);
-  }
+/// propagated to nodes outside the SCC. This requires \p CallGraphTraits
+/// to have a specialization for \p CallGraphType.
 
-  for (auto &SCCFunctions : reverse(SCCs))
-    propagateFromSCC(SCCFunctions, GetCallSiteRelFreq, GetCount, AddToCount);
+template <typename CallGraphType>
+void SyntheticCountsUtils<CallGraphType>::propagate(const CallGraphType &CG,
+                                                    GetRelBBFreqTy GetRelBBFreq,
+                                                    GetCountTy GetCount,
+                                                    AddCountTy AddCount) {
+  std::vector<SccTy> SCCs;
+
+  // Collect all the SCCs.
+  for (auto I = scc_begin(CG); !I.isAtEnd(); ++I)
+    SCCs.push_back(*I);
+
+  // The callgraph-scc needs to be visited in top-down order for propagation.
+  // The scc iterator returns the scc in bottom-up order, so reverse the SCCs
+  // and call propagateFromSCC.
+  for (auto &SCC : reverse(SCCs))
+    propagateFromSCC(SCC, GetRelBBFreq, GetCount, AddCount);
 }
+
+template class llvm::SyntheticCountsUtils<const CallGraph *>;

Modified: llvm/trunk/lib/Transforms/IPO/SyntheticCountsPropagation.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/IPO/SyntheticCountsPropagation.cpp?rev=323475&r1=323474&r2=323475&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/IPO/SyntheticCountsPropagation.cpp (original)
+++ llvm/trunk/lib/Transforms/IPO/SyntheticCountsPropagation.cpp Thu Jan 25 14:02:29 2018
@@ -102,23 +102,34 @@ PreservedAnalyses SyntheticCountsPropaga
   // Set initial entry counts.
   initializeCounts(M, [&](Function *F, uint64_t Count) { Counts[F] = Count; });
 
-  // Compute the relative block frequency for a callsite. Use scaled numbers
+  // Compute the relative block frequency for a call edge. Use scaled numbers
   // and not integers since the relative block frequency could be less than 1.
-  auto GetCallSiteRelFreq = [&](CallSite CS) {
+  auto GetCallSiteRelFreq = [&](const CallGraphNode::CallRecord &Edge) {
+    Optional<Scaled64> Res = None;
+    if (!Edge.first)
+      return Res;
+    assert(isa<Instruction>(Edge.first));
+    CallSite CS(cast<Instruction>(Edge.first));
     Function *Caller = CS.getCaller();
     auto &BFI = FAM.getResult<BlockFrequencyAnalysis>(*Caller);
     BasicBlock *CSBB = CS.getInstruction()->getParent();
     Scaled64 EntryFreq(BFI.getEntryFreq(), 0);
     Scaled64 BBFreq(BFI.getBlockFreq(CSBB).getFrequency(), 0);
     BBFreq /= EntryFreq;
-    return BBFreq;
+    return Optional<Scaled64>(BBFreq);
   };
 
   CallGraph CG(M);
   // Propgate the entry counts on the callgraph.
-  propagateSyntheticCounts(
-      CG, GetCallSiteRelFreq, [&](Function *F) { return Counts[F]; },
-      [&](Function *F, uint64_t New) { Counts[F] += New; });
+  SyntheticCountsUtils<const CallGraph *>::propagate(
+      &CG, GetCallSiteRelFreq,
+      [&](const CallGraphNode *N) { return Counts[N->getFunction()]; },
+      [&](const CallGraphNode *N, uint64_t New) {
+        auto F = N->getFunction();
+        if (!F || F->isDeclaration())
+          return;
+        Counts[F] += New;
+      });
 
   // Set the counts as metadata.
   for (auto Entry : Counts)




More information about the llvm-commits mailing list