Record whether the weights on out-edges from a MBB are normalized.
authorCong Hou <congh@google.com>
Wed, 5 Aug 2015 22:01:20 +0000 (22:01 +0000)
committerCong Hou <congh@google.com>
Wed, 5 Aug 2015 22:01:20 +0000 (22:01 +0000)
1. Create a utility function normalizeEdgeWeights() in MachineBranchProbabilityInfo that normalizes a list of edge weights so that the sum of then can fit in uint32_t.
2. Provide an interface in MachineBasicBlock to normalize its successors' weights.
3. Add a flag in MachineBasicBlock that tracks whether its successors' weights are normalized.
4. Provide an overload of getSumForBlock that accepts a non-const pointer to a MBB so that it can force normalizing this MBB's successors' weights.
5. Update several uses of getSumForBlock() by eliminating the once needed weight scale.

Differential Revision: http://reviews.llvm.org/D11442

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@244154 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/CodeGen/MachineBasicBlock.h
include/llvm/CodeGen/MachineBranchProbabilityInfo.h
lib/CodeGen/IfConversion.cpp
lib/CodeGen/MachineBasicBlock.cpp
lib/CodeGen/MachineBlockPlacement.cpp
lib/CodeGen/MachineBranchProbabilityInfo.cpp

index 5486f8a..7be5cb2 100644 (file)
@@ -65,6 +65,10 @@ class MachineBasicBlock : public ilist_node<MachineBasicBlock> {
   Instructions Insts;
   const BasicBlock *BB;
   int Number;
+
+  /// A flag tracking whether the weights of all successors are normalized.
+  bool AreSuccWeightsNormalized;
+
   MachineFunction *xParent;
 
   /// Keep track of the predecessor / successor basicblocks.
@@ -129,6 +133,9 @@ public:
   const MachineFunction *getParent() const { return xParent; }
   MachineFunction *getParent() { return xParent; }
 
+  /// Return whether all weights of successors are normalized.
+  bool areSuccWeightsNormalized() const { return AreSuccWeightsNormalized; }
+
   /// MachineBasicBlock iterator that automatically skips over MIs that are
   /// inside bundles (i.e. walk top level MIs only).
   template<typename Ty, typename IterTy>
@@ -384,6 +391,12 @@ public:
   /// Set successor weight of a given iterator.
   void setSuccWeight(succ_iterator I, uint32_t weight);
 
+  /// Normalize all succesor weights so that the sum of them does not exceed
+  /// UINT32_MAX. Return true if the weights are modified and false otherwise.
+  /// Note that weights that are modified after calling this function are not
+  /// guaranteed to be normalized.
+  bool normalizeSuccWeights();
+
   /// Remove successor from the successors list of this MachineBasicBlock. The
   /// Predecessors list of succ is automatically updated.
   void removeSuccessor(MachineBasicBlock *succ);
index 7ba7495..55b9696 100644 (file)
@@ -59,6 +59,10 @@ public:
   // adjustment. Any edge weights used with the sum should be divided by Scale.
   uint32_t getSumForBlock(const MachineBasicBlock *MBB, uint32_t &Scale) const;
 
+  // Get sum of the block successors' weights, and force normalizing the
+  // successors' weights of MBB so that their sum fit within 32-bits.
+  uint32_t getSumForBlock(MachineBasicBlock *MBB) const;
+
   // A 'Hot' edge is an edge which probability is >= 80%.
   bool isEdgeHot(const MachineBasicBlock *Src,
                  const MachineBasicBlock *Dst) const;
@@ -82,8 +86,34 @@ public:
   raw_ostream &printEdgeProbability(raw_ostream &OS,
                                     const MachineBasicBlock *Src,
                                     const MachineBasicBlock *Dst) const;
+
+  // Normalize a list of weights by scaling them down so that the sum of them
+  // doesn't exceed UINT32_MAX. Return the scale.
+  template <class WeightList>
+  static uint32_t normalizeEdgeWeights(WeightList &Weights);
 };
 
+template <class WeightList>
+uint32_t
+MachineBranchProbabilityInfo::normalizeEdgeWeights(WeightList &Weights) {
+  assert(Weights.size() < UINT32_MAX && "Too many weights in the list!");
+  // First we compute the sum with 64-bits of precision.
+  uint64_t Sum = std::accumulate(Weights.begin(), Weights.end(), uint64_t(0));
+
+  // If the computed sum fits in 32-bits, we're done.
+  if (Sum <= UINT32_MAX)
+    return 1;
+
+  // Otherwise, compute the scale necessary to cause the weights to fit, and
+  // re-sum with that scale applied.
+  assert((Sum / UINT32_MAX) < UINT32_MAX &&
+         "The sum of weights exceeds UINT32_MAX^2!");
+  uint32_t Scale = (Sum / UINT32_MAX) + 1;
+  for (auto &W : Weights)
+    W /= Scale;
+  return Scale;
+}
+
 }
 
 
index ee0532b..8896cdb 100644 (file)
@@ -1232,15 +1232,17 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
 
   bool HasEarlyExit = CvtBBI->FalseBB != nullptr;
   uint64_t CvtNext = 0, CvtFalse = 0, BBNext = 0, BBCvt = 0, SumWeight = 0;
-  uint32_t WeightScale = 0;
 
   if (HasEarlyExit) {
     // Get weights before modifying CvtBBI->BB and BBI.BB.
+    // Explictly normalize the weights of all edges from CvtBBI->BB so that we
+    // are aware that the edge weights obtained below are normalized.
+    CvtBBI->BB->normalizeSuccWeights();
     CvtNext = MBPI->getEdgeWeight(CvtBBI->BB, NextBBI->BB);
     CvtFalse = MBPI->getEdgeWeight(CvtBBI->BB, CvtBBI->FalseBB);
     BBNext = MBPI->getEdgeWeight(BBI.BB, NextBBI->BB);
     BBCvt = MBPI->getEdgeWeight(BBI.BB, CvtBBI->BB);
-    SumWeight = MBPI->getSumForBlock(CvtBBI->BB, WeightScale);
+    SumWeight = MBPI->getSumForBlock(CvtBBI->BB);
   }
 
   if (CvtBBI->BB->pred_size() > 1) {
@@ -1277,8 +1279,8 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
     // New_Weight(BBI.BB, CvtBBI->FalseBB) =
     //   Weight(BBI.BB, CvtBBI->BB) * Weight(CvtBBI->BB, CvtBBI->FalseBB)
 
-    uint64_t NewNext = BBNext * SumWeight + (BBCvt * CvtNext) / WeightScale;
-    uint64_t NewFalse = (BBCvt * CvtFalse) / WeightScale;
+    uint64_t NewNext = BBNext * SumWeight + BBCvt * CvtNext;
+    uint64_t NewFalse = BBCvt * CvtFalse;
     // We need to scale down all weights of BBI.BB to fit uint32_t.
     // Here BBI.BB is connected to CvtBBI->FalseBB and will fall through to
     // the next block.
index 5d3f7eb..e2f381e 100644 (file)
@@ -16,6 +16,7 @@
 #include "llvm/ADT/SmallString.h"
 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
 #include "llvm/CodeGen/LiveVariables.h"
+#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
 #include "llvm/CodeGen/MachineDominators.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
@@ -39,8 +40,9 @@ using namespace llvm;
 #define DEBUG_TYPE "codegen"
 
 MachineBasicBlock::MachineBasicBlock(MachineFunction &mf, const BasicBlock *bb)
-  : BB(bb), Number(-1), xParent(&mf), Alignment(0), IsLandingPad(false),
-    AddressTaken(false), CachedMCSymbol(nullptr) {
+    : BB(bb), Number(-1), AreSuccWeightsNormalized(false), xParent(&mf),
+      Alignment(0), IsLandingPad(false), AddressTaken(false),
+      CachedMCSymbol(nullptr) {
   Insts.Parent = this;
 }
 
@@ -481,8 +483,10 @@ void MachineBasicBlock::addSuccessor(MachineBasicBlock *succ, uint32_t weight) {
   if (weight != 0 && Weights.empty())
     Weights.resize(Successors.size());
 
-  if (weight != 0 || !Weights.empty())
+  if (weight != 0 || !Weights.empty()) {
     Weights.push_back(weight);
+    AreSuccWeightsNormalized = false;
+  }
 
    Successors.push_back(succ);
    succ->addPredecessor(this);
@@ -1096,7 +1100,25 @@ uint32_t MachineBasicBlock::getSuccWeight(const_succ_iterator Succ) const {
 void MachineBasicBlock::setSuccWeight(succ_iterator I, uint32_t weight) {
   if (Weights.empty())
     return;
-  *getWeightIterator(I) = weight;
+  auto WeightIter = getWeightIterator(I);
+  uint32_t OldWeight = *WeightIter;
+  *WeightIter = weight;
+  if (weight > OldWeight)
+    AreSuccWeightsNormalized = false;
+}
+
+/// Normalize all succesor weights so that the sum of them does not exceed
+/// UINT32_MAX. Return true if the weights are modified and false otherwise.
+/// Note that weights that are modified after calling this function are not
+/// guaranteed to be normalized.
+bool MachineBasicBlock::normalizeSuccWeights() {
+  if (!AreSuccWeightsNormalized) {
+    uint32_t Scale =
+        MachineBranchProbabilityInfo::normalizeEdgeWeights(Weights);
+    AreSuccWeightsNormalized = true;
+    return Scale != 1;
+  }
+  return false;
 }
 
 /// getWeightIterator - Return wight iterator corresonding to the I successor
index b77c803..ecc093e 100644 (file)
@@ -361,8 +361,7 @@ MachineBlockPlacement::selectBestSuccessor(MachineBasicBlock *BB,
   // improve the MBPI interface to efficiently support query patterns such as
   // this.
   uint32_t BestWeight = 0;
-  uint32_t WeightScale = 0;
-  uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
+  uint32_t SumWeight = MBPI->getSumForBlock(BB);
   DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
   for (MachineBasicBlock *Succ : BB->successors()) {
     if (BlockFilter && !BlockFilter->count(Succ))
@@ -378,7 +377,7 @@ MachineBlockPlacement::selectBestSuccessor(MachineBasicBlock *BB,
     }
 
     uint32_t SuccWeight = MBPI->getEdgeWeight(BB, Succ);
-    BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
+    BranchProbability SuccProb(SuccWeight, SumWeight);
 
     // If we outline optional branches, look whether Succ is unavoidable, i.e.
     // dominates all terminators of the MachineFunction. If it does, other
@@ -675,8 +674,7 @@ MachineBlockPlacement::findBestLoopExit(MachineFunction &F, MachineLoop &L,
     // FIXME: Due to the performance of the probability and weight routines in
     // the MBPI analysis, we use the internal weights and manually compute the
     // probabilities to avoid quadratic behavior.
-    uint32_t WeightScale = 0;
-    uint32_t SumWeight = MBPI->getSumForBlock(MBB, WeightScale);
+    uint32_t SumWeight = MBPI->getSumForBlock(MBB);
     for (MachineBasicBlock *Succ : MBB->successors()) {
       if (Succ->isLandingPad())
         continue;
@@ -705,7 +703,7 @@ MachineBlockPlacement::findBestLoopExit(MachineFunction &F, MachineLoop &L,
           BlocksExitingToOuterLoop.insert(MBB);
       }
 
-      BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
+      BranchProbability SuccProb(SuccWeight, SumWeight);
       BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
       DEBUG(dbgs() << "    exiting: " << getBlockName(MBB) << " -> "
                    << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] (";
index 6fbc2be..fe03d4d 100644 (file)
@@ -28,36 +28,35 @@ char MachineBranchProbabilityInfo::ID = 0;
 
 void MachineBranchProbabilityInfo::anchor() { }
 
-uint32_t MachineBranchProbabilityInfo::
-getSumForBlock(const MachineBasicBlock *MBB, uint32_t &Scale) const {
-  // First we compute the sum with 64-bits of precision, ensuring that cannot
-  // overflow by bounding the number of weights considered. Hopefully no one
-  // actually needs 2^32 successors.
-  assert(MBB->succ_size() < UINT32_MAX);
-  uint64_t Sum = 0;
-  Scale = 1;
+uint32_t
+MachineBranchProbabilityInfo::getSumForBlock(MachineBasicBlock *MBB) const {
+  // Normalize the weights of MBB's all successors so that the sum is guaranteed
+  // to be no greater than UINT32_MAX.
+  MBB->normalizeSuccWeights();
+
+  SmallVector<uint32_t, 8> Weights;
   for (MachineBasicBlock::const_succ_iterator I = MBB->succ_begin(),
-       E = MBB->succ_end(); I != E; ++I) {
-    uint32_t Weight = getEdgeWeight(MBB, I);
-    Sum += Weight;
-  }
+                                              E = MBB->succ_end();
+       I != E; ++I)
+    Weights.push_back(getEdgeWeight(MBB, I));
 
-  // If the computed sum fits in 32-bits, we're done.
-  if (Sum <= UINT32_MAX)
-    return Sum;
+  return std::accumulate(Weights.begin(), Weights.end(), 0u);
+}
 
-  // Otherwise, compute the scale necessary to cause the weights to fit, and
-  // re-sum with that scale applied.
-  assert((Sum / UINT32_MAX) < UINT32_MAX);
-  Scale = (Sum / UINT32_MAX) + 1;
-  Sum = 0;
+uint32_t
+MachineBranchProbabilityInfo::getSumForBlock(const MachineBasicBlock *MBB,
+                                             uint32_t &Scale) const {
+  SmallVector<uint32_t, 8> Weights;
   for (MachineBasicBlock::const_succ_iterator I = MBB->succ_begin(),
-       E = MBB->succ_end(); I != E; ++I) {
-    uint32_t Weight = getEdgeWeight(MBB, I);
-    Sum += Weight / Scale;
-  }
-  assert(Sum <= UINT32_MAX);
-  return Sum;
+                                              E = MBB->succ_end();
+       I != E; ++I)
+    Weights.push_back(getEdgeWeight(MBB, I));
+
+  if (MBB->areSuccWeightsNormalized())
+    Scale = 1;
+  else
+    Scale = MachineBranchProbabilityInfo::normalizeEdgeWeights(Weights);
+  return std::accumulate(Weights.begin(), Weights.end(), 0u);
 }
 
 uint32_t MachineBranchProbabilityInfo::