Reapply "blockfreq: Rewrite BlockFrequencyInfoImpl"

This reverts commit r206707, reapplying r206704.  The preceding commit
to CalcSpillWeights should have sorted out the failing buildbots.

<rdar://problem/14292693>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206766 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Analysis/BlockFrequencyInfo.cpp

@@ -11,6 +11,7 @@
 //
 //===----------------------------------------------------------------------===//
 
+#define DEBUG_TYPE "block-freq"
 #include "llvm/Analysis/BlockFrequencyInfo.h"
 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
 #include "llvm/Analysis/BranchProbabilityInfo.h"
@@ -106,6 +107,7 @@ struct DOTGraphTraits<BlockFrequencyInfo*> : public DefaultDOTGraphTraits {
 INITIALIZE_PASS_BEGIN(BlockFrequencyInfo, "block-freq",
                       "Block Frequency Analysis", true, true)
 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfo)
+INITIALIZE_PASS_DEPENDENCY(LoopInfo)
 INITIALIZE_PASS_END(BlockFrequencyInfo, "block-freq",
                     "Block Frequency Analysis", true, true)
 
@@ -120,14 +122,16 @@ BlockFrequencyInfo::~BlockFrequencyInfo() {}
 
 void BlockFrequencyInfo::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.addRequired<BranchProbabilityInfo>();
+  AU.addRequired<LoopInfo>();
   AU.setPreservesAll();
 }
 
 bool BlockFrequencyInfo::runOnFunction(Function &F) {
   BranchProbabilityInfo &BPI = getAnalysis<BranchProbabilityInfo>();
+  LoopInfo &LI = getAnalysis<LoopInfo>();
   if (!BFI)
     BFI.reset(new ImplType);
-  BFI->doFunction(&F, &BPI);
+  BFI->doFunction(&F, &BPI, &LI);
 #ifndef NDEBUG
   if (ViewBlockFreqPropagationDAG != GVDT_None)
     view();
@@ -158,7 +162,7 @@ void BlockFrequencyInfo::view() const {
 }
 
 const Function *BlockFrequencyInfo::getFunction() const {
-  return BFI ? BFI->Fn : nullptr;
+  return BFI ? BFI->getFunction() : nullptr;
 }
 
 raw_ostream &BlockFrequencyInfo::

lib/Analysis/BlockFrequencyInfoImpl.cpp

@@ -0,0 +1,932 @@
+//===- BlockFrequencyImplInfo.cpp - Block Frequency Info Implementation ---===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// Loops should be simplified before this analysis.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "block-freq"
+#include "llvm/Analysis/BlockFrequencyInfoImpl.h"
+#include "llvm/ADT/APFloat.h"
+#include "llvm/Support/raw_ostream.h"
+#include <deque>
+
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+//
+// PositiveFloat implementation.
+//
+//===----------------------------------------------------------------------===//
+#ifndef _MSC_VER
+const int32_t PositiveFloatBase::MaxExponent;
+const int32_t PositiveFloatBase::MinExponent;
+#endif
+
+static void appendDigit(std::string &Str, unsigned D) {
+  assert(D < 10);
+  Str += '0' + D % 10;
+}
+
+static void appendNumber(std::string &Str, uint64_t N) {
+  while (N) {
+    appendDigit(Str, N % 10);
+    N /= 10;
+  }
+}
+
+static bool doesRoundUp(char Digit) {
+  switch (Digit) {
+  case '5':
+  case '6':
+  case '7':
+  case '8':
+  case '9':
+    return true;
+  default:
+    return false;
+  }
+}
+
+static std::string toStringAPFloat(uint64_t D, int E, unsigned Precision) {
+  assert(E >= PositiveFloatBase::MinExponent);
+  assert(E <= PositiveFloatBase::MaxExponent);
+
+  // Find a new E, but don't let it increase past MaxExponent.
+  int LeadingZeros = PositiveFloatBase::countLeadingZeros64(D);
+  int NewE = std::min(PositiveFloatBase::MaxExponent, E + 63 - LeadingZeros);
+  int Shift = 63 - (NewE - E);
+  assert(Shift <= LeadingZeros);
+  assert(Shift == LeadingZeros || NewE == PositiveFloatBase::MaxExponent);
+  D <<= Shift;
+  E = NewE;
+
+  // Check for a denormal.
+  unsigned AdjustedE = E + 16383;
+  if (!(D >> 63)) {
+    assert(E == PositiveFloatBase::MaxExponent);
+    AdjustedE = 0;
+  }
+
+  // Build the float and print it.
+  uint64_t RawBits[2] = {D, AdjustedE};
+  APFloat Float(APFloat::x87DoubleExtended, APInt(80, RawBits));
+  SmallVector<char, 24> Chars;
+  Float.toString(Chars, Precision, 0);
+  return std::string(Chars.begin(), Chars.end());
+}
+
+static std::string stripTrailingZeros(const std::string &Float) {
+  size_t NonZero = Float.find_last_not_of('0');
+  assert(NonZero != std::string::npos && "no . in floating point string");
+
+  if (Float[NonZero] == '.')
+    ++NonZero;
+
+  return Float.substr(0, NonZero + 1);
+}
+
+std::string PositiveFloatBase::toString(uint64_t D, int16_t E, int Width,
+                                        unsigned Precision) {
+  if (!D)
+    return "0.0";
+
+  // Canonicalize exponent and digits.
+  uint64_t Above0 = 0;
+  uint64_t Below0 = 0;
+  uint64_t Extra = 0;
+  int ExtraShift = 0;
+  if (E == 0) {
+    Above0 = D;
+  } else if (E > 0) {
+    if (int Shift = std::min(int16_t(countLeadingZeros64(D)), E)) {
+      D <<= Shift;
+      E -= Shift;
+
+      if (!E)
+        Above0 = D;
+    }
+  } else if (E > -64) {
+    Above0 = D >> -E;
+    Below0 = D << (64 + E);
+  } else if (E > -120) {
+    Below0 = D >> (-E - 64);
+    Extra = D << (128 + E);
+    ExtraShift = -64 - E;
+  }
+
+  // Fall back on APFloat for very small and very large numbers.
+  if (!Above0 && !Below0)
+    return toStringAPFloat(D, E, Precision);
+
+  // Append the digits before the decimal.
+  std::string Str;
+  size_t DigitsOut = 0;
+  if (Above0) {
+    appendNumber(Str, Above0);
+    DigitsOut = Str.size();
+  } else
+    appendDigit(Str, 0);
+  std::reverse(Str.begin(), Str.end());
+
+  // Return early if there's nothing after the decimal.
+  if (!Below0)
+    return Str + ".0";
+
+  // Append the decimal and beyond.
+  Str += '.';
+  uint64_t Error = UINT64_C(1) << (64 - Width);
+
+  // We need to shift Below0 to the right to make space for calculating
+  // digits.  Save the precision we're losing in Extra.
+  Extra = (Below0 & 0xf) << 56 | (Extra >> 8);
+  Below0 >>= 4;
+  size_t SinceDot = 0;
+  size_t AfterDot = Str.size();
+  do {
+    if (ExtraShift) {
+      --ExtraShift;
+      Error *= 5;
+    } else
+      Error *= 10;
+
+    Below0 *= 10;
+    Extra *= 10;
+    Below0 += (Extra >> 60);
+    Extra = Extra & (UINT64_MAX >> 4);
+    appendDigit(Str, Below0 >> 60);
+    Below0 = Below0 & (UINT64_MAX >> 4);
+    if (DigitsOut || Str.back() != '0')
+      ++DigitsOut;
+    ++SinceDot;
+  } while (Error && (Below0 << 4 | Extra >> 60) >= Error / 2 &&
+           (!Precision || DigitsOut <= Precision || SinceDot < 2));
+
+  // Return early for maximum precision.
+  if (!Precision || DigitsOut <= Precision)
+    return stripTrailingZeros(Str);
+
+  // Find where to truncate.
+  size_t Truncate =
+      std::max(Str.size() - (DigitsOut - Precision), AfterDot + 1);
+
+  // Check if there's anything to truncate.
+  if (Truncate >= Str.size())
+    return stripTrailingZeros(Str);
+
+  bool Carry = doesRoundUp(Str[Truncate]);
+  if (!Carry)
+    return stripTrailingZeros(Str.substr(0, Truncate));
+
+  // Round with the first truncated digit.
+  for (std::string::reverse_iterator I(Str.begin() + Truncate), E = Str.rend();
+       I != E; ++I) {
+    if (*I == '.')
+      continue;
+    if (*I == '9') {
+      *I = '0';
+      continue;
+    }
+
+    ++*I;
+    Carry = false;
+    break;
+  }
+
+  // Add "1" in front if we still need to carry.
+  return stripTrailingZeros(std::string(Carry, '1') + Str.substr(0, Truncate));
+}
+
+raw_ostream &PositiveFloatBase::print(raw_ostream &OS, uint64_t D, int16_t E,
+                                      int Width, unsigned Precision) {
+  return OS << toString(D, E, Width, Precision);
+}
+
+void PositiveFloatBase::dump(uint64_t D, int16_t E, int Width) {
+  print(dbgs(), D, E, Width, 0) << "[" << Width << ":" << D << "*2^" << E
+                                << "]";
+}
+
+static std::pair<uint64_t, int16_t>
+getRoundedFloat(uint64_t N, bool ShouldRound, int64_t Shift) {
+  if (ShouldRound)
+    if (!++N)
+      // Rounding caused an overflow.
+      return std::make_pair(UINT64_C(1), Shift + 64);
+  return std::make_pair(N, Shift);
+}
+
+std::pair<uint64_t, int16_t> PositiveFloatBase::divide64(uint64_t Dividend,
+                                                         uint64_t Divisor) {
+  // Input should be sanitized.
+  assert(Divisor);
+  assert(Dividend);
+
+  // Minimize size of divisor.
+  int16_t Shift = 0;
+  if (int Zeros = countTrailingZeros(Divisor)) {
+    Shift -= Zeros;
+    Divisor >>= Zeros;
+  }
+
+  // Check for powers of two.
+  if (Divisor == 1)
+    return std::make_pair(Dividend, Shift);
+
+  // Maximize size of dividend.
+  if (int Zeros = countLeadingZeros64(Dividend)) {
+    Shift -= Zeros;
+    Dividend <<= Zeros;
+  }
+
+  // Start with the result of a divide.
+  uint64_t Quotient = Dividend / Divisor;
+  Dividend %= Divisor;
+
+  // Continue building the quotient with long division.
+  //
+  // TODO: continue with largers digits.
+  while (!(Quotient >> 63) && Dividend) {
+    // Shift Dividend, and check for overflow.
+    bool IsOverflow = Dividend >> 63;
+    Dividend <<= 1;
+    --Shift;
+
+    // Divide.
+    bool DoesDivide = IsOverflow || Divisor <= Dividend;
+    Quotient = (Quotient << 1) | uint64_t(DoesDivide);
+    Dividend -= DoesDivide ? Divisor : 0;
+  }
+
+  // Round.
+  if (Dividend >= getHalf(Divisor))
+    if (!++Quotient)
+      // Rounding caused an overflow in Quotient.
+      return std::make_pair(UINT64_C(1), Shift + 64);
+
+  return getRoundedFloat(Quotient, Dividend >= getHalf(Divisor), Shift);
+}
+
+std::pair<uint64_t, int16_t> PositiveFloatBase::multiply64(uint64_t L,
+                                                           uint64_t R) {
+  // Separate into two 32-bit digits (U.L).
+  uint64_t UL = L >> 32, LL = L & UINT32_MAX, UR = R >> 32, LR = R & UINT32_MAX;
+
+  // Compute cross products.
+  uint64_t P1 = UL * UR, P2 = UL * LR, P3 = LL * UR, P4 = LL * LR;
+
+  // Sum into two 64-bit digits.
+  uint64_t Upper = P1, Lower = P4;
+  auto addWithCarry = [&](uint64_t N) {
+    uint64_t NewLower = Lower + (N << 32);
+    Upper += (N >> 32) + (NewLower < Lower);
+    Lower = NewLower;
+  };
+  addWithCarry(P2);
+  addWithCarry(P3);
+
+  // Check whether the upper digit is empty.
+  if (!Upper)
+    return std::make_pair(Lower, 0);
+
+  // Shift as little as possible to maximize precision.
+  unsigned LeadingZeros = countLeadingZeros64(Upper);
+  int16_t Shift = 64 - LeadingZeros;
+  if (LeadingZeros)
+    Upper = Upper << LeadingZeros | Lower >> Shift;
+  bool ShouldRound = Shift && (Lower & UINT64_C(1) << (Shift - 1));
+  return getRoundedFloat(Upper, ShouldRound, Shift);
+}
+
+//===----------------------------------------------------------------------===//
+//
+// BlockMass implementation.
+//
+//===----------------------------------------------------------------------===//
+BlockMass &BlockMass::operator*=(const BranchProbability &P) {
+  uint32_t N = P.getNumerator(), D = P.getDenominator();
+  assert(D && "divide by 0");
+  assert(N <= D && "fraction greater than 1");
+
+  // Fast path for multiplying by 1.0.
+  if (!Mass || N == D)
+    return *this;
+
+  // Get as much precision as we can.
+  int Shift = countLeadingZeros(Mass);
+  uint64_t ShiftedQuotient = (Mass << Shift) / D;
+  uint64_t Product = ShiftedQuotient * N >> Shift;
+
+  // Now check for what's lost.
+  uint64_t Left = ShiftedQuotient * (D - N) >> Shift;
+  uint64_t Lost = Mass - Product - Left;
+
+  // TODO: prove this assertion.
+  assert(Lost <= UINT32_MAX);
+
+  // Take the product plus a portion of the spoils.
+  Mass = Product + Lost * N / D;
+  return *this;
+}
+
+PositiveFloat<uint64_t> BlockMass::toFloat() const {
+  if (isFull())
+    return PositiveFloat<uint64_t>(1, 0);
+  return PositiveFloat<uint64_t>(getMass() + 1, -64);
+}
+
+void BlockMass::dump() const { print(dbgs()); }
+
+static char getHexDigit(int N) {
+  assert(N < 16);
+  if (N < 10)
+    return '0' + N;
+  return 'a' + N - 10;
+}
+raw_ostream &BlockMass::print(raw_ostream &OS) const {
+  for (int Digits = 0; Digits < 16; ++Digits)
+    OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
+  return OS;
+}
+
+//===----------------------------------------------------------------------===//
+//
+// BlockFrequencyInfoImpl implementation.
+//
+//===----------------------------------------------------------------------===//
+namespace {
+
+typedef BlockFrequencyInfoImplBase::BlockNode BlockNode;
+typedef BlockFrequencyInfoImplBase::Distribution Distribution;
+typedef BlockFrequencyInfoImplBase::Distribution::WeightList WeightList;
+typedef BlockFrequencyInfoImplBase::Float Float;
+typedef BlockFrequencyInfoImplBase::PackagedLoopData PackagedLoopData;
+typedef BlockFrequencyInfoImplBase::Weight Weight;
+typedef BlockFrequencyInfoImplBase::FrequencyData FrequencyData;
+
+/// \brief Dithering mass distributer.
+///
+/// This class splits up a single mass into portions by weight, dithering to
+/// spread out error.  No mass is lost.  The dithering precision depends on the
+/// precision of the product of \a BlockMass and \a BranchProbability.
+///
+/// The distribution algorithm follows.
+///
+///  1. Initialize by saving the sum of the weights in \a RemWeight and the
+///     mass to distribute in \a RemMass.
+///
+///  2. For each portion:
+///
+///      1. Construct a branch probability, P, as the portion's weight divided
+///         by the current value of \a RemWeight.
+///      2. Calculate the portion's mass as \a RemMass times P.
+///      3. Update \a RemWeight and \a RemMass at each portion by subtracting
+///         the current portion's weight and mass.
+///
+/// Mass is distributed in two ways: full distribution and forward
+/// distribution.  The latter ignores backedges, and uses the parallel fields
+/// \a RemForwardWeight and \a RemForwardMass.
+struct DitheringDistributer {
+  uint32_t RemWeight;
+  uint32_t RemForwardWeight;
+
+  BlockMass RemMass;
+  BlockMass RemForwardMass;
+
+  DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
+
+  BlockMass takeLocalMass(uint32_t Weight) {
+    (void)takeMass(Weight);
+    return takeForwardMass(Weight);
+  }
+  BlockMass takeExitMass(uint32_t Weight) {
+    (void)takeForwardMass(Weight);
+    return takeMass(Weight);
+  }
+  BlockMass takeBackedgeMass(uint32_t Weight) { return takeMass(Weight); }
+
+private:
+  BlockMass takeForwardMass(uint32_t Weight);
+  BlockMass takeMass(uint32_t Weight);
+};
+}
+
+DitheringDistributer::DitheringDistributer(Distribution &Dist,
+                                           const BlockMass &Mass) {
+  Dist.normalize();
+  RemWeight = Dist.Total;
+  RemForwardWeight = Dist.ForwardTotal;
+  RemMass = Mass;
+  RemForwardMass = Dist.ForwardTotal ? Mass : BlockMass();
+}
+
+BlockMass DitheringDistributer::takeForwardMass(uint32_t Weight) {
+  // Compute the amount of mass to take.
+  assert(Weight && "invalid weight");
+  assert(Weight <= RemForwardWeight);
+  BlockMass Mass = RemForwardMass * BranchProbability(Weight, RemForwardWeight);
+
+  // Decrement totals (dither).
+  RemForwardWeight -= Weight;
+  RemForwardMass -= Mass;
+  return Mass;
+}
+BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
+  assert(Weight && "invalid weight");
+  assert(Weight <= RemWeight);
+  BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
+
+  // Decrement totals (dither).
+  RemWeight -= Weight;
+  RemMass -= Mass;
+  return Mass;
+}
+
+void Distribution::add(const BlockNode &Node, uint64_t Amount,
+                       Weight::DistType Type) {
+  assert(Amount && "invalid weight of 0");
+  uint64_t NewTotal = Total + Amount;
+
+  // Check for overflow.  It should be impossible to overflow twice.
+  bool IsOverflow = NewTotal < Total;
+  assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
+  DidOverflow |= IsOverflow;
+
+  // Update the total.
+  Total = NewTotal;
+
+  // Save the weight.
+  Weight W;
+  W.TargetNode = Node;
+  W.Amount = Amount;
+  W.Type = Type;
+  Weights.push_back(W);
+
+  if (Type == Weight::Backedge)
+    return;
+
+  // Update forward total.  Don't worry about overflow here, since then Total
+  // will exceed 32-bits and they'll both be recomputed in normalize().
+  ForwardTotal += Amount;
+}
+
+static void combineWeight(Weight &W, const Weight &OtherW) {
+  assert(OtherW.TargetNode.isValid());
+  if (!W.Amount) {
+    W = OtherW;
+    return;
+  }
+  assert(W.Type == OtherW.Type);
+  assert(W.TargetNode == OtherW.TargetNode);
+  assert(W.Amount < W.Amount + OtherW.Amount);
+  W.Amount += OtherW.Amount;
+}
+static void combineWeightsBySorting(WeightList &Weights) {
+  // Sort so edges to the same node are adjacent.
+  std::sort(Weights.begin(), Weights.end(),
+            [](const Weight &L,
+               const Weight &R) { return L.TargetNode < R.TargetNode; });
+
+  // Combine adjacent edges.
+  WeightList::iterator O = Weights.begin();
+  for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
+       ++O, (I = L)) {
+    *O = *I;
+
+    // Find the adjacent weights to the same node.
+    for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
+      combineWeight(*O, *L);
+  }
+
+  // Erase extra entries.
+  Weights.erase(O, Weights.end());
+  return;
+}
+static void combineWeightsByHashing(WeightList &Weights) {
+  // Collect weights into a DenseMap.
+  typedef DenseMap<BlockNode::IndexType, Weight> HashTable;
+  HashTable Combined(NextPowerOf2(2 * Weights.size()));
+  for (const Weight &W : Weights)
+    combineWeight(Combined[W.TargetNode.Index], W);
+
+  // Check whether anything changed.
+  if (Weights.size() == Combined.size())
+    return;
+
+  // Fill in the new weights.
+  Weights.clear();
+  Weights.reserve(Combined.size());
+  for (const auto &I : Combined)
+    Weights.push_back(I.second);
+}
+static void combineWeights(WeightList &Weights) {
+  // Use a hash table for many successors to keep this linear.
+  if (Weights.size() > 128) {
+    combineWeightsByHashing(Weights);
+    return;
+  }
+
+  combineWeightsBySorting(Weights);
+}
+static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
+  assert(Shift >= 0);
+  assert(Shift < 64);
+  if (!Shift)
+    return N;
+  return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
+}
+void Distribution::normalize() {
+  // Early exit for termination nodes.
+  if (Weights.empty())
+    return;
+
+  // Only bother if there are multiple successors.
+  if (Weights.size() > 1)
+    combineWeights(Weights);
+
+  // Early exit when combined into a single successor.
+  if (Weights.size() == 1) {
+    Total = 1;
+    ForwardTotal = Weights.front().Type != Weight::Backedge;
+    Weights.front().Amount = 1;
+    return;
+  }
+
+  // Determine how much to shift right so that the total fits into 32-bits.
+  //
+  // If we shift at all, shift by 1 extra.  Otherwise, the lower limit of 1
+  // for each weight can cause a 32-bit overflow.
+  int Shift = 0;
+  if (DidOverflow)
+    Shift = 33;
+  else if (Total > UINT32_MAX)
+    Shift = 33 - countLeadingZeros(Total);
+
+  // Early exit if nothing needs to be scaled.
+  if (!Shift)
+    return;
+
+  // Recompute the total through accumulation (rather than shifting it) so that
+  // it's accurate after shifting.  ForwardTotal is dirty here anyway.
+  Total = 0;
+  ForwardTotal = 0;
+
+  // Sum the weights to each node and shift right if necessary.
+  for (Weight &W : Weights) {
+    // Scale down below UINT32_MAX.  Since Shift is larger than necessary, we
+    // can round here without concern about overflow.
+    assert(W.TargetNode.isValid());
+    W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
+    assert(W.Amount <= UINT32_MAX);
+
+    // Update the total.
+    Total += W.Amount;
+    if (W.Type == Weight::Backedge)
+      continue;
+
+    // Update the forward total.
+    ForwardTotal += W.Amount;
+  }
+  assert(Total <= UINT32_MAX);
+}
+
+void BlockFrequencyInfoImplBase::clear() {
+  *this = BlockFrequencyInfoImplBase();
+}
+
+/// \brief Clear all memory not needed downstream.
+///
+/// Releases all memory not used downstream.  In particular, saves Freqs.
+static void cleanup(BlockFrequencyInfoImplBase &BFI) {
+  std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
+  BFI.clear();
+  BFI.Freqs = std::move(SavedFreqs);
+}
+
+/// \brief Get a possibly packaged node.
+///
+/// Get the node currently representing Node, which could be a containing
+/// loop.
+///
+/// This function should only be called when distributing mass.  As long as
+/// there are no irreducilbe edges to Node, then it will have complexity O(1)
+/// in this context.
+///
+/// In general, the complexity is O(L), where L is the number of loop headers
+/// Node has been packaged into.  Since this method is called in the context
+/// of distributing mass, L will be the number of loop headers an early exit
+/// edge jumps out of.
+static BlockNode getPackagedNode(const BlockFrequencyInfoImplBase &BFI,
+                                 const BlockNode &Node) {
+  assert(Node.isValid());
+  if (!BFI.Working[Node.Index].IsPackaged)
+    return Node;
+  if (!BFI.Working[Node.Index].ContainingLoop.isValid())
+    return Node;
+  return getPackagedNode(BFI, BFI.Working[Node.Index].ContainingLoop);
+}
+
+/// \brief Get the appropriate mass for a possible pseudo-node loop package.
+///
+/// Get appropriate mass for Node.  If Node is a loop-header (whose loop has
+/// been packaged), returns the mass of its pseudo-node.  If it's a node inside
+/// a packaged loop, it returns the loop's pseudo-node.
+static BlockMass &getPackageMass(BlockFrequencyInfoImplBase &BFI,
+                                 const BlockNode &Node) {
+  assert(Node.isValid());
+  assert(!BFI.Working[Node.Index].IsPackaged);
+  if (!BFI.Working[Node.Index].IsAPackage)
+    return BFI.Working[Node.Index].Mass;
+
+  return BFI.getLoopPackage(Node).Mass;
+}
+
+void BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
+                                           const BlockNode &LoopHead,
+                                           const BlockNode &Pred,
+                                           const BlockNode &Succ,
+                                           uint64_t Weight) {
+  if (!Weight)
+    Weight = 1;
+
+#ifndef NDEBUG
+  auto debugSuccessor = [&](const char *Type, const BlockNode &Resolved) {
+    dbgs() << "  =>"
+           << " [" << Type << "] weight = " << Weight;
+    if (Succ != LoopHead)
+      dbgs() << ", succ = " << getBlockName(Succ);
+    if (Resolved != Succ)
+      dbgs() << ", resolved = " << getBlockName(Resolved);
+    dbgs() << "\n";
+  };
+  (void)debugSuccessor;
+#endif
+
+  if (Succ == LoopHead) {
+    DEBUG(debugSuccessor("backedge", Succ));
+    Dist.addBackedge(LoopHead, Weight);
+    return;
+  }
+  BlockNode Resolved = getPackagedNode(*this, Succ);
+  assert(Resolved != LoopHead);
+
+  if (Working[Resolved.Index].ContainingLoop != LoopHead) {
+    DEBUG(debugSuccessor("  exit  ", Resolved));
+    Dist.addExit(Resolved, Weight);
+    return;
+  }
+
+  if (!LoopHead.isValid() && Resolved < Pred) {
+    // Irreducible backedge.  Skip this edge in the distribution.
+    DEBUG(debugSuccessor("skipped ", Resolved));
+    return;
+  }
+
+  DEBUG(debugSuccessor(" local  ", Resolved));
+  Dist.addLocal(Resolved, Weight);
+}
+
+void BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
+    const BlockNode &LoopHead, const BlockNode &LocalLoopHead,
+    Distribution &Dist) {
+  PackagedLoopData &LoopPackage = getLoopPackage(LocalLoopHead);
+  const PackagedLoopData::ExitMap &Exits = LoopPackage.Exits;
+
+  // Copy the exit map into Dist.
+  for (const auto &I : Exits)
+    addToDist(Dist, LoopHead, LocalLoopHead, I.first, I.second.getMass());
+
+  // We don't need this map any more.  Clear it to prevent quadratic memory
+  // usage in deeply nested loops with irreducible control flow.
+  LoopPackage.Exits.clear();
+}
+
+/// \brief Get the maximum allowed loop scale.
+///
+/// Gives the maximum number of estimated iterations allowed for a loop.
+/// Downstream users have trouble with very large numbers (even within
+/// 64-bits).  Perhaps they can be changed to use PositiveFloat.
+///
+/// TODO: change downstream users so that this can be increased or removed.
+static Float getMaxLoopScale() { return Float(1, 12); }
+
+/// \brief Compute the loop scale for a loop.
+void BlockFrequencyInfoImplBase::computeLoopScale(const BlockNode &LoopHead) {
+  // Compute loop scale.
+  DEBUG(dbgs() << "compute-loop-scale: " << getBlockName(LoopHead) << "\n");
+
+  // LoopScale == 1 / ExitMass
+  // ExitMass == HeadMass - BackedgeMass
+  PackagedLoopData &LoopPackage = getLoopPackage(LoopHead);
+  BlockMass ExitMass = BlockMass::getFull() - LoopPackage.BackedgeMass;
+
+  // Block scale stores the inverse of the scale.
+  LoopPackage.Scale = ExitMass.toFloat().inverse();
+
+  DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
+               << " - " << LoopPackage.BackedgeMass << ")\n"
+               << " - scale = " << LoopPackage.Scale << "\n");
+
+  if (LoopPackage.Scale > getMaxLoopScale()) {
+    LoopPackage.Scale = getMaxLoopScale();
+    DEBUG(dbgs() << " - reduced-to-max-scale: " << getMaxLoopScale() << "\n");
+  }
+}
+
+/// \brief Package up a loop.
+void BlockFrequencyInfoImplBase::packageLoop(const BlockNode &LoopHead) {
+  DEBUG(dbgs() << "packaging-loop: " << getBlockName(LoopHead) << "\n");
+  Working[LoopHead.Index].IsAPackage = true;
+  for (const BlockNode &M : getLoopPackage(LoopHead).Members) {
+    DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
+    Working[M.Index].IsPackaged = true;
+  }
+}
+
+void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
+                                                const BlockNode &LoopHead,
+                                                Distribution &Dist) {
+  BlockMass Mass = getPackageMass(*this, Source);
+  DEBUG(dbgs() << "  => mass:  " << Mass
+               << " (    general     |    forward     )\n");
+
+  // Distribute mass to successors as laid out in Dist.
+  DitheringDistributer D(Dist, Mass);
+
+#ifndef NDEBUG
+  auto debugAssign = [&](const BlockNode &T, const BlockMass &M,
+                         const char *Desc) {
+    dbgs() << "  => assign " << M << " (" << D.RemMass << "|"
+           << D.RemForwardMass << ")";
+    if (Desc)
+      dbgs() << " [" << Desc << "]";
+    if (T.isValid())
+      dbgs() << " to " << getBlockName(T);
+    dbgs() << "\n";
+  };
+  (void)debugAssign;
+#endif
+
+  PackagedLoopData *LoopPackage = 0;
+  if (LoopHead.isValid())
+    LoopPackage = &getLoopPackage(LoopHead);
+  for (const Weight &W : Dist.Weights) {
+    // Check for a local edge (forward and non-exit).
+    if (W.Type == Weight::Local) {
+      BlockMass Local = D.takeLocalMass(W.Amount);
+      getPackageMass(*this, W.TargetNode) += Local;
+      DEBUG(debugAssign(W.TargetNode, Local, nullptr));
+      continue;
+    }
+
+    // Backedges and exits only make sense if we're processing a loop.
+    assert(LoopPackage && "backedge or exit outside of loop");
+
+    // Check for a backedge.
+    if (W.Type == Weight::Backedge) {
+      BlockMass Back = D.takeBackedgeMass(W.Amount);
+      LoopPackage->BackedgeMass += Back;
+      DEBUG(debugAssign(BlockNode(), Back, "back"));
+      continue;
+    }
+
+    // This must be an exit.
+    assert(W.Type == Weight::Exit);
+    BlockMass Exit = D.takeExitMass(W.Amount);
+    LoopPackage->Exits.push_back(std::make_pair(W.TargetNode, Exit));
+    DEBUG(debugAssign(W.TargetNode, Exit, "exit"));
+  }
+}
+
+static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
+                                     const Float &Min, const Float &Max) {
+  // Scale the Factor to a size that creates integers.  Ideally, integers would
+  // be scaled so that Max == UINT64_MAX so that they can be best
+  // differentiated.  However, the register allocator currently deals poorly
+  // with large numbers.  Instead, push Min up a little from 1 to give some
+  // room to differentiate small, unequal numbers.
+  //
+  // TODO: fix issues downstream so that ScalingFactor can be Float(1,64)/Max.
+  Float ScalingFactor = Min.inverse();
+  if ((Max / Min).lg() < 60)
+    ScalingFactor <<= 3;
+
+  // Translate the floats to integers.
+  DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
+               << ", factor = " << ScalingFactor << "\n");
+  for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
+    Float Scaled = BFI.Freqs[Index].Floating * ScalingFactor;
+    BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
+    DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
+                 << BFI.Freqs[Index].Floating << ", scaled = " << Scaled
+                 << ", int = " << BFI.Freqs[Index].Integer << "\n");
+  }
+}
+
+static void scaleBlockData(BlockFrequencyInfoImplBase &BFI,
+                           const BlockNode &Node,
+                           const PackagedLoopData &Loop) {
+  Float F = Loop.Mass.toFloat() * Loop.Scale;
+
+  Float &Current = BFI.Freqs[Node.Index].Floating;
+  Float Updated = Current * F;
+
+  DEBUG(dbgs() << " - " << BFI.getBlockName(Node) << ": " << Current << " => "
+               << Updated << "\n");
+
+  Current = Updated;
+}
+
+/// \brief Unwrap a loop package.
+///
+/// Visits all the members of a loop, adjusting their BlockData according to
+/// the loop's pseudo-node.
+static void unwrapLoopPackage(BlockFrequencyInfoImplBase &BFI,
+                              const BlockNode &Head) {
+  assert(Head.isValid());
+
+  PackagedLoopData &LoopPackage = BFI.getLoopPackage(Head);
+  DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getBlockName(Head)
+               << ": mass = " << LoopPackage.Mass
+               << ", scale = " << LoopPackage.Scale << "\n");
+  scaleBlockData(BFI, Head, LoopPackage);
+
+  // Propagate the head scale through the loop.  Since members are visited in
+  // RPO, the head scale will be updated by the loop scale first, and then the
+  // final head scale will be used for updated the rest of the members.
+  for (const BlockNode &M : LoopPackage.Members) {
+    const FrequencyData &HeadData = BFI.Freqs[Head.Index];
+    FrequencyData &Freqs = BFI.Freqs[M.Index];
+    Float NewFreq = Freqs.Floating * HeadData.Floating;
+    DEBUG(dbgs() << " - " << BFI.getBlockName(M) << ": " << Freqs.Floating
+                 << " => " << NewFreq << "\n");
+    Freqs.Floating = NewFreq;
+  }
+}
+
+void BlockFrequencyInfoImplBase::finalizeMetrics() {
+  // Set initial frequencies from loop-local masses.
+  for (size_t Index = 0; Index < Working.size(); ++Index)
+    Freqs[Index].Floating = Working[Index].Mass.toFloat();
+
+  // Unwrap loop packages in reverse post-order, tracking min and max
+  // frequencies.
+  auto Min = Float::getLargest();
+  auto Max = Float::getZero();
+  for (size_t Index = 0; Index < Working.size(); ++Index) {
+    if (Working[Index].isLoopHeader())
+      unwrapLoopPackage(*this, BlockNode(Index));
+
+    // Update max scale.
+    Min = std::min(Min, Freqs[Index].Floating);
+    Max = std::max(Max, Freqs[Index].Floating);
+  }
+
+  // Convert to integers.
+  convertFloatingToInteger(*this, Min, Max);
+
+  // Clean up data structures.
+  cleanup(*this);
+
+  // Print out the final stats.
+  DEBUG(dump());
+}
+
+BlockFrequency
+BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
+  if (!Node.isValid())
+    return 0;
+  return Freqs[Node.Index].Integer;
+}
+Float
+BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
+  if (!Node.isValid())
+    return Float::getZero();
+  return Freqs[Node.Index].Floating;
+}
+
+std::string
+BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
+  return std::string();
+}
+
+raw_ostream &
+BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
+                                           const BlockNode &Node) const {
+  return OS << getFloatingBlockFreq(Node);
+}
+
+raw_ostream &
+BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
+                                           const BlockFrequency &Freq) const {
+  Float Block(Freq.getFrequency(), 0);
+  Float Entry(getEntryFreq(), 0);
+
+  return OS << Block / Entry;
+}

lib/Analysis/CMakeLists.txt

@@ -7,6 +7,7 @@ add_llvm_library(LLVMAnalysis
   Analysis.cpp
   BasicAliasAnalysis.cpp
   BlockFrequencyInfo.cpp
+  BlockFrequencyInfoImpl.cpp
   BranchProbabilityInfo.cpp
   CFG.cpp
   CFGPrinter.cpp

lib/CodeGen/MachineBlockFrequencyInfo.cpp

@@ -11,9 +11,12 @@
 //
 //===----------------------------------------------------------------------===//
 
+#define DEBUG_TYPE "block-freq"
 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Support/CommandLine.h"
@@ -112,6 +115,7 @@ struct DOTGraphTraits<MachineBlockFrequencyInfo*> :
 INITIALIZE_PASS_BEGIN(MachineBlockFrequencyInfo, "machine-block-freq",
                       "Machine Block Frequency Analysis", true, true)
 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
+INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
 INITIALIZE_PASS_END(MachineBlockFrequencyInfo, "machine-block-freq",
                     "Machine Block Frequency Analysis", true, true)
 
@@ -127,16 +131,18 @@ MachineBlockFrequencyInfo::~MachineBlockFrequencyInfo() {}
 
 void MachineBlockFrequencyInfo::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.addRequired<MachineBranchProbabilityInfo>();
+  AU.addRequired<MachineLoopInfo>();
   AU.setPreservesAll();
   MachineFunctionPass::getAnalysisUsage(AU);
 }
 
 bool MachineBlockFrequencyInfo::runOnMachineFunction(MachineFunction &F) {
   MachineBranchProbabilityInfo &MBPI =
-    getAnalysis<MachineBranchProbabilityInfo>();
+      getAnalysis<MachineBranchProbabilityInfo>();
+  MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
   if (!MBFI)
     MBFI.reset(new ImplType);
-  MBFI->doFunction(&F, &MBPI);
+  MBFI->doFunction(&F, &MBPI, &MLI);
 #ifndef NDEBUG
   if (ViewMachineBlockFreqPropagationDAG != GVDT_None) {
     view();
@@ -166,7 +172,7 @@ getBlockFreq(const MachineBasicBlock *MBB) const {
 }
 
 const MachineFunction *MachineBlockFrequencyInfo::getFunction() const {
-  return MBFI ? MBFI->Fn : nullptr;
+  return MBFI ? MBFI->getFunction() : nullptr;
 }
 
 raw_ostream &

test/Analysis/BlockFrequencyInfo/bad_input.ll

@@ -0,0 +1,50 @@
+; RUN: opt < %s -analyze -block-freq | FileCheck %s
+
+declare void @g(i32 %x)
+
+; CHECK-LABEL: Printing analysis {{.*}} for function 'branch_weight_0':
+; CHECK-NEXT: block-frequency-info: branch_weight_0
+define void @branch_weight_0(i32 %a) {
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
+entry:
+  br label %for.body
+
+; Check that we get 1,4 instead of 0,3.
+; CHECK-NEXT: for.body: float = 4.0,
+for.body:
+  %i = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  call void @g(i32 %i)
+  %inc = add i32 %i, 1
+  %cmp = icmp ugt i32 %inc, %a
+  br i1 %cmp, label %for.end, label %for.body, !prof !0
+
+; CHECK-NEXT: for.end: float = 1.0, int = [[ENTRY]]
+for.end:
+  ret void
+}
+
+!0 = metadata !{metadata !"branch_weights", i32 0, i32 3}
+
+; CHECK-LABEL: Printing analysis {{.*}} for function 'infinite_loop'
+; CHECK-NEXT: block-frequency-info: infinite_loop
+define void @infinite_loop(i1 %x) {
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
+entry:
+  br i1 %x, label %for.body, label %for.end, !prof !1
+
+; Check that the loop scale maxes out at 4096, giving 2048 here.
+; CHECK-NEXT: for.body: float = 2048.0,
+for.body:
+  %i = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  call void @g(i32 %i)
+  %inc = add i32 %i, 1
+  br label %for.body
+
+; Check that the exit weight is half of entry, since half is lost in the
+; infinite loop above.
+; CHECK-NEXT: for.end: float = 0.5,
+for.end:
+  ret void
+}
+
+!1 = metadata !{metadata !"branch_weights", i32 1, i32 1}

test/Analysis/BlockFrequencyInfo/basic.ll

@@ -1,13 +1,14 @@
 ; RUN: opt < %s -analyze -block-freq | FileCheck %s
 
 define i32 @test1(i32 %i, i32* %a) {
-; CHECK: Printing analysis {{.*}} for function 'test1'
+; CHECK-LABEL: Printing analysis {{.*}} for function 'test1':
-; CHECK: entry = 1.0
+; CHECK-NEXT: block-frequency-info: test1
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
 entry:
   br label %body
 
 ; Loop backedges are weighted and thus their bodies have a greater frequency.
-; CHECK: body = 32.0
+; CHECK-NEXT: body: float = 32.0,
 body:
   %iv = phi i32 [ 0, %entry ], [ %next, %body ]
   %base = phi i32 [ 0, %entry ], [ %sum, %body ]
@@ -18,29 +19,29 @@ body:
   %exitcond = icmp eq i32 %next, %i
   br i1 %exitcond, label %exit, label %body
 
-; CHECK: exit = 1.0
+; CHECK-NEXT: exit: float = 1.0, int = [[ENTRY]]
 exit:
   ret i32 %sum
 }
 
 define i32 @test2(i32 %i, i32 %a, i32 %b) {
-; CHECK: Printing analysis {{.*}} for function 'test2'
+; CHECK-LABEL: Printing analysis {{.*}} for function 'test2':
-; CHECK: entry = 1.0
+; CHECK-NEXT: block-frequency-info: test2
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
 entry:
   %cond = icmp ult i32 %i, 42
   br i1 %cond, label %then, label %else, !prof !0
 
 ; The 'then' branch is predicted more likely via branch weight metadata.
-; CHECK: then = 0.94116
+; CHECK-NEXT: then: float = 0.9411{{[0-9]*}},
 then:
   br label %exit
 
-; CHECK: else = 0.05877
+; CHECK-NEXT: else: float = 0.05882{{[0-9]*}},
 else:
   br label %exit
 
-; FIXME: It may be a bug that we don't sum back to 1.0.
+; CHECK-NEXT: exit: float = 1.0, int = [[ENTRY]]
-; CHECK: exit = 0.99993
 exit:
   %result = phi i32 [ %a, %then ], [ %b, %else ]
   ret i32 %result
@@ -49,37 +50,37 @@ exit:
 !0 = metadata !{metadata !"branch_weights", i32 64, i32 4}
 
 define i32 @test3(i32 %i, i32 %a, i32 %b, i32 %c, i32 %d, i32 %e) {
-; CHECK: Printing analysis {{.*}} for function 'test3'
+; CHECK-LABEL: Printing analysis {{.*}} for function 'test3':
-; CHECK: entry = 1.0
+; CHECK-NEXT: block-frequency-info: test3
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
 entry:
   switch i32 %i, label %case_a [ i32 1, label %case_b
                                  i32 2, label %case_c
                                  i32 3, label %case_d
                                  i32 4, label %case_e ], !prof !1
 
-; CHECK: case_a = 0.04998
+; CHECK-NEXT: case_a: float = 0.05,
 case_a:
   br label %exit
 
-; CHECK: case_b = 0.04998
+; CHECK-NEXT: case_b: float = 0.05,
 case_b:
   br label %exit
 
 ; The 'case_c' branch is predicted more likely via branch weight metadata.
-; CHECK: case_c = 0.79998
+; CHECK-NEXT: case_c: float = 0.8,
 case_c:
   br label %exit
 
-; CHECK: case_d = 0.04998
+; CHECK-NEXT: case_d: float = 0.05,
 case_d:
   br label %exit
 
-; CHECK: case_e = 0.04998
+; CHECK-NEXT: case_e: float = 0.05,
 case_e:
   br label %exit
 
-; FIXME: It may be a bug that we don't sum back to 1.0.
+; CHECK-NEXT: exit: float = 1.0, int = [[ENTRY]]
-; CHECK: exit = 0.99993
 exit:
   %result = phi i32 [ %a, %case_a ],
                     [ %b, %case_b ],
@@ -91,44 +92,50 @@ exit:
 
 !1 = metadata !{metadata !"branch_weights", i32 4, i32 4, i32 64, i32 4, i32 4}
 
-; CHECK: Printing analysis {{.*}} for function 'nested_loops'
-; CHECK: entry = 1.0
-; This test doesn't seem to be assigning sensible frequencies to nested loops.
 define void @nested_loops(i32 %a) {
+; CHECK-LABEL: Printing analysis {{.*}} for function 'nested_loops':
+; CHECK-NEXT: block-frequency-info: nested_loops
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
 entry:
   br label %for.cond1.preheader
 
+; CHECK-NEXT: for.cond1.preheader: float = 4001.0,
 for.cond1.preheader:
   %x.024 = phi i32 [ 0, %entry ], [ %inc12, %for.inc11 ]
   br label %for.cond4.preheader
 
+; CHECK-NEXT: for.cond4.preheader: float = 16008001.0,
 for.cond4.preheader:
   %y.023 = phi i32 [ 0, %for.cond1.preheader ], [ %inc9, %for.inc8 ]
   %add = add i32 %y.023, %x.024
   br label %for.body6
 
+; CHECK-NEXT: for.body6: float = 64048012001.0,
 for.body6:
   %z.022 = phi i32 [ 0, %for.cond4.preheader ], [ %inc, %for.body6 ]
   %add7 = add i32 %add, %z.022
-  tail call void @g(i32 %add7) #2
+  tail call void @g(i32 %add7)
   %inc = add i32 %z.022, 1
   %cmp5 = icmp ugt i32 %inc, %a
   br i1 %cmp5, label %for.inc8, label %for.body6, !prof !2
 
+; CHECK-NEXT: for.inc8: float = 16008001.0,
 for.inc8:
   %inc9 = add i32 %y.023, 1
   %cmp2 = icmp ugt i32 %inc9, %a
   br i1 %cmp2, label %for.inc11, label %for.cond4.preheader, !prof !2
 
+; CHECK-NEXT: for.inc11: float = 4001.0,
 for.inc11:
   %inc12 = add i32 %x.024, 1
   %cmp = icmp ugt i32 %inc12, %a
   br i1 %cmp, label %for.end13, label %for.cond1.preheader, !prof !2
 
+; CHECK-NEXT: for.end13: float = 1.0, int = [[ENTRY]]
 for.end13:
   ret void
 }
 
-declare void @g(i32) #1
+declare void @g(i32)
 
 !2 = metadata !{metadata !"branch_weights", i32 1, i32 4000}

test/Analysis/BlockFrequencyInfo/double_exit.ll

@@ -0,0 +1,165 @@
+; RUN: opt < %s -analyze -block-freq | FileCheck %s
+
+; CHECK-LABEL: Printing analysis {{.*}} for function 'double_exit':
+; CHECK-NEXT: block-frequency-info: double_exit
+define i32 @double_exit(i32 %N) {
+; Mass = 1
+; Frequency = 1
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
+entry:
+  br label %outer
+
+; Mass = 1
+; Backedge mass = 1/3, exit mass = 2/3
+; Loop scale = 3/2
+; Psuedo-edges = exit
+; Psuedo-mass = 1
+; Frequency = 1*3/2*1 = 3/2
+; CHECK-NEXT: outer: float = 1.5,
+outer:
+  %I.0 = phi i32 [ 0, %entry ], [ %inc6, %outer.inc ]
+  %Return.0 = phi i32 [ 0, %entry ], [ %Return.1, %outer.inc ]
+  %cmp = icmp slt i32 %I.0, %N
+  br i1 %cmp, label %inner, label %exit, !prof !2 ; 2:1
+
+; Mass = 1
+; Backedge mass = 3/5, exit mass = 2/5
+; Loop scale = 5/2
+; Pseudo-edges = outer.inc @ 1/5, exit @ 1/5
+; Pseudo-mass = 2/3
+; Frequency = 3/2*1*5/2*2/3 = 5/2
+; CHECK-NEXT: inner: float = 2.5,
+inner:
+  %Return.1 = phi i32 [ %Return.0, %outer ], [ %call4, %inner.inc ]
+  %J.0 = phi i32 [ %I.0, %outer ], [ %inc, %inner.inc ]
+  %cmp2 = icmp slt i32 %J.0, %N
+  br i1 %cmp2, label %inner.body, label %outer.inc, !prof !1 ; 4:1
+
+; Mass = 4/5
+; Frequency = 5/2*4/5 = 2
+; CHECK-NEXT: inner.body: float = 2.0,
+inner.body:
+  %call = call i32 @c2(i32 %I.0, i32 %J.0)
+  %tobool = icmp ne i32 %call, 0
+  br i1 %tobool, label %exit, label %inner.inc, !prof !0 ; 3:1
+
+; Mass = 3/5
+; Frequency = 5/2*3/5 = 3/2
+; CHECK-NEXT: inner.inc: float = 1.5,
+inner.inc:
+  %call4 = call i32 @logic2(i32 %Return.1, i32 %I.0, i32 %J.0)
+  %inc = add nsw i32 %J.0, 1
+  br label %inner
+
+; Mass = 1/3
+; Frequency = 3/2*1/3 = 1/2
+; CHECK-NEXT: outer.inc: float = 0.5,
+outer.inc:
+  %inc6 = add nsw i32 %I.0, 1
+  br label %outer
+
+; Mass = 1
+; Frequency = 1
+; CHECK-NEXT: exit: float = 1.0, int = [[ENTRY]]
+exit:
+  %Return.2 = phi i32 [ %Return.1, %inner.body ], [ %Return.0, %outer ]
+  ret i32 %Return.2
+}
+
+!0 = metadata !{metadata !"branch_weights", i32 1, i32 3}
+!1 = metadata !{metadata !"branch_weights", i32 4, i32 1}
+!2 = metadata !{metadata !"branch_weights", i32 2, i32 1}
+
+declare i32 @c2(i32, i32)
+declare i32 @logic2(i32, i32, i32)
+
+; CHECK-LABEL: Printing analysis {{.*}} for function 'double_exit_in_loop':
+; CHECK-NEXT: block-frequency-info: double_exit_in_loop
+define i32 @double_exit_in_loop(i32 %N) {
+; Mass = 1
+; Frequency = 1
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
+entry:
+  br label %outer
+
+; Mass = 1
+; Backedge mass = 1/2, exit mass = 1/2
+; Loop scale = 2
+; Pseudo-edges = exit
+; Psuedo-mass = 1
+; Frequency = 1*2*1 = 2
+; CHECK-NEXT: outer: float = 2.0,
+outer:
+  %I.0 = phi i32 [ 0, %entry ], [ %inc12, %outer.inc ]
+  %Return.0 = phi i32 [ 0, %entry ], [ %Return.3, %outer.inc ]
+  %cmp = icmp slt i32 %I.0, %N
+  br i1 %cmp, label %middle, label %exit, !prof !3 ; 1:1
+
+; Mass = 1
+; Backedge mass = 1/3, exit mass = 2/3
+; Loop scale = 3/2
+; Psuedo-edges = outer.inc
+; Psuedo-mass = 1/2
+; Frequency = 2*1*3/2*1/2 = 3/2
+; CHECK-NEXT: middle: float = 1.5,
+middle:
+  %J.0 = phi i32 [ %I.0, %outer ], [ %inc9, %middle.inc ]
+  %Return.1 = phi i32 [ %Return.0, %outer ], [ %Return.2, %middle.inc ]
+  %cmp2 = icmp slt i32 %J.0, %N
+  br i1 %cmp2, label %inner, label %outer.inc, !prof !2 ; 2:1
+
+; Mass = 1
+; Backedge mass = 3/5, exit mass = 2/5
+; Loop scale = 5/2
+; Pseudo-edges = middle.inc @ 1/5, outer.inc @ 1/5
+; Pseudo-mass = 2/3
+; Frequency = 3/2*1*5/2*2/3 = 5/2
+; CHECK-NEXT: inner: float = 2.5,
+inner:
+  %Return.2 = phi i32 [ %Return.1, %middle ], [ %call7, %inner.inc ]
+  %K.0 = phi i32 [ %J.0, %middle ], [ %inc, %inner.inc ]
+  %cmp5 = icmp slt i32 %K.0, %N
+  br i1 %cmp5, label %inner.body, label %middle.inc, !prof !1 ; 4:1
+
+; Mass = 4/5
+; Frequency = 5/2*4/5 = 2
+; CHECK-NEXT: inner.body: float = 2.0,
+inner.body:
+  %call = call i32 @c3(i32 %I.0, i32 %J.0, i32 %K.0)
+  %tobool = icmp ne i32 %call, 0
+  br i1 %tobool, label %outer.inc, label %inner.inc, !prof !0 ; 3:1
+
+; Mass = 3/5
+; Frequency = 5/2*3/5 = 3/2
+; CHECK-NEXT: inner.inc: float = 1.5,
+inner.inc:
+  %call7 = call i32 @logic3(i32 %Return.2, i32 %I.0, i32 %J.0, i32 %K.0)
+  %inc = add nsw i32 %K.0, 1
+  br label %inner
+
+; Mass = 1/3
+; Frequency = 3/2*1/3 = 1/2
+; CHECK-NEXT: middle.inc: float = 0.5,
+middle.inc:
+  %inc9 = add nsw i32 %J.0, 1
+  br label %middle
+
+; Mass = 1/2
+; Frequency = 2*1/2 = 1
+; CHECK-NEXT: outer.inc: float = 1.0,
+outer.inc:
+  %Return.3 = phi i32 [ %Return.2, %inner.body ], [ %Return.1, %middle ]
+  %inc12 = add nsw i32 %I.0, 1
+  br label %outer
+
+; Mass = 1
+; Frequency = 1
+; CHECK-NEXT: exit: float = 1.0, int = [[ENTRY]]
+exit:
+  ret i32 %Return.0
+}
+
+!3 = metadata !{metadata !"branch_weights", i32 1, i32 1}
+
+declare i32 @c3(i32, i32, i32)
+declare i32 @logic3(i32, i32, i32, i32)

test/Analysis/BlockFrequencyInfo/irreducible.ll

@@ -0,0 +1,197 @@
+; RUN: opt < %s -analyze -block-freq | FileCheck %s
+
+; A loop with multiple exits should be handled correctly.
+;
+; CHECK-LABEL: Printing analysis {{.*}} for function 'multiexit':
+; CHECK-NEXT: block-frequency-info: multiexit
+define void @multiexit(i32 %a) {
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
+entry:
+  br label %loop.1
+
+; CHECK-NEXT: loop.1: float = 1.333{{3*}},
+loop.1:
+  %i = phi i32 [ 0, %entry ], [ %inc.2, %loop.2 ]
+  call void @f(i32 %i)
+  %inc.1 = add i32 %i, 1
+  %cmp.1 = icmp ugt i32 %inc.1, %a
+  br i1 %cmp.1, label %exit.1, label %loop.2, !prof !0
+
+; CHECK-NEXT: loop.2: float = 0.666{{6*7}},
+loop.2:
+  call void @g(i32 %inc.1)
+  %inc.2 = add i32 %inc.1, 1
+  %cmp.2 = icmp ugt i32 %inc.2, %a
+  br i1 %cmp.2, label %exit.2, label %loop.1, !prof !1
+
+; CHECK-NEXT: exit.1: float = 0.666{{6*7}},
+exit.1:
+  call void @h(i32 %inc.1)
+  br label %return
+
+; CHECK-NEXT: exit.2: float = 0.333{{3*}},
+exit.2:
+  call void @i(i32 %inc.2)
+  br label %return
+
+; CHECK-NEXT: return: float = 1.0, int = [[ENTRY]]
+return:
+  ret void
+}
+
+declare void @f(i32 %x)
+declare void @g(i32 %x)
+declare void @h(i32 %x)
+declare void @i(i32 %x)
+
+!0 = metadata !{metadata !"branch_weights", i32 3, i32 3}
+!1 = metadata !{metadata !"branch_weights", i32 5, i32 5}
+
+; The current BlockFrequencyInfo algorithm doesn't handle multiple entrances
+; into a loop very well.  The frequencies assigned to blocks in the loop are
+; predictable (and not absurd), but also not correct and therefore not worth
+; testing.
+;
+; There are two testcases below.
+;
+; For each testcase, I use a CHECK-NEXT/NOT combo like an XFAIL with the
+; granularity of a single check.  If/when this behaviour is fixed, we'll know
+; about it, and the test should be updated.
+;
+; Testcase #1
+; ===========
+;
+; In this case c1 and c2 should have frequencies of 15/7 and 13/7,
+; respectively.  To calculate this, consider assigning 1.0 to entry, and
+; distributing frequency iteratively (to infinity).  At the first iteration,
+; entry gives 3/4 to c1 and 1/4 to c2.  At every step after, c1 and c2 give 3/4
+; of what they have to each other.  Somehow, all of it comes out to exit.
+;
+;       c1 = 3/4 + 1/4*3/4 + 3/4*3^2/4^2 + 1/4*3^3/4^3 + 3/4*3^3/4^3 + ...
+;       c2 = 1/4 + 3/4*3/4 + 1/4*3^2/4^2 + 3/4*3^3/4^3 + 1/4*3^3/4^3 + ...
+;
+; Simplify by splitting up the odd and even terms of the series and taking out
+; factors so that the infite series matches:
+;
+;       c1 =  3/4 *(9^0/16^0 + 9^1/16^1 + 9^2/16^2 + ...)
+;          +  3/16*(9^0/16^0 + 9^1/16^1 + 9^2/16^2 + ...)
+;       c2 =  1/4 *(9^0/16^0 + 9^1/16^1 + 9^2/16^2 + ...)
+;          +  9/16*(9^0/16^0 + 9^1/16^1 + 9^2/16^2 + ...)
+;
+;       c1 = 15/16*(9^0/16^0 + 9^1/16^1 + 9^2/16^2 + ...)
+;       c2 = 13/16*(9^0/16^0 + 9^1/16^1 + 9^2/16^2 + ...)
+;
+; Since this geometric series sums to 16/7:
+;
+;       c1 = 15/7
+;       c2 = 13/7
+;
+; If we treat c1 and c2 as members of the same loop, the exit frequency of the
+; loop as a whole is 1/4, so the loop scale should be 4.  Summing c1 and c2
+; gives 28/7, or 4.0, which is nice confirmation of the math above.
+;
+; However, assuming c1 precedes c2 in reverse post-order, the current algorithm
+; returns 3/4 and 13/16, respectively.  LoopInfo ignores edges between loops
+; (and doesn't see any loops here at all), and -block-freq ignores the
+; irreducible edge from c2 to c1.
+;
+; CHECK-LABEL: Printing analysis {{.*}} for function 'multientry':
+; CHECK-NEXT: block-frequency-info: multientry
+define void @multientry(i32 %a) {
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
+entry:
+  %choose = call i32 @choose(i32 %a)
+  %compare = icmp ugt i32 %choose, %a
+  br i1 %compare, label %c1, label %c2, !prof !2
+
+; This is like a single-line XFAIL (see above).
+; CHECK-NEXT: c1:
+; CHECK-NOT: float = 2.142857{{[0-9]*}},
+c1:
+  %i1 = phi i32 [ %a, %entry ], [ %i2.inc, %c2 ]
+  %i1.inc = add i32 %i1, 1
+  %choose1 = call i32 @choose(i32 %i1)
+  %compare1 = icmp ugt i32 %choose1, %a
+  br i1 %compare1, label %c2, label %exit, !prof !2
+
+; This is like a single-line XFAIL (see above).
+; CHECK-NEXT: c2:
+; CHECK-NOT: float = 1.857142{{[0-9]*}},
+c2:
+  %i2 = phi i32 [ %a, %entry ], [ %i1.inc, %c1 ]
+  %i2.inc = add i32 %i2, 1
+  %choose2 = call i32 @choose(i32 %i2)
+  %compare2 = icmp ugt i32 %choose2, %a
+  br i1 %compare2, label %c1, label %exit, !prof !2
+
+; We still shouldn't lose any frequency.
+; CHECK-NEXT: exit: float = 1.0, int = [[ENTRY]]
+exit:
+  ret void
+}
+
+; Testcase #2
+; ===========
+;
+; In this case c1 and c2 should be treated as equals in a single loop.  The
+; exit frequency is 1/3, so the scaling factor for the loop should be 3.0.  The
+; loop is entered 2/3 of the time, and c1 and c2 split the total loop frequency
+; evenly (1/2), so they should each have frequencies of 1.0 (3.0*2/3*1/2).
+; Another way of computing this result is by assigning 1.0 to entry and showing
+; that c1 and c2 should accumulate frequencies of:
+;
+;       1/3   +   2/9   +   4/27  +   8/81  + ...
+;     2^0/3^1 + 2^1/3^2 + 2^2/3^3 + 2^3/3^4 + ...
+;
+; At the first step, c1 and c2 each get 1/3 of the entry.  At each subsequent
+; step, c1 and c2 each get 1/3 of what's left in c1 and c2 combined.  This
+; infinite series sums to 1.
+;
+; However, assuming c1 precedes c2 in reverse post-order, the current algorithm
+; returns 1/2 and 3/4, respectively.  LoopInfo ignores edges between loops (and
+; treats c1 and c2 as self-loops only), and -block-freq ignores the irreducible
+; edge from c2 to c1.
+;
+; Below I use a CHECK-NEXT/NOT combo like an XFAIL with the granularity of a
+; single check.  If/when this behaviour is fixed, we'll know about it, and the
+; test should be updated.
+;
+; CHECK-LABEL: Printing analysis {{.*}} for function 'crossloops':
+; CHECK-NEXT: block-frequency-info: crossloops
+define void @crossloops(i32 %a) {
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
+entry:
+  %choose = call i32 @choose(i32 %a)
+  switch i32 %choose, label %exit [ i32 1, label %c1
+                                    i32 2, label %c2 ], !prof !3
+
+; This is like a single-line XFAIL (see above).
+; CHECK-NEXT: c1:
+; CHECK-NOT: float = 1.0,
+c1:
+  %i1 = phi i32 [ %a, %entry ], [ %i1.inc, %c1 ], [ %i2.inc, %c2 ]
+  %i1.inc = add i32 %i1, 1
+  %choose1 = call i32 @choose(i32 %i1)
+  switch i32 %choose1, label %exit [ i32 1, label %c1
+                                     i32 2, label %c2 ], !prof !3
+
+; This is like a single-line XFAIL (see above).
+; CHECK-NEXT: c2:
+; CHECK-NOT: float = 1.0,
+c2:
+  %i2 = phi i32 [ %a, %entry ], [ %i1.inc, %c1 ], [ %i2.inc, %c2 ]
+  %i2.inc = add i32 %i2, 1
+  %choose2 = call i32 @choose(i32 %i2)
+  switch i32 %choose2, label %exit [ i32 1, label %c1
+                                     i32 2, label %c2 ], !prof !3
+
+; We still shouldn't lose any frequency.
+; CHECK-NEXT: exit: float = 1.0, int = [[ENTRY]]
+exit:
+  ret void
+}
+
+declare i32 @choose(i32)
+
+!2 = metadata !{metadata !"branch_weights", i32 3, i32 1}
+!3 = metadata !{metadata !"branch_weights", i32 2, i32 2, i32 2}

test/Analysis/BlockFrequencyInfo/loop_with_branch.ll

@@ -0,0 +1,44 @@
+; RUN: opt < %s -analyze -block-freq | FileCheck %s
+
+; CHECK-LABEL: Printing analysis {{.*}} for function 'loop_with_branch':
+; CHECK-NEXT: block-frequency-info: loop_with_branch
+define void @loop_with_branch(i32 %a) {
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
+entry:
+  %skip_loop = call i1 @foo0(i32 %a)
+  br i1 %skip_loop, label %skip, label %header, !prof !0
+
+; CHECK-NEXT: skip: float = 0.25,
+skip:
+  br label %exit
+
+; CHECK-NEXT: header: float = 4.5,
+header:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %back ]
+  %i.next = add i32 %i, 1
+  %choose = call i2 @foo1(i32 %i)
+  switch i2 %choose, label %exit [ i2 0, label %left
+                                   i2 1, label %right ], !prof !1
+
+; CHECK-NEXT: left: float = 1.5,
+left:
+  br label %back
+
+; CHECK-NEXT: right: float = 2.25,
+right:
+  br label %back
+
+; CHECK-NEXT: back: float = 3.75,
+back:
+  br label %header
+
+; CHECK-NEXT: exit: float = 1.0, int = [[ENTRY]]
+exit:
+  ret void
+}
+
+declare i1 @foo0(i32)
+declare i2 @foo1(i32)
+
+!0 = metadata !{metadata !"branch_weights", i32 1, i32 3}
+!1 = metadata !{metadata !"branch_weights", i32 1, i32 2, i32 3}

test/Analysis/BlockFrequencyInfo/nested_loop_with_branches.ll

@@ -0,0 +1,59 @@
+; RUN: opt < %s -analyze -block-freq | FileCheck %s
+
+; CHECK-LABEL: Printing analysis {{.*}} for function 'nested_loop_with_branches'
+; CHECK-NEXT: block-frequency-info: nested_loop_with_branches
+define void @nested_loop_with_branches(i32 %a) {
+; CHECK-NEXT: entry: float = 1.0, int = [[ENTRY:[0-9]+]]
+entry:
+  %v0 = call i1 @foo0(i32 %a)
+  br i1 %v0, label %exit, label %outer, !prof !0
+
+; CHECK-NEXT: outer: float = 12.0,
+outer:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %inner.end ], [ %i.next, %no_inner ]
+  %i.next = add i32 %i, 1
+  %do_inner = call i1 @foo1(i32 %i)
+  br i1 %do_inner, label %no_inner, label %inner, !prof !0
+
+; CHECK-NEXT: inner: float = 36.0,
+inner:
+  %j = phi i32 [ 0, %outer ], [ %j.next, %inner.end ]
+  %side = call i1 @foo3(i32 %j)
+  br i1 %side, label %left, label %right, !prof !0
+
+; CHECK-NEXT: left: float = 9.0,
+left:
+  %v4 = call i1 @foo4(i32 %j)
+  br label %inner.end
+
+; CHECK-NEXT: right: float = 27.0,
+right:
+  %v5 = call i1 @foo5(i32 %j)
+  br label %inner.end
+
+; CHECK-NEXT: inner.end: float = 36.0,
+inner.end:
+  %stay_inner = phi i1 [ %v4, %left ], [ %v5, %right ]
+  %j.next = add i32 %j, 1
+  br i1 %stay_inner, label %inner, label %outer, !prof !1
+
+; CHECK-NEXT: no_inner: float = 3.0,
+no_inner:
+  %continue = call i1 @foo6(i32 %i)
+  br i1 %continue, label %outer, label %exit, !prof !1
+
+; CHECK-NEXT: exit: float = 1.0, int = [[ENTRY]]
+exit:
+  ret void
+}
+
+declare i1 @foo0(i32)
+declare i1 @foo1(i32)
+declare i1 @foo2(i32)
+declare i1 @foo3(i32)
+declare i1 @foo4(i32)
+declare i1 @foo5(i32)
+declare i1 @foo6(i32)
+
+!0 = metadata !{metadata !"branch_weights", i32 1, i32 3}
+!1 = metadata !{metadata !"branch_weights", i32 3, i32 1}

test/CodeGen/XCore/llvm-intrinsics.ll

@@ -287,9 +287,8 @@ define void @Unwind1() {
 ; CHECKFP: .LBB{{[0-9_]+}}
 ; CHECKFP-NEXT: ldc r2, 40
 ; CHECKFP-NEXT: add r2, r10, r2
-; CHECKFP-NEXT: add r0, r2, r0
+; CHECKFP-NEXT: add r2, r2, r0
 ; CHECKFP-NEXT: mov r3, r1
-; CHECKFP-NEXT: mov r2, r0
 ; CHECKFP-NEXT: ldw r9, r10[4]
 ; CHECKFP-NEXT: ldw r8, r10[5]
 ; CHECKFP-NEXT: ldw r7, r10[6]
@@ -337,9 +336,8 @@ define void @Unwind1() {
 ; CHECK-NEXT: ldc r2, 36
 ; CHECK-NEXT: ldaw r3, sp[0]
 ; CHECK-NEXT: add r2, r3, r2
-; CHECK-NEXT: add r0, r2, r0
+; CHECK-NEXT: add r2, r2, r0
 ; CHECK-NEXT: mov r3, r1
-; CHECK-NEXT: mov r2, r0
 ; CHECK-NEXT: ldw r10, sp[2]
 ; CHECK-NEXT: ldw r9, sp[3]
 ; CHECK-NEXT: ldw r8, sp[4]