//===----------------------------------------------------------------------===//
#include "llvm/Analysis/BlockFrequencyInfoImpl.h"
-#include "llvm/ADT/APFloat.h"
#include "llvm/ADT/SCCIterator.h"
#include "llvm/Support/raw_ostream.h"
-#include <deque>
+#include <numeric>
using namespace llvm;
using namespace llvm::bfi_detail;
#define DEBUG_TYPE "block-freq"
-//===----------------------------------------------------------------------===//
-//
-// ScaledNumber implementation.
-//
-//===----------------------------------------------------------------------===//
-#ifndef _MSC_VER
-const int32_t ScaledNumberBase::MaxScale;
-const int32_t ScaledNumberBase::MinScale;
-#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 >= ScaledNumberBase::MinScale);
- assert(E <= ScaledNumberBase::MaxScale);
-
- // Find a new E, but don't let it increase past MaxScale.
- int LeadingZeros = ScaledNumberBase::countLeadingZeros64(D);
- int NewE = std::min(ScaledNumberBase::MaxScale, E + 63 - LeadingZeros);
- int Shift = 63 - (NewE - E);
- assert(Shift <= LeadingZeros);
- assert(Shift == LeadingZeros || NewE == ScaledNumberBase::MaxScale);
- D <<= Shift;
- E = NewE;
-
- // Check for a denormal.
- unsigned AdjustedE = E + 16383;
- if (!(D >> 63)) {
- assert(E == ScaledNumberBase::MaxScale);
- 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 ScaledNumberBase::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 &ScaledNumberBase::print(raw_ostream &OS, uint64_t D, int16_t E,
- int Width, unsigned Precision) {
- return OS << toString(D, E, Width, Precision);
-}
-
-void ScaledNumberBase::dump(uint64_t D, int16_t E, int Width) {
- print(dbgs(), D, E, Width, 0) << "[" << Width << ":" << D << "*2^" << E
- << "]";
-}
-
-//===----------------------------------------------------------------------===//
-//
-// BlockMass implementation.
-//
-//===----------------------------------------------------------------------===//
-ScaledNumber<uint64_t> BlockMass::toFloat() const {
+ScaledNumber<uint64_t> BlockMass::toScaled() const {
if (isFull())
return ScaledNumber<uint64_t>(1, 0);
return ScaledNumber<uint64_t>(getMass() + 1, -64);
return OS;
}
-//===----------------------------------------------------------------------===//
-//
-// BlockFrequencyInfoImpl implementation.
-//
-//===----------------------------------------------------------------------===//
namespace {
typedef BlockFrequencyInfoImplBase::BlockNode BlockNode;
typedef BlockFrequencyInfoImplBase::Distribution Distribution;
typedef BlockFrequencyInfoImplBase::Distribution::WeightList WeightList;
-typedef BlockFrequencyInfoImplBase::Float Float;
+typedef BlockFrequencyInfoImplBase::Scaled64 Scaled64;
typedef BlockFrequencyInfoImplBase::LoopData LoopData;
typedef BlockFrequencyInfoImplBase::Weight Weight;
typedef BlockFrequencyInfoImplBase::FrequencyData FrequencyData;
BlockMass takeMass(uint32_t Weight);
};
-}
+
+} // end namespace
DitheringDistributer::DitheringDistributer(Distribution &Dist,
const BlockMass &Mass) {
Total = NewTotal;
// Save the weight.
- Weight W;
- W.TargetNode = Node;
- W.Amount = Amount;
- W.Type = Type;
- Weights.push_back(W);
+ Weights.push_back(Weight(Type, Node, Amount));
}
static void combineWeight(Weight &W, const Weight &OtherW) {
}
assert(W.Type == OtherW.Type);
assert(W.TargetNode == OtherW.TargetNode);
- assert(W.Amount < W.Amount + OtherW.Amount && "Unexpected overflow");
- W.Amount += OtherW.Amount;
+ assert(OtherW.Amount && "Expected non-zero weight");
+ if (W.Amount > W.Amount + OtherW.Amount)
+ // Saturate on overflow.
+ W.Amount = UINT64_MAX;
+ else
+ W.Amount += OtherW.Amount;
}
static void combineWeightsBySorting(WeightList &Weights) {
// Sort so edges to the same node are adjacent.
Shift = 33 - countLeadingZeros(Total);
// Early exit if nothing needs to be scaled.
- if (!Shift)
+ if (!Shift) {
+ // If we didn't overflow then combineWeights() shouldn't have changed the
+ // sum of the weights, but let's double-check.
+ assert(Total == std::accumulate(Weights.begin(), Weights.end(), UINT64_C(0),
+ [](uint64_t Sum, const Weight &W) {
+ return Sum + W.Amount;
+ }) &&
+ "Expected total to be correct");
return;
+ }
// Recompute the total through accumulation (rather than shifting it) so that
- // it's accurate after shifting.
+ // it's accurate after shifting and any changes combineWeights() made above.
Total = 0;
// Sum the weights to each node and shift right if necessary.
if (isLoopHeader(Resolved)) {
DEBUG(debugSuccessor("backedge"));
- Dist.addBackedge(OuterLoop->getHeader(), Weight);
+ Dist.addBackedge(Resolved, Weight);
return true;
}
return true;
}
-/// \brief Get the maximum allowed loop scale.
-///
-/// Gives the maximum number of estimated iterations allowed for a loop. Very
-/// large numbers cause problems downstream (even within 64-bits).
-static Float getMaxLoopScale() { return Float(1, 12); }
-
/// \brief Compute the loop scale for a loop.
void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
// Compute loop scale.
DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");
+ // Infinite loops need special handling. If we give the back edge an infinite
+ // mass, they may saturate all the other scales in the function down to 1,
+ // making all the other region temperatures look exactly the same. Choose an
+ // arbitrary scale to avoid these issues.
+ //
+ // FIXME: An alternate way would be to select a symbolic scale which is later
+ // replaced to be the maximum of all computed scales plus 1. This would
+ // appropriately describe the loop as having a large scale, without skewing
+ // the final frequency computation.
+ const Scaled64 InifiniteLoopScale(1, 12);
+
// LoopScale == 1 / ExitMass
// ExitMass == HeadMass - BackedgeMass
- BlockMass ExitMass = BlockMass::getFull() - Loop.BackedgeMass;
+ BlockMass TotalBackedgeMass;
+ for (auto &Mass : Loop.BackedgeMass)
+ TotalBackedgeMass += Mass;
+ BlockMass ExitMass = BlockMass::getFull() - TotalBackedgeMass;
- // Block scale stores the inverse of the scale.
- Loop.Scale = ExitMass.toFloat().inverse();
+ // Block scale stores the inverse of the scale. If this is an infinite loop,
+ // its exit mass will be zero. In this case, use an arbitrary scale for the
+ // loop scale.
+ Loop.Scale =
+ ExitMass.isEmpty() ? InifiniteLoopScale : ExitMass.toScaled().inverse();
DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
- << " - " << Loop.BackedgeMass << ")\n"
+ << " - " << TotalBackedgeMass << ")\n"
<< " - scale = " << Loop.Scale << "\n");
-
- if (Loop.Scale > getMaxLoopScale()) {
- Loop.Scale = getMaxLoopScale();
- DEBUG(dbgs() << " - reduced-to-max-scale: " << getMaxLoopScale() << "\n");
- }
}
/// \brief Package up a loop.
Loop.IsPackaged = true;
}
+#ifndef NDEBUG
+static void debugAssign(const BlockFrequencyInfoImplBase &BFI,
+ const DitheringDistributer &D, const BlockNode &T,
+ const BlockMass &M, const char *Desc) {
+ dbgs() << " => assign " << M << " (" << D.RemMass << ")";
+ if (Desc)
+ dbgs() << " [" << Desc << "]";
+ if (T.isValid())
+ dbgs() << " to " << BFI.getBlockName(T);
+ dbgs() << "\n";
+}
+#endif
+
void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
LoopData *OuterLoop,
Distribution &Dist) {
// 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 << ")";
- if (Desc)
- dbgs() << " [" << Desc << "]";
- if (T.isValid())
- dbgs() << " to " << getBlockName(T);
- dbgs() << "\n";
- };
- (void)debugAssign;
-#endif
-
for (const Weight &W : Dist.Weights) {
// Check for a local edge (non-backedge and non-exit).
BlockMass Taken = D.takeMass(W.Amount);
if (W.Type == Weight::Local) {
Working[W.TargetNode.Index].getMass() += Taken;
- DEBUG(debugAssign(W.TargetNode, Taken, nullptr));
+ DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
continue;
}
// Check for a backedge.
if (W.Type == Weight::Backedge) {
- OuterLoop->BackedgeMass += Taken;
- DEBUG(debugAssign(BlockNode(), Taken, "back"));
+ OuterLoop->BackedgeMass[OuterLoop->getHeaderIndex(W.TargetNode)] += Taken;
+ DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "back"));
continue;
}
// This must be an exit.
assert(W.Type == Weight::Exit);
OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
- DEBUG(debugAssign(W.TargetNode, Taken, "exit"));
+ DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "exit"));
}
}
static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
- const Float &Min, const Float &Max) {
+ const Scaled64 &Min, const Scaled64 &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)
+ // differentiated. However, in the presence of large frequency values, small
+ // frequencies are scaled down to 1, making it impossible to differentiate
+ // small, unequal numbers. When the spread between Min and Max frequencies
+ // fits well within MaxBits, we make the scale be at least 8.
+ const unsigned MaxBits = 64;
+ const unsigned SpreadBits = (Max / Min).lg();
+ Scaled64 ScalingFactor;
+ if (SpreadBits <= MaxBits - 3) {
+ // If the values are small enough, make the scaling factor at least 8 to
+ // allow distinguishing small values.
+ ScalingFactor = Min.inverse();
ScalingFactor <<= 3;
+ } else {
+ // If the values need more than MaxBits to be represented, saturate small
+ // frequency values down to 1 by using a scaling factor that benefits large
+ // frequency values.
+ ScalingFactor = Scaled64(1, MaxBits) / Max;
+ }
// 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;
+ Scaled64 Scaled = BFI.Freqs[Index].Scaled * 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
+ << BFI.Freqs[Index].Scaled << ", scaled = " << Scaled
<< ", int = " << BFI.Freqs[Index].Integer << "\n");
}
}
DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
<< ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
<< "\n");
- Loop.Scale *= Loop.Mass.toFloat();
+ Loop.Scale *= Loop.Mass.toScaled();
Loop.IsPackaged = false;
DEBUG(dbgs() << " => combined-scale = " << Loop.Scale << "\n");
// final head scale will be used for updated the rest of the members.
for (const BlockNode &N : Loop.Nodes) {
const auto &Working = BFI.Working[N.Index];
- Float &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
- : BFI.Freqs[N.Index].Floating;
- Float New = Loop.Scale * F;
+ Scaled64 &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
+ : BFI.Freqs[N.Index].Scaled;
+ Scaled64 New = Loop.Scale * F;
DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => " << New
<< "\n");
F = New;
void BlockFrequencyInfoImplBase::unwrapLoops() {
// Set initial frequencies from loop-local masses.
for (size_t Index = 0; Index < Working.size(); ++Index)
- Freqs[Index].Floating = Working[Index].Mass.toFloat();
+ Freqs[Index].Scaled = Working[Index].Mass.toScaled();
for (LoopData &Loop : Loops)
unwrapLoop(*this, Loop);
void BlockFrequencyInfoImplBase::finalizeMetrics() {
// Unwrap loop packages in reverse post-order, tracking min and max
// frequencies.
- auto Min = Float::getLargest();
- auto Max = Float::getZero();
+ auto Min = Scaled64::getLargest();
+ auto Max = Scaled64::getZero();
for (size_t Index = 0; Index < Working.size(); ++Index) {
// Update min/max scale.
- Min = std::min(Min, Freqs[Index].Floating);
- Max = std::max(Max, Freqs[Index].Floating);
+ Min = std::min(Min, Freqs[Index].Scaled);
+ Max = std::max(Max, Freqs[Index].Scaled);
}
// Convert to integers.
return 0;
return Freqs[Node.Index].Integer;
}
-Float
+Scaled64
BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
if (!Node.isValid())
- return Float::getZero();
- return Freqs[Node.Index].Floating;
+ return Scaled64::getZero();
+ return Freqs[Node.Index].Scaled;
+}
+
+void BlockFrequencyInfoImplBase::setBlockFreq(const BlockNode &Node,
+ uint64_t Freq) {
+ assert(Node.isValid() && "Expected valid node");
+ assert(Node.Index < Freqs.size() && "Expected legal index");
+ Freqs[Node.Index].Integer = Freq;
}
std::string
raw_ostream &
BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
const BlockFrequency &Freq) const {
- Float Block(Freq.getFrequency(), 0);
- Float Entry(getEntryFreq(), 0);
+ Scaled64 Block(Freq.getFrequency(), 0);
+ Scaled64 Entry(getEntryFreq(), 0);
return OS << Block / Entry;
}
break;
}
}
- assert(Headers.size() >= 2 && "Should be irreducible");
+ assert(Headers.size() >= 2 &&
+ "Expected irreducible CFG; -loop-info is likely invalid");
if (Headers.size() == InSCC.size()) {
// Every block is a header.
std::sort(Headers.begin(), Headers.end());
void
BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
OuterLoop.Exits.clear();
- OuterLoop.BackedgeMass = BlockMass::getEmpty();
+ for (auto &Mass : OuterLoop.BackedgeMass)
+ Mass = BlockMass::getEmpty();
auto O = OuterLoop.Nodes.begin() + 1;
for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
if (!Working[I->Index].isPackaged())
*O++ = *I;
OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());
}
+
+void BlockFrequencyInfoImplBase::adjustLoopHeaderMass(LoopData &Loop) {
+ assert(Loop.isIrreducible() && "this only makes sense on irreducible loops");
+
+ // Since the loop has more than one header block, the mass flowing back into
+ // each header will be different. Adjust the mass in each header loop to
+ // reflect the masses flowing through back edges.
+ //
+ // To do this, we distribute the initial mass using the backedge masses
+ // as weights for the distribution.
+ BlockMass LoopMass = BlockMass::getFull();
+ Distribution Dist;
+
+ DEBUG(dbgs() << "adjust-loop-header-mass:\n");
+ for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
+ auto &HeaderNode = Loop.Nodes[H];
+ auto &BackedgeMass = Loop.BackedgeMass[Loop.getHeaderIndex(HeaderNode)];
+ DEBUG(dbgs() << " - Add back edge mass for node "
+ << getBlockName(HeaderNode) << ": " << BackedgeMass << "\n");
+ if (BackedgeMass.getMass() > 0)
+ Dist.addLocal(HeaderNode, BackedgeMass.getMass());
+ else
+ DEBUG(dbgs() << " Nothing added. Back edge mass is zero\n");
+ }
+
+ DitheringDistributer D(Dist, LoopMass);
+
+ DEBUG(dbgs() << " Distribute loop mass " << LoopMass
+ << " to headers using above weights\n");
+ for (const Weight &W : Dist.Weights) {
+ BlockMass Taken = D.takeMass(W.Amount);
+ assert(W.Type == Weight::Local && "all weights should be local");
+ Working[W.TargetNode.Index].getMass() = Taken;
+ DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
+ }
+}