llvm-6502/include/llvm/Analysis/BlockFrequencyInfoImpl.h

1551 lines
54 KiB
C
Raw Normal View History

//==- BlockFrequencyInfoImpl.h - Block Frequency Implementation -*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// Shared implementation of BlockFrequency for IR and Machine Instructions.
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
#define LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/Support/BlockFrequency.h"
#include "llvm/Support/BranchProbability.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include <string>
#include <vector>
#define DEBUG_TYPE "block-freq"
//===----------------------------------------------------------------------===//
//
// UnsignedFloat definition.
//
// TODO: Make this private to BlockFrequencyInfoImpl or delete.
//
//===----------------------------------------------------------------------===//
namespace llvm {
class UnsignedFloatBase {
public:
static const int32_t MaxExponent = 16383;
static const int32_t MinExponent = -16382;
static const int DefaultPrecision = 10;
static void dump(uint64_t D, int16_t E, int Width);
static raw_ostream &print(raw_ostream &OS, uint64_t D, int16_t E, int Width,
unsigned Precision);
static std::string toString(uint64_t D, int16_t E, int Width,
unsigned Precision);
static int countLeadingZeros32(uint32_t N) { return countLeadingZeros(N); }
static int countLeadingZeros64(uint64_t N) { return countLeadingZeros(N); }
static uint64_t getHalf(uint64_t N) { return (N >> 1) + (N & 1); }
static std::pair<uint64_t, bool> splitSigned(int64_t N) {
if (N >= 0)
return std::make_pair(N, false);
uint64_t Unsigned = N == INT64_MIN ? UINT64_C(1) << 63 : uint64_t(-N);
return std::make_pair(Unsigned, true);
}
static int64_t joinSigned(uint64_t U, bool IsNeg) {
if (U > uint64_t(INT64_MAX))
return IsNeg ? INT64_MIN : INT64_MAX;
return IsNeg ? -int64_t(U) : int64_t(U);
}
static int32_t extractLg(const std::pair<int32_t, int> &Lg) {
return Lg.first;
}
static int32_t extractLgFloor(const std::pair<int32_t, int> &Lg) {
return Lg.first - (Lg.second > 0);
}
static int32_t extractLgCeiling(const std::pair<int32_t, int> &Lg) {
return Lg.first + (Lg.second < 0);
}
static std::pair<uint64_t, int16_t> divide64(uint64_t L, uint64_t R);
static std::pair<uint64_t, int16_t> multiply64(uint64_t L, uint64_t R);
static int compare(uint64_t L, uint64_t R, int Shift) {
assert(Shift >= 0);
assert(Shift < 64);
uint64_t L_adjusted = L >> Shift;
if (L_adjusted < R)
return -1;
if (L_adjusted > R)
return 1;
return L > L_adjusted << Shift ? 1 : 0;
}
};
/// \brief Simple representation of an unsigned floating point.
///
/// UnsignedFloat is a unsigned floating point number. It uses simple
/// saturation arithmetic, and every operation is well-defined for every value.
///
/// The number is split into a signed exponent and unsigned digits. The number
/// represented is \c getDigits()*2^getExponent(). In this way, the digits are
/// much like the mantissa in the x87 long double, but there is no canonical
/// form, so the same number can be represented by many bit representations
/// (it's always in "denormal" mode).
///
/// UnsignedFloat is templated on the underlying integer type for digits, which
/// is expected to be one of uint64_t, uint32_t, uint16_t or uint8_t.
///
/// Unlike builtin floating point types, UnsignedFloat is portable.
///
/// Unlike APFloat, UnsignedFloat does not model architecture floating point
/// behaviour (this should make it a little faster), and implements most
/// operators (this makes it usable).
///
/// UnsignedFloat is totally ordered. However, there is no canonical form, so
/// there are multiple representations of most scalars. E.g.:
///
/// UnsignedFloat(8u, 0) == UnsignedFloat(4u, 1)
/// UnsignedFloat(4u, 1) == UnsignedFloat(2u, 2)
/// UnsignedFloat(2u, 2) == UnsignedFloat(1u, 3)
///
/// UnsignedFloat implements most arithmetic operations. Precision is kept
/// where possible. Uses simple saturation arithmetic, so that operations
/// saturate to 0.0 or getLargest() rather than under or overflowing. It has
/// some extra arithmetic for unit inversion. 0.0/0.0 is defined to be 0.0.
/// Any other division by 0.0 is defined to be getLargest().
///
/// As a convenience for modifying the exponent, left and right shifting are
/// both implemented, and both interpret negative shifts as positive shifts in
/// the opposite direction.
///
/// Exponents are limited to the range accepted by x87 long double. This makes
/// it trivial to add functionality to convert to APFloat (this is already
/// relied on for the implementation of printing).
///
/// The current plan is to gut this and make the necessary parts of it (even
/// more) private to BlockFrequencyInfo.
template <class DigitsT> class UnsignedFloat : UnsignedFloatBase {
public:
static_assert(!std::numeric_limits<DigitsT>::is_signed,
"only unsigned floats supported");
typedef DigitsT DigitsType;
private:
typedef std::numeric_limits<DigitsType> DigitsLimits;
static const int Width = sizeof(DigitsType) * 8;
static_assert(Width <= 64, "invalid integer width for digits");
private:
DigitsType Digits;
int16_t Exponent;
public:
UnsignedFloat() : Digits(0), Exponent(0) {}
UnsignedFloat(DigitsType Digits, int16_t Exponent)
: Digits(Digits), Exponent(Exponent) {}
private:
UnsignedFloat(const std::pair<uint64_t, int16_t> &X)
: Digits(X.first), Exponent(X.second) {}
public:
static UnsignedFloat getZero() { return UnsignedFloat(0, 0); }
static UnsignedFloat getOne() { return UnsignedFloat(1, 0); }
static UnsignedFloat getLargest() {
return UnsignedFloat(DigitsLimits::max(), MaxExponent);
}
static UnsignedFloat getFloat(uint64_t N) { return adjustToWidth(N, 0); }
static UnsignedFloat getInverseFloat(uint64_t N) {
return getFloat(N).invert();
}
static UnsignedFloat getFraction(DigitsType N, DigitsType D) {
return getQuotient(N, D);
}
int16_t getExponent() const { return Exponent; }
DigitsType getDigits() const { return Digits; }
/// \brief Convert to the given integer type.
///
/// Convert to \c IntT using simple saturating arithmetic, truncating if
/// necessary.
template <class IntT> IntT toInt() const;
bool isZero() const { return !Digits; }
bool isLargest() const { return *this == getLargest(); }
bool isOne() const {
if (Exponent > 0 || Exponent <= -Width)
return false;
return Digits == DigitsType(1) << -Exponent;
}
/// \brief The log base 2, rounded.
///
/// Get the lg of the scalar. lg 0 is defined to be INT32_MIN.
int32_t lg() const { return extractLg(lgImpl()); }
/// \brief The log base 2, rounded towards INT32_MIN.
///
/// Get the lg floor. lg 0 is defined to be INT32_MIN.
int32_t lgFloor() const { return extractLgFloor(lgImpl()); }
/// \brief The log base 2, rounded towards INT32_MAX.
///
/// Get the lg ceiling. lg 0 is defined to be INT32_MIN.
int32_t lgCeiling() const { return extractLgCeiling(lgImpl()); }
bool operator==(const UnsignedFloat &X) const { return compare(X) == 0; }
bool operator<(const UnsignedFloat &X) const { return compare(X) < 0; }
bool operator!=(const UnsignedFloat &X) const { return compare(X) != 0; }
bool operator>(const UnsignedFloat &X) const { return compare(X) > 0; }
bool operator<=(const UnsignedFloat &X) const { return compare(X) <= 0; }
bool operator>=(const UnsignedFloat &X) const { return compare(X) >= 0; }
bool operator!() const { return isZero(); }
/// \brief Convert to a decimal representation in a string.
///
/// Convert to a string. Uses scientific notation for very large/small
/// numbers. Scientific notation is used roughly for numbers outside of the
/// range 2^-64 through 2^64.
///
/// \c Precision indicates the number of decimal digits of precision to use;
/// 0 requests the maximum available.
///
/// As a special case to make debugging easier, if the number is small enough
/// to convert without scientific notation and has more than \c Precision
/// digits before the decimal place, it's printed accurately to the first
/// digit past zero. E.g., assuming 10 digits of precision:
///
/// 98765432198.7654... => 98765432198.8
/// 8765432198.7654... => 8765432198.8
/// 765432198.7654... => 765432198.8
/// 65432198.7654... => 65432198.77
/// 5432198.7654... => 5432198.765
std::string toString(unsigned Precision = DefaultPrecision) {
return UnsignedFloatBase::toString(Digits, Exponent, Width, Precision);
}
/// \brief Print a decimal representation.
///
/// Print a string. See toString for documentation.
raw_ostream &print(raw_ostream &OS,
unsigned Precision = DefaultPrecision) const {
return UnsignedFloatBase::print(OS, Digits, Exponent, Width, Precision);
}
void dump() const { return UnsignedFloatBase::dump(Digits, Exponent, Width); }
UnsignedFloat &operator+=(const UnsignedFloat &X);
UnsignedFloat &operator-=(const UnsignedFloat &X);
UnsignedFloat &operator*=(const UnsignedFloat &X);
UnsignedFloat &operator/=(const UnsignedFloat &X);
UnsignedFloat &operator<<=(int16_t Shift) { shiftLeft(Shift); return *this; }
UnsignedFloat &operator>>=(int16_t Shift) { shiftRight(Shift); return *this; }
private:
void shiftLeft(int32_t Shift);
void shiftRight(int32_t Shift);
/// \brief Adjust two floats to have matching exponents.
///
/// Adjust \c this and \c X to have matching exponents. Returns the new \c X
/// by value. Does nothing if \a isZero() for either.
///
/// The value that compares smaller will lose precision, and possibly become
/// \a isZero().
UnsignedFloat matchExponents(UnsignedFloat X);
/// \brief Increase exponent to match another float.
///
/// Increases \c this to have an exponent matching \c X. May decrease the
/// exponent of \c X in the process, and \c this may possibly become \a
/// isZero().
void increaseExponentToMatch(UnsignedFloat &X, int32_t ExponentDiff);
public:
/// \brief Scale a large number accurately.
///
/// Scale N (multiply it by this). Uses full precision multiplication, even
/// if Width is smaller than 64, so information is not lost.
uint64_t scale(uint64_t N) const;
uint64_t scaleByInverse(uint64_t N) const {
// TODO: implement directly, rather than relying on inverse. Inverse is
// expensive.
return inverse().scale(N);
}
int64_t scale(int64_t N) const {
std::pair<uint64_t, bool> Unsigned = splitSigned(N);
return joinSigned(scale(Unsigned.first), Unsigned.second);
}
int64_t scaleByInverse(int64_t N) const {
std::pair<uint64_t, bool> Unsigned = splitSigned(N);
return joinSigned(scaleByInverse(Unsigned.first), Unsigned.second);
}
int compare(const UnsignedFloat &X) const;
int compareTo(uint64_t N) const {
UnsignedFloat Float = getFloat(N);
int Compare = compare(Float);
if (Width == 64 || Compare != 0)
return Compare;
// Check for precision loss. We know *this == RoundTrip.
uint64_t RoundTrip = Float.template toInt<uint64_t>();
return N == RoundTrip ? 0 : RoundTrip < N ? -1 : 1;
}
int compareTo(int64_t N) const { return N < 0 ? 1 : compareTo(uint64_t(N)); }
UnsignedFloat &invert() { return *this = UnsignedFloat::getFloat(1) / *this; }
UnsignedFloat inverse() const { return UnsignedFloat(*this).invert(); }
private:
static UnsignedFloat getProduct(DigitsType L, DigitsType R);
static UnsignedFloat getQuotient(DigitsType Dividend, DigitsType Divisor);
std::pair<int32_t, int> lgImpl() const;
static int countLeadingZerosWidth(DigitsType Digits) {
if (Width == 64)
return countLeadingZeros64(Digits);
if (Width == 32)
return countLeadingZeros32(Digits);
return countLeadingZeros32(Digits) + Width - 32;
}
static UnsignedFloat adjustToWidth(uint64_t N, int32_t S) {
assert(S >= MinExponent);
assert(S <= MaxExponent);
if (Width == 64 || N <= DigitsLimits::max())
return UnsignedFloat(N, S);
// Shift right.
int Shift = 64 - Width - countLeadingZeros64(N);
DigitsType Shifted = N >> Shift;
// Round.
assert(S + Shift <= MaxExponent);
return getRounded(UnsignedFloat(Shifted, S + Shift),
N & UINT64_C(1) << (Shift - 1));
}
static UnsignedFloat getRounded(UnsignedFloat P, bool Round) {
if (!Round)
return P;
if (P.Digits == DigitsLimits::max())
// Careful of overflow in the exponent.
return UnsignedFloat(1, P.Exponent) <<= Width;
return UnsignedFloat(P.Digits + 1, P.Exponent);
}
};
#define UNSIGNED_FLOAT_BOP(op, base) \
template <class DigitsT> \
UnsignedFloat<DigitsT> operator op(const UnsignedFloat<DigitsT> &L, \
const UnsignedFloat<DigitsT> &R) { \
return UnsignedFloat<DigitsT>(L) base R; \
}
UNSIGNED_FLOAT_BOP(+, += )
UNSIGNED_FLOAT_BOP(-, -= )
UNSIGNED_FLOAT_BOP(*, *= )
UNSIGNED_FLOAT_BOP(/, /= )
UNSIGNED_FLOAT_BOP(<<, <<= )
UNSIGNED_FLOAT_BOP(>>, >>= )
#undef UNSIGNED_FLOAT_BOP
template <class DigitsT>
raw_ostream &operator<<(raw_ostream &OS, const UnsignedFloat<DigitsT> &X) {
return X.print(OS, 10);
}
#define UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, T1, T2) \
template <class DigitsT> \
bool operator op(const UnsignedFloat<DigitsT> &L, T1 R) { \
return L.compareTo(T2(R)) op 0; \
} \
template <class DigitsT> \
bool operator op(T1 L, const UnsignedFloat<DigitsT> &R) { \
return 0 op R.compareTo(T2(L)); \
}
#define UNSIGNED_FLOAT_COMPARE_TO(op) \
UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint64_t, uint64_t) \
UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint32_t, uint64_t) \
UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int64_t, int64_t) \
UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int32_t, int64_t)
UNSIGNED_FLOAT_COMPARE_TO(< )
UNSIGNED_FLOAT_COMPARE_TO(> )
UNSIGNED_FLOAT_COMPARE_TO(== )
UNSIGNED_FLOAT_COMPARE_TO(!= )
UNSIGNED_FLOAT_COMPARE_TO(<= )
UNSIGNED_FLOAT_COMPARE_TO(>= )
#undef UNSIGNED_FLOAT_COMPARE_TO
#undef UNSIGNED_FLOAT_COMPARE_TO_TYPE
template <class DigitsT>
uint64_t UnsignedFloat<DigitsT>::scale(uint64_t N) const {
if (Width == 64 || N <= DigitsLimits::max())
return (getFloat(N) * *this).template toInt<uint64_t>();
// Defer to the 64-bit version.
return UnsignedFloat<uint64_t>(Digits, Exponent).scale(N);
}
template <class DigitsT>
UnsignedFloat<DigitsT> UnsignedFloat<DigitsT>::getProduct(DigitsType L,
DigitsType R) {
// Check for zero.
if (!L || !R)
return getZero();
// Check for numbers that we can compute with 64-bit math.
if (Width <= 32 || (L <= UINT32_MAX && R <= UINT32_MAX))
return adjustToWidth(uint64_t(L) * uint64_t(R), 0);
// Do the full thing.
return UnsignedFloat(multiply64(L, R));
}
template <class DigitsT>
UnsignedFloat<DigitsT> UnsignedFloat<DigitsT>::getQuotient(DigitsType Dividend,
DigitsType Divisor) {
// Check for zero.
if (!Dividend)
return getZero();
if (!Divisor)
return getLargest();
if (Width == 64)
return UnsignedFloat(divide64(Dividend, Divisor));
// We can compute this with 64-bit math.
int Shift = countLeadingZeros64(Dividend);
uint64_t Shifted = uint64_t(Dividend) << Shift;
uint64_t Quotient = Shifted / Divisor;
// If Quotient needs to be shifted, then adjustToWidth will round.
if (Quotient > DigitsLimits::max())
return adjustToWidth(Quotient, -Shift);
// Round based on the value of the next bit.
return getRounded(UnsignedFloat(Quotient, -Shift),
Shifted % Divisor >= getHalf(Divisor));
}
template <class DigitsT>
template <class IntT>
IntT UnsignedFloat<DigitsT>::toInt() const {
typedef std::numeric_limits<IntT> Limits;
if (*this < 1)
return 0;
if (*this >= Limits::max())
return Limits::max();
IntT N = Digits;
if (Exponent > 0) {
assert(size_t(Exponent) < sizeof(IntT) * 8);
return N << Exponent;
}
if (Exponent < 0) {
assert(size_t(-Exponent) < sizeof(IntT) * 8);
return N >> -Exponent;
}
return N;
}
template <class DigitsT>
std::pair<int32_t, int> UnsignedFloat<DigitsT>::lgImpl() const {
if (isZero())
return std::make_pair(INT32_MIN, 0);
// Get the floor of the lg of Digits.
int32_t LocalFloor = Width - countLeadingZerosWidth(Digits) - 1;
// Get the floor of the lg of this.
int32_t Floor = Exponent + LocalFloor;
if (Digits == UINT64_C(1) << LocalFloor)
return std::make_pair(Floor, 0);
// Round based on the next digit.
assert(LocalFloor >= 1);
bool Round = Digits & UINT64_C(1) << (LocalFloor - 1);
return std::make_pair(Floor + Round, Round ? 1 : -1);
}
template <class DigitsT>
UnsignedFloat<DigitsT> UnsignedFloat<DigitsT>::matchExponents(UnsignedFloat X) {
if (isZero() || X.isZero() || Exponent == X.Exponent)
return X;
int32_t Diff = int32_t(X.Exponent) - int32_t(Exponent);
if (Diff > 0)
increaseExponentToMatch(X, Diff);
else
X.increaseExponentToMatch(*this, -Diff);
return X;
}
template <class DigitsT>
void UnsignedFloat<DigitsT>::increaseExponentToMatch(UnsignedFloat &X,
int32_t ExponentDiff) {
assert(ExponentDiff > 0);
if (ExponentDiff >= 2 * Width) {
*this = getZero();
return;
}
// Use up any leading zeros on X, and then shift this.
int32_t ShiftX = std::min(countLeadingZerosWidth(X.Digits), ExponentDiff);
assert(ShiftX < Width);
int32_t ShiftThis = ExponentDiff - ShiftX;
if (ShiftThis >= Width) {
*this = getZero();
return;
}
X.Digits <<= ShiftX;
X.Exponent -= ShiftX;
Digits >>= ShiftThis;
Exponent += ShiftThis;
return;
}
template <class DigitsT>
UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
operator+=(const UnsignedFloat &X) {
if (isLargest() || X.isZero())
return *this;
if (isZero() || X.isLargest())
return *this = X;
// Normalize exponents.
UnsignedFloat Scaled = matchExponents(X);
// Check for zero again.
if (isZero())
return *this = Scaled;
if (Scaled.isZero())
return *this;
// Compute sum.
DigitsType Sum = Digits + Scaled.Digits;
bool DidOverflow = Sum < Digits;
Digits = Sum;
if (!DidOverflow)
return *this;
if (Exponent == MaxExponent)
return *this = getLargest();
++Exponent;
Digits = UINT64_C(1) << (Width - 1) | Digits >> 1;
return *this;
}
template <class DigitsT>
UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
operator-=(const UnsignedFloat &X) {
if (X.isZero())
return *this;
if (*this <= X)
return *this = getZero();
// Normalize exponents.
UnsignedFloat Scaled = matchExponents(X);
assert(Digits >= Scaled.Digits);
// Compute difference.
if (!Scaled.isZero()) {
Digits -= Scaled.Digits;
return *this;
}
// Check if X just barely lost its last bit. E.g., for 32-bit:
//
// 1*2^32 - 1*2^0 == 0xffffffff != 1*2^32
if (*this == UnsignedFloat(1, X.lgFloor() + Width)) {
Digits = DigitsType(0) - 1;
--Exponent;
}
return *this;
}
template <class DigitsT>
UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
operator*=(const UnsignedFloat &X) {
if (isZero())
return *this;
if (X.isZero())
return *this = X;
// Save the exponents.
int32_t Exponents = int32_t(Exponent) + int32_t(X.Exponent);
// Get the raw product.
*this = getProduct(Digits, X.Digits);
// Combine with exponents.
return *this <<= Exponents;
}
template <class DigitsT>
UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
operator/=(const UnsignedFloat &X) {
if (isZero())
return *this;
if (X.isZero())
return *this = getLargest();
// Save the exponents.
int32_t Exponents = int32_t(Exponent) - int32_t(X.Exponent);
// Get the raw quotient.
*this = getQuotient(Digits, X.Digits);
// Combine with exponents.
return *this <<= Exponents;
}
template <class DigitsT>
void UnsignedFloat<DigitsT>::shiftLeft(int32_t Shift) {
if (!Shift || isZero())
return;
assert(Shift != INT32_MIN);
if (Shift < 0) {
shiftRight(-Shift);
return;
}
// Shift as much as we can in the exponent.
int32_t ExponentShift = std::min(Shift, MaxExponent - Exponent);
Exponent += ExponentShift;
if (ExponentShift == Shift)
return;
// Check this late, since it's rare.
if (isLargest())
return;
// Shift the digits themselves.
Shift -= ExponentShift;
if (Shift > countLeadingZerosWidth(Digits)) {
// Saturate.
*this = getLargest();
return;
}
Digits <<= Shift;
return;
}
template <class DigitsT>
void UnsignedFloat<DigitsT>::shiftRight(int32_t Shift) {
if (!Shift || isZero())
return;
assert(Shift != INT32_MIN);
if (Shift < 0) {
shiftLeft(-Shift);
return;
}
// Shift as much as we can in the exponent.
int32_t ExponentShift = std::min(Shift, Exponent - MinExponent);
Exponent -= ExponentShift;
if (ExponentShift == Shift)
return;
// Shift the digits themselves.
Shift -= ExponentShift;
if (Shift >= Width) {
// Saturate.
*this = getZero();
return;
}
Digits >>= Shift;
return;
}
template <class DigitsT>
int UnsignedFloat<DigitsT>::compare(const UnsignedFloat &X) const {
// Check for zero.
if (isZero())
return X.isZero() ? 0 : -1;
if (X.isZero())
return 1;
// Check for the scale. Use lgFloor to be sure that the exponent difference
// is always lower than 64.
int32_t lgL = lgFloor(), lgR = X.lgFloor();
if (lgL != lgR)
return lgL < lgR ? -1 : 1;
// Compare digits.
if (Exponent < X.Exponent)
return UnsignedFloatBase::compare(Digits, X.Digits, X.Exponent - Exponent);
return -UnsignedFloatBase::compare(X.Digits, Digits, Exponent - X.Exponent);
}
template <class T> struct isPodLike<UnsignedFloat<T>> {
static const bool value = true;
};
}
//===----------------------------------------------------------------------===//
//
// BlockMass definition.
//
// TODO: Make this private to BlockFrequencyInfoImpl or delete.
//
//===----------------------------------------------------------------------===//
namespace llvm {
/// \brief Mass of a block.
///
/// This class implements a sort of fixed-point fraction always between 0.0 and
/// 1.0. getMass() == UINT64_MAX indicates a value of 1.0.
///
/// Masses can be added and subtracted. Simple saturation arithmetic is used,
/// so arithmetic operations never overflow or underflow.
///
/// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
/// an inexpensive floating-point algorithm that's off-by-one (almost, but not
/// quite, maximum precision).
///
/// Masses can be scaled by \a BranchProbability at maximum precision.
class BlockMass {
uint64_t Mass;
public:
BlockMass() : Mass(0) {}
explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
static BlockMass getEmpty() { return BlockMass(); }
static BlockMass getFull() { return BlockMass(UINT64_MAX); }
uint64_t getMass() const { return Mass; }
bool isFull() const { return Mass == UINT64_MAX; }
bool isEmpty() const { return !Mass; }
bool operator!() const { return isEmpty(); }
/// \brief Add another mass.
///
/// Adds another mass, saturating at \a isFull() rather than overflowing.
BlockMass &operator+=(const BlockMass &X) {
uint64_t Sum = Mass + X.Mass;
Mass = Sum < Mass ? UINT64_MAX : Sum;
return *this;
}
/// \brief Subtract another mass.
///
/// Subtracts another mass, saturating at \a isEmpty() rather than
/// undeflowing.
BlockMass &operator-=(const BlockMass &X) {
uint64_t Diff = Mass - X.Mass;
Mass = Diff > Mass ? 0 : Diff;
return *this;
}
/// \brief Scale by another mass.
///
/// The current implementation is a little imprecise, but it's relatively
/// fast, never overflows, and maintains the property that 1.0*1.0==1.0
/// (where isFull represents the number 1.0). It's an approximation of
/// 128-bit multiply that gets right-shifted by 64-bits.
///
/// For a given digit size, multiplying two-digit numbers looks like:
///
/// U1 . L1
/// * U2 . L2
/// ============
/// 0 . . L1*L2
/// + 0 . U1*L2 . 0 // (shift left once by a digit-size)
/// + 0 . U2*L1 . 0 // (shift left once by a digit-size)
/// + U1*L2 . 0 . 0 // (shift left twice by a digit-size)
///
/// BlockMass has 64-bit numbers. Split each into two 32-bit digits, stored
/// 64-bit. Add 1 to the lower digits, to model isFull as 1.0; this won't
/// overflow, since we have 64-bit storage for each digit.
///
/// To do this accurately, (a) multiply into two 64-bit digits, incrementing
/// the upper digit on overflows of the lower digit (carry), (b) subtract 1
/// from the lower digit, decrementing the upper digit on underflow (carry),
/// and (c) truncate the lower digit. For the 1.0*1.0 case, the upper digit
/// will be 0 at the end of step (a), and then will underflow back to isFull
/// (1.0) in step (b).
///
/// Instead, the implementation does something a little faster with a small
/// loss of accuracy: ignore the lower 64-bit digit entirely. The loss of
/// accuracy is small, since the sum of the unmodelled carries is 0 or 1
/// (i.e., step (a) will overflow at most once, and step (b) will underflow
/// only if step (a) overflows).
///
/// This is the formula we're calculating:
///
/// U1.L1 * U2.L2 == U1 * U2 + (U1 * (L2+1))>>32 + (U2 * (L1+1))>>32
///
/// As a demonstration of 1.0*1.0, consider two 4-bit numbers that are both
/// full (1111).
///
/// U1.L1 * U2.L2 == U1 * U2 + (U1 * (L2+1))>>2 + (U2 * (L1+1))>>2
/// 11.11 * 11.11 == 11 * 11 + (11 * (11+1))/4 + (11 * (11+1))/4
/// == 1001 + (11 * 100)/4 + (11 * 100)/4
/// == 1001 + 1100/4 + 1100/4
/// == 1001 + 0011 + 0011
/// == 1111
BlockMass &operator*=(const BlockMass &X) {
uint64_t U1 = Mass >> 32, L1 = Mass & UINT32_MAX, U2 = X.Mass >> 32,
L2 = X.Mass & UINT32_MAX;
Mass = U1 * U2 + (U1 * (L2 + 1) >> 32) + ((L1 + 1) * U2 >> 32);
return *this;
}
/// \brief Multiply by a branch probability.
///
/// Multiply by P. Guarantees full precision.
///
/// This could be naively implemented by multiplying by the numerator and
/// dividing by the denominator, but in what order? Multiplying first can
/// overflow, while dividing first will lose precision (potentially, changing
/// a non-zero mass to zero).
///
/// The implementation mixes the two methods. Since \a BranchProbability
/// uses 32-bits and \a BlockMass 64-bits, shift the mass as far to the left
/// as there is room, then divide by the denominator to get a quotient.
/// Multiplying by the numerator and right shifting gives a first
/// approximation.
///
/// Calculate the error in this first approximation by calculating the
/// opposite mass (multiply by the opposite numerator and shift) and
/// subtracting both from teh original mass.
///
/// Add to the first approximation the correct fraction of this error value.
/// This time, multiply first and then divide, since there is no danger of
/// overflow.
///
/// \pre P represents a fraction between 0.0 and 1.0.
BlockMass &operator*=(const BranchProbability &P);
bool operator==(const BlockMass &X) const { return Mass == X.Mass; }
bool operator!=(const BlockMass &X) const { return Mass != X.Mass; }
bool operator<=(const BlockMass &X) const { return Mass <= X.Mass; }
bool operator>=(const BlockMass &X) const { return Mass >= X.Mass; }
bool operator<(const BlockMass &X) const { return Mass < X.Mass; }
bool operator>(const BlockMass &X) const { return Mass > X.Mass; }
/// \brief Convert to floating point.
///
/// Convert to a float. \a isFull() gives 1.0, while \a isEmpty() gives
/// slightly above 0.0.
UnsignedFloat<uint64_t> toFloat() const;
void dump() const;
raw_ostream &print(raw_ostream &OS) const;
};
inline BlockMass operator+(const BlockMass &L, const BlockMass &R) {
return BlockMass(L) += R;
}
inline BlockMass operator-(const BlockMass &L, const BlockMass &R) {
return BlockMass(L) -= R;
}
inline BlockMass operator*(const BlockMass &L, const BlockMass &R) {
return BlockMass(L) *= R;
}
inline BlockMass operator*(const BlockMass &L, const BranchProbability &R) {
return BlockMass(L) *= R;
}
inline BlockMass operator*(const BranchProbability &L, const BlockMass &R) {
return BlockMass(R) *= L;
}
inline raw_ostream &operator<<(raw_ostream &OS, const BlockMass &X) {
return X.print(OS);
}
template <> struct isPodLike<BlockMass> {
static const bool value = true;
};
}
//===----------------------------------------------------------------------===//
//
// BlockFrequencyInfoImpl definition.
//
//===----------------------------------------------------------------------===//
namespace llvm {
class BasicBlock;
class BranchProbabilityInfo;
class Function;
class Loop;
class LoopInfo;
class MachineBasicBlock;
class MachineBranchProbabilityInfo;
class MachineFunction;
class MachineLoop;
class MachineLoopInfo;
/// \brief Base class for BlockFrequencyInfoImpl
///
/// BlockFrequencyInfoImplBase has supporting data structures and some
/// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
/// the block type (or that call such algorithms) are skipped here.
///
/// Nevertheless, the majority of the overall algorithm documention lives with
/// BlockFrequencyInfoImpl. See there for details.
class BlockFrequencyInfoImplBase {
public:
typedef UnsignedFloat<uint64_t> Float;
/// \brief Representative of a block.
///
/// This is a simple wrapper around an index into the reverse-post-order
/// traversal of the blocks.
///
/// Unlike a block pointer, its order has meaning (location in the
/// topological sort) and it's class is the same regardless of block type.
struct BlockNode {
typedef uint32_t IndexType;
IndexType Index;
bool operator==(const BlockNode &X) const { return Index == X.Index; }
bool operator!=(const BlockNode &X) const { return Index != X.Index; }
bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
bool operator<(const BlockNode &X) const { return Index < X.Index; }
bool operator>(const BlockNode &X) const { return Index > X.Index; }
BlockNode() : Index(UINT32_MAX) {}
BlockNode(IndexType Index) : Index(Index) {}
bool isValid() const { return Index <= getMaxIndex(); }
static size_t getMaxIndex() { return UINT32_MAX - 1; }
};
/// \brief Stats about a block itself.
struct FrequencyData {
Float Floating;
uint64_t Integer;
};
/// \brief Data about a loop.
///
/// Contains the data necessary to represent represent a loop as a
/// pseudo-node once it's packaged.
struct LoopData {
typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap;
typedef SmallVector<BlockNode, 4> MemberList;
BlockNode Header; ///< Header.
bool IsPackaged; ///< Whether this has been packaged.
ExitMap Exits; ///< Successor edges (and weights).
MemberList Members; ///< Members of the loop.
BlockMass BackedgeMass; ///< Mass returned to loop header.
BlockMass Mass;
Float Scale;
LoopData(const BlockNode &Header) : Header(Header), IsPackaged(false) {}
};
/// \brief Index of loop information.
struct WorkingData {
LoopData *Loop; ///< The loop this block is the header of.
LoopData *ContainingLoop; ///< The block whose loop this block is inside.
BlockMass Mass; ///< Mass distribution from the entry block.
WorkingData() : Loop(nullptr), ContainingLoop(nullptr) {}
bool hasLoopHeader() const { return ContainingLoop; }
bool isLoopHeader() const { return Loop; }
BlockNode getContainingHeader() const {
if (ContainingLoop)
return ContainingLoop->Header;
return BlockNode();
}
/// \brief Has ContainingLoop been packaged up?
bool isPackaged() const {
return ContainingLoop && ContainingLoop->IsPackaged;
}
/// \brief Has Loop been packaged up?
bool isAPackage() const { return Loop && Loop->IsPackaged; }
};
/// \brief Unscaled probability weight.
///
/// Probability weight for an edge in the graph (including the
/// successor/target node).
///
/// All edges in the original function are 32-bit. However, exit edges from
/// loop packages are taken from 64-bit exit masses, so we need 64-bits of
/// space in general.
///
/// In addition to the raw weight amount, Weight stores the type of the edge
/// in the current context (i.e., the context of the loop being processed).
/// Is this a local edge within the loop, an exit from the loop, or a
/// backedge to the loop header?
struct Weight {
enum DistType { Local, Exit, Backedge };
DistType Type;
BlockNode TargetNode;
uint64_t Amount;
Weight() : Type(Local), Amount(0) {}
};
/// \brief Distribution of unscaled probability weight.
///
/// Distribution of unscaled probability weight to a set of successors.
///
/// This class collates the successor edge weights for later processing.
///
/// \a DidOverflow indicates whether \a Total did overflow while adding to
/// the distribution. It should never overflow twice. There's no flag for
/// whether \a ForwardTotal overflows, since when \a Total exceeds 32-bits
/// they both get re-computed during \a normalize().
struct Distribution {
typedef SmallVector<Weight, 4> WeightList;
WeightList Weights; ///< Individual successor weights.
uint64_t Total; ///< Sum of all weights.
bool DidOverflow; ///< Whether \a Total did overflow.
uint32_t ForwardTotal; ///< Total excluding backedges.
Distribution() : Total(0), DidOverflow(false), ForwardTotal(0) {}
void addLocal(const BlockNode &Node, uint64_t Amount) {
add(Node, Amount, Weight::Local);
}
void addExit(const BlockNode &Node, uint64_t Amount) {
add(Node, Amount, Weight::Exit);
}
void addBackedge(const BlockNode &Node, uint64_t Amount) {
add(Node, Amount, Weight::Backedge);
}
/// \brief Normalize the distribution.
///
/// Combines multiple edges to the same \a Weight::TargetNode and scales
/// down so that \a Total fits into 32-bits.
///
/// This is linear in the size of \a Weights. For the vast majority of
/// cases, adjacent edge weights are combined by sorting WeightList and
/// combining adjacent weights. However, for very large edge lists an
/// auxiliary hash table is used.
void normalize();
private:
void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
};
/// \brief Data about each block. This is used downstream.
std::vector<FrequencyData> Freqs;
/// \brief Loop data: see initializeLoops().
std::vector<WorkingData> Working;
/// \brief Indexed information about loops.
std::vector<std::unique_ptr<LoopData>> Loops;
/// \brief Add all edges out of a packaged loop to the distribution.
///
/// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
/// successor edge.
void addLoopSuccessorsToDist(const BlockNode &LoopHead,
const BlockNode &LocalLoopHead,
Distribution &Dist);
/// \brief Add an edge to the distribution.
///
/// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
/// edge is forward/exit/backedge is in the context of LoopHead. Otherwise,
/// every edge should be a forward edge (since all the loops are packaged
/// up).
void addToDist(Distribution &Dist, const BlockNode &LoopHead,
const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
LoopData &getLoopPackage(const BlockNode &Head) {
assert(Head.Index < Working.size());
assert(Working[Head.Index].Loop != nullptr);
return *Working[Head.Index].Loop;
}
/// \brief Distribute mass according to a distribution.
///
/// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
/// backedges and exits are stored in its entry in Loops.
///
/// Mass is distributed in parallel from two copies of the source mass.
///
/// The first mass (forward) represents the distribution of mass through the
/// local DAG. This distribution should lose mass at loop exits and ignore
/// backedges.
///
/// The second mass (general) represents the behavior of the loop in the
/// global context. In a given distribution from the head, how much mass
/// exits, and to where? How much mass returns to the loop head?
///
/// The forward mass should be split up between local successors and exits,
/// but only actually distributed to the local successors. The general mass
/// should be split up between all three types of successors, but distributed
/// only to exits and backedges.
void distributeMass(const BlockNode &Source, const BlockNode &LoopHead,
Distribution &Dist);
/// \brief Compute the loop scale for a loop.
void computeLoopScale(const BlockNode &LoopHead);
/// \brief Package up a loop.
void packageLoop(const BlockNode &LoopHead);
/// \brief Finalize frequency metrics.
///
/// Unwraps loop packages, calculates final frequencies, and cleans up
/// no-longer-needed data structures.
void finalizeMetrics();
/// \brief Clear all memory.
void clear();
virtual std::string getBlockName(const BlockNode &Node) const;
virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
void dump() const { print(dbgs()); }
Float getFloatingBlockFreq(const BlockNode &Node) const;
BlockFrequency getBlockFreq(const BlockNode &Node) const;
raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
raw_ostream &printBlockFreq(raw_ostream &OS,
const BlockFrequency &Freq) const;
uint64_t getEntryFreq() const {
assert(!Freqs.empty());
return Freqs[0].Integer;
}
/// \brief Virtual destructor.
///
/// Need a virtual destructor to mask the compiler warning about
/// getBlockName().
virtual ~BlockFrequencyInfoImplBase() {}
};
namespace bfi_detail {
template <class BlockT> struct TypeMap {};
template <> struct TypeMap<BasicBlock> {
typedef BasicBlock BlockT;
typedef Function FunctionT;
typedef BranchProbabilityInfo BranchProbabilityInfoT;
typedef Loop LoopT;
typedef LoopInfo LoopInfoT;
};
template <> struct TypeMap<MachineBasicBlock> {
typedef MachineBasicBlock BlockT;
typedef MachineFunction FunctionT;
typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
typedef MachineLoop LoopT;
typedef MachineLoopInfo LoopInfoT;
};
/// \brief Get the name of a MachineBasicBlock.
///
/// Get the name of a MachineBasicBlock. It's templated so that including from
/// CodeGen is unnecessary (that would be a layering issue).
///
/// This is used mainly for debug output. The name is similar to
/// MachineBasicBlock::getFullName(), but skips the name of the function.
template <class BlockT> std::string getBlockName(const BlockT *BB) {
assert(BB && "Unexpected nullptr");
auto MachineName = "BB" + Twine(BB->getNumber());
if (BB->getBasicBlock())
return (MachineName + "[" + BB->getName() + "]").str();
return MachineName.str();
}
/// \brief Get the name of a BasicBlock.
template <> inline std::string getBlockName(const BasicBlock *BB) {
assert(BB && "Unexpected nullptr");
return BB->getName().str();
}
}
/// \brief Shared implementation for block frequency analysis.
///
/// This is a shared implementation of BlockFrequencyInfo and
/// MachineBlockFrequencyInfo, and calculates the relative frequencies of
/// blocks.
///
/// This algorithm leverages BlockMass and UnsignedFloat to maintain precision,
/// separates mass distribution from loop scaling, and dithers to eliminate
/// probability mass loss.
///
/// The implementation is split between BlockFrequencyInfoImpl, which knows the
/// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
/// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
/// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
/// reverse-post order. This gives two advantages: it's easy to compare the
/// relative ordering of two nodes, and maps keyed on BlockT can be represented
/// by vectors.
///
/// This algorithm is O(V+E), unless there is irreducible control flow, in
/// which case it's O(V*E) in the worst case.
///
/// These are the main stages:
///
/// 0. Reverse post-order traversal (\a initializeRPOT()).
///
/// Run a single post-order traversal and save it (in reverse) in RPOT.
/// All other stages make use of this ordering. Save a lookup from BlockT
/// to BlockNode (the index into RPOT) in Nodes.
///
/// 1. Loop indexing (\a initializeLoops()).
///
/// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
/// the algorithm. In particular, store the immediate members of each loop
/// in reverse post-order.
///
/// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
///
/// For each loop (bottom-up), distribute mass through the DAG resulting
/// from ignoring backedges and treating sub-loops as a single pseudo-node.
/// Track the backedge mass distributed to the loop header, and use it to
/// calculate the loop scale (number of loop iterations).
///
/// Visiting loops bottom-up is a post-order traversal of loop headers.
/// For each loop, immediate members that represent sub-loops will already
/// have been visited and packaged into a pseudo-node.
///
/// Distributing mass in a loop is a reverse-post-order traversal through
/// the loop. Start by assigning full mass to the Loop header. For each
/// node in the loop:
///
/// - Fetch and categorize the weight distribution for its successors.
/// If this is a packaged-subloop, the weight distribution is stored
/// in \a LoopData::Exits. Otherwise, fetch it from
/// BranchProbabilityInfo.
///
/// - Each successor is categorized as \a Weight::Local, a normal
/// forward edge within the current loop, \a Weight::Backedge, a
/// backedge to the loop header, or \a Weight::Exit, any successor
/// outside the loop. The weight, the successor, and its category
/// are stored in \a Distribution. There can be multiple edges to
/// each successor.
///
/// - Normalize the distribution: scale weights down so that their sum
/// is 32-bits, and coalesce multiple edges to the same node.
///
/// - Distribute the mass accordingly, dithering to minimize mass loss,
/// as described in \a distributeMass(). Mass is distributed in
/// parallel in two ways: forward, and general. Local successors
/// take their mass from the forward mass, while exit and backedge
/// successors take their mass from the general mass. Additionally,
/// exit edges use up (ignored) mass from the forward mass, and local
/// edges use up (ignored) mass from the general distribution.
///
/// Finally, calculate the loop scale from the accumulated backedge mass.
///
/// 3. Distribute mass in the function (\a computeMassInFunction()).
///
/// Finally, distribute mass through the DAG resulting from packaging all
/// loops in the function. This uses the same algorithm as distributing
/// mass in a loop, except that there are no exit or backedge edges.
///
/// 4. Loop unpackaging and cleanup (\a finalizeMetrics()).
///
/// Initialize the frequency to a floating point representation of its
/// mass.
///
/// Visit loops top-down (reverse post-order), scaling the loop header's
/// frequency by its psuedo-node's mass and loop scale. Keep track of the
/// minimum and maximum final frequencies.
///
/// Using the min and max frequencies as a guide, translate floating point
/// frequencies to an appropriate range in uint64_t.
///
/// It has some known flaws.
///
/// - Irreducible control flow isn't modelled correctly. In particular,
/// LoopInfo and MachineLoopInfo ignore irreducible backedges. The main
/// result is that irreducible SCCs will under-scaled. No mass is lost,
/// but the computed branch weights for the loop pseudo-node will be
/// incorrect.
///
/// Modelling irreducible control flow exactly involves setting up and
/// solving a group of infinite geometric series. Such precision is
/// unlikely to be worthwhile, since most of our algorithms give up on
/// irreducible control flow anyway.
///
/// Nevertheless, we might find that we need to get closer. If
/// LoopInfo/MachineLoopInfo flags loops with irreducible control flow
/// (and/or the function as a whole), we can find the SCCs, compute an
/// approximate exit frequency for the SCC as a whole, and scale up
/// accordingly.
///
/// - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting
/// BlockFrequency's 64-bit integer precision.
template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
BranchProbabilityInfoT;
typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
typedef GraphTraits<const BlockT *> Successor;
typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
const BranchProbabilityInfoT *BPI;
const LoopInfoT *LI;
const FunctionT *F;
// All blocks in reverse postorder.
std::vector<const BlockT *> RPOT;
DenseMap<const BlockT *, BlockNode> Nodes;
blockfreq: Rewrite BlockFrequencyInfoImpl Rewrite the shared implementation of BlockFrequencyInfo and MachineBlockFrequencyInfo entirely. The old implementation had a fundamental flaw: precision losses from nested loops (or very wide branches) compounded past loop exits (and convergence points). The @nested_loops testcase at the end of test/Analysis/BlockFrequencyAnalysis/basic.ll is motivating. This function has three nested loops, with branch weights in the loop headers of 1:4000 (exit:continue). The old analysis gives non-sensical results: Printing analysis 'Block Frequency Analysis' for function 'nested_loops': ---- Block Freqs ---- entry = 1.0 for.cond1.preheader = 1.00103 for.cond4.preheader = 5.5222 for.body6 = 18095.19995 for.inc8 = 4.52264 for.inc11 = 0.00109 for.end13 = 0.0 The new analysis gives correct results: Printing analysis 'Block Frequency Analysis' for function 'nested_loops': block-frequency-info: nested_loops - entry: float = 1.0, int = 8 - for.cond1.preheader: float = 4001.0, int = 32007 - for.cond4.preheader: float = 16008001.0, int = 128064007 - for.body6: float = 64048012001.0, int = 512384096007 - for.inc8: float = 16008001.0, int = 128064007 - for.inc11: float = 4001.0, int = 32007 - for.end13: float = 1.0, int = 8 Most importantly, the frequency leaving each loop matches the frequency entering it. The new algorithm leverages BlockMass and PositiveFloat to maintain precision, separates "probability mass distribution" from "loop scaling", and uses dithering to eliminate probability mass loss. I have unit tests for these types out of tree, but it was decided in the review to make the classes private to BlockFrequencyInfoImpl, and try to shrink them (or remove them entirely) in follow-up commits. The new algorithm should generally have a complexity advantage over the old. The previous algorithm was quadratic in the worst case. The new algorithm is still worst-case quadratic in the presence of irreducible control flow, but it's linear without it. The key difference between the old algorithm and the new is that control flow within a loop is evaluated separately from control flow outside, limiting propagation of precision problems and allowing loop scale to be calculated independently of mass distribution. Loops are visited bottom-up, their loop scales are calculated, and they are replaced by pseudo-nodes. Mass is then distributed through the function, which is now a DAG. Finally, loops are revisited top-down to multiply through the loop scales and the masses distributed to pseudo nodes. There are some remaining flaws. - Irreducible control flow isn't modelled correctly. LoopInfo and MachineLoopInfo ignore irreducible edges, so this algorithm will fail to scale accordingly. There's a note in the class documentation about how to get closer. See also the comments in test/Analysis/BlockFrequencyInfo/irreducible.ll. - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting the 64-bit integer precision used downstream. - The "bias" calculation proposed on llvmdev is *not* incorporated here. This will be added in a follow-up commit, once comments from this review have been handled. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206548 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 01:57:45 +00:00
typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
rpot_iterator rpot_begin() const { return RPOT.begin(); }
rpot_iterator rpot_end() const { return RPOT.end(); }
size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
blockfreq: Rewrite BlockFrequencyInfoImpl Rewrite the shared implementation of BlockFrequencyInfo and MachineBlockFrequencyInfo entirely. The old implementation had a fundamental flaw: precision losses from nested loops (or very wide branches) compounded past loop exits (and convergence points). The @nested_loops testcase at the end of test/Analysis/BlockFrequencyAnalysis/basic.ll is motivating. This function has three nested loops, with branch weights in the loop headers of 1:4000 (exit:continue). The old analysis gives non-sensical results: Printing analysis 'Block Frequency Analysis' for function 'nested_loops': ---- Block Freqs ---- entry = 1.0 for.cond1.preheader = 1.00103 for.cond4.preheader = 5.5222 for.body6 = 18095.19995 for.inc8 = 4.52264 for.inc11 = 0.00109 for.end13 = 0.0 The new analysis gives correct results: Printing analysis 'Block Frequency Analysis' for function 'nested_loops': block-frequency-info: nested_loops - entry: float = 1.0, int = 8 - for.cond1.preheader: float = 4001.0, int = 32007 - for.cond4.preheader: float = 16008001.0, int = 128064007 - for.body6: float = 64048012001.0, int = 512384096007 - for.inc8: float = 16008001.0, int = 128064007 - for.inc11: float = 4001.0, int = 32007 - for.end13: float = 1.0, int = 8 Most importantly, the frequency leaving each loop matches the frequency entering it. The new algorithm leverages BlockMass and PositiveFloat to maintain precision, separates "probability mass distribution" from "loop scaling", and uses dithering to eliminate probability mass loss. I have unit tests for these types out of tree, but it was decided in the review to make the classes private to BlockFrequencyInfoImpl, and try to shrink them (or remove them entirely) in follow-up commits. The new algorithm should generally have a complexity advantage over the old. The previous algorithm was quadratic in the worst case. The new algorithm is still worst-case quadratic in the presence of irreducible control flow, but it's linear without it. The key difference between the old algorithm and the new is that control flow within a loop is evaluated separately from control flow outside, limiting propagation of precision problems and allowing loop scale to be calculated independently of mass distribution. Loops are visited bottom-up, their loop scales are calculated, and they are replaced by pseudo-nodes. Mass is then distributed through the function, which is now a DAG. Finally, loops are revisited top-down to multiply through the loop scales and the masses distributed to pseudo nodes. There are some remaining flaws. - Irreducible control flow isn't modelled correctly. LoopInfo and MachineLoopInfo ignore irreducible edges, so this algorithm will fail to scale accordingly. There's a note in the class documentation about how to get closer. See also the comments in test/Analysis/BlockFrequencyInfo/irreducible.ll. - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting the 64-bit integer precision used downstream. - The "bias" calculation proposed on llvmdev is *not* incorporated here. This will be added in a follow-up commit, once comments from this review have been handled. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206548 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-18 01:57:45 +00:00
BlockNode getNode(const rpot_iterator &I) const {
return BlockNode(getIndex(I));
}
BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
const BlockT *getBlock(const BlockNode &Node) const {
assert(Node.Index < RPOT.size());
return RPOT[Node.Index];
}
void initializeRPOT();
void initializeLoops();
void runOnFunction(const FunctionT *F);
void propagateMassToSuccessors(const BlockNode &LoopHead,
const BlockNode &Node);
void computeMassInLoops();
void computeMassInLoop(const BlockNode &LoopHead);
void computeMassInFunction();
std::string getBlockName(const BlockNode &Node) const override {
return bfi_detail::getBlockName(getBlock(Node));
}
public:
const FunctionT *getFunction() const { return F; }
void doFunction(const FunctionT *F, const BranchProbabilityInfoT *BPI,
const LoopInfoT *LI);
BlockFrequencyInfoImpl() : BPI(0), LI(0), F(0) {}
using BlockFrequencyInfoImplBase::getEntryFreq;
BlockFrequency getBlockFreq(const BlockT *BB) const {
return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
}
Float getFloatingBlockFreq(const BlockT *BB) const {
return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
}
/// \brief Print the frequencies for the current function.
///
/// Prints the frequencies for the blocks in the current function.
///
/// Blocks are printed in the natural iteration order of the function, rather
/// than reverse post-order. This provides two advantages: writing -analyze
/// tests is easier (since blocks come out in source order), and even
/// unreachable blocks are printed.
///
/// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
/// we need to override it here.
raw_ostream &print(raw_ostream &OS) const override;
using BlockFrequencyInfoImplBase::dump;
using BlockFrequencyInfoImplBase::printBlockFreq;
raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
}
};
template <class BT>
void BlockFrequencyInfoImpl<BT>::doFunction(const FunctionT *F,
const BranchProbabilityInfoT *BPI,
const LoopInfoT *LI) {
// Save the parameters.
this->BPI = BPI;
this->LI = LI;
this->F = F;
// Clean up left-over data structures.
BlockFrequencyInfoImplBase::clear();
RPOT.clear();
Nodes.clear();
// Initialize.
DEBUG(dbgs() << "\nblock-frequency: " << F->getName() << "\n================="
<< std::string(F->getName().size(), '=') << "\n");
initializeRPOT();
initializeLoops();
// Visit loops in post-order to find thelocal mass distribution, and then do
// the full function.
computeMassInLoops();
computeMassInFunction();
finalizeMetrics();
}
template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
const BlockT *Entry = F->begin();
RPOT.reserve(F->size());
std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
std::reverse(RPOT.begin(), RPOT.end());
assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
"More nodes in function than Block Frequency Info supports");
DEBUG(dbgs() << "reverse-post-order-traversal\n");
for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
BlockNode Node = getNode(I);
DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
Nodes[*I] = Node;
}
Working.resize(RPOT.size());
Freqs.resize(RPOT.size());
}
template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
DEBUG(dbgs() << "loop-detection\n");
if (LI->empty())
return;
// Visit loops top down and assign them an index.
std::deque<const LoopT *> Q;
Q.insert(Q.end(), LI->begin(), LI->end());
while (!Q.empty()) {
const LoopT *Loop = Q.front();
Q.pop_front();
Q.insert(Q.end(), Loop->begin(), Loop->end());
// Save the order this loop was visited.
BlockNode Header = getNode(Loop->getHeader());
assert(Header.isValid());
Loops.emplace_back(new LoopData(Header));
Working[Header.Index].Loop = Loops.back().get();
DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
}
// Visit nodes in reverse post-order and add them to their deepest containing
// loop.
for (size_t Index = 0; Index < RPOT.size(); ++Index) {
const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
if (!Loop)
continue;
// If this is a loop header, find its parent loop (if any).
if (Working[Index].isLoopHeader())
if (!(Loop = Loop->getParentLoop()))
continue;
// Add this node to its containing loop's member list.
BlockNode Header = getNode(Loop->getHeader());
assert(Header.isValid());
const auto &HeaderData = Working[Header.Index];
assert(HeaderData.isLoopHeader());
Working[Index].ContainingLoop = HeaderData.Loop;
HeaderData.Loop->Members.push_back(Index);
DEBUG(dbgs() << " - loop = " << getBlockName(Header)
<< ": member = " << getBlockName(Index) << "\n");
}
}
template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
// Visit loops with the deepest first, and the top-level loops last.
for (const auto &L : make_range(Loops.rbegin(), Loops.rend()))
computeMassInLoop(L->Header);
}
template <class BT>
void BlockFrequencyInfoImpl<BT>::computeMassInLoop(const BlockNode &LoopHead) {
// Compute mass in loop.
DEBUG(dbgs() << "compute-mass-in-loop: " << getBlockName(LoopHead) << "\n");
Working[LoopHead.Index].Mass = BlockMass::getFull();
propagateMassToSuccessors(LoopHead, LoopHead);
for (const BlockNode &M : getLoopPackage(LoopHead).Members)
propagateMassToSuccessors(LoopHead, M);
computeLoopScale(LoopHead);
packageLoop(LoopHead);
}
template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
// Compute mass in function.
DEBUG(dbgs() << "compute-mass-in-function\n");
assert(!Working.empty() && "no blocks in function");
assert(!Working[0].isLoopHeader() && "entry block is a loop header");
Working[0].Mass = BlockMass::getFull();
for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
// Check for nodes that have been packaged.
BlockNode Node = getNode(I);
if (Working[Node.Index].hasLoopHeader())
continue;
propagateMassToSuccessors(BlockNode(), Node);
}
}
template <class BT>
void
BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(const BlockNode &LoopHead,
const BlockNode &Node) {
DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
// Calculate probability for successors.
Distribution Dist;
if (Node != LoopHead && Working[Node.Index].isLoopHeader())
addLoopSuccessorsToDist(LoopHead, Node, Dist);
else {
const BlockT *BB = getBlock(Node);
for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
SI != SE; ++SI)
// Do not dereference SI, or getEdgeWeight() is linear in the number of
// successors.
addToDist(Dist, LoopHead, Node, getNode(*SI), BPI->getEdgeWeight(BB, SI));
}
// Distribute mass to successors, saving exit and backedge data in the
// loop header.
distributeMass(Node, LoopHead, Dist);
}
template <class BT>
raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
if (!F)
return OS;
OS << "block-frequency-info: " << F->getName() << "\n";
for (const BlockT &BB : *F)
OS << " - " << bfi_detail::getBlockName(&BB)
<< ": float = " << getFloatingBlockFreq(&BB)
<< ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
// Add an extra newline for readability.
OS << "\n";
return OS;
}
}
#undef DEBUG_TYPE
#endif