#ifndef FOLLY_EXPERIMENTAL_ELIAS_FANO_CODING_H
#define FOLLY_EXPERIMENTAL_ELIAS_FANO_CODING_H
-#ifndef __GNUC__
-#error EliasFanoCoding.h requires GCC
-#endif
-
-#if !FOLLY_X64
-#error EliasFanoCoding.h requires x86_64
-#endif
-
#include <cstdlib>
#include <limits>
#include <type_traits>
-#include <boost/noncopyable.hpp>
#include <glog/logging.h>
#include <folly/Bits.h>
#include <folly/CpuId.h>
#include <folly/Likely.h>
+#include <folly/Portability.h>
#include <folly/Range.h>
+#include <folly/experimental/Instructions.h>
#include <folly/experimental/Select64.h>
+#ifndef __GNUC__
+#error EliasFanoCoding.h requires GCC
+#endif
+
+#if !FOLLY_X64
+#error EliasFanoCoding.h requires x86_64
+#endif
#if __BYTE_ORDER__ != __ORDER_LITTLE_ENDIAN__
#error EliasFanoCoding.h requires little endianness
// Requires: input range (begin, end) is sorted (encoding
// crashes if it's not).
- // WARNING: encode() mallocates lower, upper, skipPointers
- // and forwardPointers. As EliasFanoCompressedList has
- // no ownership of them, you need to call free() explicitly.
+ // WARNING: encode() mallocates EliasFanoCompressedList::data. As
+ // EliasFanoCompressedList has no ownership of it, you need to call
+ // free() explicitly.
template <class RandomAccessIterator>
static EliasFanoCompressedList encode(RandomAccessIterator begin,
RandomAccessIterator end) {
forwardPointers_(reinterpret_cast<SkipValueType*>(
result.forwardPointers)),
result_(result) {
+ memset(result.data.data(), 0, result.data.size());
}
EliasFanoEncoderV2(size_t size, ValueType upperBound)
uint8_t* buf = nullptr;
// WARNING: Current read/write logic assumes that the 7 bytes
// following the last byte of lower and upper sequences are
- // readable (stored value doesn't matter and won't be changed),
- // so we allocate additional 7B, but do not include them in size
+ // readable (stored value doesn't matter and won't be changed), so
+ // we allocate additional 7 bytes, but do not include them in size
// of returned value.
if (size > 0) {
- buf = static_cast<uint8_t*>(calloc(bytes() + 7, 1));
+ buf = static_cast<uint8_t*>(malloc(bytes() + 7));
}
folly::MutableByteRange bufRange(buf, bytes());
return openList(bufRange);
size_t forwardPointers = 0;
};
-// NOTE: It's recommended to compile EF coding with -msse4.2, starting
-// with Nehalem, Intel CPUs support POPCNT instruction and gcc will emit
-// it for __builtin_popcountll intrinsic.
-// But we provide an alternative way for the client code: it can switch to
-// the appropriate version of EliasFanoReader<> in realtime (client should
-// implement this switching logic itself) by specifying instruction set to
-// use explicitly.
-namespace instructions {
-
-struct Default {
- static bool supported(const folly::CpuId& cpuId = {}) {
- return true;
- }
- static inline uint64_t popcount(uint64_t value) {
- return __builtin_popcountll(value);
- }
- static inline int ctz(uint64_t value) {
- DCHECK_GT(value, 0);
- return __builtin_ctzll(value);
- }
- static inline uint64_t blsr(uint64_t value) {
- return value & (value - 1);
- }
-};
-
-struct Nehalem : public Default {
- static bool supported(const folly::CpuId& cpuId = {}) {
- return cpuId.popcnt();
- }
- static inline uint64_t popcount(uint64_t value) {
- // POPCNT is supported starting with Intel Nehalem, AMD K10.
- uint64_t result;
- asm ("popcntq %1, %0" : "=r" (result) : "r" (value));
- return result;
- }
-};
-
-struct Haswell : public Nehalem {
- static bool supported(const folly::CpuId& cpuId = {}) {
- return Nehalem::supported(cpuId) && cpuId.bmi1();
- }
- static inline uint64_t blsr(uint64_t value) {
- // BMI1 is supported starting with Intel Haswell, AMD Piledriver.
- // BLSR combines two instuctions into one and reduces register pressure.
- uint64_t result;
- asm ("blsrq %1, %0" : "=r" (result) : "r" (value));
- return result;
- }
-};
-
-} // namespace instructions
-
namespace detail {
template <class Encoder, class Instructions>
return skipToNext(v);
}
+ ValueType previousValue() const {
+ DCHECK_NE(position(), -1);
+ DCHECK_GT(position(), 0);
+
+ size_t outer = outer_;
+ block_t block = folly::loadUnaligned<block_t>(start_ + outer);
+ block &= (block_t(1) << inner_) - 1;
+
+ while (UNLIKELY(block == 0)) {
+ DCHECK_GE(outer, sizeof(block_t));
+ outer -= sizeof(block_t);
+ block = folly::loadUnaligned<block_t>(start_ + outer);
+ }
+
+ auto inner = 8 * sizeof(block_t) - 1 - Instructions::clz(block);
+ return static_cast<ValueType>(8 * outer + inner - (position_ - 1));
+ }
+
+ void setDone(size_t endPos) {
+ position_ = endPos;
+ }
+
private:
ValueType setValue() {
value_ = static_cast<ValueType>(8 * outer_ + inner_ - position_);
block_ &= ~((block_t(1) << (dest % 8)) - 1);
}
- typedef unsigned long long block_t;
+ typedef uint64_t block_t;
const unsigned char* const forwardPointers_;
const unsigned char* const skipPointers_;
const unsigned char* const start_;
} // namespace detail
+// If kUnchecked = true the caller must guarantee that all the
+// operations return valid elements, i.e., they would never return
+// false if checked.
template <class Encoder,
- class Instructions = instructions::Default>
-class EliasFanoReader : private boost::noncopyable {
+ class Instructions = instructions::Default,
+ bool kUnchecked = false>
+class EliasFanoReader {
public:
typedef Encoder EncoderType;
typedef typename Encoder::ValueType ValueType;
explicit EliasFanoReader(const EliasFanoCompressedList& list)
- : list_(list),
- lowerMask_((ValueType(1) << list_.numLowerBits) - 1),
- upper_(list_) {
+ : size_(list.size),
+ lower_(list.lower),
+ upper_(list),
+ lowerMask_((ValueType(1) << list.numLowerBits) - 1),
+ numLowerBits_(list.numLowerBits) {
DCHECK(Instructions::supported());
// To avoid extra branching during skipTo() while reading
// upper sequence we need to know the last element.
- if (UNLIKELY(list_.size == 0)) {
+ // If kUnchecked == true, we do not check that skipTo() is called
+ // within the bounds, so we can avoid initializing lastValue_.
+ if (kUnchecked || UNLIKELY(list.size == 0)) {
lastValue_ = 0;
return;
}
- ValueType lastUpperValue = 8 * list_.upperSize() - list_.size;
- auto it = list_.upper + list_.upperSize() - 1;
+ ValueType lastUpperValue = 8 * list.upperSize() - size_;
+ auto it = list.upper + list.upperSize() - 1;
DCHECK_NE(*it, 0);
lastUpperValue -= 8 - folly::findLastSet(*it);
- lastValue_ = readLowerPart(list_.size - 1) |
- (lastUpperValue << list_.numLowerBits);
+ lastValue_ = readLowerPart(size_ - 1) | (lastUpperValue << numLowerBits_);
}
void reset() {
upper_.reset();
- progress_ = 0;
value_ = 0;
}
bool next() {
- if (UNLIKELY(progress_ >= list_.size)) {
+ if (!kUnchecked && UNLIKELY(position() + 1 >= size_)) {
return setDone();
}
- value_ = readLowerPart(progress_) |
- (upper_.next() << list_.numLowerBits);
- ++progress_;
+ upper_.next();
+ value_ = readLowerPart(upper_.position()) |
+ (upper_.value() << numLowerBits_);
return true;
}
bool skip(size_t n) {
CHECK_GT(n, 0);
- progress_ += n;
- if (LIKELY(progress_ <= list_.size)) {
+ if (kUnchecked || LIKELY(position() + n < size_)) {
if (LIKELY(n < kLinearScanThreshold)) {
for (size_t i = 0; i < n; ++i) upper_.next();
} else {
upper_.skip(n);
}
- value_ = readLowerPart(progress_ - 1) |
- (upper_.value() << list_.numLowerBits);
+ value_ = readLowerPart(upper_.position()) |
+ (upper_.value() << numLowerBits_);
return true;
}
DCHECK_GE(value, value_);
if (value <= value_) {
return true;
- } else if (value > lastValue_) {
+ } else if (!kUnchecked && value > lastValue_) {
return setDone();
}
- size_t upperValue = (value >> list_.numLowerBits);
+ size_t upperValue = (value >> numLowerBits_);
size_t upperSkip = upperValue - upper_.value();
// The average density of ones in upper bits is 1/2.
// LIKELY here seems to make things worse, even for small skips.
}
bool jump(size_t n) {
- if (LIKELY(n - 1 < list_.size)) { // n > 0 && n <= list_.size
- progress_ = n;
- value_ = readLowerPart(n - 1) | (upper_.jump(n) << list_.numLowerBits);
+ if (LIKELY(n - 1 < size_)) { // n > 0 && n <= size_
+ value_ = readLowerPart(n - 1) | (upper_.jump(n) << numLowerBits_);
return true;
} else if (n == 0) {
reset();
if (value <= 0) {
reset();
return true;
- } else if (value > lastValue_) {
+ } else if (!kUnchecked && value > lastValue_) {
return setDone();
}
- upper_.jumpToNext(value >> list_.numLowerBits);
+ upper_.jumpToNext(value >> numLowerBits_);
iterateTo(value);
return true;
}
- size_t size() const { return list_.size; }
+ ValueType previousValue() const {
+ DCHECK_GT(position(), 0);
+ DCHECK_LT(position(), size());
+ return readLowerPart(upper_.position() - 1) |
+ (upper_.previousValue() << numLowerBits_);
+ }
+
+ size_t size() const { return size_; }
- size_t position() const { return progress_ - 1; }
+ size_t position() const { return upper_.position(); }
ValueType value() const { return value_; }
private:
bool setDone() {
value_ = std::numeric_limits<ValueType>::max();
- progress_ = list_.size + 1;
+ upper_.setDone(size_);
return false;
}
ValueType readLowerPart(size_t i) const {
- DCHECK_LT(i, list_.size);
- const size_t pos = i * list_.numLowerBits;
- const unsigned char* ptr = list_.lower + (pos / 8);
+ DCHECK_LT(i, size_);
+ const size_t pos = i * numLowerBits_;
+ const unsigned char* ptr = lower_ + (pos / 8);
const uint64_t ptrv = folly::loadUnaligned<uint64_t>(ptr);
return lowerMask_ & (ptrv >> (pos % 8));
}
void iterateTo(ValueType value) {
while (true) {
value_ = readLowerPart(upper_.position()) |
- (upper_.value() << list_.numLowerBits);
+ (upper_.value() << numLowerBits_);
if (LIKELY(value_ >= value)) break;
upper_.next();
}
- progress_ = upper_.position() + 1;
}
constexpr static size_t kLinearScanThreshold = 8;
- const EliasFanoCompressedList list_;
- const ValueType lowerMask_;
+ size_t size_;
+ const uint8_t* lower_;
detail::UpperBitsReader<Encoder, Instructions> upper_;
- size_t progress_ = 0;
+ const ValueType lowerMask_;
ValueType value_ = 0;
ValueType lastValue_;
+ uint8_t numLowerBits_;
};
}} // namespaces